A Course in Metric Geometry

Dmitri Burago
Yuri Burago
Sergei Ivanov

Department of Mathematics, Pennsylvania State University

E-mail address: burago@math.psu.edu

Steklov Institute for Mathematics at St. Petersburg

E-mail address: burago@pdmi.ras.ru

Steklov Institute for Mathematics at St. Petersburg

E-mail address: svivanov@pdmi.ras.ru
Contents

Preface vii

Chapter 1. Metric Spaces 1

§1.1. Definitions 1
§1.2. Examples 3
§1.3. Metrics and Topology 7
§1.4. Lipschitz Maps 9
§1.5. Complete Spaces 10
§1.6. Compact Spaces 13
§1.7. Hausdorff Measure and Dimension 17

Chapter 2. Length Spaces 25

§2.1. Length Structures 25
§2.2. First Examples of Length Structures 30
§2.3. Length Structures Induced by Metrics 33
§2.4. Characterization of Intrinsic Metrics 38
§2.5. Shortest Paths 44
§2.6. Length and Hausdorff Measure 53
§2.7. Length and Lipschitz Speed 55

Chapter 3. Constructions 59
§3.1. Locality, Gluing and Maximal Metrics 59
§3.2. Polyhedral Spaces 67
§3.3. Isometries and Quotients 74

iii
iv Contents

§3.4. Local Isometries and Coverings 78

§3.5. Arcwise Isometries 85
§3.6. Products and Cones 87
Chapter 4. Spaces of Bounded Curvature 101
§4.1. Definitions 101
§4.2. Examples 109
§4.3. Angles in Alexandrov Spaces and Equivalence of Definitions 114
§4.4. Analysis of Distance Functions 119
§4.5. The First Variation Formula 121
§4.6. Nonzero Curvature Bounds and Globalization 126
§4.7. Curvature of Cones 131
Chapter 5. Smooth Length Structures 135
§5.1. Riemannian Length Structures 136
§5.2. Exponential Map 150
§5.3. Hyperbolic Plane 154
§5.4. Sub-Riemannian Metric Structures 178
§5.5. Riemannian and Finsler Volumes 193
§5.6. Besikovitch Inequality 202
Chapter 6. Curvature of Riemannian Metrics 209
§6.1. Motivation: Coordinate Computations 211
§6.2. Covariant Derivative 214
§6.3. Geodesic and Gaussian Curvatures 221
§6.4. Geometric Meaning of Gaussian Curvature 226
§6.5. Comparison Theorems 237
Chapter 7. Space of Metric Spaces 241
§7.1. Examples 242
§7.2. Lipschitz Distance 249
§7.3. Gromov–Hausdorff Distance 251
§7.4. Gromov–Hausdorff Convergence 260
§7.5. Convergence of Length Spaces 265
Chapter 8. Large-scale Geometry 271
§8.1. Noncompact Gromov–Hausdorff Limits 271
§8.2. Tangent and Asymptotic Cones 275
Contents v

§8.3. Quasi-isometries 277

§8.4. Gromov Hyperbolic Spaces 284
§8.5. Periodic Metrics 298
Chapter 9. Spaces of Curvature Bounded Above 307
§9.1. Definitions and Local Properties 308
§9.2. Hadamard Spaces 324
§9.3. Fundamental Group of a Nonpositively Curved Space 338
§9.4. Example: Semi-dispersing Billiards 341
Chapter 10. Spaces of Curvature Bounded Below 351
§10.1. One More Definition 352
§10.2. Constructions and Examples 354
§10.3. Toponogov’s Theorem 360
§10.4. Curvature and Diameter 364
§10.5. Splitting Theorem 366
§10.6. Dimension and Volume 369
§10.7. Gromov–Hausdorff Limits 376
§10.8. Local Properties 378
§10.9. Spaces of Directions and Tangent Cones 390
§10.10. Further Information 398
Bibliography 405
Index 409
Preface

This book is not a research monograph or a reference book (although

research interests of the authors influenced it a lot)—this is a textbook.
Its structure is similar to that of a graduate course. A graduate course
usually begins with a course description, and so do we.

Course description. The objective of this book is twofold. First of all, we

wanted to give a detailed exposition of basic notions and techniques in the
theory of length spaces, a theory which experienced a very fast development
in the past few decades and penetrated into many other mathematical disci-
plines (such as Group Theory, Dynamical Systems, and Partial Differential
Equations). However, we have a wider goal of giving an elementary intro-
duction into a broad variety of the most geometrical topics in geometry—the
ones related to the notion of distance. This is the reason why we included
metric introductions to Riemannian and hyperbolic geometries. This book
tends to work with “easy-to-touch” mathematical objects by means of “easy-
to-visualize” methods. There is a remarkable book [Gro3], which gives a
vast panorama of “geometrical mathematics from a metric viewpoint”. Un-
fortunately, Gromov’s book seems hardly accessible to graduate students
and non-experts in geometry. One of the objectives of this book is to bridge
the gap between students and researchers interested in metric geometry, and
modern mathematical literature.

Prerequisite. It is minimal. We set a challenging goal of making the core

part of the book accessible to first-year graduate students. Our expectations
of the reader’s background gradually grow as we move further in the book.
We tried to introduce and illustrate most of new concepts and methods
by using their simplest case and avoiding technicalities that take attention

vii
viii Preface

away from the gist of the matter. For instance, our introduction to Riemann-
ian geometry begins with metrics on planar regions, and we even avoid the
notion of a manifold. Of course, manifolds do show up in more advanced sec-
tions. Some exercises and remarks assume more mathematical background
than the rest of our exposition; they are optional, and a reader unfamiliar
with some notions can just ignore them. For instance, solid background in
differential geometry of curves and surfaces in R3 is not a mandatory prereq-
uisite for this book. However, we would hope that the reader possesses some
knowledge of differential geometry, and from time to time we draw analogies
from or suggest exercises based on it. We also make a special emphasis on
motivations and visualizations. A reader not interested in them will be able
to skip certain sections. The first chapter is a clinic in metric topology; we
recommend that the reader with a reasonable idea of metric spaces just skip
it and use it for reference: it may be boring to read it. The last chapters
are more advanced and dry than the first four.

Figures. There are several figures in the book, which are added just to
make it look nicer. If we included all necessary figures, there would be at
least five of them for each page.
• It is a must that the reader systematically studying this book makes
a figure for every proposition, theorem, and construction!

Exercises. Exercises form a vital part of our exposition. This does not
mean that the reader should solve all the exercises; it is very individual.
The difficulty of exercises varies from trivial to rather tricky, and their
importance goes all the way up from funny examples to statements that
are extensively used later in the book. This is often indicated in the text.
It is a very helpful strategy to perceive every proposition and theorem as an
exercise. You should try to prove each on your own, possibly after having
a brief glance at our argument to get a hint. Just reading our proof is the
last resort.

Optional material. Our exposition can be conditionally subdivided into

two parts: core material and optional sections. Some sections and chapters
are preceded by a brief plan, which can be used as a guide through them.
It is usually a good idea to begin with a first reading, skipping all optional
sections (and even the less important parts of the core ones). Of course, this
approach often requires going back and looking for important notions that
were accidentally missed. A first reading can give a general picture of the
theory, helping to separate its core and give a good idea of its logic. Then
the reader goes through the book again, transforming theoretical knowledge
into the genuine one by filling it with all the details, digressions, examples
and experience that makes knowledge practical.
Preface ix

About metric geometry. Whereas the borderlines between mathemati-

cal disciplines are very conditional, geometry historically began from very
“down-to-earth” notions (even literally). However, for most of the last cen-
tury it was a common belief that “geometry of manifolds” basically boiled
down to “analysis on manifolds”. Geometric methods heavily relied on dif-
ferential machinery, as it can be guessed even from the name “Differential
geometry”. It is now understood that a tremendous part of geometry es-
sentially belongs to metric geometry, and the differential apparatus can be
used just to define some class of objects and extract the starting data to
feed into the synthetic methods. This certainly cannot be applied to all
geometric notions. Even the curvature tensor remains an obscure monster,
and the geometric meaning of only some of its simplest appearances (such
as the sectional curvature) are more or less understood. Many modern re-
sults involving more advanced structures still sound quite analytical. On
the other hand, expelling analytical machinery from a certain sphere of
definitions and arguments brought several major benefits. First of all, it
enhanced mathematical understanding of classical objects (such as smooth
Riemannian manifolds) both ideologically, and by concrete results. From a
methodological viewpoint, it is important to understand what assumptions a
particular result relies on; for instance, in this respect it is more satisfying to
know that geometrical properties of positively curved manifolds are based
on a certain inequality on distances between quadruples of points rather
than on some properties of the curvature tensor. This is very similar to
two ways of thinking about convex functions. One can say that a function
is convex if its second derivative is nonnegative (notice that the definition
already assumes that the function is smooth, leaving out such functions as
f (x) = |x|). An alternative definition says that a function is convex if its
epigraph (the set {(x, y) : y ≥ f (x)}) is; the latter definition is equivalent
to Jensen’s inequality f (αx + βy) ≤ αf (x) + βf (y) for all nonnegative α, β
with α + β = 1, and it is robust and does not rely on the notion of a limit.
From this viewpoint, the condition f ′′ ≥ 0 can be regarded as a convenient
criterion for a smooth function to be convex.
As a more specific illustration of an advantage of this way of thinking,
imagine that one wants to estimate a certain quantity over all metrics
on a sphere. It is so tempting to study a metric for which the quantity
attains its maximum, but alas this metric may fail exist within smooth
metrics, or even metrics that induce the same topology. It turns out that
it still may exist if we widen our search to a class of more general length
spaces. Furthermore, mathematical topics whose study used to lie outside
the range of noticeable applications of geometrical technique now turned
out to be traditional objects of methods originally rooted in differential
geometry. Combinatorial group theory can serve as a model example of this
x Preface

situation. By now the scope of the theory of length spaces has grown quite
far from its cradle (which was a theory of convex surfaces), including most
of classical Riemannian geometry and many areas beyond it. At the same
time, geometry of length spaces perhaps remains one of the most “hands-
on” mathematical techniques. This combination of reasons urged us to write
this “beginners’ course in geometry from a length structure viewpoint”.

Acknowledgements. The authors enjoyed hospitality and excellent work-

ing conditions during their stays at various institutions, including the Uni-
versity of Strasbourg, ETH Zurich, and Cambridge University. These un-
forgettable visits were of tremendous help to the progress of this book. The
authors’ research, which had essential impact on the book, was partially
supported by the NSF Foundation, the Sloan Research Fellowship, CRDF,
RFBR, and Shapiro Fund at Penn State, whose help we gratefully acknowl-
edge. The authors are grateful to many people for their help and encour-
agement. We want to especially thank M. Gromov for provoking us to write
this book; S. Alexander, R. Bishop, and C. Croke for undertaking immense
labor of thoroughly reading the manuscript—their numerous corrections,
suggestions, and remarks were of invaluable help; S. Buyalo for many useful
comments and suggestions for Chapter 9; K. Shemyak for preparing most
of the figures; and finally a group of graduate students at Penn State who
took a Math 597c course using our manuscript as the base text and cor-
rected dozens of typos and small errors (though we are confident that twice
as many of them are still left for the reader).
 Chapter 1

Metric Spaces

The purpose of the major part of the chapter is to set up notation and to
refresh the reader’s knowledge of metric spaces and related topics in point-
set topology. Section 1.7 contains minimal information about Hausdorff
measure and dimension.
It may be a good idea to skip this chapter and use it only for reference,
or to look through it briefly to make sure that all examples are clear and
exercises are obvious.

1.1. Definitions
Definition 1.1.1. Let X be an arbitrary set. A function d : X × X →
R ∪ {∞} is a metric on X if the following conditions are satisfied for all
x, y, z ∈ X.
(1) Positiveness: d(x, y) > 0 if x 6= y, and d(x, x) = 0.
(2) Symmetry: d(x, y) = d(y, x).
(3) Triangle inequality: d(x, z) ≤ d(x, y) + d(y, z).
A metric space is a set with a metric on it. In a formal language, a metric
space is a pair (X, d) where d is a metric on X. Elements of X are called
points of the metric space; d(x, y) is referred to as the distance between
points x and y.
When the metric in question is clear from the context, we also denote
the distance between x and y by |xy|.
Unless different metrics on the same set X are considered, we will omit
an explicit reference to the metric and write “a metric space X” instead of
“a metric space (X, d).”

1
2 1. Metric Spaces

In most textbooks, the notion of a metric space is slightly narrower

than our definition: traditionally one consider metrics with finite distance
between points. If it is important for a particular consideration that d
takes only finite values, this will be specified by saying that d is a finite
metric. There is a very simple relation between finite and infinite metrics,
namely a metric space with possibly infinite distances splits canonically into
subspaces that carry finite metrics and are separated from one another by
infinite distances:
Exercise 1.1.2. Show that the relation d(x, y) 6= ∞ is an equivalence
relation. Each of its equivalence classes together with the restriction of
d is a metric space with a finite metric.
Definition 1.1.3. Let X and Y be two metric spaces. A map f : X → Y is
called distance-preserving if |f (x)f (y)| = |xy| for any two points x, y ∈ X.
A bijective distance-preserving map is called an isometry. Two spaces are
isometric if there exists an isometry from one to the other.

It is clear that being isometric is an equivalence relation. Isometric

spaces share all properties that can be expressed completely in terms of
distances.

Semi-metrics.
Definition 1.1.4. A function d : X × X → R+ ∪ {+∞} is called a semi-
metric if it satisfies all properties from Definition 1.1.1 of a metric except
the requirement that d(x, y) = 0 implies x = y. This means that we allow
zero distance between different points.

There is an obvious relation between semi-metrics and metrics, namely

identifying points with zero distance in a semi-metric leads to a usual metric:
Proposition 1.1.5. Let d be a semi-metric on X. Introduce an equivalence
relation Rd on X: set xRd y iff d(x, y) = 0. Since d(x, y) = d(x1 , y1 )
whenever xRd x1 and yRd y1 , the projection dˆ of d onto the quotient space
ˆ is a metric space.
X̂ = X/Rd is well-defined. Then (X̂, d)

Proof. Trivial (exercise). ¤

ˆ
We will often abuse notation, writing (X/d, d) rather than (X/Rd , d),
ˆ
with X/d instead of X/Rd and using the same letter d for its projection d.
Example 1.1.6. Let the distance between two points (x, y), (x′ , y ′ ) in R2
be defined by d((x, y), (x′ , y ′ )) = |(x − x′ ) + (y − y ′ )|. Check that it is a
semi-metric. Prove that the quotient space (R2 /d, d) is isometric to the real
line.
1.2. Examples 3

1.2. Examples
Various examples of metric spaces will appear everywhere in the course. In
this section we only describe several important ones to begin with. For many
of them, verification of the properties from Definition 1.1.1 is trivial and is
left for the reader.
Example 1.2.1. One can define a metric on an arbitrary set X by
(
0 if x = y,
|xy| =
1 if x 6= y.
This example is not particularly interesting but it can serve as the initial
point for many constructions.
Example 1.2.2. The real line, R, is canonically equipped with the distance
|xy| = |x − y|, and thus can be considered as a metric space. There is an
immense variety of other metrics on R; for instance, consider dlog (x, y) =
log |x − y|.
Example 1.2.3. The Euclidean plane, R2 , with its standard distance,
is another familiar metric space. The distance can be expressed by the
Pythagorean formula,
p
|xy| = |x − y| = (x1 − y1 )2 + (x2 − y2 )2
where (x1 , x2 ) and (y1 , y2 ) are coordinates of points x and y. The triangle
inequality for this metric is known from elementary Euclidean geometry.
Alternatively, it can be derived from the Cauchy inequality.
Example 1.2.4 (direct products). Let X and Y be two metric spaces. We
define a metric on their direct product X × Y by the formula
p
|(x1 , y1 )(x2 , y2 )| = |x1 x2 |2 + |y1 y2 |2 .
In particular, R × R = R2 .
Exercise 1.2.5. Derive the triangle inequality for direct products from the
triangle inequality on the Euclidean plane.
Example 1.2.6. Recall that the coordinate n-space Rn is the vector space
of all n-tuples (x1 , . . . , xn ) of real numbers, with component-wise addition
and multiplication by scalars. It is naturally identified with the multiple
direct product R × · · · × R (n times). This defines the standard Euclidean
distance,
p
|xy| = (x1 − y1 )2 + · · · + (xn − yn )2
where x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ).
4 1. Metric Spaces

Example 1.2.7 (dilated spaces). This simple construction is similar to

obtaining one set from another by means of a homothety map. Let X be a
metric space and λ > 0. The metric space λX is the same set X equipped
with another distance function dλX which is defined by dλX (x, y) = dX (x, y)
for all x, y ∈ X, where dX is the distance in X. The space λX is referred to
as X dilated (or rescaled) by λ.
Example 1.2.8 (subspaces). If X is a metric space and Y is a subset of X,
then a metric on Y can be obtained by simply restricting the metric from
X. In other words, the distance between points of Y is equal to the distance
between the same points in X.

Restricting the distance is the simplest but not the only way to define a
metric on a subset. In many cases it is more natural to consider an intrinsic
metric, which is generally not equal to the one restricted from the ambient
space. The notion of intrinsic metric will be explained further in the course,
but its intuitive meaning can be illustrated by the following example of the
intrinsic metric on a circle.
Example 1.2.9. The unit circle, S 1 , is the set of points in the plane lying at
distance 1 from the origin. Being a subset of the plane, the circle carries the
restricted Euclidean metric on it. We define an alternative metric by setting
the distance between two points as the length of the shorter arc between
them. For example, the arc-length distance between two opposite points of
the circle is equal to π. The distance between adjacent vertices of a regular
n-gon (inscribed into the circle) is equal to 2π/n.
Exercise 1.2.10. (a) Prove that any circle arc of length less or equal to π,
equipped with the above metric, is isometric to a straight line segment.
(b) Prove that the entire circle with this metric is not isometric to any
subset of the plane (regarded with the restriction of Euclidean distance onto
this subset).

1.2.1. Normed vector spaces.

Definition 1.2.11. Let V be a vector space1. A function | · | : V → R is
a norm on V if the following conditions are satisfied for all v, w ∈ V and
k ∈ R.
(1) Positiveness: |v| > 0 if v 6= 0, and |0| = 0.
(2) Positive homogeneity: |kv| = |k||v|.
(3) Subadditivity (triangle inequality): |v + w| ≤ |v| + |w|.

1All normed spaces here are ones over R.

1.2. Examples 5

A normed space is a vector space with a norm on it. Finite-dimensional

normed spaces are also called Minkowski spaces. The distance in a normed
space (V, | · |) is defined by the formula
d(v, w) = |v − w|.

It is easy to see that a normed space with the above distance is a metric
space. The norm is recovered from the metric as the distance from the
origin.
The Euclidean space Rn described in Example 1.2.6 is a normed space
whose norm is expressed by
q
|(x1 , . . . , xn )| = x21 + · · · + x2n .
There are other natural norms in Rn .
Example 1.2.12. The space Rn1 is the coordinate space Rn with a norm
k · k1 defined by
k(x1 , . . . , xn )k1 = |x1 | + · · · + |xn |
(where | · | is just the absolute value of real numbers).
Example 1.2.13. Similarly, the space Rn∞ is Rn with a norm k · k∞ where
k(x1 , . . . , xn )k∞ = max{|x1 |, . . . , |xn |}.
Exercise 1.2.14. Prove that
(a) R21 and R2∞ are isometric;
(b) Rn1 and Rn∞ are not isometric for any n > 2.
Example 1.2.15. Let X be an arbitrary set. The space ℓ∞ (X) is the set
of all bounded functions f : X → R. This is naturally a vector space with
respect to pointwise addition and multiplication by scalars. The standard
norm k · k∞ on ℓ∞ (X) is defined by
kf k∞ = sup |f (x)|.
x∈X

Exercise 1.2.16. Show that Rn∞ = ℓ∞ (X) for a suitable set X. Hint: an
n-tuple (x1 , . . . , xn ) is formally a map, isn’t it?

1.2.2. Euclidean spaces. Let X be a vector space. Recall that a bilinear

form on X is a map F : X × X → R which is linear in both arguments.
A bilinear form F is symmetric if F (x, y) = F (y, x) for all x, y ∈ X. A
symmetric bilinear form F can be recovered from its associated quadratic
form Q(x) = QF (x) = F (x, x), e.g., by means of the formula 4F (x, y) =
Q(x + y) − Q(x − y).
6 1. Metric Spaces

Definition 1.2.17. A scalar product is a symmetric bilinear form F whose

associated quadratic form is positive definite, i.e., F (x, x) > 0 for all x 6= 0.
A Euclidean space is a vector space with a scalar product on it.
We will use notation h·, ·i for various scalar products.
Definition 1.2.18. Apnorm associated with a scalar product h·, ·i is defined
by the formula |v| = hv, vi. A norm is called Euclidean if it is associated
with some scalar product.
norm in Rn is associated with the scalar
For example, the standard P
product defined by hx, yi = xi yi where x = (x1 , . . . , xn ) and y =
(y1 , . . . , yn ).
Exercise 1.2.19. Prove the triangle inequality for a norm associated with
a scalar product.
Hint: First, reduce the triangle inequality to: hv, wi ≤ |v| · |w| for any
two vectors v and w. Then expand the relation hv − tw, v − twi ≥ 0 and
substitute t = hv, vi / hw, wi. Another way to prove the triangle inequality
is to combine Proposition 1.2.22 and the triangle inequality for Rn .
Since a scalar product is uniquely determined by its associated norm, a
Euclidean space could be defined as a normed space whose norm is Euclid-
ean. The following exercise give an explicit characterization of Euclidean
spaces among the normed spaces.
Exercise 1.2.20. Prove that a norm | · | on a vector space V is Euclidean
if and only if
|v + w|2 + |v − w|2 = 2(|v|2 + |w|2 )
for all v, w ∈ V .
Exercise 1.2.21. Show that Rn1 and Rn∞ are not Euclidean spaces for n > 1.
Two vectors in a Euclidean space are called orthogonal if their scalar
product is zero. An orthonormal frame is a collection of mutually orthogonal
unit vectors. Vectors of an orthonormal frame are linearly independent
(prove this!). An orthonormal frame can be obtained from any collection of
linearly independent vectors by a standard Gram–Schmidt orthogonalization
procedure.
In particular, a finite-dimensional Euclidean space V possesses an ortho-
normal basis. Let dim V = n and {e1 , . . . , en } be such a basis.
P Every vector
x ∈ V can be uniquely represented as a linear combination xi ei for some
xi ∈ R. Since all scalar products of vectors ei are known, we can find the
scalar product of any linear combination, namely
DX X E X
xi ei , yi ei = xi yi .
1.3. Metrics and Topology 7

This implies the following

Proposition 1.2.22. Every n-dimensional Euclidean space is isomorphic
to Rn . This means that there is a linear isomorphism f : Rn → V such
that hf (x), f (y)i = hx, yi for all x, y ∈ Rn . In particular, these spaces are
isometric.
P
Proof. Define f ((x1 , . . . , xn )) = xi ei where {ei } is an orthonormal basis.
¤

This proposition allows one to apply elementary Euclidean geometry to

general Euclidean spaces. For example, since any two-dimensional subspace
of a Euclidean space is isomorphic to R2 , any statement involving only two
vectors and their linear combinations can be automatically transferred from
the standard Euclidean plane to all Euclidean spaces.
Exercise 1.2.23. Prove that any distance-preserving map from one Euclid-
ean space to another is an affine map, that is, a composition of a linear map
and a parallel translation. Show by example that this is generally not true
for arbitrary normed spaces.
Exercise 1.2.24. Let V be a finite-dimensional normed space. Prove that
V is Euclidean if and only if for any two vectors v, w ∈ V such that |v| = |w|
there exists a linear isometry f : V → V such that f (v) = w.

1.2.3. Spheres.
Example 1.2.25. The n-sphere S n is the set of unit vectors in Rn+1 , i.e.,
S n = {x ∈ Rn+1 : |x| = 1}. The angular metric on S n is defined by
d(x, y) = arccos hx, yi .
In other words, the spherical distance is defined as the Euclidean angle
between unit vectors. It equals the length of the shorter arc of a great circle
connecting x and y in the sphere. Another formula for this metric is
|x − y|
d(x, y) = 2 arcsin .
2
The metric on the circle described in Example 1.2.9 is a partial case of this
example.

1.3. Metrics and Topology

Definition 1.3.1. Let X be a metric space, x ∈ X and r > 0. The set
formed by the points at distance less than r from x is called an (open metric)
ball of radius r centered at x. We denote this ball by Br (x). Similarly, a
closed ball B r (x) is the set of points whose distances from x are less than or
equal to r.
8 1. Metric Spaces

Exercise 1.3.2. Let x1 and x2 be points of some metric space, and let r1
and r2 be positive numbers. Show that
(a) if |x1 x2 | ≥ r1 + r2 , then the balls Br1 (x1 ) and Br2 (x2 ) are disjoint;
(b) if |x1 x2 | ≤ r1 − r2 , then Br2 (x2 ) ⊂ Br1 (x1 );
(c) the converse statements to (a) and (b) are not always true (give
counterexamples).

The topology associated with a metric is defined as follows: a set U in

the metric space is open if and only if for every point x ∈ U there exists an
ε > 0 such that Bε (r) ⊂ U .
It is easy to see that an open ball is an open set and a closed ball is a
closed set (i.e., its complement is open). As a consequence of the former, a
set is open if and only if is representable as a union of (possibly infinitely
many) open balls.

Exercise 1.3.3. Let X be a metric space and Y ⊂ X. Prove that two

topologies on Y coincide: the one associated with the metric restricted on
Y , and the subspace topology induced by the one of X (in which a set is
open in Y if and only if it is representable as an intersection of Y and an
open set in X).

Exercise 1.3.4. Prove that a metric product carries the standard product
topology.

Definition 1.3.5. A sequence {xn }∞ n=1 of points of a topological space X is

said to converge to a point x ∈ X if for any neighborhood U of x there is a
number n0 such that xn ∈ U for all n ≥ n0 . Notation: xn → x (as n → ∞).
The point x is called a limit of the sequence.

In a metric space, xn → x if and only if |xn x| → 0. The following

properties are also specific for metric spaces.

Proposition 1.3.6. Let X and Y be metric spaces. Then

(1) A sequence in X cannot have more than one limit.
(2) A point x ∈ X is an accumulation point of a set S ⊂ X (i.e.,
belongs to the closure of S) if and only if there exists a sequence
{xn }∞n=1 such that xn ∈ S for all n and xn → x. In particular, S
is closed if and only if it contains all limits of sequences contained
within S.
(3) A map f : X → Y is continuous at a point x ∈ X if and only if
f (xn ) → f (x) for any sequence {xn } converging to x.
1.4. Lipschitz Maps 9

1.4. Lipschitz Maps

Definition 1.4.1. Let X and Y be metric spaces. A map f : X → Y is
called Lipschitz if there exists a C ≥ 0 such that |f (x1 )f (x2 )| ≤ C|x1 x2 |
for all x1 , x2 ∈ X. Any suitable value of C is referred to as a Lipschitz
constant of f . The minimal Lipschitz constant is called the dilatation of f
and denoted by dil f . The dilatation of a non-Lipschitz function is infinity.
A map with Lipschitz constant 1 is called nonexpanding.
Exercise 1.4.2. The distance from a point x to a set S in a metric space is
defined by dist(x, S) = inf y∈S |xy|. Prove that dist(·, S) is a nonexpanding
function.
Proposition 1.4.3. (1) All Lipschitz maps are continuous.
(2) If f : X → Y and g : Y → Z are Lipschitz maps, then g ◦ f is
Lipschitz and dil(g ◦ f ) ≤ dil f · dil g.
(3) The set of real-valued Lipschitz functions on a metric space (and,
more generally, the set of Lipschitz functions from a metric space to
a normed space) is a vector space. One has dil(f +g) ≤ dil f +dil g,
dil(λf ) = |λ| dil f for any Lipschitz functions f and g and λ ∈ R.
Definition 1.4.4. Let X and Y be metric spaces. A map f : X → Y is
called locally Lipschitz if every point x ∈ X has a neighborhood U such that
f |U is Lipschitz. The dilatation of f at x is defined by
dilx f = inf{dil f |U : U is a neighborhood of x}.
Exercise 1.4.5. Let X be a metric space. Prove that dil f = supx∈R dilx f
for any map f : R → X. Prove the same statement with R replaced by S 1
with the metric described in Exercise 1.2.9. Show that it is not true for S 1
with the metric restricted from R2 .
Definition 1.4.6. Let X and Y be metric spaces. A map f : X → Y is
called bi-Lipschitz if there are positive constants c and C such that
c|x1 x2 | ≤ |f (x1 )f (x2 )| ≤ C|x1 x2 |
for all x1 , x2 ∈ X.

Clearly every bi-Lipschitz map is a homeomorphism onto its image.

Definition 1.4.7. Two metrics d1 and d2 on the same set X are called
Lipschitz equivalent if there are positive constants c and C such that
c · d1 (x, y) ≤ d2 (x, y) ≤ C · d1 (x, y)
for all x, y ∈ X.
10 1. Metric Spaces

In other words, d1 and d2 are Lipschitz equivalent if the identity is a bi-

Lipschitz map from (X, d1 ) to (X, d2 ). Clearly this is an equivalence relation
on the set of metrics in X. Lipschitz equivalent metrics determine the same
topology.
Exercise 1.4.8. Give an example of two metrics on the same set that
determine the same topology but are not Lipschitz equivalent.
Exercise 1.4.9. Let X and Y be metric spaces. Prove that the following
three metrics on X × Y are Lipschitz equivalent:
1. The metric defined in Example 1.2.4.
2. d1 ((x1 , y1 ), (x2 , y2 )) = |x1 x2 | + |y1 y2 |.
3. d∞ ((x1 , y1 ), (x2 , y2 )) = max{|x1 x2 |, |y1 y2 |}.
Exercise 1.4.10. Let X be a metric space. Prove that its metric is a
Lipschitz function on X × X where X × X is regarded to have any of the
metrics from the previous exercise.

We conclude this section with the following important theorem about

normed spaces.
Theorem 1.4.11. 1. Two norms on a vector space determine the same
topology if and only if they are Lipschitz equivalent;
2. All norms on a finite-dimensional vector space are Lipschitz equiva-
lent.

1.5. Complete Spaces

Definition 1.5.1. A sequence {xn } in a metric space is called a Cauchy
sequence if |xn xm | → 0 as n, m → ∞. The precise meaning of this is the
following: for any ε > 0 there exists an n0 such that |xn xm | < ε whenever
n ≥ n0 and m ≥ n0 .
A metric space is called complete if every Cauchy sequence in it has a
limit.

It is known from analysis (see e.g. [Mun]) that R is a complete space.

It easily follows that Rn is complete for all n. R \ {0} is an example of a
noncomplete space; a sequence that would converge to zero in R is a Cauchy
sequence that has no limit in this space. (Note that a converging sequence
is always a Cauchy one.)
Exercise 1.5.2. Prove that completeness is preserved by a bi-Lipschitz
homeomorphism. In particular, Lipschitz equivalent metrics share complete-
ness or noncompleteness.
1.5. Complete Spaces 11

Exercise 1.5.3. Show that completeness is not a topological property; i.e.,

there exist homeomorphic metric spaces X and Y such that X is complete
but Y is not.
Exercise 1.5.4. The diameter of a set S in a metric space is defined by
diam(S) = supx,y∈S |xy|. Prove that a metric space X is complete if and
only if it possesses the following property. If {Xn } is a sequence of closed
subsets of X such that Xn+1 ⊂ Xn for all n, and diam(Xn ) → 0 as n → 0,
then the sets Xn have a common point.
Show that the assumption diam(Xn ) → 0 is essential.
Proposition 1.5.5. Let X be a metric space and Y ⊂ X. Then
(1) If Y is complete, then Y is closed in X.
(2) If X is complete and Y is closed in X, then Y is complete.

The following two exercises provide useful tools for proving completeness
of some spaces.
Exercise 1.5.6. Let {xn } be a Cauchy sequence in a metric space. Prove
that
(a) If {xn } has a converging subsequence, then it converges itself.
(b) For any sequence {εn } of positive numbers there exists a subsequence
{yn } of {xn } such that |yn yn+1 | < εn for all n.
Exercise P1.5.7. Let {xn }∞
n=1 be a sequence in a metric space such that
the series ∞ |x x
n=1 n n+1 | has a finite sum. Prove that {xn } is a Cauchy
sequence.
Exercise 1.5.8 (fixed-point theorem). Let X be a complete space, 0 < λ <
1, and let f : X → X be a map such that |f (x)f (y)| ≤ λ|xy| for all x, y ∈ X.
Prove that there exists a unique point x0 ∈ X such that f (x0 ) = x0 .
Hint: Obtain x0 as the limit of a sequence {xn } where x1 is an arbitrary
point and xn+1 = f (xn ) for all n ≥ 1.

The following simple proposition is used many times in this book.

Proposition 1.5.9. Let X be a metric space and X ′ a dense subset of X.
Let Y be a complete space and f : X ′ → Y a Lipschitz map. Then there
exists a unique continuous map f˜ : X → Y such that f˜|X ′ = f . Moreover f˜
is Lipschitz and dil f˜ = dil f .

Proof. Let C be a Lipschitz constant for f . For every x ∈ X define

f˜(x) ∈ Y as follows. Choose a sequence {xn }∞ n=1 such that xn ∈ X for
′

all n, and xn → x as n → ∞. Observe that {f (xn )} is a Cauchy sequence

in Y . Indeed, we have |f (xi )f (xj )| ≤ C|xi xj | for all i, j, and |xi xj | → 0
12 1. Metric Spaces

as i, j → ∞ because the sequence {xn } converges. Therefore the sequence

{f (xn )} converges; then define f˜(x) = limn→∞ f (xn ).
Thus we have defined a map f˜ : X → Y . Then the inequality
|f (x)f˜(x′ )| ≤ C|xx′ | for x, x′ ∈ X follows as a limit of similar inequalities for
˜
f . Indeed, if x = lim xn , x′ = lim x′n , f˜(x) = lim f (xn ), f˜(x′ ) = lim f (x′n ),
then
|f˜(x)f˜(x′ )| = lim |f (xn )f (x′n )| ≤ C lim |xn x′n | = C|xy|.
n→∞ n→∞
Therefore f is Lipschitz (and hence continuous) and dil f ≤ C.
The uniqueness of f˜ is trivial: if two continuous maps coincide on a
dense set, then they coincide everywhere. ¤

Completion. Inside a metric space there is an operation of taking closure

that makes a closed subset out of an arbitrary subset. The following theorem
defines a similar operation that makes a complete metric space out of a
noncomplete one.
Theorem 1.5.10. Let X be a metric space. Then there exists a complete
metric space X̃ such that X is a dense subspace of X̃. It is essentially
unique in the following sense: if X̃ ′ is another space with these properties,
then there exists a unique isometry f : X̃ → X̃ ′ such that f |X = id.
Definition 1.5.11. The space X̃ from the above theorem is called the
completion of X.

Proof of Theorem 1.5.10. Let X denote the set of all Cauchy sequences
in X. Introduce the distance in X by the formula
d({xn }, {yn }) = lim |xn yn |.
n→∞
It is easy to check that, if {xn } and {yn } are Cauchy sequences, then {|xn yn |}
is either a Cauchy sequence of real numbers or |xn yn | = ∞ for all large
enough n. Therefore the above limit always exists. Clearly d is a semi-
metric on X. Define X̃ = X/d (see Proposition 1.1.5 and a remark after
it).
There is a natural map from X to X̃, namely let a point x ∈ X be
mapped to a point of X̃ represented by the constant sequence {x}∞ n=1 .
Since this map is distance-preserving, we can identify X with its image
in X̃ (formally, change the definition of X̃ so that points of X replace their
images). This way X becomes a subset of X̃. It is dense because a point of
X̃ represented by a sequence {xn } is the limit of this sequence (thought of
as the sequence in X ⊂ X̃).
The uniqueness part of the theorem follows from Proposition 1.5.9
applied to the inclusion maps from X to X̃ and X̃ ′ . ¤
1.6. Compact Spaces 13

Baire’s theorem.
Definition 1.5.12. A set Y in a topological space X is nowhere dense if
the closure of Y has empty interior.

Equivalently, Y is nowhere dense in X if the interior of X \Y is dense. By

plugging in the definitions of closure and interior, one obtains the following
description: Y is nowhere dense if and only if any open set U contains a ball
which does not intersect Y .
Theorem 1.5.13 (Baire’s theorem). A complete metric space cannot be
covered by countably many nowhere dense subsets. Moreover, a union of
countably many nowhere dense subsets has a dense complement.
Remark 1.5.14. An equivalent formulation is: in a complete space, an
intersection of countably many sets whose interiors are dense (in particular,
an intersection of countably many open dense sets) is dense.
Remark 1.5.15. A union of countably many nowhere dense sets may not
be nowhere dense. For example, consider Q ⊂ R as a union of single points.

Proof of the theorem. Let X be a complete metric space and {Yi }∞ i=1 be
a countable family of nowhere dense sets. We have to S show that any open
set U ⊂ X contains a point which does not belong to ∞ i=1 Yi . Since Y1 is
nowhere dense, there is a (closed) ball B1 ⊂ U which does not intersect Y1 .
Since Y2 is nowhere dense, there is a closed ball B2 ⊂ B1 which does not
intersect Y2 . And so on. This way we obtain a sequence B1 ⊃ B2 ⊃ . . . of
closed balls where each ball Bi has no common points with the respective
set Yi . We may choose the radii of the balls Bi so that they converge to
zero. Then the centers of the balls form a Cauchy sequence. The limit of
this sequence belongs to all balls and therefore does not belong to any of
the sets Yi . ¤

1.6. Compact Spaces

Recall that a topological space X is called compact if any open covering of
X (that is, a collection of open sets that cover X) has a finite sub-collection
that still covers X. The term “compact set” refers to a subset of a topological
space that is compact with respect to its induced topology.
Definition 1.6.1. Let X be a metric space and ε > 0. A set S ⊂ X is
called an ε-net if dist(x, S) ≤ ε for every x ∈ X.
X is called totally bounded if for any ε there is a finite ε-net in X.
Exercise 1.6.2. Let X be a metric space, Y ⊂ X and ε > 0. A set S ∈ X
is called an ε-net for Y if dist(y, S) ≤ ε for all y ∈ Y . Prove that, if there is
a finite ε-net for Y , then there exists a finite (2ε)-net for Y contained in Y .
14 1. Metric Spaces

Exercise 1.6.3. Prove that

(a) Any subset of a totally bounded set is totally bounded.
(b) In Rn , any bounded set (that is, a set whose diameter is finite) is
totally bounded.
Exercise 1.6.4. A set S in a metric space is called ε-separated, for an
ε > 0, if |xy| ≥ ε for any two different points x, y ∈ S. Prove that
1 If there exists an (ε/3)-net of cardinality n, then an ε-separated set
cannot contain more than n points.
2. A maximal ε-separated set is an ε-net.

The following theorem gives a list of equivalent definitions of compact-

ness for metric spaces. The last one is the most important for us.
Theorem 1.6.5. Let X be a metric space. Then the following statements
are equivalent:
(1) X is compact.
(2) Any sequence in X has a converging subsequence.
(3) Any infinite subset of X has an accumulation point.
(4) X is complete and totally bounded.

The following properties are known from general topology.

Proposition 1.6.6. Let X and Y be Hausdorff topological spaces. Then
(1) If S ⊂ X is a compact set, then S is closed in X.
(2) If X is compact and S ⊂ X is closed in X, then S is compact.
(3) If {Xn }∞n=1 is aTsequence of compact sets such that Xn+1 ⊂ Xn for
all n, then the ∞ n=1 Xn 6= ∅.
(4) A subset of Rn is compact if and only if it is closed and bounded.
(5) If X is compact and f : X → Y is a continuous map, then f (X) is
a compact set.
(6) If X is compact and f : X → Y is bijective continuous map, then
f is a homeomorphism.
(7) If X is compact and f : X → R is a continuous function, then f
attains its maximum and minimum.

The property of Rn expressed in the fourth statement is known as

boundedly compactness.
Definition 1.6.7. A metric space is said to be boundedly compact if all
closed bounded sets in it are compact.
1.6. Compact Spaces 15

Exercise 1.6.8. Prove that a metric space (with possibly infinite distances)
is compact if and only if it is a union of a finite number of compact subsets
each of which carries a finite metric.

Exercise 1.6.9. Let X be a compact metric space. Prove that

1. If the metric of X is finite, then diam X < ∞.
2. There exist two points x, y ∈ X such that |xy| = diam X.

Exercise 1.6.10. Define the distance between two subsets A and B of a

metric space X by dist(A, B) = inf{|xy| : x ∈ A, y ∈ B}. (Warning: this
kind of distance does not satisfy triangle inequality!) Prove that
1. If A and B are compact, then there exist x ∈ A and y ∈ B such that
|xy| = dist(A, B).
2. If X = Rn , the same is true under the weaker assumptions that A is
compact and B is closed.

Theorem 1.6.11 (Lebesgue’s Lemma). Let X be a compact metric space,

and let {Uα }α∈A be an open covering of X. Then there exists a ρ > 0 such
that any ball of radius ρ in X is contained in one of the sets Uα .

Proof. We may assume that the metric of X is finite and none of the sets
Uα covers the whole space. Then one can define a function f : X → R by

f (x) = sup{r ∈ R : Br (x) is contained in one of the Uα }.

Since {Uα } is an open covering, f (x) is well-defined and positive for all
x ∈ X. Clearly f is nonexpanding function and hence continuous. Therefore
it attains a (positive) minimum r0 . Define ρ = r0 /2. ¤

The number ρ from the theorem is referred to as a Lebesgue number of

the covering.

Theorem 1.6.12. Let X and Y be metric spaces and let X be compact.

Then every continuous map f : X → Y is uniformly continuous, i.e., for
every ε > 0 there is a δ > 0 such that for all x1 , x2 ∈ X such that |x1 x2 | < δ
one has |f (x1 )f (x2 )| < ε.

Proof. Every x ∈ X has a neighborhood U such that f (U ) ⊂ Bε/2 (f (x)), in

particular, diam(f (U )) < ε. Hence open sets U such that diam(f (U )) < ε
cover X. Let δ be a Lebesgue number of this covering. ¤

Exercise 1.6.13. Prove that a locally Lipschitz map from a compact space
is Lipschitz.
16 1. Metric Spaces

Isometries of compact spaces. Unlike most of this chapter (which is

rather analytical), the following two theorems have purely geometric con-
tents. The first of them is a simple special case of the second one.
Theorem 1.6.14. A compact metric space cannot be isometric to a proper
subset of itself. In other words, if X is a compact space and f : X → X is
a distance-preserving map, then f (X) = X.

Proof. Suppose the contrary, i.e., let p ∈ X \ f (X). Since f (X) is compact
and hence closed, there exists an ε > 0 such that Bε (p) ∩ f (X) = ∅.
Let n be the maximal possible cardinality of an ε-separated set in X (see
Exercise 1.6.4) and let S ⊂ X be an ε-separated set of cardinality n. Since f
is distance-preserving, the set f (S) is also ε-separated. On the other hand,
dist(p, f (S)) ≥ dist(p, f (X)) ≥ ε and therefore f (S) ∪ {p} is an ε-separated
set of cardinality n + 1. Contradiction. ¤
Theorem 1.6.15. Let X be a compact metric space. Then
(1) Any nonexpanding surjective map f : X → X is an isometry.
(2) If a map f : X → X is such that |f (x)f (y)| ≥ |xy| for all x, y ∈ X,
then f is an isometry.

Proof. 1. Suppose the contrary, i.e., that |f (p)f (q)| < |pq| for some points
p, q ∈ X. Fix p and q and pick an ε > 0 such that |f (p)f (q)| < |pq| − 5ε.
Let n be a natural number such that there exists at least one ε-net in
X of cardinality n. Consider the set N ⊂ X n of all n-tuples of points of X
that form ε-nets in X. This set is closed in X n and therefore it is compact.
Define a function D : X n → R by
n
X
D(x1 , . . . , xn ) = |xi xj |.
i,j=1

This function is continuous and therefore it attains a minimum on N. Let

S = (x1 , . . . , xn ) be an element of N at which the minimum is attained. Since
f is nonexpanding and surjective, the collection f (S) := (f (x1 ), . . . , f (xn ))
is also an element of N. Moreover D(f (S)) ≤ D(S) because |f (xi )f (xj )| ≤
|xi xj | for all i, j. But D(S) is the minimum of D on N; therefore D(f (S)) =
D(S) and |f (xi )f (xj )| = |xi xj | for all i, j.
On the other hand, there exist indices i and j such that |pxi | ≤ ε and
|qxj | ≤ ε. For these i and j we have
|xi xj | ≥ |pq| − |pxi | − |qxj | ≥ |pq| − 2ε
and
|f (xi )f (xj )| ≤ |f (p)f (q)| + |f (p)f (xi )| + |f (q)f (xj )|
≤ |f (p)f (q)| + 2ε ≤ |pq| − 3ε.
1.7. Hausdorff Measure and Dimension 17

Hence |f (xi )f (xj )| < |xi xj |. Contradiction.

2. Define Y = f (X). The argument from the proof of the previous
theorem shows that Y is dense in X. Consider the map g = f −1 : Y → X.
Since g is nonexpanding and Y is dense in X, g can be extended to a
nonexpanding map g̃ : X → X. By the first part of the theorem, g̃ is an
isometry. Hence Y = X and f is an isometry. ¤

1.7. Hausdorff Measure and Dimension

1.7.1. Measures in general. The notion of measure generalizes length,
area and volume. Roughly speaking, a measure on a space X is a nonnega-
tive function defined on a set of subsets of X and possessing the additivity
property of the area; namely, the measure of a union of disjoint sets equals
the sum of measures of these sets. In fact, a stronger requirement of count-
able additivity (or σ-additivity, see definitions below) is imposed to make
measures really useful.
Although it is commonly accepted that one can speak of area for figures
on the plane and volume of three-dimensional bodies, it is not so easy to
give correct definitions of these notions. In fact, the statement “every set
has a volume” is wrong, as the following fact shows (see [HM] for details).
Let B denote a unit ball in R3 with its center removed. Then B can be split
into four disjoint subsets, which can be rearranged (by means of rotations)
so as to form two copies of B. If volume was defined for sets such as these
four pieces (which are in fact extremely wild sets), then B and the union of
its disjoint copies would have equal volumes, implying that the volume of
all sets is zero. Therefore one has to restrict the class of sets for which the
volume (or other measure) is defined. A class of sets on which a measure
is defined (these sets are called measurable, w.r.t. the measure) must be a
σ-algebra, which is defined as follows:
Definition 1.7.1. Let X be an arbitrary set. A set A of subsets of X is
called a σ-algebra if it satisfies the following conditions:
(1) ∅ and X are elements of A;
(2) If A, B ∈ A, then A \ B ∈ A;
(3) If {Ai }i∈I isSa finite or countable collection of elements of A, then
their union i∈I Ai is also an element of A.
S T
Remark 1.7.2. Due to the formula (X \ Ai ) = X \ Ai , a σ-algebra
contains any intersection of a countable collection of its elements.
Definition 1.7.3. A measure on a σ-algebra A is a function µ : A →
R+ ∪ {+∞} such that
(1) µ(∅) = 0; and
18 1. Metric Spaces

(2) if {Ai } is a finite or countable

S collection
P of elements of A and the
sets Ai are disjoint, then µ( Ai ) = µ(Ai ).

The second
P condition is referred to as σ-additivity. Note that the
expression µ(Ai ) is either a finite sum or a series; its value is well defined
and independent of the order of terms since the terms are nonnegative.
Exercise 1.7.4. Let µ be a measure. Prove the following statements:
(a) Let {Ai }∞
i=1 be a sequence of measurable sets such that Ai ⊂ Ai+1
S for
all i. Then the sequence {µ(Ai )} is nondecreasing and lim µ(Ai ) = µ( Ai ).
(b) Let {Ai }∞
i=1 be a sequence of measurable sets such that Ai ⊃ Ai+1
for all i, and assume that µ(A1T ) < ∞. Then the sequence {µ(Ai )} is
nonincreasing and lim µ(Ai ) = µ( Ai ).
(c) The assumption µ(A1 ) < ∞ in (b) is essential.

If S is an arbitrary collection of subsets of a set X, there obviously

exists a unique minimal σ-algebra containing S (prove this!); it is called the
σ-algebra generated by S. If X is a topological space, then the σ-algebra
generated by its topology (i.e., by the set of all open sets) is called the Borel
σ-algebra of X. Elements of the Borel σ-algebra are called Borel sets. A
measure defined on the Borel σ-algebra is called a Borel measure over X.
The following theorem provides a basis for measure theory in Euclidean
spaces.
Theorem 1.7.5 (Lebesgue). There exists a unique Borel measure mn
over Rn which is invariant under parallel translations and such that
mn ([0, 1]n ) = 1.

The measure mn is called Lebesgue measure. The uniqueness part of

the theorem implies that every translation-invariant Borel measure in Rn
with a finite value for a cube is a constant multiple of mn .
Exercise 1.7.6. Prove that
1. Lebesgue measure is invariant under isometries of Rn .
2. If L : Rn → Rn is a linear map, then mn (L(A)) = | det L| · mn (A) for
any measurable set A ⊂ Rn . In particular, a homothety with coefficient C
multiplies the Lebesgue measure by C n .
Hint: Use the uniqueness part of Theorem 1.7.5.

1.7.2. Hausdorff measure. To motivate the definition of Hausdorff mea-

sures, let us recall the main idea of the construction behind the proof of
Theorem 1.7.5. It begins with choosing a class of simple sets (such as balls
or cubes). Then a set (from an appropriate σ-algebra) is covered by simple
sets; then the Lebesgue measure of the set is defined as the infimum of total
1.7. Hausdorff Measure and Dimension 19

measures of such covers. Speaking about “the total measure of a cover” one
means here that certain measure is already assigned to simple sets.
To define Hausdorff n-dimensional measure on a metric space, one could
proceed along the same lines: cover a set by metric balls such that all their
radii are less than ε. For each ball, consider a Euclidean ball of the same
radius and add their volumes for all balls from the cover: this will be the
total measure of the cover. Taking its infimum over all covers and passing
to the limit as ε approaches zero, one gets a version of Hausdorff measure of
the set. Instead of adding volumes of Euclidean balls of the same radii, one
could simply add the radii of balls from the cover raised to the power n: the
result is the same up to a constant multiplier. It turns out that arbitrary
sets and diameters are technically more convenient to use than metric balls
and radii.
Now we pass to formal definitions.

Definition 1.7.7. Let X be a metric space and d be a nonnegative real

number.
For a finite or countable
S covering {Si }i∈I of X (that is, a collection of
sets such that X ⊂ Si ), define its d-weight wd ({Si }) by the formula
X
wd ({Si }) = (diam Si )d .
i

If d = 0, substitute each (if any) 00 term in the formula by 1.

For an ε > 0 define µd,ε (X) by
© ª
µd,ε (X) = inf wd ({Si }) : diam(Si ) < ε for all i .
The infimum is taken over all finite or countable coverings of X by sets of
diameter < ε; if no such covering exists, then the infimum is +∞.
The d-dimensional Hausdorff measure of X is defined by the formula
µd (X) = C(d) · lim µd,ε (X)
ε→0

where C(d) is a positive normalization constant. This constant is introduced

for only one reason: for integer d it is convenient to choose C(d) so that the
d-dimensional Hausdorff measure of a unit cube in Rn equals 1. In fact,
almost nothing depends on the actual value of C(d).
It may be unclear from the definition what the value of µd (∅) is. We
explicitly define µd (∅) = 0 for all d ≥ 0.

Clearly µd,ε (X) is a nonincreasing function of ε. Since such a function

has a (possibly infinite) limit as ε → 0, µd (X) is well defined for any metric
space X. It may be either a nonnegative real number or +∞.
20 1. Metric Spaces

Though we have defined Hausdorff measure for a metric space, this

notion will often be applied to subsets of metric spaces. In such cases, a
subset should be considered as a metric space with the restricted metric.
(Note that the definition can be read verbatim if X is a subset of a larger
metric space; it does not matter whether covering sets Si are actually
contained in X.)
The following proposition summarizes the properties of Hausdorff mea-
sure that follow immediately from the definition.
Proposition 1.7.8. Let X and Y be metric spaces, and let A and B be
subsets of X. Then
(1) If A ⊂ B, then µd (A) ≤ µd (B).
S P
(2) µd ( Ai ) ≤ µd (Ai ) for any finite or countable collection of sets
Ai ⊂ X.
(3) If dist(A, B) > 0, then µd (A ∪ B) = µd (A) + µd (B).
(4) If f : X → Y is a Lipschitz map with a Lipschitz constant C, then
µd (f (X)) ≤ C d · µd (X).
(5) If f : X → Y is a C-homothety, i.e., |f (x1 )f (x2 )| = C|c1 x2 | for all
x1 , x2 ∈ X, then µd (f (X)) = C d · µd (X).

According to Carathéodory’s criterion, ([Fe], 2.3.1(9)), any nonnegative

function on the Borel σ-algebra of X possessing the properties 1–3 from
Proposition 1.7.8 is actually a measure. Thus we obtain
Theorem 1.7.9. For any metric space X and any d ≥ 0, µd is a measure
on the Borel σ-algebra of X.
Exercise 1.7.10. Prove that 0-dimensional Hausdorff measure of a set is
its cardinality. In other words, µ0 (X) is a number of points in X if X is a
finite set, and µ0 (X) = ∞ if X is an infinite set.
Exercise 1.7.11. Let X and Y be metric spaces and f : X → Y a locally
Lipschitz map with dilatation ≤ C. Prove that µd (f (X)) ≤ C d · µd (X)
assuming that (a) X is compact; (b) X has a countable topological base.

1.7.3. Hausdorff measure in Rn . Let I denote the interval [0, 1] of R.

Then I n = [0, 1]n is the unit cube in Rn .
Theorem 1.7.12. 0 < µn (I n ) < ∞.
Exercise 1.7.13. Prove the theorem.

Now we can define the normalization constant C(n) from the definition of
Hausdorff measure. Namely, choose C(n) so that µn (I n ) = 1. The existence
of such a constant follows from Theorem 1.7.12. Theorem 1.7.5 then implies
1.7. Hausdorff Measure and Dimension 21

that Hausdorff measure µn on Rn coincides with the standard n-dimensional

volume mn .
In most cases, the actual value of C(n) is not important. However it is
an interesting fact that C(n) equals the volume of the Euclidean n-ball of
diameter 1. The proof is based on the following theorem.
Theorem 1.7.14 (Vitali’s Covering Theorem). Let X be a bounded set in
Rn and let B be a collection of closed balls in Rn such that for every x ∈ X
and ε > 0 there is a ball B ∈ B such that x ∈ B and diam(B) < ε. Then
B contains a finite or countable subcollection {Bi } of disjoint balls which
covers X upSto a set of zero measure, i.e., such that Bi ∩ Bj = ∅ if i 6= j
and µn (X \ i Bi ) = 0.

Proof. We may assume that every ball B ∈ B contains at least one point
of X and exclude the balls with radius greater than 1. Then all these balls
are contained in the 2-neighborhood of X which is bounded and hence has
finite volume. We construct a sequence {Bi }∞ i=1 of balls by induction. If
B1 , . . . , Bm are already constructed, we choose the next ball Bm+1 as follows.
Let Bm denote the set of balls from the collection that do not intersect any
of B1 , . . . , Bm . If Bm is empty, then B1 ∪ · · · ∪ Bm covers the entire set X
and the proof is finished (this follows from the condition that every point
is covered by balls of arbitrarily small radii). If Bm is not empty, choose
Bm+1 to be any element of Bm with
1
(1.1) diam(Bm+1 ) > sup{diam(B) : B ∈ Bm }.
2
The balls Bi are disjoint by the construction. We will now show that they
cover X up to a set of zero measure. Fix an ε > 0. Since the P∞balls are disjoint
and are contained in a set of finiteP volume, we have i=0 µn (Bi ) <S ∞.
Hence there is an index m such that ∞ i=m+1 µ n (Bi ) < ε. Let x ∈ X \ i Bi
and let B be any ball from the collection that contains x S and does not
intersect the balls B1 , . . . , Bm . Note that B must intersect i Bi because
otherwise B ∈ Bm for all m which contradicts that µn (Bi ) → 0. Let k
be the minimal index such that B ∩ Bk 6= ∅. Then B ∈ Bk−1 and hence
diam(Bk ) > 12 diam(B) by (1.1). It follows that the distance from x to the
center of Bk is not greater than 5 times the radius of Bk . Hence x belongs
to the ball with the same center as Bk and radius 5 times larger. We denote
this ball by 5Bk .
S
We have just provedSthat every S∞x ∈ X \ i Bi belongs to a ball 5Bk for
some k > m. Thus X \ i Bi ⊂ i=m+1 (5Bi ); hence
[ ∞
X ∞
X
n
µn (X \ Bi ) ≤ µn (5Bi ) = 5 µn (Bi ) < 5n ε.
i i=m+1 i=m+1
22 1. Metric Spaces

S
Since ε is arbitrary, it follows that µn (X \ i Bi ) = 0. ¤
Corollary 1.7.15. The normalization constant for the n-dimensional Haus-
dorff measure equals the volume of the Euclidean n-ball of diameter 1.

Proof. Let Cn denote the constant from the formulation. Then the volume
of a Euclidean n-ball equals Cn dn where d is its diameter. Let µ′n be the
n-dimensional Hausdorff measure with normalization constant Cn . We have
to prove that µ′n = µn , i.e., that µ′n (I n ) = 1.
1. µ′n (I n ) ≤ 1. To prove this, apply Theorem 1.7.14 to the set B of all
closed Euclidean balls contained in I n . ThisSyields a countable collection
{Bi } of such balls such
P that the set Y = I nP\ Bi has zero measure. Hence
µn (I ) ≤ µn (Y ) + Cn diam(Bi ) = 0 + mn (Bi ) ≤ mn (I n ) = 1.
′ n ′ n

2. µ′n (I n ) ≥ 1. By a well-known Bieberbach inequality (cf. e.g. [BZ],

Theorem 11.2.1), a Euclidean ball has the maximal volume among the sets
with the same diameter. Hence mn (S) ≤ Cn diam(S)n for any P bounded set
S ⊂ R n . Now if {S }∞ is a covering of I n , then 1 = m (I n ) ≤ mn (Si ) ≤
P i i=1 n
n
Cn diam(S) . The statement follows. ¤

1.7.4. Hausdorff dimension. The next theorem tells us how the Haus-
dorff measure of a fixed set depends on dimension. Briefly, the measure is
zero or infinite for all dimensions except at most one. More precisely, there
is a “critical dimension” below which the measure is infinity and above
which the measure is zero. This dimension is an important characteristic
of a metric space, called the Hausdorff dimension. Warning: at the critical
dimension, all three possibilities (the measure is zero, positive number or
+∞) may take place.
Theorem 1.7.16. For a metric space X there exists a d0 ∈ [0, +∞] such
that µd (X) = 0 for all d > d0 and µd (X) = ∞ for all d < d0 .

Proof. Define d0 = inf{d ≥ 0 : µd (X) 6= ∞}. Trivially µd (X) = ∞ for all

d < d0 . If d > d0 , there is a d′ < d such that µd′ (X) = M < ∞. Therefore
for any ε > 0 there exists a covering {Si } of X such that diam Si < ε for all
P ′
i and (diam Si )d < 2M . Then
X ′
X ′ ′
(diam Si )d ≤ εd−d · (diam Si )d ≤ 2εd−d M.
′ ′
Hence µd,ε (X) ≤ 2εd −d M . Since εd −d → 0 as ε → 0, we have µd (X) =
0. ¤
Definition 1.7.17. The value d0 from Theorem 1.7.16 is called the Haus-
dorff dimension of X and denoted by dimH (X).
Remark 1.7.18. Hausdorff dimension is not necessarily integer.
1.7. Hausdorff Measure and Dimension 23

Here are some immediate properties of Hausdorff dimension.

Proposition 1.7.19. Let X be a metric space. Then
(1) If Y ⊂ X, then dimH (Y ) ≤ dimH (X).
(2) If X is covered by a finite or countable collection {Xi } of its subsets,
then dimH (X) = supi dimH (Xi ).
(3) If f : X → Y is a Lipschitz map, then dimH (f (X)) ≤ dimH (X). In
particular, bi-Lipschitz equivalent metric spaces have equal Haus-
dorff dimensions.
(4) dimH (Rn ) = dimH (I n ) = n.
Exercise 1.7.20. Let X and Y be metric spaces and f : X → Y a map
such that |f (x1 )f (x2 )| ≤ C · |x1 x2 |α for all x1 , x2 ∈ X, where C and α are
some positive constants. Prove that dimH (f (X)) ≤ dimH (X)/α.
Exercise 1.7.21. Prove that the Hausdorff dimension of the standard
Cantor set is log3 2. More generally, let X be a compact space that can
be split into n subsets X1 , . . . , Xn that can be obtained from X by dilations
with coefficients
P dc1 , . . . , cn respectively. Prove that d = dimH (X) satisfies
the equation ci = 1. (Warning: compactness is essential!)
Exercise 1.7.22. Give examples of
(a) an uncountable metric space whose Hausdorff dimension is zero;
(b) a metric space X with dimH (X) = 1 and µ1 (X) = 0;
(c) a metric space X with dimH (X) = 1 and µ1 (X) = +∞.
 Chapter 2

Length Spaces

2.1. Length Structures

First we want to informally illustrate our main concept. Imagine that
you ask a mathematician: “What is the distance between New York and
Sydney?”. Perhaps, you get the answer “about 8 thousand miles”. It is
formally correct and still absolutely useless: this is the length of a straight
tunnel through the Earth. Analogously, every mountaineer knows that
distance in mountains is a tricky thing: if you measure it by an optical
device, you get the distance “as a crow flies”. It may be relevant for a crow,
while wingless creatures confined to the surface of the Earth (like us) have
to take long detours with lots of ups and downs; see Figure 2.1.

A
B

PSfrag replacements
A
B

Figure 2.1: “A crow flies” along the segment AB; for a pedestrian it probably
takes longer.

25
26 2. Length Spaces

This little philosophical digression contains a very clear mathematical

moral: in many cases, we have to begin with length of paths as the primary
notion and only after that can we derive a distance function. Let us make
this observation slightly more precise. For every two points on a surface in
Euclidean space (you may keep thinking of the surface of the Earth) we can
measure Euclidean distance between the two points. What we do instead
is we introduce a new distance which is measured along the shortest path
between the two points. Generalizing this idea, one says that a distance
function on a metric space is an intrinsic metric if the distance between two
points can be realized by paths connecting the points (mathematically, it
must be equal to the infimum of lengths of paths between the points—a
shortest path may not exist).
If length of paths is our primary notion, one readily asks for its rigorous
definition, where it may arise from and what are the properties of such
structures. We will be occupied with these questions throughout this book.

2.1.1. Definition of length structures. Loosely speaking, a length struc-

ture consists of a class of admissible paths for which we can measure their
length, and the length itself, which is a correspondence assigning a nonnega-
tive number to every path from the class. Both the class and the correspon-
dence have to possess several natural properties; in all reasonable examples
(and in particular in all examples in this book) these requirements are au-
tomatically satisfied.
From now on we reserve the word path for maps of intervals: a path
γ in a (topological) space X is a (continuous) map γ : I → X defined on
an interval I ⊂ R. By an interval we mean any connected subset of the
real line; it may be open or closed, finite or infinite, and a single point is
counted as an interval. Since a path is a map one can speak about its image,
restrictions, etc.
A length structure on a topological space X is a class A of admissible
paths, which is a subset of all continuous paths in X, together with a map
L : A → R+ ∪ {∞}; the map is called length of path. The class A has to
satisfy the following assumptions:

(1) The class A is closed under restrictions: if γ : [a, b] → X is an

admissible path and a ≤ c ≤ d ≤ b, then the restriction γ|[c,d] of γ
to [c, d] is also admissible.
(2) A is closed under concatenations (products) of paths. Namely, if a
path γ : [a, b] → X is such that its restrictions γ1 ,γ2 to [a, c] and
[c, b] are both admissible paths, then so is γ. (Recall that γ is called
the product or concatenation of γ1 and γ2 , γ = γ1 · γ2 ).
2.1. Length Structures 27

(3) A is closed under (at least) linear reparameterizations: for an

admissible path γ : [a, b] → X and a homeomorphism ϕ : [c, d] →
[a, b] of the form ϕ(t) = αt + β, the composition γ ◦ ϕ(t) = γ(ϕ(t))
is also an admissible path.
Remark 2.1.1. Every natural class of paths comes with its own
class of reparameterizations. For example, consider the class of
all continuous paths and the class of homeomorphisms, the class
of piecewise smooth paths and the class of diffeomorphisms. We
only require that this class of reparameterizations includes all linear
maps.
Examples of such classes include: all continuous paths; piecewise smooth
paths (on a smooth manifold); broken lines in Rn ; see other examples below.
We require that L possesses the following properties:
(1) Length of paths is additive: L(γ|[a,b] ) = L(γ|[a,c] ) + L(γ|[c,b] ) for any
c ∈ [a, b].
(2) The length of a piece of a path continuously depends on the piece.
More formally, for a path γ : [a, b] → X of finite length, denote
by L(γ, a, t) the length of the restriction of γ : [a, b] → X to the
segment [a, t]. We require that L(γ, a, ·) be a continuous function.
(Observe that the previous property implies that L(γ, a, a) = 0.)
(3) The length is invariant under reparameterizations: L(γ ◦ ϕ) = L(γ)
for a linear homeomorphism ϕ.
(In fact, all reasonable length structures are invariant under
arbitrary reparameterizations: L(γ ◦ ϕ) = L(γ) for any homeomor-
phism ϕ such that both γ and γ ◦ ϕ are admissible. However, it is
not necessary to verify this in the beginning.)
(4) We require length structures to agree with the topology of X in the
following sense: for a neighborhood Ux of a point x, the length of
paths connecting x with points of the complement of Ux is separated
from zero:
inf{L(γ) : γ(a) = x, γ(b) ∈ X \ Ux } > 0.
There are several important types of length structures that will appear
in this course. When the reader meets with these structures, it is advisable
to come back to this definition and make sure that all of them belong to the
same general scheme.
Notation. We will often use the notation L(γ, a, b) introduced above.
Namely, if γ : I → X is an (admissible) path and [a, b] ⊂ I, where a ≤ b,
we will denote by L(γ, a, b) the length of the restriction of γ to [a, b], i.e.,
L(γ, a, b) = L(γ|[a,b] ). In addition, we define L(γ, b, a) = −L(γ, a, b). This
28 2. Length Spaces

convention implies that L(γ, a, b) = L(γ, a, c) + L(γ, c, b) for all a, b, c ∈ I

(verify this).

2.1.2. Length spaces. Once we have a length structure, we are ready to

define a metric (a distance function) associated with the structure. We will
always assume that the topological space X carrying the length structure is
a Hausdorff space. For two points x, y ∈ X we set the associated distance
d(x, y) between them to be the infimum of lengths of admissible paths
connecting these points:
dL (x, y) = inf{L(γ); γ : [a, b] → X, γ ∈ A, γ(a) = x, γ(b) = y}.

If it is clear from the context which length structure L gives rise to dL ,

we usually drop L in the notation dL .
Exercise 2.1.2. Verify that (X, dL ) is a metric space.

Note that dL is not necessarily a finite metric. For instance, if X is a

disconnected union of two components, no continuous path can go from one
component to the other and therefore the distance between points of different
components is infinite. On the other hand, there may be points such that
continuous paths connecting them exist but all have infinite length. One
says that two points x, y ∈ X belong to the same accessibility component if
they can be connected by a path of finite length.
Exercise 2.1.3. 1. Check that accessibility by paths of finite length is
indeed an equivalence relation. Your argument should use additivity of
length and the assumption that the concatenation of admissible paths is an
admissible path.
2. Verify that accessibility components coincide with components of
finiteness for dL .
3. Verify that accessibility components coincide with both connectivity
and path connectivity components of (X, dL ).

Did you notice that you have used the following fact?
Exercise 2.1.4. Prove that admissible paths of finite length are continuous
with respect to (X, dL ).

This exercise deals with the topology determined by the metric dL rather
than the initial topology of the space X. There really are examples where
these two topologies differ; such examples will appear later in this book.
Exercise 2.1.5. Prove that the topology determined by dL can be only
finer than that of X: any open set in X is open in (X, dL ) as well.
2.1. Length Structures 29

Definition 2.1.6. A metric that can be obtained as the distance function

associated to a length structure is called an intrinsic, or length, metric. A
metric space whose metric is intrinsic is called a length space.

Not every metric can arise as a length metric. Even if (X, d) is a length
space and A ⊂ X, the restriction of d to A is not necessarily intrinsic. For
example, consider a circle in the plane.
Moreover, not every metrizable topology can be induced by intrinsic
metrics:
Exercise 2.1.7. 1. Prove that the set of rational numbers is not homeo-
morphic to a length space.
2. Prove that the union of the graph {(x, y) : y = sin(1/x), x > 0} and
the y-axis (with its topology inherited from R2 ) is not homeomorphic to a
length space.

There can be more delicate reasons why a topological space may be not
homeomorphic to a length space:
Exercise 2.1.8. Consider the union of segments
∞
[
[(0, 0), (cos 1/i, sin 1/i)] ∪ [(0, 0), (1, 0)]
i=1
in the Euclidean plane, depicted in Figure 2.2.

Figure 2.2: The space (with the topology inherited from R2 ) is not homeomorphic
to a length space.

This set (resembling a fan made of segments) is a topological space

with its topology inherited from the Euclidean plane (this is the topology of
Euclidean distance restricted to the set). Prove that this topological space
is not homeomorphic to a length space.

As a hint to the above exercises, consider the following more general

one.
30 2. Length Spaces

Exercise 2.1.9. Prove that a length space is locally path connected: every
neighborhood of any point contains a smaller neighborhood which is path-
connected.

One uses infimum instead of simple minimum when defining dL since

there may be no shortest path between two points. For instance, consider
the Euclidean plane with an open segment removed; then for the endpoints
of the segment, the shortest path does not exist: it is just removed. Still, its
length can be approximated with a given precision by other paths connecting
the points. Such situations rarely arise in “real-life” examples; in most cases
they will be also prohibited by imposing completeness-compactness type
assumptions. On the other hand, existence of shortest paths helps to avoid
tedious and nonessential complications. For the simplicity of our exposition
we will often restrict ourselves to complete length structures, which are
defined as follows:
Definition 2.1.10. A length structure is said to be complete if for every
two points x, y there exists an admissible path joining them whose length
is equal to dL (x, y); in other words, a length structure is complete if there
exists a shortest path between every two points.

Intrinsic metrics associated with complete length structures are said to

be strictly intrinsic.

2.2. First Examples of Length Structures

To get better motivated, let us briefly meet with a few examples from the
zoo of length structures and intrinsic metrics; we will not analyze them in
this section, but we suggest that you keep them in mind and use them for
testing further definitions and concepts. Notice that in many examples the
space itself is a part of a Euclidean space, and there are two different ways
of changing the usual Euclidean length structure: we change the class of
admissible paths or change the notion of length of paths (or both).
Example 2.2.1 (“Driving in Manhattan”). The space here is the Euclidean
plane, and length of paths is the same as usual. The only difference is that
we restrict the class of admissible paths to broken lines with edges parallel
to one of the coordinate axes. (A critically thinking reader should yell that
paths are maps while broken lines are sets! This is absolutely true, and
formally we mean the paths whose images are broken lines.) Can you draw
a ball in the corresponding intrinsic metric? (It will not look round: you
should get a diamond.)
Example 2.2.2 (“Metric on an island”). The space is a connected region
in the Euclidean plane, and again length of paths is the same as usual.
2.2. First Examples of Length Structures 31

Figure 2.3: Metric balls in a nonconvex island.

Admissible paths are all (piecewise smooth) paths contained in the region. If
the region is convex, this length structure induces usual Euclidean distance.
One may think of this region as an island, and the distance is measured by
a creature who cannot swim. Drawing balls in intrinsic metrics arising this
way may be quite fun; see Figure 2.3. Is this metric strictly intrinsic? What
if we consider the closure of the region?

The reader can generalize this example for a subspace of a space with
length structure. Certainly, only sensible choices for a subspace lead to rea-
sonable examples. For instance, restricting the Euclidean length structure
in R2 to a circle leads to the angular metric. More generally, one obtains
spherical geometry by restricting the usual Euclidean length structure to a
round sphere. On the other hand, restricting the standard length structure
of R to the set of rational points we obtain a space with each accessibility
component consisting of just one point.
Example 2.2.3 (induced length structure). The formal contents of this
example is comprised in the following definition. Let f : X → Y be a
continuous map from a topological space X to a space Y endowed with a
length structure. One defines the induced length structure in X as follows.
A path in X is admissible if its composition with f is admissible in Y . The
length of an admissible path in X is set to the length of its composition with
f with respect to the length structure in Y .
(In fact, this construction may not define a length structure in X because
the new length function may fail to satisfy the fourth condition from section
2.1.1. We use the term “induced length structure” only if this is indeed a
length structure.)
32 2. Length Spaces

At first glance, the above definition may sound like a tautology. However,
the properties of an induced metric may drastically differ from the properties
of a metric we began with. For instance, the leading example of an induced
metric when f is a surface (that is an immersion f : Ω ⊂ R2 → R3 of a
two-dimensional region into R3 ) has served as the main motivating example
in metric geometry for over a century. For a reader who is already familiar
with Riemannian metrics, we mention that it is also true (though hard to
believe and not easy to prove) that every Riemannian length structure on
Rn can be induced by a map f : Rn → Rn (which makes lots of folds and is
rarely smooth).
Example 2.2.4 (“Crossing a swamp”: conformal length). The space is the
Euclidean plane, and admissible paths are all (piecewise smooth) paths. Let
f : R2 → R be a positively-valued continuous (or even L∞ ) function. Define
the length of a path γ : [a, b] → R2 by
Z b
L(γ) = f (γ(t)) · |γ ′ (t)|dt.
a
This length structure can be thought of as a weighted Euclidean distance.
For instance, a traveler who measures the length (=time needed to cover)
of a certain route would apparently assign big values to f in a territory
that is difficult to traverse (for instance, a swamp or a mountain trail).
From the mathematical viewpoint, this is the first example of a Riemannian
length structure, which will be discussed further in Chapter 5; the word
“conformal” in the title of this subsection reflects the fact that such types
of Riemannian structures are called conformally flat.
Example 2.2.5 (Finslerian length). Thinking of the previous example as
a length structure for a traveler who assigns weights to different parts of
his/her path, one notices that an important feature of real travel is not
reflected here. Namely, the difficulty of traversing a region depends not
only on the region itself but also on the direction of the route; for instance,
choosing a direction in which most ravines are oriented might essentially
simplify the trip. To incorporate this additional information, one introduces
a function f in two variables and applies it to both γ and its velocity γ ′ .
The expression for the length reads:
Z b
L(γ) = f (γ(t), γ ′ (t)) dt.
a
(A physics-oriented reader recognizes that this structure can be interpreted
as action.) In order for this expression to be invariant under bijective
reparameterizations of paths, one has to require that f satisfies f (x, kv) =
|k|f (x, v) for all scalars k, points x and vectors v (check this as an exercise for
change of variable in a definite integral). Usually a stronger requirement is
2.3. Length Structures Induced by Metrics 33

imposed on f , namely for every point x the function f (x, ·) must be a norm.
A motivation for this will be explained in section 2.4.2. Length structures
obtained from this type of constructions are called Finslerian, or Finsler.

Remark 2.2.6. A reader who seriously tests our definitions against needs
of travelers and mountaineers will notice that some features are still missing.
Namely the fact that walking downhill may be easier than climbing uphill
cannot be reflected in a length structure. Since the distance in a metric
space must be symmetric, we had to require that the length is invariant
under all changes of variable including the orientation-reversing ones like
t 7→ −t. One could modify the definitions to allow nonsymmetric length
structures and metrics. Everything in this chapter can be adapted to such
generalized settings; however, this would not make sense in the rest of the
book.

Example 2.2.7 (A “cobweb” and a “notebook”). Begin with several dis-

joint segments and glue some of their endpoints together. Such a space
may resemble a cobweb in Euclidean space. This space has a natural length
structure. All continuous paths are admissible. The space is built out of
segments, and we know how to measure the length of a path while it trav-
els within one segment. Thus to find the length of a path we restrict it to
(countably many) intervals such that the image of each interval is contained
in one segment, and add the lengths of the restrictions. This is a first ex-
ample of metric graphs, and the construction of its length structure is a
particular case of gluing, discussed in detail in section 3.1.
Another example of the same type can be made out of several copies of
a closed half-plane by attaching them together along their boundary lines.
This is an example of a polyhedral length space. It can be visualized (and
realized) in Euclidean spaces: it looks like an open book. Can you modify
the definition of the length structure on a cobweb for this case? Caution:
while we could disregard the part of a path spent in endpoints of segments
(nodes of the cobweb) since they have zero length, this is not the case for
the common edge of the half-planes.

2.3. Length Structures Induced by Metrics

2.3.1. Length of curves in metric spaces. Let us recall our motivat-
ing example from the very beginning of this chapter. We began with some
distance function (Euclidean distance “as a crow flies”) that was not satis-
factory since there might be no paths realizing this distance. By saying this
we already mean that we know how to measure length of paths, and this
length is somehow derived from the Euclidean distance!
34 2. Length Spaces

Indeed, some of the main examples of length structures are those induced
by metric structures. For admissible paths one may use just all continuous
paths; for some of them the length may be infinite. In some cases a better
choice is the class of Lipschitz paths, that is, the class of maps γ : [a, b] → X
such that dX (γ(t), γ(t′ )) ≤ C|t − t′ |, for all t, t′ ∈ [a, b]; C is a positive
constant.
How do we define the length of a path in Euclidean space? We approx-
imate the path by broken lines and define the length as the limit of their
lengths. For each of these broken lines, its vertices belong to (the image
of) the path and they are well ordered with respect to the parameter of
the path. Since all that we actually use are distances between neighboring
vertices in this list, we can mimic this definition in a general metric space
in the most straightforward way (compare also with 2.4.12).

Definition 2.3.1. Let (X, d) be a metric space and γ be a path in X, i.e.,

a continuous map γ : [a, b] → X. Consider a partition Y of [a, b]), that is,
a finite collection of points Y = {y0 , ..., yN } such that a = y0 ≤ y1 ≤ y2 ≤
· · · ≤ yN = b. The supremum of the sums
N
X
Σ(Y ) = d(γ(yi−1 ), γ(yi )).
i=1

over all the partitions Y is called the length of γ (with respect to the metric
d) and denoted Ld (γ). A curve is said to be rectifiable if its length is finite.
The length structure induced by the metric d is defined as follows: all
continuous paths (parameterized by closed intervals) are admissible, and the
length is given by the function Ld .

This definitions can be formally applied to any metric space, but one
gets sensible examples only by a wise choice of a metric space to begin with:
for instance, if we start with a discrete space, there are no nonconstant
continuous paths at all. If it is clear from the context which metric d induces
the length L, we usually drop d in the notation Ld .
The usual “Euclidean” definition uses passing to a limit as the edges of
broken lines approach zero. The following exercise shows that there is no
difference here:

Exercise 2.3.2. Prove that Σ(Y ) → L(γ) as maxi {|yi − yi+1 |} → 0.

Exercise 2.3.3. Prove that the definition of length is compatible with the
one used in differential geometry. Namely if (V, | · |) is a finite-dimensional
normed vector space and γ : [a, b] → V is a differentiable map, then
Rb
L(γ) = a |γ ′ (t)| dt.
2.3. Length Structures Induced by Metrics 35

2.3.2. Properties of the induced length. All properties of length struc-

tures hold for this structure; this length is also semi-continuous. Let us verify
(some of) them:
Proposition 2.3.4. The length structure L = Ld induced by a metric d
possesses the following properties:
(i) Generalized triangle inequality: L(γ) ≥ d(γ(a), γ(b)).
(ii) Additivity: if a < c < b, then L(γ, a, c) + L(γ, c, b) = L(γ). In
particularly, L(γ, a, c) is a nondecreasing function of c.
(iii) If γ is rectifiable, the function L(γ|[c,d] ) = L(γ, c, d) is continuous
in c and d.
(iv) L is a lower semi-continuous functional on the space of continuous
maps of [a, b] in X with respect to point-wise convergence, and hence with
respect to the uniform (i.e., C 0 -) topology. This means that if a sequence
of rectifiable paths γi (with the same domain) is such that γi (t) converges to
γ(t) (as i → ∞, for every t in the domain), then lim inf L(γi ) ≥ L(γ).

Proof. (i) Indeed, the triangle inequality implies that Σ(Y ) ≥ d(γ(a), γ(b))
for all Y ’s, and thus the inequality persists under passing to the limit.
(ii) First notice that if Y ′ is obtained from Y by adding one point, then
Σ(Y ) ≥ Σ(Y ) (by the same triangle inequality). Adding c to a partition Y
of [a, b] and then splitting it into two partitions of [a, c] and [c, b] completes
the argument.
We repeat again that the readers should try to consider such lemmas
as exercises and try to prove them on their own. If reading the proof was
needed, then drawing a figure with all notations is a must!
(iii) We prove the continuity of L in d, a < d ≤ b, from the left (the
other cases are analogous and are left to the reader). Take ε > 0 and
consider a partition Y such that L(γ) − Σ(Y ) < ε. One may suppose that
yj−1 < d = yj . Then
L(γ, yj−1 , d) − d(γ(yj−1 ), γ(d)) < ε,
and the same inequality takes place for each c such that yj−1 ≤ c ≤ d.
(iv) Let paths γj converge pointwise to γ. Take ε > 0 and fix a partition
Y for γ such that L(γ) − Σ(Y ) < ε. Now consider the sums Σj (Y ) for paths
γj corresponding to the same partition Y . Choose j to be so large that the
inequality d(γj (yi ), γ(yi )) < ε holds for all yi ∈ Y . Then
L(γ) ≤ Σ(Y ) + ε ≤ Σj (Y ) + ε + (N + 1)ε ≤ L(γj ) + (N + 2)ε.
Since ε is arbitrary, this implies (iv). ¤
Remark 2.3.5. In general, functional L is not continuous. A stairs-like
example is shown in Figure 2.4.
36 2. Length Spaces

PSfrag replacements
A
BA

Figure 2.4: L is not continuous.

2.3.3. Induced intrinsic metric. A metric d induces a length structure.

The latter, in its turn, gives rise to an intrinsic metric (on each component of
accessibility by rectifiable paths). Thus we obtain a canonical construction
of induced intrinsic metrics
(X, d) → (X, d)b

where db = dLd .
Exercise 2.3.6. Prove that the intrinsic metric induced by the restriction
of Euclidean distance to the circle x2 + y 2 = 1 is the angular metric.

Note that the topology of the induced intrinsic metric may be very poorly
connected with the original topology of the space, as can be seen from the
following examples-exercises:
Exercise 2.3.7. Find the induced intrinsic metric for the metric
p
d((x1 , y1 ), (x2 , y2 )) = |x1 − x2 | + |y1 − y2 |
on R2 . What is the topology of the resulting length space?
Answer: A continuum of disjoint real lines, each with its standard metric.
Exercise 2.3.8. Consider the union of segments
∞
[
U= [(0, 1), (1/n, 0)] ∪ [(0, 1), (0, 0)] ⊂ R2 .
n=1
The sequence of points {(1/n, 0)} converges to (0, 0) in the topology inher-
ited by U from R2 . Since all pairwise distances between these points in the
induced intrinsic metric are at least 2, this sequence diverges with respect
to the intrinsic metric. Prove these statements.
Exercise 2.3.9. Connecting the point (0, 1) ∈ R2 with all points of the
standard Cantor set in the segment [0, 1] = [0, 1] × {0} ⊂ R2 , one obtains a
2.3. Length Structures Induced by Metrics 37

compact connected set. Show that in the induced intrinsic metric this set is
noncompact although it is still connected.
Exercise 2.3.10. Begin with a simple nonrectifiable curve in R2 ⊂ R3 and
build a cone over its image by choosing a point (vertex of the cone) in
R3 and connecting the vertex with every point in the image of the curve
by a segment. The intrinsic distance in this cone induced by Euclidean
distance in R3 is finite for every two points: one can go from one point to
the vertex of the cone along a straight segment and then get to the other
point along another segment. Show that removing the vertex of the cone
makes it disconnected in the topology of induced intrinsic metric, while it is
still connected in usual topology.
You can begin with a simple curve whose restriction to any nontrivial
interval is nonrectifiable (prove that such a curve does exist). In the original
topology this cone is still homeomorphic to a disc. Prove that in the induced
intrinsic metric this cone is homeomorphic to the bouquet of a continuum
of intervals (that is, the disjoint union of segments glued at one point).
Exercise 2.3.11. For two vectors V, W ∈ R2 , set
¯ ¯ p
d(V, W ) = ¯|V | − |W |¯ + min(|V |, |W |) · ∡(V, W ),
where ∡(V, W ) denotes the angle between V and W . Prove that
(i) the topology determined by d is the standard Euclidean one;
(ii) the induced intrinsic metric db is
(¯ ¯
¯|V | − |W |¯ if ∡(V, W ) = 0,
b
d(V, W ) = ¯ ¯
¯|V | + |W |¯ otherwise;

(iii) (R2 , db) is homeomorphic to the bouquet of a continuum of rays.

One can consider the length Ldb induced by the new metric d, b and this
length in turn determines a “second-stage” intrinsic metric. The reader may
imagine how much confusion between various lengths and metrics might
arise. However, this is not the case, as the length induced by db is the same
as induced by d.
Proposition 2.3.12. Let (X, d) be a metric space and db be the intrinsic
metric induced by d.
(1) If γ is a rectifiable curve in (X, d), then Ldb(γ) = Ld (γ).
(2) The intrinsic metric induced by db coincides with d.b In other words,
inducing a length metric is an idempotent operation.

Proof. The fact that the length of every curve in (X, d) is not less than the
distance between its endpoints implies that db ≥ d. It follows immediately
38 2. Length Spaces

that Ldb(γ) ≥ Ld (γ). To prove the inverse inequality, let [a, b] be the domain
of γ and let Y = {yi } be an arbitrary partition of [a, b]. Observe that
b
d(γ(y i ), γ(yi+1 )) ≤ Ld (γ, yi , yi+1 ) because the left-hand value is the infimum
of lengths one of which is written on the right-hand side. Therefore
X
Σdb(Y ) = b
d(γ(y i ), γ(yi+1 )) ≤ Ld (γ).

Since Y is an arbitrary partition, the inequality Ldb(γ) ≤ Ld (γ) follows.

This proves the first statement of the proposition. The second one is a
trivial consequence. ¤
Remark 2.3.13. The assumption that the curve γ is rectifiable is essential,
simply because otherwise it may fail to be continuous in (X, d). b The set
of continuous curves in (X, d)b is generally a subset of the respective set for
(X, d) but it contains all rectifiable curves. See Exercises 2.1.4 and 2.1.5.

2.4. Characterization of Intrinsic Metrics

A metric was said to be intrinsic if it can be obtained by a certain construc-
tion. In this chapter we discuss properties that distinguish intrinsic metrics
among all metrics and criteria that tell whether a given metric is intrinsic
or not.
Although we usually allow infinite distances, many statements here and
in the next section are obviously valid for finite metrics only. This condition
has to be added where necessary.

2.4.1. Another definition of length spaces. The second statement of

Proposition 2.3.12 gives a “constructive” criterion to find out whether a
given metric can be obtained as an induced intrinsic one. Namely a metric is
induced if and only if it induces itself. The following proposition generalizes
this for arbitrary intrinsic metrics.
Proposition 2.4.1. Let (X, d) be a length space and db be the intrinsic metric
induced by d. Then db = d.

Proof. Let L be the length function that defines d and Ld be the length
induced by d. Observe that Ld (γ) ≤ L(γ) for any admissible curve γ of finite
length (repeat the respective argument from the proof of Proposition 2.3.12).
Obviously a smaller length function determines a smaller metric; thus db ≤ d.
On the other hand, we already know that db ≥ d; hence db = d. ¤

Note that the equality d = db automatically implies that d is an intrinsic

metric—just because the induced metric db is always intrinsic. Hence it can
be considered as an alternative definition of the term “intrinsic metric”.
In other words, (X, d) is a length space if and only if for any points
2.4. Characterization of Intrinsic Metrics 39

x, y ∈ X and any ε > 0 there exists a curve γ connecting x and y such

that Ld (γ) < d(x, y) + ε.
If one considers an intrinsic metric d and it does not matter which length
structure determines it, Proposition 2.4.1 allows us to assume that the length
structure is the one induced by d and therefore use all properties of induced
length that we established in previous sections.

2.4.2. Recovering a length structure. Now we are ready to answer

another natural question, namely: given an intrinsic metric space (X, d),
how can we recover the initial length structure L. In fact, recovering the
length structure is not possible without additional assumptions because the
same intrinsic metric usually can be obtained from many different length
structures.

Exercise 2.4.2. Give an example of a length structure on the plane for

which all continuous curves are admissible, the resulting intrinsic metric is
the standard Euclidean one, but lengths of some curves differ from their
Euclidean lengths.

Proposition 2.4.1 shows a natural candidate for the length structure.

Namely, the length structure Ld induced by the metric indeed gives rise to
the original metric. We will therefore reformulate our question: under what
assumptions is L = Ld ? This kind of question is also important for the
following reason: in many examples the initial length function L has many
specific features that one may want to exploit. For example, the definition
of Finslerian length as the integral of “speed” (see example in Section 2.2) is
much more suitable for computing actual lengths than the general definition
of induced length applied to the resulting Finslerian metric. If coincidence
of two lengths is known, one can combine specific features of the initial
function L with general properties of lengths induced by metrics.
Certainly, two length structures may be different just because they have
different classes of admissible paths. Thus, we care only for their values on
the paths admissible for L. Then we notice that, besides all properties from
the definition of length structures, Ld possesses an additional property of
lower semi-continuity (see Proposition 2.3.4). Indeed, lower semi-continuity
turns out to be the key property here:

Theorem 2.4.3. If L is a lower semi-continuous length structure, then L

coincides with the length structure induced by its intrinsic metric d = dL on
all curves admissible for L: L(γ) = Ld (γ). As usual, the semi-continuity
means that if a sequence of paths γi (t) pointwise converges to γ(t), then
lim inf L(γi ) ≥ L(γ).
40 2. Length Spaces

Proof. The inequality Ld (γ) ≤ L(γ) holds for any length structure—see
Propositions 2.3.12 and 2.4.1. Let us prove the opposite inequality. By
property 2 of length structure, the function L(t) = L(γ|[a,t] ) is uniformly
continuous in [a, b] for each rectifiable curve γ : [a, b] → X. Hence for
every ε > 0, there exists a partition a = t0 ≤ t1 ≤ · · · ≤ tk+1 = b such
that dL (γ(ti ), γ(ti+1 )) < ε for every integer i between 0 and k. According
to the definition of dL , for each i = 0, 1, . . . , k there exists a curve σi :
[ti , ti+1 ] → X with endpoints σi (ti ) = γ(ti ), σi (ti+1 ) = γ(ti+1 ) such that
L(σi ) ≤ d(γ(ti ), γ(ti+1 )) + ε/k. For the concatenation hε of the curves σi we
have
k
X k
X
L(hε ) = L(σi ) ≤ dL (γ(ti ), γ(ti+1 )) + ε ≤ Ld (γ) + ε.
i=0 i=0

From the triangle inequality one readily sees that d(γ(t), hε (t)) ≤ 3ε for
every t ∈ [a, b]. It follows that hε (t) → γ(t) since the topology determined
by d is finer than the initial one. Now the lower semi-continuity for L implies
L(γ) ≤ lim inf L(hε ) ≤ Ld (γ),
ε→0

proving the theorem. ¤

Example 2.4.4. Let us come back to the example from Section 2.2 entitled
”Finslerian length”. We have chosen a function f in two vector-valued
variables and measured the length of a path γ by the formula
Z b
L(γ) = f ((γ(t)), γ ′ (t))dt.
a
Let us take a function f that does not depend on the first argument and
such that f (x, (1, 0)) = f (x, (0, 1)) = 1/10 and f (x, (1, 1)) = 1. (The fact
that f is independent of the first argument intuitively means that the cost
of out travel depends on the direction only, and it does not depend on
a particular location.) This Finslerian length structure is not lower semi-
continuous. To see this, consider a sequence of (stairs-like) broken lines with
edges parallel to coordinate axes and approaching the segment [(0, 0), (1,√ 1)].
The length of this segment (with respect to the Finslerian structure) is 2,
while the length of each of the broken lines is 1/5. The reason why this
has happened is that we made our travel in the diagonal direction “too
inexpensive” compared with the coordinate directions.
Exercise 2.4.5. Show that the Finslerian length structure constructed by a
function f (x, v) = F (v) is lower semi-continuous if and only if F (v) satisfies
the subadditivity assumption from Definition 1.2.11 of norm: F (v + w) ≤
F (v) + F (w). Show that this is also true for general f (x, v): the lower
semi-continuity of the corresponding length structure is equivalent to the
2.4. Characterization of Intrinsic Metrics 41

assumption that f satisfies the inequality f (x, v + w) ≤ f (x, v) + f (x, w).

This is the reason why, in defining Finslerian structures, one usually requires
f (x, ·) to be a norm for each x.
Exercise 2.4.6. Let d be a Finslerian metric (that is, the metric associated
with a Finslerian length structure) on a domain D ⊂ Rn . Prove that the
topology of d coincides with the standard Euclidean one.

2.4.3. Existence of midpoints.

Definition 2.4.7. A point z ∈ X is called a midpoint between points x, y
in a metric space (X, d) if d(x, z) = d(z, y) = 12 d(x, y).

The following lemma formulates a necessary condition for a metric to be

strictly intrinsic (this condition, under mild assumptions, also turns out to
be sufficient).
Lemma 2.4.8. If d is a strictly intrinsic metric, then for every two points
x, y there exists a midpoint z.

Proof. The length of a shortest path γ : [a, b] → X between x and y

is L(γ) = d(x, y). Denote L(t) = L(γ|[a,t] ). Since L(t) is continuous
in t and L(0) = 0, there is a c ∈ [a, b] such that L(c) = 12 L(b). Now
choosing z = γ(c) and using the fact that the length of a path is not
less than the distance between its endpoints, one immediately sees that
d(x, z) = d(y, z) = 1/2d(x, y). ¤

In other words, the previous lemma tells us that if d is strictly intrin-

sic, then the (closed) balls B d(x,y)/2 (x) and B d(x,y)/2 (y) have a nonempty
intersection.
Exercise 2.4.9. Show that, for a strictly intrinsic metric d, if r1 + r2 =
d(x, y), then the balls B r1 (x) and B r2 (y) have a nonempty intersection.

For intrinsic metrics, an analogous lemma asserts:

Lemma 2.4.10. If d is an intrinsic metric, then, given a positive ε, for
every two points x, y ∈ X there exists an ε-midpoint z, that is, a point z
such that |2d(x, z) − d(x, y)| ≤ ε and |2d(y, z) − d(x, y)| ≤ ε. In other words,
if 2r > d(x, y), then the balls Br (x) and Br (y) have a nonempty intersection.

Proof. Repeat the same arguments as in the proof of the previous lemma
for a path γ connecting x and y, and such that L(γ) − d(x, y) ≤ ε. ¤
Exercise 2.4.11. Let d be an intrinsic metric. Show that, if r1 + r2 >
d(x, y), then the balls Br1 (x) and Br2 (y) have a nonempty intersection.
42 2. Length Spaces

We leave as an exercise the following corollary. In a sense it allows

one to measure distances using (sufficiently fine) “dotted lines” between
two points instead of measuring distances by means of connecting points by
paths (compare also with Definition 2.3.1).
Corollary 2.4.12. Given a positive ε and two points x, y ∈ X in a space
with a strictly intrinsic metric d, there exists a finite sequence of points
x1 = x, x2 , . . . , xk = y such that every two neighboring points in this
sequence are ε-close (that is, d(xi , xi+1 ) ≤ ε for all I = 1, . . . , k − 1) and
Pk−1
i=1 d(xi , xi+1 ) = d(x, y).

For intrinsic metrics the last formula in the corollary should be replaced
Pk−1
by i=1 d(xi , xi+1 ) − d(x, y) ≤ ε.
Exercise 2.4.13. If x and y are two points in a length space (X, d) and
r < d(x, y), then dist(y, Br (x)) = d(x, y) − r. Prove this.
Exercise 2.4.14. Let X be a length space, Y a metric space, and let a map
f : X → Y be locally Lipschitz with a Lipschitz constant C. Prove that f
is Lipschitz with the same constant.
Exercise 2.4.15. Let (X, d) be a length space and A a connected open
subset of X. Then d induces on A the (finite-valued) intrinsic metric dA .
Moreover each point p ∈ A has a neighborhood U ⊂ A such that for any
points p, q ∈ U we have d(p, q) = dA (p, q).

2.4.4. Complete intrinsic metrics. In many cases the converse of

Lemma 2.4.8 is true: the existence of midpoints (resp. ε-midpoints) im-
plies that a complete metric space is strictly intrinsic (resp. intrinsic). Thus
we have a criterion that tells us whether a complete metric space is a length
space.
Theorem 2.4.16. Let (X, d) be a complete metric space.
1. If for every x, y ∈ X there exists a midpoint, then d is strictly
intrinsic.
2. If for every x, y ∈ X and every positive ε there exists an ε-midpoint,
then d is intrinsic.

This theorem has the following immediate corollary, which can be used
as an alternative criterion for a metric to be intrinsic:
Corollary 2.4.17. A complete metric space (X, d) is a length space iff,
given a positive ε and two points x, y ∈ X, there exists a finite sequence
of points x1 = x1 , x2 , . . . , xk = y such that every two neighboring points in
this sequence are ε-close (i.e., d(xi , xi+1 ) ≤ ε for all i = 1, . . . , k − 1) and
Pk−1
i=1 d(xi , xi+1 ) < d(x, y) + ε.
2.4. Characterization of Intrinsic Metrics 43

This corollary says that a metric is intrinsic if and only if, given two
points and a positive ε, one can reach one of the points starting from the
other one and hopping with jumps shorter than ε and with the total length
of the jumps not exceeding the distance between the points plus ε.

Proof of Theorem 2.4.16. To prove that a metric is intrinsic we have to

show that for any two points x and y there are paths connecting x and
y whose lengths approximate d(x, y) with any given precision. In case of
strictly intrinsic metric, there must be a path whose length is equal to
d(x, y). We proceed with the case of strictly intrinsic metrics; modifying
this argument for the other case is left as an exercise.
We will construct a path γ : [0, 1] → X between x and y such that
γ(0) = x, γ(1) = y and L(γ) = d(x, y). First we assign the values of γ for
all dyadic rationals (rational numbers of the form k/2m for some natural
numbers k, m). Then we extend this partially defined map by continuity;
only this step of the argument will use the completeness of (X, d). Indeed, a
path with the desirable properties must pass through a midpoint between x
and y. Since such midpoints exist by the assumption of the theorem, choose
such a midpoint and assign it to be the image γ(1/2). Now we assign γ(1/4)
to be a midpoint between x = γ(0) and γ(1/2) and γ(3/4) to be a midpoint
between γ(1/2) and y = γ(1). Proceeding this way, we define γ for all dyadic
rationals between 0 and 1.
According to our construction, for every two dyadic rationals ti , tj

(2.1) d(γ(t), γ(t′ )) ≤ |t − t′ | · d(x, y).

This inequality implies that the map γ, defined on the set of dyadic
rationals, is Lipschitz. Since X is complete and the set of dyadic rationals
is dense in [0, 1], this map can be extended to the entire interval [0, 1] (cf.
Proposition 1.5.9). Thus we obtained a path γ : [0, 1] → X connecting x
and y. Then (2.1) implies that L(γ) = d(x, y). ¤

To see how these results work, we suggest several exercises:

Exercise 2.4.18. Prove that the completion of a length space is a length

space.

Exercise 2.4.19. Let X be a compact topological space and let {dn }∞ n=1
be a sequence of intrinsic metrics on X that uniformly converge to a metric
d (recall that metrics are functions on X × X, so the notion of uniform
convergence applies here). Prove that d is intrinsic too.
44 2. Length Spaces

2.5. Shortest Paths

2.5.1. Curves and natural parameterizations. It is important to re-
member that, when speaking about paths, we do mean maps and not their
images. Indeed, an image of a continuous path may fill a disc; two paths
making different number of rounds around a circle are essentially different
and in particular have different lengths. Still, changing the parameter by a
strictly increasing change of variable means that we visit the same collection
of points in the same order; in other words, we traverse the same “curve” in
the same direction. One expects that such changes of variable do not change
geometric properties.
This suggests the idea of a curve as a class of equivalent paths with
respect to the following relation: paths γ1 : I1 → X and γ2 : I2 → X are
equivalent if there exists a strictly increasing continuous map ϕ from I1 onto
I2 such that γ1 = γ2 ◦ ϕ. (Check that this indeed is an equivalence relation.)
However such a definition would be too restrictive for our purposes. If a path
is constant on some subinterval (i.e., it “stops for a while” at a point), so
does any path obtained from this one by a change of variable. We want such
a path to be equivalent to one that goes the same way except that it passes
through the point without stopping. To achieve this, we allow nonstrictly
monotone changes of variable. This is formalized by the following

Definition 2.5.1. An (unparameterized) curve is an equivalence class of

the minimal equivalence relation satisfying the following: paths γ1 : I1 → X
and γ2 : I2 → X are equivalent whenever there exists a nondecreasing
continuous map ϕ from I1 onto I2 such that γ1 = γ2 ◦ ϕ.
Paths (representatives of an equivalence class) are also called parame-
terizations of the curve and re-parameterizations of one another.

The term “curve” is used for both unparameterized curves and their
parameterizations (i.e., paths). In most cases, “curve” is formally a synonym
for “path”. However, the former is more appropriate when parameterization-
independent properties are considered.

Remark 2.5.2. Existence of a nonstrictly monotone change of variable is

not an equivalence relation by itself (due to the lack of inverse changes of
variable). This is why we consider the equivalence relation generated by it.
In other words, two paths γ and γ̄ are equivalent (represent the same curve)
if and only if there exists a finite sequence of paths γ1 , γ2 , . . . , γn such that
γ1 = γ, γn = γ̄ and for every i = 1, . . . , n − 1 either γi is obtained from γi+1 ,
or γi+1 is obtained from γi , by a nondecreasing change of variable. This
description can be simplified by means of the following exercise.
2.5. Shortest Paths 45

Exercise 2.5.3. A path γ : I → X is said to be never-locally-constant if

there exists no interval [a, b] ⊂ I such that a 6= b and the restriction of γ to
[a, b] is a constant map. Prove that
(a) Every curve admits a never-locally-constant parameterization.
Hint: Formally, for a path γ : I → X introduce an equivalence relation
∼ on I: y ∼ y ′ if and only if γ is constant on [y, y ′ ]. Show that the quotient
J = I/ ∼ is homeomorphic to an interval. Then observe that there is a
unique map γ̃ : J → X such that γ = γ̃ ◦ π where π is the canonical
projection I on J, and that γ̃ is never-locally-constant and continuous.
(Loosely speaking, J is obtained from I by cutting off all intervals where a
path is constant and gluing together the ends of each of the resulting gaps.)
(b) Two paths γ1 : I1 → X and γ2 : I2 → X are equivalent if and only
if there exist a path γ : J → X and changes of variable ϕ1 : I1 → J and
ϕ2 : I2 such that γi = γ ◦ ϕi .
Hint: Let γ be a never-locally-constant parameterization.
(c) Two paths γ1 : I1 → X and γ2 : I2 → X are equivalent if and only if
there exist an interval J and changes of variable ϕi : J → Ii (i = 1, 2) such
that γ1 ◦ ϕ1 = γ2 ◦ ϕ2 .

It is easy to see that all parameterizations of a curve have equal lengths

(check this). We will denote a curve and its parameterization by the same
letter.

Definition 2.5.4. A curve γ : [a, b] → X is called simple if the pre-image

of every point is an interval.

This means that a simple path is allowed to stop at a point for a while,
but having left a point it never comes back. Roughly speaking, the image
of a simple path is a curve without ”self-intersections”. The following very
easy exercise justifies the correctness of the definition:

Exercise 2.5.5. If one parameterization of a curve is simple, then so are

all other parameterizations.

Exercise 2.5.6. If two simple curves have the same image, then they are
equivalent up to a change of variable t 7→ −t.

Our next goal is to choose our favorite parameterization for every curve.
These parameterizations are analogous to motion with unit speed in physics
or differential geometry:

Definition 2.5.7. A parameterization γ : I → X is natural if L(γ, t, t′ ) =

t − t′ for all t, t′ ∈ I.
46 2. Length Spaces

Remark 2.5.8. To verify that a parameterization γ is natural, it suffices

to check that L(γ, a, t) = t − a for a fixed a and all t. This follows from the
formula L(γ, t, t′ ) = L(γ, a, t′ ) − L(γ, a, t).

Other names for natural parameterization are arc-length parameteriza-

tion and parameterization by arc length. In other words, a parameterization
γ(t) is natural when
d
L(γ, a, t) = 1,
dt
and one also calls it a unit speed parameterization. More generally, one says
that a parameterization γ is of constant speed v if L(γ, t, t′ ) = v(t − t′ ) for
all t, t′ .
The following proposition tells us that every curve admits a natural
parameterization.
Proposition 2.5.9. Every rectifiable curve γ : [a, b] → X can be represented
in the form γ = γ̄ ◦ ϕ where γ̄ : [0, L(γ)] → X is a natural parameterization
and ϕ is a nondecreasing continuous map from [a, b] onto [0, L(γ)].

Proof. The idea of construction of ϕ is trivial: let γ̄(τ ) be the point on (the
image of) γ such that the length of the interval of γ between its origin and
that point is equal to τ . This is formalized as follows. Define ϕ(t) = L(γ, a, t)
for all t ∈ [a, b]. Then the function ϕ is nondecreasing and continuous (by the
continuity property of length). The set of its values is the interval [0, L(γ)].
Now for every τ ∈ [0, L(γ)] pick a t ∈ [a, b] such that ϕ(t) = τ and define
γ̄(τ ) = γ(t). This definition does not depend on the choice of t. Indeed, if
ϕ(t) = ϕ(t′ ), then γ(t) = γ(t′ ) because L(γ, t, t′ ) = ϕ(t′ ) − ϕ(t) = 0.
Thus we have defined a map γ̄ : [0, L(γ)] → X. The relation γ = γ̄ ◦ ϕ
follows immediately from the definition. It remains to verify that γ̄ is
continuous and parameterized by arc length. For the former, let τ1 = ϕ(t1 )
and τ2 = ϕ(t2 ). Then γ̄(τ1 ) and γ̄(τ2 ) are the endpoints of the path γ|[t1 ,t2 ] .
The length of this path is L(γ, t1 , t2 ) = ϕ(t2 ) − ϕ(t1 ) = τ2 − τ1 . Since
the distance between endpoints is no greater than the length, we obtain
that d(γ̄(τ1 ), γ̄(τ2 )) ≤ |τ1 − τ2 |. This means that γ̄ is a nonexpanding
and hence continuous map. Furthermore γ|[t1 ,t2 ] is a re-parameterization
of γ̄|[τ1 ,τ2 ] and hence L(γ̄, τ1 , τ2 ) = L(γ, t1 , t2 ) = τ2 − τ1 . Thus γ̄ is a natural
parameterization. ¤
Exercise 2.5.10. Prove that a natural parameterization of a curve is unique
up to a translation [a, b] → [a + c, b + c] : t 7→ t + c of the variable.
Exercise 2.5.11. Given v > 0, any rectifiable curve admits a parameteri-
zation with constant speed v.

The following exercise is a (trivial) corollary of Proposition 2.5.9:

2.5. Shortest Paths 47

Exercise 2.5.12. If a length space is homeomorphic to a segment, then it

is isometric to a segment.

This corollary tells us a remarkable thing: one-dimensional intrinsic

geometry is trivial since all intrinsic metrics on a line are locally indistin-
guishable! We will see that already two-dimensional surfaces are completely
different in this respect. By the way, although there is an essentially unique
intrinsic metric on R, there are lots of different
p metrics, which are not in-
trinsic. For instance, consider d(x, y) = |x − y| (you may replace the
square root by other concave functions to get more examples). The intrinsic
distance induced by this metric is infinite everywhere.

2.5.2. Existence of shortest paths. The goal of this section is to prove

that a complete locally compact length space is strictly intrinsic, that is,
there is a shortest path between every two points.
We begin with the definition of uniform convergence for curves:
Definition 2.5.13. A sequence of curves uniformly converges to a curve
γ if they admit parameterizations (with the same domain) that uniformly
converge to a parameterization of γ.

The following theorem shows that the space of curves of uniformly

bounded lengths in a compact space is compact with respect to the above
convergence. This is a version of the Arzela–Ascoli Compactness Theorem
in functional analysis.
Theorem 2.5.14 (Arzela–Ascoli Theorem). In a compact metric space,
any sequence of curves with uniformly bounded lengths contains a uniformly
converging subsequence.

Proof. For each γi , there is a unique constant speed parameterization by

the unit interval [0, 1]. Uniform boundedness of the lengths of γi means that
the speeds of these parameterizations are uniformly bounded. In its turn,
this implies that for some C < ∞
(2.2) d(γi (t), γi (t′ )) ≤ L(γ, t, t′ ) ≤ C|t − t′ |
for every integer i and all t, t′ ∈ [0, 1].
Let S = {tj } be a countable dense subset of [0, 1]. Using the Cantor
diagonal process one can find a subsequence γni of {γi } such that for
each j ∈ N the sequence γni (tj ) converges. We plan to show that the
subsequence γni itself converges; to avoid double indices, we may assume
(without loss of generality) that this subsequence is γi itself: for every tj the
limit limi→∞ γi (tj ) exists.
48 2. Length Spaces

To prove that the sequence {γi (t)} converges for every t ∈ [0, 1], we will
show that this is a Cauchy sequence.
Given ε > 0, choose tj ∈ S such that |t − tj | < ε/C and then N ∈ N
such that d(γi (tj ), γk (tj )) < ε for all i, k > N . For these choices
d(γi (t), γk (t)) ≤ d(γi (t), γi (tj )) + d(γi (tj ), γk (tj )) + d(γk (tj ), γk (t)) ≤ 3ε.
This proves that γi (t) is a Cauchy sequence and hence we can define
γ(t) = limj→∞ γj (ti ).
Passing to the limit in (2.2) we get
(2.3) d(γ(t), γ(t′ )) ≤ C|t − t′ |,
and thus γ is a continuous map.
Let us show that γi converges to γ uniformly. Given ε > 0, choose N >
4C
ε and let M be such that d(γ(k/N ), γi (k/N )) < ε/2 for all k = 0, 1, . . . , N
and all i > M . This choice is possible since γi converges to γ point-wise.
Combining (2.2) with (2.3), for every 0 ≤ t ≤ 1 and k/N ≤ t ≤ (k + 1)/N
we have
d(γ(t), γi (t)) ≤ C|t − k/N | + ε/2 + C|t − k/N | ≤ ε
for all i > M . This concludes the proof. ¤

Although we already used the notion of shortest paths, this cornerstone

notion deserves a formal definition:
Definition 2.5.15. A curve γ : [a, b] → X is a called a shortest path if its
length is minimal among the curves with the same endpoints, in other words
L(γ1 ) ≥ L(γ) for any curve γ1 connecting γ(a) and γ(b).

It is trivial that any interval of a shortest path is a shortest path (check

this).
Remark 2.5.16. In a length space the above definition can be reformulated
as follows: a curve γ : [a, b] → X is a shortest path if and only if its length is
equal to the distance between its endpoints: L(γ) = d(γ(a), γ(b)). Shortest
paths in length spaces are also called distance minimizers.

Shortest paths in length spaces possess some nice properties that do not
hold in general metric spaces. One such property is the following
Proposition 2.5.17. If shortest paths γi in a length space (X, d) converge
to a path γ as i → ∞, then γ is also a shortest path.

Proof. Since the endpoints of γi converge to endpoints of γ and the length

of each γi is equal to the distance between its points, we conclude that
2.5. Shortest Paths 49

L(γi ) → d(x, y), where x, y are the endpoints of γ. By the lower semi-
continuity of length,
L(γ) ≤ lim L(γi ) = d(x, y).
i→∞

¤
Exercise 2.5.18. Give an example showing that a limit of shortest paths
in a metric space may fail to be a shortest path.

The example of R2 \ {0} shows that there may be no shortest path

between two points. On the other hand, there may be several different
shortest paths between the same two points: for instance, consider two
antipodal points in a sphere.
Convention. We use the notation [x, y] to denote a shortest path
between points x and y. This notation is convenient when either such
shortest path is unique or it does not matter which of the shortest paths is
considered. This notation is well compatible with our notation for segments
in Euclidean space, since the latter are just shortest paths with respect to
Euclidean distance.
Proposition 2.5.19. Let (X, d) be a compact metric space and let x, y ∈ X
be points that can be connected by at least one rectifiable curve. Then there
exists a shortest path between x and y.

Proof. Let Linf denote the infimum of lengths of rectifiable curves con-
necting x and y. Then there exists a sequence {γi } of such curves with
L(γi ) → Linf . According to Theorem 2.5.14, the sequence {γi } contains a
converging subsequence. Without loss of generality we may assume that
{γi } itself converges to a curve γ. Then γ has the same endpoints and,
by the lower semi-continuity of length, L(γ) ≤ lim L(γi ) = Linf . Thus
L(γ) = Linf . ¤
Corollary 2.5.20. Let (X, d) be a boundedly compact metric space. Then
for every two points x, y ∈ X connected by a rectifiable curve there exists a
shortest path between x and y.

Proof. Let L be a length of some rectifiable curve connecting x and y. Ob-

serve that this curve, as well as any shorter curve with the same endpoints,
is contained in the closed metric ball of radius L centered at x. So it is suf-
ficient to prove the existence of a shortest path only inside this ball. Since
the ball is compact, this follows from the previous proposition. ¤
Definition 2.5.21. A topological space X is called locally compact if every
point of X has a pre-compact neighborhood.
50 2. Length Spaces

Proposition 2.5.22. If (X, d) is a complete locally compact length space,

then every closed ball in X is compact (and hence X is boundedly compact).

Note that in this proposition it is essential that X is a length space. For

example, a space where all distances between points equal to 1 is locally
compact and complete, but a closed unit ball in such a space is not compact
unless the space has a finite cardinality.

Proof of Proposition 2.5.22. Let x ∈ X be an arbitrary point. Observe

that if B̄r (x) is compact for some r, then B̄ρ (x) is compact for any ρ < r.
Define
R = sup{r > 0 : B̄r (x) is compact}.
Since x has a pre-compact neighborhood, we have R > 0. Suppose that
R < ∞ and denote the ball B̄R (x) by B.
First let us prove that B is compact. Since B is a closed set in a complete
space, it suffices to prove that for any ε > 0 it contains a finite ε-net. We
may assume that ε < R. Let B ′ denote the ball B̄R−ε/3 (x). This ball is
compact and hence it contains a finite (ε/3)-net S. Let y ∈ B. Since X
is a length space, we have dist(y, B ′ ) ≤ ε/3. Therefore there exists a point
y ′ ∈ B ′ with d(y, y ′ ) < ε/2. On the other hand, dist(y ′ , S) ≤ ε/2; hence
dist(y, S) < ε. This means that S is an ε-net for B, and we have proven the
compactness of B.
Every point y ∈ B has a pre-compact neighborhood Uy . Pick a finite
collection
S {Uy }y∈Y of such neighborhoods that cover B. Their union U =
y∈Y Uy is a pre-compact neighborhood of B. Using the compactness of
B again, we can conclude that there exists a positive ε > 0 such that the
ε-neighborhood of B is contained in U . Since X is a length space, the ε-
neighborhood of B is the ball BR+ε (x), and its closure is B̄R+ε (x). Therefore
B̄R+ε (x) is compact. This contradicts the choice of R.
Thus the assumption R < ∞ is wrong; hence R = ∞. This means that
all balls centered at x are compact. ¤

Combining Proposition 2.5.22 and the previous Corollary 2.5.20, we

obtain the main result of this section.

Theorem 2.5.23. Let (X, d) be a complete locally compact length space.

Then this space is strictly intrinsic: for every x, y ∈ X such that d(x, y) < ∞
there exists a shortest path γ connecting x and y, i.e., a curve γ : [a, b] → X
such that γ(a) = x, γ(b) = y and L(γ) = d(x, y).

The example of R2 \ {0} shows that completeness in this theorem is

essential. So is local compactness:
2.5. Shortest Paths 51

Exercise 2.5.24. Give an example of a complete length space (with finite

metric) in which there is no shortest path between some points.
However, the condition that X is a length space in the above consider-
ations can be omitted, at the expense of more complicated formulations:
Exercise 2.5.25. Let X be a complete locally compact metric space (not
necessarily a length space). Prove that for every two points x, y ∈ X that
can be connected by a rectifiable curve, there exists a shortest path between
x and y.
Hint: Prove that for every R > 0 the set of points that can be connected
to x by a curve of length less than R is pre-compact. In other words, balls
of induced length metric are pre-compact in the topology of the original
metric (!). To prove this, modify the proof of Proposition 2.5.22.
Exercise 2.5.26. Is it true that a completion of a locally compact length
space is locally compact?
2.5.3. Geodesics and the Hopf–Rinow Theorem.
Definition 2.5.27. Let X be a length space. A curve γ : I → X is
called a geodesic if for every t ∈ I there exists an interval J containing
a neighborhood of t in I such that γ|J is a shortest path. In other words,
a geodesic is a curve which is locally a distance minimizer (i.e., a shortest
path).
The example of the sphere shows that there are geodesics that are not
shortest paths: whereas every segment of a great circle on a sphere is a
geodesic, it fails to be shortest as soon as it is longer than half of the equator.
Although the spherical example suggests that a shortest path between two
points should be unique at least locally, this is also not true in general.
To see an example, consider the surface of a cube with its intrinsic metric
induced by the Euclidean metric of its ambient space. We suggest that the
reader shows that any neighborhood of a vertex contains points with multiple
shortest paths between them. Another natural conjecture that turns out to
be wrong is that a limit of geodesics is a geodesic; give a counterexample to
this on the surface of the cube. We will see later that these phenomena are
caused by nonsmoothness of the surface and that indeed such things never
happen to smooth examples.
It is clear that a shortest path always admits a natural parameterization,
and hence so does a geodesic.
Intuitively, a space is noncomplete if a point is missing. One may suspect
that this absence of a point can be noticed by moving along a geodesic that
would end at this point: the interval for which our motion is well-defined is
not closed. The theorem below formalizes this observation.
52 2. Length Spaces

Before formulating it, we generalize the notion of a shortest path so as

to include paths defined on nonclosed intervals. Namely we say that a curve
γ : I → X (where I ⊂ R is an interval and X is a metric space) is a shortest
path, or a minimal geodesic, if its restriction to any interval [a, b] ⊂ I is a
shortest path in the sense of Definition 2.5.15.
Theorem 2.5.28 (Hopf–Rinow–Cohn-Vossen Theorem). For a locally com-
pact length space X, the following four assertions are equivalent:
(i) X is complete.
(ii) X is boundedly compact, i.e., every closed metric ball in X is
compact.
(iii) Every geodesic γ : [0, a) → X can be extended to a continuous path
γ : [0, a] → X.
(iv) There is a point p ∈ X such that every shortest path γ : [0, a) → X
with γ(0) = p can be extended to a continuous path γ : [0, a] → X.

The theorem generalizes the classical Hopf-Rinow theorem which origi-

nally was proved only in smooth situation, i.e., for Riemannian manifolds.
Remark 2.5.29. By Theorem 2.5.23, these conditions imply that every two
points a, b can be connected by a shortest path.

Proof of the theorem. Implications (ii) =⇒ (i) =⇒ (iii) =⇒ (iv) are left
as easy exercises. We will prove that (iv) implies (ii). The proof uses the
same general scheme as the one of Proposition 2.5.22. The details are slightly
more complicated because we cannot utilize completeness at this point. We
can use only the property (iv) and this requires a more delicate argument.
Since X is locally compact, sufficiently small closed balls B r (p) are
compact. Reasoning by contradiction (that is, assuming that there are
noncompact closed balls), define
R = sup{r : B r (p) is a compact set }
and assume that R < ∞. The argument consists of two steps.
1. First, we prove that the open ball BR (p) is pre-compact. To do
this, it suffices to show that every sequence {xi } in this ball contains a
converging subsequence (whose limit does not necessary belong to this ball).
Set ri = d(p, xi ). One may assume that ri → R as i → ∞; otherwise a
subsequence of {xi } is contained in a ball B r (p) for some r < R, and then
there is a converging subsequence because this smaller ball is compact (by
the choice of R).
Let γi : [0, ri ] → X be a (naturally parameterized) shortest path connec-
tion p to xi . Such a shortest path exists because xi belongs to a compact
2.6. Length and Hausdorff Measure 53

ball centered at p (see the proof of Corollary 2.5.20). We can choose a sub-
sequence of {γi } such that the restrictions of the paths to [0, r1 ] converge
(along this subsequence). From this subsequence, we choose a further sub-
sequence of paths whose restrictions to [0, r2 ] converge, and so on. Then
the Cantor diagonal procedure (that is, picking the nth element from the
nth subsequence for n = 1, 2, . . . ) yields a sequence {γin } such that for
every t ∈ [0, R) the sequence {γin (t)} converges in X. (More precisely,
points γin (t) are well-defined for all large enough n and form a converging
sequence.)
Define γ(t) = lim γin (t); then γ : [0, R) → X is a nonexpanding map and
a shortest path (see Proposition 2.5.17). By the assertion (iv), there is a
continuous extension γ : [0, R] → X. One easily sees (exercise!) that the
points xin (i.e., the endpoints of our converging curves {γin }) converge to
γ(R).
2. Since the open ball BR (p) is pre-compact, the closed ball B R (p) is
compact. (Recall that a closed ball in a length space is the closure of the
respective open ball.) Now we show that a ball BR+ε (p) is pre-compact
for some ε > 0. Since X is locally compact, for every x ∈ B R (p) there
is an r(x) > 0 such that the ball Br(x) (x) is pre-compact. Choose a finite
subcover B r(xi ) (xi ) out of the cover of B R (x) by these balls. The union of
these balls is pre-compact and contains the ball BR+ε (p) for ε = min ri > 0.
This contradicts the choice of R. ¤

2.6. Length and Hausdorff Measure

Recall that the length of a curve is independent of the parameterization.
This fact suggests that the length of a simple curve can be recovered from
its image in the space, i.e., the set of points it passes through. In this
section we show that the length of a path actually equals the one-dimensional
Hausdorff measure of the image. We assume that the normalization constant
C(1) for Hausdorff measure is 1 (for a definition and elementary properties
of Hausdorff measure see Section 1.7).

Lemma 2.6.1. If X is a connected metric space, then µ1 (X) ≥ diam X.

Proof. 1. A general observation: in the definition of Hausdorff measure

we can restrict ourselves to coverings by open sets Si . Indeed, an arbitrary
covering {Si } can be replaced by an open covering {Si′ } where

Si′ = Uδ/2i (Si ) := {x ∈ X : dist(x, Si ) < δ/2i }

54 2. Length Spaces

for a small δ > 0. Then diam Si′ ≤ diam Si + 2δ/2i and hence w1 ({Si′ }) ≤
w1 ({Si }) + 2δ. Since δ is arbitrary, it follows that the measure can be ap-
proximated by 1-weights of open coverings. (Exercise: extend the argument
to work for measures of all dimensions.)
2. Let X be a connected topological space and {Si } an open covering
of X. Then for every two points x, y ∈ X there exists a finite sequence
Si1 , . . . , Sin of sets such that x ∈ Si1 , y ∈ Sin , and Sik ∩ Sik+1 6= ∅ for all k,
1 ≤ k ≤ n − 1. To prove this, fix a point x ∈ X and consider the set Y of
all points y ∈ X for which such a sequence exists. It is clear that for every
set Si one has either Si ⊂ Y or Si ⊂ X \ Y . Therefore both Y and X \ Y
are open sets. Since X is connected, it follows that Y = X and therefore
any point y ∈ Y is “accessible” by a sequence of sets as described above.
3. Let {Si } be an open covering of X, x and y two points of X, and
Si1 , . . . , Sin a sequence from Step 2. For k = 1, . . . , n − 1 pick a point
xk ∈ Sik ∩ Sik+1 , and define x0 = x, xn = y. Then |xk−1 xk | ≤ diam Sik for
all k = 1, . . . , n because both xk−1 and xk are in Sik . Hence
X n
X n
X
diam Si ≥ diam Sik ≥ |xk−1 xk | ≥ |x0 xn | = |xy|.
k=1 k=1
Due to the observation made in step 1, it follows that µ1 (X) ≥ |xy|. Since
x and y are arbitrary points, this means that µ1 (X) ≥ diam X. ¤
Theorem 2.6.2. Let X be a metric space, γ : [a, b] → X a simple curve.
Then L(γ) = µ1 (γ([a, b])).

Proof. Let S = γ([a, b]) and L = L(γ). Without loss of generality assume
that γ is parameterized by arc length: γ : [0, L] → X. Then for every
natural number N , S is covered by N intervals of γ each of length L/N ,
L L
namely by the sets γ([i N , (i + 1) N ]), i = 0, 1, . . . , N − 1. The diameters of
these sets are no greater than L/N . Hence the sum of the diameters is no
greater than L, whereas the diameters themselves approach 0 as N → ∞.
This shows that µ1 (S) ≤ L(γ).
On the other hand, for a partition a = t0 ≤ t1 ≤ · · · ≤ tn = b of [a, b],
let Si = γ([ti , ti+1 ]) for i = 1, . . . , n − 1. The sets Si are disjoint modulo
a finite number of points γ(ti ). Since the one-dimensional
P measure of a
single point is obviously zero, one has µ1 (S) = µ1 (Si ). By Lemma 2.6.1,
µ1 (Si ) ≥ diam Si ≥ |γ(ti )γ(ti+1 )|, and hence
X X
µ1 (S) = µ1 (Si ) ≥ |γ(ti )γ(ti+1 )|
for any partition {ti }. This implies that µ1 (S) ≥ L(γ). ¤
Remark 2.6.3. If γ is not simple, the same argument shows that L(γ) ≥
µ1 (γ([a, b])).
2.7. Length and Lipschitz Speed 55

Exercise 2.6.4. Prove that, for any curve γ : [a, b] → X,

X
L(γ) = k · µ1 ({x ∈ X : #(γ −1 (x)) = k})
k∈N∪{∞}

where # denotes the cardinality and ∞ · 0 = 0.

Hint: The right-hand part is the integral of the function x 7→ #(γ −1 (x))
with respect to the measure µ1 . Combine Levy’s limit theorem for integrals
with inequalities “diam ≤ µ1 ≤ L” applied to small pieces of γ.

2.7. Length and Lipschitz Speed

This section is rather technical. We will generalize the formula “length
equals the integral of speed” which is well known and trivial for smooth
curves in Rn (cf. Exercise 2.3.3). The last theorem of this section requires
the knowledge of Lebesgue integration, and we will use without proofs some
facts from measure theory.
Definition 2.7.1. Let (X, d) be a metric space and γ : I → X a curve. The
speed of γ at t ∈ I, denoted by vγ (t), is defined by
d(γ(t), γ(t + ε))
vγ (t) := lim
ε→0 |ε|
if the limit exists.
Exercise 2.7.2. Let γ be a differentiable curve in Rn (more generally, in a
normed vector space). Prove that vγ (t) exists for all t and vγ (t) = |γ ′ (t)|.
Exercise 2.7.3. Let X be a metric space and γ : [a, b] → X a curve.
Suppose that vγ (t) exists for all t ∈ [a, b] and is continuous in t. Prove that
Z b
L(γ) = vγ (t) dt
a

(compare Exercise 2.3.3).

Of course, the speed of a curve may not exist. However, it exists almost
everywhere (i.e., except a set of zero measure) for a wide class of curves,
namely for every Lipschitz curve. (Note that every rectifiable curve admits
Lipschitz parameterization. For example, all natural parameterizations are
Lipschitz with unit Lipschitz constant.) Moreover, the length of a Lipschitz
curve equals the (Lebesgue) integral of its speed.
As a first step, we prove the following
Theorem 2.7.4. Let (X, d) be a metric space, γ : [a, b] → X a rectifiable
curve. Then for almost all t ∈ [a, b] (i.e., for all t except a set of zero
56 2. Length Spaces

measure) the following holds: either

L(γ|[t−ε,t+ε′ ] )
lim inf = 0,
′
ε,ε →0+ ε + ε′
or
d(γ(t − ε), γ(t + ε′ ))
lim = 1.
′
ε,ε →0+ L(γ|[t−ε,t+ε′ ] )

One can let ε or ε′ in the above formulas be zero. This gives the following
Corollary 2.7.5. If γ is as in Theorem 2.7.4, then for almost all t ∈ [a, b]
either
L(γ|[t,t+ε] )
lim inf = 0,
ε→0 |ε|
or
d(γ(t), γ(t + ε))
lim =1
ε→0 L(γ|[t,t+ε] )
(if ε < 0, the interval [t, t + ε] in the denominator of the last formula should
be interpreted as [t + ε, t]).

Proof of Theorem 2.7.4. Suppose the contrary. For every α > 0 let Zα
denote the set of all t ∈ [a, b] such that
L(γ|[t−ε,t+ε′ ] )
lim inf >α
′
ε,ε →0+ ε + ε′
and
d(γ(t − ε), γ(t + ε′ ))
lim inf < 1 − α.
′
ε,ε →0+ L(γ|[t−ε,t+ε′ ] )
Then Sµ1 (Zα ) > S 0 for all sufficiently small α. Indeed, otherwise the set
Z0 = α>0 Zα = n∈N Z1/n would have zero measure, and this is equivalent
to the statement of the theorem. Fix an α > 0 such that µ1 (Zα ) > 0. For
brevity, we denote Z = Zα and µ = µ1 (Z). Choose ε0 so small that for
any partition {yi }Ni=1 (a = y0 ≤ y1 ≤ · · · ≤ yN = b) of [a, b] such that
maxi (yi − yi−1 ) < ε0 , one has
N
X
L(γ) − d(γ(yi−1 ), γ(yi )) < µα2 /2.
i=1
Such an ε0 exists by Exercise 2.3.2. Consider the set B of all intervals of
the form [t − ε, t + ε′ ] such that t ∈ Z, ε + ε′ < ε0 , L(γ|[t−ε,t+ε′ ] ) > α(ε + ε′ )
and
d(γ(t − ε), γ(t + ε′ )) < (1 − α)L(γ|[t−ε,t+ε′ ] ).
By the definition of Z = Zα , every point t ∈ Z is contained in arbitrarily
short elements of B. Applying Vitali’s covering theorem (Theorem 1.7.14)
2.7. Length and Lipschitz Speed 57

we can extract from B a countable collection {[ti − εi , ti + ε′i ]}∞ i=1 of disjoint
intervals that covers Z up to a set of zero measure. In particular,
X∞
¡[ ¢
(εi + ε′i ) = µ1 [ti − εi , ti + ε′i ] ≥ µ1 (Z) = µ.
i=1
Hence for a sufficiently large M ,
M
X
(εi + ε′i ) > µ/2.
i=1

Since the intervals {[ti − εi , ti + ε′i ]}M

i=1 are disjoint, they can be included
in a partition {yj }N j=1 all whose intervals are shorter than ε0 . We denote
Lj = L(γ|[yj−1 ,yj ] ) and dj = d(γ(yj−1 ), γ(yj )). By the choice of ε0 , we have
N
X N
X
(Lj − dj ) = L(γ) − dj < µα2 /2.
j=1 j=1
In the left-hand sum above all terms are nonnegative and those for which
[yj−1 , yj ] ∈ B (i.e., yj−1 = ti − εi and yj = ti + ε′i for some i) satisfy
Lj − dj > αLj > α2 (yj−1 − yj ) = α2 (εi + ε′i ).
Therefore
N
X M
X
2
(Lj − dj ) ≥ α (εi + ε′i ) > µα2 /2.
j=1 i=1
This contradiction proves the theorem. ¤
Theorem 2.7.6. Let X be a metric space; γ : [a, b] → X is a Lipschitz
curve. Then the speed vγ (t) exists for almost all t ∈ [a, b] and L(γ) =
Rb R
a vγ (t) dt where is the Lebesgue integral.

Proof. We need the following fact ([Fe], Theorem 2.9.19): if f : [a, b] → R is

a Lipschitz function, then the derivative f ′ (t) exists for almost all t ∈ [a, b]
Rb
and a f ′ (t) dt = f (b) − f (a). (Remark: the proof of this fact is based
on the same ideas as the above proof of Theorem 2.7.4 though it is more
complicated.)
Define f (t) = L(γ|[a,t] ) for t ∈ [a, b]. Then f is a Lipschitz function and
hence is differentiable almost everywhere. We rewrite f ′ (t) as follows:
L(γ|[t,t+ε] ) L(γ|[t,t+ε] ) d(γ(t), γ(t + ε))
f ′ (t) = lim = lim · .
ε→0 |ε| ε→0 d(γ(t), γ(t + ε)) |ε|
By Corollary 2.7.5, for almost all t ∈ [a, b] either f ′ (t) = 0 or the first term
in the last product goes to 1 as ε → 0. In the first case we have
d(γ(t), γ(t + ε))
vγ (t) = lim =0
ε→0 |ε|
58 2. Length Spaces

because d(γ(t), γ(t + ε)) ≤ L(γ|[t,t+ε] ). In the second case, it follows that
d(γ(t), γ(t + ε))
vγ (t) = lim = f ′ (t).
ε→0+ |ε|
Thus vγ (t) exists and equals f ′ (t) in both cases. The theorem follows. ¤
Exercise 2.7.7. Give an example of a nonconstant curve γ in R2 for which
vγ = 0 almost everywhere.
Exercise 2.7.8. Let X be a metric space and γ : I → X a curve. For a
t ∈ [a, b] define
dilt (γ) = lim sup dil(γ|[t−ε,t+ε] ).
ε→0+

(1) Prove that dilt (γ) ≥ vγ (t) whenever vγ (t) is defined.

(2) Prove that, if vγ (t) is defined for all t and is continuous in t, then
dilt (γ) = vγ (t) for all t.
Rb
(3) Give an example where γ is Lipschitz but a dilt (γ) 6= L(γ).
 Chapter 3

Constructions

3.1. Locality, Gluing and Maximal Metrics

3.1.1. Locality. First we observe that to reconstruct an intrinsic metric it
is enough to know it only locally, as the following proposition asserts:

Lemma 3.1.1. Let a topological space X be covered by a collection (not

necessarily finite or countable) of open sets {Xα }. Assume that each Xα
is equipped with a length structure Lα and that these structures agree. The
latter assumption means that, if a curve γ belongs to the intersection of Xα
and Xβ , then Lα (γ) = Lβ (γ).
Then there exists a unique length structure L on X whose restriction
to every Xα is Lα . Moreover, if X is connected and all intrinsic metrics
induced by Lα on Xα are finite, then so is L.

Proof. Consider a curve γ : [a, b] → X. The inverse images γ −1 (Xα ) form

an open covering of [a, b]. The compactness of [a, b] implies that there is a
finite partition a = t0 ≤ t1 ≤ · · · ≤ tn = b such that every segment [ti , ti+1 ]
is contained in an element of this covering. Then every image γ([ti , ti+1 ]) is
contained in one of the sets Xα , and the length of γ|[t ,t ] is given by Lα . By
i i+1
the additivity of length, the length of γ must be equal to the sum of lengths
of its restrictions to the intervals [ti , ti+1 ]. This proves the uniqueness part
of the lemma and suggests how to define L. To complete the proof one
checks that L(γ) defined this way is independent of the choice of a partition
and satisfies the conditions from the definition of length structures. This
part is left to the reader as an exercise.
To prove the statement about finiteness, fix a point x ∈ X and consider
the set Y of all points whose distances from x are finite. Every set Xα is

59
60 3. Constructions

contained in either Y or X \ Y ; hence both Y and X \ Y are open. Since X

is connected, it follows that Y = X and the metric is finite. ¤

Intrinsic metrics also enjoy the following property of locality:

Corollary 3.1.2. Consider two intrinsic metrics d1 and d2 defined on the
same set X and inducing the same topology. Assume that every point
x ∈ X has a neighborhood Ux such that the restrictions of the metrics to
this neighborhood coincide: for every p, q ∈ Ux , d1 (p, q) = d2 (p, q). Then
d1 = d2 .

The corollary implies that, unlike general metrics, an intrinsic metric

can be recovered from local measurements.
Exercise 3.1.3. Prove the corollary.
Exercise 3.1.4. Give an example demonstrating that Corollary 3.1.2 fails
without the assumption that the metrics in question are intrinsic.

The next proposition shows that locality in fact characterizes intrinsic

metrics among all metrics:
Proposition 3.1.5. If a complete metric d on a set X is not intrinsic, then
6 d1 but every point has a
there exists another metric d1 on X such that d =
neighborhood where d and d1 coincide.

Proof. For every positive ε, introduce the metric dε :

k
X
dε (x, y) = inf d(pi , pi+1 ),
i=0
where the infimum is taken over all finite sequences of points p0 , p1 , . . . , pk+1
such that p0 = x, pk+1 = y and d(pi , pi+1 ) ≤ ε for all i = 0, 1, . . . , k.
Intuitively, dε is measured by leaping from x to y with jumps each no longer
than ε. Obviously dε (x, y) = d(x, y) if d(x, y) ≤ ε, and thus d and dε coincide
on every ball of radius ε/2. On the other hand, if dε = d for all ε > 0, then
d is intrinsic by Proposition 2.4.17. ¤

3.1.2. Gluing. Imagine that we glue together two length spaces, or even
some points in the same length space, as we do to glue a Möbius strip or a
ring out of a paper rectangle. How do we measure distances in this space?
For two points “on different sides of the gluing line”, the first thing that
occurs to one’s mind would be to look for a shortest path that consists of
two pieces: first it goes in one space to the gluing edge and then from this
edge it continues in the other space. In other words, one thinks of a path
with a “gap”, but the gap is eliminated by gluing. As the matter of fact,
3.1. Locality, Gluing and Maximal Metrics 61

this is not at all so simple. It may very well happen that a shortest path
has to cross the gluing edge several or even infinitely many times.
Before passing to rigorous definitions, let us consider several examples.
Example 3.1.6. Begin with a strip R × [0, 1] ⊂ R2 and glue together its
edges by identifying (x, 1) with (x+100, 0) for all x. We obtain a topological
cylinder.
Question. Is the distance between (0, 1/2) and (1000, 1/2) (after the
identifications are made) greater than 899?
Answer. No, it is even less than 11! Indeed, consider the path whose
itinerary is
|(0, 1/2) → (0, 1)| = 1/2, |(0, 1) → (100, 0)| = 0 (“free”),
|(100, 0) → (100, 1)| = 1, |(100, 1) → (200, 0)| = 0 (“free”),
...,
|(900, 0) → (900, 1)| = 1, |(900, 1) → (1000, 0)| = 0 (“free”),
|(1000, 0) → (1000, 1/2)| = 1/2.
The length of this path is 10.
Exercise 3.1.7. Is this a shortest path?
Exercise 3.1.8. Consider the region E = {(x, y) : 0 ≤ y ≤ e−x , x ≥ 0}.
(This is the region enclosed between the graph of e−x , x ≥ 0, and the x-axis.)
We glue together the infinite edges of this region by identifying (x, 0) with
(x + 1, ex+1 ) for all x ≥ 0. Show that the diameter of the resulting space is
finite!
Hint: Consider the path consisting of the segments {[(n, 0), (n, e−n )],
n = 1, 2, . . .}. Observe that this is indeed a continuous path and its length
is equal to
X∞ Z ∞
−m
e < e−x dx = 1.
m=1 0

Remark 3.1.9. This is not an artificial example: such examples naturally

arise in hyperbolic geometry.

Certainly, both the arguments and the questions themselves are informal:
we have not given a definition of gluing yet. To show that this notion is not
so simple and indeed needs a rigorous definition, look at the following more
striking example:
Example 3.1.10. Begin with R2 and glue each point (x, y) with (−y, 2x).
Question 1. If we forget about metric and do this identification topolog-
ically, what is the quotient space?
62 3. Constructions

Question 2. How can we find the distance between two points after this
identification?
Answer. Very simple: the distance is identically zero.
Exercise 3.1.11. Prove that, given a positive ε and two arbitrary points in
the space from the above example, there is a path between the points whose
length is less than ε.
Hint: Our gluing rule can be re-written as
³y ´ µ ¶ µ ¶ µ ¶
−x −y −y x x −y
(x, y) → , −x → , → , → , → ....
2 2 2 4 2 4 4
These examples suggest the following strategy of measuring the metric
resulting from “gluing some points together” in terms of the intact space
(before identifications are made). We consider finite sequences of paths such
that the terminating point of each of them will be glued with the starting
point of the next one. To measure the distance between two points we
require that the first path starts at one of the points and the last path ends
at the other. The infimum of lengths of such paths gives us the new distance.
In other words, to measure the distance between x and y, we connect x and
y by a finite sequence of points (a “dotted line”). If we added the distances
between neighboring points in this sequence and took the infimum of such
sums, this would be just the original distance between x and y. What we
do instead is we add only the distances between neighboring points that are
nonequivalent. Intuitively, we may jump from a point to an equivalent point
for free.
Definition 3.1.12. Let (X, d) be a metric space and let R be an equivalence
relation on X. The quotient semi-metric dR is defined as
k
X
dR (x, y) = inf{ d(pi , qi ) : p0 = x, qk = y, k ∈ N}
i=1
where the infimum is taken over all choices of {pi } and {qi } such that qi is
R-equivalent to pi+1 for all i = 1, . . . , k − 1. As usual, we associate with a
semi-metric space (X, dR ) a metric space (X/dR , dR ) by identifying points
with zero dR -distances (and keep the same notation dR for this metric; see
Section 1.1). This space is called the quotient metric space. One also says
that it results from gluing the space (X, d) along the relation R.
Exercise 3.1.13. Prove that dR is indeed a semi-metric, i.e., it is nonneg-
ative, symmetric and satisfies the triangle inequality.

We will consider quotients of length spaces only. It is easy to see that

gluing a length space yields a length space. First of all, observe that dR ≤ d,
i.e., the gluing procedure never increases distances between (equivalence
3.1. Locality, Gluing and Maximal Metrics 63

classes of) points. Hence every d-continuous curve is dR -continuous, and

the dR -length is not greater than the d-length. Now if {pi }ki=1 and {qi }ki=1
are as in the definition of dR (x, y), one can constructPa curve connecting x
and y in (X/dR , dR ) whose length is almost equal to d(pi , qi ). To do this,
simply concatenate almost shortest paths (of the original metric) connecting
pi to qi , i = 1, . . . , k. Since qi and pi+1 are identified in X/dR , this curve
P continuous in (X/dR , dR ). Its length can be made arbitrarily close to
is
d(pi , qi ), and hence to dR (x, y). Therefore (X/dr , dR ) is a length space.
Let us make a few immediate observations. First of all, if two points
are R-equivalent, then they obviously get identified when passing to X/dR .
On the other hand, it is possible that the relation dR = 0 defining the
quotient X/dR is actually stronger than R. This means that more points are
identified in X/dR than in X/R. In particular, (X/dR , dR ) is not necessarily
homeomorphic to the topological quotient X/R. To see how this can happen,
glue together all rational points in [0, 1]. The topological quotient is a very
wild (even nonmetrizable) space, while the metric quotient is just a point.
The same effect takes place in Example 3.1.10 above.
Even if no additional identifications occur by metric reasons (i.e., X/dR
and X/R coincide as sets), still the topology of the metric quotient may
differ from the topological quotient (see Example 3.1.17 below).
Exercise 3.1.14. Suppose that X/dR and X/R coincide as sets, i.e., the
relation dR = 0 coincide with R. Prove that
1. The topology of the metric quotient X/dR is weaker than that of the
topological quotient X/R.
2. If X is compact, then the two topologies coincide.
Hint: Every continuous bijection from a compact topological space to a
Hausdorff space is a homeomorphism. Observe that X/R is compact, X/dR
is Hausdorff, and the identity map from X/R to X/dR is continuous.

The procedure of passing to a metric quotient can be used to glue several

metric spaces together. First we need a definition:
Definition 3.1.15. Let (Xα , dα ) be a collection of length spaces. Consider
the disjoint union ∪α Xα . We introduce a length metric d on this disjoint
union defining d(x, y) by the following rule:
If x, y ∈ Xα for some α, then d(x, y) = dα (x, y);
Otherwise, d(x, y) = ∞. The metric d will be referred to as the length
metric of disjoint union.

Now let us begin with the simplest example of two length spaces (X, dX )
and (Y, dY ) and a bijection I : X ′ → Y ′ between subsets X ′ ⊂ X and
64 3. Constructions

Y ′ ⊂ Y . Consider the disjoint union Z = X ∪ Y of X and Y together

with the length metric of disjoint union. To glue X and Y along I, we first
introduce the equivalence relation R on Z generated by the relations x∼y
if I(x) = y. Now the result of gluing X and Y along I by definition is the
quotient space (Z/dR , dR ).
More generally, if we want to glue a collection of length spaces along
an equivalence relation defined on their disjoint union, we first endow this
disjoint union with the length metric of disjoint union and then pass to the
metric quotient.
Numerous examples of gluing will appear in this chapter and throughout
the book. Here we give a few first examples:
Example 3.1.16. Consider the segment I = [0, 1] with its canonical metric
d. Introduce the equivalence relation R generated by 0 ∼ 1. It is easy to see
that (I/dR , dR ) is isometric to a circle of length 1.
Example 3.1.17. This example may seem as elementary as the first one;
however most of the students who see it for the first time make various
mistakes when analyzing its topology. Consider a countable collection of
disjoint intervals Ii and identify together all their left ends. The space
resulting from this metric gluing looks like a point with countably many
needles sticking out.
Question. Is this a compact space?
Answer. The topology of this space depends on the lengths of intervals Ii !
In particular, if all lengths of Ii are equal to 1, this space is homeo-
morphic to the topological quotient: this is the bouquet of countably many
segments. In particular, this is a noncompact space: the right ends of inter-
vals form a countable collection of points with all pairwise distances equal
to 2. On the other hand, if the length of Ii is 1/i, the “metric bouquet”
is compact (prove this!). Since compactness is a topological property, we
see that this gluing leads to different topologies resulting from the same
topological spaces with the same equivalence relation factored out.
Exercise 3.1.18. Can you find all topological spaces that may result from
this example for different lengths of Ii ?
Example 3.1.19. Begin with the square Q = [0, 1]×[0, 1] with its Euclidean
metric. Introduce the equivalence R generated by the relations of the form
(0, x) ∼ (1, x) and (x, 0) ∼ (x, 1), x ∈ [0, 1]. The quotient space is a torus.
It may seem that, while nothing happens to the interior points of Q, the
boundary points and especially the vertices (all the four glued together) are
somewhat distinguished and the quotient metric may look different near to
these points. This is not the case:
3.1. Locality, Gluing and Maximal Metrics 65

Exercise 3.1.20. Prove that the (equivalence class of) a vertex of Q in the
quotient metric has a neighborhood isometric to a Euclidean region.

This exercise explains why this quotient space is called a flat torus.
Example 3.1.21. Let G be a group of isometries (rigid motions) of R2 .
Introduce an equivalence relation R defined as follows: xRy iff there exists
a g ∈ G such that x = gy (i.e., x and y belong to the same orbit).
Exercise 3.1.22. Prove that the quotient metric between (the equivalence
classes of) two points x and y is equal to inf g∈G |x − g(y)|, where | · | stands
for Euclidean distance.

Different choices of G lead to various examples. For instance, the group

of translations by vectors with integer coordinates leads to the same torus
as in the previous example. For G being the whole group of rotations, the
quotient space is isometric to the ray [0, ∞). The reader may wish to play
with the examples of G generated by
(a) all translations along the x-axis;
(b) the transformation (x, y) → (x + 1, −y);
(c) the translation by vector (0, 1) and the symmetry in the y-axis;
(d) the transformations (x, y) → (x + 1, −y) and (x, y) → (−x, y + 1).

3.1.3. Maximal metric. This section introduces a very useful construc-

tion that generalizes (and in a sense axiomatizes) the procedure of gluing.
To understand the punch line of the construction, let us informally formu-
late the properties that a gluing procedure must definitely possess; we will
see that it is uniquely determined by these properties.

Informal axiomatizing of gluing. Instead of saying that gluing makes

one point out of two (or more) points, we could say that we glue two
points “metrically” by setting zero distance between them. This is certainly
impossible since our definition of distance says that the distance between
different points is positive. This is, however, a minor difficulty: we relaxed
this assumption and defined semi-metrics, which will perfectly suit our
purpose. There is a more serious problem: just assigning some distances
to be zero and keeping other distances unchanged may violate the triangle
inequality, and this is what we cannot afford to sacrifice. This forces us to
modify other distances at the same time. On the other hand, it would be
very unnatural if gluing some points together could increase any distance.
We also want to decrease distances only to the extent dictated by gluing.
Summarizing this, we informally define the metric resulting from gluing as
the largest metric not exceeding the original one and such that the distances
between identified points are zero.
66 3. Constructions

Maximal metric. Now we can proceed with formal definitions. To formal-

ize the words “the largest metric...” in the informal definition of gluing, we
need the following lemma.
Lemma 3.1.23. Let b : X × X → R+ ∪ {+∞} be an arbitrary function.
Consider the class D of all semi-metrics d on X such that d ≤ b, i.e.,
d(x, y) ≤ b(x, y) for all x, y ∈ X. Then D contains a unique maximal semi-
metric dm such that dm ≥ d for all d ∈ D.

Proof. For every two points x, y ∈ X define

dm (x, y) = sup{d(x, y) : d ∈ D}.
The function dm is nonnegative, symmetric and satisfies dm ≤ b. So all we
have to prove is the triangle inequality for dm . For all x, y, z ∈ X one has
dm (x, y) = sup d(x, y) ≤ sup (d(x, z) + d(z, y))
d∈D d∈D
≤ sup d(x, z) + sup d(z, y) = dm (x, z) + dm (z, y),
d∈D d∈D
and the lemma follows. ¤
Corollary 3.1.24. Let a set X be covered by a collection of subsets {Xα }
each carrying a semi-metric dα . Consider the class D of all semi-metrics
d (possibly taking infinite values) such that d(x, y) ≤ dα (x, y) whenever
x, y ∈ Xα . Then D contains a unique maximal semi-metric dm such that
dm (x, y) ≥ d(x, y) for all d ∈ D and x, y ∈ X. If all dα are intrinsic (semi-)
metrics, then so is dm .

Proof. One may assume that the metrics dα are defined on the whole space
X by setting dα (x, y) = ∞ if x ∈ / Xα or y ∈ / Xα . To prove the existence of
dm , apply Lemma 3.1.23 to the function b(x, y) = inf α dα (x, y). To prove
that dm is intrinsic if dα are, consider the intrinsic metric dbm induced by dm .
Since all dα are intrinsic and dm ≤ dα , it follows that dbm ≤ dα , so dbm belongs
to D. Then dbm = dm because dm is maximal. ¤

An important case is a topological space X covered by a collection of

metric spaces (Xα , dα ) whose metrics agree on the intersections of their
domains and are consistent with the topology (compare with Lemma 3.1.1).
If X is connected, the sets Xα are open and the semi-metrics dα are finite,
then the maximal semi-metric is finite.
Exercise 3.1.25. Prove this statement.
Exercise 3.1.26. Let (X, d) be a metric space and set dε to be the maximum
metric w.r.t. the covering of X by all balls of radius ε equipped with the
restrictions of d. Let d0 (x, y) = supε>0 dε (x, y). Prove that if d is complete,
then d0 is the intrinsic metric induced by d.
3.2. Polyhedral Spaces 67

Now we are ready to give a definition of gluing in terms of maximal

metrics.
Theorem 3.1.27. Let (X, d) be a metric space and R an equivalence rela-
tion on X. Consider the function bR on X × X defined by
(
0 if x is R-equivalent to y,
bR (x, y) =
d(x, y) otherwise.
Then the maximal semi-metric among those not exceeding bR coincides with
the semi-metric dR obtained by gluing (X, d) along R as in Definition 3.1.12.

Proof. Let D denote the class of semi-metrics not exceeding bR . Obviously

dR ∈ D, so it suffices to prove that dR ≥ d′ for any semi-metric d′ ∈ D.
Let x, y ∈ X and {pi }ki=1 , {qi }ki=1 be as in Definition 3.1.12. Then by the
triangle inequality
k
X k−1
X k
X
′ ′ ′
d (x, y) ≤ d (pi , qi ) + d (qi , pi+1 ) ≤ d(pi , qi )
i=1 i=1 i=1
since bR (pi , qi ) ≤ d(pi , qi ) and bR (qi , pi+1 ) = 0. Hence d′ ≤ dR . ¤

Note that the function bR in the theorem is the minimum of two length
semi-metrics: the original metric d and the one which equals 0 for R-
equivalent points and ∞ otherwise. Thus we are in conditions of Corollary
3.1.24, which gives us another proof for the fact that dR is intrinsic.

3.2. Polyhedral Spaces

3.2.1. Polyhedral metrics.

Two-dimensional finite polyhedral spaces. Roughly speaking, a poly-

hedral space is a length space that can be obtained by using convex poly-
hedra as building blocks and a collection of isometries between their faces
indicating how these blocks should be attached to each other.
Two-dimensional polyhedral spaces are glued from points, segments and
polygons. Since every polygon can be cut into several triangles by its
diagonals, we can do with triangles instead of polygons.
Let us define 0-, 1- and 2-dimensional finite polyhedral spaces.
A 0-dimensional finite polyhedral space P0 is just a space consisting of
finitely many points, with all pairwise distances being infinite. These points
are called vertices.
To construct a 1-dimensional finite polyhedral space, we begin with a
0-dimensional finite polyhedral space P0 and a finite collection of segments
E = {Ei } (each viewed as a length space isometric to a segment). For each
68 3. Constructions

Ei , fix an injective map ei sending the endpoints of the segment Ei to P0 .

Gluing E to P0 along {ei } produces a 1-dimensional polyhedral space P1 .
Note that different points from the same segment from E never glue together
and hence we can regard segments from E lying in P1 . These segments are
called edges. 1-dimensional polyhedral spaces are called graphs (see the next
section).
To construct a 2-dimensional polyhedral space, begin with a 1-dimensi-
onal polyhedral space P1 and take a finite collection of polygons F = {Fi }.
For each polygon Fi fix an injective map fi from the boundary of Fi to P1
such that fi bijectively maps each side of Fi onto an edge from E of the
same length. Moreover fi must map sides of the polygon to the respective
edges isometrically in the sense that lengths of subintervals are preserved.
Gluing F to P1 along {fi } produces a 2-dimensional polyhedral space P2 .
Copies of Fi lying in P2 are called faces.
Several examples are given as the following exercises.
Exercise 3.2.1. Prove that the surface of a convex polyhedron in R3 is a
2-dimensional polyhedral space.
Exercise 3.2.2. Prove that the flat torus from Example 3.1.19 is a 2-
dimensional polyhedral space.
Exercise 3.2.3. Prove that each two-dimensional finite polyhedral space
can be glued in such a way that the distance between two points in the
same face can be measured by a segment in this face. In other words, there
is a shortest path that does not traverse other faces.
General polyhedral spaces. Here we introduce general polyhedral spaces.
We do not impose such restrictions as local finiteness, finite dimension and
even local compactness. Such spaces can be quite useful, and we will give a
few examples of polyhedral monsters. Most of other examples of polyhedral
spaces that arise further in this book are two-dimensional locally-finite (or,
better, finite) polyhedral spaces. Hence, this section can be omitted for the
first reading.
We will not repeat standard definitions from the theory of convex poly-
hedra in Euclidean space. The reader who is not comfortable with higher-
dimensional convex polyhedra can think of two-dimensional polygons and
perhaps three-dimensional polyhedra.
Since every polyhedron can be triangulated, it is actually enough to work
with simplexes. One-dimensional simplexes are segments, two-dimensional
simplexes are triangles and three-dimensional simplexes are tetrahedra.
Each tetrahedron has four triangular faces of dimension 2 and six one-
dimensional faces (edges). Vertices may be regarded as 0-dimensional faces.
The entire tetrahedron can be viewed as its three-dimensional face.
3.2. Polyhedral Spaces 69

Let us formally define a polyhedral space.

Definition 3.2.4. Let a topological space (P, d) be covered by a collection of
length spaces (Pα , dα ), each isometric to a simplex (or a convex polyhedron).
Assume that for any two simplexes Pα and Pβ their intersection Pα ∩ Pβ is
a face in both of them, and that the restrictions of the metrics of Pα and
Pβ to Pα ∩ Pβ coincide. Consider the maximal metric d majorized by all
metrics dα . The length space (P, d) is called a polyhedral space, or, more
precisely, a Euclidean polyhedral space. (The latter term stresses the fact
that P is built from Euclidean polyhedra. One may as well consider spaces
glued from simplexes equipped with some non-Euclidean length metrics.)
The polyhedra Pα are called faces of P . 0-dimensional and 1-dimensional
faces are called vertices and edges, respectively. A particular representation
of a polyhedral space as a union of faces Pα is referred to as a triangulation.
We will usually assume that every polyhedral space comes with some fixed
triangulation.

There is a more constructive reformulation of this definition. Namely,

we can begin with a collection of simplexes (convex polyhedra) {Pα } and a
collection of isometries {Iδ }, such that each isometry Iδ : Pαf → Pβf maps a
face Pαf ⊂ Pα onto a face Pβf ⊂ Pβ for some Pα and Pβ . Form the disjoint
S
union Z = α Pα and introduce an equivalence relation R on Z generated
by the relations x ∼ Iδ (x).
The reader should be already suspecting that what we want to do next
is to glue Z (endowed with the length metric of disjoint union) along the
relation R. There is, however, a little formal problem here. It may happen
that the relation R identifies two different points of the same simplex Pα .
If we want this construction to be coherent with the previous definition,
this should be prohibited. Thus, the quotient space (P, d) resulting from
gluing Z along the relation R is a polyhedral space if each Pα projects to P
injectively. Fortunately this requirement does not restrict the class of spaces
that can be produced this way:
Exercise 3.2.5. Prove that every space P glued from convex polyhedra by
means of isometries of their faces is a polyhedral space.
Hint: Subdivide polyhedra into smaller ones.

Notice that an isometry between two simplexes (as well as convex

polyhedra) is uniquely determined by the images of the vertices. Thus the
data needed to describe a polyhedral space are more or less combinatorial.
More precisely, to define a polyhedral space it suffices to fix the lengths of
edges of simplexes Pi (to determine their geometry) and give a combinatorial
scheme of gluing.
70 3. Constructions

By the dimension of a polyhedral space we mean the maximal dimension

of polyhedra used to glue it. A polyhedral space is said to be dimensionally
homogeneous is it can be glued from polyhedra of the same dimension.
Equivalently, a polyhedral space P is dimensionally homogeneous if every
point of P belongs to an n-dimensional face where n is the dimension of P .
A polyhedral space is said to be finite if is has finitely many faces, and
locally finite if every point possesses a neighborhood which intersects only
finitely many faces.
Exercise 3.2.6. Prove that a polyhedral space P is locally finite if and only
if every point of P belongs to finitely many faces.

For nice examples, we refer to the previous subsection. Here we present

a couple of a bit terrifying examples.
Example 3.2.7. Consider an infinite sequence of cubes Pi = [0, 1]i . There
are the isometries Ii : Pi−1 → Pi acting by adding 0 for the last coordinate:
Ii (x1 , x2 , . . . xi−1 ) = (x1 , x2 , . . . xi−1 , 0).
Thus we can form a polyhedral space. This is an infinite-dimensional cube
of all finite sequences of reals between 0 and 1.
Exercise 3.2.8. Can you define an infinite-dimensional simplex?

3.2.2. Metric graphs. We presume that the reader is familiar with topo-
logical graphs at least as a nice way to draw a few cities connected by several
roads. Metric graphs are just one-dimensional polyhedral spaces. Regard-
less the fact that the definition of a polyhedral space is given in the previous
section, we repeat it here to better visualize its meaning in this simplest
case. This definition generalizes the example of a “cobweb” from the first
chapter.
By a metric segment (of length a) we mean a metric space isometric to
a segment [0, a].
Definition 3.2.9. A (metrized) graph is the result of gluing of a disjoint
collection of metric segments {Ei } and points {vj } (regarded with the length
metric of disjoint union) along an equivalence relation R defined on the union
of the set {vj } and the set of the endpoints of the segments.

In other words, we consider the maximal semi-metric that is bounded by

all metrics of segments and an additional semi-metric dR (x, y) that is zero
if xRy and infinite otherwise, and then identify points that are within zero
distance from each other w.r.t. the maximal semi-metric.
The segments {Ei } are called edges, and the equivalence classes of the
endpoints are called vertices of the graph. The length of an edge is just
3.2. Polyhedral Spaces 71

the length of the segment (it may be different from the distance between
its endpoints in the graph). The cardinality of the endpoints of segments in
an equivalence class representing a vertex is called its degree. Most of the
graphs considered in this book will be locally finite and therefore with finite
degree of every vertex (one exception is a monster-graph in an example in
this section).
The most natural way to define a graph is to take a set V of vertices,
indicate which pairs of vertices are connected by edges, and specify lengths
of these edges. To recover the original definition from these data, take a
collection of segments {Ei } corresponding to the desired edges and having
the given lengths. Then, if the edge corresponding to Ei should connect
vertices a and b, let one endpoint of Ei be equivalent to a and another be
equivalent to b. This generates an equivalence relation R used to glue the
graph from V and {Ei }.
In particular, every topological graph can be turned to a metric graph
by assigning lengths to its edges. However, one should remember that (for
infinite graphs) this may change the topology of the space. See also Exercise
3.2.12 below.
Note that a pair of vertices may be connected by two or more different
edges. Definition 3.2.9 even allows a graph to have loops, that is, edges
connecting a point to itself. One can always obtain a graph without
loops and multiple edges by dividing edges into smaller ones (compare with
Exercise 3.2.5).
To get better understanding of the definitions, we suggest the following

Exercise 3.2.10. Unwind the definition of gluing and verify that for finite
graphs our definition agrees with the usual one: the distance from x to y is
the shortest way of reaching y from x by means of traversing finitely many
segments such that the terminating point of each segment is identified with
the beginning of the next one. Is this true for infinite graphs? What if we
replace “the shortest” by “the infimum”?

Caution: As well as in other cases of gluing, points that are nonequiva-

lent w.r.t. R can still get identified.

Exercise 3.2.11. Give such an example.

Exercise 3.2.12. Prove that additional identifications cannot occur if the

graph is locally finite, or if the lengths of all edges are bounded from below
by a positive constant. Moreover in these cases the topology of the metric
graph coincides with that of the topological graph (i.e., of the topological
quotient of the disjoint union by the same equivalence relation).
72 3. Constructions

On the other hand, the identification X → X/dR can never affect the
interiors of segments:
Exercise 3.2.13. Prove that points in the interiors of edges do not get
identified and thus the disjoint union of these interiors is embedded into the
graph.

In other words, every edge is a simple curve connecting two vertices and
not passing through other vertices, and the interiors of different edges have
no common points.
Exercise 3.2.14. Prove that the length of every edge as a curve in the
graph equals the length of the segment from which the edge was obtained.
Exercise 3.2.15. Let X be a length space and V ⊂ X a finite set.
Suppose that X \ V is covered by a finite collection of rectifiable curves,
each connecting two points of V , and these curves have no self-intersections
and do not intersect one another (except at endpoints). Prove that X is a
metric graph.
Exercise 3.2.16. Show that the statement of the previous exercise fails
without the assumption that the collection of curves is finite.
Hint: Every length space is a union of points, but not all length spaces
are 0-dimensional polyhedra.

Although our definitions sound quite scientific, the object is very familiar
and our everyday intuition is a good guide through simple cases of graphs.
For instance, to find intrinsic distances and shortest curves in the frame of
a cube is a good exercise for a 10-year old kid. Graphs may even seem dull
as being too discrete and combinatorial for a geometrically-thinking reader.
To relieve this feeling we suggest the following approach: one can try to
approximate an arbitrary length space by graphs whose vertices are dense
enough in the space; see Chapter 7 for more details. For the most skeptical
reader, the following example shows that every metric space can be realized
as the set of vertices of a graph, with the distance between points being
equal to that in the graph. Alas, this graph is apparently a hardly useful
monster.
Example 3.2.17. Begin with a general metric space (X, d); consider a
disjoint union of segments Ix,y parameterized by pairs x, y ∈ X. Equip each
segment Ix,y with the metric of the segment [0, d(x, y)]. These are edges.
The identifications are most natural: one identifies the right ends of two
segments Ix,y and Ix′ ,y′ if y = y and the left ends if x = x′ . We leave as an
exercise to verify that the set X can be canonically identified with the set of
vertices of the graph and that the intrinsic metric of this graph restricted to
3.2. Polyhedral Spaces 73

the set of its vertices is the original metric d in X. (To get better intuition
behind this construction, one can think of it this way: for each pair of points
x, y ∈ X, one takes a segment of length d(x, y) and attaches one of its ends
to x and the other one to y.)

3.2.3. Word metrics. There is an important class of graphs that arise

from groups. Let G be a finitely generated group; that is, it contains a finite
generating set: a collection of elements S = {g1 , . . . , gk } ⊂ G such that
every element of G can be represented as a finite product of elements of S.
We also require S to be symmetric: for every s ∈ S, s−1 is also in S.
Remark 3.2.18. If S is not symmetric, it can be symmetrized by including
all inverses of its elements. Thus we will not list all inverses of elements
in the generating sets, following a commonly accepted convention that all
inverses are also included there by default.

Given a group G and a generating set S = {g1 , . . . , gk } ⊂ G, one

constructs the Cayley graph of (G, S). Its vertices are just elements of G,
and two elements g, h ∈ G are connected by an edge if and only if gh−1 ∈ S.
In other words, every vertex g ∈ G is connected by edges to the k vertices
g1 g, g2 g, . . . , gk g. The lengths of all edges are equal to 1.
In terms of Definition 3.2.9, the Cayley graph can be described as follows.
Take a collection of unit intervals Igs labeled by the pairs (g ∈ G, s ∈ S),
and glue them along the equivalence relation generated by the following
identifications: if h = sg, identify the right end of the segment Igs with the
left ends of segments Iht for all t ∈ S.
Exercise 3.2.19. What are the Cayley graphs for the following groups and
generating sets:
1. G is Z generated by {1};
2. G is Z/mZ generated by one generator;
3. G is Z2 generated by {(0, 1), (1, 0)};
4. G is Z2 generated by {(0, 1), (1, 0), (1, 1)};
5. G is a free group generated by two generators a, b.
Answers. 1. Real line made of unit segments.
2. A regular m-gone with side 1.
3. The grid on the plane formed by lines x = n and y = n for all
integers n.
4. The grid on the plane formed by the lines x = n, y = n and x − y = n
for all integers n. To make the picture more consistent with the metric
(where all edges have unit length), draw a grid dividing the plane into regular
triangles with unit sides.
74 3. Constructions

5. An infinite tree with all nodes of degree 4. (A tree is a graph which

contains no subsets homeomorphic to the circle.)
Definition 3.2.20. The restriction to G of the intrinsic distance associated
with the length structure of a Cayley graph is called the word metric on G
(with respect to a given set of generators).
Exercise 3.2.21. Check that this is a finite metric.
Exercise 3.2.22. Write down the word metrics and draw word metric’s
balls of radius 5 for the first four examples in Exercise 3.2.19.
Exercise 3.2.23. Show that the word distance between g and h is equal to
the smallest number n such that g can be represented as g = h1 h2 . . . hn h,
with all hi ∈ S (recall that we assumed that S is symmetrical). That is why
this distance is called word metric: the distance between two elements g and
h is the length of the shortest word in generators that is equal to g −1 h.

The role of this construction in both combinatorial group theory and

modern geometry is difficult to over-estimate; we will dwell on it a little
further in Chapter 8.

3.3. Isometries and Quotients

Let us recall the notion of isometry:
Definition 3.3.1. Let X and Y be two metric spaces. A map f : X → Y
is called an isometry onto its image if it preserves distance. The latter
means that |f (x)f (y)| = |xy| for any two points x, y ∈ X (and thus it is
automatically injective).
A map that is surjective and preserves distance is called an isometry.
Two spaces are said to be isometric if there exists an isometry from one to
the other.

It is clear that an isometry is automatically a homeomorphism and being

isometric is an equivalence relation. It is clear that an isometry onto its
image maps a curve to a curve of the same length and therefore it is an
isomorphism of length structures. For an isometry, the image of a shortest
path (geodesic, sphere, ball, etc.) is again a shortest path (geodesic, sphere,
ball, etc.). We can say that metric geometry studies classes of isometric
spaces.
Isometries of a space to itself obviously form a group (check this!)
naturally called the isometry group. We will denote it by Iso(X). For
a “generic” space this group is trivial. (Can you give an example of a
length space with trivial isometry group?) If the isometry group of a length
3.3. Isometries and Quotients 75

space is nontrivial, this is often expressed by saying that the space possesses
symmetries. There is a remarkable class of very symmetrical spaces, namely
homogeneous spaces.
Definition 3.3.2. A length space X is said to be homogeneous if for every
x, y ∈ X there exists an isometry I : X → X such that I(x) = y.
Example 3.3.3. Euclidean spaces and (round) spheres are homogeneous
spaces. The cylinder x2 + y 2 = 1 in R3 and the torus from Example 3.1.19
are also homogeneous. On the other hand, the cone x2 + y 2 = z 2 and the
paraboloid x2 + y 2 = z are not homogeneous spaces.
Can you find the isometry groups for these spaces?
There is an important notion, which is relevant to introduce here.
Definition 3.3.4. A metric spaces X is said to be locally isometric to a
homogeneous space Y if each point in X possesses a neighborhood isometric
to an open set in Y . Spaces locally isometric to a Euclidean space are said
to be flat.
Flat spaces cannot be distinguished from Euclidean space by local mea-
surements: they look the same. For instance, the cylinder x2 + y 2 = 1 in R3
is flat. To realize that this is a cylinder, a two-dimensional creature would
have to undertake a long “Magellan” trip around the cylinder. The cone
x2 + y 2 = z 2 , z ≥ 0, is not flat, as it looks very different at its apex than the
Euclidean plane.
Question: What would you suggest for a two-dimensional creature living
in the cone to tell that this is not a two-plane? What about a two-
dimensional creature living in the sphere? Did we have to travel around the
Earth to figure out that it is not planar? Notice that the experiments with
a horizon are illegal in this set-up: they use measurements in the ambient
space, and we are allowed to use intrinsic measurements only. The answer
to these questions is not that easy, but the reader should be able to give
them after reading this textbook.

Quotient spaces. In this section we are going to deal with a subgroup of

an isometry group. The reader who is familiar with group actions should
rather think of a group G acting on X by isometries. We recall the definition
of group action here:
Definition 3.3.5. One says that a group G acts on a set X if there is a
map ϕ : G × X → X, which we abbreviate as ϕ(g, x) = g(x), such that
(i) gh(x) = g(h(x)) and
(ii) e(x) = x
for every g, h ∈ G, x ∈ X. Here e is the unit of G.
76 3. Constructions

Consider a subgroup G ⊂ Iso(X) of the isometry group of a length

space (X, d). Introduce an equivalence relation RG : points x and y in X
are equivalent iff x = g(y) for some g ∈ G (check that this is an equivalence
relation!)
Since we have an equivalence relation on a length space, we can glue
together equivalent points and consider the quotient metric on the space
resulting from this identification as in Definition 3.1.12 of gluing.
It turns out that, for an equivalence relation arising from an isometry
subgroup, the space resulting from the gluing construction is just X/RG
(usually denoted by just X/G) and the distance between two points in X/G
is the infimum of distances between their representatives in X.
More formally, for x, y ∈ X/G, set

d(x, y) = inf{d(x, y) : x ∈ x, y ∈ y}

for every x, y ∈ X/G.

Recall that the equivalence class of a point x is called the orbit O(x) =
{g(x) : g ∈ G}. Then the definition of d can be re-formulated as follows:

d(O(x), O(y)) = inf d(x, g(y)).

g∈G

Lemma 3.3.6. d coincides with the quotient metric dRG .

Proof. By the definition of gluing, the distance between the classes of

p, q ∈ X in the quotient metric is the infimum of sums

k
X
d(pi , qi )
i=0

for all finite collections of {pi , qi , i = 0, 1, . . . , k, k ∈ N} such that p0 = p,

qk = q and with pi being equivalent to qi−1 for all i > 1. The latter means
that there are isometries gi ∈ G such that gi (pi ) = qi−1 . We can use these
isometries to assemble all segments pi , qi together. Namely, consider a new
sequence of points

p̃0 = p0 = p, q̃0 = q0 ,
p̃1 = g1 (p1 ), q̃1 = g1 (q1 ),
p̃2 = g1 (g2 (p2 )), q̃2 = g1 (g2 (q2 )),
...
p̃k = g1 ◦ g2 ◦ · · · ◦ gk (pk ), q̃k = g1 ◦ g2 ◦ · · · ◦ gk (qk ).
3.3. Isometries and Quotients 77

Since all gi and thus their compositions g1 ◦ g2 ◦ · · · ◦ gi ) are isometries,

d(pi , qi ) = d(p̃i , q̃i ) and thus
k
X k
X
d(pi , qi ) = d(p̃i , q̃i ).
i=0 i=0

On the other hand, by the choice of the isometries gi , q̃i = p̃i+1 . Thus
k
X k
X
d(p, q̃k ) ≤ (d(p̃i , q̃i ) + d(q̃i , p̃i+1 )) = d(pi , qi ).
i=0 i=0

Recalling that
q̃k = g1 ◦ g2 ◦ · · · ◦ gk (qk ) = g1 ◦ g2 ◦ · · · ◦ gk (q)
and thus q̃k belongs to the orbit O(q) of q, we conclude that the distance
between the orbits of p and q is less than or equal to the distance between the
equivalence classes of p and q in the quotient metric. The opposite inequality
is obvious. Indeed, if q ′ ∈ O(q) and p′ ∈ O(p), then p′ and q ′ are equivalent
to p and q respectively, and hence one chooses the path p0 = q0 = p, p1 = p′ ,
q1 = q ′ and p2 = q2 = q with only one nonzero jump, and the length of this
path is equal to d(p′ , q ′ ).
¤
Exercise 3.3.7. Give a direct proof (that is, without referring to Definition
3.1.12 of gluing and the notion of maximal metric) that the metric d defined
this way is an intrinsic metric.
Hint: Here is a proof for the case when X is locally compact and
complete and all orbits are compact. To get a general proof, one has to
use approximations.
Given a, b ∈ X/G, choose their representatives x ∈ a, y ∈ b (as a and
b are equivalence classes) such that d(a, b) = d(x, y). This is possible since
we assumed that all orbits are compact. Let z be a mid-point for x, y, i.e.
d(x, z) = d(y, z) = 21 d(x, y). Obviously d(a, O(z)) ≤ d(x, z) = 12 d(x, y) =
1 1
2 d(a, b). Similarly d(b, O(z)) ≤ 2 d(a, b). The triangle inequality shows that
both inequalities turn out to be equalities, so O(z) is a mid-point for a, b. ¤
Example 3.3.8. For the subgroup G of isometries of R2 acting by integer
translations (g(x, y) = (x + k, y + l), k, l are integers), the quotient space
R2 /G is a torus (isometric to the flat torus from Example 3.1.19).
Example 3.3.9. For the subgroup G of isometries of R2 consisting of
two elements, the identity and the symmetry −id : (x, y) → (−x, −y), the
quotient space is isometric to a cone x2 + y 2 = cz 2 , z ≥ 0. Can you find this
coefficient c? What about G being a finite group of rotations?
78 3. Constructions

Example 3.3.10. For the subgroup G of isometries of R2 consisting of

two elements, the identity and the symmetry (x, y) → (x, −y), the quotient
space is a half-plane.
The reader may notice a difference between these examples: while the
torus looks “the same at every point”, both cones and half-planes have
distinguished “singular” points (the vertex of a cone and the edge of a half-
plane). In the next subsection we will see that these singular points come
from fixed points of elements from G.

3.4. Local Isometries and Coverings

3.4.1. Local isometries.
Definition 3.4.1. A map f : X → Y is called a local isometry at x ∈ X if x
has a neighborhood Ux such that (the restriction of) f maps Ux isometrically
onto an open set Uy in Y . A map which is a local isometry at every point
is called a local isometry.
Notice that in this definition we consider Ux and Uy with the restrictions
of the metrics; we do not induce a new length structure in them. This
remark is important for the spaces where points may have no “convex”
neighborhoods.
A local isometry can change distances. For example, it can map two
distinct points into one point. However, every local isometry of a length
space is a nonexpanding map; i.e., it never increases distances between
points.
Exercise 3.4.2. Prove this statement.
Example 3.4.3. The map f : R → S given by f (t) = (sin(t), cos(t)) is a
local isometry (w.r.t. the intrinsic (angular) metric on S).
Example 3.4.4. More generally, if f : X → Y is a local homeomorphism
and Y is a length space, then X admits a unique structure of a length space
such that f is a local isometry. This is the length structure induced by f
(see section 2.2 for the definition).
Exercise 3.4.5. Prove the statement formulated in this example.
The following exercise explains why the torus R2 /Z2 is flat and the cone
R2 /{id, −id} is not.
Exercise 3.4.6. Let a group G act by isometries on a length space X. (The
reader still may think of G as just a subgroup of Iso(X).) Assume that all
orbits are discrete. Prove that the projection map p : X → X/G is a local
isometry at a point x if and only if x is not fixed by a nonidentity element
of G.
3.4. Local Isometries and Coverings 79

3.4.2. Covering maps. Covering maps are an important class of local

homeomorphisms. When length spaces are considered, covering maps pro-
vide an important class of local isometries. Here we recall basic properties
of covering maps and consider relations between covering maps and length
metrics. As for topological properties, we only recall the definitions and for-
mulate the most important facts. The proofs can be found in many books
(see [Mas]).
We want to stress additional properties arising when covering maps of
length spaces are considered.

Coverings in general. Let X and Y be topological spaces and f : X → Y

a continuous map. An open set V ⊂ Y is said to be evenly covered if its
inverse image f −1 (V ) is a disjoint union of open sets Ui ⊂ X such that
the restriction of f onto each Ui is a homeomorphism from Ui to V . The
map f is a covering map, or simply covering, if every point y ∈ Y has an
evenly covered neighborhood. The space Y is called the base of the covering
and X the covering space. Classical examples of covering maps include:
the map f : R → S 1 given by f (x) = (cos x, sin x); the standard covering
of the torus by the plane, i.e., the map F : R2 → T 2 = S 1 × S 1 given by
F (x, y) = (f (x), f (y)) where f (x) = (cos x, sin x) ∈ S 1 ; the projection from
the sphere S 2 to the projective plane RP2 (recall that RP2 is the quotient
of S 2 by the equivalence relation x ∼ −x).
Every covering map is a local homeomorphism. The converse is not true;
for example, consider the inclusion map from the interval (0, 1) to R.
If X and Y are connected and f : X → Y is a covering map, then the
cardinality of f −1 (y) does not depend on y ∈ Y and is called the number of
sheets of f (this “number” may be infinity). We consider covering maps of
arcwise connected spaces only.
Suppose that f : X → Y is a covering map and Y is a length space.
Since f is a local homeomorphism, there is a unique length metric on X
such that f is a local isometry. This metric on X is called the lift of the
metric of X.
For a reader already familiar with smooth manifolds and Riemannian
metrics, we note that if Y is a Riemannian manifold, then its differential
structure and Riemannian metric can be similarly lifted to X.

Exercise 3.4.7. Let X and Y be length spaces, and let f : X → Y be a

covering map and a local isometry. Let y ∈ Y and a ball Br (y) be an evenly
covered neighborhood. Prove that the distances between distinct points of
f −1 (y) are no less than 2r.
80 3. Constructions

Exercise 3.4.8. Let X and Y be length spaces, and let f : X → Y be a

covering map and a local isometry, and suppose that Y is complete. Prove
that X is complete too.
Hint: Since the map is nonexpanding, it maps a Cauchy sequence in X
to a Cauchy sequence in Y . Consider the limit of the latter and an evenly
covered neighborhood of it.

Exercise 3.4.9. Let X and Y be length spaces, and let f : X → Y be a

covering map and a local isometry, y1 , y2 ∈ Y , ε > 0. Prove that for every
x1 ∈ f −1 (y1 ) there exists an x2 ∈ f −1 (y2 ) such that |x1 x2 | ≤ |y1 y2 | + ε. If
the metric of Y is strictly intrinsic, then the same is true with ε = 0.
Hint: Utilize the standard topological lemma about lifting curves: for
every curve γ : [a, b] → Y and every x ∈ f −1 (γ(a)) there exists a (unique)
curve γ̃ : [a, b] → X such that f ◦ γ̃ = γ and γ̃(a) = x.

Exercise 3.4.10. Prove the converse to Exercise 3.4.8. Namely if X and

Y are length spaces, f : X → Y is a covering map and a local isometry, and
X is complete, then Y is complete too.
Hint: For a given Cauchy sequence {yn } in Y construct a Cauchy
sequence {xn } in X such that f (xn ) = yn .

Universal covering. A covering map f : X → Y is called a universal

covering if X is simply connected. In this case, X is called a universal
covering space for Y .
The next theorem tells that every “not too weird” topological space
possesses a unique universal covering. To make this more precise, we give
some definitions first. Two covering maps f1 : X1 → Y and f2 : X2 → Y are
said to be equivalent if there is a homeomorphism h : X1 → X2 such that
f1 = f2 ◦ h. It is natural not to distinguish equivalent coverings.
A topological space Y is locally arcwise connected if for every point
y ∈ Y and every neighborhood U of y there is a smaller neighborhood U ′ ,
y ∈ U ′ ⊂ U , such that every two points a, b ∈ U ′ can be connected by a
path in U . A space Y is locally simply connected in the large or semi-locally
simply connected if every point y ∈ Y has a neighborhood U such that every
loop contained in U is contractible in Y . (In other words, the image of the
fundamental group π1 (U, y) under the homomorphism π1 (U, y) → π1 (Y, y)
induced by the inclusion of U to Y is trivial.)

Theorem 3.4.11. If a topological space Y is a connected, locally arcwise

connected and locally simply connected in the large, then there exists a
universal covering f : X → Y . A universal covering is unique up to an
equivalence.
3.4. Local Isometries and Coverings 81

These conditions on Y in the theorem look natural and hold for most
topological spaces except “pathological” ones. From now on, we consider
only topological spaces satisfying these conditions.
One of the most important features of universal coverings is their relation
to fundamental groups. Let f : X → Y be a covering map. Fix y0 ∈ Y and
choose an x0 ∈ f −1 (y0 ). Then every path γ in Y starting at x0 has a unique
lift γ̃ in X starting at x0 (recall that γ̃ being a lift of γ means that f ◦ γ̃ = γ).
In particular, a loop with endpoints at y0 is lifted to a curve starting at x0
and ending at some point of f −1 (y0 ). The standard “covering homotopy”
lemma tells that lifts of homotopic paths are homotopic (by a homotopy
of paths we mean a homotopy with fixed endpoints). In particular, lifts
of homotopic paths have the same endpoints; i.e., if two loops γ1 and γ2
with endpoints at y0 are homotopic, then their lifts γ̃1 and γ̃2 starting at x0
end at the same point y ∈ f −1 (y0 ). If f is a universal covering, then the
converse statement is also true: every two paths in X connecting a given
pair of points are homotopic and hence they are lifts of homotopic paths
in Y .
Thus we have a 1-1 correspondence between classes of homotopic loops
in Y with endpoints at y0 (that is, the fundamental group π1 (Y, y0 )) and
the elements of f −1 (y0 ) ⊂ X.

Regular coverings and deck transformations. Actually not all cover-

ings are equally interesting from the geometrical point of view. We want to
look more closely at the class of regular coverings. We note in advance that
all universal coverings are regular.
Recall that every continuous map f : X → Y induces a homomorphism
of fundamental groups. Namely, if x0 ∈ X and y0 = f (x0 ), there is a
natural homomorphism f∗ : π1 (X, x0 ) → π1 (Y, y0 ) which maps the class of
a loop γ in X to the class of the loop f ◦ γ in Y . If f is a covering map,
the covering homotopy lemma implies that the image of a noncontractible
curve is noncontractible. In other words, f∗ is a monomorphism (that is, an
injective homomorphism) from π1 (X, x0 ) to π1 (Y, y0 ).
So the image f∗ (π1 (X, x0 )) is a subgroup of π1 (Y, y0 ) isomorphic to
π1 (X, x0 ). This subgroup is referred to as the group of covering. Geometri-
cally, it consists of all loops with endpoints at y0 whose lifts starting at x0 are
loops. Note that this subgroup may depend on the choice of x0 ∈ f −1 (y0 )
despite the fact that all subgroups obtained this way are isomorphic.
A covering map f : X → Y is said to be regular if the following
equivalent properties hold:
• f∗ (π1 (X, x0 )) is a normal subgroup of π1 (Y, y0 ).
• f∗ (π1 (X, x0 )) does not depend on x0 ∈ f −1 (y0 ).
82 3. Constructions

A deck transformation of a covering f : X → Y is a covering equivalence

from X to itself, that is, a homeomorphism h : X → X such that f = f ◦ h.
Geometrically, a deck transformation is a homeomorphism from X to itself
which “permutes the sheets” of the covering, more formally, permutes the
points of f −1 (y) for any y ∈ Y . The deck transformations obviously form
a group. A covering f : X → Y is regular if and only if this group acts
“transitively on sheets”, i.e.,

• for every y ∈ Y , x, x′ ∈ f −1 (y) there exists a (unique) deck

transformation h : X → X such that h(x) = x′ .

This can be viewed as yet another equivalent definition of a regular covering.

It is not hard to show that the group of deck transformations is iso-
morphic to the quotient group π1 (Y, y0 )/f∗ (π1 (X, x0 )) where y0 = f (x0 ).
In particular, if f : X → Y is a universal covering, then its group of deck
transformations is isomorphic to π1 (Y ).

Example 3.4.12. For the covering f : R → S 1 given by f (x) = (cos x, sin x)

the deck transformations are maps from R to R of the form x 7→ x + 2πk
where k is an integer. So the group of deck transformations is just Z acting
on R by translations.
Similarly, the group of deck transformations of the standard covering
R2 → T 2 is Z2 acting on R2 by parallel translations.
For the standard covering S 2 → RP2 , the group of deck transformations
consists of two elements, the identity and the reflection x 7→ −x.

Now let Y be a length space and X be equipped with the length metric
lifted from Y . Then every deck transformation is an isometry from X to
itself (prove this!). In the case of a universal covering, this yields an action
of π1 (Y ) on X by isometries.
This fact (that the fundamental group of a length space acts on its
universal covering space by isometries) is very important. Some of its
numerous applications can be found in this book.

There is another (but equivalent) approach to this subject based on the

fact that the space Y and the covering f : X → Y are determined by the
group of deck transformations. Namely Y is the quotient of X by this group
action, and f is the projection to the quotient. For example, S 1 ≃ R/Z and
T 2 ≃ R2 /Z2 .
Of course, not every group action on a topological space can arise as
a group of deck transformations of a covering. A necessary and sufficient
condition for this is contained in the following definitions.
3.4. Local Isometries and Coverings 83

Definition 3.4.13. Let a group G act on a set X. This action is said to be

free if gx 6= x for all x ∈ X, g ∈ G \ {e}.

Definition 3.4.14. Let a group G act on a topological space X. This action

is said to be totally discontinuous if every point x has a neighborhood U such
that gU ∩ U = ∅ for all g ∈ G such that gx 6= x.

It is easy to see that a group of deck transformations of a covering acts

freely and totally discontinuously. Conversely, every free totally discontinu-
ous group action is a group of deck transformations of a covering:

Proposition 3.4.15. Let a group G act on a topological space X freely and

totally discontinuously. Then the projection X → X/G is a covering map.
Moreover this is a regular covering and the group of its deck transformations
coincides with G.

Proof. The proof is trivial but it is a good test for understanding the
definitions.
Let f : X → X/G be the projection. If x ∈ X and U is a neighborhood
of x from the definition of totally discontinuous action, then V = f (U ) is an
evenly covered neighborhood of f (x). Indeed, f −1 (V ) is the disjoint union
of the sets gU over g ∈ G, and f |gU is a homeomorphism from gU to V .
Obviously every element of G (considered as a homeomorphism from X
to itself) is a deck transformation. If y ∈ Y and x, x′ ∈ f −1 (y), then there is
an element g ∈ G such that gx = x′ . This means that G acts transitively on
sheets; hence f is a regular covering. On the other hand, there is at most one
deck transformation which maps x to x′ ; therefore every deck transformation
is represented by an element of G. ¤

Consider a covering f : X → X/G where G acts on X freely and totally

discontinuously. We have seen that every length metric on X/G can be lifted
to X so that f becomes a local isometry. The lifted metric is G-invariant,
which simply means that G acts by isometries with respect to this metric.
Conversely, every G-invariant length metric on X is a lift of a (unique)
length metric on X/G. This is nothing but the standard metric of the
quotient space defined in Section 3.3 (see also Exercise 3.4.6). We summarize
these observations in the following proposition.

Proposition 3.4.16. Let f : X → Y be a regular covering and G its

group of deck transformations. Then the length metrics on Y are in 1-
1 correspondence with the G-invariant length metrics on X so that for
corresponding metrics dX on X and dY on Y f is a local isometry from
(X, dX ) to (Y, dY ).
84 3. Constructions

3.4.3. Local isometries of complete spaces. In general, a local isome-

try is not necessarily a covering map; for instance, consider Example 3.4.3
and replace the domain of f by an open segment instead of R. Neverthe-
less, for a complete length space X a local isometry f : X → Y somewhat
resembles a covering map. In particular, geodesics (and shortest paths) in
Y can be lifted to X:
Lemma 3.4.17. Let f : X → Y be a local isometry of a complete length
space X. Given a geodesic (shortest path) γ : [0, a] → Y and a point x0
such that f (x0 ) = γ(0), there exists a unique geodesic (resp. shortest path)
γ̃ : [0, a] → X such that γ̃(0) = x0 and f (γ̃(t)) = γ(t).

Proof. The proof is similar to the construction for lifts of paths in the
theory of covering spaces. Assume that γ is parameterized by arc-length.
Consider the subintervals [0, t) of [0, a] such that γ|[0,t) can be lifted to X.
This means that there is a path γ̃t : [0, t) → Y such that γ|[0,t) = f ◦ γ̃
together with γ̃(0) = x0 . The set of such subintervals is not empty since the
restriction of f to some neighborhood of x0 is a homeomorphism onto its
image. Let [0, t0 ) be the union of all such subintervals, i.e., the maximum
interval that admits a lifting. We have a path γ̃t0 : [0, t0 ) → Y such that
γ|[0,t) = f ◦ γ̃t0 . Choose a sequence {ti } such that ti < t0 and ti → t0 as
i → ∞. Then γ̃t0 (ti ) is a Cauchy sequence and, since X is complete, it tends
to a point p. Clearly the choice of p is independent of the choice of {ti }.
Since the points f ◦ γ̃t0 (ti ) = γ(ti ) converge to γ(t0 ) as i → ∞, one can set
p to be the lift of γ(t0 ) and thus γ is lifted on the closed interval [0, t0 ]. If t0
is not equal to a, then γ could be lifted on a larger interval [0, t0 + ε) since
f is a local homeomorphism on a neighborhood of p. Since this contradicts
the maximality of [0, t0 ), we conclude that t0 = a. ¤

This lemma yields the following useful corollary:

Theorem 3.4.18. Let f : X → Y be a surjective local isometry of a
complete locally compact length space X. Assume that each point in Y has a
neighborhood such that every geodesic segment contained in our neighborhood
is a unique shortest path between its end-points. Then f is a covering map.
Remark 3.4.19. Actually the theorem is still correct under a weaker
condition: that there is an unique shortest part connecting every two points
of the neighborhood. (the condition allows existence of geodesics which are
not shortest paths).However the proof is more complicate in this case.

Proof. By the definition of covering maps, we need to show that every

q ∈ Y has a neighborhood Uq whose pre-image consists of pairwise disjoint
open sets each mapped homeomorphically onto Uq .
3.5. Arcwise Isometries 85

The informal idea of the proof is to represent a small ball centered at q

as a bunch of shortest paths from q leading to all points of the ball. Once
we choose a pre-image of q, each of the shortest paths can be lifted as a
shortest path starting at this pre-image. Thus the whole ball gets lifted (as
a “porcupine”) to every pre-image of q.
Choose Uq to be a metric ball B = Br (q). Choose r to be so small
that each point y ∈ B r (q) is connected with q by a unique geodesic (in
particularly, this geodesic is a shortest path). Fix a point p ∈ f −1 (q). By
Lemma 3.4.17, for a point y ∈ B, the shortest curve [q, y] can be lifted to
X as a shortest curve [p, x]. Introduce a map gp given by x = gp (y) and
let Vp = gp (B) and V p = gp (B r (q)). Obviously, Vp ⊂ Br (p). From other
hand, if x ∈ Vp , then the f -image of a shortest path [px] is a geodesic and,
therefore, the shortest path. So f (x) ∈ Vq . This mean that Vp = Br (p), in
particularly, Vp is an open set.
By construction, f (gp (y)) = y and gp (f (x)) = x for y ∈ B and x ∈ V p .
Thus the restriction of f onto V p is a bijection onto B r (q) and hence a
homeomorphism (since B r (q) is compact). Thus f|Vp is a homeomorphism
onto B.
S
It remains to show that all Vp are disjoint and f −1 (B) = p∈P Vp .
Indeed, assume that x ∈ Vp1 ∩ Vp2 . By the construction of Vp1 and
Vp2 , there are two shortest curves [p1 , x] and [p2 , x] that are the lifts of the
shortest [q, f (x)]. Lifting the shortest curve [f (x), q] “from the other end”,
that is, so that it starts at x, and using the uniqueness part of Lemma 3.4.17,
we conclude that p1 = p2 .
S
Finally, let us show that f −1 (B) = p∈f −1 B Vp . Since we know that
f (Vp ) = B for all p, it remains to show that a point x such that f (x) ∈ B
belongs to one of the sets Vp . Consider the lift [x, p] of the shortest path
[f (x), q]. Then p belongs to f −1 B and thus x ∈ Vp . ¤

3.5. Arcwise Isometries

Since length structure is our primary notion, we can look for maps that
preserve this structure. Namely,
Definition 3.5.1. A map f between two length spaces is called an arcwise
isometry if L(γ) = L(f (γ)) for every path γ.
Note that we do not require f to be bijective, since otherwise we would
get nothing but usual isometries.
Exercise 3.5.2. 1. Prove that a bijective arcwise isometry is an isometry.
2. Prove that an arcwise isometry that is also a local homeomorphism
is a local isometry.
86 3. Constructions

3. Give a definition of local arcwise isometry and prove that local arcwise
isometries are just arcwise isometries (since length structures are local).

For instance, a path parameterized by arc length is an arcwise isometry

from a segment into a length space. Here we arrive at the following
important notion.

3.5.1. Isometric embeddings. Injective arcwise isometries are also called

isometric embeddings. There is often a terminological confusion around
this notion. We stress that isometric embedding is not the same notion as
isometry onto its image. For instance, a simple curve γ : [0, 1] → R2 , γ(t) =
(cos(t), sin(t)) is an isometric embedding, although it is not an isometry
on its image (w.r.t. the restriction of Euclidean distance). Isometric
embeddings that are studied in differential geometry (such as isometric
embeddings of surfaces) are actually arcwise isometric embeddings. On the
other hand, a sphere S 2 already does not admit an embedding f : S 2 → Rn
that would be an isometry onto its image. This, however, becomes possible
if one substitutes an infinite-dimensional space instead of Rn as the target
space for f . This fascinating construction is worth explaining here:
Example 3.5.3 (embedding by an isometry onto its image). Let X be a
compact length space. Consider the space C(X) of all continuous functions
X → R. This space turns into a metric space by introducing the uniform
distance between functions d∞ (f, g) = sup |f (x) − g(x)|. The map E : X →
C(X) defined by E(x) = d(x, ·) (in other notation, E(x) evaluated at y is
d(x, y)) is an embedding that is an isometry onto its image.
Exercise 3.5.4. Verify that
(i)(C(X), d∞ ) is indeed a metric space. (In addition, this is also a
normed vector space; we use the notation d∞ since this is actually the L∞
norm on this vector space).
(ii) d∞ (d(x, ·), d(y, ·)) = d(x, y), that is, supz (|d(x, z) − d(y, z)|) =
d(x, y).

3.5.2. Surjective arcwise isometries. While bijective arcwise isometries

are just isometries, surjective arcwise isometries can be extremely wild even
for nice spaces. The reason for this is that they may crease the space many
times. The simplest example of a pleat appears already in arcwise isometries
R → R. For instance, consider f : R → R given by f (t) = t if t ≤ 0,
f (t) = −t for t ∈ [0, 1] and f (t) = t − 2 if t ≥ 1.
At first glance it may seem striking that every 2-dimensional polyhedral
surface admits an arcwise isometry onto a polygonal region. This statement,
however, becomes much less surprising after a practical experiment: make a
3.6. Products and Cones 87

model of a polyhedron using scissors, thick paper and glue, and then flatten
this model by stepping on it.
It is much more surprising that every surface homeomorphic to a sphere
admits an arcwise isometry onto a round sphere. This map is not at all
as nice as the one that can be used for polyhedra: for a generic surface, it
has everywhere dense folds and singular points. Although the proof of the
existence of such arcwise isometries for surfaces is beyond the frame of this
course, we sketch two absolutely elementary arguments proving that every
2-dimensional polyhedral space can be folded onto a polygon.
Proposition 3.5.5. Every (locally-)finite 2-dimensional polyhedral space
admits an arcwise isometry onto a planar polygon. In other words, every
2-dimensional polyhedral metric can be induced by a map to R2 .

Proof. Let P be a 2-polyhedron, and let A = {ai } be a finite set of its

points. Define Qi = {x ∈ P : |xai | ≤ |xaj | for all j 6= i}. (These sets Qi
are called Dirichlet-Voronoi regions.) It is easy to see that every set Qi is
bounded by a graph Li whose edges are geodesic curves.
The set A can be chosen so that each Qi possesses a triangulation with
the one vertex ai inside Qi . To achieve this, let A be a sufficiently fine net
including all vertices of P and a fine net in each edge of P . Still some pairs
of points ai ∈ A may be connected with more than one shortest path. One
can add new points to A so that, for the new set A, points ai , aj ∈ A are
connected with a unique shortest curve if Li and Lj have common points.
S
Adding to the graph L = Li new edges ai xj , where xj are vertices
of Li , we obtain a triangulation of P such that every pair of triangles with
different vertices ai either have no more than one common point (vertex) or
have a common edge which belongs to L.
Observe that all triangles with a common edge [bc] ⊂ L are isometric
(by the definition of Qi ). Fix a ray l in R2 starting at a point a. Imagine
that our polyhedron P is made of paper. Then we can cut each triangle
△ai bc out of P and then bend it around the bisector of the angle ai so that
directions of ai b and ai c coincide. Now we place all bent triangles in the
plane R2 on one side of l and such that all the vertices ai coincide with a
and all the edges [ai b], [ai c] go along l. So all the edges we cut P along are
placed in l and start at a. Therefore we can glue together again all such
edges and this gives us the desired arcwise isometry. ¤

3.6. Products and Cones

3.6.1. Direct products. Among various mathematical constructions for
building new objects (such as products or quotients in algebra), there are
88 3. Constructions

several natural ways of constructing new length spaces. The simplest of

them is direct product. Let (X, dX ) and (Y, dY ) be length spaces.
Equip the direct product Z = X × Y with the metric
q
d((x1 , y1 ), (x2 , y2 )) = d2X (x1 , x2 ) + d2Y (y1 , y2 ),

where x1 , x2 ∈ X and y1 , y2 ∈ Y . This formula is motivated by the

Pythagorean theorem.
It is easy to see that d is a metric. This metric is called the product
metric, and the metric space (Z, d) is called the direct metric product of X
and Y .
Proposition 3.6.1. The direct product of strictly intrinsic metrics is a
strictly intrinsic metric. The direct metric product of two length spaces is a
length space.

Proof. We will carry out the argument for strictly intrinsic spaces; the
second part is left as an exercise. By Theorem 2.4.16, it is sufficient to
show that there is a midpoint between two given points z1 = (x1 , y1 ) and
z2 = (x2 , y2 ) in Z. For the midpoints xm and ym between x1 , x2 and y1 , y2 ,
resp., one sees that

d2 ((xm , ym ), z1 ) = d2X (xm , x1 ) + d2Y (ym , y1 )

1 1
= (d2X (x1 , x2 ) + d2Y (y1 , y2 )) = d2 (z1 , z2 ),
4 4
and hence d((xm , ym ), z1 ) = 12 (d(z1 , z2 )). Analogously d((xm , ym ), z2 ) =
1
2 (d(z1 , z2 )). This shows that (xm , ym ) is a midpoint for z1 , z2 . ¤
Exercise 3.6.2. To put your hand on this notion, we suggest you consider
the examples of R×S 1 , R×S 2 , S 2 ×S 2 . In particular, try to find all shortest
curves in and all isometries of these spaces. The two latter product spaces
contain a whole bunch of (nicely embedded) surfaces (tori) that are locally
isometric to the Euclidean plane. Finding these tori may be very helpful for
understanding the sequel.

The product metric d constructed this way enjoys the following property:
consider a fiber Sy = {(x, y) : y = const} together with the restriction of the
metric d. Then the projection map of Sy to X is an isometry. (In particular,
the restriction of d is an intrinsic metric.) This follows immediately from the
definition of the product metric. Evidently, the same holds for the “vertical”
fibers Sx .
In addition, the group of isometries of (Z, d) is at least as rich as the
product of the isometry groups of X and Y .
3.6. Products and Cones 89

More precisely, every isometry IX of X extends on Z as IX × id.

Moreover, for isometries IX of X and IY of Y , one canonically assigns an
isometry IX × IY : (x, y) → (IX (x), IY (y)) of Z, and therefore the isometry
group of Z contains an isomorphic copy of the product of the isometry groups
of X and Y . The reader can check that, for X = Y = R, the isometry group
of the product is much richer than the product of isometry groups.

Remark 3.6.3. One should not think that our formula for product metrics
is the only possible definition of a product metric on Z enjoying these nice
properties. For instance, by setting d((x, y), (x′ , y ′ )) = dX (x, x′ ) + dY (y, y ′ )
one gets an intrinsic metric that possesses all the properties listed above.
Prove this!
More generally, every norm k , k on R2 such that its restrictions to the
rays {x0 , y > 0} and {x > 0, y0 } are monotone give rise to a construction of
product metrics by assigning d(z, z ′ ) = kdX (x, x′ ), dY (y, y ′ )k. Moreover, if
one axiomatizes the desirable properties of product metrics, then all possible
constructions arise this way. The reason we used the Pythagorean theorem
and thus the standard Euclidean norm for k , k in our preferred definition
will become clear later (as a glimpse ahead, let us mention that this is the
only definition that respects the boundedness of curvature).

We leave as a not difficult exercise to justify the following description of

all shortest paths of the metric space (Z, d):

Lemma 3.6.4. A constant-speed path in Z is a shortest path (a geodesic)

if and only if it is the product of two shortest paths (geodesics) in X and Y
with constant-speed parameterizations.

Notice that the projections of a shortest curve γ in Z to both X and

Y are shortest paths regardless of whether the parameterization of γ is
constant-speed or not. However, the other implication essentially uses this
assumption: in the product R2 = R × R, every curve (x(t), y(t)) with
increasing coordinates x(t), y(t) is the product of two shortest paths x(t)
and y(t).

Convex subsets. There is an important notion that we want to discuss in

connection with the statement of the lemma. While involving this notion
might seem a little bit superfluous in this simple situation, it proves very
useful in more complicated cases.
We begin with the following important definition:

Definition 3.6.5. A subset A in a metric space (X, d) is said to be convex

if the restriction of d to A is strictly intrinsic and finite.
90 3. Constructions

If the metric of X is strictly intrinsic, a set A is convex if and only if for

every two points x, y ∈ A, there is a shortest path between x and y which
entirely belongs to A. (Prove this!) This explains the name convex.
Being convex is a “global” property of a set. Let us introduce an
analogous local property.
Definition 3.6.6. Suppose that the metric of X is strictly intrinsic. A set
A in X is said to be locally convex if every point x ∈ A has a neighborhood
U in A such that for every two points y, z ∈ U , there is a shortest path
between y and z contained in A.
Exercise 3.6.7. Give a definition of local convexity for general metric
spaces.
Remark 3.6.8. For a reader familiar with differential geometry, we mention
that there is a special term for locally convex submanifolds (in particular,
surfaces) in Riemannian and Finsler manifolds. Such submanifolds are said
to be totally geodesic and may be alternatively defined as submanifolds
in which all geodesics (of the induced metric) are geodesics of the ambient
space at the same time. A submanifold of a Riemannian manifold is totally
geodesic if and only if if its second fundamental form vanishes, i.e., all
principal curvatures are zero.

The following lemma follows immediately from the definition:

Lemma 3.6.9. If X is a space with strictly intrinsic finite metric and
F : X → Y is a distance-preserving map, then the image Im(F ) := F (X) is
convex in Y .
Exercise 3.6.10. Prove the lemma.
Proposition 3.6.11. Let X and Y be length spaces, and α : [a, b] → X,
β : [c, d] → Y shortest paths. Then the product of their images R =
Im(α) × Im(β) is convex in X × Y and isometric to a Euclidean rectangle.

Proof. We may assume that both α and β are parameterized by arc length.
Introduce F : [a, b] × [c, d] → Z given by F (t, s) = (α(t), β(s)). This map is
an isometry:
d2 (F (t, s), F (t′ , s′ )) = d2X (g(t), g(t′ )) + d2Y (h(t), h(t′ )) = (t − t′ )2 + (s − s′ )2 .
Applying Lemma 3.6.9 finishes the proof. ¤

3.6.2. Cone over a metric space. A cone Con(X) over a topological

space X is the quotient of the product X ×[0, ∞) obtained by gluing together
(identifying) all points in the fiber X × {0}. This point is called the origin
(or apex) of the cone.
3.6. Products and Cones 91

How should one equip a cone with a metric? To get an idea, imagine
that X is a subset of the unit sphere S 2 = {(x, y, z) : x2 + y 2 + z 2 = 1} ⊂ R3
equipped with the spherical (angular) metric. To build a cone over X, we
draw a ray from the origin through every point x ∈ X. Thus a point a in
the cone can be described by as (x, r), where x is a point in X that belongs
to the ray Oa and r = |aO| the distance from the origin; the latter is a
nonnegative number. Thinking of a and x as vectors, we can write a = rx.
How can we express the Euclidean distance between two points a = (x, t)
and b = (y, s) in terms of x, y, t, s? Consider the triangle △Oab. We have
|Oa| = t, |Ob| = s, and the angle ∡aOb between the sides equals the angular
distance d(x, y) in X. Hence by the cosine formula
p
|ab| = t2 + s2 − 2ts cos(d(x, y)).
We will use this formula to define cone metrics in general:

Definition 3.6.12. Let X be a metric space with diam(X) ≤ π. The cone

metric dc on Con(X) is given by the formula
p
(3.1) dc (p, q) = t2 + s2 − 2ts cos(d(x, y)),
where p, q ∈ Con(X), p = (x, t), q = (y, s).

Proposition 3.6.13. If X is a metric space with diam(X) ≤ π, then dc is

a metric.

Proof. Positiveness and symmetry for dc are trivial. Let us prove the
triangle inequality. Consider three points y1 = (x1 , r1 ), y2 = (x2 , r2 ) and
y3 = (x3 , r3 ) in Con(X). Denote α = d(x1 , x2 ) and β = d(x2 , x3 ). Construct
three points y 1 , y 2 , y 3 ∈ R2 so that their distances from the origin O equal
r1 , r2 and r3 respectively, ∡y 1 Oy 2 = α, ∡y 2 Oy 3 = β, and the rays Oy 1 and
Oy 3 are in different half-planes with respect to Oy 2 . Then |y 1 y 2 | = dc (y1 , y2 )
and |y 2 y 3 | = dc (y2 , y3 ). Now we have two cases: α + β ≤ π and α + β > π.
If α+β ≤ π, then ∡y 1 Oy 3 = α+β ≥ d(x1 , x3 ); hence |y 1 y 3 | ≥ dc (y1 , y3 ).
Then the triangle inequality for y1 , y2 and y3 follows from the triangle
inequality in R2 :
dc (y1 , y2 ) + dc (y2 , y3 ) = |y 1 y 2 | + |y 2 y 3 | ≥ |y 1 y 3 | ≥ dc (y1 , y3 ).
If α + β > π, the planar triangle inequality does not help, but there is a
better estimate for |y 1 y 2 | + |y 2 y 3 |. Indeed, since the broken line y 1 y 2 y 3 lies
outside the sector y 1 Oy 3 , we have |y 1 y 2 | + |y 2 y 3 | ≥ |y 1 O| + |Oy 3 |. Then
dc (y1 , y2 ) + dc (y2 , y3 ) ≥ |y 1 O| + |Oy 3 | = r1 + r3 ≥ dc (y1 , y3 )
(the last inequality follows from the definition of dc ). ¤
92 3. Constructions

Exercise 3.6.14. Let γ̃ : [a, b] → Con(X) be a curve in the cone, γ̃(t) =

(γ(t), r(t)) where γ is a curve in X. Prove that
p
L(γ̃) ≥ r(a)2 + r(b)2 − 2r(a)r(b) cos(L(γ))
if L(γ) ≤ π, and
L(γ̃) ≥ r(a) + r(b)
if L(γ) ≥ π.
Hint: Repeat the argument proving the triangle inequality.

If the metric of X is intrinsic, then dc is intrinsic too. To prove this

one could write an explicit expression for a midpoint and a shortest path
between two points, as we did in case of products. For cones, this is a tedious
approach with cumbersome formulas. It is more convenient to begin with
flat convex surfaces like those we observed in direct products.
To see where they come from, let us come back to the example of X
being a subset of the unit sphere. For simplicity assume that X is the whole
sphere; then the cone is the whole space R3 . Since a shortest path γ is
nothing but an arc of a greater circle, the cone over γ : [0, a] → S 2 is a
planar sector. A point in this sector has cone coordinates (γ(τ ), t). For γ(τ )
being parameterized by arc length, τ and t are the usual polar coordinates in
the planar sector with the angular coordinate measured from the direction
of the ray [O, γ(0)).
Let us see how this works in general. Let γ : [0, L] be a shortest path
in X. Introduce a polar coordinate system (r, ϕ) on the Euclidean plane and
denote by Q the set of points in the plane whose ϕ-coordinate is between 0
and L. Consider the map F : Q → Con(X) given in polar coordinates by
the formula F (r, ϕ) = (γ(ϕ), r). The image of this map is the cone over γ.
A simple calculation shows that the map F is distance-preserving. In-
deed,
2
d2c (F (r, ϕ), F (r′ , ϕ′ )) = r2 + r′ − 2rr′ cos dX (γ(ϕ), γ(ϕ′ ))
2
= r2 + r′ − 2rr′ cos(ϕ − ϕ′ ) = dR2 ((r, ϕ), (r, ϕ)).
This immediately implies that F (Q) is flat and convex (Lemma 3.6.9). In
particular, the image of every segment containing Q is a shortest segment
in Con(X) and hence dc is an intrinsic metric.
Notice that we have actually proved the following useful lemma:

Lemma 3.6.15. If X is a length space with diam(X) ≤ π, γ is a shortest

segment in X, then the cone over the image of γ is a convex flat surface in
the cone Σ(X) over X.
3.6. Products and Cones 93

Conversely, consider a shortest path γ : [a, b] → Con(X) in the cone not

passing through the origin. We can write γ(t) = (γ(t), r(t)) where r(t) ∈ R
and γ is a curve in X (γ is called the projection of γ to X). Looking at the
proof of the triangle inequality in the cone (Proposition 3.6.13), one sees
that the triangle inequality must turn to equality for any three points γ(t1 ),
γ(t2 ), γ(t3 ) such that t1 < t2 < t3 . This implies that L(γ) = d(γ(a), γ(b));
hence γ is a shortest path in X.
Thus we have a 1-1 correspondence between shortest paths in X of length
strictly less than π and shortest paths in Con(X) not passing through the
origin. As for shortest paths passing through the origin, it is easy to see the
following. Every point (x, r) ∈ Con(X) is connected to the origin by a unique
shortest path, namely the “segment” {(x, t)}t∈[0,r] . The concatenation of two
such segments with endpoints (x1 , r1 ) and (x2 , r2 ) is a shortest path if and
only if d(x1 , x2 ) = π.

Cone over a large space. Now we drop the assumption that diam(X) ≤
π. The formula (3.1) is not suitable for defining a metric on Con(X) if
diam(X) > π; for example, the triangle inequality for dc may fail. How
does one define the cone metric in the general case? The guidelines are the
following: we want the formula (3.1) to hold for “small” distances in X, and
we want a cone over a length space be a length space.
If X is a length space, the existence and uniqueness of such a metric
is guaranteed by Lemma 3.1.2. In fact, this metric has a simple explicit
description:
Definition 3.6.16. Let (X, d) be a metric space. The cone distance dc (a, b)
between points a = (x, t) and b = (y, s) in Con(X) is defined as
(p
t2 + s2 − 2ts cos(d(x, y)), d(x, y) ≤ π,
dc (a, b) =
t + s, d(x, y) ≥ π.

Alternatively, one can introduce a new distance d on X by

d(a, b) = min{d(a, b), π}
and define Con(X, dc ) = Con(X, d) where the metric on Con(X, d) is given
by Definition 3.6.12. Since d is a metric, Proposition 3.6.13 implies that dc
is a metric. We summarize the above observations about intrinsic metrics
in the following theorem.
Theorem 3.6.17. The metric dc on Con(X, d) is intrinsic (resp. strictly
intrinsic) if and only if the metric d is intrinsic (resp. strictly intrinsic)
at distances less than π. The latter means that for any x, y ∈ X such
that d(x, y) < π there is a curve in X connecting x and y whose length is
arbitrarily close (resp. is equal to) d(x, y).
94 3. Constructions

Proof. First suppose that d is strictly intrinsic at distances less than π. Let
x, y ∈ X, a, b ∈ Con(X), a = (x, t), b = (y, s). If d(x, y) < π, apply Lemma
3.6.15 to a shortest path γ connecting x and y. It follows that there is a curve
of length dc (a, b) connecting a and b. If d(x, y) ≥ π, then dc (a, b) = t + s and
there is a curve of length t + s connecting a and b, namely, the union of the
two segments connecting a and b to the origin. Thus dc is strictly intrinsic.
Conversely, suppose that the metric dc is strictly intrinsic. For any two
points x, y ∈ X with d(x, y) < π apply the result of Exercise 3.6.14 to a
shortest path γ̃ connecting the points a = (x, 1) and b = (y, 1) in the cone.
Since L(γ) = dc (a, b) < 2, γ̃ does not pass through the origin and hence
has a well-defined (and continuous) projection γ in X. The inequality from
Exercise 3.6.14 implies that L(γ) = d(x, y).
The proof for non-strictly intrinsic metrics is similar and is left to the
reader. ¤

Exercise 3.6.18. Let X be a metric space with diam(X) = π. Suppose

that Con(X) is a length space but X is not. Prove that there are three
distinct points x, y, z ∈ X such that |xy| = |xz| = π.

Exercise 3.6.19. Let X be a line segment of length α, 0 < α < 2π. Prove
that Con(X) is isometric to the planar sector of angular measure α with its
intrinsic metric.

This exercise is a partial case of the next proposition which says that
every polyhedral space locally looks like a cone. For simplicity we restrict
ourselves to 2-dimensional polyhedral spaces. In higher dimensions, one
should consider cones over spherical polyhedral spaces.

Proposition 3.6.20. Let p ∈ P be a point in a two-dimensional polyhedral

space P . For all sufficiently small r > 0, the ball Br (p) is isometric to an
r-ball in Con(G) centered at the origin of the cone over a graph G.

Exercise 3.6.21. Prove the proposition.

Hint: First prove that the construction of a cone “commutes” with
gluing; i.e., if a space space X is obtained by gluing from a space Y , then the
cone Con(X) can be obtained by gluing respective rays in the cone Con(Y ).

The graph G mentioned in the proposition is called the link of P at p. If

p belongs to the interior of a face, the link is just a circle of length 2π. For a
point in the interior of an edge, the link is a graph consisting of two vertices
connected by several segments of length π (each segment corresponds to a
face of P adjacent to the edge containing p). If p is a vertex, any graph may
appear as the link at p.
3.6. Products and Cones 95

3.6.3. Spherical suspensions. There are other constructions with famil-

iar Euclidean ancestors. For instance, let us simulate the procedure building
S n out of its equator S n−1 by means of adding two poles and drawing a semi-
circle (meridian) through every point of S n−1 . Begin with the direct product
X × I of a topological space X and a segment I = [0, a] and contract the
fibers X × 0 and X × a each to a point. The resulting space Σ(X) is called
the spherical cone, or the suspension, over X.
If (X, d) is a length space with diam X ≤ π, we choose a = π and define
the distance dΣ by the equation
cos dΣ (p, q) = cos t cos s + sin t sin s cos d(x, y)
for all points p = (x, t), q = (y, s) in Σ(X). This formula is certainly
suggested by the cosine theorem in spherical geometry.

Exercise 3.6.22. 1. For the standard sphere S n , prove that Σ(S n ) is

isometric to the standard sphere S n+1 .
2. Similarly as for cones and products, the suspension over a geodesic
in X is a convex (totally geodesic) surface in Σ(X). In this case, however,
it is not flat. This surface is locally isometric to the standard unit sphere.
Prove these statements.
3. Arguing as in the cone case, prove that dΣ is an intrinsic metric.
Notice: This exercise assumes some knowledge of spherical geometry,
such as the spherical cosine formula.

3.6.4. Warped products. Let us mention a construction which general-

izes direct (metric) products, cones, and suspensions.
Let X and Y be two (complete) length spaces and f : X → R a
positive continuous function. Consider a Lipschitz curve γ in X × Y whose
projections to X and Y are γ1 : [a, b] → X and γ2 : [a, b] → Y . To equip
X × Y with a warped product metric, define the length of γ by the formula
Z bq
L(γ) = |γ1′ |2 (t) + f 2 (γ1 (t))|γ ′2 |2 (t) dt,
a

where |γ1′ |, |γ2′ |

are defined a.e. (see 2.7.6). The metric on X × Y induced
by this length structure is called the warped product metric. One denotes
by M ×f N the space X × Y equipped with this metric.
It is obvious that direct metric product is a particular case of warped
product (with f ≡ 1).
Instead of a positive function f , it is sometimes natural to consider f
vanishing at certain points. Then one obtains X × Y with a warped product
semi-metric. As usual, this semi-metric in its turn gives rise to a metric
96 3. Constructions

after points at zero distance from each other are identified. This new space
with the resulting metric will also be called a warped product.
Exercise 3.6.23. Let Con(X) be the cone over a length space X, and
diam X < π. Prove that Con(X) = [0, ∞) ×f X, where f (t) = t.
Exercise 3.6.24. Let Σ(X) be the spherical suspension over a length space
X. Show that Σ(X) = [0, π] ×f X, where f (t) = sin t.

3.6.5. Definition of angle. Since the notion of angle belongs to basic

notions of Euclidean geometry, it is desirable to define angles in our metric
context. To figure out how we could measure angles in a metric space, let
us first express usual Euclidean angles in purely metric terms. Consider two
rays α : [0, ∞[→ R2 and β : [0, ∞[→ R2 emanating from the same point
a = α(0) = β(0). Picking two positive numbers t, s and applying the cosine
theorem to the triangle aα(t)β(s), we express the angle between the rays as
|aα(t)|2 + |aβ(s)|2 − |α(t)β(s)|2
arccos .
2|aα(t)||aβ(s)|
The fact that this expressions is independent of the choice of t and s is
certainly a lucky coincidence: should we use two smooth curves instead of
rays, we would get a nonconstant function in t and s; thinking of angle as
an infinitesimal notion suggests passing to the limit as t and s tend to zero.
Definition 3.6.25. Let x, y, z be three distinct points in a metric space
e
(X, d). The comparison angle xyz, denoted by ∡xyz e y, z), is defined
or ∡(x,
by
e d(x, y)2 + d(y, z)2 − d(x, z)2
∡xyz = arccos .
2 d(x, y) d(y, z)
The geometric meaning of this definition is the following. Let △xyz be
a triangle in R2 whose sides |xy|, |yz| and |xz| equal the respective distances
d(x, y), d(y, z) and d(x, z). (Such a triangle is uniquely defined up to a rigid
motion.) Then ∡xyze = ∡xyz.
Definition 3.6.26. Let α : [0, ε) → X and β : [0, ε) → X be two paths in a
length space X emanating from the same point p = α(0) = β(0). We define
the angle ∡(α, β) between α and β as
e
∡(α, β) = lim ∡(α(s), p, β(t))
s,t→0

if the limit exists.

Let us introduce some notation. Assuming the curves α and β fixed,

define
e
θ(s, t) = ∡(α(s), p, β(t)).
3.6. Products and Cones 97

In this notation,
∡(α, β) = lim θ(s, t).
s,t→0,s→0

If α and β are shortest paths parameterized by arc length, then d(p, α(s)) =
s, d(p, β(t)) = t, and θ(s, t) is determined by the distance d(α(s), β(t)):
s2 + t2 − d(α(s), β(t))2
θ(s, t) = arccos .
2st
Then the definition of angle (for unit-speed curves) can be rewritten as
follows: the angle ∡(α, β) exists and equals θ0 ∈ [0, π] if and only if

d(α(s), β(t))2 = s2 + t2 − 2st cos θ0 + o(st), s, t → 0.

The notion of angle will be mainly used for shortest paths, but let us first
analyze what we get for different paths in usual Euclidean space. It turns
out that the existence of angle is closely connected with differentiability:

Proposition 3.6.27. Let α : [0, ε) → R2 and β : [0, ε) → R2 be two paths

parameterized by arc length and emanating from the same point a = α(0) =
β(0). Then
1. If both paths are differentiable at t = 0, then the angle ∡(α, β) is
equal to the angle between their velocity vectors.
2. If at least one of the paths α, β is not differentiable at t = 0, then the
angle ∡(α, β) does not exist.

Exercise 3.6.28. Prove the proposition.

There are lots of examples of length spaces where the angle ∡(α, β) does
not exist even for two shortest curves α, β. Perhaps the simplest example
of this type can be made out of two copies of R by gluing them together at
the origin and at all points of the form 2−n with integer n.
Another example is given in the following exercise:

Exercise 3.6.29. Let (V, | · |) be a two-dimensional normed space. Prove

that either (V, | · |) is a Euclidean space or it contains two rays such that the
angle between them is not defined.
Hint: A normed space is Euclidean if and only if

|w + v|2 + |v − w|2 = 2(|v|2 + |w|2 )

for all vectors v, w.

The following natural properties of angles follow immediately from the

definition and are left as easy exercises:
98 3. Constructions

Proposition 3.6.30. 1. Every shortest path forms zero angle with itself.
2. If two shortest segments [a, b] and [b, c] are such that their concatena-
tion (“[abc]”) is also a shortest path, then the angle between [b, a] and [b, c]
is π.
Remark 3.6.31. In this course we will mainly deal with length spaces with
certain curvature bounds. In such spaces the angle between two shortest
paths is always well-defined. For general spaces one may consider upper
angles, which always exist. Namely,
Definition 3.6.32. The upper angle ∡U (α, β) is defined as
e
∡U (α, β) = lim sup ∡(α(s), p, β(t)).
s,t→0
Remark 3.6.33. Historically the notion of angle played a very important
role in the development of metric geometry. It is now well understood that
the usage of angles can be eliminated (without essential complications) from
most arguments dealing with singular (non-Riemannian) spaces. On the
other hand, involving angles often helps to visualize arguments and evokes
Riemannian analogies.
3.6.6. Space of directions. As we mentioned above, the notion of angle
does not play a very important role in modern metric geometry. Neverthe-
less, we will use it to define the cornerstone notion of the space of directions.
In a sense, this notion replaces the concept of tangent space in the theory
of smooth manifolds, and the advanced theory of Alexandrov spaces begins
with this notion.
Let us fix a point p and consider the space of all curves starting from
p. Our goal is to choose a subspace of “nicely behaved curves” and put an
angular semi-metric on this space.
To do this we need a version of the triangle inequality for angles.
Important remark: Please pay very close attention to the proof of
this theorem. This is the first example of an argument based on comparing
certain configurations of points in a length space with in some sense anal-
ogous configurations in the Euclidean plane. This type of arguments is as
basic and widely used in metric geometry as ε-δ arguments in analysis.
Theorem 3.6.34. Consider three curves γ1 , γ2 and γ3 starting at p. Assume
that the angles α1 = ∡(γ2 , γ3 ), α2 = ∡(γ1 , γ3 ) exist. If the angle α3 =
∡(γ1 , γ2 ) also exists, then it satisfies the following triangle inequality:
α3 ≤ α1 + α2 .
Remark 3.6.35. One of the assumptions of this theorem is the existence
of the angles α1 and α2 . In fact, the same inequality holds for upper angles,
which always exist.
3.6. Products and Cones 99

Proof. The statement is trivial if α1 + α2 ≥ π. So suppose that this is not

the case.
Each angle is the limit of an appropriate function θ. Thus, given a
positive ε, for all sufficiently small s, t, r one has
|α1 − θ(b, c)| ≤ ε, |α2 − θ(a, c)| ≤ ε, |α3 − θ(a, b)| ≤ ε,
where a = a(s) = γ1 (s), b = b(t) = γ2 (t) and c = c(r) = γ3 (r).
Here the comparison part begins! Let us develop the two “triangles”
△pac and △pbc on the Euclidean plane. This means that we pick four
points p, a, b and c in the Euclidean plane R2 so that
|pa| = d(p, a), |pb| = d(p, b), |pc| = d(p, c), |ca| = d(c, a), |cb| = d(c, b)
and the points a and b are situated on opposite sides of the line (pc).
Let us fix a, b and move c towards p. Formally this means that we fix s
and t and decrease r. For c very close to p (while a and b are fixed) it is easy
to see that p and c are situated on the same side of the line ab. (Drawing a
picture here is a must!)
On the other hand, fixing r and making s and t sufficiently small we
obtain a configuration with p and c lying on opposite sides of the line ab.
By continuity, we can find such values of the parameters s, t, r that c
belongs to the segment [a, b]. To formalize this, we can fix the choice of
p; for instance choose p = (0, 0) ∈ R2 and choose c in a fixed ray, say
c = (d(p, c), 0). Then the choices of a and b are uniquely determined by
and continuously depend on the distances d(p, a), d(p, b), d(c, a), d(c, b) and
d(p, c). Thus all four points move continuously (well, p does not move at
all) as we vary s, t, r.
For c ∈ [a, b], we obviously have |ac| + |cb| = |ab|, and hence
|ab| = |ac| + |cb| = d(a, c) + d(c, b) ≥ d(a, b).
Add to our planar configuration a point b̃ such that
|pb̃| = |pb| = d(p, b), |ab̃| = d(a, b)
and b̃ lies on the same side of the line (pa) as b.
Recall that θ(a, b) is by definition equal to the angle ∡apb̃ of the triangle
△b̃pa at p, and analogously θ(a, c) = ∡apc, θ(b, c) = ∡bpc. Therefore
θ(a, c) + θ(b, c) = ∡apb.
Comparing the triangles △bpa and △b̃pa, we see that they have two equal
sides and |ab| ≥ |ab̃|. Hence their angles also satisfy the inequality ∡apb ≥
∡apb̃. Thus we obtain that
θ(a, c) + θ(b, c) ≥ θ(a, b).
100 3. Constructions

Combining this with

|α1 − θ(b, c)| ≤ ε, |α2 − θ(a, c)| ≤ ε, |α3 − θ(a, b)| ≤ ε,
we see that α3 ≤ α1 + α2 + 3ε for all positive ε. This completes the proof of
the theorem. ¤
Exercise 3.6.36. Prove the following proposition strengthening the theo-
rem:
Proposition 3.6.37. 1. Theorem 3.6.34 holds for upper angles.
2. Under the assumptions of Theorem 3.6.34, suppose that at least one
of the angles α1 and α2 is equal to zero. Then α3 also exists.
Exercise 3.6.38. Prove that the sum of adjacent angles is at least π.
More precisely, if two shortest segments [a, b] and [b, c] are such that their
concatenation (“[abc]”) is also a shortest path, and the angles ∡dba and
∡dbc exist, then their sum is greater than or equal to π.
Definition 3.6.39. We say that a curve γ (starting at p) has a direction
at p if the angle ∡(γ, γ) does exist. We say that two curves α, β have the
same direction at p if the angle ∡(α, β) exists and is equal to 0.
Exercise 3.6.40. (i) Give an example of a curve γ : [0, ε) → R2 such that
the angle ∡(γ, γ) (at γ(0)) does not exist.
(ii) Prove that ∡(γ, γ), if it exists, is equal to zero.
(iii) Prove that if γ is a geodesic, then the angle ∡(γ, γ) always exists.

Now we consider only curves having direction and observe that the
property of having the same direction is an equivalence relation. A class
of equivalent curves is called a direction. One notices that this space of
directions is a metric space with respect to the upper angle.
Unfortunately, this seemingly very general notion has not proven useful
(as perhaps it is indeed too general). As a glimpse ahead, let us informally
explain what we will use instead of this space. We will mainly study such
spaces where the angle between shortest paths always exists. We will restrict
ourselves to such curves that are equivalent to a shortest curve. In other
words, we care only for directions arising from shortest paths. The space of
such directions will be a metric space with respect to angle (vs. upper-angle
in the general case). Alas, this is not yet the space of directions, as this
space may be noncomplete, and only its completion is called the space of
directions or directional space.
 Chapter 4

Spaces
of Bounded Curvature

4.1. Definitions
4.1.1. Introduction. General length spaces can be extremely nasty, and
there are almost no nontrivial results which do not impose additional re-
strictions on spaces in question. Among such geometric restrictions, which
serve as sort of regularity assumptions, the main role is played by curvature
bounds. Loosely speaking, curvature bounds guarantee a certain degree of
convexity or concavity for distance functions. We begin with the two most
important curvature bounds: spaces of nonpositive curvature and spaces of
nonnegative curvature. These are two classes of spaces which enjoy very dif-
ferent and distinct features, and whose analysis however involves very similar
machinery. While other classes of curvature bounds are also very important,
most of their main features can be traced in the two main classes. Spaces
with a curvature bound (either above or below) will be called Alexandrov
spaces.
At the first glance, spaces of bounded curvature may have neither
topologic nor metric resemblance with Euclidean spaces. Nevertheless, their
definition is based on comparisons with the Euclidean world and (in certain
respects) they are less monstrous than, for instance, normed spaces. For
instance, the notion of angle makes perfect sense in spaces of bounded
curvature, while there are very serious reasons why this notion cannot be
generalized to normed spaces.
We give several equivalent definitions of nonpositively (resp. nonnega-
tively) curved spaces. These definitions formalize that:

101
102 4. Spaces of Bounded Curvature

• Distance functions for such spaces are not less convex (resp. not
less concave) than for the Euclidean plane;
• Geodesics emanating from one point diverge at least as fast as (resp.
not faster) than in the Euclidean plane;
• Triangles are not thicker (resp. not thinner) than Euclidean trian-
gles with the same side lengths.
Very vaguely, these spaces are not smaller (resp. not larger) than the
Euclidean space: for instance, they admit more nonexpanding (resp. ex-
panding) maps.
The sphere is an example of a positively curved space where the above
properties are seen easily. Spherical triangles look “fat” compared to their
Euclidean counterparts (for example, the angles in spherical triangles are
greater than those in Euclidean ones). Spherical geodesics tend to “attract”
one another rather than diverge linearly (more precisely, the distance be-
tween two geodesics emanating from one point in the sphere is a concave
function). These are typical features of spaces having positive curvature.
More generally, every convex surface in R3 (that is, a boundary of a
convex body) is a nonnegatively curved space, and every smooth saddle
surface (locally looking like a hyperbolic paraboloid) is a nonpositively
curved space. For a reader with a background in Riemannian geometry, we
mention in advance that a Riemannian manifold is a nonnegatively (resp.
nonpositively) curved length space if and only if its sectional curvatures are
nonnegative (resp. nonpositive).

Convention. Unless otherwise stated, all length spaces in this chapter

are assumed connected (or, equivalently, with finite metrics), and their
metrics are assumed strictly intrinsic, i.e., such that any two points can
be connected by a shortest path.

With this convention we sacrifice generality for the sake of simplicity

of our exposition; more general cases will be considered later (namely in
Chapters 9 and 10).

4.1.2. Comparisons for distance functions. Fix a point p ∈ X; call

p a reference point. Distance to p is a real-valued function defined on X:
dp (x) = d(x, p). It is easier to deal with functions R → R, and to study dp
restricted to various shortest segments in X. More precisely, for a shortest
path [ab] = γ : [0, T ] → X parameterized by arc length, we introduce a
function g(t) = d(p, γ(t)) = dp ◦ γ(t), which represents the distance function
dp restricted to the segment γ. We will call such functions 1-dimensional
distance functions.
4.1. Definitions 103

We want to compare g with an appropriate 1-dimensional Euclidean

distance function. To do so, we choose a segment of the same length in the
Euclidean plane and a reference point “positioned in the same way as p is
positioned with respect to γ”. Formally, we choose a Euclidean comparison
segment [ab] of the same length T as [a, b] and a reference point p such that
|ap| = dp (a) = d(a, p) and |bp| = dp (b) = d(b, p). Of course, this comparison
configuration is unique up to a rigid motion. Regard this segment as a path
γ0 (t) parameterized by arc length; γ0 (0) = a, γ0 (T ) = b.
Definition 4.1.1. The comparison function for g is g0 (t) = |pγ0 (t)|, the
Euclidean distance from p restricted to a comparison segment [ab].

Following our ideology that distance functions for nonpositively (resp.

nonnegatively) curved spaces must be more convex (resp. more concave)
than for the Euclidean plane, we are going to define such spaces by the
inequality g0 (t) ≥ g(t) (resp. g0 (t) ≤ g(t)). In order to make our definition
local, we do not impose this condition on all pairs (p, γ); we require this to
hold only in a sufficiently small neighborhood of each point.
Definition 4.1.2 (“Distance condition”). We say that (X, d) is nonposi-
tively (resp. nonnegatively) curved if every point in X has a neighborhood
such that, whenever a point p and a segment γ lie within this neighbor-
hood, the comparison function g0 for the 1-dimensional distance function
g(t) = d(p, γ(t)) satisfies g0 (t) ≥ g(t) (resp. g0 (t) ≤ g(t)) for all t ∈ [0, T ].

We will use the name Alexandrov space for all spaces with a curvature
bound, and in particular for spaces of nonpositive and nonnegative curva-
ture.

4.1.3. First examples. Though we choose very simple examples to begin

with, the proofs of their properties are rather long (especially for the second
one). If your intuition tells you that these examples are correct, you can
postpone proofs for a while. The reason is that in the proofs we use the
definition straightforwardly, working by “bare hands”, not having yet any
machinery to shorten them.
Example 4.1.3. Attach together three copies of the ray [0, ∞) ⊂ R by
gluing at the point 0. The resulting space R(3) has nonpositive curvature.
Example 4.1.4. Let K be the cone over a circle of length L (see 3.6.2).
Then K is a space of nonnegative curvature if L ≤ 2π and K is a space of
nonpositive curvature if L ≥ 2π.
Example 4.1.5. Consider R2 with the norm kvk = |x| + |y| where x
and y are Cartesian coordinates of v. This normed space X is neither
nonnegatively curved nor nonpositively curved.
104 4. Spaces of Bounded Curvature

Now we pass to the proofs for these examples.

Proof for Example 4.1.3. Denote by O the common point of the three
rays. Every shortest path in R(3) is either a segment in one of the rays, or a
concatenation of two segments in two different rays. Let γ : [0, T ] → R(3) be
a shortest path and p ∈ R(3) . If two of the points γ(0), γ(T ) and p belong
to the same ray, then the statement is trivial because γ and p are contained
in a union of two rays and this union is isometric to R. So we consider only
the case when the three points p, a = γ(0), and b = γ(T ) belong to different
rays. For every x ∈ [Oa] one has |px| = |pa| − |ax|. For the function g
from Definition 4.1.2, this means that g(t) = g(0) − t if γ(t) ∈ [Oa]. For
the function g0 , one has g0 (t) ≥ g0 (0) − t by the triangle inequality. Since
g0 (0) = g(0), the desired inequality g(t) ≤ g0 (t) follows for γ(t) ∈ [Oa]. The
case γ(t) ∈ [Ob] is similar. ¤

Proof for Example 4.1.4. The cone over a circle is flat outside the vertex;
every sub-cone over a segment of length α ≤ max{L/2, π} is convex and
isometric to a planar sector of angular measure α. (Recall the discussion
of cones in Section 3.6.2.) For a shortest path γ : [0, T ] → K and a point
p ∈ K, consider a triangle △ composed of three shortest paths between the
points p, a = γ(0) and b = γ(T ). There are two possibilities:
• The triangle △ bounds a region not containing O, or one of the
points a, b, p coincides with O.
• The triangle △ bounds a region containing O, or some of its sides
pass through O.
In the first case, the bounded region is isometric to (the region bounded
by) a triangle in the plane, and functions g and g0 from Definition 4.1.2
coincide.
We consider the second case separately for L < 2π and L > 2π. (The
case L = 2π is trivial because the cone is isometric to R2 .)
1. Suppose that L < 2π. Then one can represent the cone K as a
trihedral cone in R3 with edges passing through the vertices of △. (The
three-hedral cone is regarded with its induced length metric; “represent”
means that the two cones are isometric.) To do this, cut K into three
sectors by the rays Oa, Ob and Op, and place these sectors in R3 so as to
form a three-hedral angle (this is possible due to the triangle inequality for
angles).
Now the sides of △ are straight segments in R3 . These three segments
are contained in some plane P ⊂ R3 , and one can use the same triangle,
regarded as a subset of P , to define the function g0 from Definition 4.1.2.
In other words, g0 (t) equals the distance in R3 from p to γ(t). Since the
 4.1. Definitions 105

b b

PSfrag replacements
a
O
b c
p c
Oa a

Figure 4.1: Three comparison triangles do not overlap.

distances in the cone’s intrinsic metric are greater than or equal to those in
the ambient space R3 , it follows that g ≥ g0 . Therefore the cone is a space
of nonnegative curvature.
2. Suppose that L > 2π. The triangles △abO, △apO and △bpO are flat,
i.e., isometric to planar ones. Consider the triangles △abO and △bpO and
place their isometric copies △abO and △bpO in the plane at different sides
of the common side Ob. (If a shortest path [ab] or [bp] passes through O, its
isometric copy degenerates to a segment.) Observe that ∡aOb + ∡bOp > π
and ∡aOp ≤ ∡aOc; hence |ap| ≤ |ap|. Let us rotate the triangle △abO
around b until |ap| becomes equal to |ap| (see the left part of Figure 4.1).
This shows that isometric copies of △abO and △bpO lie without overlapping
in a planar triangle △abp whose sides are equal to those of △abp. This
arguments work for any pair of triangles △aOb, △bOp, △pOa; hence their
isometric copies lie in △abp without overlapping, as shown in the right part
of Figure 4.1.
Then it is clear that all distances between points in the sides of △abp
are less than or equal to distances between corresponding points of the
comparison triangle △abp, i.e., g ≤ g0 . Therefore the cone is a space of
nonpositive curvature. ¤
Remark 4.1.6. From the above proof one can see that the converse state-
ments are also true: if a cone over a circle is nonnegatively (resp. nonpos-
itively) curved, then the length of the circle is not greater (resp. not less)
than 2π.

The following exercise tells one about curvature of a cone over a segment
(note that a cone over a sufficiently short segment is just a planar sector).
We will use the result of this exercise later in this chapter.
Exercise 4.1.7. Let X be a cone over a segment of length L. Prove that
106 4. Spaces of Bounded Curvature

1. X is nonpositively curved for any L.

2. X is nonnegatively curved if and only if L ≤ π.

Proof for Example 4.1.5. First note that straight lines in a normed vec-
tor space are always shortest paths, because the length of a straight segment
equals the distance between its endpoints. Note that there may be other
shortest paths as well.
The unit sphere in X is the Euclidean square with vertices (1, 0), (0, 1),
(−1, 0), (0, −1). Let p = (0, 0). First consider the shortest path connecting
the points (1, 0) and (0, 1), parameterized by the interval [−1, 1]. For this
shortest path, g(t) ≡ 1 while g0 (t) = |t|, so g > g0 except endpoints, contrary
to the definition of nonpositive curvature. Then consider the shortest path
connecting points ( 21 , 12 ) and ( 21 , − 12 ) and parameterized by the interval [0, 1].
Here g(0) = g0 (0) = 1 and g(1) = g0 (1) = 1. However g( 21 ) = 12 while
√
g0 ( 12 ) = 23 > 21 . So in this case g( 21 ) < g0 ( 21 ), contrary to the definition of
nonnegative curvature.
The definitions of nonpositive and nonnegative curvature are local, but
one readily sees that the same construction can be repeated in an arbitrary
neighborhood of the origin since Euclidean homotheties are homotheties
w.r.t. the norm as well. ¤

4.1.4. Distance comparison for triangles. Probably the reader has not
yet gained much insight into the geometry of nonpositively (nonnegatively)
curved spaces from our analysis of distance functions. Let us re-formulate
our definition in more geometric terms. Traditionally the assertion of these
definitions is abbreviated as the CAT(0)-condition, and spaces fitting these
definitions are called CAT(0)-spaces. Here CAT stands for the comparison
of Cartan–Alexandrov–Toponogov, and (0) means that we impose zero as
the curvature bound, reflecting the fact that we compare our space with
the Euclidean plane. Comparing with other spaces (such as spheres) one
defines other CAT(k)-spaces, k ∈ R. This abbreviation is usually used for
upper curvature bounds only, whereas in case of lower curvature bounds one
just speaks of Alexandrov spaces of curvature bounded below by k. Note
that the term “Alexandrov space” can often be confusing since it does not
indicate whether k is the upper bound or the lower bound, and this must
be specified.
By a triangle in X we mean a collection of three points, a, b and c
(vertices) connected by three shortest paths (sides). For brevity we denote
these shortest segments by [ab], [bc], [ca] and their lengths by |ab|, |bc|, |ca|,
respectively. Recall that the vertices alone may not define a triangle uniquely
since there may be several different shortest paths between the same pair of
4.1. Definitions 107

vertices. By ∡abc we denote the angle between the shortest paths [ba] and
[bc] at b (if this angle is well-defined).
For each triangle △abc in X, we construct a triangle △abc in the
Euclidean plane with the same lengths of sides, i.e.
|ab| = |ab|, |bc| = |bc|, |ac| = |ac|.
Definition 4.1.8. Such a triangle △abc is called a comparison triangle for
the triangle △abc.

It is clear that a comparison triangle is uniquely defined up to a rigid mo-

tion of Euclidean plane. Now we can re-formulate the comparison condition
for distance functions as follows.
Definition 4.1.9 (“Triangle condition”). X is a space of nonpositive (resp.
nonnegative) curvature if in some neighborhood of each point the following
holds:
For every △abc and every point d ∈ [ac], one has |db| ≤ |db| (resp.|db| ≥
|db|) where d is the point on the side [ac] of a comparison triangle △abc such
that |ad| = |ad|.

Such a neighborhood is called a normal region. One can always choose

a normal region G so small that all shortest paths with endpoints in G are
still contained in a (possibly larger) normal region.
Exercise 4.1.10. Check that Definition 4.1.9 is indeed a word-by-word
reformulation of Definition 4.1.2 in new notation.
Exercise 4.1.11. Show that it is sufficient to require the triangle condition
only for d being the midpoint of [ab]; i.e., such a weakened definition is still
equivalent to the original one.
Exercise 4.1.12. Show that the following property is implied by but not
equivalent to the triangle condition:
For any triangle △abc, and midpoints d and e of its sides [ab] and [bc]
the inequality 2|de| ≤ |ac| holds.
Hint: Consider a normed vector space.
Example 4.1.13. The direct metric product (see 3.6.1) of nonpositively
(nonnegatively) curved spaces is a nonpositively (resp. nonnegatively)
curved space.
Exercise 4.1.14. Prove this.
Hint: Let △X be a triangle in X = X1 × X2 , △1 and △2 its projections
to X1 and X2 . Denote by △a1 b1 c1 and △a2 b2 c2 comparison triangles for △X
and △Y respectively. Choose these triangles in the (x, y)- and (z, t)-planes
 108 4. Spaces of Bounded Curvature

in Euclidean space R4 with Cartesian coordinates (x, y, z, t). Now observe

that the triangle in R21 × R22 with vertices (a1 , a2 ), (b1 , b2 ), and (c1 , c2 ) is
a comparison triangle for △X . Now a short straightforward computation
finishes the proof.

Roughly speaking, all sufficiently small triangles in a space of nonposi-

tive (resp. nonnegative) curvature are not thicker (resp. not thinner) than
corresponding Euclidean triangles; in a nonnegatively curved space, a tri-
angle may be thought of as “fat” (or “swollen”) as in Figure 4.2, while a
triangle in nonpositive curvature is skinny (with sides “sucked inside”).
PSfrag replacements
B B B
A
B K>0 K=0 K<0

C
K<0
K=0
K > 0A C A C A C

Figure 4.2: Comparison of triangles.

Though the definitions of nonpositively and nonnegatively curved spaces

look similarly, we will see later that their properties are very different.
Please note that although one can speak about “a space having nonpos-
itive (resp. nonnegative) curvature”, we do not define a notion of curvature
alone, and neither do we assign to it a numerical value.

4.1.5. Angle comparison for triangles. Looking at the figures of “fat”

and “skinny” triangles, one naturally suspects that “fat” triangles should
have large angles, while the angles of “skinny” triangles are expected to be
small. Indeed, this observation leads to a definition of nonpositively curved
spaces via comparisons of angles: X is a space of nonpositive curvature if the
angles of every sufficiently small triangle exist and they are not greater than
the corresponding angles of a comparison triangle in the Euclidean space.
In case of a nonnegatively curved space we have to change the words
“not greater” to “not less” and assume, in addition, that the sum of adjacent
angles is equal to π. 1 More formally, the new definition reads as follows.

Definition 4.1.15 (“Angle condition”). A length space X is a space of

nonpositive curvature if every point of X has a neighborhood such that, for

1We do not know if the last condition is really necessary or not.

4.2. Examples 109

every triangle △abc contained in this neighborhood, the angles ∡bac, ∡cba
and ∡abc are well defined and satisfy the inequalities
e
∡bac ≤ ∡bac e
∡abc ≤ ∡abc e
∡bca ≤ ∡bca
e
(recall that ∡abc e
denotes the comparison angle, i.e., ∡abc = ∡abc where
△abc is a comparison triangle, cf. Definition 3.6.25).
A length space X is a space of nonnegative curvature if every point of X
has a neighborhood such that, for every triangle △abc contained in this
neighborhood, the angles ∡bac, ∡cba and ∡abc are correctly defined and
satisfy the inequalities
e
∡bac ≥ ∡bac e
∡abc ≥ ∡abc e
∡bca ≥ ∡bca,
and, in addition, the following holds: for any two shortest path [pq] and [rs]
where r is a inner point of [pq], one has ∡prs + ∡srq = π.

While the triangle condition 4.1.9 is just a reformulation of Definition

4.1.2, proving their equivalence to the angle condition 4.1.15 requires some
work. We prove the equivalence later in Section 4.3.2.

4.2. Examples
With the three definitions in mind, we are ready to give more examples
of Alexandrov spaces. Apparently the examples that historically motivated
the notion of curvature are convex and saddle surfaces. Unfortunately, we
need to develop some machinery to prove that these examples satisfy our
definition; however we suggest the reader keep them in mind.
Here we consider examples of Alexandrov spaces whose justification does
not involve any techniques, namely a few trivial examples and polyhedral
spaces in low dimensions (1 and 2).
Example 4.2.1. Euclidean spaces are obviously Alexandrov spaces (both
nonpositively and nonnegatively curved at the same time).
Example 4.2.2. A convex set in an Alexandrov space is obviously an
Alexandrov space (with the same sign of curvature).
Example 4.2.3. An open set in an Alexandrov space, regarded with the
induced length metric, is an Alexandrov space (with the same sign of
curvature). This is so because our definitions are local, and the induced
length metric of an open set locally coincides with the metric restricted
from the ambient length space.
Example 4.2.4. A “fan” made of several segments glued together at one
end is a space of nonpositive curvature. The proof could be the same as for
Example 4.1.3. But now, using the angle definition, we are able to prove
110 4. Spaces of Bounded Curvature

this in a few words: every triangle in our space either has three zero angles
(in which case the angle condition obviously holds) or is degenerate (that
is, contained in one of its sides). In the latter case its comparison triangle
is degenerate as well.
Since every point in a locally-finite graph has a neighborhood that is
either a segment or a bunch of segments attached to one point, all locally-
finite connected graphs are nonpositively curved spaces.

Warning: Unlike graphs, not all two-dimensional locally-finite polyhedra

are Alexandrov spaces.
Example 4.2.5. The union of the xy-plane and the z-axis in R3 (with the
induced intrinsic metric) is another example of nonpositively curved space.
This example can be obtained (and this is a better “intrinsic” description)
by attaching together a plane and a line by gluing them at one point, that
is, forming a metric bouquet (see Definition 4.2.7 below).

Proof. We will show that every triangle in our space satisfies the angle
condition. Note that every shortest path with endpoints in the plane is
entirely contained in the plane. And a shortest path starting at a point in
the line can leave the line only through the origin O. Now one can see that
the angle condition for a triangle is trivial if all its vertices belong to the
plane or to the line, or if two vertices are in the line and one is in the plane.
In the only nontrivial case when vertices a, b reside in the plane and c in the
line, the triangle △abc consists of the flat triangle △abO and the “tail” Oc.
It is then an elementary exercise to check that the angles of this triangle are
less than the corresponding angles of a comparison triangle. ¤
Example 4.2.6. Consider several copies of the plane R2 and identify their
origins. This gives a nonpositively curved space. For two copies of R2 , this
space topologically looks like the cone x2 + y 2 = z 2 .
Warning: This cone together with its intrinsic metric induced from R3
is not an Alexandrov space!
The proof is essentially the same as in the previous example.

Generalizing the above examples, let us give the following

Definition 4.2.7. Let {Xi } be a collection of length spaces and a point xi
is chosen in every space Xi . The metric bouquet of spaces XiS(with marked
points xi ) is a length space obtained from the disjoint union Xi by gluing
all points xi together.
Exercise 4.2.8. Prove that gluing length spaces into a bouquet does not
change their metrics. In other words, the spaces Xi are projected isometri-
cally onto their images in the bouquet.
4.2. Examples 111

The above three examples are partial cases of the following general
statement.

Proposition 4.2.9. A metric bouquet of nonpositively curved spaces is a

nonpositively curved space.

Exercise 4.2.10. Prove the proposition.

Example 4.2.11. “Notebook” example 2.2.7 from Section 2.2 (several half-
planes glued together along their edges) is a nonpositively curved space.

Proof. Consider a triangle △abc in the “notebook” space X. The only

interesting case is when the vertices a, b, c of the triangle lie in different half-
planes, say, in half-planes A, B and C respectively. Let L be the common
edge of the half-planes. The union of the half-planes A ∪ B is a Euclidean
plane, and the shortest path [ab] is a segment in this plane (this shortest
path cannot visit the interior of C: such a visit makes a path longer because
in this case [ab] intersects L at least twice and an interval of the path in C
is longer that the corresponding segment of L). Shortest paths [bc] and [ac]
have similar properties. Now consider the isometry f of B to C fixing L.
(Imagine that we turn B around L to C till they coincide.) Denote by d the
intersection of [bc] and L. Then the quadrangle (a, b, d, f (c)) has the same
angles at points a, b, f (c) as the triangle △abc. It is easy to check (now it is
elementary Euclidean geometry) that angles of the quadrangle are less than
the angles of a comparison triangle for △abc. ¤

Remark 4.2.12. In Chapter 9 we will prove a general Reshetnyak’s theorem

(Theorem 9.1.21), which covers the above nonpositively curved examples as
well as many other spaces obtained by gluing.

It turns out that upper bounds for curvature are less restrictive and
therefore there are fewer different types of examples of nonnegatively curved
spaces. In particular, all two-dimensional spaces of nonnegative curvature
are topological manifolds, possibly with boundary. We will see later that
all convex surfaces are nonnegatively curved spaces. Here we prove this for
polyhedral surfaces. More generally, we can carry out a complete analysis
of which two-dimensional polyhedral spaces are Alexandrov spaces. It is
advisable to refresh the basics of polyhedral metrics given in 3.2. We need
the following important lemma, whose proof is already contained in the
proof for Example 4.1.4 above.

Lemma 4.2.13. The cone over a circle is a nonpositively curved space iff the
length of the circle is at least 2π. The cone over a circle is a nonnegatively
curved space iff the length of the circle is less than or equal to 2π.
112 4. Spaces of Bounded Curvature

Using this lemma, we can characterize all two-dimensional polyhedral

Alexandrov spaces. Recall that polyhedral length spaces were defined and
considered in 3.2. Each point of a two-dimensional polyhedral space has a
neighborhood isometric to a neighborhood of the vertex in a cone over a
graph. This graph is called the link of the point.

Theorem 4.2.14. 1. A two-dimensional polyhedral space is a space of non-

negative curvature iff it is a topological manifold (possibly with boundary),
and the sum of angles around every vertex is not greater than 2π, and more-
over is not greater than π if the vertex belongs to the boundary. Equivalently,
this means that the link of every vertex is a circle of length at most 2π or a
segment of length at most π.
2. A two-dimensional polyhedral space is a space of nonpositive curvature
iff the link of each vertex does not contain a subspace isometric to a circle
of length less than 2π.

Proof. Since being nonnegatively (resp. nonpositively) curved is a local

property, it suffices to consider only a small neighborhood of each point. If
the neighborhood is small enough, then there is no difference between the
polyhedron and the cone over the link. In other words, a polyhedral space X
is nonnegatively (resp. nonpositively) curved if and only if, for every x ∈ X,
the cone over the link of x is a nonnegatively (resp. nonpositively) curved
space.
1. A cone over a circle of length no greater than π is nonnegatively curved
by Lemma 4.2.13. A cone over a segment of length α ≤ π is isometric to
a planar sector of angular measure α; this sector is nonnegatively curved
as a convex subset of the plane. Thus, if the link of every point is a circle
of length at most 2π or a segment of length at most π, then the space is
nonnegatively curved.
Now assume that the space is nonpositively curved, and let us rule out
all cases except these two types of links.
(a) Suppose that the link of some point is not connected. Then removal
of this point vertex makes X locally disconnected. Considering a triangle
△abc where a belongs in one component and b, c in another one, one can see
that X cannot have nonnegative curvature (compare with Example 4.2.6).
(b) Suppose that more than two faces are adjacent to some edge. Then a
neighborhood of a point in that edge looks like the “notebook” in Example
4.2.11, which is not nonnegatively curved.
(c) Now we know that the link is connected, and that every point in the
link has degree at most 2 (otherwise in X there are more than two faces
adjoint to one edge). Therefore the link is either a segment or a circle. If it
4.2. Examples 113

is a circle, Lemma 4.2.13 implies that its length is not greater than 2π. If
the link is a segment, its length is not greater than π by Exercise 4.1.7.
2. Now we pass to the case of nonpositive curvature. We have to prove
the following:
The cone K over a graph Γ is a space of nonpositive curvature if and
only if the length of each nontrivial loop in Γ is not less than 2π.
First assume that K is nonpositively curved, and let us show that Γ
does not contain nontrivial loops of length less than 2π. Let γ be a shortest
nontrivial loop in Γ. Then, for every two points x, y ∈ γ, one of the two
intervals of γ between x and y is a shortest path in Γ. It follows that γ is a
convex set in Γ; hence the sub-cone over γ is a convex set in K (by Lemma
3.6.15). Therefore this sub-cone is nonpositively curved as long as the cone
is; hence L(γ) ≥ 2π by Lemma 4.2.13. Since γ is a shortest loop, all loops
have length no less than 2π.
Now assume that all nontrivial loops in Γ are not shorter than 2π, and
let us prove that K is nonpositively curved. Consider a triangle △abc in K.
First suppose that the sides of the triangle do not pass through O. Consider
the projection of △abc to Γ. This projection is a triangle △a′ b′ c′ whose
sides are shortest paths in Γ (recall the discussion of shortest path in cones
in Section 3.6.2). Consider two cases.
(a) Suppose that △a′ b′ c′ (as a subset of Γ) does not contain a simple
loop (a simple loop is a set homeomorphic to the circle). Then it is easy
to see that all three sides of △a′ b′ c′ have a common point d ∈ Γ; more
precisely, there are shortest paths [a′ d], [b′ d] and [c′ d] such that the triangle
△a′ b′ c′ , as a subset of Γ, coincides with the “fan” [a′ d] ∪ [b′ d] ∪ [c′ d]. This
fan is a convex set in Γ (because every pair of its points belong to a shortest
path contained in the fan). Hence the original triangle △abc is contained in
the cone over the fan which is a convex set in K consisting of three sectors
glued together along a ray. Then absolutely the same argument as for the
“notebook” Example 4.2.11 finishes the proof.
(b) Now suppose that △a′ b′ c′ does contain a simple loop. Then the
perimeter L = |a′ b′ | + |b′ c′ | + |a′ c′ | is no less than 2π. Consider the cone K1
over the circle S of length L. Fix a length-preserving map g from S onto
the triangle △a′ b′ c′ ; namely, split S into three arcs of lengths |a′ b′ |, |b′ c′ |
and |a′ c′ |, and let g map these arcs onto the respective sides of the triangle.
This map g induces the map g : K1 → K sending a point (x, t) ∈ K1 to
the point (g(x), t) ∈ K. The map g is an arcwise isometry (and hence is
nonexpanding); and the triangle △abc is the image of some triangle △a′′ b′′ c′′
in K1 (again, recall the structure of shortest paths in cones discussed in
Section 3.6.2).
114 4. Spaces of Bounded Curvature

Since the length of S is not less than 2π, K1 is a space of nonpositive

curvature. Since the triangles △abc in K and △a′′ b′′ c′′ in K1 have equal
lengths of sides, they have a common comparison triangle △abc in R2 . Then
one has |ad| ≤ |a′′ d′′ | ≤ |ad| for every point d ∈ [bc] and corresponding points
d′′ ∈ [b′′ c′′ ] and d ∈ [bc]. Here the inequality |ad| ≤ |a′′ d′′ | follows from the
fact that g is a nonexpanding map, and the inequality |a′′ d′′ | ≤ |ad| from the
fact that K1 is nonpositively curved. This finishes the proof in the case (b).
It remains to consider the case when at least one side of △abc passes
through the origin O of the cone. Suppose that the side [ab] passes through
O. Note that in this case the distance in Γ between the projections of a
and b is greater than or equal to π.
If one or both of the sides [bc] and [ac] also pass through O, the proof is
similar to Examples 4.2.5 and 4.2.4 and is left as an exercise. We consider
only “the most general case” when neither [ac] nor [bc] passes through O.
Similarly to the above, the projections of [ac] and [bc] to Γ are shortest paths
[a′ c′ ] and [b′ c′ ] in Γ, and the triangle △abc is contained in the sub-cone over
the union [a′ c′ ] ∪ [b′ c′ ]. Then, similarly to the case (b) above, the distance
condition for the triangle △abc is reduced to that for a triangle in the cone
K1 over a segment of length |a′ c′ | + |b′ c′ |. Now the remaining part of the
theorem follows from the fact that a cone over a segment is nonpositively
curved (Exercise 4.1.7). ¤

4.3. Angles in Alexandrov Spaces and

Equivalence of Definitions
4.3.1. Monotonicity of angles. Let us formulate one more definition of
Alexandrov spaces. Then we will prove that all the definitions are equivalent.
Let α and β be two shortest paths (parameterized by arc length) starting
at the same point p. Such configuration will be also referred to as “a hinge”
e
α, β. As in Section 3.6.5, introduce the notation θ(x, y) = ∡α(x)pβ(y); i.e.,
θ(x, y) is the angle at p in a comparison triangle for △α(x)pβ(y).
Definition 4.3.1 (“Monotonicity condition.”). X is a space of nonnegative
(resp. nonpositive) curvature if it can be covered by neighborhoods such
that, for two any shortest segments α and β contained in this neighborhood
(and starting from the same point p), the corresponding function θ(x, y) is
nonincreasing (resp. nondecreasing) in each variable x and y (with the other
one remaining fixed).

We have an immediate corollary:

Proposition 4.3.2. If X is an Alexandrov space in the sense of Definition
4.3.1, then the angle between any two shortest paths in X is well-defined.
 4.3. Equivalence of definitions 115

4.3.2. Equivalence of the definitions. We need an elementary fact in

Euclidean geometry, the so-called Alexandrov’s lemma (this lemma was first
used in a similar context in A.D.Alexandrov’s book [Al]). Before giving its
formal statement, let us explain its meaning from an engineer’s viewpoint.
Such associations may be very helpful in understanding geometric theorems.
Consider a plane quadrangle made of four rods connected by joints.
Denote two opposite (diagonal) joints by b and d and assume that all angles
of the quadrangle, possibly except at d, are less than π. Regarding the two
sides attached to the joint b as a hinge, open this hinge up to straighten the
hinge formed by the sides attached to d (if this is possible). If the angle at d
were less than π (that is, if the quadrangle was convex), then this procedure
would cause d to move closer to b, and it moves away from b otherwise.
Perhaps, drawing a figure may replace the proof and convince most
readers.
Formally, it could be stated in this way:
Lemma 4.3.3 (Alexandrov’s lemma). Let a, b, c, d be points in the plane
such that a and c are in different half-planes with respect to the line bd.
Consider a triangle △a′ b′ c′ in R2 such that
|ab| = |a′ b′ |, |bc| = |b′ c′ |, |ad| + |dc| = |a′ c′ |
and let d′ be a point in the side [a′ c′ ] such that |ad| = |a′ d′ |.
Then ∡adb + ∡bdc < π if and only if |b′ d′ | < |bd|. In this case, one also
has ∡b′ a′ d′ < ∡bad and ∡b′ c′ d′ < ∡bcd; see Figure 4.3.2,on left.
And ∡adb + ∡bdc > π if and only if |b′ d′ | > |bd|, in which case one has
∡b′ a′ d′ > ∡bad and ∡b′ c′ d′ > ∡bcd; see Figure 4.3.2,on right.
PSfrag replacements
a a’ d’=d c’
a′ a c
b
b′ c d’=d
a’ c’
c
a b’
c′
d′ = d b
∡adb + ∡bdc > π < >
∡adb + ∡bdc < π b b’

Figure 4.3: Distance bd decreases or increases depending on ∡adb + ∡bdc.

Proof. We use only the following fact: if two sides of a planar triangle are
fixed, then the angle between them is a monotone (increasing) function of
116 4. Spaces of Bounded Curvature

the third side. In other words, if |xy| = |x′ y ′ | and |yz| = |y ′ z ′ | for two planar
triangles △xyz and △x′ y ′ z ′ , then ∡xyz > ∡x′ y ′ z ′ if and only if |xz| > |x′ z ′ |,
and vice versa.
Take a point c1 on the ray ad so that d is between a and c1 , and
|dc| = |dc1 |. Suppose that ∡adb + ∡bdc > π; then ∡bdc1 < ∡bdc. Hence
|bc1 | < |bc| = |b′ c′ | from the triangles △bdc and △bdc1 . Now applying the
same observation to the triangles △abc1 and △a′ b′ c′ (for which |ab| = |a′ b′ |
and |ac1 | = |a′ c′ |), we obtain that ∡bac1 < ∡b′ a′ c′ . Therefore |bd| < |b′ d′ |
(from the triangles △bad and △b′ ad′ ).
The case ∡adb+∡bdc < π is similar, up to reversing the inequalities. ¤
Remark 4.3.4. The lemma is true (and the same proof works) if the
triangles are placed in a sphere or a hyperbolic plane (the latter is defined
in Chapter 5) instead of the Euclidean plane. We will use this remark later
when we define more general curvature bounds.
Theorem 4.3.5. All the definitions (distance 4.1.2, triangle 4.1.9, angle
4.1.15, and monotonicity 4.3.1) are equivalent.
Remark 4.3.6. Our definitions are local. Their equivalence means that if
one of the definitions holds in a region U , then others hold in some region
V , possibly smaller than U . Nevertheless we will refer to all such regions as
normal regions.

Proof of Theorem 4.3.5. The proofs for spaces of nonpositive and non-
negative curvature are similar up to reversing the inequalities. To avoid rep-
etition, we prove the equivalence for nonpositively curved spaces and then
indicate the necessary modifications for the case of nonnegative curvature.
(1) The distance and triangle conditions are obviously equivalent.
(2) Assume the triangle condition holds, and let us show that then
the monotonicity condition holds as well. Consider a hinge of two
shortest paths α = [p, a], [p, b] and a point a1 in α starting at p.
Consider comparison triangles △pab and △pa1 b for the triangles
△pab and △pa1 b. Let ã be a point in the side pa1 such that
|pã| = |pa|. Then the triangle condition implies |ãb| ≥ |ab| = |ab|.
This means that ∡a1 pb ≥ ∡apb, which is the monotonicity of
angles.
(3) The monotonicity condition 4.3.1 implies the angle condition 4.1.15.
To prove this, consider a triangle △abc. Let its sides [ba], [bc] be
the shortest paths α, β, α(0) = β(0) = b. Monotonicity of angles
implies that
∡abc ≡ ∡(α, β) = lim θ(t, t) ≤ θ(|ab|, |bc|)
t→0
4.3. Equivalence of definitions 117

where θ is as in 4.3.1. Since θ(|ab|, |bc|) = ∡abc, the angle condition

follows.
(4) It remains to prove that the angle condition 4.1.15 implies the
triangle condition 4.1.9. Consider a triangle △abc and a point d in
its side [ac]. Note that ∡bda + ∡bdc ≥ ∡adc = π by the triangle
inequality for angles. Place comparison triangles △abd and △cbd in
different half-planes with respect to the line bd in R2 . Then by the
angle condition we have ∡adb + ∡cdb ≥ π. Now by Alexandrov’s
lemma 4.3.3 it follows that |bd| = |bd| ≤ |b1 d1 |, where △a1 b1 c1 is a
comparison triangle for △abc and d1 is the point in [a1 c1 ] such that
|a1 d1 | = |ad|. This is the triangle condition for △abc and d, and the
equivalence of the definitions (of nonpositive curvature) follows.
To prove the theorem for nonnegatively curved spaces, just reverse all
inequalities. Besides inequalities originating from the definitions of bounded
curvature, we used the fact that the sum of adjacent angles is not less than
π. This is true in any length space but the opposite inequality does not hold
automatically. However the requirement that the sum of adjacent angles
equals π was included in the angle condition for nonnegative curvature (recall
Definition 4.1.15). To complete the proof of the theorem, we have to show
that this property follows from the monotonicity condition. We formulate
this as a separate lemma (it has numerous applications).
Lemma 4.3.7. If a space X has nonnegative curvature in the sense of
monotonicity definition 4.3.1, then, for any shortest path, the sum of adja-
cent angles is equal to π. In other words, if q0 is an inner point of a shortest
path p0 r0 and q0 s0 is a shortest path, then ∡p0 q0 s0 + ∡s0 q0 r0 = π.

Proof. We have ∡p0 q0 s0 +∡s0 q0 r0 ≥ ∡p0 q0 r0 ≥ π by the triangle inequality

for angles. To prove the opposite inequality, consider arbitrary points p, s,
r in the shortest paths [p0 q0 ], [s0 q0 ], [r0 q0 ], respectively. Place comparison
triangles △pqs and △sqr on different sides of the line qs in R2 . And let
△p1 s1 r1 be a comparison triangle for the triangle △psr. The monotonicity
condition 4.3.1 implies that ∡spq ≥ ∡s1 p1 r1 . Then by Alexandrov’s Lemma
4.3.3 the quadrangle pqrs is convex, i.e., ∡pqs + ∡rqs ≤ π. Passing
to a limit as points p, r, s approach q we obtain the desired inequality
∡p0 q0 s0 + ∡s0 q0 r0 ≤ ∡p0 q0 r0 ≤ π. ¤

With this lemma, the proof of Theorem 4.3.5 is finished. ¤

Exercise 4.3.8. Prove that our conditions of nonpositivity (resp. nonneg-
ativity) of curvature are equivalent to the following: let △pqr be a (suffi-
ciently small) triangle in X and △pqr be its comparison triangle. Then for
points p1 , r1 in the sides [pq], [rq] and points p1 , r1 in the sides [pq], [rq]
 118 4. Spaces of Bounded Curvature

such that |qp1 | = |qp1 |, |qr1 | = |qr1 | the inequality |p1 r1 | ≤ |p1 r1 | (resp.
|p1 r1 | ≥ |p1 r1 |) holds.

4.3.3. Semi-continuity of angles. In the plane, angles of a triangle

depend continuously on its vertices. This is no longer true in general length
spaces. However, angles in Alexandrov spaces enjoy certain semi-continuity
properties.
Suppose that the sequences of shortest segments [ai bi ] and [ai ci ] converge
uniformly to shortest paths [ab] and [ac], respectively. (All shortest segments
are considered as paths parameterized with constant speed.) Note that the
uniform convergence of paths is much stronger than its implication that
ai → a, bi → b and ci → c. Nevertheless, in general it is impossible to
claim anything about relations between the angle ∡bac and the limit of the
sequence ∡bi ai ci even if this limit exists.

PSfrag replacements ci c
a c
aai ci
ai b
b bi b a
bi bi ai
c
ci

Figure 4.4: The limit angle can both collapse or explode.

Example 4.3.9. Let [ab] and [bc] be edges of a cube (we consider the surface
of the cube but not the interior). Consider segments [ai bi ] and [bi ci ] parallel
to [ab] and [bc] resp., at distances 1/i and placed in different faces of the
cube; see Figure 4.4.
The hinges ([ai bi ], [bi ci ]) converge to the hinge ([ab], [bc]) as i → ∞.
However ∡abc = π/2 while ∡ai bi ci = π for all i.
Example 4.3.10. Consider the coordinate plane R2 without the coordinate
quadrant {x > 0, y > 0}. Let a = (1, 0), ai = (1, −1/i), b = (0, 0),
bi = (−1/i, −1/i), c = (0, 1), ci = (−1/i, 1). Then the hinges ([ai bi ], [bi ci ])
converge to the hinge ([ab], [bc]), but ∡abc = π while limi→∞ ∡ai bi ci =
1/2π.

Note that the space in Example 4.3.9 has nonnegative curvature while
the one in Example 4.3.10 has nonpositive curvature. This illustrates the
following property of semi-continuity of angles
4.4. Analysis of Distance Functions 119

Theorem 4.3.11. Let X be a space of nonpositive (resp. nonnegative)

curvature. Suppose that the sequences of shortest paths {[ai bi ]}∞ i=1 and
∞
{[ai ci ]}i=1 converge uniformly to shortest paths [ab] and [ac], respectively.
Then ∡bac ≥ lim supi→∞ ∡bi ai ci (resp. ∡bac ≤ lim inf i→∞ ∡bi ai ci ).

Proof. Let α and αi denote the angles ∡bac and ∡bi ai ci respectively. For
a small positive x, let b′ ∈ [a, b], c′ ∈ [a, c] and b′i ∈ [ai , bi ], c′i ∈ [ai , ci ] be
points at the distance x from a and ai respectively. Denote by θ(x) and
θi (x) the comparison angles ∡b e ′ ac′ and ∡b e ′ ai c′ , respectively. Notice that
i i
′ ′ ′ ′
|ai bi | = |ab|, |ai ci | = |ac| and |bi ci | → |bc| as i → ∞ (for fixed x), and hence
limi→∞ θi (x) = θ(x).
By the definition of the angle, α = limx→0 θ(x) and αi = limx→0 θi (x).
If X is nonpositively curved, then θ and θi are nondecreasing functions by
the monotonicity condition. Therefore θi (x) ≥ αi for all x, whence
θ(x) = lim θi (x) ≥ lim sup αi .
i→∞ i→∞

It follows that α = limx→0 θ(x) ≥ lim supi→∞ αi .

If X is nonnegatively curved, the proof is similar: θi is a nonincreasing
function, hence θi (x) ≤ αi , then θ(x) ≤ lim inf i→∞ αi for all x, and therefore
α = limx→0 θ(x) ≤ lim inf i→∞ αi . ¤

4.4. Analysis of Distance Functions

This section is optional, and nothing in the book relies on it. We already
mentioned that distance functions in Alexandrov spaces curvature are “more
convex” or “more concave” than those in the Euclidean plane (for nonposi-
tive and nonnegative curvature, respectively). Here we put a formal wrap-
ping around this claim. This approach to the definition of Alexandrov spaces
is analytic rather than geometric, but it proves very useful in some cases.
As in Definition 4.1.2, consider a length space (X, d), a point p ∈ X, and
a unit-speed shortest path γ : [a, b] → X, and introduce the function g(t) =
d(γ(t), p) and the corresponding comparison function g0 for the Euclidean
comparison segment. Note that the function g0 is uniquely determined by
the function g; it is an easy trigonometric exercise to find a precise expression
for g0 in terms of A = d(γ(a), p) = g(a) and B = d(γ(b), p) = g(b).
Hence, we defined spaces of bounded curvature by restricting the class
of functions that may arise as restrictions of distance functions to shortest
paths. Let us make this description more explicit.
Notice that even without any curvature restrictions not every continuous
function can arise as a 1-dimensional distance function. First of all, g
obviously must be nonnegative; furthermore, it must be a nonexpanding
120 4. Spaces of Bounded Curvature

function. (Recall that a function g is called nonexpanding if |g(t) − g(s)| ≤

|t − s| for all s, t. For a smooth g, this is equivalent to the statement that
|g ′ (t)| ≤ 1 for all t.)

Exercise 4.4.1. Prove that g is nonexpanding.

Hint: This is a trivial consequence of the triangle inequality.

We need a complete list of all possible functions g0 in one variable

that can be obtained by restriction of the Euclidean distance function to
a segment (in other words, the list of all functions we compare with). If
p ∈ R2 and γ0 is a straight 2
p line in R parameterized with unit speed, then
2 2
g0 (t) = |p − γ0 (t)| = (t + c) + h where c is the parameter such that
γ0 (−c) is the orthogonal projection of p to γ0 , and h is the distance from p to
this projection. Thus the set of 1-dimensional
p Euclidean distance functions
is the set of functions of the form t 7→ (t + c)2 + h2 , where c and h are
arbitrary real constants.
To develop a convenient language for expressing comparison inequalities,
let us give a general definition.

Definition 4.4.2. Let F be a class of continuous functions. A continuous

function g : R → R is called F -convex (resp. F -concave) if for every f ∈ F
such that g(x) = f (x) and g(y) = f (y), one has f (z) ≥ g(z) (resp.
f (z) ≤ g(z)) for all z ∈ [x, y].

Example 4.4.3. For F = {f (t) = at + b, a, b ∈ R} consisting of all linear

functions, we get the usual notions of convex and concave functions.

Example 4.4.4. For Fλ = {f (t) = λt2 + at + b, a, b ∈ R}, we get the

notion of λ-convexity. This notion means that one can touch the graph of
g from below by a translation of the parabola y = λx2 . In other words,
while smooth convex functions are characterized by the inequality g ′′ ≥ 0,
for λ-convex functions this inequality turns into g ′′ ≥ λ.

We are certainly concerned with the class of functions

p
E = {f (t) = (t + c)2 + h2 , c, h ∈ R}.
Indeed, the additional restriction imposed by the distance condition 4.1.2 is
that all 1-dimensional distance functions g are E-convex (resp. concave).

Remark 4.4.5. To be more precise, our requirement is assumed to hold only

locally. Thus, for a 1-dimensional distance function g, we should require that
the restriction of g to an interval is E-convex (or E-concave) if the maximum
value of this restriction is sufficiently small. For the sake of simplicity we
ignore this circumstance and consider E-convex functions R → R.
4.5. The First Variation Formula 121

The next exercise is an analog of the fact that a convex function possesses
a “supporting linear function” at every point.
Exercise 4.4.6. Prove that a nonnegative nonexpanding function g : R → R
is E-convex (resp. E-concave) if and only if for every x ∈ R there exists a
function f ∈ E such that f (x) = g(x) and f ≤ g (resp. f ≥ g) everywhere.
Notice that in particular all nonnegative nonexpanding E-convex func-
tions are convex in the usual sense.
The class E is the set of nonnegative solutions of the differential equation
g ′′ (t)g(t) = 1 − (g ′ (t))2 . Analogously to the case of convex functions, a
smooth nonnegative nonexpanding function g is E-convex (resp. E-concave)
if and only if
1 − (g ′ (t))2 1 − (g ′ (t))2
g ′′ (t) ≥ (resp. g ′′ (t) ≤ ).
g(t) g(t)
Exercise 4.4.7. Prove this.
Convex functions, possibly nonsmooth, enjoy nice properties: they have
right and left derivatives everywhere; they are differentiable except at no
more than countably many points, and their derivatives have positive (resp.
negative) jumps at points of nonsmoothness. The same is true for E-convex
and E-concave functions:
Exercise 4.4.8. Let g : R → R be a nonnegative nonexpanding E-convex
(resp. E-concave) function. Prove that
1. g is continuous.
2. g has right and left derivatives everywhere, and the left derivative is
not greater (resp. not less) than the right one.
3. The set of points where g is not differentiable (i.e., where the right
and left derivatives are not equal) is finite or countable.
4. The derivative of g is continuous on the set where it is defined.
Hint: g(t)2 − t2 is a convex (resp. concave) function.

4.5. The First Variation Formula

The term “first variation formula” comes from differential geometry and
means a rule for differentiating the length of a variable curve. In this
section we prove such a formula for shortest paths in Alexandrov spaces;
furthermore, we restrict ourselves to the case when one endpoint of a shortest
path is fixed and the other one is moving along a geodesic. (For more general
cases, see exercises in the end of the section.) Since the length of a shortest
path equals the distance between its endpoints, this also gives us a rule for
differentiating distance functions in Alexandrov spaces.
122 4. Spaces of Bounded Curvature

Let us first consider the Euclidean case. Let γ : [0, a] → R2 be a smooth

unit-speed curve in R2 and p ∈ R2 a point not belonging to (the image of) γ.
Consider the distance function l(t) = |pγ(t)|. Then
dl
= − cos ∡(p − γ(t), γ ′ (t)),
dt
where γ ′ is, of course, the velocity of γ.
Exercise 4.5.1. Prove this formula.

We are going to show that similar formulas are valid for spaces of
nonpositive and nonnegative curvature.
Later on we use the following notation. Let X be a length space,
γ : [0, T ] → X a unit-speed shortest path, a = γ(0), d = γ(T ), and
p ∈ X \ {a}. For each t ∈ [0, T ], set l(t) = |pγ(t)| and fix a shortest
path σt connecting γ(t) to p.
The following proposition is a general fact; it is valid for any length space
without any curvature restrictions.
Proposition 4.5.2. If there exist the angle α = ∡pad between the shortest
paths γ and [ap] = σ0 , then
l(t) − l(0)
(4.1) lim sup ≤ − cos α.
t→+0 t
Remark 4.5.3. The reader who likes the traditional notation for infini-
tesimal quantities will probably prefer the following form of the inequality
(4.1):
l(t) ≤ l(0) − t cos α + o(t), t → +0.
Remark 4.5.4. Since the left-hand side of (4.1) does not depend on σ0 ,
one can write
l(t) − l(0)
lim sup ≤ inf (− cos α),
t→0 t σ0
or, equivalently,
l(t) − l(0)
lim sup ≤ − cos αmin
t→0 t
where αmin is the infimum of angles between γ and all possible shortest
paths from a to p. (See also Corollary 4.5.7 below.)

We need the following elementary

Lemma 4.5.5. Let △abc be a triangle in R2 , α = ∡bac, t = |ac|. Then
¯ ¯
¯ ¯
¯cos α − |ab| − |bc| ¯ ≤ t .
¯ t ¯ |ab|
4.5. The PSfrag replacements
First Variation Formula 123
a
α
b
b
t x z
c al
ya t
c
z

Figure 4.5: Lemma 4.5.5.

Proof. Denote |ab| = y and |bc| = z (see Figure 4.5). By the cosine rule we
have
t2 + y 2 − z 2 y−z y+z t
cos α = = + .
2ty t 2y 2y
Then
¯ ¯ ¯ ¯
¯ y − z ¯ ¯y − z y + z t y − z ¯
¯cos α − ¯=¯ + − ¯
¯ t ¯ ¯ t 2y 2y t ¯
¯ ¯ ¯ ¯
¯y − z ¯ ¯y + z ¯ t t t t
¯
≤¯ ¯ ¯
·¯ − 1¯¯ + ≤1· + ≤
t ¯ 2y 2y 2y 2y y
¯ ¯ ¯ ¯
¯y − z ¯ ¯y + z ¯ t by the triangle inequality.
because ¯ t ¯ ≤ 1 and ¯ 2y − 1¯ ≤ 2y ¤

Proof of Proposition 4.5.2. Consider two “variable” points: a point b in

the shortest path [ap] = σ0 and a point c = γ(t). The triangle inequality
implies that
|ab| − |bc| = |ap| − (|bp| + |bc|) ≤ l(0) − l(t).
Then apply Lemma 4.5.5 to the comparison triangle for △abc. This yields
e |ab| − |bc| t l(t) − l(0) t
cos ∡abc ≤ + ≤− + .
t |ab| t |ab|
One can let points b and c converge to a so that t/|ab| → 0. Then the
proposition follows by passing to the limit in the last inequality. ¤

The following theorem is the main result of this section. We keep on

using the notation introduced before Proposition 4.5.2.
Theorem 4.5.6 (First variation theorem). Let X be a space of nonpositive
or nonnegative curvature, let γ, σt and l(t) be as above, and assume that a
sequence σti converges to σ0 for some sequence {ti }∞ i=1 , ti → 0 as i → ∞.
Then there exists a limit
l(ti ) − l(0)
(4.2) lim = − cos α,
ti →0 ti
where α is the angle at a between σ0 and γ.
124 4. Spaces of Bounded Curvature

Proof of Theorem 4.5.6. By Proposition 4.5.2 we have

l(ti ) − l(0)
lim sup ≤ − cos α.
i→∞ ti
Hence it suffices to prove that
l(ti ) − l(0)
lim inf ≥ − cos α.
i→∞ ti
Fix an r > 0 so small that |ap| > 5r and the ball B5r (a) is a normal region
for nonpositive or nonnegative curvature (this ball contains all triangles used
in the constructions below). We may assume that γ(ti ) ∈ Br (a) for all i. For
each i, set ci = γ(ti ) and let bi be the point in the shortest path [ci p] = σti
such that |bi ci | = r. We will prove that
(4.3) e i bi ≤ π − α.
lim sup ∡ac
i→∞

First let us show how the theorem follows from (4.3). Applying Lemma
4.5.5 to a comparison triangle for △aci bi yields
2
e i bi + ti .
l(0) = |pa| ≤ |pbi | + |bi a| ≤ |pbi | + |bi ci | − ti cos ∡ac
|bi ci |
Since |pbi | + |bi ci | = l(ti ), it follows that
l(ti ) − l(0) e i bi − ti = cos ∡ac
e i bi − ti .
≥ cos ∡ac
ti |bi ci | r
Then by (4.3),
l(ti ) − l(0) e i bi ) ≥ cos(π − α) = − cos α,
lim inf ≥ lim inf (cos ∡ac
i→∞ ti i→∞
and the theorem follows.
The proof for (4.3) is different for nonpositively and nonnegatively
curved spaces.
1. Let X be a space of nonnegative curvature. Then
e i bi ≤ ∡aci bi = π − ∡bi ci d
∡ac
by the angle condition. Then (4.3) follows because lim inf i→∞ ∡bi ci d ≥ α
by semi-continuity of angles (Theorem 4.3.11).
2. Let X be a space of nonpositive curvature. Denote by b the point in
[ap] = σ0 such that |ab| = r. Then ∡babi ≤ ∡babe i , and ∡bab
e i → 0 as i → ∞
because |bi b| → 0 while |ab| and |abi | stay bounded away from zero. Hence
∡babi → 0 as i → ∞.
By the triangle inequality for angles it follows that ∡ci abi → α as i → ∞.
Then
e i abi ≥ lim inf ∡ci abi = α
lim inf ∡c
i→∞ i→∞
4.5. The First Variation Formula 125

e i abi + ∡ac
by the angle condition. On the other hand, ∡c e i bi → π as i → ∞
e i abi + ∡ac
because ∡c e i bi + ∡ab
e i ci = π and ∡ab
e i ci → 0. Thus
e i bi = π − lim inf ∡c
lim sup ∡ac e i abi ≤ π − α.
i→∞ i→∞

This proves (4.3) for nonpositive curvature. ¤

Theorem 4.5.6 obviously implies the following rule for differentiating the
length: if {σt }t∈[0,T ] is a continuous family of shortest paths connecting p
to points γ(t), then there exists the right derivative dl/dt|t=0 = − cos α.
Moreover, one can differentiate the distance from p to γ(t), even if shortest
paths connecting p to γ(t) are not unique. Namely, the following holds:

Corollary 4.5.7. Let X be a nonpositively or nonnegatively curved complete

locally compact space, γ : [0, T ] a geodesic parameterized by arc length,
p ∈ X, p 6= γ(0). Then the function t 7→ l(t) = |pγ(t)| has the right
derivative and
l(t) − l(0)
lim = − cos αmin
t→+0 t
where αmin is the infimum (in fact, minimum) of angles between γ and
shortest paths connecting γ(0) to p.

Proof. By Proposition 4.5.2, we have

l(t) − l(0)
lim sup ≤ − cos αmin .
t→+0 t
Choose a sequence {ti } converging to 0 such that
l(ti ) − l(0) l(t) − l(0)
−→ lim inf
ti t→+0 t
as i → ∞, and fix shortest paths σti connecting p to γ(ti ). By the Arzela–
Ascoli Theorem 2.5.14, {σti } contains a subsequence converging to some
shortest path σ0 . We may assume that the sequence {σti } itself converges
to σ0 . Then by Theorem 4.5.6,
l(ti ) − l(0)
lim = − cos α
i→∞ ti
where α is the angle between γ and σ0 . Thus
l(t) − l(0)
lim inf = − cos α ≥ − cos αmin
t→+0 t
and the formula for the right derivative follows. Note that the equality α =
αmin also follows, so the minimum of angles is indeed attained (at σ0 ). ¤
126 4. Spaces of Bounded Curvature

Remark 4.5.8. Theorem 4.5.6 and Corollary 4.5.7 imply a restriction on

a shortest path σ0 that can be obtained as a limit of shortest paths {σti },
ti → 0. Namely, if σ0 is such a limit, then the angle α = ∡(γ, σ0 ) equals
αmin from Corollary 4.5.7. Indeed, the limit (l(0) − l(ti ))/ti equals cos α by
the theorem, and the same limit equals cos αmin by the corollary. Therefore
α = αmin .
Exercise 4.5.9. Prove that Theorem 4.5.6 is valid for an arbitrary curve γ
(not necessarily a shortest path) having a direction at the initial point a.
Exercise 4.5.10. Generalize Theorem 4.5.6 to the case when both end-
points of shortest paths {σt } are variable and move along two geodesics γ1
and γ2 with constant speeds v1 and v2 . Namely prove that, if σt denotes a
shortest path connecting γ1 (t) and γ2 (t) and a sequence {σ(ti )} converges
to σ0 , then
l(ti ) − l(0)
lim = −v1 cos α1 − v2 cos α2
i→∞ ti
where l(t) = L(σt ) = |γ1 (t)γ2 (t)| and αj = cos ∡(σ0 , γj ) for j = 1, 2.
Exercise 4.5.11. Generalize Corollary 4.5.7 to distance functions of sets.
Namely prove the following. If γ : [0, T ] → X is a shortest path, A is a closed
set not containing γ(0), and l(t) = dist(γ(t), A), then
l(t) − l(0)
lim = − cos αmin
t→+0 t
where αmin is the minimum of angles between γ and shortest paths of length
l(0) connecting γ(0) to A.
Remark 4.5.12. Theorem 4.5.6 is valid for general Alexandrov spaces of
curvature bounded below or above (defined in Section 4.6), and the same
proof works with minimal modifications.

4.6. Nonzero Curvature Bounds and

Globalization
4.6.1. Nonzero curvature bounds. So far, we considered only spaces of
nonpositive or nonnegative curvature. This was a simplification intended
to help the reader understand the subject. Now we are going to define
Alexandrov spaces of curvature not greater than k and no less than k,
for an arbitrary k ∈ R. These generalizations are not just a pursuit of
generality. The reason is that spaces of curvature, say, ≥ 1 or ≤ −1, enjoy
new important properties for which comparison with zero curvature is not
sufficient.
In fact, it is sufficient to consider only two cases, k = 1 and k = −1, in
addition to the case k = 0, because all other cases may be reduced to these
4.6. Nonzero Curvature Bounds and Globalization 127

three by re-scaling. We still assume that all metrics in question are strictly
intrinsic (more general definitions can be found in Chapters 9 and 10).
Historically, the notion of curvature comes from differential geometry in
the form of Gaussian curvature for two-dimensional surfaces or Riemannian
manifolds, and sectional curvatures in higher dimensions. (At this point,
it does not matter what these terms mean; they are just classical objects
of differential geometry. We will consider Riemannian manifolds and their
curvatures in Chapters 5 and 6. For now, it suffices to mention that Gauss-
ian and sectional curvatures are real-valued functions defined by means of
certain differential expressions.) A Riemannian manifold is a nonpositively
(resp. nonnegatively) curved Alexandrov space if and only if its sectional cur-
vatures are nonpositive (resp. nonnegative) everywhere. These two classes
of Riemannian manifolds play an important role in Riemannian geometry,
and this is one of the motivations for studying Alexandrov spaces.
Similarly, Alexandrov spaces of curvature ≥ k and ≤ k (that we are
about to define) include all Riemannian manifolds whose sectional curva-
tures are ≥ k (resp. ≤ k) everywhere.
To stress the connection between our previous classes (of nonnegative
and nonpositive curvature) and new ones (not defined yet), we mention the
following facts that will be proved later:
1. If X is a length space of nonnegative curvature (or, more generally,
of curvature bounded below), then its spaces of directions are spaces of
curvature ≥ 1.
2. If X is a length space of nonpositive curvature (or, more generally, of
curvature bounded above), then its spaces of directions have curvature ≤ 1.

Definitions. The essence of the generalization is that we compare with

other “model spaces” instead of the Euclidean plane. In fact, these model
spaces are standard two-dimensional spaces of constant Gaussian curvature
√
k. For k > 0, the model space is the Euclidean sphere of radius 1/ k. For
k < 0, the model space is the hyperbolic plane of curvature k. Hyperbolic
planes are defined later in Chapter 5. For now, you may restrict yourself to
the case k ≥ 0; then revisit this section after getting familiar with hyperbolic
planes.
To avoid considering the three cases k < 0, k = 0 and k > 0 separately,
we introduce the notion of k-plane:

Definition 4.6.1. Let k be a real number. The k-plane is one of the

following spaces, depending on the sign of k:

• R2 , if k = 0;
128 4. Spaces of Bounded Curvature

√
• the Euclidean sphere of radius 1/ k (with its intrinsic metric), if
k > 0;
• the hyperbolic plane of curvature k, that is, the
√ standard Loba-
chevsky plane with the metric multiplied by 1/ −k, if k < 0.

Note that the k-plane is bounded (i.e., has finite diameter) if k > 0, and
not bounded if k ≤ 0. Denote the diameter of the k-plane by Rk , i.e.,
( √
π/ k, k > 0,
Rk =
∞, k ≤ 0.

We need the following elementary properties of k-planes: for any a, b, c > 0

such that a + b + c < 2Rk , there exists a triangle in the k-plane with sides
a, b, c; moreover such a triangle is unique up to a rigid motion (that is, an
isometry from the k-plane to itself). Therefore for every sufficiently small
triangle in a length space, there is a unique (up to a rigid motion) comparison
triangle in the k-plane. (The words “sufficiently small” here can be omitted
if k ≤ 0.)
Now one can define spaces of curvature ≥ k and of curvature ≤ k in the
same words as we defined nonnegatively and nonpositively curved spaces,
but with the k-plane instead of the plane.
For instance, here is the new “triangle definition”:

Definition 4.6.2. Let k be a real number. A length space X (with a strictly

intrinsic metric) is a space of curvature ≥ k (resp. ≤ k) if every point x ∈ X
has a neighborhood U such that for any triangle △abc containing in U and
any point d ∈ [ac] the inequality |bd| ≥ |bd| (resp., |bd| ≤ |bd|) holds, where
△abc is a comparison triangle in the k-plane and d ∈ [ac] is the point such
that |ad| = |ad|.

All the other definitions can be re-formulated in the similar way. Except
a few inessential points, all statements and proofs related to the equivalence
of the definitions and simplest properties remain the same as for k = 0. Later
we will see that local properties of spaces of curvature ≤ k and of curvature
≥ k (well, not all local properties but all that we are interested in) do not
depend on k. But there are important global properties (properties “in the
large”) that do depend on k. More precisely, spaces of curvature ≥ k where
k > 0 and spaces of curvature ≤ k where k < 0 definitely have additional
interesting nonlocal properties.

Exercise 4.6.3. Formulate distance, angle and monotonicity definitions of

spaces of curvature ≥ k and ≤ k for any k. Prove equivalence of these
definitions.
4.6. Nonzero Curvature Bounds and Globalization 129

Exercise 4.6.4. If k1 > k2 , then every space of curvature ≥ k1 is a space

of curvature ≥ k2 , and every space of curvature ≤ k2 is a space of curvature
≤ k1 . Prove this.

The following definition introduces Alexandrov spaces with “variable”

curvature bounds.
Definition 4.6.5. A length space X is a space of curvature bounded above
(resp. below) if every point x ∈ X has a neighborhood which is a space of
curvature ≤ k (resp. ≥ k) for some k ∈ R. (A neighborhood is regarded
with its induced length metric; k may depend on x.)

4.6.2. Globalization. There is an important class of Alexandrov spaces

for which triangle comparison works “in the large”, i.e., for all triangles no
matter how big they are.
Definition 4.6.6. We say that a length space X is a space of curvature
≥ k or ≤ k globally, or in the large if the triangle condition from Definition
4.6.2 is satisfied for all triangles △abc in X for which a comparison triangle
in the k-plane is well-defined, i.e., exists and is unique up to a rigid motion.

In other words, having curvature ≤ k or ≥ k in the large means that the

whole space is a normal region.
It may be unclear how to apply this definition to nonconnected length
spaces (in which some distances are infinite). The answer is the same as
with local definitions: a length space X has curvature ≥ k or ≤ k in the
large if and only if every connected component of X is. Below we implicitly
assume that the spaces under consideration are connected.
The condition that a comparison triangle is well-defined requires some
clarification. The matter is that if k > 0, then in √ the k-plane there are
no triangles with perimeter greater than 2Rk = 2π/ k, so no √ comparison
triangle exists if the√perimeter of △abc is greater than 2π/ k. If the
perimeter equals 2π/ k, then there are two cases:
√
(1) All sides are shorter than π/ k. In this case, a comparison triangle
is unique: it is a great circle with three points marked as vertices.
√ √
(2) One of the sides equals π/ k, say |ab| = π/ k. In this case,
there are many different comparison triangles. One can take two
opposite points in the sphere for a and b and connect them by two
arbitrary great half-circles, then place c in one of these half-circles
and consider the other half-circle as the side [ab].
Thus, the condition that a comparison triangle for △abc is well-defined is
equivalent
√ √ of inequalities: max{|ab|, |ac|, |bc|} <
to the following system
π/ k and |ab| + |ac| + |bc| ≤ 2π/ k.
130 4. Spaces of Bounded Curvature

√
In fact, it is not necessary to consider triangles of perimeter 2π/ k:

Exercise 4.6.7. Let k > 0, and X is a length space such that the distance
condition from Definition 4.6.2 (either for curvature ≥ k or for√curvature
≤ k) holds for all triangles whose perimeters are less than 2π/ k. Prove
that X is a space of curvature respectively ≥ k or ≤ k in the large.
case of curvature ≤ k, the distance condition for triangles
Hint: In the √
of perimeter 2π/ k simply reduces to nothing. √In the case of curvature ≥ k,
approximate a given triangle of perimeter 2π/ k by triangles with smaller
perimeters and obtain the desired inequalities by passing to a limit.

Similarly to Definition 4.6.6, one can “globalize” other curvature com-

parison conditions, and the global definitions are equivalent just like local
ones.

Exercise 4.6.8. Formulate the global versions of distance, angle and

monotonicity definitions of spaces of curvature ≥ k and ≤ k for any k,
and prove equivalence of these definitions.

Exercise 4.6.9. If k1 > k2 , then every space of curvature ≤ k2 in the large

is a space of curvature ≤ k1 in the large, and every space of curvature ≥ k1
in the large is a space of curvature ≥ k2 in the large. Prove this.

The reader probably has noticed that the second part of the above
exercise is much harder than the first one. The reason is, of course, that
in the case k1 > 0 the definition of curvature bounded by k2 involves more
triangles than that of curvature bounded by k1 . Namely the latter
√ does not
tell us anything about triangles of perimeter greater than 2π/ k1 while the
former
√ requires considering triangles of perimeters up to the greater value
2π/ k2 (or infinity if k1 < 0).
In fact, this difficulty does not exist: if a space has curvature√ ≥ k in
the large, then all triangles have perimeters no greater than 2π/ k. Well,
this rule has some exceptions among one-dimensional spaces (intervals and
circles). We will give a precise formulation and prove this fact in Chapter 10.

Globalization theorems. There are two theorems saying that, under some
conditions, local curvature bounds imply global ones. These theorems are
very important; actually they are central facts of the whole theory. Though
we prove these theorems in later chapters, it is worthwhile to formulate them
here.
1. Globalization theorem for nonpositive curvature (Theorem 9.2.9):
every complete simply connected space of curvature ≤ k, where k ≤ 0, is a
space of curvature ≤ k in the large.
4.7. Curvature of Cones 131

2. Toponogov’s globalization theorem (Theorem 10.3.1): for any k ∈ R,

every complete space of curvature ≥ k is a space of curvature ≥ k in the
large.
You see that the second theorem is very general, while the first one
includes two additional assumptions: namely, it does not work for positive
k, and it requires that the space is simply connected. It is easy to see that
the topological requirement is necessary; for example, the product S 1 × R
is nonpositively curved (because it is locally isometric to R2 ) but is not
nonpositively curved in the large (consider a triangle with vertices in the
circle S 1 × {0} ⊂ S 1 × R).
We suggest you prove these two theorems as exercises in the relatively
simple case when k = 0 and the space is homeomorphic to R2 . (If you
wish, assume in addition that the space is polyhedral; however this does not
simplify the proof much.)

4.7. Curvature of Cones

Here we add one more example to our list of Alexandrov spaces. Namely,
we carry out a complete analysis of when a cone over a length space has
nonpositive or nonnegative curvature. The cone over a metric space was
defined in Section 3.6.2. Recall that points of the cone over a space X
are pairs (x, t), where x ∈ X and t ∈ [0, +∞), and all the pairs (x, 0) are
identified (the resulting point is called the origin, or the vertex, of the cone).
The metric on the cone is defined by an analog of the Euclidean cosine law;
in particular, the cone over the unit sphere S 2 is Euclidean space.
The sphere is the standard space of constant curvature 1, and Euclidean
space has zero curvature (in our context, it is better to say that it is both
nonpositively and nonnegatively curved). It is natural to expect that, if a
length space X is more curved than the sphere (i.e., has curvature ≥ 1), then
the cone over X is more curved than Euclidean space (i.e., has curvature
≥ 0), and vice versa. In fact, more assumptions about X are required to
make this statement correct. The issue is that the definitions of bounded
curvature are local, but the local structure of the cone near its vertex
depends on the global structure of the base space X.
For instance, recall cones over circles (Example 4.1.4). Outside the
origin, every cone over a circle is locally isometric to R2 , but the geometry
near the origin depends on the total length of the circle. In particular, cones
over long circles are nonpositively (but not nonnegatively) curved, and cones
over short circles are nonnegatively (but not nonpositively) curved.
For general cones, the following theorem holds:
132 4. Spaces of Bounded Curvature

Theorem 4.7.1. Let K be a cone over a (possibly not connected) length

space X and O its origin. Then
1. K \ {O} is a space of curvature ≥ 0 (resp. ≤ 0) if and only if X is a
space of curvature ≥ 1 (resp. ≤ 1).
2. K is a space of curvature ≤ 0 if and only if X is a (possibly not
connected) space of curvature ≤ 1 in the large.
3. K is a space of curvature ≥ 0 if and only if either X consists of
exactly two points, or X is a connected space of curvature ≥ 1 in the large
and no triangle in X has perimeter greater than 2π.
Remark 4.7.2. As we already mentioned, the last condition on the perime-
ters is satisfied automatically for “almost all” spaces of curvature ≥ 1 (more
precisely, for all spaces except a few one-dimensional counter-examples).
This allows one to simplify the last statement of the theorem. We will
discuss this further in Section 10.2.1.
Remark 4.7.3. A local curvature bound for K immediately implies the
corresponding global one. To prove this, observe that any triangle in K can
be moved to a normal neighborhood of O by means of re-scaling.

Proof of the theorem. 1. Consider a triangle △abc in K whose sides do

not pass through O. Its projection to X is a triangle △a′ b′ c′ in X with side
lengths less than π. The sides of △abc are contained in convex flat sectors,
namely in the sub-cones over the sides of △a′ b′ c′ . (See Figure 4.6. The base
of the cone is depicted as being flat for the sake of clarity of the picture, but
the reader should rather think of it as a part of a sphere.)
O

PSfrag replacements
a
b B
c
d
O C
A D
a′
b′ B’
c′
X
d′ A’
C’
D’
X
Figure 4.6: Cone is a space of K ≥ 0 if and only if the space itself has K ≥ 1.
4.7. Curvature of Cones 133

We are going to prove that △abc in K satisfies the triangle condition

for curvature ≥ 0 or ≤ 0 if and only if △a′ b′ c′ in X does so for curvature
≥ 1 or ≤ 1, respectively, provided that the perimeter of △a′ b′ c′ is no greater
than 2π.
′
Consider a comparison triangle △a′ b c′ for △a′ b′ c′ in the standard unit
sphere S 2 ⊂ R3 (centered at the origin O ∈ R3 ). Place points a, b, c in the
′
rays Oa′ , Ob , Oc′ ⊂ R3 respectively so that |Oa| = |Oa|, |Ob| = |Ob| and
|Oc| = |Oc|. The resulting triangle △abc in R3 has the same side lengths as
△abc; in other words, △abc is a comparison triangle for △abc.
Pick a point d ∈ [ac] ⊂ K and let d′ ∈ [a′ c′ ] ⊂ X be the projection of d.
′
Let d and d be the corresponding points in the Euclidean segment [ac] and
′
the spherical segment [a′ d ], respectively. The sub-cone over [a′ c′ ] in K is
isometric to the planar sector in R3 spanned by the spherical segment [a′ c′ ].
An isometry from this sub-cone to this sector sends [ac] isometrically to the
segment [ac]; in particular, it sends d to d. Furthermore, this isometry is
naturally related to the isometry from [a′ c′ ] to the spherical segment [a′ c′ ],
′ ′
and the latter sends d′ to d . It follows that d belongs to the ray Od and
|Od| = |Od|.
Assuming the distances |Ob| and |Od| fixed, the distance |bd| in K is an
increasing function in |b′ d′ | (recall the formula for the distance in a cone). If
′ ′
|b′ d′ | = |b d |, then |bd| = |bd| because |Ob| = |Ob| and |Od| = |Od|. Hence
′ ′
|b′ d′ | ≥ |b d | if and only if |bd| ≥ |bd|, and vice versa. This proves the desired
statement about distance conditions for △abc and △a′ b′ c′ .
2. To finish the proof of the theorem, it remains to connect the property
proved in the first step to the definitions, and consider various marginal
cases.
First, observe that every triangle with side lengths less than π is a
projection of some triangle in K (which can be placed in an arbitrarily
small neighborhood of the origin). Thus, by the result of the first step, if K
has curvature ≥ 0 (resp. ≤ 0), then X has curvature ≥ 1 (resp. ≤ 1) in the
large.
Similarly, the projection of any normal region in K \ {O} (for curvature
≥ 0 or ≤ 0) is a normal region in X (for curvature ≥ 1 or ≤ 1 respectively),
and, conversely, a sub-cone over a (sufficiently small) normal region in X is
a normal region in K \ {O}. (“Sufficiently small” means, for example, of
diameter less than 2π/3.) This proves the first statement of the theorem.
3. Now let us consider “large” triangles in X. Namely, let △abc be
a triangle in K whose sides do not pass through O but whose projection
to X, △a′ b′ c′ , has perimeter L > 2π. Then the sub-cone over △a′ b′ c′ in K
(composed of three convex flat planar sectors) is an image under an arcwise
134 4. Spaces of Bounded Curvature

isometry of the cone K1 over the circle of length L. Moreover, △abc is an

image of a triangle in K1 . Since an arcwise isometry is a nonexpanding map
and K1 has nonpositive curvature, the triangle △abc satisfies the triangle
condition for curvature ≤ 0 (without any conditions on X).
On the other hand, observe that the triangle in K1 which corresponds
to △abc does not satisfy the triangle condition for curvature ≥ 0 because
the origin of K1 is “inside” this triangle. Neither does the triangle △abc.
It follows that, if K has curvature ≥ 0, then X contains no triangles with
sides less than π and perimeter greater than 2π.
Triangles △a′ b′ c′ in X with some of sides greater than or equal to π are
considered similarly. Such a triangle corresponds to a triangle in K some
of whose sides pass through O, and the latter is an image under an arcwise
isometry of a triangle in a (nonpositively but not nonnegatively curved) cone
over a segment of length L > π. (Compare with the last part of the proof
of Theorem 4.2.14 about polyhedral spaces.)
The conclusion is: for a connected X, existence of triangles with perime-
ter greater than 2π does not affect nonpositive curvature of K but prevents
K from being nonnegatively curved. This proves the theorem in the case
when X is connected.
4. Finally, consider the case when X is not connected. Then X is a
disjoint union of its components, and K is a metric bouquet of cones over
the components. By Proposition 4.2.9, K has curvature ≤ 0 if and only if
the cone over each component has curvature ≤ 0. This finishes the proof of
the second statement of the theorem.
And K cannot have curvature ≥ 0 unless X is a two-point space (in
which case K ∼ = R). Indeed, let x and y be two points in different
components of X, and let z ∈ X be an arbitrary third point. We may assume
that x and z belong to different components (otherwise interchange x and
y). Let x′ , y ′ and z ′ be the corresponding points in K; namely, x′ = (x, 1),
y ′ = (y, 1) and z ′ = (z, 1). Then the sides [x′ y ′ ] and [x′ z ′ ] of △x′ y ′ z ′ pass
through O and share a common segment [x′ O]. Hence ∡y ′ x′ z ′ = 0, so
△x′ y ′ z ′ does not satisfy the angle condition.
Gathering all implications proved during this long argument, one can
easily see that the theorem follows. ¤
 Chapter 5

Smooth Length
Structures

This chapter deals with certain length spaces whose definition and analysis
involve smooth structures (calculus of variations). We begin with a basic
discussion of Finsler and Riemannian manifolds, followed by a metric intro-
duction to hyperbolic geometry. Then we consider some other interesting
examples. Here is a brief plan of the chapter along with some guidance
through it.
Section 5.1 describes Riemannian metrics in coordinates and promotes
an important concept that a Riemannian metric locally is almost Euclidean.
For the sake of simplicity of notations we mainly stick to two-dimensional
regions; generalizations to higher-dimensional manifolds are obvious. In
Section 5.2 we exploit an observation that shortest paths in a Riemannian
metric satisfy a particular second-order differential equation. This equation
enables us to draw geodesics with given initial positions and velocity vectors.
Using this, we can build normal coordinate systems. They are analogous
to polar, spherical, and Cartesian coordinates: one family of coordinate lines
is formed by geodesics parameterized by arc length, and the other family
consists of equidistant curves orthogonal to the curves from the first family.
There are two ways of regarding normal coordinates. These two ap-
proaches correspond to two dual viewpoints in mechanics and optics: prop-
agation of rays and wavefronts. Namely, one can first begin with a family
of geodesics (for instances, emanating from one point—thus obtaining nor-
mal coordinates centered at the point). These geodesics can be perceived
as rays, and their orthogonal trajectories, which happen to be equidistant

135
136 5. Smooth Length Structures

curves, correspond to wavefronts. Alternatively, one can start with a smooth

curve (initial wavefront), and draw a family of its equidistants—curves lying
at a fixed distance from the initial wavefront. This family represents prop-
agation of the initial wavefront, and orthogonal trajectories to the family of
wavefronts happen to be geodesics. To generalize the notion of normal co-
ordinates to higher dimensions, one considers a family of geodesics forming
orthogonal trajectories for a family of equidistant hypersurfaces.
There is no essential difference between Finsler and Riemannian met-
rics till here, and results of the first two sections hold (with appropriate
modifications) for Finsler manifolds as well.
There is a remarkable one-parameter family of fully homogeneous two-
dimensional Riemannian metrics formed by the spheres, the Euclidean space
and the hyperbolic spaces. These are two-dimensional metrics of constant
curvature, and their geometry is relatively well understood. Geometric
meaning of the Gaussian curvature transpires via comparison with these
spaces. Section 5.3 gives a metric introduction to two-dimensional hyper-
bolic geometry. Hyperbolic geometry can be regarded from an axiomatic
viewpoint, as well as by means of Riemannian geometry. Its study is, how-
ever, assisted a lot by exploiting an auxiliary Euclidean structure: we use a
model of the hyperbolic plane based on Euclidean geometry.
Section 5.4 introduces a remarkable example of sub-Riemannian length
structures. Whereas Riemannian metrics are obtained by modifying the
Euclidean length functional, sub-Riemannian length structures are defined
by restricting the class of admissible curves. We limit ourselves to presenting
some distinctive features of a model example of sub-Riemannian metrics.
This section is optional, and it assumes some knowledge of Lie bracket and
differential forms.
Two last sections of the chapter deal with volumes in smooth length
spaces.

5.1. Riemannian Length Structures

5.1.1. Tangent vectors and coordinate systems. There is a good tra-
dition that almost every textbook in geometry begins with an introduction
to the theory of smooth manifolds. We will try to minimize this introduction,
reducing it to notation and terminology accompanied by a brief ideological
discussion. All we need is to introduce the space T Ω of vectors tangent to
a planar region Ω. For a reader who finds it difficult to get used to notions
introduced in this section, we advise skipping it for the first time and just
thinking of T Ω as Ω × R2 , with a Cartesian coordinate system (x, y) in Ω,
and two base vectors X = (1, 0), Y = (0, 1) in the second factor R2 . This
5.1. Riemannian Length Structures 137

noninvariant coordinate viewpoint is quite convenient as long as we do not

change coordinates.

Tangent vectors. For a point p ∈ Ω, consider all smooth curves γ : [0, 1] →

Ω starting from p, that is, with γ(0) = p. Two curves are said to be
equivalent if their derivatives (velocities) at 0 are the same, that is, if they
start from p with the same velocity. A tangent vector v at p is an equivalence
class of such curves v = [γ].
Of course, since all equivalent curves start from p with the same velocity
vector, it appears as that there is no essential difference between tangent
vectors and usual vectors. Hence a reader may wonder why we make an
easy notion difficult. The reason for that becomes clear when one wants to
change a coordinate system. Indeed, though we stated that Ω is a planar
region, we do not want to fix such an identification. For instance, if Ω is
a surface in R3 , it can be identified with a region in R2 , but there are lots
of such identifications (coordinate systems), and possibly no preferred one.
Secondly, even if Ω is already a subset of R2 , re-imbedding it in a more
convenient way may still be helpful.
As a matter of fact, by introducing tangent vectors we rather make a
difficult object easy. When studying high school geometry, it takes a serious
effort to get used to the concept that two vectors applied at different points
“are equal” if their magnitudes and directions are. The main message now
is that tangent vectors at different points cannot be equal, nor can they be
added.
Even a consistent concept of “the same direction at different points”
cannot exist in general—because otherwise one could easily produce a non-
vanishing vector field tangent to a sphere (and this is a classical theorem
that it is impossible to “comb the hair on a ball”).
The set of tangent vectors at p will be denoted by Tp Ω; the set of all
tangent vectors will be denoted by T Ω. Since tangent vectors at p can
be identified with usual vectors, one can add them and multiply them by
scalars. These operations are invariant, that is, the result is independent
of the coordinate system. Thus Tp Ω is a (two-dimensional) vector space.
Although tangent vectors at different points formally can be added as well
(if we identify them with Euclidean vectors), we will not do this. The reason
is again that this “addition” does not persist under coordinate changes.
Let ϕ : Ω → Ω′ be a smooth map. For a curve γ starting at p, its
image ϕ ◦ γ starts at ϕ(p); moreover, if two curves emanate from x at
the same velocity, so do their images from ϕ(p). Hence ϕ induces a map
dp ϕ : Tx Ω → Tϕ(p) Ω′ , and more generally a map dϕ : T Ω → T Ω′ . This map
138 5. Smooth Length Structures

is called the derivative of ϕ. One can check that it is a linear map on each
Tp Ω.

Coordinate systems. By a coordinate system we understand a diffeomor-

phism ϕ : Ω′ → Ω from a region Ω′ ⊂ R2 in the coordinate Euclidean plane
to Ω. For two reals x, y (with (x, y) ∈ Ω′ ) we say that the point ϕ(x, y) ∈ Ω
is the point with coordinates (x, y).
When a coordinate system is fixed, we will usually abuse notation and
denote a point ϕ(x, y) by (x, y), thus identifying Ω and Ω′ !
Once we fix a coordinate system in Ω, we get the corresponding coor-
dinate system in T Ω. Indeed, for a point p = (x, y) in Ω, there are two
designated curves starting from p: γx (t) = (x + t, y) and γy (t) = (x, y + t).
These curves are called coordinate lines (the x-line and the y-line) passing
through p, and the corresponding tangent vectors are called the coordinate
vectors. We denote them by X and Y (or by X(p) and Y (p) if we want to
emphasize their dependence on p).
Recall that a vector field is a correspondence (a map) that assigns for
every point a tangent vector at this point. Thus the coordinate vectors form
two coordinate vector fields.
If our coordinate system is given by a map ϕ, the coordinate vectors X
and Y are just the images of the vectors of the standard basis i = (1, 0), j =
(0, 1) ∈ R2 under dϕ. The coordinate vector fields are also often denoted
by ∂/∂x and ∂/∂y. The reason for the latter notation becomes obvious in
the next paragraph. Now for a vector V ∈ Tp Ω we define its coordinates
(vx , vy ) to be such (uniquely defined) reals that V = vx X + vy Y . We will
often write V = (vx , vy ), identifying again an object and its coordinates.
What are tangent vectors good for? The most important usage of a
vector is to differentiate functions: if f : Ω → R is a smooth function and
V = [γ] is a tangent vector at p, we define its derivative V f along V to be
¯
df (γ(t)) ¯¯
Vf = .
dt ¯t=0
Differentiation with respect to the coordinate vector fields X and Y is
nothing but taking the well-familiar partial derivatives with respect to x
∂ ∂
and y, thus explaining the notation X = ∂x , Y = ∂y . Moreover, tangent
vectors can even be defined axiomatically as “derivations”, that are maps
from the space of the smooth functions to reals satisfying certain properties
(such as linearity over scalars, and the product rule).
Another way of using vectors is to “move with the velocity prescribed
by a vector field”. More formally, for a (sufficiently smooth) vector field V ,
one can consider its integral curves. An integral curve is a curve γ(t) such
5.1. Riemannian Length Structures 139

that γ ′ (t) = dγ(t)/dt = V (γ(t)) for all t. The existence and uniqueness of an
integral curve with a fixed initial condition (γ(0) = p) is a principal theorem
in the theory of ordinary differential equations.

Degenerate coordinates. It is convenient to consider more general “de-

generate” coordinate systems. For instance, polar coordinates (r, ρ) are
degenerate at r = 0: the map (r, ρ) −→ (x = r cos ρ, y = r sin ρ) is not
invertible, and its derivative is degenerate at r = 0. By a degenerate coor-
dinate system we understand a smooth surjective map ϕ : Ω′ ⊂ R2 → Ω.
We assume that the degenerations of ϕ are nice, as in polar coordinates:
they take place on a few smooth connected submanifolds, each of which is
mapped to a point, and ϕ is a local diffeomorphism elsewhere. In many
cases, such as polar or normal coordinates, a “degenerate coordinate sys-
tem” is traditionally called a coordinate system; we will follow this abuse
of terminology. One can work with degenerate coordinate systems in al-
most the same way as with usual coordinate systems. The only differences
are that there may be a few points in U that have many pre-images, and
that the coordinate vector fields should be regarded as vector-valued maps
X, Y : Ω′ → T Ω given by
∂
X(x, y) = dϕ(x,y) ∈ Tϕ(x,y) Ω,
∂x
∂
Y (x, y) = dϕ(x,y) ∈ Tϕ(x,y) Ω.
∂y
This means that one may have to assign multiple values of a coordinate
∂
vector field at a degenerate point. For instance, ∂r looks like a porcupine
at the origin of polar coordinates: it consists of unit vectors sticking in all
∂
directions (on the other hand, ∂ρ is zero at the origin). Notice that if two
vector fields V, W are coordinate vector field for a coordinate system, this
in particular means that they are linearly independent at every point. This
restriction is no more in effect for two vector fields arising from a degenerate
coordinate system. Nevertheless, only a special class of pairs of vector fields
can be obtained this way. These are “commuting vector fields”, which will
be discussed later.

5.1.2. Finsler surfaces. Let us recall that Finsler length structures, as

well as their particular case of Riemannian and conformal length structures,
are motivated by the idea that we think of the distance as the time (or
other resources) needed to travel between two points. We start from an
intuitive assumption that the rate per “unit of Euclidean distance” depends
on where we travel and in what direction. This is formalized by a smooth
nonnegative “cost” function λ : T Ω → R, where Ω ∈ R2 is a region. We
think of (p, V ) ∈ T Ω as a point p and a “velocity vector” V at p. We can
140 5. Smooth Length Structures

think of λ(p, V ) = λp (V ) as having two arguments representing location and

velocity (we would love to denote this function by L, as this is nothing but
the Lagrangian in mechanics; alas, L is already used to denote the length
functional). Finslerian length structure generated by λ is given by all piece-
wise smooth curves together with a length functional defined by
Z b
(5.1) L(γ, a, b) = λγ(t) (γ ′ (t))dt.
a

Recall that we impose an important requirement on λ: for each p, the

function λp (V ) = λ(p, V ) is a norm on Tp Ω. Its positive homogeneity
guarantees that L is invariant under reparameterizations of curves, and its
convexity (the triangle inequality) implies the semi-continuity of L.

Exercise 5.1.1. Show that the Finsler length structure defined by (5.1) is
indeed a lower semi-continuous length structure, and therefore it is induced
by its intrinsic metric.

As we are going to do local analysis, we assume that all considerations

take place in a neighborhood that is remote enough from the boundary of
Ω; for instance, we may assume that Ω = R2 . This assumption is made
to stay away from boundary effects, such as the situation when a shortest
curve “bends around a peninsula of the complement of the region”.
As usual, the length structure L gives rise to an intrinsic metric d. It
is easy to show that the length of a curve induced by d coincides with L
for all smooth curves (Exercise 5.1.1). However, whereas the existence of a
shortest path for d connecting any two points follows from Theorem 2.5.23,
its smoothness is less obvious. Moreover, a shortest path for d can fail to be
smooth unless a certain additional assumption of strict convexity is imposed
on λp .

Exercise 5.1.2. To see how a shortest curve can be nonsmooth, consider

a function λp which is given by the following norm (independent of p):
λp (v1 , v2 ) = |v1 | + |v2 |. Describe all shortest paths in the corresponding
metric space. In particular, notice that every broken line consisting of two
segments with end-points (0, 0), (1, 1) and an intermediate vertex (x, y),
where x ≥ 0, y ≥ 0, x + y ≤ 1, is a shortest path.

Such a pathology (nonsmoothness and nonuniqueness of shortest paths,

even in an arbitrary small neighborhood of a point) is due to the fact that
balls of the norm (which are “diamonds”) are not strictly convex. An
interested reader can prove that for strictly convex norms all shortest paths
are smooth, and a shortest path connecting two sufficiently close points is
unique.
5.1. Riemannian Length Structures 141

Hint: Use an analog of Lemma 5.1.13 with an appropriate two-dimensi-

onal normed space instead of Euclidean space.
Now we are going to leave general Finsler metrics and concentrate on
their special case of Riemannian metrics. Whereas the geometry of Finsler
spaces may prove to be even more interesting than Riemannian geometry,
it is now not nearly as developed. This is a rapidly growing field, its basic
methods have not been settled yet, and it is not all clear what kind of
problems will determine the progress of Finsler geometry even in the near
future.

5.1.3. Riemannian metrics.

Metric coefficients. For Riemannian metrics, the functions λp are given

by positive definite quadratic forms:
q
λp (V ) = Qp (V, V ),
where, for each p, Qp is a positive definite quadratic form. We will use the
notation hV, W ip = Qp (V, W ) for the symmetrical bilinear form correspond-
p p
ing to Qp . The quantity |V | = hV, V i = Q(V, V ) is called the length of
a tangent vector V .
Let us see how a Riemannian metric can be described using a coordinate
system (x, y). To fix a quadratic form Qp on Tp Ω, it is enough to know the
values
E(p) = Qp (X(p), X(p)) = hX(p), X(p)i ,
F (p) = Qp (X(p), Y (p)) = Qp (Y (p), X(p)) = hX(p), Y (p)i ,
G(p) = Qp (Y (p), Y (p)) = hY (p), Y (p)i ,
which are the entries of the matrix of the bilinear form h, ip in the basis
formed by the coordinate vectors X(p) and Y (p). Indeed,

(5.2) Q(V, W ) = Q(vx X + vy Y, wx X + wy Y )

= Evx wx + F vx wy + F vy wx + Gvy wy
(to avoid cumbersome notation, we drop the dependence on p in this and
many other formulas). Hence a Riemannian metric on Ω can be given by
its metric coefficients, which are three functions E, F, G : Ω → R. These
functions have to satisfy the inequalities E > 0, G > 0, EG − F 2 > 0 since
Q is a positive definite quadratic form. A reader who is not familiar with
quadratic forms can take the formula (5.2) as a coordinate definition of a
Riemannian metric. Of course, coordinate vectors of a degenerate coordi-
nate system may be linearly dependent; one can still define E = hX, Xi,
F = hX, Y i and G = hY, Y i, but it is possible to claim only that E ≥ 0,
G ≥ 0, EG − F 2 ≥ 0.
142 5. Smooth Length Structures

Exercise 5.1.3. Show that if one introduces new coordinates (u, v) such
that x = x(u, v), y = y(u, v) the rule of converting E to the new coordinates
is given by the following formula:
∂x ∂x ∂y ∂y
Enew = E( )2 + 2F + G( )2 .
∂u ∂u ∂u ∂u
Find analogous formulas for F and G.
Note: Recall that a coordinate system is formally a map ϕ : Ω1 → Ω.
Changing coordinates to (u, v) such that x = x(u, v), y = y(u, v), where
x(u, v) and y(u, v) are two given functions (it is just convenient to denote
them by the same letters as coordinates) formally means that we consider
a map ψ : Ω2 → Ω1 , ψ(u, v) = (x(u, v), y(u, v)), and our new coordinate
system is a map ϕ ◦ ψ : Ω2 → Ω.

Once a coordinate system (x, y) is chosen, the length of a curve γ(t) =

(x(t), y(t)) is given by
Z bp
(5.3) L(γ, a, b) = E(x′ )2 (t) + 2F x′ (t)y ′ (t) + G(y ′ )2 (t) dt
a
where E = E(x(t), y(t)) and similarly for F and G.

5.1.4. Isometries and Riemannian manifolds. Let us give the follow-

ing (obvious) definition:
Definition 5.1.4. Two regions Ω, Ω′ with Riemannian metrics Q, Q′ are
said to be isometric if there exists a diffeomorphism ϕ : Ω → Ω′ such that
Q′ (dϕ(V ), dϕ(V )) = Q(V, V ) for every tangent vector V ∈ T Ω.

Equivalently, two Riemannian metrics are isometric if there exist coor-

dinate systems in Ω and Ω′ such that the metric coefficients E, F , G for the
metrics are the same at points with the same coordinates.
Remark 5.1.5. Isometric Riemannian regions are obviously isometric as
length spaces. The converse is also true: if Riemannian regions Ω and Ω′ are
isometric as length spaces, then they are isometric in the sense of Definition
5.1.4 (that is, there exists a smooth isometry that respects their Riemannian
structures). Moreover, every isometry map from Ω to Ω′ is smooth, and (as
a consequence) can be taken as ϕ in Definition 5.1.4.
We leave this fact as an (not so obvious!) exercise. The easiest proof we
know is based on the results of Section 5.2 (namely, smoothness of shortest
paths and properties of exponential maps).

Smoothness of isometries allows us to give a metric definition of a

Riemannian manifold (for we were so far looking only at regions with
Riemannian metrics):
5.1. Riemannian Length Structures 143

Definition 5.1.6. A Riemannian manifold is a length space such that every

point has a neighborhood isometric to a region with Riemannian metric.
Remark 5.1.7. This definition is not standard. In most textbooks Rie-
mannian manifolds are defined as smooth manifolds equipped with Rie-
mannian structures. The definitions are equivalent: this easily follows from
Remark 5.1.5 (in particular, a length space that is locally isometric to a Rie-
mannian region naturally carries a structure of a smooth manifold). If you
are familiar with smooth manifolds, we recommend you prove this equiva-
lence as an exercise.

If you are not happy with these abstract definitions, you may think of
a Riemannian manifold as a smooth surface in a Euclidean space. With
this simplification, one does not even sacrifice generality: according to the
famous Nash’s Embedding Theorem, every Riemannian manifold is isometric
to a smooth embedded surface in a Euclidean space of some (sufficiently
large) dimension. Two-dimensional examples of this kind are discussed in
the next subsection.

Embedded surfaces. An important example of Riemannian metrics comes

from embedded surfaces. This is one of the main motivating examples for
the theory; it goes back to K. F. Gauss. If r : Ω → R3 is an embedding,
one defines λp (V ) = |dp r(V )|. The corresponding bilinear form is given
by the pull-back of the Euclidean scalar product under the embedding:
hV, W ip = hdp r(V ), dp r(W )iE , where h, iE is the Euclidean scalar product
in R3 . This definition has a very clear geometric meaning: if γ is a smooth
curve in Ω, the length of γ is equal to the Euclidean length LE of its image
under r. Indeed,
Z b
L(γ, a, b) = λγ(t) (γ ′ (t)) dt
a
Z b Z b¯ ¯
= |dγ(t) r(γ ′ (t))| dt = ¯ d r(γ(t))¯ dt = LE (r ◦ γ, a, b).
a a dt
A reader may very well think of Ω as surface in R3 and regard r as a
coordinate system; then dr disappears from the formula.
This computation is valid only for smooth curves. Notice that we have Ω
with its Riemannian metric, and we have its image in R3 with the intrinsic
metric dI induced from the ambient Euclidean distance. We want to show
that r is an isometry between the two metric spaces. In other words, we
introduce a metric on r(Ω) via the class of all piece-wise smooth curves. Of
course, r is an isometry with respect to this metric, so we will also denote
it by d (this would not even be an abuse of notation if we think of Ω as
already a subset of R3 with a coordinate system given by r). We want to
144 5. Smooth Length Structures

see that d = dI . There is an obvious inequality d ≥ dI . The other inequality

is contained in the following exercise:
Exercise 5.1.8. 1. For two (sufficiently close) points p, q ∈ r(Ω), let ρ(p, q)
be the length of the curve γpq = r(Ω) ∩ P , where P is the plane containing
the segment [pq] and the normal to the surface r(Ω) at p (notice that γpq
is actually a smooth curve provided that q is sufficiently close to p). Show
that
ρ(p, q) − |pq| = o(|pq|)
as |pq| → 0, where |pq| is Euclidean distance in R3 .
2. Show that, given a curve of length L in r(Ω) and ε > 0, there is a
piece-wise smooth curve with the same endpoints and whose length is at
most L − ε.
Hint: Use directly the definition of an induced metric to find the induced
length of a nonsmooth curve; in this definition, replace each segment [pi pi+1 ]
of a Euclidean broken line inscribed in the curve by a smooth curve γpq , and
use the first part of the exercise to estimate how this could increase the
length of the broken line.

It is still not clear yet that the shortest paths are smooth. This issue is
discussed in the next section. We finish this section with a brief comment
on the relationship between intrinsic and extrinsic geometries of a surface.
By intrinsic geometry we understand the properties that depend only
on the Riemannian metric. In case of an embedded surface the Riemannian
metric is induced by an embedding r.
It is easy to give examples of different embeddings that induce the same
Riemannian structure.
Exercise 5.1.9. Verify that the embeddings ϕ, ψ : [0, 1] × [0, 1] → R3 given
by
ϕ(u, v) = (u, v, 0),
ψ(u, v) = (sin u, cos u, v)
induce the same Riemannian metric.

A more interesting example is given in the next exercise. It shows that

there is a whole variety of surfaces whose Riemannian metrics are isometric
to those of planar regions:
Exercise 5.1.10. Let γ : R → R3 be a curve parameterized by arc length,
and assume that the curvature of γ does not vanish. Consider an embedding
ϕ : U → R3 from a sufficiently small neighborhood U ⊂ R2 of a point (u0 , v0 )
with v0 6= 0 given by
ϕ(u, v) = γ(u) + v γ̇(u).
5.1. Riemannian Length Structures 145

Show that U with the Riemannian metric induced by this embedding is

isometric to a flat region.
Hint: Consider an analogous embedding for a planar curve γ1 with the
same curvature (as a function of its natural parameter) as that of γ. It is
clear that this embedding sends U to a planar region; show that it induces
the same Riemannian metric as ϕ.

To show that two Riemannian metrics are not isometric, one looks for a
metric property that can tell them apart. Here is an example:

Exercise 5.1.11. Give an elementary argument showing that a region of a

sphere is not isometric to a planar region. More precisely, let r : Ω → R3
be an embedding whose image belongs to a sphere (for instance, r(ϕ, ρ) =
(sin ϕ sin ρ, sin ϕ cos ρ, cos ϕ) defined on a small neighborhood in the
(ϕ, ρ)-plane not intersecting the ϕ-axis). Prove that Ω with the Riemannian
metric induced by r is not isometric to a Euclidean region.
Hint: Perhaps the easiest metric invariant that distinguishes the spher-
ical regions from the Euclidean ones is the length of a circle (as a function
depending on the radius).

This is quite a delicate problem to determine what properties of an

embedding (extrinsic properties) can be determined by the intrinsic geom-
etry of a surface. The most famous statement of this type is a theorem
of K. F. Gauss (which delighted him so much that he called it “Theorema
Egregium”, that is,“Magnificent Theorem”) stating that the product of prin-
cipal curvatures of an embedded surface is an invariant of its intrinsic geom-
etry. This invariant is called the Gaussian curvature, and in our exposition
it will be introduced in terms of the intrinsic geometry. In particular, “The-
orema Egregium” implies that the product of two principal curvatures does
not change under bendings of a surface (a bending of a surface is a continuous
family of embeddings inducing the same lengths of curves in the surface).

5.1.5. Riemannian: infinitesimally Euclidean. This section is devoted

to a technically useful (and ideologically very important) observation that
locally a Riemannian metric is almost Euclidean.
Choosing a coordinate system means choosing an identification between
Ω and a planar region. Hence every coordinate system induces an auxiliary
Euclidean structure in Ω. We are going to show that the coordinates can
be chosen in such a way that the auxiliary Euclidean metric induced by the
choice of the coordinates is close to the Riemannian one.
We will need the following lemma, whose proof is left to the reader (as
a trivial exercise; this is just a linear coordinate change):
146 5. Smooth Length Structures

Lemma 5.1.12. Given a point p, one can choose coordinates such that the
metric coefficients at p are E = G = 1, F = 0. More precisely, one can
choose a coordinate system ϕ : Ω′ ⊂ R2 → Ω, ϕ(x0 , y0 ) = p such that the
metric coefficients at p are E(x0 , y0 ) = G(x0 , y0 ) = 1, F (x0 , y0 ) = 0 in this
coordinate system.

Thus one can choose a coordinate system such that, at one given point
p, the metric coefficients look the same as the Euclidean metric in Cartesian
coordinates. This means that the coordinate vectors form an orthonormal
basis of (Tp Ω, h, ip ). In other words, the derivative d(x0 ,y0 ) ϕ of the coordinate
map is a linear isometry
d(x0 ,y0 ) ϕ : (T(x0 ,y0 ) R2 , Euclidean Scalar Product) → (Tp Ω, h, ip ).

The following lemma shows that if we identify Ω with a Euclidean region

by a coordinate system as in Lemma 5.1.12, then the Riemannian and
Euclidean distance functions are close to each other in a small neighborhood
of p. Intuitively, this means that locally a Riemannian metric is “almost
Euclidean”:
Lemma 5.1.13. Let a Riemannian metric on Ω ⊂ R2 be such that its
metric coefficients with respect to the standard Euclidean coordinates (x, y)
in R2 are E(p) = G(p) =p 1, F (p) = 0. Then, for every vector V at q, its
Riemannian length |V | = hV, V i is close to its Euclidean length |V |E :
|V |
lim = 1.
|pq|E →0 |V |E
Furthermore, the Riemannian distance function is close to the Euclidean
one in a small neighborhood of p, namely,
d(q, r)
lim = 1,
|pq|+|pr|→0 |qr|

where |pq| is the Euclidean distance between p and q.

The lemma follows from the continuity of the Riemannian metric Q,

that is, from the continuity of its metric coefficients E, F , G. In its turn,
the lemma easily implies that the Riemannian angle coincides with that in
the length space induced by the Riemannian metric:
Lemma 5.1.14. Let γ1 , γ2 be two smooth paths in Ω starting from p with
nonzero velocities, that is γ1 (0) = γ2 (0) = p and γ1′ (0) 6= 0, γ2′ (0) 6= 0.
Then the (metric) angle between γ1 and γ2 at p does exist and is equal to
the Riemannian angle
hγ ′ , γ ′ i
arccos p ′ ′1 p2 ′ ′
hγ1 , γ1 i hγ2 , γ2 i
5.1. Riemannian Length Structures 147

between their velocity vectors γ1′ = γ1′ (0) and γ2′ = γ2′ (0) at t = 0.
Exercise 5.1.15. Prove Lemmas 5.1.13 and 5.1.14.
Remark. The arguments we know are absolutely straightforward but some-
what tedious.

There is the following convenient reformulation of Lemma 5.1.13. Let

ϕ : Ω′ ⊂ R2 → Ω be a coordinate system in Ω, and p ∈ Ω. Let us introduce
a scalar product h, iP on R2 such that dp ϕ : (R2 , h, iP ) → (Tp Ω, h, ip ) is an
isometry. Recall that by definition the coordinate vectors X and Y at p are
given by
∂ ∂
X(p) = dp ϕ( ) = dp ϕ(1, 0), Y (p) = dp ϕ( ) = dp ϕ(0, 1).
∂x ∂y
Hence in Euclidean coordinates the scalar product h, iP is given by
h(1, 0), (1, 0)iP = E(p), h(1, 0), (0, 1)iP = F (p), h(0, 1), (0, 1)iP = G(p),
where E,F ,G are the metric coefficients of the Riemannian metric.
Lemma 5.1.16. For every vector V at q, its Riemannian
p p length |V | =
hV, V i is close to its Euclidean length |V |P = hV, V iP with respect to
the Euclidean structure given by h, iP :
|V |
lim = 1.
|pq|P →0 |V |P

Lemma 5.1.16 suggests the following way of computing (or defining)

the Riemannian area in coordinates. Let us look at a coordinate rectangle
[x0 , x0 + ∆x] × [y0 , y0 + ∆y], where p = (x0 , y0 ), and ∆x, ∆y are two
(small) positive numbers. Then Lemma 5.1.16 suggests that a definition
of Riemannian area should be such that the area of this rectangle is close to
its Euclidean area with respect to h, iP . It is a standard exercise in Euclidean
vector computations to show thatqthe area of a parallelogram spanned
by vectors V and W is equal to hV, V i hW, W i − hV, W i2 . Hence the
p
Euclidean h, iP -area of the rectangle is equal to E(p)G(p) − F 2 (p) ∆x∆y.
Thus one defines the Riemannian area of Ω by integration as
Z p
(5.4) Area(Ω) = E((x, y))G(x, y) − F 2 (x, y) dxdy,
Ω′
where Ω′ is the corresponding region in the (x, y)-plane.
Exercise 5.1.17. 1. Prove Lemmas 5.1.16.
2. Show that the Riemannian area defined by (5.4) coincides with the
2-dimensional Hausdorff measure.
3. Generalize the formula (5.4) to higher dimensions.
148 5. Smooth Length Structures

Smoothness of shortest paths. This subsection is optional; we will

discuss the smoothness of shortest paths, which will follow independently
from further results as well. It is not difficult to prove that all naturally
parameterized shortest paths in Ω are differentiable. Unfortunately, we
know only a rather tedious, though a straightforward argument. We will
not present a complete proof here, limiting ourselves to its punchline.
We want to prove that a naturally parameterized shortest path γ is
differentiable at t0 . Denote γ(t0 ) = p.
Let us introduce a coordinate system in Ω, thus introducing an auxiliary
Euclidean structure in Ω. Applying Lemma 5.1.12, we may assume without
loss of generality that E(p) = G(p) = 1, F (p) = 0. There is a routine
argument that shows that the differentiability of γ is equivalent to the fact
that “γ has a certain direction and a certain (unit) speed”, namely, that it
satisfies the following two conditions:
(5.5) lim sup ∡(p, γ(u), γ(v)) = 0,
u,v→t0 ,uv>0

where ∡(A, B, C) is the Euclidean angle at A in the triangle ∆ABC, and

d ¯¯
(5.6) |pγ(t)| = 1,
dt t0
where |pq| is the Euclidean distance.
Lemma 5.1.13 immediately implies Condition (5.6).
To check Condition (5.5), one considers a triangle ∆(p, γ(u), γ(v)), where
u > v > 0. It is a standard exercise in planimetry to check that, if the angle
∡(p, γ(u), γ(v)) ≥ α > 0, there exists a positive constant C = C(α) such
that
|pγ(u)| + |γ(v)γ(u)| − |pγ(v)| ≥ C|pγ(v)|.
Combining this observation with the fact that all ratios
d(p, γ(u)) d(p, γ(v)) d(γ(u), γ(v))
, ,
|pγ(u)| |pγ(v)| |γ(u)γ(u)|
converge to 1 as u, v → 0, one arrives at a contradiction to the fact that
d(p, γ(u)) = d(p, γ(v)) + d(γ(u), γ(v)).
Notice that our further exposition will not rely on this argument; we
will derive the smoothness of the shortest paths as an easy corollary of the
Gauss Lemma 5.2.8.

5.1.6. Important examples. Let us explicitly compute the metric coef-

ficients E, F, G for some “model” metrics in “natural” coordinates.
Let us begin with the Euclidean metric in polar coordinates (which
degenerate at the origin). The coordinates are given by the map (r, ρ) −→
∂
(r cos ρ, r sin ρ), and hence the coordinate vectors are ∂r = (cos ρ, sin ρ),
5.1. Riemannian Length Structures 149

∂
∂ρ = (−r sin ρ, r cos ρ) (where the right-hand side expressions are given in
Cartesian coordinates). Hence
(5.7) E(r, ρ) = 1, F (r, ρ) = 0, G(r, ρ) = r2 .

Notice that, although the Euclidean plane is absolutely “homogeneous”,

that is, it looks “the same” at every point, this is not immediately seen from
(5.7): it is only obvious that the Euclidean plane is rotationally symmetric,
for its metric coefficients do not depend on ρ, and hence a transformation
(r, ρ) −→ (r, ρ + const) is an isometry.
Exercise 5.1.18. Represent a parallel translation in polar coordinates and
verify that it is an isometry directly by checking that it preserves the length
structure given by (5.7).

Now let us consider the sphere of radius R with the (degenerate) coor-
dinates given by the map
(ϕ, ρ) −→ (x = R sin(ϕ/R) cos ρ, y = R sin(ϕ/R) sin ρ, z = R cos(ϕ/R)).
This is almost the usual spherical coordinate system, and the only difference
is that the ϕ-coordinate is rescaled to have the ϕ-lines parameterized by arc
length. A trivial computation yields:
(5.8) E = 1, F = 0, G = R2 sin2 (ϕ/R).

Similarly to the previous case of the Euclidean plane, only the rotational
symmetry (ϕ, ρ) −→ (ϕ, ρ + const) of the metric is obvious directly from
the formulas. Of course, we know that a sphere is a perfectly symmetric
space from Euclidean considerations: there is a rigid motion of R3 that
maps the sphere to itself and sends any given point to any other given
point. It is, however, not at all transparent from the intrinsic viewpoint
when we look at the formula (5.8); if we defined a sphere as a surface whose
Riemannian metric is given by (5.8), producing formulas for all isometries
would be quite a task (compare with the previous exercise; a reader who
likes spherical geometry may enjoy converting to spherical coordinates a
3-dimensional rotation about a line other than the z-axis).
Now we define the hyperbolic plane of curvature k (where k < 0!) as a
plane with Riemannian metric given by the following formulas for its metric
coefficients:
1 √
(5.9) E = 1, F = 0, G = sinh2 ( −k r).
−k
Other definitions of the hyperbolic plane are given in Section 5.3. It will be
shown later that the hyperbolic planes are as homogeneous as the spheres
and the Euclidean plane.
150 5. Smooth Length Structures

Exercise 5.1.19. Using (5.8) and (5.9), compute the length of a circle
of radius r in the sphere of a radius R and in the hyperbolic plane of a
curvature k.

5.2. Exponential Map

The main objective of this section is to show that one can introduce a coordi-
nate system in a neighborhood of every point such that E is identically 1 and
F is identically 0. Such coordinate systems are called normal coordinates.
It is easy to see that one family of coordinate lines in such coordinates has
to be formed by shortest paths (parameterized by arc length). We introduce
geodesics as curves satisfying a certain differential equation, and use them
to build normal coordinates.

5.2.1. Geodesics. Normal coordinates. We define geodesics as the

curves satisfying certain second-order differential equations. We will see
that geodesics are locally shortest paths, and that every shortest path
parameterized by arc length is a geodesic.
Definition 5.2.1. A geodesic is a smooth curve (x(t), y(t)) satisfying the
following differential equations:
(5.10)
µ µ ¶2 µ ¶2 µ ¶¶
d2 x d2 y 1 dx ∂E dx dy ∂E dy ∂F 1 ∂G
E 2 +F 2 =− + + − ,
dt dt 2 dt ∂x dt dt ∂y dt ∂y 2 ∂x
(5.11)
µµ ¶2 µ ¶ µ ¶ ¶
d2 x d2 y dx ∂F 1 ∂E dx dy ∂G 1 dy 2 ∂G
F 2 +G 2 =− − + + .
dt dt dt ∂x 2 ∂y dt dt ∂x 2 dt ∂y

This definition may look strange at first glance. To motivate the choice of
this differential equation, in the next chapter we will use variational methods
to derive it as an equation that is satisfied by every smooth shortest path.
A reader can notice that we actually will not use the result obtained by a
variational argument. Instead, we will give a simple proof based on certain
properties of the equation that the geodesics are the locally shortest curves.
It will also become obvious later that this (coordinate) equation defines the
same geometric object if we change the coordinate system.
The reader familiar with classical mechanics may also note that these
equations describe a free particle whose kinetic energy is given by E ẋ2 +
2F ẋẏ + Gẏ 2 (the quadratic form of the metric).
Exercise 5.2.2. Prove that every geodesic (i.e., every solution of equations
(5.10) and (5.11)) except a constant map is a curve parameterized propor-
tionally to arc-length.
5.2. Exponential Map 151

d d
Hint: Multiply equations (5.10) and (5.11) by dt x and dt y, respectively,
and then sum them up. After that compare the result with the equation
µ µ ¶2 µ ¶2 ¶
d dx dx dy dy
E + 2F +G = 0.
dt dt dt dt dt
Exercise 5.2.3. Show by a direct computation that if a curve is a geodesic
in a coordinate system, then it is a geodesic with respect to every coordinate
system.
Hint: Use Exercise 5.1.3.

Now the main theorem of the theory of Ordinary Differential Equations

tells us that, given an initial point p = (x, y) and a unit vector V = (vx , vy )
at p, there exists a unique geodesic γ(t) that starts from p at t = 0 with
the velocity γ̇(0) = V . (Prove that such a geodesic γ is really an arc-length
parameterized curve.) The geodesic γ is defined for all sufficiently small |t|
(since the system is not linear, one cannot guarantee that such a solution is
extendible to a bigger interval).
Definition 5.2.4. For a tangent vector W ∈ Tp Ω, we define the exponential
map expp (W ) by
expp (W ) = γ(|W |),
p
where γ is the geodesic with γ(0) = p, γ̇(0) = W/|W |, and |W | = hW, W i.

The geometric meaning of expp is very transparent: to arrive at expp (W ),

one travels along the geodesic starting from p in the direction of W for dis-
tance |W |.
Of course, expp (W ) is not well-defined if γ(t) cannot be extended for
t = |W |, but at least it is defined for all sufficiently short vectors W . Hence
we have constructed a map expp : Θp ⊂ Tp Ω → Ω, defined on a neighborhood
Θp of the origin in Tp Ω. Regarding p as the second S variable, we will also
consider the map exp : Θ ⊂ T Ω → Ω, Θ = p∈Ω Θp . This is a smooth
map (by the theorem that guarantees that the solution of the initial value
problem smoothly depends on the initial condition).
Notice that the derivative of expp at 0 ∈ Tp Ω is the identity map by
the construction of expp . Indeed, the image expp (tV ) of a curve tV is a
geodesic whose velocity vector at t = 0 is just V . (In particular, expp
preserves the angles between rays emanating from the origin—a fact that
we will extensively use later!) By the inverse function theorem expp is a
diffeomorphism between a neighborhood Up of the origin 0 ∈ Tp Ω in the
tangent space and its image Vp = expp (Up ) in Ω.

Injectivity radius. The maximum “size” of a neighborhood where expp

remains a diffeomorphism is an important characteristic of the metric near p.
152 5. Smooth Length Structures

Definition 5.2.5. The injectivity radius rp at p is the maximum real

number rp (or ∞) such that expp : B(rp ) → Ω is a diffeomorphism to its
image, where B(r) is the disc of radius r centered at 0 ∈ Tp Ω.
Exercise 5.2.6. Prove that inf{rp : p ∈ K} 6= 0 for any compact K
contained in Ω.
Remark. This is a rather unpleasant exercise; though this is just some
version of a compactness argument, a neat proof is rather involved. Perhaps
the most straightforward way to do the exercise is to look at the proof of the
implicit function theorem, and extract an effective estimate on the size of a
neighborhood where the map is invertible. There are also nice geometrical
proofs (using the sort of geometrical thinking that we are trying to promote
by this book), but they are somewhat tricky.

Normal coordinates. Introduce a polar coordinate system ϕ : R2 → B(rp )

in the ball of the radius rp centered at 0 ∈ Tp Ω. Combining it with expp , we
get a new coordinate system (x, y) → expp ϕ(x, y) in Vp = expp B(rp ). To be
more precise, this is a degenerate coordinate system, for polar coordinates
are “singular at r = 0”, but as usual this singularity does not cause any
trouble. To find the coordinates of a point q ∈ Vp ⊂ Ω, one connects p
and q by a geodesic segment contained in Vp (its existence and uniqueness
follow from the fact that expp : Up → Vp is a diffeomorphism). The length
of this segment is the x-coordinate of q, and the ρ-coordinate of the velocity
vector of this geodesic segment at p is the y-coordinate of q. This is just a
Riemannian analog of polar coordinates. Such coordinates are called normal
coordinates centered at p. Notice that the x-lines γy (t) = (t, y) are geodesics
parameterized by arc length.
More generally, normal coordinates is a coordinate system (x, y) such
that its x-lines are geodesic parameterized by arc length, and the coordinate
vector fields are orthogonal for x = 0. This means that the y-lines are
geodesics orthogonal to the y-axis x = 0. In case of “generalized polar
coordinates” described in the previous paragraph, the line x = 0 degenerates
into one point p; to distinguish such “generalized polar coordinates” we
called them normal coordinates centered at p.
The next section implies that alternatively normal coordinates can be
defined as a coordinate system in which E = 1, F = 0.
Remark. The reader who follows a higher-dimensional version of the theory
can define normal coordinates as a coordinate system such that the length of
the first coordinate vector is 1, and it is orthogonal to the other coordinate
vectors. This will imply that the coordinate lines corresponding to the first
coordinate are geodesics. Intuitively it may be convenient to still think of
two coordinates, the second one being (n − 1)-dimensional; then one sees
5.2. Exponential Map 153

a family of geodesic coordinate lines and a family of equidistant surfaces

orthogonal to them. This approach is promoted in the next chapter (Section
6.4.1).

5.2.2. Gauss Lemma and local minimality of geodesics. There is a

very simple criterion for a coordinate system to be normal:
Lemma 5.2.7. If the metric coefficients with respect to a coordinate system
satisfy E = 1, F = 0, then the coordinate system is normal.

Proof. Equations (5.10) and (5.11) are satisfied by the x-lines (x = t, y =

const). Hence the lemma follows. ¤

We are going to show that the converse is also true, namely, that in
normal coordinates the x-lines and the y-lines are mutually orthogonal (not
only at x = 0). This fact is called the Gauss Lemma:
Lemma 5.2.8. In normal coordinates the metric coefficient F = hX, Y i is
identically zero.

Proof. Consider a coordinate line (x(t), y(t)) = (t, y0 ). It has to satisfy

equations (5.10) and (5.11) (note that we are using the result of Exercise
5.2.3). These equations for coordinate lines take a very simple form since
x′ = 1, x′′ = 0, y ′ = 0, y ′′ = 0, where, as usually, prime mean derivative with
respect to t. Thus we get two equations: 0 = −∂E/∂x and 0 = −∂F/∂x.
The first equation carries no new information. Indeed, the fact that the
x-lines are parameterized by arc length is equivalent to E ≡ 1. The second
equation implies that F does not change along the x-lines. Hence it is
identically zero since it is 0 along the y-axis x = 0. ¤

By this lemma, in normal coordinates a Riemannian scalar product has

only one nontrivial coefficient G(x, y) since E = 1 and F = 0, and the
coordinate lines form an orthogonal web.
We will employ this fact to show that a sufficiently short segment of a
geodesic is indeed a shortest path:
Lemma 5.2.9. Let γ(t) be a geodesic with γ(0) = p, and let b < rp , where
rp is the injectivity radius at p. Then γ is the shortest path between p = γ(0)
and q = γ(b).

Proof. Let us consider a piecewise smooth curve σ(t) = (x(t), y(t)), t ∈

[0, b], connecting p = (0, 0) and q = (x0 , y0 ), where (x, y) are normal
coordinates centered at p. Notice that, by the construction of normal
coordinates centered at p, x0 = b < rp . Let b1 be the first value of t
154 5. Smooth Length Structures

such that x(t) = x0 . Then the segment σ|[0,b ] is contained in our coordinate
1
system, and we can estimate its length by integration:
Z b1 p Z b1 Z b1
′
L(σ) ≥ ′ 2 ′ 2
(x ) + G(y ) dt ≥ |x | dt ≥ x′ dt = x0 .
0 0 0
Hence the length of any path from p to q is at least b. Analyzing the equality
case in our inequalities, the reader can easily verify that it can happen only
if σ = γ. ¤
Exercise 5.2.10. Show that a by-product of the argument above is the fact
that all shortest paths are smooth!

Combining this lemma with Exercise 5.2.6 yields a very useful

Lemma 5.2.11. Every point p ∈ Ω possesses a neighborhood such that every
two points from this neighborhood can be connected by exactly one shortest
path. Every geodesic segment contained in this neighborhood is the shortest
path between its endpoints.
Exercise 5.2.12. Prove the lemma.

5.3. Hyperbolic Plane

Hyperbolic planes are defined in Section 5.1 by their metric coefficients (see
(5.9)). This definition does not look too motivated (apart from the fact
that it is obtained simply by replacing the sine and cosine in the metric
coefficients of a sphere by their hyperbolic versions). Nevertheless, there
are important reasons to study hyperbolic planes. This section contains a
metric introduction to two-dimensional hyperbolic geometry.

5.3.1. Motivations.

Comparison spaces. Many geometric ideas, conjectures and examples

arise from studying very symmetric spaces, which are often better under-
stood than other length spaces. Many results are formulated as “comparison
theorems” stating that certain quantities in a metric space are bounded by
those of a relevant comparison space. For instance, our definitions of non-
positively and nonnegatively curved spaces are based on comparisons with
the Euclidean plane. The three most important types of comparison spaces
are the Euclidean plane, spheres and hyperbolic planes (of course, there are
other important comparison spaces, such as symmetric spaces). Apparently
the reader is well acquainted with the first two types of comparison spaces:
we deal with the Euclidean plane since kindergarten, and some results in
the intrinsic geometry of round spheres in R3 are proven in high school
(via the three-dimensional geometry of the ambient space R3 ). A good
5.3. Hyperbolic Plane 155

knowledge of spherical geometry is a must in astronomy and geography. In

contrast, geometry of the hyperbolic plane H2 often remains obscure and
mysterious, although it by no means is more difficult than spherical geom-
etry. This section gives an introduction to hyperbolic geometry. We will
see that the spheres, the plane and the hyperbolic planes together form a
nice 1-parameter family of spaces. This parameter is called curvature; it is
zero for the Euclidean plane and 1/r2 for a sphere of radius r. Using this
parameter, all geometric formulas can be written in such a way that they
are valid for all spaces in this family. On the other hand, qualitative geo-
metric phenomena may be drastically different for Euclidean, hyperbolic or
spherical geometries.

Meta-mathematical discussion: why hyperbolic spaces. It is quite

natural that Euclidean space is the first choice for a space to compare with:
it may be viewed as a mathematical counterpart of the space where we derive
our everyday experience from. There are several meta-mathematical reasons
why we add spherical and hyperbolic spaces to the list of model spaces.
1. These are the only isotropic 2-dimensional spaces. A space is called
isotropic if it “looks the same at every point and in every direction” (that
is, for every two given unit tangent vectors, there exists an isometry whose
derivative maps one of them to the other). Speculating on a physical
analogy, one says that the fundamental laws of space should be the same
for all locations and directions (we mean the space itself, as opposed to the
influence of objects such as the Earth, etc.).
2. The idea of a straight line, which is a cornerstone notion of Euclidean
geometry, has been motivated by looking at trajectories of a free motion
of a particle, or by light rays. In a slightly more general situation (in the
presence of a potential, or in a curved space as in general relativity) a bunch
of trajectories may feature two main types of local behavior: elliptic and
hyperbolic. Model examples of such behaviors are exhibited by lines in
spheres and hyperbolic planes.
3. In fact, the previous remark reflects a more general phenomenon: here
and there in mathematics (and physics) one sees hyperbolic and elliptic ob-
jects: hyperbolas and ellipses, hyperbolic and elliptic differential equations,
sine and hyperbolic sine, and even the real component and the imaginary
component of a complex number. From this viewpoint, Euclidean geometry
is a borderline (parabolic) case between spherical and hyperbolic geometries.
One can easily visualize a sphere and see how it looks “flatter and flatter”
as its radius grows. But to forget about the family of hyperbolic spaces here
is the same as to forget about complex roots of a quadratic equations. In
fact, one can very well think of hyperbolic planes as spheres of imaginary
radii.
156 5. Smooth Length Structures

4. The last but still important reason is that historically the discovery
and study of hyperbolic geometry had tremendous impact on geometry and
mathematical ideology in general.

Axioms and models. Random historical remarks. We are so used to

the Euclidean plane that we rarely ask ourselves about the origin of the very
fundamental geometric laws. There are two approaches to the definition of
Euclidean geometry. Following Euclid, one says that the Euclidean plane
is an object satisfying a list of axioms; the list suggested by Euclid was
supposed to reflect the most basic properties of the real world given in
everyday experience. There are many elementary textbooks in geometry
that use this approach and systematically build Euclidean geometry out
of a set of axioms (the latter are more-or-less the same Euclid’s axioms
formulated in accordance with the modern requirements of mathematical
rigor). Such an approach is used for teaching geometry at schools in most
countries, and it takes a few grades for school children to build Euclidean
geometry (more or less rigorously and systematically) out of its axioms. An
axiomatic approach usually seems very natural for those who have undergone
such a training.
Another approach is to construct a Euclidean plane (that is, its “model”)
out of other mathematical objects that are believed to be better understood.
For instance, one may define a Euclidean plane as the set of all pairs
(x, y) of reals. The pairs are called points, every set of points satisfying
a linear equation ax + by + c = 0 is pronounced to be a line, and the
distance
p between two points is introduced by the formula |(x1 , y1 )(x2 , y2 )| =
(x1 − x2 )2 + (y1 − y2 )2 .
An axiomatic approach is convenient as it begins with listing the key
properties of the object; but one may doubt why such an object exists. It is
wise to keep both approaches in mind and switch between them depending
on a particular consideration.
When it comes to spherical geometry, we hardly can suggest a textbook
that develops it out of axioms, although a set of axioms for spherical geom-
etry can be obtained from Euclidean axioms by a few minor modifications
(and to find which axioms of Euclidean geometry should be changed to de-
fine a sphere instead of Euclidean plane is a good exercise). We usually
study spheres as subsets in R3 on the basis of Euclidean geometry. We
might axiomatically define the fundamental objects (points, lines, angles,
distances) by imposing certain relations given in axioms. Instead, we tradi-
tionally begin with their models using a Euclidean round sphere: we define
the points in spherical geometry to be the points of R3 that lie at a fixed
distance from the origin; we define the lines as the great circles (central
5.3. Hyperbolic Plane 157

cross-sections of the sphere). The spherical angle between two lines is de-
fined as the Euclidean angle between the Euclidean planes containing these
lines. Finally the “spherical distance” is defined to be the Euclidean angle
at the origin. Note that the “ spherical straight lines” do not look straight
from the viewpoint of the ambient Euclidean space, and as well the spherical
distance is different from the distance in the ambient space. A reader may
say: “Aha, but these are just geodesics, angles and distances in the induced
intrinsic metric!” This is true, and this fact had its impact on the history of
hyperbolic geometry. However, we are just lucky to have such a nice model
for spherical geometry—not all models can be explained in such a natural
way.
After this general digression, let us turn back to hyperbolic geometry.
There is a very short axiomatic definition of the hyperbolic planes:
consider Euclid’s set of axioms and substitute the Fifth Postulate (stating
that “two lines parallel to the same line and having a point in common must
coincide”) by its negation. Certainly, Euclid’s formulations are not rigorous
enough by the modern standards, so one may use their version updated by
Hilbert. A space satisfying this new set of axioms is a hyperbolic plane.
There is, however, a long historical path behind this definition.
The story goes back to Euclid, who demonstrated ingenious intuition by
writing that after a few unsuccessful attempts to derive the fifth postulate
from other axioms, he gave up since he felt it could lead him too far.
Indeed, numerous attempts to prove the Fifth Postulate (based on the other
postulates) led quite far! There were many “proofs”, which usually drew a
“contradiction”, but a contradiction with the everyday intuition rather than
the other axioms. For instance, one can prove that, if the Fifth Postulate
were incorrect, then there is a number such that the area of every triangle
is smaller than this number. Is not it a “contradiction”, for we “know” that
one can build as huge a triangle as she wishes? The only problem is that
this “knowledge” does not follow from Euclid’s axioms.
Perhaps the first mathematician who systematically studied the system
of axioms with the negation of the Fifth Postulate was K. F. Gauss. (The
mathematical world described by this collection of axioms is called “non-
Euclidean geometry”.) Gauss seriously suspected that there may be no
contradiction in this axiomatic system, although many properties of non-
Euclidean geometry are very different from what the common sense based
on physical experience would prompt us to expect. In this respect Gauss
went as far as Lobachevsky and Bolyai, and the reason why the latter two are
known as “the inventors of non-Euclidean geometry” is that Gauss had not
published his work. Surprisingly enough, it was just “too much knowledge”
that kept Gauss from publishing his discovery: he realized that one had to
158 5. Smooth Length Structures

build a model to prove that there was no contradictions in non-Euclidean

geometry (though neither Lobachevsky nor Bolyai had had such models,
which were constructed much later by Beltrami and Poincaré). In his letters,
Gauss did express his personal belief that there is no contradiction in the
axioms of non-Euclidean geometry. He had an ill-fated, though an extremely
wise idea of how to construct a model: he wanted to realize hyperbolic
geometry as the intrinsic geometry in some surface in R3 —the same way
as spherical geometry is realized by Euclidean spheres. Gauss even found
small embedded regions with desired properties (so-called pseudo-spheres),
but he could not realize the whole plane. This forced him to suspect that
this might be an indication that a contradiction was still hidden somewhere.
It was later shown by D. Hilbert that the entire hyperbolic plane cannot be
realized by an embedded surface. We will see that it can be represented by a
Riemannian metric on the plane; the problem is that it cannot be embedded
isometrically into R3 .
Finishing this historical digression, let us mention that if one wants to
develop hyperbolic geometry from its axioms, it is the same long way as we
took in school to arrive at interesting results in Euclidean geometry. Per-
haps having this experience once is quite enough, and one usually studies
hyperbolic geometry by means of its models based on Riemannian or even
Euclidean geometry. Notice, however, that if one kept track of the axioms
used to derive each particular statement in Euclidean geometry, the state-
ments can be divided into two classes: those that can be proved without
involving the Fifth Postulate, and the statements that are essentially based
on it. The former class of statements is called absolute geometry; they are
true in both Euclidean and hyperbolic planes. For instance, the fact that
the altitudes of a triangle intersect in one point is an absolute statement,
whereas the fact that the medians in a triangle meet in one point is purely
Euclidean: it does not hold in hyperbolic or spherical geometries.

5.3.2. Elementary hyperbolic geometry via Poincaré models. Let

us describe the Poincaré model of hyperbolic geometry. We will postpone a
proof of the fact that it is isometric to a hyperbolic plane defined before, and
proceed with a relatively detailed study of the hyperbolic plane given by this
model. Note that by speaking about “models” we only mean that our study
of a Riemannian metric is assisted by an auxiliary Euclidean structure, and
hence Riemannian objects can also be regarded from a Euclidean viewpoint.
Unlike Euclidean geometry with only one Euclidean plane (up to an
isometry), there are many pairwise nonisometric hyperbolic planes (one
for each negative number k, thus forming a one-parameter family). All
hyperbolic planes can be obtained from each other by dilations (multiplying
5.3. Hyperbolic Plane 159

a metric by a constant factor). Let us begin with the hyperbolic plane with
k = −1, which will be referred to as the hyperbolic plane.

Poincaré model in the upper half-plane.

Definition 5.3.1. Consider the open upper half-plane

H2 = {(x, y) ∈ R2 : y > 0}
of the coordinate xy-plane equipped with a Riemannian metric given by the
scalar product
1
(5.12) hV, W i(x,y) = hV, W iE ,
y2
where h, iE is the Euclidean scalar product. Hence the metric coefficients of
the hyperbolic metric are given by
1
E(x, y) = G(x, y) = , F = 0,
y2
and the hyperbolic length structure consists of all piecewise smooth curves
together with a length functional L defined by
Z b ¯ ¯
1 ¯¯ dγ(t) ¯¯
(5.13) L(γ, a, b) = ¯ dt ¯ dt,
a y
d
where y = y(γ(t)) is just the y-coordinate of γ and | dt γ(t)| stands for the
Euclidean magnitude of the velocity vector. This length space is called the
hyperbolic (or Lobachevsky) plane, and it is usually denoted by H or H2 .
As usual, when speaking about a length space, we mean the whole class of
isometric spaces. The particular representation of this space given in this
definition will be referred to as the Poincaré model.

Note that the hyperbolic length structure is a particular case of the

conformal length structures (that is, its quadratic form Q(V, V ) is a scalar
multiple of the Euclidean one).
Let us describe the primary objects of this geometry in terms of the
Euclidean structure of the xy-plane (for elementary properties of inversions,
see Appendix 5.3.6 at the end of this section):
• Hyperbolic lines (geodesics) are Euclidean semi-circles orthogonal
to and vertical rays starting from the x-axis.
• Hyperbolic rigid motions (isometries) are translations parallel to
the x-axis, symmetries w.r.t. vertical lines, dilations (homotheties)
and inversions with centers in the x-axis, and their (finite) com-
positions. If we regard H as the set of complex numbers z with
Im(z) > 0, the orientation preserving hyperbolic rigid motions are
160 5. Smooth Length Structures

complex transformations of the form

az + b
T (z) = ,
cz + d
where a, b, c, d are arbitrary reals with ad − bc = 1. To get all
hyperbolic motions one should add transformations of the form
az̄ + b
T (z) = ,
cz̄ + d
with ad − bc = −1.
• The distance between two points p = (0, y1 ) and q = (0, y2 ) on the
y-axis is
d(p, q) = | ln y1 − ln y2 |.
It will be shown below that, given two points a and b, one can
choose a rigid motion that maps a and b to two points p and q in
the y-axis. Then we can use the y-axis as a “ruler” since d(a, b) is
supposed to be equal to d(p, q).
• The hyperbolic angle between two hyperbolic lines is the usual
(Euclidean) angle between (the tangents to) Euclidean semi-circles
(or a semi-circle and a ray) representing the lines in the Poincaré
model.
• The hyperbolic area of a region Ω is defined by integration as
Z
1
h2 (Ω) = 2
dxdy.
Ω y
Certainly, a reader understands that these claims require proofs.
Warning. Dealing with the hyperbolic plane in the Poincaré model,
it is often very tempting to appeal to Euclidean notions. This may be
dangerous: a definition of a hyperbolic notion should be invariant under the
hyperbolic isometries. For instance, Euclidean angle does persist under the
hyperbolic isometries, and hence our definition of hyperbolic angle is at least
not meaningless. On the other hand, Euclidean area in the Poincaré model
does not have any meaning in hyperbolic geometry, for it gets distorted by
hyperbolic rigid motions.
Terminological remark: To emphasize that a notion is regarded with
respect to the structure of the hyperbolic plane (as opposed to the Euclidean
structure of the Poincaré model), we will add an adjective “hyperbolic”. For
instance, we may speak about hyperbolic lines (which may be Euclidean
circles), hyperbolic distances, etc.

It may seem that the hyperbolic plane defined via the Poincaré model is
rather nonhomogeneous: it even has two types of lines: Euclidean semi-
circles and rays. When we approach the x-axis of the Poincaré model,
5.3. Hyperbolic Plane 161

hyperbolic distances are huge compared to Euclidean ones; and they become
small as we move towards y → ∞. However, intrinsically the hyperbolic
plane is quite homogeneous; this becomes obvious by looking at hyperbolic
rigid motions. Indeed, given two points, it is easy to find a hyperbolic rigid
motion that maps one of the points to the other. Notice that one can do
this even without using inversions! They become handy when we want to
map a given ray to another given ray (to prove that hyperbolic geometry is
isotropic).

Poincaré model in the disk. There is a convenient modification of the

Poincaré model, which represents the hyperbolic plane by the interior of
a disc. Apparently, the wisest strategy is to keep all models in mind and
switch between them depending on the needs of a particular argument.
Passing to the model in the disc is nothing but choosing a new coordinate
system in the upper half-plane.
√ Let us apply an inversion ϕ with respect
to the circle of a radius 2 centered at p = (0, −1). We could choose any
inversion whose center does not belong to H as well. The reason behind
this particular choice is that the image of the upper half-plane is exactly the
(open) disc D = {(x, y) : x2 + y 2 = 1}. This model of the hyperbolic plane
will be referred to as the disc model.
Let us see how hyperbolic objects are represented in the disc model.
• The length of a curve γ is given by
Z b
2 |γ ′ (t)|
(5.14) L(γ, a, b) = 2
dt,
a 1 − |γ(t)|
where |γ(t)| denotes the Euclidean distance from γ(t) to the origin
(the center of the disc). Hence the metric coefficients of the
hyperbolic metric in the disc model in the Cartesian xy-coordinates
are
4
E(x, y) = G(x, y) = , F (x, y) = 0.
(1 − x − y 2 )2
2

• Hyperbolic lines (geodesics) are represented by the diameters of the

disc and the Euclidean circular arcs orthogonal to the boundary
circle of the disc.
• Hyperbolic rigid motions are the rotations about the center of
the disc, the symmetries with respect to lines passing through the
center of the disc, the inversions mapping the boundary circle of
the disc to itself, and the products of these.
• The hyperbolic angle between two hyperbolic lines is the usual
(Euclidean) angle between (the tangents to) Euclidean circular arcs
(or segments) representing the lines in the disc model.
162 5. Smooth Length Structures

• The hyperbolic area of a region Ω is defined by integration as

Z
4
h2 (Ω) = 2 2
dxdy.
Ω 1−x −y
Exercise 5.3.2. Describe which of the arcs and diameters representing
hyperbolic lines in the disc model correspond to the vertical rays in the
Poincaré model.
Exercise 5.3.3. Verify that these descriptions are correct.
Hint: To convert a hyperbolic rigid motion to the disc model, we have
to conjugate it by ϕ. Namely, let I be a hyperbolic rigid motion given as a
map in the Poincaré model. To see what it does to a point a ∈ D, we first
map a back to the Poincaré model by ϕ−1 = ϕ, then apply I, and then map
it back to D by ϕ. The formula reads ϕ ◦ I ◦ ϕ.
Exercise 5.3.4. Find a complex representation for hyperbolic rigid motions
in the disc model.
Hint: Use the hint to the previous exercise together with a complex
representation of the inversion I.
Exercise 5.3.5. Check that the metric coefficients of the hyperbolic metric
in the disc model with respect to the standard polar coordinate system (r, ρ)
in the disc are given by the following expressions:
4 4r2
(5.15) E= , F = 0, G= .
(1 − r2 )2 (1 − r2 )2
Let us introduce a new coordinate system (d, ρ), where the second
coordinate of a point is the same as in polar coordinates, and the first
coordinate is equal to the hyperbolic distance from the point to the center of
the disc. This is just a direct analog of polar coordinates for the hyperbolic
plane. Recall that the r-lines (r = t, ρ) of the polar coordinate system are
also hyperbolic lines, and hence all we have to do is to re-parameterize them
by hyperbolic arc length. The hyperbolic distance from the origin to a point
(r, ρ) is Z r
1+r
d(r) = 2(1 − r2 )−1 dr = ln .
0 1−r
Hence the map
ex − 1
(x, y) → (r = x , ρ = y)
e +1
converts the hyperbolic coordinates (d, ρ) to the polar coordinates (r, ρ).
Lemma 5.3.6. The metric coefficients of the hyperbolic metric in the disc
model with respect to (d, ρ)-coordinates are
E = 1, F = 0, G = sinh2 d.
Hence (by Lemma 5.2.7) (d, ρ) is a normal coordinate system.
5.3. Hyperbolic Plane 163

Proof. We have E = 1 (since the d-lines are parameterized by hyperbolic

arc length), and F = 0 (since the ρ-lines (which are Euclidean concentric
circles) and the d-lines (which are Euclidean diameters of the disc) are
orthogonal (in both the Euclidean and the hyperbolic metrics, for Euclidean
and hyperbolic angles are equal). Finally, we get
¿ À
∂ ∂ 4r2
G= , = = sinh2 d.
∂ρ ∂ρ hyp (1 − r2 )2
¤

Note that the metric coefficients are given by exactly the same expression
as in the formula (5.9) for k = −1. Hence we proved that the definition of
the hyperbolic plane via the Poincaré model is equivalent to the one given
in Section 5.1 by (5.9).

Hyperbolic planes of different curvatures. Together with the hyper-

bolic plane given by the Poincaré model, we can consider a continuum of
other spaces just by multiplying the hyperbolic length by a positive con-
stant. We will see that the Gaussian curvature of the hyperbolic plane is
identically equal to −1; after rescaling its metric by multiplying it by a we
will obtain a Riemannian metric of Gaussian curvature −1/a2 . Hence one
can consider a hyperbolic plane of curvature −k for each positive k.
To better understand the meaning of the procedure of rescaling a Rie-
mannian metric, let us first apply it for the Euclidean plane. Namely let Ec
be the length space whose points are points of R2 , but the distance between
every two points is multiplied by c: dc (x, y) = c · |xy|.
Exercise 5.3.7. Show that all Ec are isometric to each other and thus to
E1 = R2 .
Hint: Consider the map hc : Ec → R2 , where hc (x) = cx. This map is
an isometry.

This example might prompt a completely wrong idea that this procedure
always leads to an isometric space. Actually, this happens only in excep-
tional cases. Indeed, let Sc be the length space whose points are points of
the unit 2-dimensional sphere S 2 , and the distance function dc is obtained
√
from the distance function of S 2 by multiplication by c. Then Sc is not
√
isometric to S1 = S 2 : Sc is isometric to a sphere of the radius c. Thus
we obtain the family of the spheres. The number c−1 is called the curvature
of the sphere (and it is equal to the Gaussian curvature of its Riemannian
metric; see the next chapter).
Analogously, for each positive c we consider a space Hc obtained from
√
H by multiplying all distances by c. The formula for its length structure
164 5. Smooth Length Structures

reads: ¯ Z ¯
b
1 ¯¯ dγ(t) ¯¯
√
L(γ, a, b) = c ¯ dt ¯ dt.
a y
We will see that all spaces Hc are different (pairwise nonisometric). The
number −1/c2 is called the curvature of Hc (and we will also see later that
it is equal to its Gaussian curvature).
Remark 5.3.8. Let us fix a number r, and let Bk be a ball of radius r in
Hk . Then the spaces Bk “converge” to B0 as k → 0, and thus on any fixed
scale the hyperbolic metric with k → 0 looks more and more like a Euclidean
one (one can formalize this by choosing two normal coordinate systems and
comparing distances between points with the same coordinates, as in the
proof of Lemma 5.1.16).1
Remark 5.3.9. In fact, already the choice of the rigid motions of the
hyperbolic plane in the Poincaré model determines the Riemannian length
structure up to a multiplicative constant. Indeed, let us try to choose
expressions for the metric coefficients. Since the group of isometries contains
the translations along the x-axis, the expressions for metric coefficients
cannot depend on the x-coordinate, and thus they are functions of y. Since
Euclidean homotheties are hyperbolic isometries, the metric coefficients
must be homogeneous of order −2; that is, they have to satisfy f (cy) =
f (y)/c2 for all positive c. Hence each metric coefficient has to be of the
form const · y −2 . Now one notices that the derivatives of the inversions that
fix a given point generate all rotations around this point, and hence the
hyperbolic scalar product has to be a multiple of the Euclidean one. This
uniquely determines E = G = const · y −2 , F = 0.

Rigid motions and hyperbolic lines. We want to make sure that our
definitions of geometric notions introduced via the Poincaré model are con-
sistent. First, let us verify that the hyperbolic rigid motions are isometries.
Since all transformation called “the hyperbolic rigid motions” are bijective,
it suffices to check that they preserve the hyperbolic length structure.
Parallel translations along the x-axis obviously leave the integrand in
(5.13) unchanged, for they do not change the magnitude of γ ′ , nor the y-
coordinate. The same applies to the symmetries in vertical lines.

1It is a historical fact that Gauss tried to experimentally verify whether the physical space
is non-Euclidean. He did this by measuring angles in a huge triangle (with the tops of three
mountains as vertices, and using light rays as the sides of his triangle). He observed that, with
available precision, the sum of the angles was π. This allowed him only to conclude that, even
if that was a hyperbolic triangle, the curvature had to be very small. As a matter of fact, the
measurements of Gauss did not have nearly enough precision to detect effects caused by general
relativity, or otherwise he indeed would have noticed that the sum of the angles is different from
π, indicating that physical space is indeed curved!
5.3. Hyperbolic Plane 165

Consider a homothety with a coefficient c. Without loss of generality we

assume that its center is at the origin, for we can always compose it with
a translation along the x-axis. For a curve γ(t) = (x(t), y(t)), the velocity
of its image cγ(t) = (cx(t), cy(t)) under the homothety is cγ ′ (t). Thus the
magnitude of the velocity gets multiplied by c. Since the y coordinate of
γ(t) also gets multiplied by c, the integrand in (5.13) remains unchanged.
For an inversion mapping p to q, one can see that its differential is
the composition of the differential of the (unique) homothety with the same
center sending p to q, and a line symmetry. Thus the above argument applies
to inversions as well.
One can also see this from a straightforward computation as well. Con-
sider an inversion I. Without loss of generality we assume that I is the
inversion in the unit circle centered at the origin, for we can always achieve
this by composing it with an appropriate homothety and a translation along
the x-axis. The inversion I acts by the formula
x y
I(x, y) = ( 2 , 2 ).
x + y x + y2
2

For a curve γ(t) = (x(t), y(t)), the length of its composition with I is

Z b ¯ dI(γ(t)) ¯
¯ ¯
L(I(γ, a, b)) = Iy (γ(t))−1 ¯ ¯ dt
a dt
Z bµ ¶−1 sµ ¶2 µ ¶2
y d x d x
= + dt = L(γ, a, b).
a x2 + y 2 dt x2 + y 2 dt x2 + y 2
Exercise 5.3.10. Re-prove these statements using a complex representation
az + b
T (z) = , a, b, c, d ∈ R, ad − bc = 1,
cz + d
for a hyperbolic rigid motion T .
Exercise 5.3.11. Let us denote by Ta,b,c,d a hyperbolic rigid motion given
by a complex formula
az + b
Ta,b,c,d (z) = , a, b, c, d ∈ R, ad − bc = 1.
cz + d
Verify that the composition of transformations Ta,b,c,d and Ta′ ,b′ ,c′ ,d′ corre-
sponds to the product of matrices with entries a, b, c, d and a′ , b′ , c′ , d′ :
Ta,b,c,d ◦ Ta′ ,b′ ,c′ ,d′ = Taa′ +bc′ ,ab′ +bd′ ,ca′ +dc′ ,cb′ +dd′ .
Hence the group of orientation preserving hyperbolic rigid motions is iso-
morphic to SL2 (R)—the group of 2 × 2-matrices with determinant 1.
166 5. Smooth Length Structures

As was mentioned already, the hyperbolic plane is homogeneous, and

actually a very small part of its isometry group makes it homogeneous.
Indeed, in the Poincaré model, if two points have the same y-coordinates,
there exists a parallel translation along the x-axis that maps one of the
points onto the other. If the y-coordinates of two points are different, then
the line through these points intersects the x-axis. There is a homothety
with its center at the intersection point that maps one of the points onto
the other.
Definition 5.3.12. A Riemannian manifold is called isotropic with respect
to a group of transformation G if, for every two geodesic rays, there exists
a transformation from G that maps one ray to the other one.
Proposition 5.3.13. The hyperbolic plane is isotropic (with respect to its
isometry group).

Proof. The proof becomes obvious if we use the model in the disc. Indeed,
both vertices of the rays can be mapped to the center of the disc (by
homogeneity). Now the rays are represented by two semi-open Euclidean
segments (two radii of the disc), and there is a rotation that maps one of
them to the other. ¤

The proposition immediately implies:

Lemma 5.3.14. Given two points p, q in the disc model, there exists a
hyperbolic rigid motion that maps p and q to two points that lie in the same
diameter of the disc.
Of course, the Euclidean plane and the spheres are isotropic as well as
the hyperbolic planes, and this fundamental property is usually included in
Euclid-type lists of axioms.
As a matter of fact, the spheres, the Euclidean plane, and the hyperbolic
planes give an exhaustive list of two-dimensional isotropic spaces.
Exercise 5.3.15. 1. Prove the previous proposition directly in the Poincaré
model (without using the model in the disc).
2. Show that the orientation-preserving isometry I that maps a given
ray to a given ray is unique (as well as the one that reverses the orientation).
3. Show that our list of rigid motions of the Poincaré model is complete
(that is, it includes all isometries of the hyperbolic plane).
Definition 5.3.16. A metric space (X, d) is said to be fully homogeneous
if, for every two subsets A, B ⊂ X, and an isometry f : A → B (with respect
to the restrictions of the distance function to A and B), f can be extended
to an isometry of the entire X: there exists an isometry f˜: X → X such
that f˜|A = f ).
5.3. Hyperbolic Plane 167

Exercise 5.3.17. Give an example of a homogeneous space which is not

fully homogeneous.
Exercise 5.3.18. Show that the hyperbolic plane is fully homogeneous. In
particular, for two hyperbolic triangles with pairwise equal lengths of sides,
there exists a rigid motion that maps one of the triangles to the other one.
Remark. Notice that, while there is no essential difference between full
homogeneity and isotropy in dimension 2, full homogeneity is much stronger
in higher dimensions. Can you give an example of an isotropic Riemannian
manifold which is not fully homogeneous?

We will use the richness of the group of hyperbolic rigid motions to show
that the hyperbolic lines are shortest paths. Notice that we will prove not
only that hyperbolic lines are geodesics (and hence locally shortest paths):
we will show that every segment of a hyperbolic line is the shortest path
between its endpoints!
First of all, let us prove a particular case of this statement:
Lemma 5.3.19. A segment that belongs to a diameter of the disc in the
disc model is the unique shortest path between its points.

Proof. The segment belongs to a d-line in the normal coordinate system

(ρ, d) (see Lemma 5.3.6). Hence it is a geodesic. Now the statement follows
from Lemma 5.2.9 and a trivial observation that the injectivity radius is ∞
(since the coordinates are defined in the entire disc of the disc model). ¤

Now, given two points p and q (in the disc model), by Lemma 5.3.14 we
can map them by a rigid motion I onto two points I(p), I(q) lying on the
same diameter of the disc. Since I is an isometry of the hyperbolic plane, the
image of a shortest path between p and q is a shortest path between I(p)
and I(q). According to Lemma 5.3.19, the unique shortest path between
I(p) and I(q) is the Euclidean segment [I(p), I(q)]. Therefore its pre-image,
which is an arc of the circle passing through p and q and orthogonal to the
boundary of the disc, is the unique shortest path between p and q.
Along the way we verified one of the principal axioms of both Euclidean
and hyperbolic geometries: there is exactly one line passing through any pair
of distinct points. Of course, this statement becomes obvious in the Poincaré
model: move the two points to the same vertical line of the Poincaré model
by a (hyperbolic) rigid motion. Now it is clear that two distinct points in the
same vertical line cannot belong to the same circle centered at the x-axis,
and hence the only line that passes through the points is the vertical ray.
Exercise 5.3.20. Draw an example of two intersecting lines that do not
intersect another line (a counter-example to the Fifth Postulate).
168 5. Smooth Length Structures

Exercise 5.3.21. Show that, for every two intersecting lines, there exists
a line that is perpendicular to one of the lines and does not intersect the
other one.

Angles in Poincaré model. Let us pay attention to an important prop-

erty of the Poincaré model. The identity map of the upper semi-plane H2
considered as a map of the Euclidean upper semi-plane onto the hyperbolic
plane has a remarkable property: it preserves angles between lines; i.e., the
Euclidean angle between curves is equal to the hyperbolic angle between
their images. This is clear from (5.12).
The traditional term for angle-preserving maps in Riemannian geometry
is conformal maps. Thus the identity map of the upper semi-plane onto
H2 is a conformal map. This makes the Poincaré model very convenient:
although Euclidean (straight) lines rarely remain straight with respect to
the hyperbolic metric, we can still analyze angles between curves as if they
were usual Euclidean angles. The disc model is also conformal, for it is
obtained by applying an inversion to the Poincaré model, and inversions are
conformal maps.
Remark. There are models representing the hyperbolic plane as subsets
of the Euclidean plane and such that the hyperbolic lines are represented by
Euclidean lines (or segments). These models, however, are never conformal,
and they prove to be less convenient in most cases.

5.3.3. Ideal boundary. Recall that the boundary circle {(x, y) :

x2 + y 2 = 1} does not belong to the disc model. It is useful to adjoin
it to the hyperbolic plane, thus obtaining a compact space homeomorphic
to a disc. Let us call the boundary disc of the disc model the ideal boundary
of the hyperbolic plane; we will denote it by Γ. The inversion that maps
the disc model to the Poincaré model maps the boundary of the disc to the
x-axis, and one point of the boundary disc has no image (one can imagine
that it is mapped to a “point” with y = ∞). Hence, if working with the
Poincaré model, one should think of the ideal boundary as consisting of the
x-axis and one more “infinitely remote point y = ∞”. To avoid this com-
plication, we suggest sticking to the disc model until we give an intrinsic
definition of the ideal boundary.
There is no reasonable way to extend the hyperbolic metric to the ideal
boundary. However, there is a natural topology (a notion of converging
sequences) in the union H2 ∪ Γ, which makes it a compact space (and this
compactification is finer than the one-point compactification). By definition,
a sequence of points ai ∈ H2 converges to a point a ∈ Γ if it converges to a in
the Euclidean topology of the disc model. Two oriented lines in the disc are
said to be asymptotic if they are represented by circular arcs (or diameters)
5.3. Hyperbolic Plane 169

whose closures have one endpoint in Γ the same. The property of lines to
be asymptotic is an equivalence relation (prove this). Nonintersecting and
not asymptotic lines are said to be hyperparallels.

Intrinsic description of the ideal boundary. A disadvantage of these

definitions is that they appeal to a model, and hence their intrinsic meaning
(if any) remains obscure. Here is an intrinsic way of defining Γ.
One says that two geodesics γ(t), γ1 (1) are asymptotic if the (hyperbolic)
distance d(γ(t), γ1 (t)) is bounded for t ∈ (0, ∞).
Exercise 5.3.22. 1. Prove that this definition is consistent with the one
given before.
Hint: We need to estimate the hyperbolic distance between two Euclid-
ean circular arcs meeting at the same point of the boundary circle of the
disc model. Without loss of generality we can assume that this point is
(0, −1) (by homogeneity). Converting these lines to the Poincaré model by
the inversion I, we get two vertical rays. Now this is an easy computation.
2. Prove that if geodesics γ(t), γ1 (t) are asymptotic, then there exists a
constant t0 such that limt→∞ d(γ(t), γ1 (t + t0 )) = 0.
3. Show that, for this choice of t0 , the limit
lim e−t d(γ(t), γ1 (t + t0 ))
t→∞

is a positive (finite) number.

Hint: Use the hint to the first part.

Now we can define Γ as the set of equivalence classes of asymptotic

lines. One says that a sequence of points pi ∈ H converges to [γ] ∈ Γ
if d(pi , γ(0)) → ∞ and the angle αi between γ and the geodesic segment
[γ(0), pi ] at γ(0) tends to 0 as i → ∞.
Exercise 5.3.23. Verify that this definition is equivalent to the definition
via the disc model.

For two asymptotic lines ab, ac with a common point a ∈ Γ, we define

the angle between them at a to be zero. This definition is motivated by the
fact that, for two points b, c ∈ H2 and a sequence ai ∈ H2 , ai → a as i → ∞,
the angle ∡bai c tends to zero as i → ∞ (prove this!).
Now, in addition to “bounded” triangles, we will consider ideal triangles.
We say that a triangle is ideal if one or more of its vertices belong to the
ideal boundary. According to our convention, the angle at such a vertex is
equal to zero by definition. For example, if all vertices of a triangle are in
the ideal boundary, then all its angles are equal to zero.
170 5. Smooth Length Structures

It is clear (from the disc model) that, for an oriented line l and a point
p∈/ l, there is exactly one line lp+ passing through p and asymptotic to l. If
we reverse the orientation of l, we also get one line lp− that passes through a
and is asymptotic to l with reversed orientation. All lines passing through
p “between” lp+ and lp− are hyperparallel to l (i.e., do not intersect it and
are not parallel to it). We suggest the reader make a sketch to visualize this
observation.

Horocycles. Busemann functions.

Definition 5.3.24. Let a = [γ] be a point in the ideal boundary Γ. A curve

orthogonal to all geodesics asymptotic to γ is called a horocycle centered
at a.

Exercise 5.3.25. Prove that in the Poincaré model the horocycles are
represented by the circles tangent to the x-axis and the horizontal lines.
Prove that in the disc model the horocycles are represented by the circles
tangent to the boundary circle.

To understand the metric nature of horocycles, let us introduce a Buse-

mann function Bγ : H2 → R associated with a geodesic γ:
Bγ (p) = lim (d(p, γ(t)) − t).
t→∞

Exercise 5.3.26. Prove that the limit does exist.

Hint: Notice that the expression (d(p, γ(t)) − t) is monotone decreasing
in t (by the triangle inequality).

For a fixed t, one can think of the function dt (p) = d(p, γ(t)) − t as the
distance function to γ(t) normalized by subtracting t to get zero at γ(0).
Hence a level curve Sτ,t = {p ∈ H : d(p, γ(t)) − t = −τ } of dt is a circle
centered at γ(t) and passing through a point γ(τ ) (orthogonally to γ ′ (τ )).

Exercise 5.3.27. Prove that the level curves Sτ = {p ∈ H : Bγ (p) = τ } are

horocycles.

Now one can think of a horocycle Sτ passing through γ(τ ) as a limit

of spheres “whose centers tend to infinity along γ”. This motivates us to
define a horoball as a set Dτ = {p ∈ H : Bγ (p) ≤ τ }.

Exercise 5.3.28. Prove that a horoball Dτ is the union of the balls

Ballγ(t) (t − τ ), t ≥ τ .

Exercise 5.3.29. What would this construction produce in the Euclidean

plane? Answer: a semi-plane orthogonal to γ.
5.3. Hyperbolic Plane 171

There is an example of a web formed by a family of asymptotic lines and

a family of horocycles orthogonal to them that is very easy to visualize in
the Poincaré model. It consists of the families of vertical rays and horizontal
lines, which are coordinate lines in Cartesian coordinates of the xy-plane.
(Note that Cartesian coordinates are not normal for the hyperbolic metric,
for the y lines are not parameterized by the arc length).
Exercise 5.3.30. Convert Cartesian coordinates into hyperbolic normal
coordinates by reparameterizing y1 = ϕ(y), and compute metric coefficients
in this coordinate system.

The reader probably has already realized that a web formed by a family
of asymptotic geodesics and the family of horocycles orthogonal to them can
always be represented by coordinate lines in a normal coordinate system. In
particular, horocycles form an equidistant family.

Busemann compactification. Let us describe a general construction of

the Busemann compactification of a (complete locally compact) length
space. This compactification is finer than the one-point compactification,
and adding the ideal boundary to the hyperbolic plane is an example of this
compactification.
Consider a locally compact and complete length space X. Denote
by Lip(X) the space of all Lipschitz-1 functions on X with C 0 -metric
dC 0 (f, g) = sup |f − g| on it. Note that C 0 -distance between two Lipschitz
functions on a noncompact space can be infinite—but this is not a problem
since our notion of a metric space allows infinite distances. If one uses a
more restricted notion of a metric space, Lip(X) can be decomposed into
a (usually a continuum) family of metric spaces with C 0 -metric on each of
them, and infinite supremum of the difference between any two functions
from different spaces.
One can isometrically embed X into Lip(X) by x → d(x, ·). (Prove
yourself that this map really is an isometry.) It is more convenient to fix a
reference point y ∈ X and define an embedding by x → dx = d(x, ·)−d(x, y).
This embedding puts into correspondence to a point x a distance function
dx (z) = d(x, z) − d(x, y) normalized (by subtracting a constant d(x, y))
to make it vanish at y. Note that C 0 -distance between any two distance
function is finite. We say that a function g ∈ Lip(X) is a distance-like
function if dC 0 (g, f ) < ∞ for some (and hence for every) distance function
f . An example of a distance-like function is a function d(x, ·) + const.
For a noncompact X, the image of the embedding x → dx = d(x, ·) −
d(x, y) is not closed in Lip(X). Let us consider its closure, which is called
the Busemann compactification of X. It contains an isometric copy of X
(whose points are represented by distance functions normalized to be zero
172 5. Smooth Length Structures

at a reference point y), and a set of limit points of the image of X. This
set of limit points, denoted by X(∞), is the Busemann boundary at infinity
of X, and its elements are called Busemann functions. These are distance-
like Lipschitz-one functions that arise as limits (uniform on compacts) of
sequences of normalized distance functions dn (z) = d(z, xn ) − d(xn , y).
There is another notion of the boundary at infinity (there are at least
five different and useful notions of boundaries at infinity). The ray boundary
consists of equivalence classes of rays starting from a reference point (and
two rays are equivalent if they are asymptotic, that is, stay within bounded
distance from each other). This set is equipped with the topology of uniform
convergence on open sets. For the hyperbolic plane, ray boundary and
Busemann boundary coincide (check this!).

Exercise 5.3.31. 1. What is the Busemann boundary at infinity for the

Euclidean plane? For a free group with two generators (with respect to a
word metric)?
2. Show that Busemann functions of rays in the hyperbolic plane are a
particular case of general Busemann functions described here (for a sequence
of points xn escaping to infinity along a ray).
3. Show that, for the hyperbolic plane, its Busemann compactification
coincides with its compactification we described earlier.
4. Give an example of a length space whose Busemann boundary at
infinity is different from its ray boundary.

5.3.4. Gauss-Bonnet Formula. It is well-known that the sum of angles

of every planar triangle is equal to π. This property is equivalent to the
Fifth Postulate. (It has been known for centuries that the Fifth Postulate
follows from the existence of just one triangle whose sum of angles is π.)
This is a particular case of a more general Gauss-Bonnet Formula, which
in its turn is a particular case of Theorem 6.3.17. In a certain sense, this
formula gives a qualitative version of Exercise 6.5.7
As usual, by a triangle in a length space we understand three points
a, b, c connected by three shortest segments [ab], [bc], and [ca]. Now we are
going to consider a triangle that bounds a region T in a hyperbolic plane or
a sphere (ideal triangles in a hyperbolic plane are not excluded). Then the
following Gauss-Bonnet Formula holds.

Theorem 5.3.32.

∡a + ∡b + ∡c − π = k · Area(T ),

where k is a curvature of the space.

5.3. Hyperbolic Plane 173

Proof. It is enough to consider two cases: k = 1 and k = −1, and then use
a dilation.
Case 1: k = 1. A spherical triangle can be represented as the inter-
section of three semi-spheres. Our argument will consist of an elementary
application of the inclusion-exclusion formula.
Let a′ , b′ , c′ be the points of the unit sphere opposite to a, b, c,
respectively. For each pair of opposite points, say a, a′ , one of the two
bi-gons B(aa′ ), B ′ (aa′ ) with the vertices a and a′ covers the triangle △abc,
and the other one covers the antipodal triangle △a′ b′ c′ . Notice that the area
of each of these bi-gons is equal to twice the angle α of △abc at a.
All the six bi-gons cover the entire sphere with multiplicity 1 (each point
is covered by exactly one of them) except that the triangles △abc and △a′ b′ c′
are covered three times each. By the inclusion-exclusion formula we get

4π = Area (S 2 )
= 2(Area (B(aa′ )) + Area (B(bb′ )) + Area (B(cc′ ))) − 4Area (T ).

Together with the observation that the area of each of the bi-gons is twice
the corresponding angle of △abc this proves case 1.
Case 2: k = −1. Observe that if a triangle (maybe an ideal one) is cut
into two triangles and the conclusion of Theorem 5.3.32 holds for two of the
three triangles, then it also holds for the third one. Every triangle can be
obtained by cutting an ideal triangle off another ideal triangle. Hence it is
enough to prove 5.3.32 for ideal triangles.
Consider an ideal triangle △abc, a ∈ Γ. Using the Poincaré model, one
can map it (by a hyperbolic rigid motion) to such an ideal triangle that two
of its sides belong to two vertical rays, and two vertices lie on the circle
x2 + y 2 = 1 (and the third one is mapped to the “infinitely remote point
y = ∞). An elementary (Euclidean) geometric consideration shows that if
the angles of this triangle are α and β (α is at the left vertical side and β is at
the right one), then its vertices have Cartesian coordinates (− cos α, sin α)
and (cos β, sin β).
Now the argument is based on a straightforward integration. Indeed,
the hyperbolic area of the triangle T is equal to
Z cos β Z ∞
dy
Area (T ) = dx √ 2
= π − α − β.
− cos α 1−x2 y

Exercise 5.3.33. Formulate and prove an analog of the Gauss-Bonnet

Formula for (hyperbolic and spherical) polygons.
174 5. Smooth Length Structures

Exercise 5.3.34. Show that the area of a hyperbolic triangle does not
exceed π.
Remark 5.3.35. There is a remarkable trick that allows one to prove many
statements in hyperbolic geometry by their spherical analogs. For example,
in the proof of the Gauss-Bonnet Formula we could not apply the inclusion-
exclusion formula in Case 2 since the hyperbolic plane has infinite area.
However, this trick allows one to apply the result of Case 1 to claim that
the Gauss-Bonnet Formula is correct in general. For instance, one can also
use this trick to show that the altitudes of a hyperbolic triangle intersect in
one point.
Here is the idea of the trick. There are many ways to formalize the
claim that spheres, the Euclidean plane and hyperbolic planes form a one-
parameter family. For instance, formulas (5.9), (5.8), and (5.7) √ for the metric
coefficients can be represented as one formula G(r) = (Re sin( k r))2 using
complex numbers. Now if an identity that can be expressed by a formula
via metric coefficients is valid for all positive values of k, then by analyticity
it is true for all k (for a nonzero analytic function has only isolated zeroes).
All details are left to the interested reader.

5.3.5. Selected features of hyperbolic geometry. We finish this sec-

tion by formulating a few distinctive geometrical properties of the hyperbolic
plane. The proofs are left to the reader. These features are mainly related
to large-scale properties of the hyperbolic plane, and many of them will be
generalized in Section 8.4. We deal with the hyperbolic plane (of curvature
k = 1), leaving it as a useful exercise to figure out how our statements should
be modified for other values of k.

Growth of balls. Everybody knows that the length of a Euclidean circle

of radius r is 2πr, and the area of the ball enclosed in this circle is πr2 .
By Lemma 5.1.16, small circles and balls in a Riemannian metric look
almost like their Euclidean counterparts: in particular, the length of a small
hyperbolic circle is 2πr +o(r2 ). In contrast, for large r hyperbolic circles and
balls grow exponentially fast: the length of a circle of radius r is π(er − e−r ).
Surprisingly enough, for large r the area of the ball enclosed in this circle is
smaller than the length of the circle: it is π(er + e−r − 2). In particular, the
annulus between two concentric circles of radii r and r + 1 contains more
area than the ball enclosed in the circle of radius r.

Large triangles. First of all, the Gauss-Bonnet Formula implies that the
area of a triangle is at most π. Since the area of a circle of radius 1 is greater
than π, the radius of a circle inscribed in a triangle is less than 1 (of course,
this is a very rough upper bound; can you give a better estimate?) This
5.3. Hyperbolic Plane 175

means that if one chooses three points in the hyperbolic plane (arbitrarily
far apart), the three segments connecting the points “almost stick together
in the center of the triangle”: they intersect the same disc of radius 1! Hence
each side of a hyperbolic triangle is contained in the 2-neighborhood of the
union of the two other sides. When seen from far away, a big hyperbolic
triangle looks like a union of three rays emanating from the center of the
triangle: the triangle is contained in the 1-neighborhood of this union.
Loosely speaking, the triangle looks very “slim”, as if its sides have been
“sucked inside”.
Using these considerations, it is easy to observe the following remarkable
phenomenon. Consider a triangle ∆abc, and let p be a point on the
side [bc]. Then at least one of the triangles ∆abp and ∆acp is almost
degenerate, in the sense that at least one of the following inequalities holds:
d(a, b) + d(b, p) ≤ d(a, p) + 4 or d(a, c) + d(c, p) ≤ d(a, p) + 4. This property
will be used to define Gromov hyperbolic (δ-hyperbolic) spaces.
Note that, unlike Euclidean geometry, there is a criterion for hyperbolic
triangles to be congruent “by three angles”: two triangles are congruent
(one of them can be obtained from the other by a rigid motion) if each
angle of one triangle is equal to the corresponding angle of the other one.
On the other hand, funny enough, but there are no rectangles in hyperbolic
geometry!

Morse lemma. A quasi-geodesic with a quasi-geodesic constant c is a curve

γ such that any segment of γ is at most c times longer than the distance
between its endpoints:
L(γ, t1 , t2 ) ≤ cd(γ(t1 ), γ(t2 ))
for all t1 , t2 . Of course, c cannot be smaller than 1 unless γ is a constant
curve; if c = 1, then γ is a shortest path. We are concerned with the following
question: for a fixed quasi-geodesic constant c, how far can a quasi-geodesic
deviate from a geodesic between its √ endpoints? In the Euclidean case this
deviation can be at most const · d, where d is the distance between the
endpoints of the quasi-geodesic. (Prove √ this and give an example when this
deviation is indeed as large as const · d.)
In contrast to the Euclidean case, the Morse Lemma asserts that there
exists a constant D (depending on c) such that every quasi-geodesic in the
hyperbolic plane lies within distance D from the geodesic segment between
its endpoints. For a proof of the Morse Lemma (in a much more general
form) see Section 8.4.
There is the following striking reformulation of the Morse Lemma.
Suppose that we change the Riemannian metric of the hyperbolic plane
in such a way that the ratio of the two metrics remains bounded. Then
176 5. Smooth Length Structures

every shortest path of the new metric stays within a bounded distance from
a hyperbolic line. This suggests that the large-scale structure of geodesics
in the hyperbolic plane is very stable; this phenomenon is closely related to
structural stability of hyperbolic (Anosov) flows.

Curvature of horocycles. A Euclidean circle of a large radius locally

looks like a straight line: its curvature tends to zero as the radius grows.
As opposed to the Euclidean case, (geodesic) curvature of a circle in the
hyperbolic plane tends to 1 as the radius grows to infinity (see Exercise
6.4.9).
This fact makes it possible for the length of a circle to grow exponentially.
Indeed, recall that the derivative of the length of a curve under an equidistant
variation is equal to the integral of the geodesic curvature of the curve. The
geodesic curvature of a large circle is approximately 1, and this means that
the derivative of the length is approximately equal to the length itself.
Since circles passing through a fixed point at a fixed direction converge
to a horocycle (as the radius tends to infinity), this means that all horocycles
have geodesic curvature 1. Recall that horizontal lines in the Poincaré model
are horocycles. Thus we see that intrinsically they are curved with curvature
1. Which way do they bend: up or down?
Developing this idea, let us consider a curve of constant geodesic curva-
ture. If this curvature is greater than 1, then the curve will close up (forming
a circle), exactly like a Euclidean curve of constant curvature. However, if
the curvature of a curve is at most 1, the curve will not close up! Imag-
ine that you are driving a “hyperbolic car”, and you keep turning with the
same curvature (less than 1). Then, against all intuition, you will neverthe-
less drive away, and moreover you will drive away staying within bounded
distance from a straight line!

Tessellations by regular polygons. As usual, we say that a polygon is

regular if all its sides are equal to each other and all its angles are equal to
each other. In other words, an n-gon is regular if its group of isometries
contains Zn . In the disc model the vertices of a regular Euclidean polygon
centered at the origin also form the set of vertices of a regular hyperbolic
polygon (certainly the sides are different, for Euclidean segments are not
hyperbolic geodesics). One can see that, whereas the angles of a small
hyperbolic regular n-gon are almost equal to those of a Euclidean one, in
the hyperbolic case the angles tend to zero as the length of sides of the n-
gon tends to infinity. In particular, if n > 4, one can choose a regular n-gon
whose angles are 2π/n. Then the entire hyperbolic plane can be tessellated
by isometric copies of this polygon, with n polygons meeting at each vertex.
This tessellation is similar to a tessellation of the Euclidean plane by square
5.3. Hyperbolic Plane 177

tiles. However, the only regular tiles that can be used to pave the Euclidean
plane are triangles, squares and hexagons, whereas the hyperbolic plane can
be paved by regular n-gons for all n.

Discrete groups of hyperbolic isometries. The isometry group of the

Euclidean plane does not contain many interesting discrete subgroups: every
infinite discrete subgroup G contains in its turn a finite-index subgroup
isomorphic to Z or Z2 . If we impose an additional restriction that the
quotient space R2 /G is a manifold, there are only two examples: G = Z2 ,
in which case R2 /G is a torus, and an extension of Z2 by one line symmetry
(in this case R2 /G is a (nonorientable!) Klein bottle). The group of
hyperbolic isometries is much richer, and it contains a whole world of discrete
subgroups. Recall that, by Exercise 5.3.11, the group of hyperbolic rigid
motions is isomorphic to SL2 (R). Hence for instance SL2 (Z) acts on H by
hyperbolic isometries.
Once we have a group G acting by isometries on a hyperbolic or Euclid-
ean plane, there is a tessellation by polygons associated with this action.
To construct this tessellation, choose a point p, and consider the set of its
images {g(p), q ∈ G}. Then the plane is tessellated by Voronoi regions Vg(p) ,
where
Vg(p) = {q : d(q, g(p)) ≤ d(q, g ′ (p)) for all g ′ ∈ G}.
Tessellations of the hyperbolic plane by regular 4k-gons, k > 1, can be
obtained this way by groups acting by isometries without fixed points. The
corresponding quotient spaces are two-dimensional closed surfaces (of higher
genus) together with Riemannian metrics of constant curvature −1.

5.3.6. Appendix: inversion. By definition, an inversion Ip,r with respect

to a circle of radius r centered at p ∈ R2 maps a point q to a point Ip,r (q)
on the ray [pq) uniquely determined by the relation |pq| · |pIp,r (q)| = r2 .
Note that Ip,r is not defined at p. Every inversion Ip,r can be obtained as
a composition of the inversion I = I(0,0),1 , with homothety and a parallel
translation. If one regards points in R2 as complex numbers, the inversion
I is given by a very simple formula
1
(5.16) I(z) = ,
z̄
where z̄ denotes the complex conjugation. We list the properties of inversions
that are important for our exposition. The proofs, which can be considered
as exercises in high-school geometry (or in operations with complex numbers
using (5.16)), are omitted.
2 = id. Each inversion changes
1. An inversion is an involution, i.e., Ip,r
the orientation.
178 5. Smooth Length Structures

2. An inversion is a conformal map, i.e., it preserves angles between

curves. This means that the angle between the images of two curves is
equal to the angle between the curves (we need this property for circles and
lines only).
3. If a circle does not pass through the center of an inversion, its image
is a circle. Otherwise it is a line. If a line does not pass through the center
of an inversion, its image is a circle. Otherwise it is the same line.
4. The fixed points of an inversion Ip,r are the points of the circle of
radius r centered at p (the circle of the inversion). The derivative of the
inversion at each point of the circle is a line symmetry.

5.4. Sub-Riemannian Metric Structures

All length structures we considered so far were constructed by introducing
new length functionals. The class of admissible paths was rather standard:
we used continuous or piecewise smooth paths. This section describes a
remarkable type of length structure defined by modifying a class of admis-
sible paths. Our introduction to sub-Riemannian geometry amounts to a
definition of its length structure and a brief consideration of one model ex-
ample. For further reading we recommend [BR]. Whereas this section may
be omitted for the first reading, it is worth returning to it later. We hope
that some readers will be intrigued enough by it to get interested in studying
sub-Riemannian geometry.
We consider the simplest case, starting from a region in Ω ⊂ R3 with its
standard Euclidean metric; replacing it by a Riemannian metric would not
change anything essential but require slightly more cumbersome notation.
There are, however, nontrivial phenomena that show up only in higher
dimensions; they remain beyond the scope of our exposition.
Later parts of this section assume that the reader is familiar with Lie
bracket and differential forms.

5.4.1. Carnot–Carathéodory metrics. In a wide class of examples ad-

missible curves are defined by imposing restrictions on their velocity at
each point. For instance, for every point p in Ω ∈ R3 , choose a two-plane
Hp ⊂ Tp Ω (through 0 ∈ Tp Ω). Assume that Hp smoothly depends on p.
This object is called a distribution (of two-planes), or a two-plane field. Let
us call the planes Hp admissible. We say that a (piecewise smooth) curve
γ is H-admissible if its velocity vector γ ′ (t) at every point always lies in the
admissible plane at the point: γ ′ (t) ∈ Hγ(t) . It may sound unbelievable,
but for a generic distribution every two points in Ω can be connected by an
admissible curve!
5.4. Sub-Riemannian Metric Structures 179

The following considerations can only add skepticism to this statement.

First of all, for a two-dimensional region Ω an analog of a distribution is a
line field; then every accessible curve lies in an integral curve of the line field.
Hence the locus of points accessible from a given point is one-dimensional,
exactly as intuition prompts.
Furthermore, let us try to look at an example when we choose admissible
planes parallel to each other: for instance, we choose a horizontal plane at
every point. Of course, in this case every admissible curve is confined to a
horizontal plane. Hence two points with different z-coordinates cannot be
connected by an admissible curve. A criterion for connectivity via admissible
curves will be formulated later, see 5.4.6, but it is already clear that the
planes should not be tangent to a family of surfaces (unlike vector fields, it
is known that a general field of planes is not even locally tangent to a family
of surfaces).
Now a sub-Riemannian metric structure associated with a distribution
H is given by the class of H-admissible curves with Euclidean length. The
resulting intrinsic metric space is often called a Carnot–Carathéodory space.

Note that Carnot–Carathéodory metrics can be regarded as limits of

Riemannian metrics. Indeed, choose a quadratic form Qp on each Tp Ω so
that its kernel is Hp . Consider the following family of Riemannian metrics,
which depend on a real parameter h:
hV, V iR = hV, V i + h · Qp (V, V ),
where hV, V i is the usual scalar square of V ∈ Tp Ω. Denote the correspond-
ing Riemannian distance function by dh (p, q). Loosely speaking, the term
h · Qp (V, V ) is a penalty for going in a direction that does not belong to Hp
(recall that Hp is the kernel of Qp ).

Exercise 5.4.1. Prove that limh→∞ dt (p, q) is the Carnot–Carathéodory

distance between p and q.

The reader familiar with constraints in mechanics may notice a similarity

here: to confine a particle to a surface, one introduces a potential which
becomes huge as one moves away from the surface. However for a generic
distribution there are no surfaces tangential to it; hence such constraints
are called nonholonomic. A classical example of a ball rolling on a plane
without skidding is briefly discussed below.

Geometric control theory. There is a slight modification of this con-

struction. Let Vi , i = 1, 2, . . . , k, be several smooth vector fields on Ω. We
say that a path γ(t) is admissible if there are real-valued functions vi (t),
180 5. Smooth Length Structures

i = 1, . . . , k, such that
X
γ ′ (t) = vi (t)Vi (γ(t)).
This relation means that γ is allowed to move along any linear combination
of Vi ’s. One can think of Vi ’s as available controls to direct a moving point;
then vi ’s would be control functions. For two vector fields V1 and V2 that
are linearly independent at every point, one can introduce a distribution H
of the planes spanned by V1 and V2 , thus reducing this construction to the
previous one. A difference arises if the vector fields happen to be linearly
dependent at some points.
Exercise 5.4.2. Show that, for V1 (x, y, z) = (1, 0, 0) and V2 (x, y, z) =
(0, 1, x), every two points are connected by an admissible path.

Geometric control: parking a car. Such classes of admissible paths arise

in many applications; the fields Vi ’s are often called controls. For instance,
consider a bicycle (on a plane). Riding a bicycle, a cyclist combines two
ways of controlling her bicycle: steering with the handlebars and pushing
pedals to move forward. On the other hand, the dimension of the space of all
possible position of the bicycle is three! Indeed, a position of the rear wheel
can be described by its coordinates (x, y); and we still need one number α
for the angle between the bicycle and the x-axis. Thus a position of the
bicycle corresponds to a point (x, y, α) in R2 × S 1 . This subspace is called
the configuration space for our model. Note that, although with only two
controls, the cyclist can reach every point in the 3-dimensional configuration
space.
Let us describe the corresponding distribution. When a bicycle moves,
its rear wheel follows the front one. In other words, the velocity of the path
traced by the rear wheel remains proportional to the vector connecting the
wheels. In coordinates (x, y, α) introduced above this condition is equivalent
to ẋ sin α = ẏ cos α. In other words, the velocity vector (ẋ, ẏ, α̇) of the point
moving in the configuration space at a point (x, y, α) must belong to the
plane orthogonal to (− sin α, cos α, 0). (Here ẋ means differentiation with
respect to time t.) Geometrically, at every point of the three-dimensional
configuration space R2 ×S 1 we have chosen a two-dimensional plane of those
(tangent) vectors which correspond to the velocities of actually possible
motions of the bicycle: the other vectors represent dragging the bicycle in
its side direction.
To produce a length structure, we need to fix a length function, though
its particular choice is not so important here: the restriction of the class
of admissible curves results in very unusual qualitative properties of the in-
duced sub-Riemannian metric, whereas modifications of the length function
cause relatively modest changes of the length structure. For instance, let us
5.4. Sub-Riemannian Metric Structures 181

use a Euclidean length given by

Z p
L(γ, a, b) = ẋ2 + ẏ 2 + α̇2 dt.

Let’s try to informally discuss the sub-Riemannian metric arising in this

example. Consider two points which are very close to each other. If the
vector between the two close points “lies close to an admissible direction”,
one should not expect that the new distance is essentially different from
Euclidean one. But if the vector between the points is perpendicular to
admissible planes, then an admissible path between the points has to go a
long way; for such points the distance is stretched. Everybody observes this
when doing parallel parking (and our model can be applied to cars as well as
to bicycles): you want to displace your car just one foot to its side direction,
but you have to drive 50 feet moving it back and forth. Perhaps parking a
car is not such a difficult task, but parking a car with two trailers requires
very advanced mastery. Geometrically, a car with two trailers gives rise to
the same type of intrinsic metric, just the dimension of the configuration
space is higher, while there are still only two controls.

Nonholonomic mechanical constraint: rolling a ball. Another exam-

ple of the same kind is rolling a ball. A position of a ball lying on a plane
requires five coordinates: two reals to characterize the point in the plane,
another two coordinates to characterize the point of the ball which touches
the plane, and the last one for spinning the ball around its vertical axis.
When one rolls the ball without sliding, there are only three admissible di-
rections (controls): two to choose a direction where to roll it and the third
one for spinning it. Still, one can get to any position regardless of the initial
position: the reader can check that experimentally rolling a globe on the
floor. Nevertheless, for certain positions which are visually close, it takes
quite a bit of rolling to get from one to the other: position the globe such
that it touches the floor by Washington, and try to roll it back to the same
point on the floor but such that now its bottom point is at New-York.

5.4.2. Connectibility and Ball-Box Theorem.

Lie bracket. Here we briefly recall the notion of Lie bracket; one can find
details and proofs in many books (see for example [BC]).
Two main objects associated with a smooth vector field V are the cor-
responding derivation of smooth functions f → V f and the one-parameter
group of diffeomorphisms ϕtV generated by V . Recall that, for each t, ϕtV is
a diffeomorphism, and for a fixed p,
d t
ϕ (p) = V (ϕtV (p)).
dt V
182 5. Smooth Length Structures

In other words, for each x, γ(t) = ϕtV (p) is an integral curve of V that passes
through p at t = 0. (A curve γ is an integral curve of a vector field V if
d
dt γ(t) = V (γ(t)). )
Here are three equivalent definitions of Lie bracket [V, W ] of vector fields
V , W.
Definition 5.4.3. Lie bracket [V, W ] is a vector field defined by the condi-
tion: [V, W ]f = V W f − W V f for every smooth function f .

Of course, one has to prove that there exists a unique vector field [V, W ]
that differentiates functions according to the formula in the definition. The
definition immediately implies that Lie bracket is linear in each argument
and skew-symmetric. It is easy to check that the following property holds:
[f V, W ] = f [V, W ] − (W f )V , where f is a smooth function.
The next definition is especially important in our geometric context:
d ¯¯ √ √ √ √
Definition 5.4.4. [V, W ](p) = ¯ (ϕV t ◦ ϕWt ◦ ϕV− t ◦ ϕ− t
W (p)).
dt t=0+
To better understand its geometric meaning, one can use a coordinate
chart (to be able to “subtract points”) and rewrite it in a less invariant form
as
1
[V, W ](p) = lim 2 (ϕtV ◦ ϕtW ◦ ϕ−t −t
V ◦ ϕW (p) − p).
t→0 t
We suggest the reader make a sketch with a “quadrilateral” formed by four
segments of integral curves of V and W traversed back and forth. The
“quadrilateral” has its first vertex at p, then its first side follows the integral
curve of W backwards for time t, and so on. This “quadrilateral” does
not quite “close up” (unless we are dealing with coordinate vector fields,
in which case we actually get a closed “quadrilateral”): when one follows
all four sides of the “quadrilateral”, he gets to a point p′ close to p; this
displacement p′ − p is of second order in t; dividing it by t2 and sending
t → 0 gives the Lie bracket.
The last definition stresses similarity between Lie brackets and usual
derivatives:
Definition 5.4.5.
1 d¯
[V, W ](p) = lim (dϕ−tV W (ϕtV (p)) − W (p)) = ¯t=0 dϕ−t
V W (γ(t)),
t→0 t dt
where γ is the integral curve of v with γ(0) = p.

The expression on the right-hand side of the formula in this definition

is called the Lie derivative of W with respect to the flow ϕtV . To justify
this name, imagine that you follow the integral curve of V passing through
p and study how W changes: you want to compare W (p) = W (γ(0)) and
5.4. Sub-Riemannian Metric Structures 183

W (γ(t)). These vectors belong to different tangent spaces, so to subtract

them we map W (γ(t)) to the tangent space at p by the differential dϕ−t
V of
−t
ϕV . Then, as usual, we subtract W (p) from it, divide the difference by t
and send t to 0.

Chow–Rashevsky’s Theorem. In this section we give a sufficient con-

dition for a Carnot–Carathéodory metric to be finite. This criterion is
known as Chow–Rashevsky Connectivity Theorem. Here is its simplest
3-dimensional case:
Theorem 5.4.6. Let Ω ⊂ R3 be a connected region, and V1 , V2 two smooth
vector fields such that V1 (p), V2 (p), and their Lie bracket [V1 , V2 ](p) form
a basis of Tp Ω for every p ∈ Ω. Then every two points p, q ∈ Ω can be
connected by an admissible path, that is, a piecewise smooth path γ whose
velocity vector at every (smooth) point is a linear combination of V1 and V2 :
γ ′ (t) = v1 (t)V1 (γ(t)) + v2 (t)V2 (γ(t)).

Proof. Let us prove that, for every p ∈ Ω, the set of points that can be
connected with p by admissible paths is open.
Let ϕt1 , ϕt2 be the one-parameter groups of diffeomorphisms generated
d t
by V1 , V2 : dt ϕi (p) = Vi (ϕti (p)), i = 1, 2. Recall that
d ¯¯ √ √ √ √
[V1 , V2 ](p) = ¯ (ϕ1 t ◦ ϕ2 t ◦ ϕ− 1
t
◦ ϕ−
2
t
(p)).
dt t=0+
Consider a map F : R3 → Ω:
√ √ √ √
|τ | sign(τ ) |τ | − |τ | −sign(τ ) |τ |
(5.17) F (u, v, τ ) = ϕu1 ◦ ϕv2 ◦ ϕ1 ◦ ϕ2 ◦ ϕ1 ◦ ϕ2 (p).
This map is at least C 1 -smooth (check this!) and
∂ ∂ ∂
∂u F (0) = V1 (0), ∂v F (0) = V2 (0), ∂τ F (0) = [V1 , V2 ](0),
where 0 = (0, 0, 0).
Hence by the Inverse Function Theorem, the image of F contains a
neighborhood of p. Now it remains to observe that a point F (u, v, τ ) is
connected with p by the following admissible path γ : [0, 6] → Ω, which
alternates moving along V1 and V2 (that is, it is built out of segments of
integral curves of V1 and V2 ):
√
−sign(τ ) |τ |t
γ(t) = ϕ2 (p), t ∈ [0, 1],
√ √
− |τ |(t−1) −sign(τ ) |τ |
γ(t) = ϕ1 ◦ ϕ2 (p), t ∈ [1, 2],
√ √ √
sign(τ ) |τ |(t−2) − |τ | −sign(τ ) |τ |
γ(t) = ϕ2 ◦ ϕ1 ◦ ϕ2 (p), t ∈ [2, 3],
√ √ √ √
|τ |(t−3) sign(τ ) |τ | − |τ | −sign(τ ) |τ |
γ(t) = ϕ1 ◦ϕ ◦ϕ ◦ϕ (p), t ∈ [3, 4],
√ 2 √ 1 √ 2 √
v(t−4) |τ | sign(τ ) |τ | − |τ | −sign(τ ) |τ |
γ(t) = ϕ2 ◦ ϕ1 ◦ ϕ2 ◦ ϕ1 ◦ ϕ2 (p), t ∈ [4, 5],
184 5. Smooth Length Structures

√ √ √ √
|τ | sign(τ ) |τ | − |τ | −sign(τ ) |τ |
γ(t) = ϕu(t−5) ◦ϕv2 ◦ϕ1 ◦ϕ2 ◦ϕ1 ◦ϕ2 (p), t ∈ [5, 6].
Hence every point in the image of F can be connected with p; since the image
of F contains a neighborhood of p, its connectivity component is open.
Since connectivity components are open and either coincide or do not
intersect, the theorem immediately follows from connectedness of Ω. ¤
Exercise 5.4.7. Show that under the assumptions of Theorem 5.4.6 every
two points in Ω can be connected by a smooth admissible path (as opposed
to a piecewise smooth one).

In full generality the Chow–Rashevsky Theorem reads as follows: let Vi ,

i = 1, 2, . . . , k, be smooth vector fields on a connected manifold Ω. Assume
that the Vi ’s, their Lie brackets, and their iterated Lie brackets (such as
[[[V1 , V2 ], V3 ], V4 ]) span Tp Ω for every p. Then every two points can be
connected by an admissible curve.
Chow’s Theorem for distributions can be formulated in the following
elegant way: Let Hp ∈ Tp Ω be a smooth distribution of linear subspaces on
a connected manifold Ω. Denote by H the space of vector fields contained
in the distribution, that is, V ∈ H if V (p) ∈ Hp for all p. Assume that the
Lie sub-algebra of vector fields on Ω generated by H is the whole algebra of
vector fields. Then every two points in Ω can be connected by an admissible
path. Note that if iterated Lie brackets of vector fields from H generate all
vector fields, then it is enough to use at most n = dim(Ω) iterations.
Exercise 5.4.8. In each of the following examples check that the assump-
tions of Chow’s Theorem are satisfied, and hence the corresponding Carnot–
Carathéodory metrics are finite.
(1) V1 = (1, 0), V2 = (y, x) on R2 in Cartesian coordinates (x, y). Can
you sketch a small sub-Riemannian ball centered at (0, 0)?
(2) V1 = (1, 0, 0), V2 = (0, 1, x) on R3 in Cartesian coordinates (x, y, z),
(3) V1 = (0, 0, 1), V2 = (sin α, − cos α, 0) on a three-dimensional cylinder
R2 ×S 1 with coordinates (x, y, α) (the distribution of the admissible motions
of a bicycle in the example above),
(4) V1 ,V2 are any two linearly-independent left invariant fields on SO(3).
Exercise 5.4.9. Prove the general Chow–Rashevsky Theorem.

Ball-Box Theorem. Now we want to get some insight into infinitesimal be-
havior of a sub-Riemannian metric generated by a distribution Hp in Ω. We
will restrict ourselves to the case dim Ω = 3; unlike Chow’s Theorem, whose
general version does not involve new ideas and requires only more tedious
and cumbersome notations, even the formulation of the Ball-Box Theorem
5.4. Sub-Riemannian Metric Structures 185

in higher dimensions relies on a more delicate technique (of privileged coor-

dinates). However, already the three-dimensional Ball-Box Theorem gives
a pretty good geometric idea of what is going on.
Denote by Prc the parallelepiped (“box”)
[−cr, cr] × [−cr, cr] × [−cr2 , cr2 ] ⊂ R3 .
Theorem 5.4.10. Let V1 , V2 be two smooth vector fields in a neighborhood
of the origin in R3 , and let Hp be the corresponding distribution of two-
dimensional planes spanned by V1 , V2 ; assume that H(0,0,0) is the xy-plane
z = 0. Suppose that the assumptions of Chow’s Theorem are satisfied. Let
d be the sub-Riemannian metric induced by the distribution H. Denote by
Br the d-ball of radius r centered at the origin. Then there are constants
0 < c < C such that, for all sufficiently small r, one has
Prc ⊂ Br ⊂ PrC .
Hence in the xy-directions Br looks approximately like a Euclidean ball,
whereas in the z-direction it is very thin (the length of its intersection with
the z-axis is of order r2 ). One can think of this ball as approximately a
Euclidean ball squeezed in r times in the z-direction. Note, however, that it
is not known even whether it is a topological ball for all sufficiently small r!

Proof. I. First we prove that there is a positive c such that Prc ⊂ Br for all
sufficiently small r. To prove this inclusion, it suffices to show that every
point from Prc can be connected with the origin by an admissible path of
length at most r. Notice that W = [V1 , V2 ](0, 0, 0) does not lie in the xy-
plane (by Chow’s condition).
We will use the map F (for p = (0, 0, 0)) defined in (5.17) and an
admissible curve γ connecting (0, 0, 0) and F (u, v, τ ) constructed in the proof
of the Chow–Rashevsky Theorem 5.4.6. Let us estimate the length of γ. Let
M be such that |V1 (p)| ≤ M , |V2 (p)| ≤ M for all p in a neighborhood p of
the origin. Then the length of γ is bounded by M |u| + M |v| + 4M |τ |.
Indeed, the velocity
p of γ pfor the p intervals p t ∈ [0, 1], t ∈ [1, 2], t ∈ [2, 3],
t ∈ [3, 4] is − |τ |V2 , − |τ |V1p, |τ |V1 , |τ |V2 respectively. Hence its
speed for t ≤ 4 is bounded by |τ |M , and thus the length of γ|[0,4] is at
p
most 4M |τ |. Analogously for t ∈ [4, 5] and t ∈ [5, 6], the velocity of γ
is vV1 and uV2 respectively, and hence the length of these segments of γ
is at most M |u| + M |v|. Thus we conclude that the Carnot–Carathéodory
p
distance from the origin to F (u, v, τ ) is at most M (|u| + |v| + 4 |τ |). Thus
it remains to show that (there exists a positive c such that) c
p Pr is contained
in the image under F of the set {(u, v, τ ) : |u| + |v| + 4 |τ | ≤ r/M }. This
image in its turn contains the image of the rectangle
{(u, v, τ ) : |u| ≤ δr, |v| ≤ δr, |τ | ≤ δ 2 r2 },
186 5. Smooth Length Structures

where δ = 1/6M .
Now all is left is to recall that
∂ ∂ ∂
F (0, 0, 0) = V1 (0, 0, 0), F (0, 0, 0) = V2 (0, 0, 0), F (p) = W.
∂u ∂v ∂τ
Since V1 (0, 0, 0), V2 (0, 0, 0), W are linearly independent, there is a constant
c > 0 such that
¡ ¢
Prc ⊂ F [−δr, δr] × [−δr, δr] × [−δ 2 r2 , δ 2 r2 ] ,
provided that r is small enough. This proves part I.

II. Now we need to show that there is a positive C such that Br ⊂ PrC
for all sufficiently small r; in other words, we need to estimate Carnot–
Carathéodory distance from below. Let γ be an admissible path connecting
the origin and (u, v, τ ). We want to estimate from below the (Euclidean)
length L of γ. It is enough to show that
p
(5.18) CL ≥ max(|u|, |v|, |τ |).

Let us choose (and fix) a nonvanishing differential 1-form ω in a neigh-

borhood of the origin such that ω = dz at the origin, and ω(V1 ) = ω(V2 ) = 0;
i.e., the kernel of ω at every point p coincides with Hp = span(V1 (p), V2 (p)).
Let ω = dz + ν. Then |νp | ≤ A|p| for some A since ν vanishes at the origin.
Choose M such that ||dω|| ≤ M in a neighborhood of the origin.
Note that L is obviously greater than the Euclidean distance from the
√ to (u, v, τ ). Hence 2L ≥ (|u| + |v|). Thus, for C, say, greater than
origin
100 p 1 + A, the
√inequality (5.18) is automatically satisfied for points (u, v, τ )
with |τ | ≥ 2 1 + A (|u| + |v|). We willpchoose C > 100(1 + A), and hence
√
it is enough to handle the points with |τ | ≥ 2 1 + A (|u| + |v|), that is,
with |τ | ≥ 4(1 + A)(|u| + |v|)2 ≥ 2(1 + A)(u2 + v 2 ).
Form a cycle α by completing γ to a closed curve by connecting (u, v, τ )
with the origin by a segment σ. Then
Z Z Z
ω = ω + ω.
α γ σ
R
Notice that γ ω = 0 since γ′
is a linear combination of V1 and V2 , and
′
therefore ω(γ ) = 0. Thus
¯Z ¯ ¯Z ¯ ¯Z ¯ Z
¯ ¯ ¯ ¯ ¯ ¯
¯ ω ¯ = ¯ ω ¯ = ¯ (dz + ν)¯ = | − τ + ν| ≥ |τ | − A(u2 + v 2 + τ 2 ).
¯ ¯ ¯ ¯ ¯ ¯
α σ σ σ

Recalling that we need to deal only with points with |τ | ≥ 2(1 + A)(u2 + v 2 ),
we see that for such points
¯Z ¯
¯ ¯
(5.19) ¯ ω ¯ ≥ 1 |τ | − Aτ 2 ≥ 1 |τ |
¯ ¯ 2 3
α
5.4. Sub-Riemannian Metric Structures 187

for all sufficiently small |τ |.

Now we will make use of an isoperimetric inequality in R3 ; indeed, the
length of the cycle α is at most 2L, and hence it can be represented as
the boundary of a two-chain β whose area is at most (1/π)L2 . (1/π is
a sharp constant, and the proof of this statement may be tricky; for our
argument a constant, say, 10, would do as well, and the corresponding crude
isoperimetric inequality is a simple exercise.) Thus choose a β with ∂β = α
and ||β|| ≤ 10L2 . By Stokes’ formula,
Z Z
| ω| = | dω| ≤ M ||β|| ≤ 10M L2 .
α β

Together with (5.19), this

p yields √ L2 ≥ |τ |/(30M ). Recall that we are
dealing
p with points with |τ | ≥ 2 1 + A (|u| + |v|), and in particular
|τ | ≥ max(|u|, |v|). Thus we get that
√ p p
30M L ≥ |τ | = max(|u|, |v|, |τ |),
√
which proves the theorem (for C = max( 30M , 100(1 + A))). ¤
Exercise 5.4.11. Find the Hausdorff dimension of a (finite) Carnot–
Carathéodory metric induced by a two-dimensional distribution in R3 .
Answer: 4.
188 5. Smooth Length Structures

5.4.3. Model example: Connections in fiber bundles.

Model example. We complete our discussion of sub-Riemannian length

structures by looking at a model example. This example will allow us to
see many important features of Carnot–Carathéodory spaces, as well as to
demonstrate important connections between sub-Riemannian geometry and
other concepts.
The example is given by the following distribution Hp of 2-planes in
R3 : H(x,y,x) = span(V (x, y, z), W (x, y, z)), where V (x, y, z) = (1, 0, 0),
W (x, y, z) = (0, 1, x). This distribution can be regarded as a linearized
approximation to the one describing the admissible motions of a bicycle.
The Lie bracket of V and W is [V, W ] ≡ (0, 0, 1). Indeed,
∂f ∂f ∂f ∂f
Vf = , Wf = +x , V Wf − V Wf = = (0, 0, 1)f.
∂x ∂y ∂z ∂z
Hence the conditions of the Chow–Rashevsky Theorem are satisfied: every
two points can be connected by an admissible curve. For instance, for any
choice of a finite length structure on smooth curves in R3 , Hp induces a finite
sub-Riemannian length structure. We could use the usual Euclidean length
of curves, but it is more convenient to define the length of an admissible
curve γ(t) = (x(t), y(t), z(t)) by
Z p
(5.20) L(γ) = x′ 2 (t) + y ′ 2 (t) dt.

Note that admissible curves cannot go vertically, and hence the integrand
in this formula never vanishes.
We will look at this example in various clothing, which will give different
insights into the geometry of sub-Riemannian length structures.

Connections in fiber bundles. Let us think of R3 as B × F , where the

base B = R2 is the xy-plane and the fiber F = R is the z-axis. Denote the
natural projection to the B-factor (base) by π. Note that we have chosen
our length structure in such a way that this projection preserves the length
of admissible curves; this is a reason why we preferred the length structure
given by (5.20) to the Euclidean one.
The results of this section apply (after proper modifications) to a general
distribution on a fiber bundle; we use the existence of a group of translations
(along the z-axis) that preserve the distribution, and the fact that (the
derivative of) the projection restricted to each plane of the distribution
is bijective. Restricting ourselves to one-dimensional fibers allows us to
avoid vector-valued forms, but apart from that the more general case is not
essentially more difficult.
5.4. Sub-Riemannian Metric Structures 189

The most important feature of this situation is that every smooth curve
in the base can be lifted to R3 as an admissible path. Once the lift of
one point is fixed, the lifting of the whole curve is unique. Indeed, the
restriction of the differential dp π of the projection to the base to Hp is a
linear isomorphism; hence for a vector field X tangent to the base B, there
is a unique lift of X, that is, a vector field X̃ that lies in the distribution
H and dπ(X̃) = X. From here, it is easy to see that given a smooth path
γ : R → B, and a point p with π(p) = γ(0), there exists a unique lift of
γ, that is, an admissible curve γ̃ with γ̃(0) = p and π(γ̃) = γ. Of course,
all admissible curves come this way. Thus if one wants to connect two
points p, q ∈ R3 by an admissible curves, one can present it by drawing
its projection connecting the projections p0 , q0 of the points to the base.
Notice, however, that if one connects p0 and q0 by some path γ : [0, 1] → B,
its lift γ̃ with γ̃(0) = p will connect p with some point in the fiber containing
q, γ̃(1) ∈ q0 × F , though there is no reason to expect that it will be just q.
This suggests the following simple but useful definition.

Holonomy group. For a path γ : [0, 1] → B with π(γ(0)) = p, π(γ(1)) = q

define by Gγ : F × {p} → F × {q} the following transformation: Gγ (a) = b
if γ̃(1) = b for the lift γ̃ of γ with γ̃(0) = a.
Exercise 5.4.12. Show that each Gγ is a diffeomorphism.
Exercise 5.4.13. Show that, for a fixed p ∈ B, the collection of Gγ for
the loops γ with γ(0) = γ(1) = p forms a group of diffeomorphisms of
F = F × {p}, and concatenation of paths corresponds to composition of
transformations. This group is called the holonomy group at p.

Note that R acts on R3 by translations in the z-direction: ga (x, y, z) =

(x, y, z + a). Both the distribution H and a lift X̃ are invariant under this
action.
Exercise 5.4.14. Verify that for our distribution every Gγ is just a trans-
lation.

Holonomy groups at different points p and q are isomorphic, and every

path γ between p and q gives rise to an isomorphism: for a loop σ at p, we see
that Gσ → G−1γ ◦ Gσ ◦ Gγ comes from the loop formed by traversing γ, then
σ and then again γ in the reversed direction. This construction actually
establishes a conjugacy between actions of holonomy groups at different
points.
Now the problem of finding an admissible path connecting two points p1
and q1 reduces to finding a path γ connecting their projections p = π(p1 )
and q = π(q1 ) with Gγ (p1 ) = q1 . In particular, every two points can be
190 5. Smooth Length Structures

connected if the holonomy group at a point (and hence at every point) acts
transitively (verify this assertion!)

Curvature form. The following construction plays a crucial role in under-

standing holonomy groups.
Given a vector X ∈ T(x,y,z) R3 , it can be uniquely represented by XH +
XV , where XH ∈ H(x,y,x) , and pr(XV ) = 0. The vectors XH and XV can be
called the horizontal and the vertical components of X; XV is tangential to
the copy of the fiber F (a vertical line) passing through (x, y, z). Of course,
a curve is admissible if the vertical component of its velocity is identically
zero.
Now let us consider two vector fields X, Y on the base, and lift them to
X̃, Ỹ . Denote by ω(X, Y ) the vertical component of the Lie bracket [X̃, Ỹ ].
ω(X, Y ) is a vector field tangent to the fibers; since in our case all fibers are
identified with R, ω(X, Y ) can be regarded as just a real-valued function.
Exercise 5.4.15. Prove that ω(X, Y )(p) depends only on the vectors
X(π(p)) and Y (π(p)). Show that ω is a two-form on the base.
Hint: Note that the vertical component of X is nothing but hX, Zi,
where Z = (0, 0, 1).

This may be somewhat surprising, for the Lie bracket used to define Z
involves derivatives; this is similar to the situation with the curvature tensor
(see Chapter 6, 6.3.3), which is defined using vector fields, but at the end it
happens to depend only on their values at one point. As a matter of fact,
the curvature tensor is a particular case of this construction.
The geometric meaning of the curvature form ω is given by the following
important exercise:
Exercise 5.4.16. Let γ be a loop at p: γ(0) = γ(1) = p. Assume that γ
bounds a simple region Γ, whose orientation agrees with that of γ. Then
the transformation Gγ is a translation z → z + A, where
Z
A= ω.
Γ

Hint: There is a straightforward hand-crafted argument, which is very

good to reveal the geometric meaning of the construction. One considers
two commuting vector fields X and Y , and first proves the statement for
an infinitesimal rectangle formed by integral curves of the fields just from
the definition of the Lie bracket. Then one replace the projection of γ by
a piecewise path consisting of integral curves of X or Y . Another proof,
which is much shorter but not that visual, can be conducted analogously to
the reasoning in the section “Contact Viewpoint” below.
5.4. Sub-Riemannian Metric Structures 191

Shortest path and isoperimetric problem. Finally, let us look at our

particular distribution V (x, y, z) = (1, 0, 0), W (x, y, z) = (0, 1, x). To apply
ω to X = (1, 0), Y = (0, 1), we note that their lifts are just V and W ; there
Lie bracket [V, W ] = (0, 0, 1) has been computed at the beginning of this
section; hence ω(X, Y ) = 1, and therefore ω is just the standard area form.
This allows us to explicitly find the shortest paths of this sub-Riemannian
structure. For instance, let us connect p = (0, 0, 0) and q = (0, 0, z) by
a shortest path. The projection of any curve connecting these points is a
loop at (0, 0) enclosing oriented area z, and the length of this projection is
the same as the length of the curve (by (5.20)). Hence the shortest path
between the points is a lift of the shortest closed curve enclosing area z.
This is an isoperimetric problem, and it is known that the shortest curve
enclosing a given area is a circle. Hence shortest paths connecting p and q
√
are lifts of circles of a radius π −1 z passing through the origin. Note that,
unlike any Riemannian metric, there are infinitely many different shortest
paths between p and q for every value of z.

Exercise 5.4.17. Show that every shortest path in this sub-Riemannian

structure is a lift of a circular arc. Note that shortest paths may split: there
are infinitely many shortest paths emanating from every point in a given
(admissible) direction.

Exercise 5.4.18. Using the description of shortest paths, give an explicit

formula for the sub-Riemannian distance function, and explicitly verify the
conclusion of the Ball-Box Theorem.

Contact viewpoint. The distribution Hp can be regarded as a contact

structure: it is the distribution of kernels of a contact 1-form ω1 = xdy − dz.
Let us use an argument analogous to the one exploited to obtain a lower
bound in the Ball-Box Theorem. As a model example, let us again consider
an admissible path γ connecting two points p = (0, 0, 0) and q = (0, 0, z).
We can form a 1-cycle α by closing γ by a segment qp. Let β be a (smooth)
two-chain whose boundary is α. By Stokes’ Formula,
Z Z
ω1 = dω1 .
α β

The exterior differential of ω1 is dω1 = ω2 = dx ∧ dy, whichRis nothing but

the oriented area of the projection to the xy-plane. Hence α ω1 measures
the oriented area enclosed in the projection of γ (for the segment qp projects
to a point). On the other hand,
Z Z Z
ω1 = ω1 + ω1 = 0 + z.
α γ qp
192 5. Smooth Length Structures

Hence we obtain the same conclusion as at the end of the previous section:
the shortest curve connecting p and q projects to the shortest curve bounding
a region of area z.

Heisenberg group. Let us introduce the following group structure on R3 :

(x, y, z) · (x1 , y1 , z1 ) = (x + x1 , y + y1 , z + z1 + xy1 ).
This is “almost” the usual addition, with a “twisting term” xy1 in the z-
component of the result. Note H is a left-invariant distribution with respect
to this group structure. Indeed, a left translation L(x0 ,y0 ,z0 ) that sends
(0, 0, 0) to a point (x0 , y0 , z0 ) is a multiplication by (x0 , y0 , z0 ):
L(x0 ,y0 ,z0 ) (x, y, z) = (x + x0 , y + y0 , z + z0 + x0 y),
and its differential dL(x0 ,y0 ,z0 ) acts by L(x0 ,y0 ,z0 ) (x, y, z) = (x, y, z + x0 y).
Thus
dL(x0 ,y0 ,z0 ) (1, 0, 0) = (1, 0, 0) = V (x0 , y0 , z0 ),
dL(x0 ,y0 ,z0 ) (0, 1, 0) = (0, 1, x0 ) = W (x0 , y0 , z0 ).
Moreover, the length structure given by (5.20) is also left-invariant (verify
this!) Thus Hp determines a left-invariant Carnot–Carathéodory metric dH
on the Heisenberg group.
It is interesting to compare this metric with a left-invariant Riemannian
metric d whose quadratic form at the origin (0, 0, 0) coincides with the
standard coordinate Euclidean structure.
Exercise 5.4.19. Show that d and dH are asymptotically the same:
d(p, q)
lim = 1.
d(p,q)→∞ dH (p, q)
(As a matter of fact, the difference between d and dH actually is
uniformly bounded; i.e., there exists a constant C such that |d(p, q) −
dH (p, q)| ≤ C for all p, q).

This provides us with valuable information about large-scale structure

of left-invariant metrics on the Heisenberg group, for we have an explicit pde-
scription of dH . In particular, d((0, 0, 0), (0, 0, z)) is approximately π −1 |z|.
We see that the z-axis does not look at all like a geodesic with respect to
d—the distance between two far-away points in the z-axis is much smaller
than the length of its segment.
Let us note that dH possesses an additional symmetry as compared to d.
Consider a transformation At : R3 → R3 (t is a positive parameter) acting
by At (x, y, x) = (tx, ty, t2 z). It follows from the following exercise that At
is a homothety (dilation) with respect to dH , that is, dH (At (p), At (q)) =
tdH (p, q).
5.5. Riemannian and Finsler Volumes 193

Exercise 5.4.20. Check that At leaves the distribution Hp invariant, and

it multiplies the length of curves given in (5.20) by t.

This implies that the (metric) tangent cone of (R3 , dH ) is isometric to

R3 , dH itself, and it in its turn is also isometric to its asymptotic cone (the
cone at infinity). Using Exercise 5.4.19, we conclude that the asymptotic
cone of the Heisenberg group with a left-invariant metric d is isometric to
(R3 , dH ).
Information: The tangent cone of a Carnot–Carathéodory space at
a generic point is isometric to a nilpotent group with a left-invariant sub-
Riemannian metric; at “nongeneric” points it is isometric to a homogeneous
space of a nilpotent group.

5.5. Riemannian and Finsler Volumes

In this section we take a more detailed and formal look at the notion of
volume in Riemannian manifolds. We have already defined the area of a two-
dimensional Riemannian region by an explicit formula (5.4). The motivation
for this formula remains valid in any dimension. We will turn the motivation
for this formula into a formal argument showing that the volume must equal
a similar integral expression provided that it depends monotonically on the
metric.
In other words, only one “reasonable” notion of Riemannian volume
exists—no matter how one defines it, the result is the same. This is no
longer true for Finsler metrics; as a result, there exist many different notions
of Finsler volume. These issues are discussed in subsection 5.5.3.
We fix a dimension n ≥ 1; all vector spaces and manifolds in this section
are implicitly assumed to be n-dimensional.

5.5.1. Riemannian volume and Jacobians.

Definition 5.5.1. The Riemannian volume in an n-dimensional Riemann-

ian manifold is the n-dimensional Hausdorff measure (cf. Section 1.7) deter-
mined by its Riemannian metric.

We denote the Riemannian volume by Vol. In case of ambiguity we add

the manifold or Riemannian structure as an index (e.g., VolM ).
In fact, we need only two properties of Riemannian volume:
(1) Riemannian volume in Rn is the standard Euclidean volume (the
Lebesgue measure).
194 5. Smooth Length Structures

(2) Volume is monotone with respect to the metric. That is, if M

and N are Riemannian manifolds and f : M → N is a distance-
nonexpanding diffeomorphism, then VolN (f (Ω)) ≤ VolM (Ω) for
any measurable set Ω ⊂ M .
We will show that the Riemannian volume is uniquely determined by these
properties. The purpose of the above definition was to make sure that such
a volume functional exists.
According to the definition, the Riemannian volume is a Borel measure
over a Riemannian manifold. Then one can use Lebesgue integration: if
M is a Riemannian manifold and f : Ω → R is a nonnegative measurable
(or, moreR generally, integrable) function on a measurable set Ω ⊂ M , the
integral Ω f d VolM is defined. We will use only elementary and natural
properties of integration such as being linear and monotone with respect to
the function being integrated. We refer to Chapter 2 of [Fe] for the general
theory of measure and integration.
Jacobians. Let M and N be Riemannian manifolds of the same dimension
n. Let f : M → N be a map differentiable at an x ∈ M . The derivative
dx f is a linear map from Tx M to Tf (x) N . Riemannian structures of M
and N define Euclidean scalar products in Tx M and Tf (x) N making these
two tangent spaces isometric to Rn . In particular, Tx M and Tf (x) N are
naturally equipped with n-dimensional Lebesgue (or Hausdorff) measure.
Since every linear map from Rn to Rn multiplies the volumes by a constant
depending only on the map (see Exercise 1.7.6), so does the linear map
dx f : Tx M → Tf (x) N . The respective constant is called the Jacobian:
Definition 5.5.2. Let M , N , f and x be as above. The Jacobian of f at x,
denoted Jac f (x), is a real number such that
µn (dx f (X)) = Jac f (x) · µn (X)
for all measurable sets X ⊂ Tx M .

If Tx M and Tf (x) N are equipped with orthonormal bases and A is the

matrix of dx f in these bases, then Jac f (x) = | det A| (cf. Exercise 1.7.6).
Remark 5.5.3. Our definition implies that the Jacobian is always non-
negative. There is a slightly different notion of Jacobian which is a signed
quantity and equals the determinant of the corresponding matrix (not its
absolute value). This is suitable for maps from Rn to Rn but the sign makes
little sense for Jacobians in general (possibly nonorientable) Riemannian
manifolds. We will never use signed Jacobians.

Jacobians determine how a map changes the Riemannian volume. The

respective formula is known in analysis as the change-of-variable formula.
5.5. Riemannian and Finsler Volumes 195

Theorem 5.5.4 (change of variable formula). If M and N are Riemannian

manifolds, f : M → N is a diffeomorphism, and Ω ⊂ M is a measurable
set, then
Z
Vol(f (Ω)) = Jac f d VolM .
Ω
R
In particular, Vol(N ) = M Jac f .

Proof. The formula is based on the following local statement: given x ∈ M ,

then for (small) measurable sets X containing x one has
VolN (f (X))
(5.21) → Jac f (x) as diam(X) → 0.
VolM (X)
To prove (5.21), fix a sufficiently small neighborhood U of x and a map
ϕ : U → Tx M which is a diffeomorphisms from U to a neighborhood of 0
in Tx M such that ϕ(x) = 0 and dx ϕ : Tx M → Tx M is the identity map. (For
example, let ϕ be the inverse of expx .) This map introduces local coordinates
in U . Compare the Riemannian structure of M written in these coordinates
with the Euclidean structure in Tx M . They coincide at the origin; hence
their corresponding metrics are equal up to the first order near the origin.
Or, equivalently, ϕ and ϕ−1 restricted to sufficiently small neighborhoods of
0 and x have Lipschitz constants arbitrarily close to 1. Since the volume is
monotone with respect to the metric, it follows that
VolTx M (ϕ(X))
→1
VolM (X)
as diam(X) → 0. Now let y = f (x), V = f (U ) and a map ψ : V → Ty N
be defined by the formula ψ ◦ f |U = dx f ◦ ϕ|U . Then ψ(y) = 0 and
dy ψ : Ty N → Ty N is the identity. Similarly to the above, we obtain that
VolTy N (ψ(f (X)))
→1
VolN (f (X))
as diam(X) → 0. Since ψ(f (X)) = dx f (ϕ(X)), the desired formula (5.21)
now follows from the relation
VolTy N (dx f (ϕ(X)))
= Jac f (x)
VolTx M (ϕ(X))
which is nothing but the definition of Jacobian.
How to derive the theorem from (5.21) depends on the reader’s back-
ground. The most elementary method is to split X into small subsets
{X
P i }, choose points xi ∈ Xi , and compare
P the volume VolN (f (X)) =
VolN (f (Xi )) with the integral sum Jac f (xi ) VolM (Xi ). The details
are left as an exercise. ¤
196 5. Smooth Length Structures

As a corollary to Theorem 5.5.4 we obtain the coordinate formula for

the Riemannian volume which is often taken as the definition.
Theorem 5.5.5. Let M be an n-dimensional Riemannian manifold, U an
open set in Rn and ϕ : U → M a coordinate system. For an x ∈ U ,
let (gij (x)) be the matrix composed of the scalar products of the coordinate
tangent vectors at ϕ(x), that is, gij (x) = hdx ϕ(ei ), dx ϕ(ej )iM where {ei } is
the standard basis of Rn . Then
Z q
VolM (ϕ(Ω)) = det(gij ) dmn
Ω
for any measurable Ω ⊂ U .

Proof. Consider U as a Riemannian manifold whoseR scalar product is

standard Euclidean. By Theorem 5.5.4, VolM (ϕ(U )) = U Jac ϕ. It remains
p
to prove that Jac ϕ(x) = det(gij (x)) for every x ∈ U . Let (vi ) be an
orthonormal basis of Tϕ(x) M and A be the matrix of dx ϕ : Rn → Tϕ(x) M
with respect to this basis. Then Jac ϕ(x) = | det A|. On the other hand,
(gij (x)) = AT A where AT is the transpose matrix.
p Therefore det(gij (x)) =
det(AT ) det(A) = det(A)2 ; hence | det A| = det(gij (x)). The theorem
follows. ¤

Note that Theorem 5.5.5 uniquely defines the volume of any measurable
set in a Riemannian manifold. On the other hand, in its proof (including
the proof of the change of variable formula) we only used the two basic
properties mentioned after Definition 5.5.1. Hence these two properties
uniquely determine the volume, and we obtain the following
Theorem 5.5.6. Let V be a function associating to every Riemannian
manifold M a Borel measure over it. Suppose that V is monotone with
respect to the metric and yields the standard Euclidean volume in Rn . Then
V coincides with the Riemannian volume.

5.5.2. Volume of Lipschitz maps. Though we proved the change of vari-

able formula only for diffeomorphisms, it works for arbitrary Lipschitz home-
omorphisms. Moreover, a simple modification makes it valid for Lipschitz
maps that are not bijective. Below we give the respective formulations with-
out proofs.
Theorem 5.5.7 (Rademacher’s theorem; see Theorem 3.1.6 in [Fe]). Every
Lipschitz map is differentiable almost everywhere. That is, if M and N
are Riemannian manifolds and f : M → N is a Lipschitz map, then the
derivative dx f exists for all x ∈ M except a set of zero measure.

Note that the word “Riemannian” in this theorem can be replaced by

“smooth” because the class of Lipschitz maps and the class of sets of zero
5.5. Riemannian and Finsler Volumes 197

measure does not depend on the choice of Riemannian structure (in fact,
both notions can be defined in terms of local coordinates).
Rademacher’s theorem implies that for a Lipschitz map f : M → N the
JacobianR Jac f (x) is defined for almost all x ∈ M . Therefore the Lebesgue
integral M Jac f is defined. This integral is called the volume (or area) of f .
If f is injective, it equals the volume of f (M ) in N . If f covers some parts
of N multiple times, the multiplicity should be taken into account:
Theorem 5.5.8 ([Fe], Theorem 3.2.3). If M and N are n-dimensional
Riemannian manifolds and f : M → N is a Lipschitz map, then
Z Z
Jac f (x) d VolM (x) = #(f −1 (y)) d VolN (y)
M N
where # denotes the cardinality.

The integral on the right-hand side is, of course, the Lebesgue integral of
the function y 7→ #(f −1 (y)) which is always measurable. Since the function
takes only integer values and the value infinity, its integral can be written
as
Z X
#(f −1 (y)) d VolM (y) = k · VolN ({y ∈ N : #(f −1 (y)) = k})
N k∈N∪{∞}

where the term ∞ · 0, if it appears, equals 0.

One of the most common applications of the above theorem is the
following upper bound for the volume of a map’s image.
Corollary 5.5.9. If M and N are n-dimensional Riemannian manifolds
and f : M → N is a Lipschitz map, then
Z
VolN (f (M )) ≤ Jac f (x) d VolM (x).
M
In particular, if Jac f ≤ 1 almost everywhere, then VolN (f (M )) ≤ Vol(M ).

Proof. #(f −1 (y)) ≥ 1 if y ∈ f (M ). Substitute this into Theorem 5.5.8. ¤

The next proposition is used in Section 5.6.

Proposition 5.5.10. Let M be an n-dimensional Riemannian manifold,
f : M → Rn be a Lipschitz
Q map and fi (1 ≤ i ≤ n) be its coordinate
functions. Then Jac f ≤ mi=1 dil(fi ) wherever f is differentiable.

Proof. Let f be differentiable at a point x ∈ M . It is easy to see

that |dx fi (v)| ≤ dil(fi ) · |v| for all v ∈ Tx M , i.e., kdx fi k ≤ dil(fi ). Let
A = (aij ) be a matrix of dx f with respect to an orthonormal basis of Tx M
and the standard basis of Rn . The ith row of this matrix is composed
of the numbers dx fi (vj ), 1 ≤ j ≤ n. Let us interpret this row as a
198 5. Smooth Length Structures

vector ai ∈ Rn ; then |ai | = kdx fi k ≤ dil(fi ). Since | det A| equals the

Euclidean volume of Q the parallelotope
Qn spanned by the vectors ai , we have
n
Jac f (x) = | det A| ≤ i=1 |ai | ≤ i=1 dil(fi ). ¤

5.5.3. Finslerian volumes. Theorem 5.5.6 tells us that there is only one
reasonable notion of volume for Riemannian manifolds. This is not the case
with Finsler metrics. One can define Finslerian volume in different ways
and obtain essentially different results. Some examples are given below in
this section.
For now, we assume that some Finslerian volume functional is fixed.
That is, every Finsler manifold is equipped with a Borel measure (depending
on its Finsler structure) that we denote by Vol. We require that Vol satisfy
the properties listed after Definition 5.5.1, namely, the Euclidean compat-
ibility and monotonicity with respect to metric. Note that monotonicity
implies that (smooth) isometries preserve the volume.
Let k · k be a norm in Rn . As a particular case of a Finsler manifold, the
normed space (Rn , k · k) carries a Finslerian volume, Volk·k . Let | · | be the
Euclidean norm. Since k · k and | · | are bi-Lipschitz equivalent (Theorem
1.4.11), there exist positive constants c and C such that c|x| ≤ kxk ≤ C|x|
for all x ∈ Rn . The monotonicity of volume then implies that cn mn ≤
Volk·k ≤ C n mn . (To prove this, consider the identity map as a map from
(Rn , k · k) to (Rn , c| · |) and from (Rn , C| · |) to (Rn , k · k).) In particular, the
k · k-volume of a unit cube is finite and positive. We denote this quantity
Volk·k ([0, 1]n ) by ν(k · k).
The volume Volk·k is preserved by parallel translations since they are
isometries of (Rn , k · k). By Lebesgue’s theorem (1.7.5) it follows that the
Volk·k
measure ν(k·k) coincides with the Lebesgue measure mn . Thus

Volk·k (X) = ν(k · k) mn (X)

for any measurable set X ⊂ Rn .

Recall that a Finsler structure Φ in a region U ⊂ Rn is a continuous
function on T U whose restriction to every tangent space Tx U (where x ∈ U )
is a vector-space norm. We denote this restriction by Φx .

Proposition 5.5.11. If Φ is a Finsler structure in a region U ⊂ Rn , then

Z
VolΦ (Ω) = ν(Φx ) dmn (x)
Ω

for any measurable set Ω ⊂ U .

Proof. Similar to Theorems 5.5.4 and 5.5.5. ¤

5.5. Riemannian and Finsler Volumes 199

This proposition provides a sort of “general form” of a Finslerian volume.

One can define a Finslerian volume functional by specifying a value ν(k · k)
for every norm k · k in Rn . In other words, an n-dimensional Finslerian
volume is determined by its values on flat normed spaces. More precisely,
the following proposition holds.
Proposition 5.5.12. Suppose that every normed space (V, k · k) is equipped
with a translation-invariant measure Vol(V,k·k) which is finite and positive
on open bounded sets, so that the following conditions are satisfied.
(1) Euclidean compatibility: Vol(Rn ,standard Euclidean norm) = mn .
(2) Affine invariance: any linear isometry between normed spaces pre-
serves the measure.
(3) Monotonicity: if k · k ≤ k · k′ , then Vol(V,k·k) ≤ Vol(V,k·k′ ) .
Then this family of measures can be extended to all Finsler manifolds as a
Euclidean-compatible and monotone volume functional.

A plan of proof. Having a measure over every normed space allows one
to define the Jacobian of a map from one Finsler manifold to another,
similarly to Definition 5.5.2. Define Finslerian volume by the formula from
Proposition 5.5.11; then the change of variable formula (Theorem 5.5.4)
follows easily. The conditions 2 and 3 imply that a linear nonexpanding map
does not increase the measure. Observe that derivatives of a nonexpanding
maps between Finsler manifolds are nonexpanding linear maps between
their tangent spaces (equipped with norms restricted from the Finslerian
structures). Then the change of variable formula implies that the Finslerian
volume is well-defined (i.e., does not depend on the choice of a coordinate
system) and monotone with respect to metric. ¤

In order to define a translation-invariant measure over a vector space, it

is sufficient to specify its value on a single set. Above we used the unit cube
in Rn for this purpose (recall the definition of ν(k · k)). In most cases it is
more convenient to fix a volume of a set which is naturally associated with
the norm, for example, the norm’s unit ball.
Example 5.5.13 (Hausdorff measure). Let αn denote the Euclidean volume
of the standard unit ball in Rn . For every normed space (V, k · k) define a
measure Vol(V,k·k) so that the measure of the norm’s unit ball equals αn .
Equivalently, define
αn
ν(k · k) =
mn (unit ball of k · k)
for a norm k · k in Rn . It is easy to see that the conditions of Proposition
5.5.12 are satisfied, so we obtain some Finsler volume functional. In fact,
200 5. Smooth Length Structures

this functional coincides with the n-dimensional Hausdorff measure. The

proof of this coincidence is similar to determining the precise value of the
normalization constant of the Hausdorff measure (see Theorem 1.7.14 and
its corollary).

The definition of volume in the following example requires choosing a set

of minimal volume from a certain class of sets in a vector space. This looks
like a vicious circle because the volume itself is not yet defined. However,
one does not need to fix a volume functional to compare volumes of sets.
Fix an arbitrary Euclidean structure and compare the Lebesgue measures;
the result does not depend on that auxiliary structure because all measures
are the same up to a multiplicative constant.

Example 5.5.14 (comass, [Gro1]). Let (V, k · k) be a normed space, B its

unit ball, and Q is an affine cube of minimal volume containing B. (By
affine cube we mean an image of the cube [−1, 1]n under a linear map.)
Then define the volume by the formula Vol(V,k·k) (Q) = 2n . This defines the
Finslerian volume functional called comass.

Example 5.5.15 (inscribed Riemannian volume). Similarly to the above,

let Q be an ellipsoid of maximal volume contained in B. Let the volume
Vol(V,k·k) (Q) equal αn , the Euclidean volume of the standard Euclidean ball.
An equivalent definition is: for a given norm k · k, find a Euclidean norm | · |
whose volume form is minimal among all Euclidean norms which are greater
or equal to k · k; then set Vol(V,k·k) = Vol(V,|·|) . The resulting Finslerian
volume equals the minimal Riemannian volume of a Riemannian metric
which is greater or equal to the given Finsler one.
Similarly, one can consider the maximal Riemannian volume of a Rie-
mannian metric which is less than or equal to a given Finsler one. This
definition of volume is related to the minimal ellipsoid containing the norm’s
unit ball.

Example 5.5.16 (symplectic volume). Let k · k be a norm in Rn and B its

unit ball. A polar set to B is defined by
B ∗ = {x ∈ Rn : hx, yi ≤ 1 for all y ∈ B}.
Define the volume functional by ν(k · k) = mn (B ∗ )/αn where αn is the
Euclidean volume of the Euclidean unit ball (compare Example 5.5.13). The
resulting Finsler volume functional is called the Holmes–Thompson volume.

This particular Finsler volume is related to the so-called symplectic

volume (which is defined over the co-tangent space of the manifold and does
not depend on the metric). Precise formulation: the Holmes–Thompson
5.5. Riemannian and Finsler Volumes 201

volume equals a constant multiplied by the projection of the symplectic

volume from the set of co-vectors having less than unit norm.
Exercise 5.5.17. Prove that all Finsler volume functionals in the above
examples are mutually different.

While there are many different natural Finsler volume functionals, they
are nevertheless “not too different”. Namely the ratio of any two of them is
bounded by a constant depending only on dimension:
Theorem 5.5.18. Let Vol and Vol′ be two n-dimensional Euclidean-compa-
tible monotone Finsler volume functionals. Then Vol(Ω) ≤ n3n/2 Vol′ (Ω) for
any measurable set Ω is any Finsler manifold.

Proof. By Proposition 5.5.11, it is sufficient to prove the inequality Vol ≤

n3n/2 Vol′ in normed vector spaces. We need the following lemma (also
used in Subection 8.5.3). Roughly speaking, it claims that every symmetric
convex body in Rn can be approximated, with a bounded relative precision,
by an affine cube. Recall that an affine cube is an image of the cube [−1, 1]n
under a nondegenerated linear map.
Lemma 5.5.19. Let D be a unit ball of a norm k · k in an n-dimensional
vector space V . Then there exists an affine cube Q ⊂ V such that n1 Q ⊂
D ⊂ Q.

Proof. We may assume that V = Rn . Consider the function F on D × · · · ×

D (n times) defined by F (v1 , . . . , vn ) = | det[v1 , . . . , vn ]| where [v1 , . . . , vn ]
is the n × n matrix composed of the coordinates of the vectors vi . In other
words, F (v1 , . . . , vn ) is the volume of the parallelotope spanned by these
vectors. Since F is continuous and D × · · · × D is compact, F attains a
maximum. We assume that this maximum is achieved on the standard basis
(e1 , . . . , en ) of Rn . This can be achieved by applying a linear transformation
to D. Then one can let Q = [−1, 1]n . Indeed, sinceP D contains P the vectors
ei , one has kei k ≤ 1; hence k(x1 , . . . , xn )k = k xi ei k ≤ |xi | ≤ 1
1 1 n 1
whenever (x1 , . . . , xn ) ∈ [− n , n ] . Therefore n Q ⊂ D. On the other hand,
if v = (x1 , . . . , xn ) ∈ D, then |xi | ≤ 1 for all i; otherwise one could replace
ei by v in the expression F (e1 , . . . , en ) and obtain a greater value. Thus
D ⊂ [−1, 1]n . ¤

Now let (V, k · k) be an n-dimensional normed space. Let Q be an

affine cube from Lemma 5.5.19. Identify V with Rn by means of a linear
isomorphism so that Q = [−1, 1]n . Then the relation n1 Q ⊂ D ⊂ Q implies
√
that n1 B ⊂ D ⊂ nB where B is the Euclidean unit ball in Rn . This is in
turn equivalent to an inequality between norms: √1n | · | ≤ k · k ≤ n| · |, where
202 5. Smooth Length Structures

| · | is the standard Euclidean norm. Hence

1
Vol(Rn ,|·|) ≤ Vol(Rn ,k·k) ≤ nn Vol(Rn ,|·|)
nn/2
by monotonicity of volume (note that Vol(Rn ,c|·|) = cn Vol(Rn ,|·|) for any
constant c > 0). Similarly, the same inequalities hold for Vol′ in place
of Vol. Since Vol(Rn ,|·|) = Vol′(Rn ,|·|) , the desired inequality Vol ≤ n3n/2 Vol′
follows. ¤

5.6. Besikovitch Inequality

The Besikovitch inequality is one of the simplest facts that allows one to
estimate the volume of a Riemannian metric if certain (very limited) infor-
mation about distances is known. It belongs to the “curvature-free” part of
metric geometry in the sense that its formulation does not include curvature
bounds or similar local conditions. Many other results and problems in this
area can be found in [Gro1]. We formulate and prove the Besikovitch in-
equality in subsection 5.6.2. The proof relies on the theory of degree (which
is a part of differential topology) that we briefly discuss in the first sub-
section. While this theory is far from the topic of this book, it provides
important tools commonly used in metric geometry.

5.6.1. Degree of a map. In this subsection we introduce an important

homotopy invariant of a map, the degree. We mainly discuss the degree
modulo 2 which is sufficient for our purposes. For details and proofs see
[Mi]. We recommend the reader prove all the statements below as exercises
in the one-dimensional case (i.e., for circles or intervals).
Definition 5.6.1. Let M and N be smooth manifolds and f : M → N a
smooth map. An x ∈ M is called a regular point for f if the rank of dx f
equals the dimension of N , i.e., dx f : Tx M → Tf (x) N is surjective. A point
y ∈ N is called a regular value of f if every point x ∈ f −1 (y) is a regular
point for f .

The well-known Sard–Brown Theorem says that regular values of f are

dense in N (moreover, the complement of the set of regular values in N is a
set of zero measure).
Note that any point y ∈ N such that f −1 (y) = ∅ is a regular value. If
dim M < dim N the set of regular values coincides with N \ f (M ). The
definition is more interesting if dim M ≥ dim N . In this case the implicit
function theorem implies that, if y is a regular value of f , then f −1 (y) is
a smooth submanifold of M whose dimension equals dim M − dim N . In
particular, in the case dim M = dim N the inverse image of a regular value
5.6. Besikovitch Inequality 203

is discrete (consists of isolated points). If M is compact, the inverse image

of a regular value is a finite set.
The theory of degree applies only to maps between manifolds of equal
dimensions. Below M and N always denote compact smooth manifolds of
dimension n ≥ 1.
Definition 5.6.2. Let f : M → N be a smooth map and y ∈ N be a regular
value of f . Then the quantity deg2 (f ; y) is a residue modulo 2 defined by
deg2 (f ; y) = #(f −1 (y)) mod 2.
In other words deg2 (f ; y) equals 0 if the number of points in f −1 (y) is even,
and 1 otherwise.
Proposition 5.6.3. deg2 (f ; y) does not depend on the choice of a regular
value y ∈ N .

This proposition, along with the fact that at least one regular value
exists, allows us to introduce the following
Definition 5.6.4. Let f : M → N be a smooth map. Then deg2 (f ; y),
where y ∈ N is an arbitrary regular value of f , is called the degree modulo 2
of f and denoted deg2 (f ).
Proposition 5.6.5. If smooth maps f1 and f2 from M to N are homotopic,
then deg2 (f1 ) = deg2 (f2 ).

This proposition allows us to apply the notion of degree to nonsmooth

maps. It can be shown that every continuous map from M to N is homotopic
to a smooth map. If f : M → N is an arbitrary continuous map and f1 is
a smooth map which is homotopic to f , one can define deg2 (f ) = deg2 (f1 ).
Proposition 5.6.5 implies that deg2 (f1 ) does not depend on the choice of f1 ,
so deg2 (f ) is well-defined by this formula. Moreover, deg2 (f ) depends only
on the homotopy class of f .
Manifolds with boundary. In the above considerations we assumed that M
and N have no boundary. This assumption can be dropped if one restricts
the class of maps to ones that map the boundary of M to (a subset of) the
boundary of N .
Here by a manifold with boundary we mean a smooth manifold with a
piecewise-smooth boundary. In fact, the degree can be defined without any
differential structure (i.e., it applies to topological manifolds).
Proposition 5.6.6. Suppose that M and N are compact smooth manifolds,
possibly with boundaries, and f is a smooth map such that f (∂M ) ⊂ ∂N .
Then deg2 (f ; y) does not depend on a regular value y ∈ N . Hence Definition
5.6.4 applies to such a map f .
204 5. Smooth Length Structures

Furthermore, if f0 and f1 are connected via a homotopy {ft }t∈[0,1] such

that ft (∂M ) ⊂ ∂N for all t ∈ [0, 1], then deg2 (f0 ) = deg2 (f1 ).

Proof. We reduce the statements to the case of closed manifolds (Proposi-

tions 5.6.3 and 5.6.5) by means of the doubling construction. Let M̄ be the
double of M , that is, M̄ consists of two copies of M glued along the bound-
ary, and N̄ is the double of N . Then a map f : M → N naturally defines
a map f¯ : M̄ → N̄ which maps each half-manifold of M̄ to the respective
half-manifold of N̄ in the same way as f does. If f (∂M ) ⊂ ∂N , the new
map f¯ is continuous and it is easy to see that deg2 (f ; y) = deg2 (f¯; y ′ ) where
y ′ is one of the two points corresponding to y. Then Propositions 5.6.3
implies the first statement. Similarly, Proposition 5.6.5 implies the second
one. The details (including the issue of introducing a differential structure
on the doubles) are left to the reader. ¤

As we mentioned after Proposition 5.6.5, the homotopy invariance of

deg2 allows us to define deg2 for nonsmooth maps. The same argument
applies to manifolds with boundaries and the class of maps f such that
f (∂M ) ⊂ ∂N .

Proposition 5.6.7. If M and N are compact smooth n-manifolds, possibly

with boundaries, f : M → N is a continuous map such that f (∂M ) ⊂ ∂N
and deg2 f 6= 0, then f is surjective, i.e., f (M ) = N .

Proof. First, reduce the proposition to the case of no boundary just like
Proposition 5.6.6. Then suppose f is not surjective. If f is smooth,
consider a y ∈ N \ f (M ). Since f −1 (y) = ∅, y is a regular value and
deg2 (f ) = deg2 (f ; y) = 0. In the general case, replace f by a smooth
approximation f1 . If f1 is sufficiently close to f , it is homotopic to f and
still nonsurjective; hence deg2 (f ) = deg2 (f1 ) = 0. ¤

Remark 5.6.8. If M and N are oriented manifolds, one can define the
integer-valued degree of a map f : M → N , denoted deg(f ). Namely, if
y ∈ N is a regular value of f , let
X
deg(f ; y) = ε(x)
x∈f −1 (y)

where (
+1, if dx f is orientation-preserving,
ε(x) =
−1, otherwise.
All properties of deg2 that we formulated hold for deg with obvious modifi-
cations. It is clear that deg2 (f ) = deg(f ) mod 2.
5.6. Besikovitch Inequality 205

5.6.2. Besikovitch inequality. We will use the following notation. Let

I = [0, 1]; then I n = [0, 1]n ⊂ Rn is the standard n-dimensional cube.
The boundary ∂I n consists of points where at least one of the coordinates
equals 0 or 1. This boundary is the union of faces that we denote Fi0 and
Fi1 ; namely, Fi0 (resp. Fi1 ) is the set of points in I n whose ith coordinate
equals 0 (resp. 1).
Theorem 5.6.9 (Besikovitch inequality). Let g be a Riemannian structure
in I n . For i = 1, . . . , n let di denote the Riemannian
Q distance in this metric
between the faces Fi0 and Fi1 . Then Volg (I n ) ≥ ni=1 di .

Proof. During this proof all distances, Jacobians, etc on I n are taken with
respect to g. For each i = 1, . . . , n define a function fi : I n → R by
fi (x) = min{di , dist(x, Fi0 )}
and let f : I n → Rn be the function whose coordinate functions are fi , i.e.,
f (x) = (f1 (x), . . . , fn (x)). Observe the following trivial facts.
(1) The functions fi are nonexpanding.
(2) fi (x) ∈ [0, di ] for all x ∈ I n .This means that f maps I n to the
parallelotope P = [0, d1 ] × [0, d2 ] × . . . [0, dn ].
(3) fi (x) = 0 if x ∈ Fi0 and fi (x) = di if x ∈ Fi1 . In other words, f
maps each face of I n to the corresponding face of the parallelotope
P.
By Proposition 5.5.10 the first of the above facts implies that Jac f ≤ 1
almost everywhere. By Corollary 5.5.9 it follows that Volg (I n ) ≥ µn (f (I n )).
We will show that f (I n ) = P .
The second of the above facts allows us to consider f as a map from
I n to P . The third one implies that f (∂I n ) ⊂ ∂P ; hence the notion of
degree applies. We will show that the degree of f (modulo 2) as a map
from I n to P equals 1. Consider the linear map A : Rn → Rn defined by
A(x1 , . . . , xn ) = (x1 /d1 , . . . , xn /dn ). Its restriction A|P is a homeomorphism
from P to I n . Hence it is sufficient to prove that A ◦ f has degree 1 as a
map from I n to I n (then apply Proposition 5.6.7). We will prove this by
showing that A ◦ f is homotopic to the identity via a homotopy {ϕt }t∈[0,1]
such that ϕt (∂I n ) ⊂ ∂I n for all t (see Proposition 5.6.6). Let {ϕt } be a linear
homotopy between A◦f and the identity, i.e., ϕt (x) = (1−t)A(f (x)+tx. Let
x ∈ ∂I n and let F be a face of I n to which x belongs. Then y := A(f (x)) ∈ F
since A ◦ f |I n maps each face of I n to itself. It follows immediately that
ϕt (x) = tx + (1 − t)y ∈ F ⊂ ∂I n for all t.
Thus ϕt (∂I n ) ⊂ ∂I n for all t. Therefore deg2 (f ) = deg2 (A ◦ f ) =
deg2 (IdI n ) = 1 as we claimed above. In particular, the image of f is the
206 5. Smooth Length Structures

entire P . Then
n
Y
Volg (I n ) ≥ Vol(f (I n )) = Vol(P ) = di .
i=1
¤
Exercise 5.6.10. Prove the following generalizations of Theorem 5.6.9.
1. Let M be an n-dimensional
Qn Riemannian manifold with ∂M = ∂I n .
Prove that Vol(M ) ≥ i=1 distM (Fi0 , Fi1 ).
2. Let M be an n-dimensional Riemannian manifold and f : ∂M → ∂I n
be
Qn a continuous map with nonzero degree modulo 2. Then Vol(M ) ≥
−1 (Fi0 ), f −1 (Fi1 )).
i=1 distM (f
Exercise 5.6.11. Let M be a Riemannian manifold homeomorphic to
the sphere S 2 . Suppose that there are four points a, b, c, d ∈ M with
the following distances between them: |ab| = |bc| = |cd| = |da| = 1,
|ac| = |bd| = 3/2. Prove that the area of M is at least 1/2.
Hint: Prove that the distance between a shortest path from a to b and
one from c to d is not less than 1/2, and similarly for shortest paths from b
to c and from d to a. Then apply the Besikovitch inequality to each of the
two quadrilaterals into which these four paths divide the sphere.
Exercise 5.6.12. Let M be a Riemannian manifold homeomorphic to the
projective plane RP2 . Suppose that the length of any noncontractible loop
in M is at least 1. Prove that the area of M is at least 1/16.
Hint: Cut M along a shortest noncontractible loop. The resulting space
is homeomorphic to a disc. Divide its boundary into four arcs of equal
lengths and prove that the distance between “opposite” arcs is not less
than 1/4. Then apply the Besikovitch inequality to this disc (which is
homeomorphic to I 2 ).
Remark 5.6.13. In fact, the constant 1/16 above can be replaced by 2/π
(this is known as the theorem of Pu, see [Gro1]). The constant 2/π is
optimal and is achieved for the metric on RP2 space obtained from the
sphere of radius 1/π (by identifying opposite points).
Exercise 5.6.14. Let M be a Riemannian manifold homeomorphic to the
torus T 2 = S 1 × S 1 . Suppose that the length of any noncontractible loop in
M is at least 1. Prove that the area of M is at least 1/4.
Remark √ 5.6.15. The optimal constant in the statement of the above
exercise is 3/2 (Loewner’s theorem; see [Gro1]). Can you find an example
for which this is achieved?
Finally, let us mention that there is no “inverse Besikovitch inequality”
(i.e., no similar upper bound for the volume):
5.6. Besikovitch Inequality 207

Exercise 5.6.16. Prove that for every C > 0 and every integer n ≥ 2 there
exists a Riemannian structure g in the n-cube I n such that the Hausdorff
distance (in (I n , distg )) between Fi0 and Fi1 is less than 1 for all i, but
Volg (I n ) > C.

5.6.3. Generalization: systoles. Exercises 5.6.12 and 5.6.14 are partic-

ular cases of a general problem which can be formulated as follows.
Let (M, g) be a compact n-dimensional Riemannian manifold. Its one-
dimensional systole, denoted by sys1 (M, g), is the infimum of lengths of
noncontractible loops in M . (The definition of k-dimensional systoles for
k > 1 is more technical; see below.) The problem is: given the topological
type of M and the value of sys1 (M, g), can one estimate the volume of
(M, g) from below? For example, if M is homeomorphic to the √
torus T 2 or
the projective plane RP2 , one can prove that Vol(M, g) ≥ 23 sys1 (M, g)2
2
and Vol(M, g) ≥ vπ sys1 (M, g)2 , respectively (see remarks after Exercises
5.6.12 and 5.6.14).
More formally, for an M , one defines the isosystolic constant c1 (M ) of
an n-dimensional manifold M by

c1 (M ) = inf {Vol(M, g) : sys1 (M, g) ≥ 1},

g

or, equivalently,
Vol(M, g)
c1 (M ) = inf
g sys1 (M, g)n
where the infimum is taken over all Riemannian metrics g on M . One then
asks, for a given topological type of M , what is the value of this constant,
or at least whether it is positive or not.
Though the precise values of c1 (M ) are currently known only in a few
simplest cases (all of which are two-dimensional), a vast class of manifolds
M with c1 (M ) > 0 has been found (see [Gro1]).
The notion of systole can be generalized to higher dimensions in a
variety of ways. One possible approach is that the k-dimensional systole
of a Riemannian manifold M is the infimum of k-volumes of homologically
nontrivial closed k-dimensional “films” in M . Here one has to specify what
is meant by a “film”. It is not good to consider only k-submanifolds as
films; one of the reasons is that not all homology classes can be represented
by submanifolds. Instead, one takes the infimum over all singular Lipschitz
chains.
Let us explain this in more detail. Let τ be a k-dimensional Lipschitz
singular simplex, that is, a Lipschitz map of the standard k-simplex into M .
For such a simplex, the k-dimensional volume Vol(τ ) is well-defined (see
208 5. Smooth Length Structures

Section 5.5.2). This allows one to introduce the notion of volume for k-
PN
dimensional Lipschitz chains. Namely for c = , where xi ∈ R
i=1 xi τiP
N
and τi are Lipschitz singular simplices, define Vol(c) = i=1 |xi | Vol(τi ).
The boundary of a chain, cycles and homology groups are defined as usual.
Finally one defines
sysk (M, g) = inf{Vol(c) : c is a k-dimensional cycle not homological to 0}
and
Vol(M, g)
ck (M ) = inf ,
g sys (M, g)n/k
k
and asks whether ck (M ) is positive or not.
It is a bit surprising that the answer is absolutely different for k = 1
and k > 1. While c1 (M ) > 0 in many cases, it has been proved recently
([KS] and [Bab]) that one always has sysk (M ) = 0 for 1 < k < dim M .
This property is called systolic softness. Actually in [Bab] a more general
fact on so-called inter-systoles is proved; furthermore, M is allowed to be a
polyhedral space instead of a manifold.
 Chapter 6

Curvature
of Riemannian Metrics

The main conclusion of this chapter is the fact that a Riemannian manifold of
sectional curvature bounded by k is a length space with the same curvature
bound in the sense of Chapter 4. Hence the reader who is ready to take this
statement on faith (or who is familiar with Riemannian geometry) can omit
this chapter. It is included mainly to make this book more self-contained.
In our introduction to Riemannian geometry we basically stop where
substantial Riemannian geometry begins. There are many textbooks con-
taining various accounts of methods of differential and Riemannian geome-
try. In this course we regard Riemannian manifolds (as well as other length
spaces whose definitions rely on calculus of variations) just as one of the
important (or even leading) sources of examples. From this viewpoint, all
differential methods can retire after they produce certain basic informa-
tion to feed into the synthetic machinery. A shortest path leading from
infinitesimal definitions to local metric properties is surprisingly short, and
we intentionally make it somewhat longer to provide better motivations.
Basically all we need at the end is to convert a Riemannian metric to nor-
mal coordinates, and then prove the (local) Cartan-Alexandrov-Toponogov
Comparison Theorems 6.5.6.
For the sake of simplicity of exposition we restrict our introduction to
Riemannian geometry mainly to the case of two-dimensional manifolds; this
also allows us to give explicit formulas avoiding cumbersome indices and
to eliminate linear algebra. Higher-dimensional generalizations are mostly
straightforward and left to the reader (as an important exercise). We point

209
210 6. Curvature of Riemannian Metrics

out a construction where passing to the higher-dimensional case may cause

some difficulty and indicate how to overcome it.
One can introduce sectional (Gaussian) curvature, as well as understand
its geometric meaning, even without defining covariant differentiation at
all. To use this notion introduced this way one would have, however,
to take a certain (rather technical) statement on faith. Recall that the
cornerstone notion in the previous chapter was normal coordinates, and
we had two (dual) approaches to them. If one begins with an equidistant
family of curves (propagation of wavefronts), the geodesic curvature of a
curve in this family is defined as the derivative of its length element. Then
the Gaussian curvature governs the derivative of this geodesic curvature
across the equidistant family. If one thinks of normal coordinates as a
family of rays (geodesics), then their divergence is governed by the Gaussian
curvature. In the two cases, Gaussian curvature enters as a coefficient
in certain very simple differential equations (first-order and second-order,
respectively). This coefficient is independent of the choice of a particular
normal coordinate system (and depends on a point only, or on a point and a
two-dimensional direction in the higher-dimensional case)—and we use the
covariant derivative just to prove this fact (of course, it also can be obtained
by quite cumbersome coordinate computations).
Since our considerations are local, we deal with a metric defined in a
region. Of course a reader familiar with smooth manifolds will be able to
immediately reformulate our statements for Riemannian metrics on smooth
manifolds.
Here is a brief plan of the chapter.
First we begin with a coordinate computation. Using a direct variational
argument, we show that every shortest path satisfies the equation of geodes-
ics. This is the Euler-Lagrange equation, which is satisfied by minimizers
for a wide class of integral functionals. We need this computation for moti-
vations only, and we advise the reader who hates coordinate computations
to skip it. Another possibility is to go directly to Section 6.2, and then come
back to Section 5.2 and restate its material in invariant form.
Our next step is to substitute a noninvariant coordinate-wise derivative
for vector fields by covariant differentiation. One also defines covariant
differentiation for Finsler metrics, but its algebraic structure is much more
complicated, and it is not nearly as useful and convenient as the Riemannian
one. Using covariant differentiation, we rewrite the equation for geodesics
and introduce the Gaussian (sectional) curvature.
Sections 6.3 and 6.4 are devoted to the geometric meaning of Gaussian
curvature. Note that we suggest two alternative approaches: via divergence
of geodesics and via equidistant variations of curves (hypersurfaces). The
6.1. Motivation: Coordinate Computations 211

reader looking for geometric intuition and applications can take the results
of these sections for the definition of sectional curvature and completely skip
covariant differentiation. The latter is, however, needed for computations
in actual examples, as well as to show the correctness of such “synthetic”
definitions of Riemannian curvature.

6.1. Motivation: Coordinate Computations

6.1.1. Differential equations of geodesics in coordinates. Computa-
tions in this section are mainly used for motivation only, and they can be
omitted or referred to later.

Variations of curves. In this chapter we will often deal with one-parameter

families of curves, (also called a variation of a curve, especially if there is a
designated curve in the family). A one-parameter family of curves will be
usually denoted by γε . Formally this is nothing but a map (ε, t) −→ γε (t)
from (a region in) the coordinate (ε, t)-plane into Ω. The image of the co-
ordinate vector field ∂/∂t under this map forms the velocity field of curves
γε . The image of the other coordinate field ∂/∂ε will be called the variation
field. Indeed, when we vary the parameter ε keeping t fixed, γε (t) moves
(away from γ0 (t)) with the velocity (∂/∂ε)γε (t). Notice that neither the
velocity field nor the variation field is formally a vector field unless the map
γε (t) is a diffeomorphism from its domain in the (ε, t)-plane into Ω: they
may have “multiple values” at some points of Ω (when curves from the fam-
ily intersect), and they are not defined at the points of Ω which are not in
the image of the variation. Hence these “vector fields” should rather be un-
derstood as vector-valued maps from the domain of the variation (the same
way as we treat the coordinate “vector fields” for a degenerate coordinate
system).
In many cases we will use one-parameter families of coordinate lines of
a certain (possibly degenerate) coordinate system; in this case one of the
coordinate vector fields is the field of velocities, and the other one is the
variation field.

Variational computation in coordinates. Let γ(t) = (x(t), y(t)),

γ : [a, b] → Ω be a smooth shortest path parameterized by arc length,
γ̇(t) = dγ/dt = T (t), hT, T i = 1. Here and later on differentiation with
d
respect to t will be denoted by a dot interchangeably with dt . We will in-
clude γ into a family of paths with the same endpoints and make use of
the fact that the length functional restricted to this family attains its min-
imum at γ. Choose a vector function V (t) = (m(t), n(t)) , V : [a, b] → R2 ,
V (a) = V (b) = 0, which will be our variation field. Fix a coordinate
system (x, y) and consider the family of curves γε = γ + εV given by
212 6. Curvature of Riemannian Metrics

γε (t) = (x(t) + εm(t), y(t) + εn(t)). Of course, this is a very noninvariant

construction. It is based on a particular choice of coordinates (and hence a
vector structure) on Ω; but making it invariant would bring us too far away
from the topic of this section. A reader familiar with variational methods
on manifolds can easily convert our arguments into a coordinate-free form.
All curves γε have the same endpoints, and the fact that γ is a shortest
path between its endpoints implies that (d/dε)|ε=0 L(ε) = 0, where L(ε) is
the length of γε . Thus we have
d
0= L(ε)¯¯ .
dε ε=0

Denote Tε = dγ dTε
dt and note that dε (t, ε) = V (t). Now differentiating under
ε

the symbol of integration and substituting hT, T i |ε=0 = 1, we get:

Z bD E 1 ³ ∂E
d ¯ ∂F ∂G 2 ´
(6.1) 0 = L(ε)¯ = T (t), V̇ (t) + ẋ2 + 2 ẋẏ + ẏ dt,
dε ε=0 a 2 ∂ε ∂ε ∂ε
where dot over a letter means differentiation with respect to t.
Now we want to use integration by parts. It is very tempting to use the
usual Product Rule for differentiating the Euclidean scalar product:
d D E D E
hT (t), V (t)iEucl = Ṫ (t), V (t) + T (t), V̇ (t) ,
dt Eucl Eucl

but for our “scalar product” h, i this formula is just incorrect! Recall that
we use h, i to denote a bilinear form Qp (·, ·) on (tangent) vectors. The
coefficients of this form depend on p, and hence one has to use the Chain
Rule to differentiate it. There is an invariant and very convenient way
of doing this by introducing covariant derivatives (see Section 6.2). As a
matter of fact, already the expressions V̇ (t), Ṫ (t) themselves depend on
the choice of a coordinate system! Covariant derivatives are designed to
substitute this “noninvariant” coordinate-wise derivative by an “invariant”
(coordinate-independent) operation.
We will proceed here with a coordinate computation now. A geometrical
meaning of what is going on here will become clear in Section 6.2.
Recall that

hT (t), V (t)i = E(x(t), y(t))ẋm(t) + F (x(t), y(t))ẋn(t)

+ F (x(t), y(t))ẏm(t) + G(x(t), y(t))ẏn(t).

Differentiating this expression with respect to t (and omitting arguments

to avoid too cumbersome formulas), we get
D E d D E
T, V̇ = hT, V i − Ṫ , V − Ė ẋm − Ḟ (ẋn + ẏm) − Ġẏn.
dt
6.1. Motivation: Coordinate Computations 213

Substitute the last expression in (6.1) and take into account that Eε =
dE ∂E ∂E
dε = ∂x ẋ + ∂y ẏ (the Chain Rule) and simultaneously for Fε , Gε . After
some intermediate calculations we get:

Z bD E b DZ E
(6.2) 0 = T (t), V̇ (t) dt = hT, V i |ba
− ( Ṫ , V + mA + nB) dt =
a a
Z bµ 2 2
¶
d x d y d2 x d2 y
= hT, V i |ba − m(A + E 2 + 2 F ) + n(B + F 2 + G 2 ) dt,
a dt dt dt dt
where
1 ∂E ∂E ∂F 1 ∂G
A = − ẋ2 − ẋẏ − ẏ 2 ( − ),
2 ∂x ∂y ∂y 2 ∂x
∂F 1 ∂E ∂G 1 2 ∂G
B = −ẋ2 ( + ) − ẋẏ − ẏ .
∂x 2 ∂y ∂x 2 ∂y
Note that hT, V i |ba = 0 since we supposed that all curves γε have the
same endpoints. Recall that this identity is true for all choices of functions
m, n, with the only restriction that they vanish at a and b. Using a standard
analytical argument one shows that this implies that both expressions
A + E ẍ + F ÿ and B + F ẍ + Gÿ
must be identically zero! Indeed, if for instance the first expression was
nonzero at some point t0 , one chooses n = 0 and a nonnegative function m
such that m(t0 ) = 1 and m vanishes outside a small neighborhood of t0 . It
is easy to see that (6.2) is not satisfied for this choice of m, n.
Hence we see that our path γ satisfies just the system of two nonlinear
second-order differential equations (5.10) and (5.11) that we used to define
geodesics:
1 ∂E ∂E ¡ ∂F 1 ∂G ¢
E ẍ + F ÿ = ẋ2 + ẋẏ + ẏ 2 − ,
2 ∂x ∂y ∂y 2 ∂x
¡ ∂F 1 ∂E ¢ ∂G 1 2 ∂G
F ẍ + Gÿ = ẋ2 − + ẋẏ + ẏ .
∂x 2 ∂y ∂x 2 ∂y
Exercise 6.1.1. Repeat the same computation for a conformal metric in Ω
(given by a function λ(p)) for a shortest curve γ parameterized by Euclidean
arc length. Show that the Euclidean
¿ curvature k(t)À of γ satisfies the relation
∆λ(γ(t))
k(t) = T (t), .
λ(γ(t))
This equation says that a shortest path must bend towards the direction
where λ increases. Although this may sound like nonsense at first glance
(for a shortest path should try to go through small values of λ), the following
consideration relieves this feeling. Imagine that you walk around a swamp
(with λ being very big at the center of the swamp). Suppose that the swamp
214 6. Curvature of Riemannian Metrics

is on your left. Notice that you also keep turning to your left, that is towards
the swamp.

6.2. Covariant Derivative

The goal of this section is to substitute noninvariant component-wise differ-
entiation (which caused us a lot of trouble due to the absence of a product
rule) by an invariant operation of covariant differentiation.
When we were deducing a differential equation for geodesics, our compu-
tations became rather messy once we had to differentiate a scalar product.
Moreover, we noticed that we were doing coordinate computations, and cer-
tain objects had no geometrical meaning.
Our main difficulty was that we needed to differentiate vector functions
that take values in the tangent space. Component-wise differentiation of
such functions is not invariant under coordinate changes.
Let us describe this situation in more detail. Let γ(t) be a curve starting
from p ∈ Ω with velocity vector T , that is, γ(0) = p, γ̇(0) = T . Here and
later on in this section we always mean that the derivatives are taken at
p, that is, for t = 0. In order to simplify notation we do not indicate this
explicitly in our formulas.
If f : Ω → R is a function, we differentiate it by
df f (γ(t)) − f (γ(0))
(6.3) Tf = = lim .
dt t→0 t
Now consider a vector function along γ, that is, a function V : R → T Ω
such that V (t) ∈ Tγ(t) Ω. The most important example of such a vector
function is a restriction V (t) = W (γ(t)) of a vector field W : Ω → T Ω. If
we want to differentiate V , we cannot use formula (6.3): we cannot subtract
tangent vectors at different points, since the result would change if we choose
a different coordinate system. Of course, if we fix a coordinate system,
V (t) can be differentiated component-wise, as we did in Subsection 6.1.1.
However, this derivative has no geometric meaning, as again it will depend
on our choice of coordinates. The objective of this section is to construct an
“invariant”way of differentiating vector functions. The next section contains
an axiomatic definition of covariant differentiation. It is followed by two
sections with motivations; these sections are optional.

6.2.1. Axiomatic definition of covariant differentiation. Let us de-

D
note the covariant differentiation that we want to construct by dt . Let us
make a list of its desirable properties and take them as axioms. These
properties are counterparts of familiar properties of the derivative for vector
6.2. Covariant Derivative 215

functions in Rn . We will see that these properties uniquely determine the

operation of covariant differentiation.
(1). Linearity: For any two real constants a, b,
D D D
(aV + bW ) = a V + b V.
dt dt dt
(2). Product rule: For a real-valued function f (t),
D D df
(f V ) = f (p) V + V.
dt dt dt
(3). Differentiation of restrictions of vector fields: If V is a restriction
D
of a vector field W : V (t) = W (γ(t)), we want the value of dt V to be the
same for all curves γ with the same tangent vector T = γ ′ (t0 ) at p. This
means that the following operation of (directional) covariant differentiation
is well-defined: given a vector field W : Ω → T Ω and a vector T ∈ Tp Ω, we
set
D
∇T W = W (γ(t)),
dt
where γ is a curve emanating from p with velocity T .
Formally we can apply ∇T to W only if W is a vector field defined in a
neighborhood of p. However, by (3) the result ∇T W is determined by the
restriction of W to any curve with velocity T at p. The notation ∇T W and
DW
dt = D/dt will be used interchangeably once W is defined along a curve
emanating from p with velocity T . It is obvious that the operations ∇ and
D/dt uniquely determine each other.
(4). Linearity of covariant directional derivative: For constants a, b,
∇aT +bS W = a∇T W + b∇S W.

(5). Product rule for scalar multiplication:

¿ À ¿ À
d DV DW
hV, W i = , W + V, .
dt dt dt
Unfortunately, the covariant derivative is not uniquely defined by this
list of five very natural properties. We will have to add one more property,
which is a counterpart of the fact that partial derivatives commute. We will
precede it by a short digression.

Commuting vector fields. Even if two vector fields form a basis at

every point, they usually cannot be represented as coordinate fields of some
coordinate system (even locally). Coordinate vector fields have to satisfy an
additional relation: their Lie bracket must be identically zero. We will not
even define the Lie bracket here; instead, when considering two vector fields
simultaneously, we will restrict ourselves to a certain class of commuting
216 6. Curvature of Riemannian Metrics

vector fields. Two vector fields V, W : Ω → T Ω are said to be commuting

if they can be represented as the images of two coordinate fields under a
smooth map. More precisely, we require that there exists a smooth map
∂
ϕ : Ω′ ⊂ R2 → Ω such that, for every q ∈ Ω′ , V (ϕ(q)) = X(q) = dq ϕ( ∂x )
∂
and W (ϕ(q)) = Y (q) = dq ϕ( ∂y ). Loosely speaking, a pair of commuting
vector fields is a pair of “coordinate vector fields in a coordinate system
with possible degenerations”.
To justify the term commuting vector fields we suggest the following easy
exercise:

Exercise 6.2.1. For two commuting vector fields V, W , and a function f ,

show that V W f = W V f ( that is, they commute as differential operators).
The converse result (stating that if V W f = W V f for every smooth
function f , then V and W commute) is also correct (it follows from the
Frobenius Theorem). Thus this property can be used as an alternative
definition of commuting vector fields.

Symmetry of covariant derivative. Now we can formulate the last

axiom of covariant derivative:
(6). Symmetry: ∇V W = ∇W V for every two commuting vector fields
V, W .
To see why this axiom actually generalizes the commutativity property
of partial derivatives, we suggest the following two exercises:

Exercise 6.2.2. Two vector fields V = (v1 (x, y), v2 (x, y)) and W =
(w1 (x, y), w2 (x, y)) on R2 . commute if and only if their components v1 ,
v2 , w1 , w2 satisfy certain differential relation. Find this relation and prove
this.
Hint: You may use the previous exercise applied to coordinate functions.

Exercise 6.2.3. The component-wise derivative (used in vector calculus)

DV W of a vector field W = (w1 (x, y), w2 (x, y)) on R2 along a vector
V = (v1 , v2 ) is defined as
∂w1 ∂w1 ∂w2 ∂w2
DV W = (v1 + v2 , v1 + v2 ).
∂x ∂y ∂x ∂y
Show that for commuting vector fields DV W = DW V .

Levi-Civita Lemma. The following simple Levi-Civita Lemma is consid-

ered as the main theorem in the foundations of Riemannian geometry:

Lemma 6.2.4. The operations ∇V W and DV /dt are uniquely defined by

properties (1)–(6).
6.2. Covariant Derivative 217

A detailed proof of this lemma can be found in numerous textbooks, and

we will limit ourselves to a sketch.
Introduce a coordinate system (x, y). For two vector fields V = v1 X +
v2 Y , W = w1 X + w2 Y , where as usual X, Y are coordinate vector fields,
properties (1), (2) and (4) imply that
∂w1 ∂w1 ∂w1 ∂w2
∇V W = (v1 + v2 )X + (v1 + v2 )Y
∂x ∂y ∂x ∂y
+ v1 w1 ∇X X + v1 w2 ∇X Y + v2 w1 ∇Y X + v2 w2 ∇Y Y.

Hence ∇ is uniquely determined if we know three vector fields: ∇X X,

∇X Y = ∇Y X (since X and Y commute!), and ∇Y Y . Thus it is enough to
determine six real-valued functions:
Γ11,1 = hX, ∇X Xi , Γ11,2 = hY, ∇X Xi , Γ21,1 = Γ12,1 = hX, ∇X Y i ,

Γ21,2 = Γ12,2 = hX, ∇X Y i , Γ22,1 = hX, ∇Y Y i , Γ11,2 = hY, ∇Y Y i .

To find these functions (in terms of E, F, G and their partial derivatives)
one computes partial derivatives of the functions E = hX, Xi, F = hX, Y i,
and G = hY, Y i using (5), thus obtaining a nondegenerate system of linear
equations for Γij,k . This proves the uniqueness part. Now it is easy to check
that ∇, determined by this choice of Γij,k , indeed satisfies axioms (1)–(6).
The details are left to the reader.
The following exercise involves rather long computations:
Exercise 6.2.5. Using formulas (5.8) and (5.9), repeat the argument to find
explicit coordinate formulas for the spheres and hyperbolic planes. Using
(5.7), convert the formula for the component-wise Euclidean derivative (see
Exercise 6.2.3) to polar coordinates.

6.2.2. Analytical motivation. Let us look again at the problem that

we ran into in Subsection 6.1.1. The problem was that we could not use
the product rule to differentiate the Riemannian scalar product of two
vector fields. Let us pick a point p and fix coordinates (x, y) such that
E = G = 1, and F = 0 at the point p (to avoid cumbersome calculations).
Suppose that p = γ(0) and consider two vector fields V (t) = (v1 (t), v2 (t)),
W (t) = (w1 (t), w2 (t)) along γ. Next calculation we will do at the point p.
Then, applying the chain rule as in Subsection 6.1.1, we have:
d D E D E
hV, W i = V̇ , W + V, Ẇ + 2S(V, W ),
dt
where
dE dF dF dG
2S(V, W ) = v1 w1 + v1 w2 + v2 w1 + v2 w2 ,
dt dt dt dt
218 6. Curvature of Riemannian Metrics

and V̇ , Ẇ are the component-wise derivatives V ′ = (v̇1 (t), v̇2 (t)), W ′ =

(ẇ1 (t), ẇ2 (t)).
We notice that this formula differs from the product rule by the term
2S(V, W ), which does not involve derivatives of V and W: S is just a
symmetric bi-linear form. Hence it is tempting to make a correction by
adding to V̇ = dV /dt and Ẇ = dW/dt certain linear expressions in such a
way that the product rule would hold. It is natural to “send one-half of 2S
to V̇ and one-half to Ẇ ”, thus making our formula as symmetric as possible
(this is the reason why we put a coefficient 2 in front of S). Thus we define
covariant differentiation by the following formula:

µ µ ¶ µ ¶¶
DV 1
dE dF 1 dF dG
(6.4) = v̇1 + v1 + v2 , v̇2 + v1 + v2
dt 2dt dt 2 dt dt
¿ À
DV
(recall that E = G = 1, F = 0 at p!). Then of course ,W =
D E dt
V̇ , W + S(V, W ), and hence we get a nice product rule:
¿ À ¿ À
d D D
hV, W i = V, W + V, W .
dt dt dt

It is not at all clear if such a definition of DV /dt has any geometric

meaning (that is, whether it is independent of the choice of a coordinate
system). However, one can check that, if defined this way, DV /dt satisfies
the properties (1)–(6) from the previous section (do this as an exercise!)
Hence, by the uniqueness part of the Levi-Civita Lemma (6.2.4), DV /dt is
indeed coordinate-independent.

6.2.3. Another motivation: Covariant differentiation on embedded

surfaces. Let us consider our favorite example of an embedded surface.
Thus we have an embedding r : Ω → R3 , and h, i is given by hV, W i =
hdr(V ), dr(V )i, where the scalar product in the right-hand side of the
formula is the usual scalar multiplication in R3 . One may prefer to think of
Ω as a subset in R3 (i.e., replace Ω by r(Ω)), and regard r as a coordinate
system. Then tangent vectors to Ω can be viewed just as vectors in the
ambient space, and the Riemannian scalar product is nothing but the
restriction of the Euclidean scalar product to vectors tangent to Ω.
Let us repeat the variational argument from Subsection 6.1.1 in this set-
up. All scalar products here mean just the Euclidean scalar multiplication
in R3 . Recall that we begin with a smooth shortest path γ : [a, b] → Ω
parameterized by arc length, dγ/dt = T (t), hdr(T ), dr(T )i = 1. Of course,
dr(T ) is nothing but the velocity vector of our curve r(γ) in R3 . We choose
6.2. Covariant Derivative 219

a vector function V (t) = (m(t), n(t)), V : [a, b] → R2 , V (a) = V (b) = 0, and

consider a family of curves γε = γ + εV given by
γε (t) = (x(t) + εm(t), y(t) + εn(t)).
d
Since the length of γε assumes its minimum at ε = 0 we have L(ε)|ε=0 =0 ,
dε
where L(ε) is the length of γε . Thus we have
d
0= L(ε)|ε=0
dε
Z br
d ­ d d ®
= dr(T ) + ε (dr(V (t))), dr(T ) + ε (dr(V (t))) dt.
dε a dt dt
Differentiating under the symbol of integration and substituting ε = 0, we
get:

Z b¿ À
d
(6.5) dr(T (t)), dr((V (t)) dt = 0.
a dt
Now we can use integration by parts, for we can apply the product rule
to the Euclidean scalar product! We get
Z b
d b d
0 = L(ε)|ε=0 = hdr(T ), dr(V )i |a − hdr(T (t)), dr(V (t))i dt.
dε a dt

Notice that not every vector in R3 can be represented as dr(V (t)) for
some V : dr(V (t)) has to be tangent to our surface. Now arguing as in
d
Subsection 6.1.1 one concludes that dt dr(T ) has to be orthogonal to all
vectors tangent to the surface at r(γ(t)). In other words, the acceleration
of a shortest path parameterized by arc length remains orthogonal to the
surface at all times. If thinking of a geodesic as a free motion of a particle
confined to the surfaces, one can recognize a well-known mechanical principle
that the centrifugal acceleration is orthogonal to the surface.
The only disadvantage of this argument is that it does not have
much meaning in terms of intrinsic geometry of the surface. Imagine two-
dimensional creatures living in the surface. Vectors in R3 other than those
tangent to the surface have no meaning in their physical world. There were
two objects that stuck out of two dimensions in our argument, namely,
d d
dt dr(T (t)) and dt dr(V (t)). Both of them are of the same nature: they came
from differentiating a tangent vector field, and this derivative might have
a nontrivial component orthogonal to the surface. Notice that both vec-
tors come into our formulas multiplied by a vector tangent to the surface,
and hence they can be replaced by their orthogonal projections to the plane
tangent to the surface at r(γ(t)). This suggests the following strategy of
defining an “intrinsic differentiation” of tangent vector fields: begin with
220 6. Curvature of Riemannian Metrics

the usual derivative in the ambient space, and then get rid of its component
orthogonal to the surfaces by projecting it to the tangent plane:
D d
V (t) = dr−1 Proj dr(Tγ(t) Ω) dr(V (t)).
dt dt
If one wants to forget about the embedding r and think of Ω as a subset of
R3 , this formula takes a simpler form:
D dV (t)
V (t) = Proj Tγ(t) Ω .
dt dt
D
It is not clear a priori that, if defined this way, dt would not change
3
if we consider another isometric embedding Ω into R , that is, another
embedding that induces the same Riemannian metric on Ω. To prove that
D
dt depends only on the intrinsic geometry (Riemannian metric) of Ω, one
D
can easily check that dt satisfies axioms (1)–(6) of covariant derivative, and
then apply Lemma 6.2.4 (do this as an exercise!).

Second fundamental form. As a concluding remark, let us notice that

the normal component of the coordinate-wise derivative is
¿ À ¿ À
d D
(6.6) V, W − V, W = S(V, W ).
dt dt
Comparing this with (6.4), one concludes that the normal component is a
symmetric bilinear form on Tγ(0) Ω; its value depends only on the values
of V (0) and W (0). This form is called the second fundamental form of
the surface, and its eigenvectors and eigenvalues are called the principal
directions and principal curvatures. While the intrinsic geometry of a
surface is determined by h, i (which is often called the first fundamental
form), it is a well-known theorem in the differential geometry of surfaces
that S and h, i together determine the shape of a surface in R3 (up to rigid
motions).

6.2.4. Parallel transport. Though our exposition will not rely on the
notion of a parallel transport of vectors, we find it relevant to give its
definition and list its main properties.
When we were introducing the concept of a tangent vector, we empha-
sized that there is no way to define the notion of “the same direction at
different points” in a consistent way. However, if we consider a region Ω
with a Riemannian metric, a choice of a path γ between two points p = γ(a)
and q = γ(b) induces a natural correspondence between Tp Ω and Tq Ω. It is
enough to assume that γ is piecewise smooth.
We say that a vector field along γ is parallel if it satisfies the following
first-order linear differential equation: DV (t)/dt = 0. Every initial value
V (a) = W uniquely determines a parallel vector field V , and thus one can
6.3. Geodesic and Gaussian Curvatures 221

define a map Iγ by Iγ (W ) = V (b), Iγ : Tp Ω → Tq Ω. This map is called

the parallel transport along γ. It is obvious that Iγ is a linear map since
solutions of a linear differential equations depend linearly on the initial data.
Moreover, Iγ is a linear isometry between (Tp Ω, h, ip ) and (Tq Ω, h, iq ). This
follows from the fact that the scalar product of two parallel vector fields is
a constant function:
¿ À ¿ À
d D D
hV (t), W (t)i = V (t), W (t) + V (t), W (t) = 0.
dt dt dt
Let us note that the Gaussian curvature (which will be introduced in the
next section) is a certain measure of the dependence of Iγ on γ. Using the
parallel transport, one can compute the covariant derivative in the same way
as one computes the derivative of a vector-valued function to R3 in vector
calculus:
D d ³ −1 ´
V = Iγ[0,t] (V (γ(t)) .
dt dt
Note that this is a direct analog of the formula (6.3) for differentiating scalar
functions.

6.3. Geodesic and Gaussian Curvatures

6.3.1. Invariant equation of geodesics. In order to rewrite the differ-
ential equation of geodesics using the covariant derivative, let us repeat the
computation from Subsection 6.1.1. Recall that we have a family of curves
∂
γε with a variation vector field V = Vε (t) = ∂ε γε (t), and γ = γ0 is a shortest
∂
path parameterized by arc length. As usual, T = Tε (t) = ∂t γε (t) is the
velocity field for the family of curves γε ; here we do not suppose so far that
curves γε have the same endpoints. We have
Z b p
d d
L(ε)|ε=0 = hT, T i dt
dε a dε |ε=0

(because hT0 (t), T0 (t)i = 1)

Z b¿ À Z b
D
= T, T dt = h∇V T (t), T (t)i dt.
a dε |ε=0 a
Since V and T commute, by axiom (6) of the covariant derivative, we have
∇V T = ∇T V , and hence
Z b¿ À
d D
L(ε)|ε=0 = T (t), V (t) dt = 0.
dε a dt
d
­ D
® ­ D
®
Using dt hV, T i = dt T, V + T, dt V , we get
Z b¿ À
d b D
(6.7) L(ε)|ε=0 = hV, T i |a − T (t), V (t) dt.
dε a dt
222 6. Curvature of Riemannian Metrics

This formula is called the first variation formula. Now suppose that curves
d
γε have the same endpoints. Then hV, T i |ba = 0. In this case dε L(ε)|ε=0 = 0
for any choice of V (since γ0 is a shortest path). Arguing as in Subsection
6.1.1, we obtain a very simple differential equation:
D
(6.8) T (t) = 0, or in alternative notation ∇T T = 0.
dt
Definition 6.3.1. A path γ is called a geodesic if its velocity vector field T
satisfies the equation (6.8).

This is very similar to the Euclidean case: a straight line is a curve traced
by a motion with constant velocity. Here for a velocity to be “constant”
means that its covariant derivative is zero.
Owing to invariant notations, equation (6.8) makes sense in any dimen-
sion.
Note that every geodesic
­D is® parameterized proportionally to arc-length
d
(because dt hT, T i = 2 dt T, T = 0).
Exercise 6.3.2. Prove that in dimension two Definitions 5.2.1 and 6.3.1 are
equivalent.
Exercise 6.3.3. Re-prove the results of subsections 5.2.2 and 5.2.1 using
the new definition of geodesics.

6.3.2. First variation of length. Geodesic curvature. Pay attention

that the right side of expression (6.7) depends only on a vector field V but
does not depend on a family γε directly. Expression (6.7) is often called
the formula of the first variation of length, for it describes how the length
of a curve changes under a variation produced by “pulling the curve by a
vector field V ”. Recall that it is valid only for a curve parameterized by arc
length. Note that (6.7) for a Riemannian manifold has the same form as for
the Euclidean plane. This reflects the fact that Riemannian manifolds are
“Euclidean in the first order”.
This formula takes the following simple form if γ is a geodesic:
d
(6.9) L(ε)|ε=0 = hV, T i |ba .
dε
This equation carries particularly nice information if we restrict ourselves
to a neighborhood such that there is a unique shortest path between any
two points in the neighborhood. Let σ(τ ) be a smooth path, and denote
d
its velocity by V = dτ σ. Let ρ(τ ) = d(p, σ(τ )) be the distance from p to a
“moving point” σ(τ ). Applying formula (6.7) to the family of shortest paths
γτ connecting p and σ(τ ) (and noticing that the corresponding variation
vector field vanishes at p, for this endpoint of the geodesics γτ does not
6.3. Geodesic and Gaussian Curvatures 223

move), we obtain
dρ
= hV, T i ,
dτ
where T is the tangent vector of γτ at its end σ(τ ). In other words, the
derivative of the distance from p to a point moving at unit speed is equal to
the negative cosine of the angle between the velocity vector of the moving
point and the direction from the moving point towards p (since the latter is
just −T ).
Exercise 6.3.4. Let σ(t) be a smooth closed curve, and let σ(t0 ) be the
point of σ closest to p. Show that every shortest path connecting p and
σ(t0 ) is orthogonal to (the velocity of) σ at t0 .

Let us now look at the case when a curve γ is not necessarily a geodesic.
D
Notice that the vector ∇T T = dt T (t) is always orthogonal to T (t). Indeed,
¿ À
d D
0= hT (t), T (t)i = 2 T (t), T (t) .
dt dt
D
The number | dt T (t)| = kg (t) is called the geodesic curvature of γ; in
dimension 2 one can assign it a sign + or − by choosing an orientation.
D
In higher dimensions one often calls S(t) = dt T (t) the curvature vector. As
a matter of fact, this is nothing but the value of the second fundamental
form of our curve (regarded as a one-dimensional submanifold) applied to
T : compare this definition with formula (6.6). This give a new geometric
insight into the meaning of the second fundamental form: it describes how
the metric of a submanifold changes under variations of the submanifold.
For a variation with fixed endpoints (more generally, if the variation field is
orthogonal to the curve at it ends), the double substitution term hV, T i |ba
disappears, and (6.7) takes the following form:
Z b
d
(6.10) L(ε)|ε=0 = h−S(t), V (t)i dt.
dε a

Exercise 6.3.5. 1. Show that for a general parameterization of γ (not

necessarily by arc length)
(∇T T )N
S(t) = ,
hT, T i
where (∇T T )N is the component of the vector ∇T T orthogonal to T .
2. If the variation field V is orthogonal to T , use a change of variables
in (6.7) to show that the first variation of length is
Z b* +
d ∇T T
L(ε)|ε=0 = p , V (t) dt.
dε a hT, T i
224 6. Curvature of Riemannian Metrics

Exercise 6.3.6. Carry out these computations for curves in R2 and R3 with
the Euclidean metric. Compare S with a well-known notion of curvature in
differential geometry. Study the length of equidistants of a closed planar
curve by shrinking it towards its inward normal.

6.3.3. Curvature tensor. Choose a coordinate system (x, y), and let
X and Y be (as usual) the coordinate fields. (The reader who follows
higher-dimensional generalizations can take two first coordinate fields of
a coordinate system (x, y, · · · , ).) Then of course X and Y commute as
differential operators on functions: XY f = Y Xf for every smooth function
D
f . However, unlike the Euclidean case, in general the operators ∇X = dx
D
and ∇Y = dy do not commute: for a vector field Z, ∇X ∇Y Z 6= ∇Y ∇X Z.
The difference R(X, Y )Z = ∇Y ∇X Z − ∇X ∇Y Z is called the curvature
operator. A real-valued expression hR(X, Y )Z, W i, which depends on four
vector arguments X, Y, Z, W , is called the curvature tensor. The definition
of R suggests that it depends on four vector fields (and the first two of them
are supposed to be coordinate vector fields for some coordinate system). It
is clear that R(X, Y )Z is a linear function in each of its three arguments.
The following exercises shows that R(X, Y )Z at a point p depends only on
the values of X, Y , and Z at p ! The proofs of the exercises consist of a
straightforward computation based on axioms of the covariant derivative:
Exercise 6.3.7. Show that, for a smooth function f :
(1) ∇X ∇Y (f Z) − ∇Y ∇X (f Z) = f · (∇X ∇Y Z − ∇Y ∇X Z),
(2) ∇X ∇(f Y ) Z − ∇(f Y ) ∇X = f · (∇X ∇Y Z − ∇Y ∇X Z).
Exercise 6.3.8. Show that:
(1) R(X, Y )Z = −R(Y, X)Z; in particular, R(X, X)Z = 0,
(2) hR(X, Y )Z, W i = hR(Z, W )X, Y i.
Exercise 6.3.9. Combining the two previous exercises, show that R(X, Y )Z
= 0 at p if at least one of the fields X, Y, Z vanishes at p. Conclude that
R(X, Y )Z at p depends only on the values of X, Y , and Z at p!

For two vectors X, Y ∈ Tp Ω, we denote K(X, Y ) = hR(X, Y )X, Y i. Let

us choose two vectors X, Y ∈ Tp Ω such that the area of the parallelogram
spanned by them is 1 (where the area is taken with respect to h, ip ).
For instance, we can choose two orthogonal unit vectors. The quantity
K(p) = K(X, Y ) = hR(X, Y )X, Y i is called the sectional curvature at p.
In dimension two it is the same as the Gaussian curvature of a surface.
Exercise 6.3.10. Using symmetry properties of R (and in particular
R(X + αY, Y )Z = R(X, Y )Z; see Exercise 6.3.8), show that K(X, Y ) de-
pends only on the area of the parallelogram spanned by X, Y , and hence
K(p) is well-defined.
6.3. Geodesic and Gaussian Curvatures 225

Exercise 6.3.11. Show that if a Riemannian metric is multiplied by a con-

stant c (that is, its metric coefficients are multiplied by c2 ), then its Gaussian
curvature gets divided by c2 .

Exercise 6.3.12. Use Exercise 6.2.5 to compute the Gaussian curvature of

the spheres and the hyperbolic planes.
Exercise 6.3.13. Let X and Y be two orthonormal vectors. Prove that we
get R(X, Y )Z if we rotate |Kσ |Z on the angle π/2 in a direction depending
on sign of Kσ . Can you generalize this exercise to dimensions greater two?
The importance of the Gaussian (sectional) curvature will be explained
in the following section. We conclude this section with a remark related to
the higher-dimensional case. The terms “sectional curvature” and “Gaussian
curvature” are more or less interchangeable in our 2-dimensional exposition
and the former is perhaps even more common. On the other hand, the
term “sectional curvature” sounds silly in our two-dimensional context. The
reason why this curvature is called “sectional” in higher dimensions becomes
clear from the following three exercises, which are addressed to a reader who
keeps track of a multi-dimensional theory parallel to our two-dimensional
exposition. The three exercises consider the case when the dimension of Ω
is greater than two.
Exercise 6.3.14. Show that K(X, Y ) = K(X1 , Y1 ), provided that the
parallelograms spanned by X, Y and X1 , Y1 have equal areas and belong
to the same 2-plane. Hence K(X, Y ) depends only on X ∧ Y .
Let σ be a two-dimensional plane σ ⊂ Tp Ω. Define Kσ by Kσ =
K(X, Y ), where X and Y are two orthonormal vectors X, Y ∈ σ. Equiva-
K(X,Y )
lently, Kσ = ||X∧Y ||2
.
Exercise 6.3.15. Consider a two-dimensional embedded surface obtained
by restricting expp to a small neighborhood of the origin in σ; equip it
with the metric induced from Ω. One can think of this surface as a two-
dimensional section of Ω formed by rotating a geodesic passing through p so
that its velocity vectors sweep σ. Show that the sectional curvature Kσ (p) is
equal to the sectional (Gaussian) curvature of this two-dimensional surface
at p.
Exercise 6.3.16. Use linear properties of the curvature operator to find
how the sectional curvature depends on a (two-dimensional) direction σ in
a metric product S 2 × S 2 of two unit spheres (assume that you know that
the curvature of the unit sphere is 1). Do the same for other products,
such as the product of Euclidean and hyperbolic planes, and a sphere and
a hyperbolic plane.
226 6. Curvature of Riemannian Metrics

Parallel transport and curvature. Gauss-Bonnet Formula. There is

the following description of the curvature using the parallel transport. Let
X, Y ∈ Tp Ω be two vectors such that the area of the parallelogram spanned
by them is 1. Consider a closed path γ, γ(a) = γ(b) = p. Then Iγ defines a
linear isometry from Tp Ω to itself. Then
(6.11) R(X, Y )Z = (Iγ (Z) − Z)Area(γ) + o(Area(γ)),
where Area(γ) is the (oriented) Riemannian area enclosed by γ. Notice that
formula (6.11) easily follows from the definition of the curvature operator
if γ is a rectangle formed by coordinate lines of a coordinate system whose
coordinate vectors at p are X and Y . This formula means that, when a
vector is carried along a small loop back to the same point, the vector gets
rotated by an angle that is close to the product of the Gaussian curvature
at the point and the area enclosed in the loop. formula (6.11) can be easily
generalized to higher dimensions (in terms of differential two-forms, or by
imposing an assumption formalizing that “γ encloses a region in a two-
dimensional surface tangent to X and Y ”). For two-dimensional surfaces
formula (6.11) implies the following Gauss-Bonnet Formula:
Theorem 6.3.17. Let a smooth closed curve γ enclose a simply-connected
region Ω. Then Z Z
K(p) dA(p) = kg (t) dt − 2π,
Ω γ
R
where γ is assumed to be parameterized by arc length, and Ω K(p) dA(p) is
the integral of the Gaussian curvature with respect to the Riemannian area.
If γ is a piecewise smooth curve, then one has to add π − αi for each
corner with interior angle αi to the right-hand side of the formula.

To get a better grasp of this theorem, apply it to (1) a smooth planar

curve, (2) a Euclidean triangle (3), to a greater circle in a sphere, and (4)
to each of the two regions whose boundary is a small circle in a sphere.
We will not use formula (6.11) and Theorem 6.3.17, and their proofs are
omitted. A proof of the Gauss-Bonnet Formula for hyperbolic and spherical
triangles can be found in Section 5.3.

6.4. Geometric Meaning of Gaussian Curvature

The main property that distinguishes a Riemannian length structure from
other length spaces is that locally it is “almost Euclidean”. More precisely,
Lemma 5.1.13 asserts that every point p admits a neighborhood with coor-
dinate system such that the metric distortion (with respect to the auxiliary
Euclidean metric induced by the coordinate system) is of a smaller order
than the distance to p. We are going to refine this assertion by showing that
6.4. Geometric Meaning of Gaussian Curvature 227

(appropriately chosen) normal coordinates do not distort the metric in the

second order, and the third-order distortions are controlled by the Gaussian
curvature.
There are two ways of looking at metric distortions in normal coordi-
nates. Recall that one family of coordinate lines is formed by geodesics,
and hence the metric along these lines is preserved. Hence we can con-
clude that the metric distortion arises from the difference in the behavior
of the distance between near-by geodesics in the Riemannian and Euclid-
ean spaces. We will see that the divergence of geodesics is governed by a
second-order differential equation whose difference from the corresponding
Euclidean equation is a curvature term. This approach is realized in the
subsection “Jacobi equation”. An advantage of this approach is that gen-
eral definitions of curvature bounds for length spaces are given in the same
terms.
Alternatively, we can recall that the other family of coordinate lines is
orthogonal to geodesic coordinate lines. We will see that this is a family
of equidistant curves, with a natural correspondence between the points
on different equidistant curves along the geodesic coordinate lines. Thus
the metric distortion can be measured by the distortion of the length of
equidistant curves as we move along the geodesic family of coordinate lines.
We will see that the first derivative of length of an equidistant curve is
given by its geodesic curvature, and the second derivative is again governed
by the Gaussian curvature of the metric. This approach is realized in
the subsection “Equidistant families”. An advantage of this approach is
that its generalization to higher dimensions (replacing equidistant lines by
hypersurfaces) is almost straightforward. Unfortunately, there is no obvious
analog of this method in more general length spaces.
Our further geometric conclusions can be derived from either of the
methods.

6.4.1. Jacobi equation. Let us consider a normal coordinate system (x, y)

with coordinate vector fields X, Y . Recall that the x-lines are geodesics
parameterized by arc length and orthogonal to the y-lines. This means that
D
∇X X = dx X = 0 (we use both notations for the covariant derivative to
refresh notations). Let us differentiate this identity with respect to y:
D D
∇Y ∇X X = X = 0.
dy dx
We want to change the order of differentiations. The covariant derivatives
with respect to x and y do not commute, and the curvature operator has
been introduced in the previous section just as their commutator. Thus
R(X, Y )X = ∇Y ∇X X − ∇X ∇Y X = −∇X ∇Y X = −∇X ∇X Y,
228 6. Curvature of Riemannian Metrics

where the last identity follows from axiom (6) of covariant differentiation:
∇X Y = ∇Y X for commuting (coordinate) vector fields X, Y . We will denote
D D D2
∇X ∇X = dx dx by dx2 .
Thus we get the following Jacobi equation, which is a second-order linear
differential equation for Y :
D2
(6.12) Y = −R(X, Y )X.
dx2
Notice that Y is nothing but a variational vector field for a family
of geodesics (the x-lines); it is of no importance that these lines form a
nondegenerate coordinate system, and the same argument can be repeated
verbatim for a general variation. Let us give the corresponding definition:
Definition 6.4.1. A vector field Y (t) along a geodesic γ(t) is called a Jacobi
vector field if it satisfies equation (6.12).
Exercise 6.4.2. Show that every Jacobi field along γ can be represented
as a variational vector field by including γ into a family of geodesics.
Hint: Since a Jacobi vector field Y satisfies a second-order linear differ-
ential equation, it is uniquely determined by its Cauchy data, that is, its
D
value Y (t0 ) and its derivative dt Y (t0 ) at a point t0 . On the other hand, it
is easy to include γ into a geodesic variation whose variation vector field Y1
has the same value and the same derivative at t0 as Y . This implies that
X = Y , for Y has to satisfy the same equation (6.12), since it is a variation
field of a family of geodesics.

Let us consider a Jacobi field Y (t) along a geodesic γ(t). As usual,

denote the velocity vector of γ by T . Denote by g(t) the length of Y (t).
Suppose that g(t) 6= 0. Then one can write g(t) = |Y (t)| = hY (t), N (t)i,
where N = |Y |−1 Y is a unit vector field orthogonal to γ. The function g
characterizes how geodesics from a family with variational vector Y diverge
from γ: the distance between γ(t) and γε (t) is approximately εg(t) (up to
o(ε)).
Remark. Our definition of N assumes that |Y | 6= 0, and it can be smoothly
extended to the closure of the interval where g and Y do not vanish. Hence
all further conclusions in this subsection are valid only as long as t belongs
to this interval.

The following obvious lemma will be quite handy:

Lemma 6.4.3. Covariant derivatives of a unit vector field are orthogonal
to this vector field: if W is a unit vector field, then h∇Z W, W i = 0, where
Z is any vector field.
6.4. Geometric Meaning of Gaussian Curvature 229

Proof. By differentiating the expression hW, W i = 1 we get

0 = Z hW, W i = 2 h∇Z W, W i .
¤

Let us make some preliminary computations. By differentiating the

expression hN, T i = 0, we obtain that
¿ À ¿ À ¿ À
d D D D
0= hN, T i = N, T + T, N = N, T ,
dt dt dt dt
D
since γ is a geodesic and hence by definition dt T = 0. Thus we have
¿ À ¿ À
D D
(6.13) N, T = 0 and N, N = 0,
dt dt
where the last identity follows from Lemma 6.4.3.Since Y is proportional to
D
N , it follows from the latter equation (6.13) that dt hN, T i = 0.
We want to compute the second derivative of g. Using the last equation,
we get
¿ À
dg d D
= hY, N i = Y, N .
dt dt dt
One more differentiation yields
¿ À ¿ 2 À ¿ À
d2 g d D D D D
= Y, N = Y, N + Y, N .
dt2 dt dt dt2 dt dt
Using D
dt Y = D
= dg
dt (gN )
D
dt N + g dt (N ), we can rewrite this identity as
¿ À ¿ 2 À ¿ À
d2 g d D D D D
= Y, N = Y, N + g N, N .
dt2 dt dt dt2 dx dx
Notice that, by the Jacobi equation (6.12),
¿ 2 À
D
Y, N = − hR(T, Y )T, N i
dt2
= − hR(T, gN )T, N i = −g hR(T, N )N, T i = −gK,
where K is the Gaussian (sectional) curvature (and the last equality uses
the fact that T and N are orthogonal unit vectors). Hence we finally get
that the Jacobi equation implies the following differential equation for the
scalar function g:
¿ À
d2 g D D
(6.14) = −g(K + N, N ),
dt2 dx dx
where K = K(γ(t)) is the Gaussian curvature along γ.
230 6. Curvature of Riemannian Metrics

D D
Let us notice now that actually dt N = 0, for dt N is orthogonal to both
T and N , and we are dealing with the two-dimensional case! Hence in the
two-dimensional case equation (6.14) turns into
d2 g
(6.15) = −gK.
dt2
The reason why we postponed this observation till now and derived the more
complicated equation (6.14) is explained in the following remark:
Important remark. Now we have arrived at the point where there
is an essential difference between two-dimensional surfaces and the higher-
dimensional case. In the two-dimensional case the direction of Y is deter-
mined by the fact that Y is orthogonal to T . Hence we were able to transform
the Jacobi equation (6.12) into one scalar equation for the magnitude of Y .
In higher dimensions the Jacobi equation (6.12) involves both the magnitude
and the direction of Y . The higher-dimensional situation is complicated by
the following phenomenon: geodesics from our family “twist around γ”. It
is important to notice that the equation for the length of Y that we obtained
for the two-dimensional case is incorrect in higher dimensions! However, the
main geometrical corollary, the Rauch Comparison Theorem 6.5.1, is still
true. One of the proofs consists of a computation based on decomposing y
with respect to a basis of parallel vector fields along γ, that is, such vector
fields that their covariant derivatives along γ are zero. One such vector field
along γ is T , and in our two-dimensional argument N just happened to be
another parallel vector field along γ. The next section gives an alternative
approach via equidistant variations. This approach can be easily generalized
to higher dimensions, and it can be used instead of the Rauch Comparison
Theorem to prove the Cartan-Alexandrov-Toponogov Comparison Theorem.
Note also that many geometric corollaries can be extracted directly from the
d2 g
inequality ≥ −gK implied by (6.14).
dx2
Equidistant variations. The previous section described the Gaussian cur-
vature as a quantitative way of measuring how a family of geodesics diverges.
There is an alternative approach, which avoids using the Jacobi equation.
It has certain advantages for Riemannian geometry (and in particular it
avoids the complication in higher dimensions that we ran into when deriv-
ing (6.15)). On the other hand, this approach can hardly be used for length
spaces other than manifolds.
The idea of this approach is that, instead of studying a bunch of
geodesics, we will concentrate our attention on the family of lines orthogonal
to them. In a normal coordinate system this is just the family of the y-lines.
The distance between two nearby geodesic lines is approximately the length
of the segment of the orthogonal line enclosed between them. This suggests
6.4. Geometric Meaning of Gaussian Curvature 231

that we study how the length along an orthogonal line changes within its
family. A convenient way of doing so is to consider this family of orthogonal
lines as a variation of a curve (in higher dimensions one studies variations
of hypersurfaces).
Let γε be a family of curves whose variation vector field V is a unit vector
field orthogonal to the curves of the family: hV, V i = 1, and hV, T i = 0
d
(where as usual T = dt γ, and we drop the arguments t, ε to simplify
notation). Such variation is said to be equidistant. The reason is explained
in the following exercise.
Exercise 6.4.4. Prove that, for all sufficiently small ε and for any fixed t0 ,
inf d(γε (t0 ), γ0 (t))} = ε.
t
In other words, points of γε lie at the same distance ε from (the image of)
γ0 .
Hint: Use the fact that the lines σ(ε) = γε (t) are geodesics (see below),
and argue as in the proof of the Gauss Lemma.

One should not think of the curves γε as geodesics, but rather as lines
orthogonal to a family of geodesics. Indeed, statement (ii) of Lemma 6.4.6
will imply that the orthogonal curves σ(ε) = γε (t) are geodesics.
Let us assume that the line γ = γ0 is parameterized by arc length.
According to the formula of the first variation (6.10),
Z b Z b
d
L(γε , a, b) = h−S, V i dt = −kg (t) dt,
dε a a
where S = ∇T T , and kg = h∇T T, V i is the geodesic curvature of γ at t (and
its sign is chosen with respect to the orientation prescribed by V ). Hence
a curve in our family shrinks (if kg > 0) or expands (kg < 0) at a rate
proportional to its curvature.
Exercise 6.4.5. Show that for a variation with a variation field f V , where
V is a unit normal to the curve and f is a scalar function,
Z b
d
L(γε , a, b) = −f (t)kg (t) dt.
dε a

We will need the following computational lemma:

Lemma 6.4.6. In notation before Exercise 6.4.4, one has:
(i) h∇T V, V i = h∇V V, V i = 0;
(ii) ∇V V = 0, (and therefore the lines orthogonal to the family γε are
geodesics);
(iii) h∇T ∇V T, V i = − h∇T V, ∇T V i ;
232 6. Curvature of Riemannian Metrics

(iv) h∇T V, ∇T V i = h∇T T, ∇T T i = kg2 at γ0 .

Proof. (i) This is true just because V is a unit vector field (Lemma 6.4.3).
(ii) Since hV, T i = 0, we have
d
0= hV, T i = h∇V V, T i + hV, ∇V T i
dε
1d
= h∇V V, T i + hV, ∇T V i = h∇V V, T i + hV, V i = h∇V V, T i .
2 dt
Together with (i) this implies that the vector ∇V V is zero, for it is orthogonal
to both V and T .
(iii) Differentiating h∇T V, V i = 0 (which follows from (i)), we get
d
0= h∇T V, V i = h∇T ∇V T, V i + h∇T V, ∇T V i .
dt
Together with ∇T V = ∇V T (since T and V are commuting vector fields)
this implies (iii).
(iv) Since |kg | = |∇T T |, the equality h∇T T, ∇T T i = kg2 is trivial.
Differentiating hV, T i = 0, we get
d
0= hT, V i = h∇T T, V i + hT, ∇T V i .
dt
Since V and T are unit vectors, ∇T T is proportional to V , and ∇T V is
proportional to T , this implies that |∇T T | = |∇T V |, which proves (iv). ¤
Remark. Although it may seem that the argument for (ii) uses the fact
that dim(Ω) = 2 in the same way as in the computation used to derive
(6.15), it is not the case. We leave as an exercise to show that the integral
curves of a variation field V of an equidistant variation of a hypersurface are
geodesics. To do this, one just shows that ∇V V is orthogonal to V and to
all vectors tangent to the hypersurfaces. This is a reason why this approach
is sometimes more convenient in higher dimensions.

Now we want to look at the second derivative of length; that is, we want
to know how kg changes as we change the parameter ε. We have
dkg d
= h∇T T, V i = h∇V ∇T T, V i + h∇T T, ∇V V i = h∇V ∇T T, V i ,
dε dε
since ∇V V = 0 by Lemma 6.4.6, (ii).
Now we want to change the order of differentiations in h∇V ∇T T, V i,
and thus the curvature operator shows up. We get
dkg
= h∇T ∇V T, V i + hR(T, V )T, V i .
dε
6.4. Geometric Meaning of Gaussian Curvature 233

The second term is Gaussian curvature (since V, T form an orthonormal

base), and the first term is −kg2 by Lemma 6.4.6, (iii) and (iv). Thus we
proved the following differential equation for kg (called a “Ricatti equation”).
Lemma 6.4.7.
dkg
= −kg2 + K.
dε
Remark. The statement of Lemma 6.4.7 can be used as an alternative
definition of the Gaussian curvature.
Exercise 6.4.8. Apply Lemma 6.4.7 to the inwards equidistant variation
of a Euclidean circle (remember that the choice of the sign for kg depends
on the direction of the variation). Notice that when ε is equal to the radius
of the circle, kg = ∞ (the circle shrinks to a point).
Exercise 6.4.9. Apply Lemma 6.4.7 to the outwards equidistant variation
of a circle in a plane of constant Gaussian curvature −1. What is the limit
of kg as ε → ∞? Do the same for a circle on a sphere.

The next exercise allows us to compare the geodesic curvatures of

equidistant families in two regions provided that it is known that the Gauss-
ian curvature in one region does not exceed that in the other:
Exercise 6.4.10. Let γε (t), σε (t) be two equidistant families of smooth
curves in Ω, Ω1 (with the variation vectors orthogonal to the curves).
We assume that the variations are defined in the same domain in the
(ε, t)-plane. Assume that the geodesic curvatures of γ0 and σ0 are the same
at t0 , and that the Gaussian curvature in Ω and Ω1 satisfies the inequality
K(γε (t0 )) ≥ K(σε (t0 )) for all ε in the domain of the variations. Prove that,
for every ε from the domain of the variations, the geodesic curvature of γε
is less than or equal to the geodesic curvature of σε at t0 .

This exercise is a counterpart of Theorem 6.5.1, and it can be used

instead of it to prove the main Theorem 6.5.4. Notice that Theorem 6.5.1
has a condition that Y (t) does not vanish; the statement of the theorem
does not hold past the point where Y (t) vanishes (this point is said to be
conjugate to γ(0) along γ). A counterpart of this condition in Exercise 6.4.10
is the assumption that all curves in the equidistant family are smooth; that
is, kg does not become infinite, as happens for the inwards equidistants of
a circle when the equidistants collapse to a point (which is the center of
the circle). A point where an equidistant of a curve acquires a singularity
kg = ∞ is said to be focal.
Exercise 6.4.11. Show that, for a planar curve γ(t), its focal points form
the curve γ(t) + (1/kg (t)) N (t), where N (t) is a unit normal vector field.
The number 1/kg (t) is called the curvature radius of γ at γ(t).
234 6. Curvature of Riemannian Metrics

There is the following fruitful strategy of studying geometry (or topol-

ogy) of a Riemannian manifold with certain curvature bounds using Lemma
6.4.7 (or the result of Exercise 6.4.10): one sweeps through a manifold by
a family of equidistant hypersurfaces (curves in dimension two), and then
compares this family with an appropriate equidistant family in a model space
(usually of constant curvature).

6.4.2. Metric distortion of the exponential map. The results of this

subsection can be based on either of the two previous subsections. We
will use the Jacobi equation, leaving carrying out the other approach as an
exercise.
Let us apply equation (6.15) to the family of the x-lines of a normal
coordinate system centered at p. Recall that our metric takes a very
simple form in normal coordinates: E = hX, Xi = 1, F = hX, Y i = 0,
and G = √hY, Y i is the only nontrivial coefficient. Using the notation
g = |Y | = G, we conclude that
√
∂2 G √
= −K G.
∂x2

This equation describes the divergence of two nearby coordinate lines

from each other.
More generally, if a geodesic γ(t) is included into a family of geodesics
∂
γε (t), γ = γ0 , and the variation vector Y (t) = ∂ε γε (t)|ε=t=0 is orthogonal to
γ̇(0), then the length g(t) of this vector satisfies

d2 g(t)
(6.16) = −K(t)g(t),
dt2
where K(t) = K(γ(t)) is the Gaussian curvature along γ. Recall that this
equation is valid only as long as g(t) does not vanish; vanishing of g(t)
geometrically means that nearby geodesics from the family meet γ (at least
up to the second order).
For a normal coordinate system centered at p we have g(0) = G(0, y) = 0,
for the coordinate vector Y vanishes at p. Notice that dg/dt(0) = 1. This
is a more delicate statement: it follows from Lemma 5.1.13 (two x-lines
emanating from p diverge in the first order at the same rate as two Euclidean
rays emanating from one point and forming the same angle). Hence, by
(6.16), d2 g/dt2 (0) = 0 and therefore g(t) = t + o(t2 ). Plugging this again in
(6.16), we get

d2 g d3 g
(t) = −Kt + o(t2 ) and hence (0) = −K.
dt2 dt3
6.4. Geometric Meaning of Gaussian Curvature 235

In other words, in normal coordinates

p 1
(6.17) G(x, y) = x − Kx3 + o(x3 ).
6
This conclusion gives a very good idea of a geometric meaning of curvature:
in
p normal coordinates the metric is Euclidean up to the second order (since
G(x, y) = x for a Euclidean metric), and the third-order distortion is
measured by K.
Let us give another reformulation of the assertion that a Riemannian
metric is locally Euclidean up to a third order error. Recall that a Rie-
mannian metric locally can be approximated by a Euclidean metric as in
Lemma 5.1.13. This lemma guarantees that one can choose a (coordinate)
map from a Euclidean region to a neighborhood of a given point p in Ω in
such a way that its metric distortion in a ball of (a small) radius r centered
at p is o(r). Now we can essentially refine this result. The tangent space
Tp Ω together with h, ip is a Euclidean space, and expp maps a neighborhood
of the origin in this space to Ω. The following lemma asserts that this map
differs from an isometry by at most a third-order term (in the distance from
p):
Lemma 6.4.12. We have
d(q, s) − | exp−1 −1 3
p (q) expp (s)| = o(max(d(p, q), d(p, s)) ),

where | · | is the distance in (TΩ , h, i).

Proof. Introduce a polar coordinate system (r, ρ) in Tp Ω and choose a

normal coordinate system in a neighborhood of p such that expp maps points
in Tp Ω to points in Ω with the same coordinates: expp (r, ρ) = (x = r, y = ρ).
Then the metric coefficients of the Euclidean metric h, ip in Tp Ω take the form
E = 1, F = 0, G(r, ρ) = r2 , while the metric coefficients of our Riemannian
metric at the corresponding point x = r, y = ρ are E = 1, F = 0, G(r, ρ) =
(r + oρ (r2 ))2 . Now the lemma follows from formula (5.3) for the Riemannian
length of a curve in coordinates. ¤

Let us notice again that the Gaussian curvature measures just the third-
order distortion of metric.
The construction used to prove this lemma is very useful, and it will
be applied (with minor modifications) in this chapter several times! Notice
also that expp does not distort distances from p, and hence it maps spheres
centered at the origin of Tp Ω to spheres centered at p.

Surfaces of constant curvature. Let us consider an important model

situation of a region of a constant Gaussian curvature K(p) = k for all
p ∈ Ω. There are three cases: k = 0, k > 0 and k < 0, which correspond to
236 6. Curvature of Riemannian Metrics

surfaces that are locally isometric to the Euclidean plane, a sphere, and a
hyperbolic plane. Indeed, if k = 0, the Jacobi equation (6.16) tells us that
d2
dt2
g(t) = 0, and hence g(t) is a linear function. Indeed, Euclidean geodesics
are straight lines, and the distance between straight lines is a linear function
(as long as it does not vanish!) If k > 0, one explicitly √ solves the Jacobi
√
equation (6.16), obtaining its general solution g(t) =√A sin k t + B cos k t.
√ k t. It is easy to check
In particular, if g(0) = 0, ġ(0) = 1, we get g(t) = sin
that if one rotates a meridian of a sphere of radius k around its pole at the
unit angular speed, the √speed of a point of this meridian lying at a distance
t from the pole is sin K √ t; in other words, nearby meridians diverge at a
speed proportional to sin√ k t. If k < √ 0, the general solution of equation
(6.16) is g(t) = A sinh −k t + B cosh −k t. This formula describes the
divergence of a bunch of lines in the hyperbolic plane of curvature k. Let us
formalize these considerations:

Lemma 6.4.13. Let K(q) = k for all q ∈ Ω. Let p ∈ Ω. Then there

is a neighborhood of p that is isometric to a region in the Euclidean plane
(k
√ = 0), the hyperbolic plane of curvature k (k < 0), or the sphere of radius
k (k > 0)

We consider the case k = 0; the two other cases are similar. The
argument is very similar to the proof of Lemma 6.4.12. Choose a normal
coordinate system (x, y) centered at p. Let γ(t)p= (t, y) be a coordinate
line emanating from p. Denote, as usual, g(t) = G(t, y). As in the proof
of Lemma 6.4.12, g(0) = 0 and dg dt (0) = 1. Solving the Jacobi equation
(6.16) subject to the initial conditions g(0) = 0, g ′ (0) = 1, we immediately
get g(t) = t, and hence G(x, y) = y 2 . Comparing these expressions for
metric coefficients E, F, G with the metric coefficients of the Euclidean
metric in polar coordinates (r, ρ) (formula (5.7)), we see that the map
(x, y) → (r = x, ρ = y) is an isometry to a neighborhood of the origin
in the Euclidean plane in polar coordinates. ¤

Exercise 6.4.14. Using formulas (5.8) and (5.9), repeat the same argument
to show that a surface of constant curvature is locally isometric to a sphere
or a hyperbolic plane.

Remark. Notice that we proved that the Gaussian curvature of the hyper-
bolic plane of curvature k is k; moreover, our argument also implies that the
hyperbolic planes are homogeneous: notice that we constructed an isometry
mapping any given point in a region of constant negative curvature to the
origin of the polar coordinates used in formula (5.9)!
6.5. Comparison Theorems 237

6.5. Comparison Theorems

6.5.1. Rauch Comparison Theorem. The following (two-dimensional
case of the) Rauch Comparison Theorem (or its analog Exercise 6.4.10)
plays a crucial role for our metric considerations.
Theorem 6.5.1. Let γ1 (t), γ0 (t) be geodesics in Ω1 , Ω0 respectively. Let
Y1 (t), Y0 (t) be Jacobi fields along γ1 and γ0 such that
D D
Y1 (0) = Y0 (0) = 0 and |
Y1 (0)| = | Y0 (0)| = 1.
dt dt
Let us assume that Y1 does not vanish on an interval [0, T ], and that
the Gaussian curvature K1 (t) = K(γ1 (t)) of Ω1 along γ1 is less than or
equal to the Gaussian curvature K0 (t) = K(γ0 (t)) of Ω0 along γ0 , that is,
K0 (t) ≤ K1 (t) for all t ∈ [0, T ].
Then |Y1 (t)| ≤ |Y0 (t)| for all 0 ≤ t ≤ T .

Proof. Let g1 (t) = |Y1 (t)|, g0 (t) = |Y0 (t)|. Then g1 (t) satisfies the equation
d2 g1
(t) = −K1 (t)g1 (t) subject to g1 (0) = 0, ġ1 (0) = 1,
dt2
and g0 (t) satisfies the similar equation.
We want to prove that g1 (t) ≤ g0 (t) for all 0 ≤ t ≤ T .
The idea is to consider a function ϕ(t) = gg10 (t) (t)
and prove that it
is (non-strictly) monotone increasing on the interval ]0, T [. Then, since
limt→0 ϕ(t) = 1 by the L’Hopital rule, it will follow that ϕ(t) ≥ 1 and hence
g0 (t) ≥ g1 .
To prove the monotonicity of ϕ, it suffices to verify that ϕ̇(t) ≥ 0 for
all t. We have
ġ0 (t)g1 (t) − g0 (t)ġ1 (t)
ϕ̇(t) = .
g1 (t)2
Denote the numerator of the last formula by ψ(t). Since the denominator
g1 (t)2 is positive, we have to prove that ψ(t) ≥ 0. Observe that ψ(0) = 0
because g0 (0) = g1 (0) = 0. So again it suffices to prove that ψ̇(t) ≥ 0 for
all t. From the equations for g0 and g1 one gets
ψ̇ = g̈0 g1 − g0 g̈1 = (K1 − K0 )g0 g1 .
Thus ψ̇ > 0 wherever g0 ≥ 0 (recall that g1 > 0 on ]0, T [ by assumption of
the Theorem).
Thus, it remains to show that g0 does not change the sign on ]0, T [.
Suppose the contrary, and let t0 ∈ ]0, T [ be the first point where g0 vanishes.
Then restrict the above argument to the interval ]0, t0 ]: since g0 (t) ≥ 0 for
all t ∈ ]0, t0 ], one has ψ̇(t) ≥ 0 for all t ∈ ]0, t0 ], therefore g0 ≥ g1 on
238 6. Curvature of Riemannian Metrics

]0, t0 ]. In particular, g0 (t0 ) ≥ g1 (t0 ) > 0, contrary to the assumption that

g0 (t0 ) = 0. The theorem follows. ¤
Remark. Notice that this proof is purely two-dimensional, for it uses
equation (6.15) instead of (6.14); see the remark after equation (6.15).
Remark. The condition that the Jacobi field does not vanish for t ≤ T is
essential. Two points γ(a) and γ(b) such that there is a nonzero Jacobi field
vanishing at both points are said to be conjugate along γ. A segment of
a geodesic that contains a pair of conjugate points in the interior is never
a shortest path between its endpoints. (A nice proof uses the index form,
which is beyond the scope of this book. A hand-crafted proof is tedious,
though elementary.) Hence such segments are of no interest for distance
measurements. Conversely, a geodesic segment that does not contain a pair
of conjugate points is always a locally shortest path (it is the shortest path
among sufficiently close paths). This statement can be easily proven arguing
as in the proof of Lemma 5.2.9.

6.5.2. Cartan-Alexandrov-Toponogov Comparison Theorem. Now

we are going to exploit a construction which is very similar to the one used
in the proof of Lemma 6.4.13. Consider two exponential maps expp : Up →
Ωp = expp (Up ) and expq : Uq → Ωq = expq (Uq ), where Up and Uq are chosen
small enough so that both exponential maps are diffeomorphisms.
It is convenient to assume that Up and Uq are balls of the same radius
r. Choose a linear isometry I : (Tq Ωq , h, iq ) → (Tp Ωp , h, ip ). Clearly I sends
Uq to Up . Define σ = expp ◦I ◦ exp−1
q , σ : Ωq → Ωp .
Alternatively, one can choose normal coordinate systems centered at p
and q, and define σ to map every point from Ωq to the point in Ωp with the
same coordinates.
Let us list some obvious properties of σ. First of all, σ preserves the
radial distances:
dΩq (q, s) = dΩp (p, σ(s)).
Indeed, σ maps geodesics starting from q and parameterized by arc length
to naturally parameterized geodesics starting from p.
Next, by construction dq σ is a linear isometry between tangent spaces,
and hence σ preserves angles between curves starting from q (notice that,
since we do not specify whether we mean Riemannian angles or the general
notion of an angle between two curves in a length space, we implicitly use
Lemma 5.1.14).
Exercise 6.5.2. Reasoning as in Lemmas 6.4.12 and 6.4.13, show that if
K(p) = K(q), then σ distorts the metric in no more than the fourth order
(as we move away from q).
6.5. Comparison Theorems 239

Exercise 6.5.3. Use Exercise 5.1.19 to show that the length of a circle of
radius r centered at p is equal to πr2 (1 − 61 K(p)r2 + o(r2 )).

The proof of Lemma 6.4.13 was based on the observation that, for a
region Ω of constant curvature k, σ happens to be an isometry. We will
generalize this as the following property of σ, which perhaps best explains
the geometric meaning of Gaussian curvature:
Theorem 6.5.4. If the Gaussian curvature of Ω satisfies k ≤ K(s) for all
s ∈ Ω, then σ is a nonexpanding map in some neighborhood of q; that is,
there exists an ε > 0 such that
d(σ(s1 ), σ(s2 )) ≤ d(s1 , s2 ) for all s1 , s2 ∈ Bε (q).
If the Gaussian curvature of Ω satisfies k ≥ K(s) for all s ∈ Ω, then σ
is a noncontracting map in some neighborhood of q; that is,
d(σ(s1 ), σ(s2 )) ≥ d(s1 , s2 ) for all s1 , s2 ∈ Bε (q).

Proof. Let us choose normal coordinates centered at q and p such that σ

maps every point in Ωq to the point in Ωp with the same coordinates. The
Riemannian metrics in Ω and Ωk are given by given by
q
(6.18) hV, V i = vx2 + G(x, y)vy2 in Ω,
q
(6.19) hV, V i = vx2 + Gk (x, y)vy2 in Ωk ,
where G and Gk are the metric coefficients of the metrics in Ω and Ωk .
p
As usual, denote by g(t) = G(t, y) the length of the coordinate vector
field Y , p
which is a Jacobi field along a coordinate line in Ω. Similarly
gk (t) = Gk (t, y) is the length of the coordinate Jacobi field Yk along a
coordinate line in Ωk .
Consider the case k ≥ K(s) for all s ∈ Ω; the other one is similar.
Applying the Rauch Comparison Theorem 6.5.1 to Y and Yk , we conclude
that G(x, y) ≥ Gk (x, y). Now it is obvious from (6.18) and (6.19) that σ
does not increase lengths of curves. Now it remains to choose ε small enough
so that a shortest curve between any two points in Bε (q) is contained in the
domain Ωq of σ (it is enough to choose ε = 21 rq , where rq is the injectivity
radius at q). ¤
Exercise 6.5.5. Give another proof of this theorem that uses Lemma 6.4.7
and Exercise 6.4.10 instead of Theorem 6.5.1

Recall that every point p of a region Ω has a neighborhood U such that

there is only one shortest path between any two points in U (Exercise 5.2.11
guarantees this).
240 6. Curvature of Riemannian Metrics

Let Ωk be the Euclidean plane if k = 0, the sphere of the Gaussian

curvature k if k > 0, and the hyperbolic plane of curvature k if k < 0.
Let a, b, c be three points in U , and let a′ , b′ , c′ be three points in Ωk such
that d(a, c) = d(a′ , c′ ), d(b, c) = d(b′ , c′ ) and the angle between the shortest
paths [ca] and [cb] at c is equal to that between the shortest paths [c′ a′ ] and
[c′ b′ ] at c′ .
Together with the fact that σ preserves angles at p, Theorem 6.5.4
immediately implies that a Riemannian metric in Ω enjoys the following
version of the Cartan-Alexandrov-Toponogov Comparison Theorem for
small triangles:
Theorem 6.5.6. Let U , Ωk and triangles △abc, △a′ b′ c′ be as above. Then
d(a, b) ≥ d(a′ , b′ ) provided that the Gaussian curvature in Ω satisfies
k ≥ K(s) for all s ∈ Ω, and
d(a, b) ≤ d(a′ , b′ ) if the Gaussian curvature in Ω satisfies k ≤ K(s) for
all s ∈ Ω.

Proof. The possibility to choose I in the construction of σ leaves us enough

freedom to have σ map a to a′ and b to b′ . Now the theorem follows from
Theorem 6.5.4. ¤
Exercise 6.5.7. (i) Show that if the Gaussian curvature in a region is
nonnegative (nonpositive), then every angle of a triangle is less than or equal
to (resp. greater than or equal to) the corresponding angle of a Euclidean
triangle with the same lengths of sides.
(ii) Show that if the Gaussian curvature in a region is nonnegative
(nonpositive), then the sum of the angles in every triangle is at most (resp.
at least) π.

The comparison property given by Theorem 6.5.6 is taken as (one among

several) definition of the length spaces of curvature bounded below and
above. Hence we have made the necessary preparation for our synthetic ma-
chinery by showing that the Riemannian manifolds with bounded sectional
curvature are length spaces with the same curvature bounds (in the sense
of Chapter 4).
 Chapter 7

Space of Metric Spaces

Most of our discussions in this course are about relations between various
properties and characteristics of a metric space. In other words, we study one
metric space at a time. However it may be useful to consider every particular
space as a representative of a class of similar objects, which could be called
the “space of metric spaces”. There is a similar idea in the foundation of the
mathematical analysis: instead of talking about a single number, one studies
the entire real line. This brings new notions like continuity, derivatives,
etc., and they appear to be powerful tools that give new information about
the original objects, numbers. For example, a common method of proving
inequalities is finding maxima and minima, and these notions make no sense
without considering a set of numbers rather than a single number.
Another example of using this “global” approach is given by the theory
of convex sets in Rn . Introducing Hausdorff distance (discussed in Sub-
section 7.3.1 below) turns the set of all compact convex sets into a metric
space. This opens a way to apply “analytic” techniques to convex sets just
like numbers. For example, one can use maxima and minima because the
space of compact convex sets is boundedly compact just like R (cf. Theo-
rem 7.3.8). Also, a simple argument shows that polyhedra are dense in this
space, and this allows us to extend certain statements “by continuity” from
polyhedra to arbitrary convex sets.
In this chapter we extend this approach further and introduce a distance
between abstract metric spaces. In fact, we will define several distances
suitable for different purposes. Let us make some remarks in advance.
First, in most cases the distance itself is not essential; what matters is the
topology that it defines. Concerning the “space of metric spaces”, we will
mainly study converging sequences and similar notions, not exact numerical

241
242 7. Space of Metric Spaces

values of the distance. Second, to make use of the topological structure one
usually needs a collection of quantities that are continuous, properties that
are preserved through passing to the limit, etc. The finer the topology is, the
more such quantities and properties exist. However, the topology being finer
means that there are fewer converging sequences and fewer compact sets,
and this makes the topology less useful. For this reason different topologies
may be needed. This is illustrated by functional analysis where every natural
class of functions comes with its own topology—recall C 0 , C 1 , L1 , L2 and
so on.
To ignite curiosity, we begin with a mathematical fairy tale, or a science
fiction story—for this preparation level. It was a famous open problem
whether every group of polynomial growth is virtually nilpotent (since this
is a fairy tale, we do not need to define either of the notions). Here is an
idea of Gromov’s solution of this problem. Gromov suggested considering
a sequence of metric spaces (G, hd) where G is our finitely generated group
with a word metric d, and h is a positive number tending to zero. Using
his criterion for pre-compactness of certain spaces of metric spaces, Gromov
specified a subsequence converging to some length space, along with an
action of the group on this space (a reader may think of the example (Z2 , hd),
where d is the word metric for the standard choice of two generators,
as a sequence of finer and finer grids converging to R2 with the norm
||(x, y)|| = |x| + |y|). The limit space happened to be a Finsler manifold.
Thus Gromov was able to reduce the conjecture to the case of an isometry
group of a manifold, and in this case the answer was to be affirmative.

7.1. Examples
We start with several examples, without formal definitions and proofs. Not
all these examples are covered by notions discussed later in this course; some
of them will be used for motivations and comparisons only.

7.1.1. Tangent cone of a convex set. Consider a closed convex set in

the plane. If this set is essentially two-dimensional, i.e., not contained in a
line, its boundary is a continuous curve. Playing with examples one makes
the following observations, which are not difficult to formalize and prove.
There are two kinds of points on a convex curve. There are points where
the curve is “smooth” in the sense that it admits a unique tangent line,
and points where the curve “breaks” and has different tangents at two sides.
Similarly, a three-dimensional convex body near a point of its boundary may
look (up to the first order) like a half-space, or like a dihedral angle with
an edge, or like a “sharp” solid angle. Furthermore, a dihedral angle has
a certain angular measure, and sharp angles may vary in shape. All these
7.1. Examples 243

first-order properties are encoded in the tangent cone, which is defined as

follows.
Let X be a convex set in Rn and p ∈ X. For every λ > 0, consider the
homothety with the coefficient λ centered at p. Assuming for convenience
that p is the origin of Rn , one obtains the dilated set λX = {λx : x ∈ X}.
The tangent cone of X at p is by definition the limit of these sets as λ → ∞.
Since X is convex, it is fairly easy to define what is meant by a limit here.
Namely, the family of sets {λX} is increasing in the sense that λ1 X ⊂ λ2 X
if λ1 < λ2 (this trivially follows
S from convexity). So these sets naturally
“converge” to their union, λ>0 λX. It is more convenient to have a closed
cone, so one usually takes the closure instead of a bare union.S Thus the
tangent cone of X at p is defined as the closure of the set λ>0 λX, where
the homotheties X 7→ λX are centered at p. It is easy to check that it is
indeed a cone (i.e., it consists of entire rays emanating from p, or equivalently
is invariant under homotheties). In fact, it is the minimal convex cone (with
its vertex at p) that contains X. Note that p can be an internal point of the
set; in this case the tangent cone is the entire space. The construction of the
tangent cone is pictured in Figure 7.1 (dilated spaces have lighter color).

PSfrag replacements
p

Figure 7.1: Spaces with dilated metrics converge to the tangent cone.

The tangent cone is closely related to the “space of directions” intro-

duced in Subsection 3.6.6. Namely, a ray emanating from p is contained in
the tangent cone if and only if its direction belongs to the space of direc-
tions at p. This gives an alternative definition for the space of directions.
It is more convenient in many cases, because some properties of a set can
be automatically transferred from a set to its tangent cone “by continuity”.
A trivial example is the property of convexity. This kind of definition of a
tangent cone (and a space of direction) can be applied to other spaces, not
only convex sets; however one needs a suitable definition of limit.

7.1.2. Asymptotic cone. Tangent cones grasp the behavior of a set in

a small neighborhood of a point. There is a similarly defined notion of
244 7. Space of Metric Spaces

asymptotic cone that does this “near infinity”. Given a (closed) convex set
X ⊂ Rn and a point p ∈ X, the asymptotic cone is the limit of homothetic
sets λX as λ → 0 (the homotheties are centered at p). As in the case of
tangent cone, there is no problem with a definition of the limit. T Since {λX}
is a nested family of sets, the “limit” is just their intersection, λ>0 λX (see
Figure 7.2). It is easy to see that the asymptotic cone is just the union of
all rays emanating from p and contained in X. The shape of the asymptotic
cone does not depend on the choice of p in the sense that asymptotic cones
with different choices for p are parallel translations of one another.

PSfrag replacements
p p

Figure 7.2: Spaces with dilated metrics converge to the asymptotic cone.

Loosely speaking, the asymptotic cone is what one sees observing the
set from far away (assuming that moving the observation point stretches
the visual image homothetically). A similar association applies to tangent
cones: the tangent cone at p is what an observer sees looking at p under
a microscope. The surface of the Earth looks planar to us; indeed, the
tangent cone of the sphere is a plane (more precisely, that of a solid ball is
a half-space).
We will extend the above definition of the asymptotic cone to general
metric spaces in the next chapter (Section 8.2).

7.1.3. Asymptotic cone of a lattice. Consider the group Z2 which can

be thought of as the lattice of integer points in R2 . In Subsection 3.2.3
we described how to define a word metric for any given set of generators.
Namely, the word metric is the intrinsic metric of the group’s Cayley graph.
(More precisely, it is the restriction of that metric to the group, but we
prefer to keep the graph in this example.) Take the simplest possible set
of generators, namely two generators (1, 0) and (0, 1). Then the Cayley
graph is naturally identified with the grid of horizontal and vertical lines
passing through entire points in the plane. The intrinsic metric of the graph
coincides with the one induced from the plane.
Let’s try to apply (informally) the definition of the asymptotic cone to
this graph. Let the graph be denoted by X. The dilated space λX is the grid
of all vertical and horizontal lines whose coordinates are integer multiplies of
7.1. Examples 245

λ. In other words, these lines divide R2 into squares of side λ. As λ goes to

zero, the grid becomes finer and finer, and it is natural to think that its limit
should include all points in the plane. (This is only a speculation appealing
to the reader’s visual intuition. We don’t have a suitable definition of limit
yet.)
However, the limit as a metric space is not the Euclidean plane R2
because the distances in the grids are essentially different from the Euclidean
one. In fact, the limit is the normed space R21 which is R2 with the norm
k(x, y)k1 = |x| + |y|. Indeed, looking at the shortest paths in the grid λX
one can see that the distance between two nodes (x1 , y1 ) and (x2 , y2 ) equals
|x1 − x2 | + |y1 − y2 |. It is the same as the distance in R21 . And for a small
λ the distance between nonnodes is only slightly different (by no more than
4λ) from the distance in R21 (because for every point in the grid there is a
node within a distance ≤ λ from it). Thus if we want the distances in the
grid to converge (in any sense) to the distances in the limit space, we have
to define the latter as R21 .

PSfrag replacements A
A
B

Figure 7.3: A fine grid and a shortest path.

In the above considerations we used an isometric embedding of the

Cayley graph into R2 . Such an embedding may fail to exist for other sets of
generators, but a reasonable asymptotic cone can be defined for every word
metric. In fact, the asymptotic cone is always a two-dimensional normed
space. Its unit ball can be identified with a polygon in R2 obtained as a
convex hull of the set of generators and vectors opposite to generators. We
will discuss this further in Subsection 8.5.1.

7.1.4. Converging surfaces. Smooth surfaces in R3 can be (at least

locally) parameterized by a domain D ⊂ R2 , either as graphs of smooth
functions from D to R, or as images of embeddings from D to R3 . Hence a
246 7. Space of Metric Spaces

topology on a set of surfaces can be derived from a topology on the set of

maps. Namely, a sequence of surfaces converges if a sequence of their suitable
parameterizations do. Of course, if a surface is covered by several patches,
each with its own parameterizations, certain compatibility conditions should
be imposed. Anyway, one can speak, for instance, about C k -convergence of
surfaces (k = 0, 1, . . . ).
Remark 7.1.1. If some characteristic of a surface is defined in terms of
derivatives up to the kth order, one usually has to choose at least the C k
topology in order to make the characteristic continuous. However, imposing
geometric restrictions sometimes allows one to use a weaker topology.
For example, consider regular smooth curves in the plane. The length
and curvature (which are defined in terms of first and second derivatives,
respectively) are not continuous with respect to C 0 topology. The length is
lower semi-continuous, and the curvature is not semi-continuous from either
side. But if the class of curves is restricted to convex ones (a planar curve is
called convex if it is contained in a boundary of a convex set), the situation
improves. Namely on the set of convex regular curves equipped with C 0
topology, the length is continuous and the minimum of curvature is upper
semi-continuous.
This is a partial case of a general phenomenon: convexity and lower
curvature bounds are surprisingly stable with respect to, say, passing to a
limit in a relatively weak topology (like the C 0 one in the example with
convex curves).
Exercise 7.1.2. Prove the statements about regular curves from the above
remark.

A “functional” convergence discussed above is hardly suitable for our

purposes since we mainly deal with abstract metric spaces rather than
subsets of Rn . One can modify this notion using the following idea. If two
surfaces are parameterized by the same domain, the two parameterizations
determine a one-to-one correspondence between points of the surfaces. In
other words, they define a homeomorphism from one of them to the other.
If the surfaces are close enough to each other, this homeomorphism only
slightly changes certain data such as distances, metric tensors, or their
derivatives. One can turn this to a definition saying that two spaces have
small distance between them if there is a homeomorphism from one to
the other which “almost preserves” certain geometric characteristics, for
example the distance. Later we will give several precise definitions of this
sort.
This approach has a disadvantage: it requires spaces to be homeomor-
phic. Sometimes this requirement is too restrictive. Let X be the standard
7.1. Examples 247

two-dimensional sphere and Xn be same sphere with a small handle attached

to it. (Let the diameter of the handle be less than 1/n.) As n grows, handles
become smaller and smaller and spaces Xn look more and more similar to
X (see Figure 7.4). One could say that handles vanish to a point and thus
Xn converge to X. However Xn is not homeomorphic to X, so a different
notion of convergence is needed here.

Figure 7.4: Spheres with vanishing handles.

Furthermore, even for two homeomorphic and similarly looking spaces,

it may happen that the “similarity” is not realized by a homeomorphism.
For example, consider two spheres of essentially different radii connected by
a long thin tube. One can attach a tiny little handle to either the larger
or the smaller sphere; let X and Y be the resulting spaces (see Figure 7.5).
Then X and Y are homeomorphic, but any homeomorphism (in fact, any
continuous map) from X to Y essentially distorts distances between some
points, no matter how small the handles are. Nevertheless X and Y are
“close” to each other in the same sense as in the previous example.

Figure 7.5: A homeomorphism has to change distances essentially.

The reader already familiar with Hausdorff distance may notice that it
provides some way to formalize the above examples: the distance between
these sets is small because each of them is contained in a small neighborhood
of the other in R3 . However, if we consider two surfaces as length spaces
with close intrinsic metrics, and it is not at all clear that these metrics can
be realized by close embedding (even if they can be realized by embeddings
248 7. Space of Metric Spaces

at all). Intrinsic geometry of a surfaces is in a complicated relation with its

“visual image” in R3 and measuring the distance in R3 makes little sense.
Nevertheless, measuring the distance in an ambient space is a step in the
right direction. Later in this chapter we will introduce a very important
Gromov–Hausdorff distance between metric spaces. As the name suggests,
this notion is derived from the classic Hausdorff distance.

7.1.5. Uniform convergence. Recall that a sequence {fn } of real-valued

functions on a set X is said to uniformly converge to a function f if
sup |fn (x) − f (x)| → 0
x∈X

as n → ∞. Since every metric on X is a real-valued function (defined on

X × X), the notion of uniform convergence applies to metrics: a sequence
{dn } of metrics on X uniformly converges to a metric d (or, dn ⇒ d) if
sup |dn (x, x′ ) − d(x, x′ )| → 0
x,x′ ∈X

as n → 0. We have already seen (Exercise 2.4.19) that passing to a uniform

limit of metrics preserves the property of being intrinsic: if the metrics
dn are intrinsic, the limit metric d is intrinsic too. Nevertheless, uniform
convergence is a relatively weak type of convergence. For example, a uniform
limit of Riemannian metric can be non-Riemannian.
Exercise 7.1.3. Let D2 denote the standard unit ball in R2 . Prove that
the metric of R21 (that is, R2 with the norm k(x, y)k = |x| + |y|) restricted
to D2 can be obtained as a uniform limit of Riemannian metrics.
Hint: Fill the cells in the fine grid from Subsection 7.1.3.

In fact, any intrinsic metric on D2 can be obtained as a uniform limit of

Riemannian metrics. In view of this, it seems unbelievable that a uniform
convergence preserves some information about the curvature. However it
does: if the curvatures of converging metrics are uniformly bounded from
below, this curvature bound (in the Alexandrov sense) is inherited by the
limit metric! In fact, this holds for a more general kind of convergence,
namely, the Gromov–Hausdorff one. We will prove this later in the course.
The uniform convergence of metrics has one unnatural restriction: the
metrics should be defined on the same set. Taking into account the un-
derlying set on which a metric is defined would make us distinguish metric
spaces that are otherwise identical (i.e., isometric). To avoid this, one could
introduce the following (preliminary) definition of uniform convergence of
metric spaces: a sequence {Xn } of metric spaces uniformly converges to a
metric space (X, d) if there is a sequence {dn } of metrics on X such that
(X, dn ) is isometric to Xn for all n, and dn ⇒ d.
7.2. Lipschitz Distance 249

This definition is “preliminary” because it can be reformulated in a more

convenient way. Our formulation, when saying that (X, dn ) is isometric to
Xn , implicitly includes an isometry map from Xn to (X, dn ). Constructing
an isometry for two metrics (that are known to be isometric) may be difficult,
whereas the metric dn can be trivially recovered from such an isometry map.
This observation suggests that our definition would be more usable if it dealt
with maps from Xn to X rather than with metrics on X.
Definition 7.1.4. Let X and Y be metric spaces and f : X → Y an
arbitrary map. The distortion of f is defined by
dis f = sup |dY (f (x1 ), f (x2 )) − dX (x1 , x2 )|
x1 ,x2 ∈X
where dX and dY are the metrics of X and Y .
The definition of distortion resembles that of dilatation of a Lipschitz
map. The only difference is that the latter measures relative change of
distances while the former measures absolute ones. The notion of distortion
can be applied to noncontinuous maps.
Definition 7.1.5. We say that a sequence {Xn } of metric spaces uniformly
converges to a metric space X if there exist homeomorphisms fn : Xn → X
such that dis(fn ) → 0 as n → ∞.
Exercise 7.1.6. Prove that Definition 7.1.5 is equivalent to our “prelimi-
nary” definition of uniform convergence.
Exercise 7.1.7. Define a distance between metric spaces such that the
convergence with respect to this distance is equivalent to the uniform con-
vergence in the sense of Definition 7.1.5.
Hint: See the definition of Lipschitz distance in the next section.

7.2. Lipschitz Distance

The idea of Lipschitz distance is the following: two metric spaces X and Y
are close to each other if there is a homeomorphism f : X → Y such that
the ratios dY (f (x), f (x′ ))/dX (x, x′ ) are close to 1. In other words, Lipschitz
distance measures relative difference between metrics. Note that relative
differences and relative errors is just what people care about concerning the
metric of the physical universe. It is a good achievement to find the distance
between the Sun and the Earth with an error of about a thousand miles,
but measuring one’s apartment with such a precision is not a good idea.
Let X and Y be two metric spaces. Recall that the dilatation of a
Lipschitz map f : X → Y is defined by
dY (f (x), f (x′ ))
dil f = sup
x,x′ ∈X dX (x, x′ )
250 7. Space of Metric Spaces

where dX and dY are the metrics of X and Y . A homeomorphism f is called

bi-Lipschitz if both f and f −1 are Lipschitz maps.
Definition 7.2.1. The Lipschitz distance dL between two metric spaces X
and Y is defined by
¡ ¢
dL (X, Y ) = inf log max{dil(f ), dil(f −1 )}
f : X→Y
where the infimum is taken over all bi-Lipschitz homeomorphisms f : X →
Y.
A sequence {Xn }∞n=1 of metric spaces is said to converge in the Lipschitz
sense to a metric space X if dL (Xn , X) → 0 as n → ∞.
If there are no bi-Lipschitz homeomorphisms from X to Y , then one sets
dL (X, Y ) = ∞. Thus the Lipschitz distance is not suitable for comparing
metric spaces that are not (bi-Lipschitz) homeomorphic.
Example 7.2.2. Let M be a smooth manifold and {Fn }∞ n=1 be a sequence
of Finslerian structures on M . Assume that it converges to a Finslerian
structure F in the sense that Fn (v)/F (v) → 1 as n → ∞ uniformly over
all nonzero vectors v ∈ T M (recall that Finslerian structures are functions
on T M ). Then dL ((M, dn ), (M, d)) → 0 where dn and d are the Finslerian
metrics corresponding to Fn and F .
Example 7.2.3. Let {Mt }t∈R be a family of smooth surfaces in R3 para-
meterized by maps ft : Ω → R3 where Ω is a region in R2 . Assume that
the family itself is smooth (at least C 1 ); i.e., the function F : Ω × R → R3
defined by F (x, t) = ft (x) is smooth. Then dL (Mt , M0 ) → 0 as t → 0. This
conclusion may fail if the family is only continuous.
Theorem 7.2.4. dL is nonnegative, symmetric and satisfies the triangle
inequality. For compact spaces X and Y , dL (X, Y ) = 0 if and only if X and
Y are isometric.

Proof. If f : X → Y is a homeomorphism, then at least one of the dilata-

tions dil f and dil f −1 is greater than or equal to 1 because f and f −1 cannot
both decrease the distances. Hence dL (X, Y ) ≥ 0. The symmetry of dL is
trivial.
To prove the triangle inequality, let X, Y , Z be metric spaces and
f : X → Y and g : Y → Z be bi-Lipschitz homeomorphisms. Then h =
g ◦ f is a bi-Lipschitz homeomorphism from X to Z. Moreover dil(h) ≤
dil(f ) · dil(g); hence log(dil h) ≤ log(dil f ) + log(dil g). Together with the
similar inequality for h−1 = f −1 ◦ g −1 , this implies the triangle inequality
dL (X, Z) ≤ dL (X, Y ) + dL (Y, Z).
If X and Y are isometric, substituting an isometry f : X → Y in the
definition of dL shows that dL (X, Y ) = 0. The converse statement is more
7.3. Gromov–Hausdorff Distance 251

delicate. By the definition of dL , the relation dL (X, Y ) = 0 implies that

there exists a sequence of maps fn : X → Y , n ∈ N, such that dil fn → 1
and dil(f −1 ) → 1 as n → ∞. Since the dilatations of fn are uniformly
bounded, by the Arzela–Ascoli Theorem a subsequence of {fn } uniformly
converges to a map f : X → Y . We may assume (without loss of generality)
that the sequence fn itself converges to f . Since dil fn → 1 as n → ∞, for
all x, x′ ∈ X one has |fn (x)fn (x′ )|/|xx′ | → 1 and hence |f (x)f (x′ )| = |xx′ |.
Thus f is a distance-preserving map. Similarly, there is a distance-preserving
map g : Y → X. The composition f ◦ g is a distance-preserving map from Y
to itself. Since Y is compact, the map f ◦ g is bijective by Theorem 1.6.14.
Hence f is surjective and therefore is an isometry. ¤

This theorem tells us that the Lipschitz distance is a metric on the

“space” of isometry classes of compact metric spaces. This “space” is not a
good object from the rigorous point of view of the Set Theory (just like the
“set of all sets”). However we will refer to this space in our formulations.
For us, this space is just a collection of elements (which are isometry classes
of metric spaces) and not an object of the Set Theory, so no set-theoretic
paradoxes can occur. Furthermore, all our statements about this “space of
metric spaces” can be reformulated in terms of its elements (like the above
theorem), so one can think of this space as a way to shorten formulations.
Remark 7.2.5. Alternatively, one can formally justify “the space of com-
pact metric spaces” using the following observation: this “space” is of car-
dinality continuum, i.e., there can be no more than continuum of mutually
nonisometric compact spaces. (Exercise: prove this.) Therefore, choosing a
representative from each isometry class one obtains a collection of cardinal-
ity continuum, which is a legitimate set.
Exercise 7.2.6. Prove that Lipschitz convergence of compact metric spaces
implies uniform convergence (Definition 7.1.5).
Exercise 7.2.7. Prove that Lipschitz convergence is equivalent to the uni-
form convergence within the class of finite metric spaces (that is, consisting
of finitely many points).
Remark 7.2.8. The two types of convergence are not equivalent in general.
There are sequences of compact metric spaces that converge uniformly but
do not converge w.r.t. the Lipschitz distance (can you give an example?)

7.3. Gromov–Hausdorff Distance

The Gromov–Hausdorff distance is similar to the Lipschitz distance in
the sense that it is a distance between compact metric spaces, where the
latter ones are considered up to an isometry. However it determines a
252 7. Space of Metric Spaces

weaker “topology”; in particular, the distance is always finite (and can be

arbitrarily small) even for nonhomeomorphic spaces. The difference between
the Lipschitz and the Gromov–Hausdorff distances is somewhat similar to
the one between C 1 and C 0 norms in functional spaces.

7.3.1. Hausdorff distance. First we introduce the ordinary Hausdorff

distance. This is a distance between subsets of a metric space, not between
abstract metric spaces.
We denote by Ur (S) the r-neighborhood of a set S in a metric space,
i.e.,
S the set of points x such that dist(x, S) < r. Equivalently, Ur (S) =
x∈S Br (s).

Definition 7.3.1. Let A and B be subsets of a metric space. The Hausdorff

distance between A and B, denoted by dH (A, B), is defined by
dH (A, B) = inf{r > 0 : A ⊂ Ur (B) and B ⊂ Ur (A)}.

The next exercise contains convenient reformulations of the definition.

Exercise 7.3.2. Let A and B be subsets of a metric space and r > 0.
1. Prove that dH (A, B) = max{supa∈A dist(a, B), supb∈B dist(b, A)}.
2. Prove that dH (A, B) ≤ r if and only if dist(a, B) ≤ r for all a ∈ A
and dist(b, A) ≤ r for all b ∈ B.
3. Show that the previous statement would fail if one replaces ≤ by <.
Proposition 7.3.3. Let X be a metric space. Then
(1) dH is a semi-metric on 2X (the set of all subsets of X).
(2) dH (A, A) = 0 for any A ⊂ X where A denotes the closure of A.
(3) If A and B are closed subsets of X and dH (A, B) = 0, then A = B.

Proof. 1. Nonnegativity and symmetry are obvious. The triangle inequality

follows from the following fact: for any A ⊂ X and any r1 , r2 > 0 one has
Ur1 +r2 (A) ⊂ Ur1 (Ur2 (A)). This is in turn an immediate consequence of the
triangle inequality in X.
2. dist(x, A) = 0 for all x ∈ A because S ⊂ A. For an x ∈ A one has
dist(x, A) = 0 by the definition of closure. Hence dH (A, A) = 0.
3. Suppose the contrary, e.g., there is an x ∈ A \ B. Since B is a closed
set, there is an r > 0 such that the ball Br (x) does not intersect B. Then
x∈/ Ur (B); hence dH (A, B) ≥ r > 0. ¤

Let M(X) denote the set of closed subsets of X equipped with Hausdorff
distance. The above proposition tells us that the M(X) is a metric space.
Moreover every element of the quotient 2X/dH can be represented by a closed
set and therefore 2X/dH is naturally identified with M(X).
7.3. Gromov–Hausdorff Distance 253

Exercise 7.3.4. Let a sequence of sets Ai ∈ M(X) converge to a set

A ∈ M(X) with respect to the Hausdorff distance (in short: Ai → A in
M(X)). Prove that
1. A is the set of limits of all converging sequences {an } in X such that
an ∈ An for all n.
T S∞
2. A = ∞ n=1 (closure of m=n Am ).

Exercise 7.3.5. Let X be a compact metric space and {Ai }∞

i=1 a sequence
of its compact subspaces. Prove that:
T 1. If Ai+1 ⊂ Ai for all i, then {Ai } converges in M(X) to the intersection
i Ai .
i ⊂ Ai+1 for all i, then {Ai } converges in M(X) to the closure of
2. If AS
the union i Xi .
Exercise 7.3.6. Let Ai → A in M(Rn ) and all sets Ai are convex. Prove
that A is convex. In other words, the set of compact convex sets is closed
in M(Rn ).
Proposition 7.3.7. If X is complete, then M(X) is complete.

Proof. Let {Sn }∞ n=1 be a Cauchy sequence in M(X). Let S denote the set
of all points x ∈ X such that for any neighborhood U of x one has U ∩Sn 6= ∅
for infinitely many n. We will prove that Sn → S. Fix ε > 0 and let n0
be such that dH (Sn , Sm ) < ε for all m, n ≥ n0 . It suffices to show that
dH (S, Sn ) < 2ε for any n ≥ n0 .
1. dist(x, Sn ) < 2ε for every x ∈ S. There exists an m ≥ n0 such
that Bε (x) ∩ Sm 6= ∅. In other words, there is a point y ∈ Sm such that
|xy| < ε. Since dH (Sm , Sn ) < ε, one has dist(y, Sn ) < ε and therefore
dist(x, Sn ) ≤ |xy| + dist(y, Sn ) < 2ε.
2. dist(x, S) < 2ε for every x ∈ Sn . Let n1 = n and for every integer
k > 1 chose an index nk such that nk > nk+1 and dH (Sp , Sq ) < ε/2k for
all p, q ≥ nk . Then define a sequence of points {xk }, where xk ∈ Snk , as
follows. Let x1 = x, and xk+1 be a point of Snk+1 such that |xk xk+1 | < ε/2k
for all k. Such a point can be found because dH (Snk , Snk+1 ) < ε/2k . Since
P ∞
k=1 |xk xk+1 | < 2ε < ∞, the sequence {xk } is a Cauchy sequence
P and hence
it converges to a point y ∈ X. Then |xy| = lim |xxn | ≤ |xk xk+1 | < 2ε.
Since y ∈ S by construction, it follows that dist(x, S) < 2ε. ¤
Theorem 7.3.8 (Blaschke). If X is compact, then M(X) is compact.

Proof. By the previous proposition, M(X) is complete. Therefore it suffices

to prove that M(X) is totally bounded. Let S be a finite ε-net in X. We
will prove that 2S is an ε-net in M(X). Let A ∈ M(X). Consider the set
254 7. Space of Metric Spaces

SA ∈ 2S defined by
SA = {x ∈ S : dist(x, A) ≤ ε}.
Since S is an ε-net in X, for every y ∈ A there exists an x ∈ S such that
|xy| ≤ ε. Since dist(x, A) ≤ |xy| ≤ ε, this point x belongs to SA . Therefore
dist(y, SA ) ≤ ε for all y ∈ A. Since dist(x, A) ≤ ε for any x ∈ SA (by the
definition of SA ), it follows that dH (A, SA ) ≤ ε. Since A is arbitrary, this
proves that 2S is an ε-net in M(X). ¤

Remark 7.3.9. In the theory of convex sets, the following statement

is known as Blaschke’s Theorem: the set of all compact convex subsets
contained in any fixed closed ball in Rn is compact with respect to the
Hausdorff distance. This follows from Theorem 7.3.8 and Exercise 7.3.6.

7.3.2. Gromov–Hausdorff distance. Now we are in position to define

the Gromov–Hausdorff distance between metric spaces. The idea behind
the definition is the following. First, the distance between subspaces in
the same metric space is no greater than the Hausdorff distance between
them. In other words, if two subspaces of the same space are close to each
other in the sense of Hausdorff distance in the ambient space, they must be
close to each other as abstract metric spaces. Second, one definitely wants
the distance between isometric spaces to be zero. The Gromov–Hausdorff
distance is in fact the maximum distance satisfying these two requirements.

Definition 7.3.10. Let X and Y be metric spaces. The Gromov–Hausdorff

distance between them, denoted by dGH (X, Y ), is defined by the following
relation. For an r > 0, dGH (X, Y ) < r if and only if there exist a metric
space Z and subspaces X ′ and Y ′ of it which are isometric to X and Y
respectively and such that dH (X ′ , Y ′ ) < r. In other words, dGH (X, Y ) is
the infimum of positive r for which the above Z, X ′ and Y ′ exist. Here dH
denotes the Hausdorff distance between subsets of Z.

Note that X ′ and Y ′ in the above definition are regarded with the
restriction of the metric of the ambient space Z (as opposed to the induced
intrinsic metric). For example, if X is a sphere with its standard Riemannian
metric, one cannot take Z = R3 and X ′ = S 2 ⊂ R3 because X and X ′ would
be only path-isometric but not isometric.
It is trivial that the Gromov–Hausdorff distance between isometric
spaces is zero. Later we will show that dGH is a metric in the same sense as
dL ; i.e., it is a metric on the space of the isometry classes of compact met-
ric spaces (Theorem 7.3.30). Naturally one says that a sequence {Xn }∞ n=1
of (compact) metric spaces converges in the Gromov–Hausdorff sense to a
(compact) metric space X if dGH (Xn , X) → 0. For noncompact spaces
7.3. Gromov–Hausdorff Distance 255

a slightly more general notion of convergence is used; we will define it in

Section 8.1.
Example 7.3.11. If Y is an ε-net in a metric space X, then dGH (X, Y ) ≤ ε.
Indeed, one can take Z = X ′ = X and Y ′ = Y .
Remark 7.3.12. Definition 7.3.10 deals with a huge class of metric spaces,
namely, all metric spaces Z that contain subspaces isometric to X and Y .
It is possible to reduce this class to disjoint unions of X and Y . More
precisely, the Gromov–Hausdorff distance between two metric spaces (X, dX )
and (Y, dY ) is the infimum of r > 0 such that there exists a (semi-)metric
d on the disjoint union X ∪ Y such that the restrictions of d to X and Y
coincide with dX and dY and dH (X, Y ) < r in the space (X ∪ Y, d). In
other words, dGH (X, Y ) = inf{dH (X, Y )}, where the infimum is taken over
all (semi-)metrics on X ∪ Y extending the ones of X and Y .
To prove this, simply identify X ∪ Y with X ′ ∪ Y ′ ⊂ Z (the notation
is from Definition 7.3.10). More formally, fix isometries f : X → X ′ and
g : Y → Y ′ , then define the distance between x ∈ X and y ∈ Y by
d(x, y) = dZ (f (x), g(y)). This yields a semi-metric on X ∪ Y for which
dH (X, Y ) < r (if X ′ ∩ Y ′ 6= ∅, it may happen that d(x, y) = 0). The
quotient metric space (X ∪ Y )/d is isometric to X ′ ∪ Y ′ . To obtain a metric
(not a semi-metric) on X ∪ Y , define d(x, y) = dZ (f (x), g(y)) + δ where δ is
an arbitrary positive constant. Then dH (X, Y ) < r + δ.
Exercise 7.3.13. Prove that dGH (X, Y ) < ∞ if X and Y are bounded
metric spaces.
Hint: Choose a sufficiently large constant C > 0 and set d(x, y) = C for
all x ∈ X and y ∈ Y .
Exercise 7.3.14. Let X and Y be two metric spaces and diam X < ∞.
Prove that dGH (X, Y ) ≥ 12 | diam X − diam Y |.
In other words, the diameter, as a function of a metric space, is a
Lipschitz function with a Lipschitz constant 2.
Exercise 7.3.15. Let P be a metric space consisting of one point. Prove
that dGH (X, P ) = diam(X)/2 for any metric space X.
Proposition 7.3.16. dGH satisfies the triangle inequality, i.e.,
dGH (X1 , X3 ) ≤ dGH (X1 , X2 ) + dGH (X2 , X3 )
for any metric spaces X1 , X2 , X3 .

Proof. Let d12 and d23 be metrics on X1 ∪ X2 and X2 ∪ X3 , respectively,

that extend metrics of X1 , X2 and X3 . Define the distance between x1 ∈ X1
and x3 ∈ X3 by d13 (x1 , x3 ) = inf x2 ∈X2 {d12 (x1 , x2 ) + d23 (x2 , x3 )}. It is easy
256 7. Space of Metric Spaces

to check (exercise!) that d13 , together with the metrics of X1 and X3 ,

satisfies the triangle inequality and hence is a metric on X1 ∪ X3 . The
definition of d13 yields that dH (X1 , X3 ) ≤ dH (X1 , X2 ) + dH (X2 , X3 ), where
dH (Xi , Xj ) is taken with respect to the metric dij (i, j = 1, 2, 3). Taking
the infimum over all metrics d12 and d23 we obtain the desired inequality
dGH (X1 , X3 ) ≤ dGH (X1 , X2 ) + dGH (X2 , X3 ). ¤

7.3.3. Reformulations. A direct application of the definition of Gromov–

Hausdorff distance requires constructing a new metric space Z (or a metric
on X ∪Y ) and verifying the triangle inequality. As the reader could observe,
this involves cumbersome details even in simple cases. It would be more
convenient if we could compute or estimate dGH (X, Y ) by comparing the
distances within X and Y to each other, as we did in the cases of uniform
and Lipschitz distance. We will give several criteria of this sort. The first
and the most general one is based on the notion of correspondence.
Roughly speaking, having a correspondence between two metric spaces
(or just sets) X and Y means that for every point of X there are one or
more “corresponding” points in Y , and vice versa. The criterion that we
are going to prove is the following: dGH (X, Y ) < r if and only if there is a
correspondence between X and Y such that if x, x′ ∈ X and y, y ′ ∈ Y are
corresponding pairs of points, then |dX (x, x′ ) − dY (y, y ′ )| < 2r. The precise
definitions and formulations follow.
Definition 7.3.17. Let X and Y be two sets. A correspondence between
X and Y is a set R ⊂ X × Y satisfying the following condition: for every
x ∈ X there exists at least one y ∈ Y such that (x, y) ∈ R, and similarly for
every y ∈ Y there exists an x ∈ X such that (x, y) ∈ R.

In other words, a correspondence is just a relation between points of X

and Y such that every point of X and Y is in the relation to at least one
point in the other set. If R is clear from the context, we will say that x
and y “correspond” to each other instead of writing (x, y) ∈ R.
Example 7.3.18. Any surjective map f : X → Y be defines a correspon-
dence R between X and Y given by
R = {(x, f (x)) : x ∈ X}.
We will call this R the correspondence associated with f .

Not every correspondence is associated with a map. A closer analog

of a correspondence is a “multiple-valued” map, for which a single point is
allowed to have more than one image. We will not discuss multiple-valued
maps. There is an alternative method, given in the next example, to define
a correspondence by means of (ordinary) maps.
7.3. Gromov–Hausdorff Distance 257

Example 7.3.19. Let f : Z → X and g : Z → Y be two surjective maps

from some set Z. Then a correspondence R can be defined by
R = {(f (z), g(z)) : z ∈ Z}.
Exercise 7.3.20. Prove that any correspondence can be obtained via the
construction from Example 7.3.19.
Definition 7.3.21. Let R be a correspondence between metric spaces X
and Y . The distortion of R is defined by
dis R = sup{|dX (x, x′ ) − dY (y, y ′ )| : (x, y), (x′ , y ′ ) ∈ R}
where dX and dY are the metrics of X and Y respectively.
Exercise 7.3.22. Prove that for a correspondence R associated with a map
f : X → Y as in Example 7.3.18, one has dis R = dis f . (The distortion of
a map was introduced in Definition 7.1.4.)
Exercise 7.3.23. Let R be obtained from maps f : Z → X and g : Z → Y
as in Example 7.3.19. Prove that
dis R = sup |dX (f (z), f (z ′ )) − dY (g(z), g(z ′ ))|.
z,z ′ ∈Z

Exercise 7.3.24. Let R be a correspondence between metric spaces X and

Y . Prove that dis R = 0 if and only if R is associated with an isometry map
from X to Y .
Theorem 7.3.25. For any two metric spaces X and Y ,
1
dGH (X, Y ) = inf (dis R)
2 R
where the infimum is taken over all correspondences R between X and Y .
In other words, dGH (X, Y ) is equal to the infimum of r > 0 for which
there exists a correspondence between X and Y with dis R < 2r.

Proof. 1. For any r > dGH (X, Y ), there exists a correspondence R with
dis R < 2r. Indeed, since dGH (X, Y ) < r, we may assume that X and Y are
subspaces of some metric space Z and dH (X, Y ) < r in Z. Define
R = {(x, y) : x ∈ X, y ∈ Y, d(x, y) < r}
where d is the metric of Z. That R is a correspondence follows from the
fact that dH (X, Y ) < r. The desired estimate dis R < 2r follows from the
triangle inequality: if (x, y) ∈ R and (x′ , y ′ ) ∈ R, then
|d(x, x′ ) − d(y, y ′ )| ≤ d(x, y) + d(x′ , y ′ ) < 2r.

2. dGH (X, Y ) ≤ 12 dis R for any correspondence R. Let dis R = 2r. To

avoid confusion, we use the notation dX and dY for the metrics of X and Y ,
258 7. Space of Metric Spaces

respectively. It suffices to show that there is a semi-metric d on the disjoint

union X ∪ Y such that d|X×X = dX , d|Y ×Y = dY , and dH (X, Y ) ≤ r in
(X ∪ Y, d). The idea is to set the distance from x to y equal to r whenever
x and y correspond to each other, and then take the minimal metric d
generated by this condition. This is formally achieved as follows: for an
x ∈ X and y ∈ Y define

d(x, y) = inf{dX (x, x′ ) + r + dY (y ′ , y) : (x′ , y ′ ) ∈ R}

(the distances within X and Y are already defined by dX and dY ). Verifying

the triangle inequality for d and the fact that dH (X, Y ) ≤ r is left as an
exercise to the reader. ¤

Exercise 7.3.26. Let X, Y and Z be metric spaces, R1 be a correspondence

between X and Y , and R2 be a correspondence between Y and Z. A
composition of R1 and R2 , denoted by R1 ◦ R2 , is set of all (x, z) ∈ X × Z
for which there is a y ∈ Y such that (x, y) ∈ R1 and (y, z) ∈ R2 .
1. Prove that R1 ◦ R2 is indeed a correspondence.
2. Prove that dis(R1 ◦ R2 ) ≤ dis R1 + dis R2 .
3. Using the above inequality, give an alternative proof of the triangle
inequality for the Gromov–Hausdorff distance.

The next corollary gives us more techniques for handling the Gromov–
Hausdorff distances. While it does not give explicit expressions for the
distance, it provides another quantity which differs from the distance by
no more than two times. Note that an estimate of this type is sufficient
to study the topology (on the space of metric spaces) determined by the
Gromov–Hausdorff distance.

Definition 7.3.27. Let X and Y be metric spaces and ε > 0. A (possibly

noncontinuous!) map f : X → Y is called an ε-isometry if dis f ≤ ε and
f (X) is an ε-net in Y .

Corollary 7.3.28. Let X and Y be two metric spaces and ε > 0. Then
1. If dGH (X, Y ) < ε, then there exists a 2ε-isometry from X to Y .
2. If there exists an ε-isometry from X to Y , then dGH (X, Y ) < 2ε.

Proof. 1. Let R be a correspondence between X and Y with dis R < 2ε.

For every x ∈ X, choose f (x) ∈ Y such that (x, f (x)) ∈ R. This defines
a map f : X → Y . Obviously dil f ≤ dil R < 2ε. Let us show that
f (X) is a 2ε-net in Y . For a y ∈ Y , consider an x ∈ X such that
(x, y) ∈ R. Since both y and f (x) are in correspondence with x, one has
d(y, f (x)) ≤ d(x, x) + dis R < 2r. Hence dist(y, f (X)) < 2r.
7.3. Gromov–Hausdorff Distance 259

2. Let f be an ε-isometry. Define R ⊂ X × Y by

R = {(x, y) ∈ X × Y : d(y, f (x)) ≤ ε}.
Then R is a correspondence because f (X) is an ε-net in Y . If (x, y) ∈ R
and (x′ , y ′ ) ∈ R, one has
|d(y, y ′ ) − d(x, x′ )| ≤ |d(f (x), f (x′ )) − d(x, x′ )| + d(y, f (x)) + d(y ′ , f (x′ ))
≤ dis f + ε + ε ≤ 3ε.

Hence dis R ≤ 3r, and Theorem 7.3.25 implies dGH (X, Y ) ≤ 32 r < 2r. ¤
Remark 7.3.29. It is important that we do not require continuity of ε-
isometries. Even if two spaces are very close with respect to the Gromov–
Hausdorff distance, it can happen that there are no continuous maps with
small distortion—recall the spheres with small handles in Subsection 7.1.4.

Now we are in a position to prove the analog of Theorem 7.2.4 for

the Gromov–Hausdorff distance. Note that, unlike Lipschitz distance, the
Gromov–Hausdorff one defines a finite metric.
Theorem 7.3.30. Gromov–Hausdorff distance defines a finite metric on
the space of isometry classes of compact metric spaces. In other words,
it is nonnegative, symmetric and satisfies the triangle inequality; moreover
dGH (X, Y ) = 0 if and only if X and Y are isometric.

Proof. We have already proven all statements of this theorem except the
claim that dGH (X, Y ) = 0 implies that X and Y are isometric. Let X and Y
be two compact spaces such that dGH (X, Y ) = 0. By Corollary 7.3.28, there
exists a sequence of maps fn : X → Y such that dis fn → 0. Fix a countable
dense set S ⊂ X. Using the Cantor diagonal procedure, one can choose a
subsequence {fnk } of {fn } such that for every x ∈ S the sequence {fnk (x)}
converges in Y . Without loss of generality we may assume that this holds for
{fn } itself. Then one can define a map f : S → Y as the limit of fn , namely,
set f (x) = lim fn (x) for every x ∈ S. Since |d(fn (x), fn (y)) − d(x, y)| ≤
dis fn → 0, we have d(f (x), f (y)) = lim d(fn (x), fn (y)) = d(x, y) for all
x, y ∈ S. In other words, f is a distance-preserving map from S to Y . Then
f can be extended to a distance-preserving map from the entire X to Y
(by Proposition 1.5.9). Now we can finish the proof in the same way as in
Theorem 7.2.4. Namely there is a similar distance-preserving map from Y
to X, and it follows that X and Y are isometric. ¤

The above theorem allows us to consider compact metric spaces as points

in the so-called Gromov–Hausdorff space, keeping in mind that isometric
spaces represent the same “point”. The topology of this space (determined
260 7. Space of Metric Spaces

by the Gromov–Hausdorff distance) is called the Gromov–Hausdorff topol-

ogy. (The set-theoretic remarks that we made after Theorem 7.2.4 apply
here as well.)
Exercise 7.3.31. Prove the following generalization of Theorem 7.3.30: if
X and Y are metric spaces with dGH (X, Y ) = 0, X is compact and Y is
complete, then X and Y are isometric.
Hint: Show that Y is compact. To do that, construct a finite ε-net in Y
for every ε > 0.

7.4. Gromov–Hausdorff Convergence

In this section we consider converging sequences in the Gromov–Hausdorff
space of compact metric spaces. By definition, a sequence {Xn }∞
n=1 of com-
pact metric spaces converges to a compact metric space X if dGH (Xn , X) →
0 as n → ∞. In this case, we will write Xn −→
GH X and call X the Gromov–
Hausdorff limit of {Xn }. Since dGH is a metric (Theorem 7.3.30), the limit
is unique up to an isometry.

7.4.1. Examples and properties.

Example 7.4.1 (Hausdorff convergence). For subspaces in the same metric
space, Gromov–Hausdorff distance by definition is not greater than the
Hausdorff distance. Thus Hausdorff convergence of subsets of a given space
implies Gromov–Hausdorff convergence (but not vice versa).

Corollary 7.3.28 yields the following criterion for Gromov–Hausdorff

convergence: a sequence {Xn } of metric spaces converges to a metric space
X if and only if there are a sequence {εn } of numbers and a sequence of
maps fn : Xn → X (or, alternatively, fn : X → Xn ) such that every fn is
an εn -isometry and εn → 0. This observation includes “functional” types of
convergence in our list of examples.
Example 7.4.2 (uniform convergence). If a sequence {Xn } of metric spaces
uniformly converges (in the sense of Definition 7.1.5) to a metric space X,
then Xn −→
GH X.

Example 7.4.3 (Lipschitz convergence). As we have seen in Exercise 7.2.6,

the Lipschitz convergence is a particular case of uniform convergence, and
hence a particular case of the Gromov–Hausdorff convergence. In other
words, the Gromov–Hausdorff topology is weaker than the Lipschitz one.
Example 7.4.4 (uniform convergence to a semi-metric). Let {dn } be a
sequence of metrics on a fixed set X which converges uniformly to some
function d : X ×X → R. Then d is obviously a semi-metric, and the quotient
7.4. Gromov–Hausdorff Convergence 261

metric space X/d (cf. Proposition 1.1.5) is the Gromov–Hausdorff limit of

the spaces (X, dn ).
Indeed, the distortions of the projection X → X/d with respect to the
metrics dn go to zero.
Note that if X is a finite set, it suffices to require that dn (x, y) → d(x, y)
for every pair x, y ∈ X because a pointwise convergence of functions on a
finite set implies uniform convergence.
Exercise 7.4.5. Prove that any converging sequence {Xn } of finite metric
spaces of a fixed cardinality N can be represented in the form given in the
above example. In other words, there is a sequence of metrics {dn } on a fixed
set X such that (X, dn ) is isometric to Xn and {dn } uniformly converges to
a semi-metric.
In particular, the limit of {Xn } is a finite space of cardinality no greater
than N .
Exercise 7.4.6. Prove that a sequence {Xn } converges to a one-point space
if and only if diam(Xn ) → 0.

The next exercise gives a general form of a sequence converging to a

finite space.
Exercise 7.4.7. Let {Xn }∞ n=1 be a sequence of metric spaces, and let X be
a finite metric space of cardinality N , X = {xi : 1 ≤ i ≤ N }.
1. Assume that Xn −→ GH X. Prove that, for all sufficiently large n, the
cardinality of Xn is at least N .
2. Prove that Xn −→ GH X if and only if the following holds. For all
sufficiently large n, Xn can be split into a disjoint union of N nonempty sets
Xn,1 , Xn,2 , . . . , Xn,N so that for all i, j
diam(Xn,i ) → 0, dist(Xn,i , Xn,j ) → |xi xj | (n → ∞).
Exercise 7.4.8. Let N be a fixed natural number. Prove that the Lipschitz,
uniform, and Gromov–Hausdorff distances determine the same topology on
the class of finite metric spaces of cardinality N . Give an explicit description
of this topology.

The reader may wonder why we pay attention to the trivial case of finite
spaces. One of the reasons is that finite spaces form a dense set in the
Gromov–Hausdorff space:
Example 7.4.9. Every compact metric space X is a limit of finite spaces.
Indeed, take a sequence εn → 0 of positive numbers and choose a finite
εn -net Sn in X for every n. Then Sn −→
GH X, simply because dGH (X, Sn ) ≤
dH (X, Sn ) ≤ εn .
262 7. Space of Metric Spaces

Moreover, taking appropriate ε-nets one can essentially reduce conver-

gence of arbitrary compact metric spaces to convergence of their finite sub-
sets. The details follow.
Definition 7.4.10. Let X and Y be two compact metric spaces, and ε, δ > 0
We say that X and Y are (ε, δ)-approximations of each other if there exist
finite collections of points {xi }N N
i=1 and {yi }i=1 in X and Y , respectively, such
that:
(1) The set {xi : 1 ≤ i ≤ N } is an ε-net in X, and {yi : 1 ≤ i ≤ N } is
an ε-net in Y .
(2) |dX (xi , xj ) − dY (yi , yj )| < δ for all i, j ∈ {1, . . . , N }.
An ε-approximation is a synonym for (ε, ε)-approximation (i.e., we omit
the δ if δ = ε).
Proposition 7.4.11. Let X and Y be compact metric spaces.
(1) If Y is an (ε, δ)-approximation of X, then dGH (X, Y ) < 2ε + δ.
(2) If dGH (X, Y ) < ε, then Y is a 5ε-approximation of X.

Proof. 1. Let X0 = {xi }N N

i=1 and Y0 = {yi }i=1 be as in Definition 7.4.10. The
second condition in the definition means that the natural correspondence
{(xi , yi ) : 1 ≤ i ≤ N } between X0 and Y0 has distortion less than δ. It
follows that dGH (X0 , Y0 ) < δ/2. Since X0 and Y0 are ε-nets in X and Y ,
respectively, we have dGH (X, X0 ) ≤ ε and dGH (Y, Y0 ) ≤ ε. The statement
follows by the triangle inequality for dGH .
2. By Corollary 7.3.28, there is a 2ε-isometry f : X → Y . Let X0 =
{xi }N
i=1 be an ε-net in X and yi = f (xi ). Then |d(xi , xj ) − d(yi , yj )| < 2ε <
5ε for all i, j. It remains to prove that Y0 = {yi : 1 ≤ i ≤ N } is a 5ε-net
in Y . Pick a y ∈ Y . Since f (X) is a 2ε-net in Y , there is an x ∈ X such
that d(y, f (x)) ≤ 2ε. Since X0 is an ε-net in X, there exists an xi ∈ X0
such that d(x, xi ) ≤ ε. Then
d(y, yi ) = d(y, f (xi )) ≤ d(y, f (x)) + d(f (x), f (xi ))
≤ 2ε + d(x, xi ) + dis f ≤ 2ε + ε + 2ε ≤ 5ε.
Hence dist(y, Y0 ) ≤ d(y, yi ) ≤ 5ε. ¤

The above proposition yields a criterion for convergence: Xn −→ GH X if

and only if, for any ε > 0, Xn is an ε-approximation of X for all large enough
n. There is a more elegant formulation of this kind:
Proposition 7.4.12. For compact metric spaces X and {Xn }∞ n=1 , Xn −→
GH
X if and only if the following holds. For every ε > 0 there exist a finite
ε-net S in X and an ε-net Sn in each Xn such that Sn −→
GH S.
7.4. Gromov–Hausdorff Convergence 263

Moreover these ε-nets can be chosen so that, for all sufficiently large n,
Sn have the same cardinality as S.

Proof. If such ε-nets exist, then Xn is an ε-approximation of X for all

sufficiently large n. Then Xn −→ GH X by the previous proposition. To
prove the converse implication, take a finite (ε/2)-net S in X and construct
corresponding nets Sn in Xn . Namely, pick a sequence of εn -approximations
fn : X → Xn where εn → 0 and define Sn = fn (S). Then Sn −→ GH S and, as
in the previous proposition, Sn is an ε-net in Xn for all large enough n. ¤

Let us mention one important implication of this construction. Let

Xn −→GH X and S be a finite subset of X (not necessary an ε-net for a
small ε). As in the above proof, construct sets Sn ⊂ Xn corresponding
to S. Then Sn converge to S; i.e., the distances in Sn converge to the
corresponding distances in S. It follows that all (reasonable) geometric
characteristics of the sets Sn converge to those of S. This opens a way to
prove various continuity statements about the Gromov–Hausdorff space: if
some property of spaces Xn can be formulated in terms of finite collections
of points, then this property is automatically inherited by the limit space X.
To see how this abstract scheme works, consider the property of a
metric to be intrinsic. We have a midpoint criterion (Theorem 2.4.16)
expressing this property in terms of triples of points. Since triples of points
in converging spaces correspond to almost isometric triples in the limit space,
it follows that a limit of compact length spaces is a length space. We will
repeat this proof in more formal words in Section 7.5 (see Theorem 7.5.1).

7.4.2. Compactness theorem. Since the Gromov–Hausdorff topology is

a relatively weak one (compared to, say, the topology of the Lipschitz
distance), one may expect that it has relatively many compact sets. In-
deed, many natural classes of metric spaces form (pre-)compact sets in the
Gromov-Hausdorff space. In this section we do not prove any statements of
this sort, but give a general criterion for pre-compactness.
Proposition 7.4.12 implies that members of a sequence {Xn } converging
in the Gromov–Hausdorff space must contain ε-nets of uniformly bounded
cardinality (for every given ε > 0). It follows that, if a class X of metric
spaces is pre-compact in the Gromov–Hausdorff topology, then for every
ε > 0 the size of a minimal ε-net is uniformly bounded over all elements of
X. It turns out that this property of X, along with the fact that the diameters
of its members are uniformly bounded, is sufficient for pre-compactness.
Definition 7.4.13. We say that a class X of compact metric spaces is
uniformly totally bounded if
(1) There is a constant D such that diam X ≤ D for all X ∈ X.
264 7. Space of Metric Spaces

(2) For every ε > 0 there exists a natural number N = N (ε) such that
every X ∈ X contains an ε-net consisting of no more than N points.

Exercise 7.4.14. Prove that the first condition of the above definition is
redundant (i.e., is implied by the second one) if all elements of X are length
spaces.

Theorem 7.4.15. Any uniformly totally bounded class X of compact metric

spaces is pre-compact in the Gromov–Hausdorff topology. That is, any
sequence of elements of X contains a converging subsequence.

Proof. Let D and N (ε) be as in Definition 7.4.13. Define N1 = N (1) and

Nk = Nk−1 + N (1/k) for all k ≥ 2. Let {Xn }∞ n=1 be a sequence of metric
spaces from X. In every space Xn , consider a union of (1/k)-nets for all
k ∈ N. This is a countable dense collection Sn = {xi,n }∞ i=1 ⊂ Xn such that
for every k the first Nk points of Sn form a (1/k)-net in Xn . The distances
|xi,n xj,n | do not exceed D, i.e., belong to a compact interval. Therefore,
using the Cantor diagonal procedure, we can extract a subsequence of {Xn }
in which {|xi,n xj,n |}∞
n=1 converge for all i, j. To simplify the notation, we
assume that they converge without passing to a subsequence.
Now let us construct the limit space X̄ for {Xn }. Pick an abstract
countable set X = {xi }∞
i=1 and define a semi-metric d on X by

d(xi , xj ) = lim |xn,i xn,j |.

n→∞

The quotient construction from Proposition 1.1.5 gives us a metric space

X/d. We will denote by x̄i the point of X/d obtained from xi . This quotient
space may not be complete, so let X̄ be the completion of X/d. We will
prove that {Xn } converges to X̄.
For a k ∈ N , consider the set S (k) = {x̄i : 1 ≤ i ≤ Nk } ⊂ X̄. It is a
(k)
(1/k)-net in X̄. Indeed, every set Sn = {xi,n : 1 ≤ i ≤ Nk } is a (1/k)-net
in the respective space Xn . Hence for every xi,n ∈ Sn there is a j ≤ Nk such
that |xi,n xj,n | ≤ 1/k. Since Nk do not depend on n, for every fixed i ∈ N
there is a j ≤ Nk such that |xi,n xj,n | ≤ 1/k for infinitely many indices n.
Passing to the limit we obtain that |x̄i x̄j | ≤ 1/k for this j. Thus S (n) is a
(1/k)-net in X/d and hence in X̄. Since X̄ is complete and has a (1/k)-net
for any k ∈ N , it is compact.
(k)
Furthermore, the set S (k) is a Gromov–Hausdorff limit of the sets Sn as
n → ∞, because these are finite sets consisting of Nk points (some of which
may coincide) and the distances converge. Thus for every k ∈ N we have
a (1/k)-net in X̄ which is a Gromov–Hausdorff limit of some (1/k)-nets in
the spaces Xn . By Proposition 7.4.12 it follows that Xn −→
GH X̄. ¤
7.5. Convergence of Length Spaces 265

Exercise 7.4.16. Let M n be a compact Riemannian manifold. Prove that

for any sufficiently small ε > 0 there is an ε-net in M containing no more
than ε−n C(n) Vol(M ) points (for some constant C(n) independent of M ).
Here Vol denotes the Riemannian volume.

The result of the above exercise does not mean that any class of compact
Riemannian n-manifolds with uniformly bounded volumes is pre-compact
in the Gromov–Hausdorff topology. The catch is in the phrase “sufficiently
small”— the actual bound for being “small” depends on M .

Exercise 7.4.17. Give an example of a set of two-dimensional compact

Riemannian manifolds with areas no greater than 1 which is not pre-compact
in the Gromov–Hausdorff topology.

Information. Here are some important examples (without proofs) of pre-

compact classes of Riemannian manifolds.
Bounded volume and injectivity radius. For any n ∈ N and any r, V > 0,
the class of all n-dimensional Riemannian manifolds with volume ≤ V and
injectivity radius ≥ r, is pre-compact.
Bounded diameter and curvature. For any n ∈ N and any κ ∈ R, D > 0,
the class of all n-dimensional Riemannian manifolds with diameter ≤ D and
sectional curvature ≥ κ is pre-compact. We will prove this in Chapter 10
as a part of a more general statement about Alexandrov spaces (Theorem
10.7.2). Moreover, the same is true with Ricci curvature instead of sectional
one.

7.5. Convergence of Length Spaces

We already mentioned that a Gromov–Hausdorff limit of length spaces is
a length space. In other words, the length spaces form a closed set in the
Gromov–Hausdorff topology. Here we state this important fact as a theorem
and give a more detailed proof (which is nothing but a formalization of the
argument given at the end of Subsection 7.4.1).

Theorem 7.5.1. Let {Xn }∞

n=1 be a sequence of length spaces, X a complete
metric space, and Xn −→
GH X. Then X is a length space.

Proof. By the criterion for a complete length metric (Theorem 2.4.16), it

suffices to prove that any two points x, y ∈ X possess an ε-midpoint for any
ε > 0. Let n be such that dGH (X, Xn ) < ε/10. Then, by Theorem 7.3.25,
there is a correspondence R between X and Xn whose distortion is less than
ε/5. Take points x̃, ỹ ∈ Xn corresponding to x and y. Since Xn is a length
space, there is a z̃ ∈ Xn which is an (ε/5)-midpoint for x̃ and ỹ. Let z ∈ X
266 7. Space of Metric Spaces

be a point corresponding to z̃. Then

¯ ¯ ¯ ¯
¯|xz| − 1 |xy|¯ ≤ ¯|x̃z̃| − 1 |x̃ỹ|¯ + 2 dis R < ε/5 + 2ε/5 < ε.
2 2
¯ ¯
Similarly ¯|yz| − 21 |xy|¯ < ε. Thus z is an ε-midpoint for x and y. ¤

7.5.1. First examples. One should be careful when verifying convergence

of length spaces, because distances in length metrics is a complicated subject.
This warning message is explained by the following exercises, which should
help the reader to understand what kind of difficulties can arise.
Exercise 7.5.2. 1. Let Xn be the sphere S 2 with a ball of radius 1/n
removed. Prove that the spaces Xn (regarded with their intrinsic metrics)
converge to S 2 .
2. Let Xn be obtained the same way from the circle S 1 . Show that Xn
do not converge to S 1 .
Exercise 7.5.3. 1. Let X be a straight line segment in R3 , and let Xn be
the boundary of its (1/n)-neighborhood (it is a two-dimensional surface),
equipped with the length metric induced from R3 . Prove that Xn −→
GH X as
n → ∞.
2. Let X be a planar disc in R3 , and again let Xn be the boundary of its
(1/n)-neighborhood equipped with the induced length metric. Prove that
the sequence {Xn }∞ n=1 converges in the Gromov–Hausdorff sense, but the
limit is not isometric to X.
Exercise 7.5.4. Justify the convergence of “spheres with small handles”
described in Subsection 7.1.4. Namely, let Xn be a length space obtained
from the standard unit sphere by removing a round disc of diameter less
than 1/n and attaching a handle whose (intrinsic!) diameter is less than
1/n. Prove that Xn −→ 2
GH S .

While any compact metric space can be obtained as a limit of finite

spaces (Example 7.4.9), these finite spaces do not carry length metrics. For
length spaces, the role of finite (zero-dimensional) spaces is played by the
one-dimensional ones, i.e., graphs. Recall that a finite metric graph is a
length space obtained by gluing together several spaces isometric to line
segments in such a way that only endpoints may be shared between the
segments. Equivalently, a finite metric graph is a finite topological graph
equipped with a length metric.
Proposition 7.5.5. Every compact length space can be obtained as a
Gromov–Hausdorff limit of finite graphs.

Proof. Let X be the length space in question. Pick small positive numbers
ε and δ (where δ is much smaller than ε), and choose a finite δ-net S in X.
7.5. Convergence of Length Spaces 267

Then consider the following graph G: the set of vertices of G is S, two points
x, y ∈ S are connected by an edge if and only if |xy| < ε, and the length of
this edge equals |xy|.
Let us show that G is an ε-approximation for X (cf. Definition 7.4.10) if δ
is small enough, say, δ < 14 ε2/diam(X). We consider S both as a subset of X
and a subset of G. Obviously S is an ε-net in both spaces, and |xy|G ≥ |xy|
for all x, y ∈ S where | · |G denotes the distance in G. It remains to show
that |xy|G ≤ |xy| + ε.
Let γ be a shortest path in X connecting x and y. Choose n points
x1 , . . . , xn , where n ≤ 2L(γ)/ε, dividing γ into intervals of lengths no
greater than ε/2. For every i = 1, . . . , n, there is a point yi ∈ S such
that |xi yi | ≤ δ. In addition, set x0 = y0 = x and xn+1 = yn+1 = y. Note
that |yi yi+1 | ≤ |xi xi+1 | + 2δ < ε for all i = 0, . . . , n. In particular, so yi and
yi+1 are connected by an edge in G provided that δ < ε/4. Then
n
X n
X
|xy|G ≤ |yi yi+1 | ≤ |yi yi+1 | + 2δn = |xy| + 2δn.
i=0 i=0

Recall that n ≤ 2L(γ)/ε ≤ 2 diam(X)/ε; hence

4 diam(X)
|xy|G ≤ |xy| + δ · < |xy| + ε
ε
if δ < 14 ε2/diam(X).
Thus we have a finite graph which is an ε-approximation for X. Passing
ε to zero yields a sequence of graphs converging to X. ¤

Exercise 7.5.6. Prove that every compact length space X can be repre-
sented as a Gromov–Hausdorff limit of finite graphs isometrically embedded
in X, i.e., of topological graphs embedded in X and equipped with their
induced length metrics.
Hint: Utilize the same construction as in the above proof, but draw
the edges in X (as shortest paths) adding new vertices when these shortest
paths intersect one another. To get rid of weird cases when one has to add
infinitely many vertices, show that shortest paths can be chosen so that the
intersection of any two of them is the empty set, or a single point, or an
interval of both paths.

Note that the number of vertices and edges of graphs constructed in the
proof of Proposition 7.5.5 tends to infinity as ε approaches zero. In other
words, the graphs get more and more complex. This is not a defect of this
particular construction but a consequence of a general obstacle:
268 7. Space of Metric Spaces

Exercise 7.5.7. 1. Let N be a natural number and {Xn }∞n=1 a sequence of

graphs each having no more than N edges. Prove that the limit of {Xn }, if
one exists, is a finite graph.
2. Let {Xn }∞
n=1 be a sequence of finite graphs each having no more than
N vertices. Prove that the limit of {Xn }, if one exists, is a (possibly infinite)
graph.

7.5.2. Topology of Gromov–Hausdorff limits. This subsection con-

sists of a series of exercises, some of which are quite complicated. They
give some examples of what can and what cannot happen to the Gromov–
Hausdorff limit of length spaces.
A general rule of thumb: in dimensions 1 and 2, one can expect nice
topological relations between converging spaces and their limit. In higher
dimensions, there are relations between fundamental groups but not much
can be said beyond this.
The first exercise is a general fact about Gromov–Hausdorff convergence
(not specific to length spaces). To get better understanding of it, recall
again the spheres with vanishing handles from Section 7.1.4. There are no
continuous maps with small distortion from the sphere to a sphere with a
small handle, but such a map in the opposite direction exists: just project
the surface onto the sphere.

Exercise 7.5.8. (a) Let X be the unit ball in Rn . Let {Xn }∞

n=1 be a
sequence of compact metric spaces, Xn −→
GH X. Prove that there exists a
sequence of continuous maps fn : Xn → X such that each fn is an εn -
isometry for some sequence εn → 0.
Prove the same if X is
(b) the unit circle S 1 ;
(c) the sphere S n−1 ;
(d) the n-dimensional torus T n = S 1 × · · · × S 1 ;
(e) a metric space homeomorphic to one of the above;
(f) a metric space homeomorphic to a compact smooth manifold.
(g) Prove the same under some weaker (as weak as possible) topological
restrictions on X.
Hint: Use Corollary 7.3.28 to obtain a sequence of possibly discontinuous
maps, then approximate these maps by continuous ones.

Exercise 7.5.9. Let {Xn } be a sequence of compact length spaces, Xn −→

GH
S 1 . Prove that, for all large enough n, the spaces Xn are not simply
connected.
7.5. Convergence of Length Spaces 269

Hint: Let {fn } be a sequence of maps from the previous exercise.

Construct an “approximate lift” of the circle into Xn , namely, a closed
curve in Xn that is mapped onto the circle “almost isometrically” in some
sense. Then show that this curve is not contractible.
Remark 7.5.10. It is essential that Xn in the above exercise are length
spaces. To see why, consider a sequence {Xn } where Xn is an arc in S 1
of length 2π − n1 equipped with the (nonintrinsic) restriction of the circle’s
metric.
Exercise 7.5.11. Let {Xn } be a sequence of compact locally simply con-
nected length spaces, Xn −→GH X, and let fn be as in Exercise 7.5.8. Prove
that, for all large enough n, fn induces a surjective homomorphism of fun-
damental groups.
In particular, if the spaces Xn are simply connected, then X is simply
connected too.
Hint: See the previous exercise.
Exercise 7.5.12. Let B be the unit two-dimensional disk with its standard
Euclidean metric. Let {Bn } be a sequence of length spaces homeomorphic
to B, Bn −→
GH B, and let ε be a positive number. Prove that for all large
enough n, there is a point pn ∈ Bn such that dist(pn , ∂Bn ) > 1 − ε.
Hint: Prove that there is a Jordan curve in Bn whose image in B is close
to ∂B (in the sense of uniform distance). This curve bounds a region in Bn .
Within that region, there must be a point corresponding to the center of B
in the sense of Gromov–Hausdorff approximation. This point can be taken
for pn .
Exercise 7.5.13. Prove that a sequence of length spaces homeomorphic to
the sphere S 2 cannot converge to (a) the standard two-dimensional disc B;
(b) a space homeomorphic to B.
Exercise 7.5.14. Let {Xn } be a sequence of length spaces homeomorphic
to S 2 , Xn −→ 2
GH X, and X is also homeomorphic to S . Prove that there is a
sequence of homeomorphisms fn : Xn → X with dis(fn ) → 0.
Exercise 7.5.15. 1. Let {Xn } be a sequence of length spaces homeomorphic
to S 2 converging to some compact space X. Prove that X cannot contain a
subset homeomorphic to the three-dimensional disc.
2. Prove the same for a sequence {Xn } of compact two-dimensional
manifolds with length metrics, provided that the genus of Xn (the number
of handles or films) is uniformly bounded.
3. Show that, without the condition of uniformly bounded genus, the
limit of {Xn } can be any compact length space. (Compare this with
Proposition 7.5.5 and Exercise 7.5.7.)
270 7. Space of Metric Spaces

Hint to 1 and 2: There exists a graph that can be topologically embedded

into a three-dimensional disc (any graph can), but not into Xn .

Most of the properties given in the above group of exercises are only
valid in the two-dimensional case. Below are some counterexamples to their
higher-dimensional counterparts.
Exercise 7.5.16. Let B be the standard three-dimensional ball. Show that,
for any given ε > 0, there exists a length space B ′ homeomorphic to B such
that dGH (B, B ′ ) < ε and ∂B ′ is an ε-net in B ′ . (Compare with Exercise
7.5.12.)
Hint: Push the boundary inside the ball so that it becomes an ε-net but
the intrinsic distances in the ball do not change much.
Exercise 7.5.17. Construct a sequence {Xn } of length spaces homeomor-
phic to S 3 and converging to the three-dimensional ball. (Compare with
Exercise 7.5.13.)
Hint: Consider doublings of the spaces B ′ from the previous exercise (i.e.,
the result of gluing together two isometric copies of B ′ along the boundary.)
Exercise 7.5.18. Construct a sequence {Xn } of length spaces homeomor-
phic to S 3 such that Xn −→ 3 3
GH S but no homeomorphism from Xn to S has
distortion less than 1/10. (Compare with Exercise 7.5.15.)
 Chapter 8

Large-scale Geometry

To get an idea of what we are going to talk about in this chapter, imagine
a device measuring distances with a precision, say, one mile. Such a tool
is useless for an engineer investigating the shape of a car, but it is more
than excellent for learning geometry of the solar system. In this chapter
we consider metric properties for which such a measuring device is perfectly
good. In other words, we are not going to distinguish two metrics if the
difference between them is bounded by a constant. More precisely, the
properties that we will discuss are the same for spaces lying within finite
Gromov–Hausdorff distance from one another. For example, a Euclidean
space and a lattice in it (regarded with the restriction of the Euclidean
metric) look the same from this point of view. Of course, no local properties
survive through such a transformation of a space but many global and
asymptotic ones remain.
In some cases, we will admit even less precise “measurement of dis-
tances”. For example, consider a measuring instrument which may give the
result ten times greater or smaller than the actual distance, plus the same
one-mile error. Quite surprisingly, this instrument allows one, for example,
to tell the Euclidean plane from the hyperbolic one. (A question of this sort,
concerning the physical universe, is a famous problem in modern cosmogony
and physics. Alas, our real instruments are not that good.)

8.1. Noncompact Gromov–Hausdorff Limits

For noncompact spaces, convergence with respect to the Gromov–Hausdorff
distance is not very useful. There is an analogy with the uniform convergence
of functions on a fixed domain. While the domain is compact, uniform

271
272 8. Large-scale Geometry

convergence is a powerful and widely used notion, but it becomes too

restrictive once noncompact domains come into question. For example, a
sequence {λn f }, where f is a continuous function and {λn } is a converging
sequence of real numbers, may fail to converge uniformly due to the fact
that f may not be bounded. Instead, one should utilize the wider notion of
uniform convergence on compact sets: a sequence of functions converges if
it converges uniformly on every compact subset of the domain.
A similar approach is taken to the Gromov–Hausdorff convergence when
noncompact metric spaces are involved. Roughly speaking, a sequence {Xn }
of metric spaces converges to a space X if for every r > 0 the balls of radius
r in Xn centered at some fixed points converge (as compact metric spaces)
to a ball of radius r in X. The actual definition (Definition 8.1.1 below) is
more complicated, but in most cases it is equivalent to this description.
To see the difference from the ordinary Gromov–Hausdorff convergence,
consider a sequence of Euclidean spheres of radii growing to infinity. For
every given r > 0, the sets of diameter no greater than r in these spheres
look more and more similar to subsets of the Euclidean plane, but the whole
spheres do not get close to the plane in any sense. These spheres have no
limit with respect to the Gromov–Hausdorff distance, but they converge to
the plane in the extended sense that we are about to define.
Definition 8.1.1. A pointed metric space is a pair (X, p) consisting of a
metric space X and a point p ∈ X.
A sequence {(Xn , pn )}∞
n=1 of pointed metric spaces converges in the
Gromov–Hausdorff sense to a pointed metric space (X, p) if the following
holds. For every r > 0 and ε > 0 there exists a natural n0 such that for
every n > n0 there is a (not necessarily continuous) map f : Br (pn ) → X
such that the following hold:
(1) f (pn ) = p;
(2) dis f < ε;
(3) the ε-neighborhood of the set f (Br (pn )) contains the ball Br−ε (p).
We will use the notation (Xn , pn ) −→
GH (X, p) for this type of convergence.

Exercise 8.1.2. Prove that for compact metric spaces the convergence of
pointed spaces is equivalent to the ordinary Gromov–Hausdorff convergence
in the following sense:
1. (Xn , pn ) −→
GH (X, p) implies that Xn −→
GH X.

2. If Xn −→
GH X and p ∈ X, then one can choose a point pn in every Xn
so that (Xn , pn ) −→
GH (X, p).

The first two requirements in the above definition imply that the image
f (Br (pn )) is contained in the ball of radius r + ε centered at p. This and
8.1. Noncompact Gromov–Hausdorff Limits 273

the third requirement imply (by Corollary 7.3.28) that the ball Br (pn ) in
Xn lies within the Gromov–Hausdorff distance of order ε from a subset of
X between the balls of radii r − ε and r + ε centered at p (here “between”
means that the set contains one ball and is contained in the other).
If X is a length space, this remains true for the r-ball centered at p
instead of the unknown subset; in other words, for every r > 0 the r-balls in
Xn centered at pn converge (with respect to the Gromov–Hausdorff distance)
to the r-ball in X centered at p. In the general case (for nonlength spaces)
such a simplification is not possible (see exercises below).
Exercise 8.1.3. Let (Xn , pn ) −→
GH (X, p) and let X be a length space. Prove
that Br (pn ) −→
GH Br (p) for every r > 0.
Exercise 8.1.4. Show that the statement of the previous exercise fails
without the assumption that X is a length space. To do that, construct
a sequence {Xn } of compact metric spaces converging to a compact metric
space X such that no sequence of closed unit balls in Xn converges to a
closed unit ball in X.
This property that the balls converge does not yet imply convergence of
pointed spaces. The first requirement in Definition 8.1.1 includes additional
information that puts the points pn and p into a special position. Roughly
speaking, not only should the balls converge, but also their distinguished
central points should converge at the same time.
Exercise 8.1.5. Construct a compact metric space X and two points
p, q ∈ X such that for every r > 0 the balls Br (p) and Br (q) in X are
isometric, but there is no isometry map from X to itself that maps p to q.
Hint: There are examples among finite spaces.
Given such X, p and q, let Xn = X and pn = q for all n ≥ 1. Prove
that the sequence {(Xn , pn )} of pointed spaces does not converge to (X, p)
despite the fact that Br (pn ) −→
GH Br (p) for every r > 0.

Obviously if a sequence of pointed metric spaces converges to a pointed

space (X, p), it also converges to its completion. We will only consider
complete metric spaces as Gromov–Hausdorff limits. Then, similarly to the
case of ordinary convergence, a Gromov–Hausdorff limit of pointed spaces
is essentially unique.
Definition 8.1.6. We say that two pointed metric spaces (X, p) and (X ′ , p′ )
are isometric if there is an isometry f : X → X ′ such that f (p) = p′ . Such
a map f is called a pointed isometry from (X, p) to (X ′ , p′ ).
Theorem 8.1.7. Let (X, p) and (X ′ , p′ ) be two (complete) Gromov-
Hausdorff limits of a sequence {(Xn , pn )}∞n=1 , and assume X is boundedly
compact. Then (X, p) and (X ′ , p′ ) are isometric.
274 8. Large-scale Geometry

Proof. We only outline the proof here; the details are similar to those of the
proof of Theorem 7.3.30. Given an r > 0 and ε > 0, the map whose existence
is guaranteed by Definition 8.1.1 can be converted into a correspondence Rr,ε
between sets of Yr,ε ⊂ X and Yr,ε ′ ⊂ X ′ such that: Y ′
r,ε and Yr,ε contain the
balls of radius r − ε and are contained in the balls of radius r + ε centered at
p and p′ , respectively; p and p′ are in correspondence with each other; and
dis Rr,ε < ε.
Choosing one corresponding point for every point of Yε yields a map
′ that maps p to p′ and has distortion < ε. Using the Cantor
fr,ε : Yr,ε → Yr,ε
diagonal procedure first for ε → 0 and then for r → ∞, one can end up with
a distance-preserving map from a dense subset of X to X ′ , which extends to
a distance-preserving map f : X → X ′ with f (p) = p′ . Since f is distance-
preserving, it maps every ball Br (p) in X to the corresponding ball Br (p′ )
in X ′ .
In addition, the images of maps fr,ε are ε-nets in the respective subsets of
X ′ , and this implies that the balls Br (p′ ) in X ′ are compact as well (compare
with Exercise 7.3.31). Hence a similar distance-preserving map f ′ : X ′ → X
exists. Due to compactness of balls, this implies that the restriction of f to
Br (p) is an isometry onto Br (p′ ) for every r > 0. Hence f is an isometry
onto X ′ . ¤

In the sequel, we always assume that the pointed spaces under consid-
eration are boundedly compact. As the next exercise shows, this property
is inherited by limit spaces.
Exercise 8.1.8. Suppose that (Xn , pn ) −→
GH (X, p) where the spaces Xn are
boundedly compact and X is complete. Prove that X is boundedly compact.

Most properties of Gromov–Hausdorff convergence of compact spaces

have their counterparts for pointed spaces. We collect the most important
ones in the next two theorems.
Theorem 8.1.9. Let (Xn , pn ) −→
GH (X, p) where Xn are length spaces and
X is complete. Then X is a length space.

Proof. Repeat the proof of Theorem 7.5.1. ¤

The following is the version of the compactness theorem for pointed

convergence.
Theorem 8.1.10. Let X be a class of pointed metric spaces. Suppose that
for every r > 0 and ε > 0 there exists an N = N (r, ε) such that for every
(X, p) ∈ X the ball Br (p) in X admits an ε-net of no more than N (r, ε)
points. Then the class X is precompact in the sense that any sequence of
spaces in X contains a converging subsequence.
8.2. Tangent and Asymptotic Cones 275

Proof. Similar to that of Theorem 7.4.15. Again, one has to apply the
Cantor diagonal procedure twice, for ε → 0 and for r → ∞. ¤

8.2. Tangent and Asymptotic Cones

In this section we extend the notion of tangent and asymptotic cones
discussed in Section 7.1, to abstract metric spaces.
Recall that for a metric space X and a λ > 0 one can consider a rescaled
metric space λX which is the same set of points equipped with the original
metric multiplied by λ. Rescaling a pointed space (X, p) naturally yields a
pointed space (λX, p).
Definition 8.2.1. A pointed metric space (X, p) is called a cone if it is
invariant under rescaling, i.e., if (λX, p) is isometric to (X, p) as a pointed
space for any λ > 0.

Note that a cone is not necessarily a cone over a metric space as defined
in Subsection 3.6.2.

8.2.1. Tangent cone. The tangent cone is a local notion which does not
belong to large-scale geometry. Nevertheless we define it here because the
definition is in some sense similar to that of the asymptotic cone.
Definition 8.2.2. Let X be a (boundedly compact) metric space, p ∈ X. A
Gromov–Hausdorff limit of pointed spaces (λX, p) as λ → ∞, if one exists,
is called the Gromov–Hausdorff tangent cone of X at p.
As usual, the “limit as λ → ∞” can be interpreted as the limit through
any sequence of values for λ tending to infinity (which should exist and be
the same for all sequences).

Note that the tangent cone is a pointed metric space. Its distinguished
point (the natural ancestor of p) is called the origin or the apex of the cone.
The tangent cone is indeed a cone in the sense that it is isometric to any
dilatation of itself (via a pointed isometry). The tangent cone is a local
invariant: it is determined by any small neighborhood of the point. More
precisely, if U is a neighborhood of p in X, then the tangent cones of U and
X at p are isometric. This follows immediately from the definition.
The tangent cone is indeed a cone in the sense of Definition 8.2.1. This
follows from the following simple fact:
Exercise 8.2.3. Let {(Xn , pn )}∞n=1 be a sequence of pointed metric spaces,
(Xn , pn ) −→
GH (X, p). Prove that (λXn , pn ) −→
GH (λX, p) for any λ > 0.

Exercise 8.2.4. Let M be an n-dimensional Riemannian manifold. Prove

that the tangent cone of M at any point exists and is isometric to Rn .
276 8. Large-scale Geometry

Exercise 8.2.5. Prove that the tangent cone of a convex set, as defined in
Subsection 7.1.1, is also the Gromov–Hausdorff tangent cone.

Remark 8.2.6. In Subsection 3.6.6 we introduced another local construc-

tion, the space of directions at a point. In fact, for “good” spaces the two
constructions carry the same information; moreover the Gromov–Hausdorff
tangent cone is nothing but the metric cone over the space of directions.

8.2.2. Asymptotic cone.

Definition 8.2.7. Let X be a (boundedly compact) metric space and p ∈ X.

A Gromov–Hausdorff limit of pointed spaces (λX, p) as λ → 0, if one exists,
is called the Gromov–Hausdorff asymptotic cone of X, or a cone of X at
infinity.

Proposition 8.2.8. The asymptotic cone does not depend on the choice of
the reference point p.

Proof. Let (C, o) be the asymptotic cone of X defined with a point p ∈ X.

Let p′ ∈ X be another point. We have to prove that (λX, p′ ) → (C, o)
as λ → 0. Let f : λX → C be a map from Definition 8.1.1 of pointed
convergence (for some ε > 0). Let us replace f by a map f ′ : λX → C
defined as follows: f ′ (x) = f (x) for all x 6= p′ , and f ′ (p′ ) = o. We have
moved the image of one point by a distance no greater than λ|pp′ | + ε; hence
the conditions from Definition 8.1.1 remain satisfied for p′ instead of p with
ε replaced by 2ε + λ|pp′ | (check this!). Since ε and λ are arbitrarily small
and |pp′ | is fixed, the desired convergence follows. ¤

Exercise 8.2.9. Prove that the asymptotic cone of a convex set, as defined
in Subsection 7.1.2, is also its Gromov–Hausdorff asymptotic cone.

Exercise 8.2.10. Let X and Y be metric spaces and dGH (X, Y ) < ∞.
Prove that, if X has an asymptotic cone, then Y has one too, and the two
cones are isometric.
In particular, if a metric space X lies within a finite Gromov–Hausdorff
distance from some cone Y , then Y is an asymptotic cone of X.

Exercise 8.2.11. Prove that the grid described in Subsection 7.1.3 has
asymptotic cone isometric to R21 .

The next exercise generalizes the previous one. There are further gen-
eralizations in Subsection 8.5.1.

Exercise 8.2.12. Consider a group Zn and a symmetric finite set S of

generators in it. Let d be the word metric associated with S, and k · k be
8.3. Quasi-isometries 277

the norm on Rn whose unit ball is the convex hull of S. Prove that there
exists a constant C (depending on S) such that

kx − yk ≤ d(x, y) ≤ kx − yk + C

for all x, y ∈ Zn .
In particular, the Gromov–Hausdorff distance between (Zn , d) and the
normed space (Rn , k · k) is finite (not greater than C), and hence (Rn , k · k)
is the asymptotic cone of (Zn , d).

The next exercise shows that the property of having a Gromov–Hausdorff

asymptotic cone is not as common as one might think.

Exercise 8.2.13. Prove that the hyperbolic plane H2 does not have an
asymptotic cone.
1
Hint: The sequence of rescaled balls n Bn (p), n ∈ N, p ∈ H2 , is not
uniformly totally bounded.

In plain words, the reason why the hyperbolic plane does not have an
asymptotic cone is that its metric balls grow too fast when the radius goes
to infinity. There are other constructions encoding asymptotic properties
of such “fast growing” spaces and in other cases when Gromov–Hausdorff
asymptotic cones are not applicable. One possible approach is to use a
weaker type of limit than the Gromov–Hausdorff one. We do not discuss
such generalized definitions in this book. Let us only mention that the
“generalized” asymptotic cone of the hyperbolic plane is an infinite (and
not locally finite) tree.
Another construction serving the same purposes is the ideal boundary
of a space; cf. Subsection 5.3.3 for the case of hyperbolic plane. While the
asymptotic cone is “the tangent cone at infinity” in some sense, the ideal
boundary is a sort of “space of directions” at infinity.

8.3. Quasi-isometries
8.3.1. Definitions and first examples. Quasi-isometries are a large-
scale analog of bi-Lipschitz maps. Two metric spaces are quasi-isometric if
they are bi-Lipschitz equivalent up to a finite Gromov–Hausdorff distance.
This is formally described as follows:

Definition 8.3.1. Metric spaces X and Y are said to be quasi-isometric

if there exist metric spaces X ′ and Y ′ such that dGH (X, X ′ ) < ∞ ,
dGH (Y, Y ′ ) < ∞, and the spaces X ′ and Y ′ are bi-Lipschitz homeomorphic
(i.e., there exists a bi-Lipschitz homeomorphism between them).
278 8. Large-scale Geometry

We will soon see that X ′ and Y ′ can be chosen among subsets of X

and Y , moreover, among separated nets.
Recall that a subset S of a metric space X is called a ρ-net if the
Hausdorff distance between S and X is not greater than ρ.

Definition 8.3.2. Let X be a metric space. A set S ⊂ X is called a net

in X if the Hausdorff distance between X and S is finite. In other words, S
is a ρ-net in X for a sufficiently large ρ.
A separated net is a net which is ε-separated for some ε > 0 (recall that
being an ε-separated set means that |xy| ≥ ε for any two distinct points
x, y ∈ S).

Every metric space X contains an ε-separated net for any given ε.

Indeed, by Zorn’s Lemma there exist an ε-separated set S ⊂ X which is
maximal by inclusion; i.e., if S ( S ′ , then S ′ is not ε-separated. Such S is
an ε-net; indeed, if there is a point x ∈ X with dist(x, S) ≥ ε, then the set
S ′ = S ∪ {x} is again ε-separated, contrary to the maximality of S.
Let X, Y, X ′ , Y ′ be as in Definition 8.3.1. Since dGH (X, X ′ ) < ∞ and
dGH (Y, Y ′ ) < ∞, there are maps f1 : X → X ′ and f2 : Y ′ → Y with finite
distortions and such that the images f1 (X) and f2 (Y ′ ) are nets in X ′ and
Y , respectively. Let g : X ′ → Y ′ be a bi-Lipschitz homeomorphism and
λ = max{dil g, dil g −1 }. Define a map f = f2 ◦ g ◦ f1 from X to Y . Then
1
(8.1) |xy| − C ≤ |f (x)f (y)| ≤ λ|xy| + C
λ
for all x, y ∈ X where C = dis(f2 ) + λ dis(f1 ).

Definition 8.3.3. Let X and Y be metric spaces. A map f : X → Y is

called a quasi-isometry if there are constants C ≥ 0 and λ ≥ 1 such that the
inequality (8.1) holds for all x, y ∈ X.

Observe that the quasi-isometry f : X → Y constructed above possesses

an additional property: its image f (X) is a net in Y . Moreover the image
of any net in X is a net in Y . To prove this, observe that each of the maps
f1 , f2 and g sends nets to nets.
Now choose a ∆-separated net S ∈ X for a large enough ∆, namely,
∆ > (2λ + 1)C where λ and C are from (8.1). Then (8.1) implies that
1
|xy| ≤ |f (x)f (y)| ≤ (λ + 1)|xy|
2λ
for all x, y ∈ S. Hence f is a bi-Lipschitz homeomorphism between S and
f (S). Note that S and f (S) can be used instead of X ′ and Y ′ in Definition
8.3.1. Thus we have proved the following
8.3. Quasi-isometries 279

Proposition 8.3.4. For any metric spaces X and Y , the following three
assertions are equivalent:
(i) X and Y are quasi-isometric;
(ii) there is a quasi-isometry f : X → Y whose image f (X) is a net in
Y;
(iii) X and Y contain bi-Lipschitz homeomorphic separated nets. ¤
Corollary 8.3.5. Being quasi-isometric is an equivalence relation.

Proof. Symmetry is obvious from the definition. Transitiveness follows

from (ii) in the above proposition. ¤
Example 8.3.6. Euclidean spaces Rn and Rm are not quasi-isometric for
m 6= n. To prove this, consider a separated net S in Rn and for every R > 0
denote by N (R) the number of elements of S within the R-ball centered at
the origin. Since S is a separated net, say, an ε-separated r-net, the (ε/2)-
balls centered at points of S do not intersect one another, and the r-balls
cover the space. Since the volume of a ball is proportional to the nth power
of its radius, this leads to estimates for N (R):
(R − r)n (R + ε/2)n
≤ N (R) ≤ ,
rn (ε/2)n
or, in a simple form,
cRn ≤ N (R) ≤ CRn
for all large enough R, where c and C are some positive constants. This
property of a net is obviously preserved by bi-Lipschitz maps, but no
separated net in Rm satisfies it if m 6= n (because for the same reasons
the respective quantity for Rm lies between cRm and CRm ).
Similarly, Euclidean spaces are not quasi-isometric to the hyperbolic
plane (and hyperbolic spaces) because the areas (or volumes) of hyperbolic
balls grows exponentially as radii go to infinity.
Exercise 8.3.7. Prove that hyperbolic spaces Hm and Hn are not quasi-
isometric if m 6= n.
Hint: Supposing that they are quasi-isometric, prove that there is
a continuous quasi-isometry from Hm to Hn . Then consider this quasi-
isometry restricted to a large sphere in Hm (assuming that m > n).

8.3.2. Groups and orbits. We already discussed metrics invariant under

a group action in Section 3.3; below we revise this notion in a new context.
Let G be a group, X a set, and ϕ : G × X → X an action of G on X
(cf. Definition 3.3.5). As usual, we write g(x) instead of ϕ(g, x). A metric
d on X is said to be G-invariant (under this action) if G acts by isometries,
280 8. Large-scale Geometry

i.e., if d(g(x), g(y)) = d(x, y) for all x, y ∈ X, g ∈ G. We will consider only

isometric actions (or, equivalently, only G-invariant metrics).
An action is said to be co-compact if the quotient space X/G is compact,
and is said to be co-bounded if X/G is bounded. Obviously if X is boundedly
compact (in particular, is a complete locally compact length space), then
every co-bounded action of a group on X is co-compact. An orbit of an
element x ∈ X under an action of G, denoted by Gx, is by definition the set
{gx : g ∈ G}.

Exercise 8.3.8. Prove that the orbit of a point is a net if and only if the
action is co-bounded.

An important example of a group action is its action on itself by multipli-

cation. Namely, one lets X = G and defines the action by g(h) = gh for all
g, h ∈ G. Metrics on G invariant under this action are called left-invariant
metrics (reflecting the fact that the group acts by left multiplication). If
G is an abelian group, the prefix “left-” can be omitted. Note that word
metrics discussed in Subsection 3.2.3 are left-invariant.
A left-invariant metric on G is determined by the distances from the
identity of the group. Indeed, if we define |g| = d(e, g), then d(g, g1 ) =
|g −1 g1 |. Conversely, a function | · | : G → R defines a metric if and only if

(1) |e| = 0, |g| > 0 for all g 6= 0;

(2) |g −1 | = |g| for all g;
(3) |g1 g2 | ≤ |g1 | + |g2 |.

(This is a trivial exercise.)

Note that these conditions on | · | are very similar but not identical to
those in the definition of a norm on a vector space. For a word metric, the
corresponding function | · | is the length of a shortest word representing a
given element.
Here is a more general source of examples of left-invariant metrics on
groups. Let X be an arbitrary set along with an action of G on it. Assume
that the action is free, i.e. g(x) 6= x for all x ∈ X and g ∈ G unless g = e.
Then every G-invariant metric dX on X and every point x ∈ X determine
a left-invariant metric dG on G by the formula

(8.2) dG (g1 , g2 ) = dX (g1 (x), g2 (x)).

This can be interpreted as follows: one identifies G with the orbit Gx and
uses the metric d restricted on Gx. The metric dG defined by the above
formula will be referred to as an orbit metric.
8.3. Quasi-isometries 281

Exercise 8.3.9. Let X, dX and G be as above, x1 and x2 be two points

of X, d1 and d2 be the corresponding orbit metrics on G. Prove that the
function d1 − d2 is bounded.
This construction becomes more interesting if we restrict ourselves to
length metrics on X. Assume that the action of G is free and totally
discontinuous, or, equivalently, the projection p : X → X/G is a covering
map. (See Subsection 3.4.2 for a general discussion of coverings and length
metrics.) Then there is a natural 1-1 correspondence between length metrics
on X/G and G-invariant length metrics on X (cf. Proposition 3.4.16). Thus,
given a length metric on X/G and a point y ∈ X/G, one can define a left-
invariant metric on G by picking any point x such that p(x) = y and taking
the corresponding orbit metric. Note that the result does not depend on the
choice of x (up to inner automorphisms).
Since the orbit is just the set p−1 (y), the group with this metric is
isometric to p−1 (y) with the metric restricted from X. If the action is
co-compact, then p−1 (y) is a net in X, so the Gromov–Hausdorff distance
between X and the group is finite (so one can replace X by the group for
all large-scale considerations).
Exercise 8.3.10. Prove the above statement: an orbit of a co-compact
action is a separated net (of course, provided that the action is free, totally
discontinuous and by isometries, and X is a length space).
Example 8.3.11. Word metrics (see Subsection 3.2.3) can be represented
as orbit metrics of length spaces. To produce a word metric on a group
G, let X be a Cayley graph (with the obviously defined action of G), and
choose the identity of the group as the initial point of an orbit.
One can construct orbit metrics the other way round, starting from the
quotient space. Let Y be a length space, and assume that Y is locally simply
connected. Then there is a universal covering p : X → Y . Furthermore, the
fundamental group of Y naturally acts on X by isometries (as the group of
deck transformations). Thus every length metric on Y and a point y ∈ Y
determine a left-invariant metric on the fundamental group π1 (Y, y), namely,
the metric of X restricted to the orbit p−1 (y). Here is another description
of this metric:
Definition 8.3.12. Let Y be as above, y ∈ Y , g ∈ π1 (Y, y). The length of
g is defined by
length(g) = inf{L(γ) : γ is a loop representing g in π1 (Y, y)}.
The distance between two elements g1 , g2 ∈ π1 (Y, y) is the length of g1−1 g2 .
Exercise 8.3.13. Prove that the above definition defines the same metric
on π1 (Y, y) as in the construction with the universal covering.
282 8. Large-scale Geometry

Remark 8.3.14. One can apply the same construction to other coverings,
not only to the universal one; however the covering must be regular. (This
is necessary in order to represent Y as the quotient space of the group of
deck transformations.) In this case, the group of deck transformation is a
factor-group of π1 (Y ).

We will soon see that all possible orbit metrics of co-compact actions on a
given (finitely generated) group are bi-Lipschitz equivalent to one another.
In other words, the quasi-isometry class of a metric on a group depends
only on the group, not on a length space and an action used to define the
metric. This yields the following remarkable corollary: if two compact length
spaces have isomorphic fundamental groups, then their universal coverings
are quasi-isometric. Indeed, each universal covering has finite Gromov–
Hausdorff distance from the fundamental group equipped with an orbit
metric, and two orbit metrics are bi-Lipschitz equivalent (as we will prove).
Since all metrics on a fixed group are bi-Lipschitz equivalent, we can talk
about quasi-isometries between abstract finitely generated groups without
referring to a particular metric. Namely, two groups are quasi-isometric if
they are quasi-isometric when equipped with arbitrary word metrics.

Example 8.3.15. The free groups F2 and F3 with two and three generators
(equipped, say, with word metrics) are quasi-isometric. (Compare this with
the fact that Zm and Zn are not quasi-isometric if m 6= n, as we have seen
in Example 8.3.6.)
To prove this, recall that F2 is the fundamental group of a bouquet
of two circles (say, of unit length). This bouquet X admits a two-sheeted
covering by a graph Y consisting of three circles S1 , S2 , S3 connected as a
chain: S1 has one common point with S2 , S2 has one common point with
S3 , and S1 ∩ S2 = ∅. Metrically, Y is a circle of length 2 glued with two
circles of unit length; the smaller circles are attached to opposite points of
the larger one. Constructing the covering map from Y to X is left as an
exercise.
Since Y is a covering space for X, the universal covering of Y coincides
with the universal covering of X. This covering space is quasi-isometric to
both fundamental groups π1 (X) ∼ = F2 and π1 (Y ) ∼
= F3 . (The latter group is
F3 because Y is homotopy equivalent to a bouquet of three circles. To see
this, just contract the edges.) Thus F2 and F3 are quasi-isometric.

Exercise 8.3.16. Construct an explicit quasi-isometry from F2 onto a net

in F3 .

Exercise 8.3.17. Prove that all free groups Fn , n ∈ N, are quasi-isometric.

8.3. Quasi-isometries 283

Now we turn to the proof of the above claimed equivalence of orbit

metrics. We begin with the case of word metrics.
Proposition 8.3.18. Let G be a finitely generated group. Then all word
metrics on G are bi-Lipschitz equivalent to one another.

Proof. Let |·|1 and |·|2 be the distance functions of the identity in two word
metrics defined by generating sets S1 and S2 , respectively. Recall that the
sets S1 and S2 must be finite. Let g1 . . . gn be a shortest word in generators
from S1 representing a given element g ∈ G. Then |g|1 = n and
|g|2 = |g1 . . . gn |2 ≤ |g1 |2 + · · · + |gn |2 ≤ C1 n = C1 |g|1 ,
where C1 = maxh∈S1 |h|2 . (Note that we have not yet used that | · |2 is a
word metric.) Similarly, |g|1 ≤ C2 |g|2 for some constant C2 not depending
on g. Thus | · |1 and | · |2 are bi-Lipschitz equivalent. ¤
Theorem 8.3.19. Let G be a finitely generated group and d be an orbit
metric of a free co-compact action of G by isometries on a length space X.
Then d is bi-Lipschitz equivalent to a word metric.
In particular, all such orbit metrics on G are bi-Lipschitz equivalent to
one another.
Corollary 8.3.20. All length spaces X admitting a free totally discontinu-
ous co-compact action of a given group G are quasi-isometric to one another,
and are quasi-isometric to the group G equipped with any word metric.
In particular, if Y is a length space and X its universal covering with the
metric lifted from Y , then X is quasi-isometric to the group π1 (Y ) equipped
with any word metric.

Proof of the theorem. Let | · | denote the distance from the identity in
the metric d. If | · |w is a word metric, then | · | ≤ C| · |w for some constant
C, similarly to the previous proof. Thus it suffices to prove that | · |w ≤ C| · |
for some constant C and some word metric | · |w .
Let x ∈ G be the point whose orbit is used to define the metric d (i.e.,
d(g, h) = dX (gx, hx) for all g, h ∈ G). Since the action is co-compact,
X is locally compact and complete and the orbit Gx is a separated net.
Therefore every metric ball in X contains only finitely many elements of the
orbit. Hence every ball in (G, d) is a finite set.
Let D be a number such that the orbit Gx is a D-net in X, and let r
be so large that r > 2D + 1 and the r-ball in (G, d) centered at the identity
contains a set of generators. We may assume that this ball itself is chosen
as the set of generators S defining the word metric | · |w . (We have an
option to choose a word metric to be compared with d; then the result will
284 8. Large-scale Geometry

follow for any word metric because all word metrics are already proven to
be equivalent.)
Pick a g ∈ G and let γ be a shortest path in X connecting x to gx. We
are going to show that |g|w ≤ C|g| = C ·L(γ) for a constant C not depending
on g. Divide γ into intervals of length 1 or less by points x1 , . . . , xn , where
n ≤ L(γ) ≤ n + 1. For every i = 1, . . . , n, find a yi ∈ Gx such that
dX (xi , yi ) ≤ D. In addition, set y0 = x and yn+1 = Gx. Let gi ∈ G be such
−1
that gi x = yi (note that g0 = e and gn+1 = g), and set hi = gi gi−1 . Then
hi xi−1 = yi ; hence |hi | = dX (xi−1 , xi ) ≤ 2D + 1. Therefore hi belongs to our
set of generators (recall that it is just the r-ball in G, and r > 2D + 1). On
the other hand, the product (word) h1 . . . hn+1 equals g; hence |g|w ≤ n + 1.
Thus |g|w ≤ L(γ) + 1 = |g| + 1. Since the distances |g| are separated
away from zero (i.e., there is an ε > 0 such that |g| ≥ ε for all ε > 0),
this estimate can be written in the desired form |g|w ≤ C|g|, namely,
|g|w ≤ (1 + 1/ε)|g|. ¤
Remark 8.3.21. One can remove the condition that the action is free from
the theorem. The only problem is that an orbit metric is no longer a metric;
it is only a semi-metric. As a consequence, it is not bi-Lipschitz equivalent
to word metrics, but it is still quasi-isometric to them. The same proof
works with obvious modifications.

The condition that a given metric is an orbit metric of some action seems
hard to verify and somewhat restrictive (it is not a ”large-scale” one, after
all). There are various possible replacements for it; one of such conditions
is given in the following exercise.
Exercise 8.3.22. Prove the following statement which is slightly more
general than the above theorem. Let d be a left-invariant metric on a finitely
generated group G such that every ball of d is a finite set, and there is a
constant C > 0 satisfying the following: for any x, y ∈ G there is a z ∈ G
with d(x, z) < 12 d(x, y)+C and d(y, z) < 21 d(x, y)+C. Then d is bi-Lipschitz
equivalent to the word metric.

8.4. Gromov Hyperbolic Spaces

In this section we will deal with coarse geometry; this means that we will
study our metric spaces on the large scale of distances (large-scale geome-
try), neglecting all phenomena related to distances smaller than a certain
value. From this viewpoint, a space can be substituted by a discrete subset
forming an ǫ-net and carrying discrete topology; hence all local topology
is irrelevant for our considerations. A model example of this situation is
the universal cover of a compact space X; the universal cover can be sub-
stituted by the fundamental group of X with the induced metric. We will
8.4. Gromov Hyperbolic Spaces 285

be mainly discussing δ-hyperbolicity—a version of “coarsely negative cur-

vature” introduced by Gromov. Loosely speaking, δ-hyperbolicity reflects
large-scale features of hyperbolic planes (or, more generally, of spaces of
curvature bounded above by a negative number; such spaces are discussed
in Chapter 9). (One notices that these properties have to do with large
distances only.)

8.4.1. δ-hyperbolicity. Let us recall that a triangle △a1 a2 a3 in the hy-

perbolic plane has the following property: there exists a “center”, that is,
a point c such that all distances from this point to the sides of the triangle
are less than 1 (see Subsection 5.3.5). Thus from our “coarse viewpoint” the
triangle is “slim”: it looks as if all its sides pass through one point c, and
the whole triangle looks like a bouquet of three segments [ca1 ], [ca2 ], [ca3 ].
In other words, every side of a triangle belongs to 1-neighborhood of the
union of the two other sides. One can use this property to define large-scale
hyperbolicity:
Definition 8.4.1. A length space (X, d) with a strictly intrinsic metric d is
said to be δ-hyperbolic, δ ≥ 0, if all triangles in (X, d) satisfy the following
property: each side of a triangle belongs to the δ-neighborhood of the union
of the two other sides.

Throughout this section by default we assume that all constructions take

place in a δ-hyperbolic space (X, d).
Clearly if a length space is δ-hyperbolic for a particular δ, it is also δ ′ -
hyperbolic for all δ ′ ≥ δ. Following our large-scale ideology, we will not
be concerned with a particular value of δ; a space is said to be Gromov
hyperbolic if it is δ-hyperbolic for some δ > 0.
Note that, in particular, this definition implies that two shortest paths
between the same points in a δ-hyperbolic space are within distance δ from
each other (to see this, one just regards the bi-gone formed by the shortest
path as a degenerate triangle with two coinciding vertices). Moreover, the
following proposition holds (we leave the proof to the reader as a trivial
exercise):
Lemma 8.4.2. A shortest path [ab] belongs to the (δ + d(b, c))-neighborhood
of a shortest path [ac].

For three points p, q, r, we will use the notation

1
(p, q)r = (d(r, p) + d(r, q) − d(p, q)).
2
Thus (p, q)r measures “to what extent the triangle inequality for the triangle
△prq is far from being an equality”. Let us say that a triangle △prq is δ-
almost degenerate if the triangle inequality is almost an equality: (p, q)r ≤ δ.
286 8. Large-scale Geometry

Consider a point p in the side [bc] of a triangle △abc in a Gromov

hyperbolic space. It immediately follows from the definition that at least
one of the triangles △apb and △apc is δ-almost degenerate; in other words,
the following inequality holds:
(8.3) 0 ≥ min((a, b)p , (a, c)p ) − δ.
Now we can see that in a δ-hyperbolic space (b, c)a is approximately the
distance from a to [bc]. Indeed, by the triangle inequality we obviously have
(b, c)a ≤ d(a, [bc]). Moving p continuously from b to c along [bc], we find a
value for p such that both triangles △pac and △pbc are δ-almost degenerate.
Hence
2δ ≥ d(b, p) + d(p, a) − d(a, b),
2δ ≥ d(c, p) + d(p, a) − d(a, c).
Adding these inequalities (and using
d(b, p) + d(p, c) = d(b, c) = d(a, c) + d(a, b) − 2(b, c)a ,
d(a, [bc]) ≤ d(a, p)),
we get d(a, [bc]) ≤ (b, c)a + 2δ, and hence
(8.4) (b, c)a ≤ d(a, [bc]) ≤ (b, c)a + 2δ.
Reasoning analogously to the argument used to obtain inequalities (8.4),
one proves the following useful lemma:
Lemma 8.4.3. Let [ab] and [ac] be two shortest paths. Let b′ ∈ [ab] and
c′ ∈ [ac] be such that d(a, b′ ) = d(a, c′ ) ≤ (b, c)a . Then d(b′ , c′ ) ≤ 3δ.
Moreover, there is a point p ∈ [bc] such that [bp] lies in δ-neighborhood of
[ba], and [cp] lies in δ-neighborhood of [ca].
Exercise 8.4.4. Prove the lemma.
Though easy to prove, geometrically this lemma is quite striking. Con-
sider two shortest paths [ab] and [ac] of length R starting from the same
point. Let us think of this point a as a reference point (center), and R a
large number (compared to δ). For instance, let us assume now that the
distance between b and c is also R. In the Euclidean plane such shortest
paths would form an angle of 60 degrees. However, in a δ-hyperbolic space
they go very close (within distance 3δ) to each other until at least half-way!
Moreover, a shortest path [bc] connects b and c in a pretty weird way: first
it goes (within distance δ) with the radial segment [ba]; then at some point
[ba] and [ca] get very close to each other, and the shortest path [bc] “turns
back” and goes within distance δ along with [ac]. We suggest that the reader
make a sketch of this situation, for we will exploit it a lot (in particular in
the proof of the Morse Lemma).
Using (8.4), which has been derived solely from the inequality (8.3), it
is easy to show that (8.3) alone implies Gromov hyperbolicity:
8.4. Gromov Hyperbolic Spaces 287

Exercise 8.4.5. Show that if the inequality (8.3) is satisfied for all triangles
△a1 a2 a3 and all choices of p ∈ [a2 a3 ] for a strictly intrinsic metric d on X,
then (X, d) is Gromov hyperbolic.

Consider a triangle △abc and a point p (in a δ-hyperbolic space). Since

[bc] is contained in the δ-neighborhood of the union of [ab] and [ac], we
have d(p, [bc]) ≥ min(d(p, [ab]), d(p, [ac])) − δ. Combining this with (8.4), we
obtain that
(8.5) (b, c)p ≥ min((a, b)p , (a, c)p ) − 3δ.
Notice that the left-hand side of the inequality (8.3) (i.e., zero) can be
represented as (b, c)p (for in (8.3) p is supposed to belong to [bc]), and hence
(8.5) in its turn implies (8.3) with δ ′ = 3δ in place of δ.
Now we have arrived at Gromov’s original definition of Gromov hyper-
bolic spaces:
Definition 8.4.6. A metric space (X, d) is said to be Gromov hyperbolic
if there exists a δ ≥ 0 such that for every four points a, b, c, p ∈ X, the
following inequality holds:
(b, c)p ≥ min((a, b)p , (a, c)p ) − δ.

This elegant definition has an important advantage: it does not involve

shortest paths (sides). It uses only distance measurements and defines
Gromov hyperbolic metric spaces (which can even fail to be length spaces).
However, it is more difficult to visualize, and this is the reason why we began
with an alternative (though less general) definition. Definition 8.4.6 is also
sometimes harder to verify (because (8.3) is a particular case of (8.5)).
The reader can see that solving the following exercises, which state
that spaces with strictly negative curvature are Gromov hyperbolic, whereas
Euclidean spaces (except the real line) are not:
Exercise 8.4.7. Give a δ = δ(k) such that the hyperbolic plane of curvature
k is δ-hyperbolic.
Exercise 8.4.8. Show that a complete simply connected length space of
curvature bounded √ above by −k in the large, where k > 0, is a δ-hyperbolic
space for all δ ≥ 2/ k. Can you suggest a better value for δ?
Exercise 8.4.9. Prove that the Euclidean plane is not Gromov hyperbolic.

To get more insight into the geometric meaning of Definition 8.4.6, let
us look what it means for a 4-point space:
Lemma 8.4.10. Let X be a 4-point 0-hyperbolic space. Then X is isometric
to the set of leaves of a tree (with at most six vertices). Here the tree carries
288 8. Large-scale Geometry

an intrinsic metric (it is just a finite graph, by the way isometric to an H-

shaped subset of Euclidean plane formed by five segments, with its intrinsic
metric).
Lemma 8.4.11. Let X be a 4-point δ-hyperbolic space. Then X can be
mapped to the set of leaves of a tree (with at most six vertices) so that all
(the six) distances are distorted by no more than 2δ.
We leave the proofs to the reader as an exercise. To better visualize the
statement of Lemma 8.4.11, the following consideration maybe helpful. Let
a, b, c, d be four points (in a δ-hyperbolic length space). Choose a point e in
the intersection of δ-neighborhoods of [ab], [ac], and [bc]; similarly choose f
in the intersection of δ-neighborhoods of [ad], [ac], and [dc]. Let d(a, e) ≤
d(a, f ). Then all shortest paths connecting a, b, c, d lie in 2δ-neighborhood
of the tree formed by shortest segments [ae], [ef ], [f c], [be], [df ].
Exercise 8.4.12. Show that a 0-hyperbolic length space is a topological
tree (every two points are ends of exactly one subset homeomorphic to a
segment).
Furthermore, every Gromov hyperbolic space “on the large scale looks
like a tree”. The idea becomes clear from the following trivial observation:
if (X, d) is δ-hyperbolic, then (X, c · d), c > 0, is (δ/c)-hyperbolic (check it!).
To formalize this we need the following definition:
Definition 8.4.13. Let K = {p1 , p2 , . . . , pn } be a finite metric space. K is
said to be a finite subcone at infinity for X if there is a sequence Ni → ∞,
and a sequence of n-tuples of points q1i , q2i , . . . , qni ∈ X with
d(pil , pim )
Ni → 1 for all l, m ∈ {1, 2, . . . , n} as i → ∞.
d(qli , qm
i )

A metric space Y is said to be a subcone at infinity for X if every finite

subset of Y is a finite subcone at infinity for X.
Lemma 8.4.14. Let K be a finite subcone at infinity for X. Then K is
isometric to a subset of a finite tree with intrinsic metric on it. If Y is a
subcone for X at infinity, then Y is a topological tree.
The proof is an obvious corollary of the following important proposition,
which we leave as an exercise. This proposition characterizes Gromov
hyperbolic spaces by generalizing Lemma 8.4.11 from 4-point sets to all
finite subsets of a Gromov hyperbolic space:
Lemma 8.4.15. Let X be a Gromov hyperbolic space, and n ∈ N. There
exists a C ≥ 0 such that any n-element subset of X can be mapped to the
set of leaves of a finite tree so that all distances are distorted by no more
than C.
8.4. Gromov Hyperbolic Spaces 289

8.4.2. Bi-Lipschitz equivalence, quasi-geodesics, and Morse lemma.

It follows immediately from Definition 8.4.6 that if a metric space is within
finite Gromov-Hausdorff distance from a Gromov hyperbolic space, it is also
Gromov hyperbolic (perhaps for a different value of δ). Indeed, bounded
additive change in distance can be compensated by appropriately increasing
δ in the inequality (8.5) (we suggest that the reader carry out a rigorous
argument). Note that if one looks at Definition 8.4.1, it is not at all obvious
that the property of being Gromov hyperbolic persists if the distances are
changed by a uniformly bounded additive function.
Amazingly enough, for length spaces (and even for a more general class of
quasi-geodesic spaces) Gromov hyperbolicity also persists under bi-Lipschitz
equivalence. Recall that two metric spaces (X, d) and (X1 , d1 ) are said to
be bi-Lipschitz equivalent if there exist a bijective map f : X → X1 and a
positive constant C such that
C −1 d(x, y) ≤ d1 (f (x), f (y)) ≤ Cd(x, y)
for all x, y ∈ X.

Theorem 8.4.16. Let (X1 , d1 ) be a length space bi-Lipschitz equivalent to

a Gromov hyperbolic length space (X1 , d1 ). Then (X1 , d1 ) is also Gromov
hyperbolic.

Corollary 8.4.17. Gromov hyperbolicity persists under quasi-isometries.

Roughly speaking, this theorem asserts that if we multiply an intrinsic

metric of a Gromov hyperbolic length space by a bounded factor and obtain
a new length space, it is also Gromov hyperbolic. Thus we have a huge
source of examples of Gromov hyperbolic spaces: we can begin with the
hyperbolic plane and multiply its metric by a function bounded between two
positive constants (note that this new metric can now have a lot of positive
curvature!) We suggest that the reader have another look at Definition
8.4.6 to realize that this is a highly nontrivial phenomenon: the definition
requires that certain inequalities for distance functions hold up to an additive
constant δ; at the same time, we are allowed to multiply distances, even
though the latter can be arbitrarily large, by a bounded factor.
In the proof below we work with strictly intrinsic metrics. Using “almost
shortest paths”, the reader can easily adopt the argument to the situation
when shortest paths may fail to exist.
The proof of Theorem 8.4.16 is based on a very important phenomenon:
shortest paths in a Gromov hyperbolic space are stable (up to a uniformly
bounded displacement) under bi-Lipschitz changes of metric. This property
turns out to be equivalent to Gromov hyperbolicity. We precede it with the
following definition.
290 8. Large-scale Geometry

Definition 8.4.18. A path γ (in a length space (X, d)) is called C-quasi-
geodesic if L(γ)[s,t] ≤ C ·d(γ(s), γ(t)) for all s, t in the domain of γ. In other
words, the length of every segment of γ is at most C times longer than the
distance between its endpoints.

Of course, 1-quasi-geodesics are just shortest paths.

Definition 8.4.19. A length space X is said to be quasi-geodesically
stable if, for every C, there exists M with the following property: If
γ : [a, b] → X is a C-quasi-geodesic segment and [γ(a)γ(b)] a shortest path
between its endpoints, then (the image of) γ belongs to the M -neighborhood
of [γ(a)γ(b)].

Now Theorem 8.4.16 is immediately implied by the following assertion,

known as the Morse Lemma:
Theorem 8.4.20 (Morse lemma). Gromov hyperbolic spaces are quasi-
geodesically stable. More precisely, let (X, d) be a δ-hyperbolic length space
and C ∈ R. Then there exists a constant M = M (δ, C) with the following
property: If γ : [a, b] → M is a C-quasi-geodesic segment and [γ(a)γ(b)]
a shortest path between its endpoints, then (the image of ) γ belongs to the
M -neighborhood of [γ(a)γ(b)].

In Euclidean plane every semi-circle is (π/2)-quasi-geodesic; and a semi-

circle of a radius R gets as far as R away from the segment between its
endpoints. In contrast, for a Gromov hyperbolic world the Morse Lemma
asserts that if we want to travel in such a way that our path is never too
long a detour (not longer than C times the distance between points), we
are confined to travel within a bounded distance M from some shortest
path! Another striking corollary of the Morse Lemma is the following fact:
one cannot change large-scale behavior of geodesics in the hyperbolic plane
by bounded distortions of metric: if a metric is multiplied by a function
bounded away from 0 and infinity, new lines stay within a bounded distance
from hyperbolic lines.
As a matter of fact, the property expressed by Theorem 8.4.20 is equiv-
alent to Gromov hyperbolicity (see [BM]) as follows:
Definition 8.4.21. A length space (X, d) is said to be Gromov hyperbolic
if it is quasi-geodesically stable.

Proof of the Morse Lemma. The proof consists of several steps, and we
advise the reader to try to use them as exercises before reading our proofs.
Though the proof contains ugly estimates, they are needed only to formally
show that certain distances play no role for our large-scale consideration.
8.4. Gromov Hyperbolic Spaces 291

For each distance, it is very useful to realize whether it is “of order δ” or it

can become comparable to the size of the large object we are looking at.
First we want to look at the following situation: there are two points b
and c on the boundary of a ball of radius R (centered at a point a), and the
points are connected by a path γ never entering the ball. As usual, we think
of R as a large number (when compared to δ). We are going to show that if
the distance d(b, c) is not too small, then γ is much longer than the distance
d(b, c) (and hence cannot be a C-quasi-geodesic for a small C). The reason
is that large spheres remain “rather convex” (similarly to the hyperbolic
plane, and unlike the Euclidean case when large spheres get more and more
flat). More precisely:
Lemma 8.4.22. In the above notation, let L be the length of γ. Assume
that L ≥ R ≥ k 2 δ for a natural k. Then d(b, c) ≤ 10k −1 L, and in particular
γ cannot be a 10k −1 -quasi-geodesic.

Proof. Choose points b0 = b = γ(t0 ), b1 = γ(t1 ), . . . , bn = c = γ(tn ) along γ

so that the length of each segment γ|[ti ,ti+1 ] is between R R
k and 2k . Obviously
the number of points n is bounded by
2kL
(8.6) n≤ .
R
Let b′i ∈ [bi a] with d(b′i , a) = R(1 − k2 ). By the triangle inequality
d(b′i , a)
≤ (bi , bi+1 )a . By Proposition 8.4.3, segments d(b′i , b′i+1 ) ≤ 3δ. Now
consider the following broken line connecting b and c:
b = b0 , b′0 , b′1 , b′2 , . . . , b′n , bn = c.

The length of segments [bb′0 ] and [b′n b] is 2R

k , and the length of each
segment [b′i b′i+1 ] is at most 3δ. Thus the total length of the broken line,
which gives us an upper bound on the distance from b to c, is at most
4R
k + 3nδ. Combining this with (8.6), we get
4R 6kL
d(b, c) ≤ + δ.
k R
δ
Using that L ≥ R and R ≥ k 2 δ (and hence R ≤ k1 ), we re-write this as
4L 1 10
d(b, c) ≤ + 6L = L.
k k k
¤

The next proposition is motivated by the following consideration. Let

γ1 , γ2 be two hyperbolic rays orthogonal to the segment [γ1 (0), γ2 (0)]. As-
sume that the length of this segment is, say, 10: d(γ1 (0), γ2 (0)) = 10. Then
the distance d(γ1 (T1 ), γ2 (T2 )) is at least T1 + T2 .
292 8. Large-scale Geometry

Exercise 8.4.23. Prove this assertion.

Lemma 8.4.24. Let b, c, b1 , c1 ∈ X be such that d(b, b1 ) = R1 , d(c, c1 ) = R2 ,
d(b1 , c1 ) ≥ 6δ, and b1 , c1 are closest points to b, c in the segment [b1 , c1 ].
Then d(b, c) ≥ R1 + R2 − 6δ.

Proof. Let us make a preliminary observation: the intersection of [b1 c1 ] and

3δ-neighborhoods of [b1 b] is contained in a ball of radius 3δ centered at b1
(this immediately follows from the assumption that b1 is a point closest to b
in [b1 c1 ]). Similarly, the intersection of [c1 b1 ] and 2δ-neighborhoods of [c1 c]
is contained in a ball of radius 2δ centered at c1 .
By definition of δ-hyperbolicity, [b1 c] is contained in the δ-neighborhood
of [bb1 ] ∪ [bc]. Hence the δ-neighborhood of [b1 c] is contained in the 2δ-
neighborhood of [bb1 ] ∪ [bc]. Now [b1 c1 ] is contained in the δ-neighborhood
of [b1 c] ∪ [c1 c]. Hence [b1 c1 ] is contained in the 3δ-neighborhood of [b1 b] ∪
[bc] ∪ [cc1 ]. Since the intersection of [b1 c1 ] with [b1 b] ∪ [cc1 ] is contained in
the union of two balls of radius 3δ, whereas d(b1 c1 ) ≥ 6δ, we see that [b1 c1 ]
intersects the 3δ-neighborhood of [bc]. Let p ∈ [bc] be such that d(p, q) ≤ 3δ
for some q ∈ [b1 c1 ]. Then
d(b, c) = d(b, p) + d(p, c) ≥ d(b, q) − 3δ + d(c, q) − 3δ ≥ d(b, b1 ) + d(c, c1 ) − 6δ,
where the last inequality uses d(b, b1 ) ≤ d(b, q) and d(c, c1 ) ≤ d(c, q), since
b1 and c1 are closest points to b and c respectively in [b1 c1 ]. ¤

Now we are ready to prove the Morse Lemma (Theorem 8.4.20). Con-
sider a C-quasi-geodesic γ parameterized by arc length and connecting
p = γ(0) and q = γ(T ). Let 2R = maxt∈[0,T ] d(γ(t), [pq]), and choose τ
with d(γ(τ ), [pq]) = 2R.
Reasoning by contradiction, assume that R > k 2 δ, where k > 20C + 8
is a natural number (so in particular R > 400δ).
Let [t′ , t′′ ] be the biggest interval containing τ and such that
d(γ|[t′ , t′′ ], [pq]) ≥ R (in particular, d(γ(t′ ), [pq]) = d(γ(t′′ ), [pq]) = R).
It is obvious that |t′′ − t′ | (which is equal to L(γ, t′ , t′′ )) is at least 2R.
Choose t0 = t′ > t1 > t2 · · · > tn = t′′ so that R/2 ≤ ti − ti−1 ≤ R. Then
|t′′ − t′ | ≥ nR/2. Denote b = b0 = γ(t′ ) = γ(t0 ), bi = γ(ti ). Let bi1 be a point
closest to bi in [pq].
By Lemma 8.4.24 one has d(bi1 , bi+1 1 ) ≤ 6δ (recall that R ≫ δ). Thus
d(γ(t′ ), γ(t′′ )) ≤ 4R + 6nδ, and therefore

|t′′ − t′ | nR
C≥ ′ ′′
≥ .
d(γ(t ), γ(t )) 8R + 12nδ
Since R > 400Cδ, the last inequalities imply n < 9C.
8.4. Gromov Hyperbolic Spaces 293

Now set a = b01 ; then d(a, γ(t′ )) = R. Obviously γ|[t′ t′′ ] lies outside a
ball of radius R centered at a. Let c be the point in a segment [aγ(t′′ )] at
distance R from a. Since d(bn1 , γ(t′′ ) = R) and d(b01 , bn1 ) ≤ 4nδ ≤ 36Cδ, by
the triangle inequality
(8.7) d(c, γ(t′′ )) ≤ 36Cδ < d(γ(t′ ), γ(t′′ ))
and

1
(8.8) d(b, c) ≥ d(γ(t′ ), γ(t′′ )) − 36Cδ > d(γ(t′ ), γ(t′′ )).
2
(Note that these inequalities are very crude; but all we want is to show that
[cγ(t′′ )] is short and plays no role in our large-scale considerations.)
Then a path consisting of γ|[t′ t′′ ] continued by a segment [γ(t′′ )c] also
lies outside the ball of radius R centered at a. Applying Lemma 8.4.22 to
this path, we can estimate the length of γ|[t′ t′′ ] as
k
|t′′ − t′ | ≥ d(b, c) − d(γ(t′′ ), c).
10
Combining this with (8.7) and (8.8), we get
k
|t′′ − t′ | ≥ (
− 1) d(γ(t′ ), γ(t′′ )).
20
Since k/20 − 1 > C, this contradicts the assumption that γ is a C-quasi-
geodesic. ¤

8.4.3. Hyperbolic groups. A hyperbolic group is finitely generated group

which is a Gromov hyperbolic metric space with respect to some word metric.
Then such a group happens to be Gromov hyperbolic for any word metric,
and hence hyperbolicity is an algebraic property of a group. One can very
well think of the notion of a hyperbolic group as a generalization of basic
features of fundamental groups of negatively curved compact manifolds. Our
goal is not to give a systematic introduction to the theory of hyperbolic
groups; we will only discuss some of their main properties. In particular, we
will give three different definitions of hyperbolic groups and verify them for
the fundamental groups of negatively curved manifolds.
Let Γ be a finitely generated group; that is, there is a finite symmetrical
subset S ⊂ Γ such that every element of Γ can be represented as a finite
product of elements of S (here symmetry means that s ∈ S implies s−1 ∈ S).
Recall that the word distance between two elements γ1 , γ2 ∈ Γ is defined as
the smallest k such that γ1 γ2−1 can be represented as γ1 γ2−1 = s1 s2 . . . sk ,
si ∈ S. Of course, this is not an intrinsic metric, for it takes only integer
values, and topologically Γ with this metric is just a discrete space. However,
Γ with a word metric is isometrically embedded into a Cayley graph C(Γ)
294 8. Large-scale Geometry

of Γ. Recall that the latter has Γ as its set of vertices, and two vertices
γ1 and γ2 are connected by an edge (of length 1) if and only if γ1 s = γ2
for some s ∈ S. Then the word distance between two elements of Γ is the
natural intrinsic distance between them in the Cayley graph. Of course,
the Gromov-Hausdorff distance between Γ and C(Γ) is finite, and thus they
share all large-scale properties. We will usually prefer to deal with C(Γ)
since it is a length space.
Definition 8.4.25. A group Γ is said to be hyperbolic (with respect to some
finite generating set S) if the Cayley graph C(Γ) (equivalently, just Γ) is a
Gromov hyperbolic metric space.
Recall that any two word metrics are bi-Lipschitz equivalent (Proposi-
tion 8.3.18): if dS , dT are associated to generating sets S, T , then there is a
positive C such that
1
dS (x, y) ≤ dT (x, y) ≤ CdS (x, y).
C
Indeed, every element in S can be represented as a product of elements of
T , and vice versa—then we can choose for C the maximum length of such
representations.
Therefore, by Theorem 8.4.16, if Γ is hyperbolic with respect to some
word metric, then it is hyperbolic with respect to every word metric, and
we can speak of a hyperbolic group without specifying a metric. In other
words, to be hyperbolic is an algebraic property of a group, and it does not
depend on which metric we choose.
The first and dull example of a hyperbolic group is Z. Of course, if
Γ is hyperbolic, then so is Γ × G for a finite group G (just because the
Gromov-Hausdorff distance between Γ and Γ × G is finite—check this!)
Hence Z × G is also hyperbolic for any finite G. These examples, however,
lack the most interesting features possessed by all other hyperbolic groups.
However, already a free (nonabelian) group Fk with k ≥ 2 generators is a
characteristic representative of the hyperbolic world.
Exercise 8.4.26. Prove that a tree is a 0-hyperbolic space, and hence Fk
is a hyperbolic group.
Every hyperbolic group “on the large scale looks like a tree” (see Propo-
sition 8.4.14). This property happens to be characteristic for hyperbolic
groups, and the latter can also be defined as follows:
Definition 8.4.27. A finitely-generated group Γ is said to be hyperbolic if
every subcone at infinity for C(Γ) is a topological tree.
It is clear from Proposition 8.4.14 that this definition follows from
Definition 8.4.25; we omit the proof of the converse in our exposition.
8.4. Gromov Hyperbolic Spaces 295

In a sense, hyperbolic groups resemble free groups: when one looks at a

hyperbolic groups “from far away”, its relations “cannot be seen on the
large scale”. One can speculate that a “randomly chosen” group has very
high probability to be hyperbolic (for a large number of generators and
polynomially many random relations).
A huge source of examples of hyperbolic groups is given by the following
proposition:
Theorem 8.4.28. Let M be a compact Riemannian manifold with sectional
curvature bounded above by a negative number (more generally, a compact
Alexandrov space of curvature bounded above by −k, k > 0). Then the
fundamental group π1 (M ) is hyperbolic.

Proof. The Globalization Theorem for nonpositively curved spaces (Theo-

rem 9.2.9 in Chapter 9) implies that the universal covering space M̃ of M
(with the metric lifted from M ) has curvature ≤ −k in the large. Then by
Exercise 8.4.8, M̃ is Gromov hyperbolic. By Corollary 8.3.20, the funda-
mental group π1 (M ) is quasi-isometric to M̃ . Then Corollary 8.4.17 implies
that π1 (M ) is also Gromov hyperbolic. ¤

Linear isoperimetric inequality. It turns out that hyperbolic groups

possess an unexpected property, which is also characteristic for this class
of groups. This property also generalizes a feature of negatively curved
manifolds, and we motivate our definitions by the following proposition:
Proposition 8.4.29. Let M be a simply-connected Riemannian manifold
whose sectional curvature is bounded from above by −k, k > 0. Then there
is a number C = C(k) such that, for every closed curve α of length L > 1,
there exists an immersed disc β whose Riemannian surface area is at most
CL and whose boundary is α: ∂β = α. More generally, if M is a simply-
connected length space of curvature bounded above by −k, k > 0, then there
is a number C = C(k) such that every map of a circle to M with length L > 1
can be extended to the disc enclosed by the circle so that the 2-dimensional
Hausdorff measure of the image is at most CL.

Note that, in contrast with this inequality, a Euclidean circle of length

1 2
L cannot be filled by a surface whose area is less than L : this is the area
4π
of the flat disc bounded by the circle.
We leave the proof of the proposition to the reader as an exercise. Here
is the idea of an argument. It is based on the following easy observation.
Consider a curve of length 1 in the hyperbolic plane of curvature k, and
choose a point p. Shortest paths connecting p with points of the curve form a
“cone-shaped” region. Then the area of this region is bounded by a number
296 8. Large-scale Geometry

independent of the choice of p (the reader used to “hyperbolic” intuition

should not be surprised that while one moves p arbitrarily far away from
the curve, the area of this region does not increase to infinity). Indeed, if the
curve is a segment, the “cone-shaped” region is a triangle, and we remember
that the area of a triangle cannot exceed π/k by the Gauss-Bonnet formula.
Now the reader easily obtains an upper bound for an arbitrary curve by
enclosing it into a polygon (fixed for each value of k).
Now to prove the proposition one can choose a point p in M and again
consider a cone with its vertex at p by connecting it with all points of
the curve by shortest paths. This gives us a topological disc filling in
the curve. To estimate its area, one can compare it with a cone over an
appropriately chosen curve of the same length in the hyperbolic plane of
curvature k (this curve will not be closed any more, but it would have the
same distance function to some point as α to p, both with respect to their
natural parameters).
Let us formulate a “discrete analog” of the linear isoperimetric inequality
given in the proposition. Let Γ be a group with a symmetrical generating
set G = {g1 , g1−1 , . . . , gk , gk−1 } and relations R1 , . . . , Rl (where each Ri is
a word in letters g1 , g1−1 , . . . , gk , gk−1 whose product in Γ is identity). Let
ω = w1 w2 . . . wn be a word in letters g1 , g1−1 , . . . , gk , gk−1 . As usual, by its
value we mean the product w1 · w2 · · · · · wn in Γ. The value of the empty
word is set to be identity.
By a simple modification of ω we mean one of the following three
operations: inserting one of the words R1 , . . . , Rl anywhere in ω, crossing
out a subword of ω identical to one of the words R1 , . . . , Rl , or crossing out
a generator and its inverse (gi and gi−1 ) if they appear next to each other
in ω. Of course, simple modifications do not change the value of ω. It is a
standard fact in the theory of finitely-presented groups that two words have
the same value if and only if one of them can be transformed into the other
by a sequence of simple modifications (the reader can prove this as an easy
exercise).

Definition 8.4.30. Γ is said to satisfy a linear isoperimetric inequality if

there is a constant C such that every word ω = w1 w2 . . . wn whose value is
the identity can be transformed into an empty word in at most Cn simple
modifications.

It turns out that every hyperbolic group satisfies a linear isoperimetric

inequality. It is more difficult to show that this property actually charac-
terizes hyperbolic groups. We formulate this as the following alternative
definition of hyperbolic groups:
8.4. Gromov Hyperbolic Spaces 297

Definition 8.4.31. A finitely-presented group is hyperbolic if it satisfies a

linear isoperimetric inequality.

We will not prove here that this definition is equivalent to the two
previous definitions (while showing that hyperbolic groups satisfy a linear
isoperimetric inequality is not difficult and can be suggested to the reader as
an exercise, a proof of the converse is somewhat involved). Let us, however,
discuss isoperimetric inequality for groups in a more geometrical context.
Recall that every finitely presented group is the fundamental group of a
certain finite two-dimensional cell complex. Indeed, begin with the bouquet
of k circles. Choose an orientation for each of the circles and label the
circles by pairs of generators (gi , gi−1 ). Then glue a two-cell for each relation
following the word representing this relation. This means that the boundary
circle of the cell is glued to a curve traversing the circles in the bouquet in the
same order as their labels follow in the word representing the relation (and
choosing in which direction we follow a circle labeled by (gi , gi−1 ) depending
on whether we met gi or its inverse). The Van Kampen theorem tells us
that the fundamental group of this cell complex K is Γ. If each circle comes
with an intrinsic metric, and each two-cell is represented by a polygon glued
along isometries of its sides, K turns into a length space. Its universal cover
K̃ is quasi-isometric to Γ, and hence it is Gromov hyperbolic (note that
showing that K̃ is Gromov hyperbolic, for instance by introducing a metric
of negative curvature on K̃, is in its turn a method of arguing that Γ is
hyperbolic; see Exercise 8.4.32).
Let us fix a vertex (1-cell) p in K̃. Then a word ω in generators
g1 , g1−1 , . . . , gk , gk−1 determines a path γ in K̃ starting from p and following
the edges of the 1-dimensional skeleton of K̃ with labels (lifted from K)
corresponding to letters in ω. The value of ω is the identity if and only if
this path γ is closed (returns back to p).
What is the meaning of a simple modification of ω in terms of γ?
Removing a subword identical to one of the relations means contracting
a loop of γ going around one 2-cell; inserting a relation means adding such
a loop; finally, crossing out a generator and its inverse next to each other
corresponds to contracting a trivial loop (when γ follows an edge and then
immediately comes back). Therefore a sequence of simple modifications can
be realized by a homotopy of γ, and the number of 2-cells swept by γ in the
course of this homotopy is no more than the number of simple modifications
in this sequence. Recall that contracting a closed curve to a point sweeps a
topological disc. Now if γ is closed (that is, the value of ω is the identity), the
number of simple modifications required to transform ω into an empty word
is the same as the combinatorial area (the number of 2-cells) in a topological
disc bounded by γ. Hence we see that a linear isoperimetric inequality for
298 8. Large-scale Geometry

Γ exactly corresponds to a geometrical linear isoperimetric inequality (like

the one we began this section with) for K̃.
Exercise 8.4.32. Show that a group defined by six generators a1 , a2 , . . . , a6
(here the inverse generators are not listed) and one relation a1 a2 a3 a4 a5 a6 = e
is hyperbolic.
Hint: Construct K̃ as in the discussion above. Note that geometrically it
is just a collection of hexagons glued along their sides. Choose the hexagons
to be copies of a regular hyperbolic hexagon. Assuming that the hexagon is
small (and hence its angles are close to 23 π) argue that K̃ has Alexandrov
curvature bounded from above by −1.
Exercise 8.4.33. Use a direct argument to show that the fundamental
group of a negatively-curved compact manifold satisfies a linear isoperimetric
inequality.
Hint: This can be done by a refinement of the argument suggested for the
proof of Proposition 8.4.29 by representing the manifold as a cell complex
(and looking at its 2-dimensional skeleton embedded into the universal cover
of the manifold).

We conclude this section with a few remarks, which can perhaps ignite
the reader’s curiosity and push her to systematic study of this subject. It
is well known that the word problem (determining whether the value of
a given word is the identity) is algorithmically undecidable for a general
finitely presented group. Using a linear isoperimetric inequality, it is easy to
see that this problem is always decidable for hyperbolic groups; moreover, a
more delicate analysis involving the Morse Lemma shows that it can always
be decided in linear time. As a matter of fact, hyperbolic groups belong
to the class of automatic groups, the groups whose multiplication can be
checked by a finite automaton. The reader familiar with this notion can try
to re-prove this (elementary, however tricky and very elegant) result of W.
Thurston.

8.5. Periodic Metrics

8.5.1. Asymptotic cones of abelian groups. Let G be a finitely gen-
erated abelian group equipped with an invariant metric d. We will write
d(g) instead of d(0, g) for g ∈ G. It is known from the group theory that
G can be decomposed into a sum ZN ⊕ G0 where N ≥ 0 is the rank of
G and G0 is some finite group (actually the set of finite-order elements of
G, called the torsion). Then G is at finite Hausdorff distance from its ZN
component. Such spaces are equivalent for the purposes of the large-scale
geometry, which we are concerned with this section. Thus we simply assume
that G ≃ ZN . Then G can be represented as a lattice in an N -dimensional
8.5. Periodic Metrics 299

vector space V . The reader may keep in mind the picture of ZN as the set
of integer points in RN , but remember that the Euclidean structure is irrele-
vant. (In formal algebraic language, the ambient vector space V is obtained
from G by tensor multiplication, V = G ⊗ R.)
d(nv)
Proposition 8.5.1. For any v ∈ G, the limit limn→∞ n exists (n is
assumed natural).

Proof. The statement follows from the following lemma, which is a baby
version of the subadditive ergodic theorem:
Lemma 8.5.2. Let {xn }∞n=1 be a sequence of nonnegative real numbers such
that xm+n ≤ xm + xn for all m, n. Then the limit limn→∞ n1 xn exists.

Proof of the lemma. Define α = inf n xnn . We will show that xn /n → α

as n → ∞. Fix an ε > 0. Then there exist an m such that xm /m < α + ε.
Every natural number n can be represented in the form n = km + r where
k is a nonnegative integer and 0 ≤ r < m. From the assumptions of
the lemma, we have xkn ≤ kxn for all k, n (by induction). Hence xn ≤
kxm + xr ≤ km(α + ε) + xr ≤ n(α + ε) + C(m) where C(m) = max0≤r<m xr .
Hence lim supn→∞ xn /n ≤ α + ε. Since ε is arbitrary, we conclude that
lim supn→∞ xn /n ≤ α = inf n (xn /n), and the lemma follows. ¤

To derive the proposition from the lemma, let xn = d(nv). The condition
xm+n ≤ xm + xn follows from the triangle inequality. ¤

The following proposition expresses an important property of the func-

tion v 7→ limn→∞ d(nv)
n . Namely, this function can be extended to a semi-
norm on the ambient vector space V .
Proposition 8.5.3. There exists a unique semi-norm k · k on V such that
kvk = limn→∞ d(nv)
n for all v ∈ G.

Proof. We may assume that G = ZN and V = RN . Let | · |E denote

the Euclidean norm in RN and {ei }N i=1 be the standard set of generators.
d(nv)
The formula kvk = limn→∞ n defines a function on ZN . Moreover this
function is positively homogeneous and satisfies the triangle inequality on
ZN . We can extend this function to QN defining kv/nk = kvk/n for all
v ∈ ZN , n ∈ N. The extended function is also positive homogeneous
and satisfies the triangle inequality. Moreover it is a Lipschitz function
on QN N
P⊂ R (with respect to P the Euclidean norm). Indeed, for an x ∈ QN ,
x = xi ei , we have kxk ≤ |xi |kei k ≤ N maxi {kei k} · |x|E . Since QN
is dense in RN , this function has a unique continuous extension to RN .
Positive homogeneity and the triangle inequality remain true by continuity.
Hence k · k extended to RN is a semi-norm. ¤
300 8. Large-scale Geometry

The semi-norm k · k from the above proposition is called the asymptotic

norm, or the stable norm of the metric d. In all interesting cases, the stable
is really a norm (i.e., is strictly positive on nonzero vectors). For example,
let d be an orbit metric of a co-compact free totally discontinuous action of
G on a length space. Then by Theorem 8.3.19 d is bounded below by the
standard Euclidean norm multiplied by some positive constant. This bound
is inherited by the stable norm; thus in this case the stable norm is indeed
a norm.
Theorem 8.5.4. Let k · k be the stable norm of an invariant metric d on
G ≃ ZN . Assume that k · k is a norm. Then
(1) d(v)/kvk → 1 as kvk → ∞ uniformly in v ∈ G.
(2) The asymptotic cone of (G, d) is isometric to (V, k · k).

Proof. First observe that the first statement implies the second one. In-
deed, one has to prove that (G, λd) −→ GH (V, k · k) as λ → 0 in the sense
of pointed space convergence (the distinguished point in 0). In fact, the
maps fλ : (G, λd) → (V, k · k) given by fλ (v) = λv satisfy the requirements
of the definition of pointed space convergence (Definition 8.1.1). The only
nontrivial part is that for every R > 0 the distortion of fλ within the ball of
radius R goes to zero (as λ → 0), and this follows from the first statement
of the theorem.
It remains to prove the first statement. It is trivial that kvk ≤ d(v) for all
v ∈ G. On the other hand, by Theorem 8.3.19 we have d ≤ Ck · kE for some
constant C > 0. Fix an ε > 0 and let S be a finite ε-net of rational vectors
(i.e., S ⊂ QG) in the unit ball of (V, k · k). There exists a large natural
number M such that M v ∈ G and moreover d(M v) ≤ (1 + ε)M kvk ≤
(1 + ε)M for all v °∈ S (cf. Proposition
° 8.5.1). For every w ∈ G there exists
a v ∈ S such that °w/kwk − v ° < ε.
Let k be the integer such that k ≤ kwk < k+1, then kw−kvk ≤ εkwk+1,
then
d(M w) ≤ d(kM v) + d(M w − kM v) ≤ (1 + ε)kM + CM (εkwk + 1),
hence d(M w)/kM wk ≤ 1 + ε + Cε + M/kwk. Note that this is an upper
bound on d/k · k for all vectors divisible by M in G. There is a constant
C1 = C1 (M ) such that every v ∈ G has a vector divisible by M in C1 -
neighborhood of it. It follows easily that d(v)/kvk ≤ 1 + ε(C + 2) for all
v ∈ G outside the ball of some radius R = R(M, ε). Since ε is arbitrary, the
first statement of the theorem follows. ¤
Remark 8.5.5. Let d be an orbit metric of a co-compact totally discontin-
uous action of G on a length space X. Recall that the same action defines
a variety of orbit metrics depending on the choice of the orbit. However,
8.5. Periodic Metrics 301

the difference between any two orbit metrics is bounded by a constant, so

the stable norms of all these metrics coincide. Since the Hausdorff distance
between an orbit and the entire space X is finite, (V, k · k) is also the as-
ymptotic cone of X. The norm k · k is also called the stable norm of X, or
of the quotient space X/G (with respect to a given action or covering).

Remark 8.5.6. The first statement of the theorem remains valid without
the assumption that k · k is a norm. (The proof is essentially the same.) The
second statement requires an obvious correction if k·k is not a norm. Namely
one has to apply the standard factorization to make (V, k · k) a metric space.
The result is a normed vector space of a smaller dimension.

Exercise 8.5.7. Let d be a word metric on G defined by a (finite) set S of

generators. Prove that the unit ball of the stable norm is the convex hull of
the set S ∪ (−S) in V .

Exercise 8.5.8. Let d be a word metric on G and k · k be its stable norm.

Prove that there is a constant C such that kxk ≤ d(x) ≤ kxk + C for all
x ∈ G.

Remark 8.5.9. The statement of the last exercise implies that the Gromov–
Hausdorff distance between (G, d) and its asymptotic cone (V, k · k) is finite.
This holds not only for word metrics on G ≃ ZN but for any length space
X admitting a co-compact action of ZN by isometries ([Bur]). The proof
of this fact in full generality is beyond the scope of this book. In the next
subsection we prove it in the special case X = R2 and G = Z2 (see Corollary
8.5.13).

8.5.2. Periodic metrics on the plane. In this section we consider uni-

versal covers of two-dimensional tori. The two-dimensional torus T 2 is rep-
resented as the quotient R2 /Z2 where Z2 acts on R2 by parallel translations.
The projection p : R2 → R2 /Z2 = T 2 is a covering map. As we have seen in
the previous sections, it defines a 1-1 correspondence between length met-
rics on T 2 and length metrics on R2 that are invariant under the action of
Z2 . We call the latter class of metrics Z2 -periodic ones. A typical example
is a Riemannian metric on R2 defined by a bi-periodic metric tensor (the
reader may visualize it by thinking of a “wavy surface” whose bumps are
arranged periodically by repeating the same pattern). We are going to see
that the large-scale geometry of such a surface is the same as of a certain two-
dimensional normed space; moreover, this surface cannot be distinguished
from the space by distance measurements whose error exceeds some number
(notice that we mean the absolute precision, as opposed to the relative one).
This conclusion is actually true in any dimension, although we will use some
essentially two-dimensional arguments for the sake of simplicity.
302 8. Large-scale Geometry

The fundamental group of T 2 is Z2 . If γ : [a, b] → T 2 is a closed curve

(loop) and γ̃ is a lift of γ in R2 , then the vector [γ] := γ̃(b) − γ̃(a) is the
element of Z2 represented by γ. (Note that this vector does not depend on
the choice of γ̃.) It is called the rotation vector of γ. Two loops have equal
rotation vectors if and only if they belong to the same free homotopy class.
Let T 2 be equipped with a length metric d and let d˜ denote the corre-
sponding Z2 -periodic metric on R2 . For every v ∈ Zn we denote by ℓ(v) the
length of a shortest loop in T 2 whose rotation vector is v. In terms of the
˜ this can be written as follows:
metric d,
˜ x + v).
ℓ(v) = min d(x,
x∈R2
˜ x+v) equals the length of a shortest path from x to x+v whose
Indeed, d(x,
projection to T 2 is a shortest loop with rotation vector v and endpoints at
p(x).
˜ x + v) is a continuous bi-periodic function on R2 ;
The function x 7→ d(x,
hence it attains its maximum and minimum. This validates the right-hand
side of the formula for ℓ(v) and also implies that there is a constant C such
˜ x + v) ≤ ℓ(v) + C for all x ∈ R2 .
that d(x,
˜ Theorem 8.5.4 can be reformulated
Let k·k be the stable norm of (R2 , d).
as follows:
˜ y)
d(x,
lim =1 (x, y ∈ R2 ).
kx−yk→∞ kx − yk
˜ x + v) ≤ ℓ(v) + C implies that
The definition of k · k and the estimate d(x,
˜ x + nv)
d(x, ℓ(nv)
kvk = lim = lim
n→∞ n n→∞ n
2
for a v ∈ Z . One of the purposes of this section is to improve these formulas.
The above considerations could apply to tori and periodic metrics in any
dimension. The following theorem is valid only in dimension 2.
Theorem 8.5.10. For every v ∈ Z2 and n ∈ N, one has ℓ(nv) = nℓ(v).
Remark 8.5.11. The theorem can be interpreted as follows: if γ is a
minimal loop in its free homotopy class in T 2 , then γ n (the same loop
passed around n times) is also minimal in its free homotopy class. Note
that the lift of a minimal loop is a shortest path in (R2 , d). ˜ Suppose that
γ is parameterized by arc length, γ : [0, L] → T 2 . Let γ0 : R → T 2 be an
L-periodic curve (i.e., such that γ0 (t + L) = γ0 (t) for all t) whose restriction
on [0, L] is γ. Then a lift γ̃0 of γ0 is a geodesic line—every interval of γ̃0 is
a shortest path.

Before proving the theorem, we formulate two immediate corollaries. As

we already mentioned, Corollary 8.5.13 can be generalized to any dimension.
8.5. Periodic Metrics 303

Corollary 8.5.12. kvk = ℓ(v) for all v ∈ Z2 .

¯ ¯
˜ y) − kx − yk¯ ≤ C
Corollary 8.5.13. There is a constant C such that ¯d(x,
for all x, y ∈ R2 .

Proof. Under an¯ additional condition

¯ that x − y ∈ Z2 , the existence of an
˜ y)−kx−yk¯ follows from the previous corollary. Since
upper bound for ¯d(x,
R /Z is compact, the statement follows for all x, y ∈ R2 .
2 2 ¤

Proof of Theorem 8.5.10. We will need the following topological lemma

Lemma 8.5.14. Let v ∈ Z2 be a nonzero vector and γ be a loop in T 2 with
rotation vector nv where n ∈ N, n > 1. Then γ is not simple, i.e., has self-
intersections. Moreover, γ can be decomposed into two loops (with endpoints
at one of the self-intersections) whose rotation vectors equal v and (n − 1)v.

Proof. This fact is well-known in elementary topology. Here we outline one

of the possible proofs.
Let A denote the quotient of R2 by the group of integer multiplies of nv.
This quotient is homeomorphic to the cylinder S 1 ×R so that the translation
by v in A corresponds to the rotation by 2π/n. We denote the rotation by
R. There is a natural covering map from A to T 2 that commutes with the
two projections from R2 to A and T 2 . Let γ̃ be a lift of γ in A. Then γ̃ is a
closed noncontractible curve. Such a curve separates the cylinder’s “ends”
from each other; i.e., points of S 1 × R with the last coordinate near +∞ and
those with the last coordinate near −∞ belong to different components of the
curve’s complement. Consider the curve R(γ̃) in S 1 × R. We will show that
it intersects γ at some point. Suppose the contrary; then R(γ̃) is contained
in one connected component of the complement of γ̃. Let x+ and x− be two
points on (the image of) γ̃ in S 1 × R where the last coordinate attains its
maximum and minimum, respectively. The points R(x+ ) and R(x− ) are on
the curve R(γ̃). On the other hand, R(x+ ) and R(x− ) belong to different
components of the complement of γ̃ because they can be connected to the
+∞ and −∞ ends, respectively. This is a contradiction.
Thus R(γ̃) intersects γ̃. In other words, there is a point x in γ̃ such that
y = R(x) also belongs in γ̃. These two points split γ̃ into two arcs whose
projections to T 2 are closed loops with rotation vectors ±v and (n ∓ 1)v.
If these rotation vectors are v and (n − 1)v, the lemma follows. Otherwise,
we can apply the same argument to the loop with the rotation vector
(n + 1)v, and iterate this until we obtain a sub-loop with rotation vector v
(not −v). This must eventually happen because otherwise γ have sub-loops
with arbitrary large rotation vectors and this contradicts the compactness
of its lift in R2 . ¤
304 8. Large-scale Geometry

Now the theorem follows by induction in n. It is trivial for n = 1. If

n > 1, let γ be a minimal loop representing the vector nv. By the lemma, we
can decompose γ into two loops γ1 and γ2 with rotation vectors v and (n−1)v
respectively. Then ℓ(nv) = L(γ) = L(γ1 )+L(γ2 ) ≥ ℓ(v)+ℓ((n−1)v) = nℓ(v)
using induction. The inverse inequality ℓ(nv) ≤ nℓ(v) is trivial. ¤
Exercise 8.5.15. Let d be a smooth Riemannian Z2 -periodic metric on R2 .
Prove that the unit ball of its stable norm k · k is strictly convex in the sense
that its boundary does not contain straight line segments. Equivalently,
kv + wk < kvk + kwk whenever v and w are linearly independent.
Hint: Any two loops in T 2 with linearly independent rotation vectors
have an intersection point.
Exercise 8.5.16. Let ℓ1 , ℓ2 and ℓ3 be straight lines in R3 that are parallel
to the three coordinate axes, intersect the open unit cube (0, 1)n , and do not
intersect one another. Prove that there is a Z3 -periodic Riemannian metric
in R3 such that
(1) ℓ1 , ℓ2 and ℓ3 are minimal geodesics.
(2) There are no other minimal geodesics except these lines and their
integer parallel translations.
(3) The stable norm of the metric is k · k1 , where k(x, y, z)k =
|x| + |y| + |z|.
Hint: Let the metric tensor be standard Euclidean at these lines and
sufficiently large outside appropriate neighborhoods of them. (This example
and related topics are discussed in [Ban].)

8.5.3. Asymptotic volumes of periodic metrics. Consider a Zn -

periodic metric d in Rn . This metric, or its restriction to Zn , defines
the stable norm k · k in Rn , and Theorem 8.5.4 implies that the quantity
d(x, y)/kx − yk tends to 1 as kx − yk or d(x, y) go to infinity. In other
words, the metric d “at large scale” is equivalent to the distance function
defined by k · k. (In fact, their difference is bounded above by a constant;
see Remark 8.5.9.) We denote the unit ball of the stable norm by D, i.e.,
D = {v ∈ Rn : kvk ≤ 1}.
We will study the volumes of balls of large radii in the metric d.
Definition 8.5.17. Let d be as above. Its asymptotic volume, denoted Ω(d),
is defined by
Vold (Br (x))
Ω(d) = lim
r→∞ rn
n
where x ∈ R is an arbitrary point and Br (x) is the ball of radius r in the
metric d.
8.5. Periodic Metrics 305

The first thing to prove is that the limit in the definition exists.
Proposition 8.5.18. The asymptotic volume Ω(d) exists and does not
depend on x ∈ Rn . Moreover, in the notation introduced above
Ω(d) = Vold (I n ) · µn (D)
where I n = [0, 1]n and µn is the standard Lebesgue (or Hausdorff ) measure
in Rn . In particular, Ω(d) is finite and positive.

Proof. We give a sketch of a proof leaving the details to the reader.

Theorem 8.5.4 implies that d(x, y) ∼ kx − yk as kx − yk → ∞. Hence
a ball Bx (r) of d is close to the set rD = {rv : v ∈ D} (i.e., the r-
ball of the norm k · k) in the following sense: for every given ε > 0,
(1 − ε)rD ⊂ Bx (r) ⊂ (1 + ε)D for all sufficiently large r. This allows
us to replace Br (x) by rD in the definition of the asymptotic volume.
The space Rn is split into cells obtained by translating the cube I n by
all integer vectors. All their volumes are equal because the metric is
periodic. Let Q(r) denote the union of cells contained entirely in rD.
Then Vold (rD) ∼ Vold (Q(r)) and µn (rD) ∼ µn (Q(r)) as r → ∞ because
the number of cells that intersect the boundary of rD is negligible. Since
Vold (Q(r))/µn (Q(r)) = Vold (I n )/µn (I n ), it follows that
Vold (I n ) Vold (I n )
Vold (Br (x)) ∼ µn (rD) = µn (D)rn , r → ∞.
µn (I n ) µn (I n )
The proposition follows. ¤
Exercise 8.5.19. Prove the above proposition for any Riemannian manifold
M whose metric is invariant under a proper co-compact action of Zn .
The formula for Ω(d) should be changed as follows: replace Vold (I n ) by
¯ is the quotient space of the action.
Vold¯(M/Zn ) where (M/Zn , d)

Our purpose is to prove that the asymptotic volume can be bounded

below by a constant depending only on dimension but not on the metric.
In fact, the minimal possible asymptotic volume equals the volume of the
Euclidean unit ball and is achieved for flat Euclidean metrics (this is proved
in [BI]). The lower bound given in the following theorem is not optimal;
however the mere existence of a common lower bound for all metrics is not
obvious.
Theorem 8.5.20. Ω(d) ≥ (2/n)n for any Zn -periodic Riemannian metric
d in Rn .

Proof. Let k · k be the stable norm of d, D its unit ball and Q be an affine
cube such that n1 Q ⊂ D ⊂ Q (such a Q exists due to Lemma 5.5.19). Since
D ⊂ Q, one has kv − v ′ k ≥ 2 whenever v and v ′ belong to opposite faces
306 8. Large-scale Geometry

of Q. The idea of the proof is the following: we let x = 0 replace the ball
Br (x) ≈ rD in the definition of Ω(d) by a smaller set nr Q. Then the volume
of nr Q is estimated below by means of the Besikovitch inequality.
Now we pass to a formal argument. Fix an ε > 0 and let r be so large
that
(1 − ε)kx − yk ≤ d(x, y) ≤ (1 + ε)kx − yk
whenever kx − yk ≥ r/n or d(x, y) ≥ r/n (see Theorem 8.5.4 and the
beginning of this section). Then the distance (in d) between the opposite
faces of nr Q is not less than 2(1 − ε)r/n. By the Besikovitch inequality it
follows that ³r ´
Vold Q ≥ (1 − ε)n (2r/n)n .
n
On the other hand, since nr Q ⊂ rD, we have kxk ≤ r and hence d(0, x) ≤
(1 + ε)r for all x ∈ nr Q. This means that our affine cube nr Q is contained in
the ball B(1+ε)r (0). Thus
Vold (B(1+ε)r (0)) Vold ( nr Q)
Ω(d) = lim ≥ lim inf ≥ (2(1 − ε)/n)n/(1 + ε)n .
r→∞ (1 + ε)n rn r→∞ (1 + ε)n r n

Since ε is arbitrary, the desired inequality Ω(d) ≥ (2/n)n follows. ¤

The following exercise improves the constant (2/n)n in the theorem.

Exercise 8.5.21. For a unit ball D of a norm in Rn define λ(D) by
µn (D)
λ(D) = sup{ : Q is an affine cube containing D}.
µn (Q)
Prove that Ω(d) ≥ 2n λ(D) for any periodic Riemannian metric in Rn where
D is the unit ball of the stable norm.
Hint: Combine the proofs of Proposition 8.5.18 and Theorem 8.5.20.
Namely, show that
1. Vold (rQ) ∼ rn Vold (I n ) · µn (Q) as r → ∞ for any affine cube Q;
2. Vold (rQ) ≥ (2r)n − o(rn ) as r → ∞ if an affine cube Q contains the
unit ball of the stable norm.
Exercise 8.5.22. Prove that for every integer n ≥ 2 and every C > 0 there
exists a periodic Riemannian metric d in Rn with Ω(d) ≥ C (compare with
Exercise 5.6.16).
 Chapter 9

Spaces of Curvature
Bounded Above

In this chapter we concentrate on properties specific for spaces with upper

curvature bounds. In Section 9.1 we deal with local properties of such spaces,
and Section 9.2 contains an introduction to the global geometry of Hadamard
spaces (that is, complete simply connected spaces of nonpositive curvature).
We will see that the local properties of spaces of curvature ≤ k almost do
not depend on k. However, the global properties can be very different for the
cases k ≤ 0 and k > 0. Actually not much is known about global structure
of spaces with arbitrary (possibly positive) curvature bounds, even if the
space is a Riemannian manifold or a surface in R3 . One of the reasons is
that these classes of spaces are defined in terms of local properties which
may not hold “in the large” (see e.g. Example 9.1.13.) In particular, these
classes are not stable w.r.t. Gromov–Hausdorff convergence.
In the case k ≤ 0 the situation is completely different: for simply
connected spaces the local curvature conditions imply that the same distance
and angle comparison properties hold for all (arbitrarily large) triangles.
This is the contents of Globalization Theorem 9.2.9. This theorem allows us
to obtain much more information about the global geometry than we have
in the general case.
For readers familiar with Riemannian geometry we point out that almost
all classical theorems on Riemannian manifolds of nonpositive (and strictly
negative) curvature have generalizations for general spaces of curvature ≤ 0
(resp. ≤ k < 0). These are theorems about the fundamental group, group of
isometries, existence of flat planes and asymptotic behavior at infinity (ideal
boundary). The proofs are very similar to ones in Riemannian geometry,

307
308 9. Spaces of Curvature Bounded Above

and this indicates that the differential technique in the Riemannian proofs
is not that essential. On the contrary, convexity arguments often play the
key role.

9.1. Definitions and Local Properties

9.1.1. Definitions. We begin with the definition of a space of curvature
≤ k, where k is an arbitrary real number. All components of this definition
can be found somewhere in Chapter 4; here we just collect them together.
Recall that the k-plane is the two-dimensional model space of constant
curvature k. Depending √ on the sign of k, this is either the Euclidean plane,
a sphere (of radius 1/ k) with its length metric, or a hyperbolic plane√of
curvature k. We denote by Rk the diameter of the k-plane, i.e., Rk = π/ k
if k > 0 and Rk = ∞ if k ≤ 0.
A comparison triangle for a triangle △abc in a length space is a triangle
△abc in the k-plane such that |ab| = |ab|, |ac| = |ac| and |bc| = |bc|.
When talking of comparison triangles, we always assume that the triangle’s
perimeter |ab| + |bc| + |ac| is less than 2Rk (of course, this imposes no
restriction if k ≤ 0). This assumption guarantees that a k-comparison
triangle △abc exists and is unique up to an isometry.
Definition 9.1.1. A space of curvature ≤ k is a length space X which can
be covered by a family of open sets {Ui }i∈I so that every Ui satisfies the
following:
1. Every two points in Ui can be connected by a shortest path in Ui .
2. For any a, b, c ∈ Ui such that |ab| + |bc| + |ac| < 2Rk and a point d
in any shortest path [ac], the “triangle condition” holds: |db| ≤ |db|, where
△abc is a comparison triangle for △abc in the k-plane and d is the point in
[ac] such that |ad| = |ad|.
A space of curvature bounded above is a length space in which every
point has a neighborhood where these two conditions are satisfied for some
k (possibly varying from one point to another). Equivalently, every point
has a neighborhood whose induced length metric is of curvature ≤ k for
some k depending on a point.

In fact, it is sufficient to verify the triangle condition only for d being a

midpoint between a and c (Exercise 4.1.11). Note that the definition does
not require the uniqueness of a shortest path [a, c] (or the uniqueness of a
midpoint). However this uniqueness follows automatically as we will see in
Subsection 9.1.3.
Definition 9.1.1 is a modification of Definition 4.1.9. Other definitions
of a nonpositively curved space (4.1.2, 4.1.9, and 4.3.1) can be similarly
9.1. Definitions and Local Properties 309

modified for the case of curvature ≤ k, and the four resulting definitions are
equivalent (Exercise 4.6.3).
An unpleasant part of our definition is the requirement that any two
points (in a neighborhood U ) can be connected by a shortest path. While
we mainly restrict ourselves to locally compact spaces where shortest paths
always exist, not locally compact spaces can arise in certain constructions.
In order to handle them, it is natural to have a more general version
of Definition 9.1.1 where midpoints are replaced by “almost midpoints”
(similarly to the “intrinsic versus strictly intrinsic” case in Subsection 2.4.3).

Definition 9.1.2. A space of curvature ≤ k is a length space that can be

covered by a family of open sets {Ui } so that each Ui satisfies the following:
For every three points a, b, c ∈ Ui , every ε > 0 and every ε-midpoint d
between a and b, the distance |bd| ≤ |bd| + f (ε), where d is the midpoint of
the side [ac] in a comparison triangle abc in k-plane, and f (ε) → 0 as ε → 0
(for a, b, c fixed).

In fact, just Definition 9.1.2 is the “standard” one, but we stick to

Definition 9.1.1 to avoid technical complications. Fortunately, the two
definitions are equivalent for complete spaces:

Proposition 9.1.3. For complete (more generally, locally complete) spaces

Definitions 9.1.1 and 9.1.2 are equivalent.

Proof. Suppose that a complete length space X has curvature ≤ k in the

sense of Definition 9.1.2. We only need to prove the local existence of shortest
paths in X. By Theorem 2.4.16 (more precisely, its “localized” version—
formulate and prove it yourself) it is sufficient to prove the existence of
midpoints. Let a, b ∈ U where U = Ui is from the formulation. We will
construct a midpoint between a and b as a limit of a Cauchy sequence of
“almost midpoints”.
Let {εn } be a sequence of positive numbers, εn → 0, and let dn be an
εn -midpoint for a, b. We apply triangle comparison to the triangle △abdn
and εm -midpoint dm . Then |dn dm | ≤ |dn d| + f (εm ), where d is the midpoint
in a comparison triangle △adn b. It is easy to see that |dn d| → 0.
(Indeed, the point dn in the k-plane belongs to the intersection of the
balls of radius |ab|/2 + εn centered at a and b. The diameter of this
intersection goes to zero since |ab| is fixed and εn → 0.)
Therefore we can choose a subsequence of indices n such that the
corresponding points dn form a Cauchy sequence. Since X is complete,
the subsequence converges to a point which is obviously a midpoint between
a and b.
310 9. Spaces of Curvature Bounded Above

Note that this midpoint d is unique: otherwise we could apply the

curvature condition to the triple a, b, d and once more midpoint d˜ to come
to a contradiction. ¤

9.1.2. Examples. In Chapter 4 we have already discussed several exam-

ples of graphs and polyhedra of nonpositive curvature. Even these first
examples show that a nonpositively curved space may have a very compli-
cated structure. For example (unlike in the case of nonnegative curvature),
the local dimension of a nonpositively curved space may vary from point to
point. (Though the notion of dimension is not defined yet, its meaning is
clear in simple examples like two-polyhedra).
Example 9.1.4. An open subset of a space of curvature ≤ k with its induced
length metric is itself a space of curvature ≤ k. In particular, an open subset
of Rn is a nonpositively curved space.
Note that the word “open” is essential. For example, if one removes
closed balls Brk (1/k) with rk = 1/(k + 1)2 , k = 1, 2, . . . , from the plane,
the resulting length space is not a space of nonpositive curvature. See also
Example 9.1.7 below.
Example 9.1.5. The direct product of spaces of curvature not greater than
k is a space of curvature not greater than max{k, 0}.

The proof is straightforward. Compare with Exercise 4.1.13

Note that a product of spaces of curvature ≤ k is not a space of curvature
≤ k if k < 0 unless one of the multiplied spaces is a single point. Indeed,
consider two arbitrary shortest paths in spaces X and Y . These shortest
paths (as subspaces of X and Y ) are isometric to intervals of R; hence their
product in X × X is a convex set isometric to a subset of R2 . Therefore
X × Y cannot have strictly negative curvature.
Example 9.1.6. The complement of an open round disc in the plane R2
(with its induced length metric) is nonpositively curved. (Exercise: verify
this.)
More generally, any locally simply connected subset X ⊂ R2 (with its
induced length metric) is nonpositively curved. In outline, the proof goes
as follows. By Jordan Curve Theorem, any triangle in X bounds a region
homeomorphic to the disc. Local simple-connectedness implies that, if the
triangle is small enough, the region that it bounds is contained in X. It then
follows that the sides of the triangle are concave arcs (in the usual planar
sense) because otherwise one could shorten them in the bounded region.
And the angle comparison property for such “concave” triangles is more or
less trivial (if one pulls the vertices apart to straighten the sides, angles
increase).
9.1. Definitions and Local Properties 311

The above example is a purely two-dimensional phenomenon. The next

one shows that similar constructions in R3 may fail.
Example 9.1.7. The complement of an open ball in R3 is not a space
of nonpositive curvature. For example, consider a triangle whose vertices
belong to the boundary of the removed ball. Its sides are shortest paths in
the sphere and its angles are greater than ones of the comparison triangle in
the plane.
Exercise 9.1.8. Is the complement of an open ball in R3 a space of curvature
≤ k for some k?

Answer: yes, for k = 1.

Example 9.1.7 shows that the completion of a nonpositively curved space
may fail to be nonpositively curved. In the following example the completion
of a nonpositively curved space does not have any upper curvature bound
at all.
Example 9.1.9. A circular cone in R3 with its origin removed is a flat
space and hence is nonpositively curved. Its completion—a cone with the
origin—is not a space of curvature bounded above.
Example 9.1.10. As explained in Chapter 5, a two-dimensional surface
(more generally, a Riemannian manifold) is a space of curvature ≤ k if and
only if its Gauss curvature is less than or equal to k everywhere.
Example 9.1.11. (Space forms) Every orientable closed surface of genus
n ≥ 2 can be equipped with a metric of constant negative Gauss curvature.
A surface of genus n can be (topologically) glued from a polygon with 4n
vertices as follows: mark the (cyclically ordered) sides of a 4n-gon by the
symbols a1 , b1 , a−1 −1 −1 −1
1 , b1 , · · · , an , bn , an , bn ; then glue together the sides
marked by the same letters, say a and a−1 . To obtain a surface equipped
with a metric, apply this gluing construction to a convex hyperbolic 4n-
gon satisfying the following conditions: the sides glued together have equal
lengths, and the sum of the polygon’s angles equals 2π.
It is easy to see that such a polygon exists; for example, it can be found
among regular polygons. Indeed, if a polygon is small, the sum of its angles
approximately equals that in a Euclidean polygon: (4n − 2)π > 2π. And
if it is large (i.e., if the vertices go to the ideal boundary), the angles tend
to zero. By continuity, there exists a regular 4n-gon whose sum of angles
equals 2π, see also 5.3.5. (In fact, there is a continuum of different polygons
suitable for this constructions.)
Gluing sides of such a polygon according to the rule yields an orientable
surface of constant curvature −1 which is locally isometric to the hyperbolic
312 9. Spaces of Curvature Bounded Above

plane (prove this). A constant

√ curvature k < 0 can be then obtained by
multiplying the metric by 1/ −k.
Closed surfaces of constant negative curvature are called (two-dimensi-
onal) space forms.
Exercise 9.1.12. Modify this construction
a) to obtain closed nonorientable surfaces of any genus n ≥ 3 with a
metric of constant negative curvature;
b) to obtain complete noncompact surfaces (with cusps) having finite
area.
Example 9.1.13. Take two unit spheres, make a small round hole in each
of them, and glue them to a thin cylinder attaching its boundary circles to
the boundaries of the holes. (We assume the lengths of the boundary cycles
are equal.) This yields a closed surface of curvature ≤ 1 which looks like a
dumbbell. It can be approximated by a smooth one if desired. Note that
such a surface can contain an arbitrarily short closed geodesic provided that
the cylinder is sufficiently thin.

The following nontrivial result allows us to obtain new examples of

nonnegatively curved spaces
Theorem 9.1.14 (S. Alexander, R. Bishop [AB1]). Let X and Y be
complete simply connected spaces of nonpositive curvature, and f : X → R
be a concave function. Then the warped product X ×f Y is a (complete
simply connected) space of nonpositive curvature.

9.1.3. Elementary properties.

Definition 9.1.15. Let X be a space of curvature ≤ k. A normal ball in X is
a metric ball U of radius less than Rk /2 satisfying the comparison conditions
from Definition 9.1.1. (Sometimes we also call such a ball a normal region
or a normal neighborhood.)
For a point p ∈ X, the supremum of numbers r such that Br (p) is
a normal ball is called the convexity radius at p and is denoted by r(p).
It is easy to prove (do this yourself) that the number r(K) defined as
inf{r(p) : p ∈ K} is positive for each compact set K. The number r(K) is
called the convexity radius of K.

The term “convexity radius” is motivated by the following

Proposition 9.1.16. Every normal ball is convex.

Proof. Let Br (p) be a normal ball, a, b ∈ Br (p). Since r < Rk /2, a ball
Brk (p) in the k-plane is a convex set. By the comparison property for a
9.1. Definitions and Local Properties 313

comparison triangle △pab, one has |pd| < r for any point d ∈ [ab]. Hence
|pd| ≤ |pd| < r for any d ∈ [a, b], and this means that [ab] ⊂ Br (p). ¤

All the statements in the next proposition follow almost immediately

from the definitions.

Proposition 9.1.17. Let X be a complete locally compact space of curvature

≤ k and U ⊂ X be a normal ball. Then
(1) For every two points a, b ∈ U there is a unique shortest path
connecting these points, and this path is contained in U .
(2) If γ1 , γ2 are curves in U , then the shortest path [γ1 (t)γ2 (s)] depends
continuously on t, s.
(3) Every ball Br (p) ⊂ U is convex.
(4) Let [ab], [ac] be two shortest paths in U starting at a. If ∡bac = π,
then the curve bac is a shortest path.
(5) Every geodesic contained in U is a shortest path.

Proof. (1) Suppose there are two shortest paths, [ab]1 and [ab]2 . For every
point c ∈ [ab]1 one can consider the triangle △abc and its comparison triangle
△abc. The latter one is degenerate: c ∈ ab. Let d be a point of [ab]2 such
that |ac| = |ad|. Then it has to be |cd| ≤ |cd| = 0, so c = d and [ab]1 = [ab]2
since c is arbitrary.
Since a limit of shortest paths is a shortest path, (2) follows from (1).
The proof of (3) is the same as for Proposition 9.1.16. And (4) follows
from the fact that a comparison triangle for △abc is degenerate with ∡bac =
π.
(5) Let γ : [0, L] → U be a geodesic but not a shortest path. Let
t ∈ [0, L] be the maximal value of the parameter such that the restriction
of γ to [0, t] is still a shortest path. Such a maximal t exists because the
restriction to [0, t] is a shortest path if and only if |γ(0)γ(t)| = t, and the
latter condition defines a closed set. Since γ is a geodesic, its restriction
to [t, t + ε] is a shortest path for a sufficiently small ε > 0. Applying the
statement (4) to the shortest paths γ|[0,t] and γ[t,t+ε] yields a contradiction
with the maximality of t. ¤

Remark 9.1.18. Proposition 9.1.17 implies that a normal ball U is con-

tractible. Indeed, fix a point p ∈ U ; then for every x ∈ U there is a unique
shortest path γx : [0, 1] → U (parameterized with constant speed) such that
γx (0) = x and γx (1) = p. Then a map H : U × [0, 1] → U defined by
H(x, t) = γx (t) is a homotopy between the identity map idU and the map
which sends all points of U to p. Geometrically, this homotopy moves every
314 9. Spaces of Curvature Bounded Above

point x ∈ U towards p along the corresponding shortest path. (Exercise:

prove that H is continuous.)
In particular, every space of nonpositive curvature is locally simply
connected and therefore has a universal covering space (cf. Theorem 3.4.11).

Simple examples, like flat tori or hyperbolic space forms, show that, in
general, the local curvature conditions do not imply analogous properties
in the large. This is related to the fact that these spaces contain a closed
geodesic. Tori and surfaces of higher genus are not simply connected, so
closed geodesics can be found by minimizing the length in a free homotopy
class of noncontractible closed curve. The space in Example 9.1.13 is simply
connected (it is homeomorphic to the sphere) but its convexity radius can
be made arbitrarily small while the curvature bound stays the same. We
will see later that this effect (small convexity radius in a simply connected
space) is possible only for k > 0.

9.1.4. Extreme cases of comparison conditions. The definition of a

space of curvature ≤ k says that distances and angles in (sufficiently small)
triangles must satisfy certain inequalities. The following proposition tells us
what happens if these inequalities turn into equalities.
Proposition 9.1.19. Let △abc be a triangle in a normal region of a space
X of curvature ≤ k and △abc a comparison triangle in the k-plane. Then
the following four conditions are equivalent:
(i) The angle at a of the triangle △abc is equal to the angle at a of △abc.
(ii) For some point x ∈ [ac], x 6= a, c, the equality |bx| = |bx| holds,
where x ∈ [ac] is a point such that |ax| = |ax|.
(iii) All angles of △abc are equal to the corresponding angles of △abc.
(iv) △abc spans a totally geodesic surface isometric to the triangle abc
with its interior (i.e., the region in the k-plane bounded by △abc). More
formally, there exists a distance-preserving map from the “full” triangle abc
to X which maps the sides of △abc to the respective sides of △abc.

Proof. Let α, β, γ denote the angles at a, b, c of △abc and α, β, γ the

respective angles of the comparison triangle.
1. It is obvious that (iv) implies the assertions (i)–(iii) and (iii) implies
(i).
2. (i) =⇒ (ii). Suppose α = α. Let x and x be as in the formulation. For
the triangle △abx and its comparison triangle △a0 b0 x0 one has ∡b0 a0 x0 ≥
α = ∡bax. Since |a0 x0 | = |ax| and |b0 x0 | = |bx|, it follows that |bx| =
|b0 x0 | ≥ |bx|. On the other hand, the curvature condition implies that
|bx| ≤ |bx|. Therefore |bx| = |bx|.
9.1. Definitions and Local Properties 315

3. (ii) =⇒ (iii). Suppose that |bx| = |bx| for some x ∈ [a, c] \ {a, c}.
First we show that the equality |by| = |by| holds for all y ∈ [a, c]. Since
|bx| = |bx|, the triangles bxa and bxc are comparison triangles for △bxa
and △bxc. Therefore ∡bxa ≤ ∡bxa and ∡bxc ≤ ∡bxc. On the other
hand, ∡bxa + ∡bxc ≥ ∡axc = π. Since ∡bxa + ∡bxc = π, it follows that
∡bxa = ∡bxa and ∡bxc = ∡bxc, so the assertion (i) holds for triangles
△bxa and △bxc. Since we have already proved that (i) implies (ii), we can
conclude that |by| = |by| for all y ∈ [ax] ∪ [cx].
Now applying the first variation formula (for k = 0 it is Theorem 4.5.6;
its generalization to arbitrary k is trivial) to the distance from b along the
geodesic [ac], we obtain that α = α and β = β. The remaining equality
γ = γ follows by a combination of the implication (i) =⇒ (ii) and the
already proved part of the implication (ii) =⇒ (iii).
4. Thus the assertions (i)–(iii) are equivalent. It remains to prove that
they imply (iv). The desired totally geodesic surface can be obtained by
sweeping it by a family of shortest paths connecting b and [ac].
Let x, y ∈ [ac]. Construct corresponding points x, y in [ac]. We assume
that y ∈ [x, c]. We have
β ≥ ∡abx + ∡xby + ∡ybc ≥ ∡abx + ∡xby + ∡ybc ≥ β = β.
Hence all these inequalities turn out to be equalities, in particular, ∡xby =
∡xby. Hence the assertions (i)–(iii) hold for △xby implying that the dis-
tances between points of [bx] and points of [by] are the same as between the
respective points in [bx] and [by] in the k-plane. Therefore the union of the
shortest paths {[bx]}x∈[a,c] is a desired “full” triangle isometric to △abc. ¤
Exercise 9.1.20. Let △ be a triangle contained in a normal region of a
space of curvature ≤ 0. Let a, b, c denote the lengths of the sides of △ and
α, β, γ the respective (opposite) angles. Prove the following inequalities:
α + β + γ ≤ π,
c ≥ a2 + b2 − 2ab cos γ,
2

c ≤ b cos α + a cos β.
Show that the equality in any of these inequalities implies that the assertions
from the above proposition holds.
Hint: Observe that these inequalities turn to equalities for a comparison
triangle in R2 ; then apply the angle condition.

9.1.5. Reshetnyak’s Gluing Theorem. [Resh] The following theorem

is a simple but very useful tool for verifying upper curvature bounds of
“compound” spaces. It also allows us to construct many nontrivial examples
of spaces with curvature bounded above.
 316 9. Spaces of Curvature Bounded Above

Theorem 9.1.21. (Reshetnyak’s Gluing Theorem.) Let {(Xi , di )}, i =

1, 2, be two (complete locally compact) spaces of curvature ≤ k. Suppose
that there are convex sets Ci ∈ Xi and an isometry f : C1 → C2 . Attach
these spaces together along the isometry f .
Then the resulting space (X, d) is a space of curvature ≤ k.
Remark 9.1.22. Iterating the construction described in Reshetnyak’s the-
orem one can obtain examples glued from more than two (or even countably
many) components.

Proof. To simplify notations we will not distinguish spaces Xi and their

images in X under natural projections. In particular, let C ⊂ X be the
common projection of Ci , i = 1, 2, to X. Convexity of C implies that
restrictions of d to Xi coincide with di , i = 1, 2. In particular, every point
p ∈ X has a neighborhood U ⊂ X such that Ui = U ∩ Xi is a normal ball in
Xi . Consider a triangle △abc in U . One may assume that a, b ∈ X1 , c ∈ X2 .
PSfrag replacements
C X1 c
X2 C
X1 c1 c
a c2 c2 c1
b b
c
c2 a X2
c1 a b
Figure 9.1: Reshetnyak’s theorem.

Then there are points c1 , c2 on the shortest paths [ac], [bc], resp. contained
in C. Decompose △abc into three triangles △ac1 c2 , △bac2 , △cc1 c2 . Place
their comparison triangles in the plane “in the natural” way (see Figure 9.1).
Usual angle comparison arguments show that the angles at c1 and c2 in the
polyhedron ac1 cc2 b are concave (i.e., not less than π). Loosely speaking, one
can consider this polyhedron like a triangle abc having two “concave sides”,
ac1 c and bc2 c. “Straightening this triangle” (compare with Alexandov’s
Lemma 4.3.3) we see that the angles of △abc are not greater than the angles
of its comparison triangle. ¤

9.1.6. Reshetnyak’s Majorization Theorem. [Resh] There is one more

important theorem due to Reshetnyak, the Majorization Theorem. This
theorem gives better insight into the nature of spaces of curvature bounded
above. We present it here without a proof because we never use it hereafter.
9.1. Definitions and Local Properties 317

Theorem 9.1.23. (Reshetnyak’s Majorization Theorem.) Let U be a nor-

mal ball in a space of curvature not greater k and γ be a closed rectifiable
2π
curve in U . If k > 0, then we suppose in addition that L(γ) ≤ √ k
. Then
there exists a closed convex set Ω in the k-plane bounded by a closed curve
Γ and a nonexpanding map f of Ω into U such that the restriction f˜ of f
on Γ is a length-preserving map onto γ.
Note that the converse assertion is correct as well: if the statement of
the theorem holds for every (sufficiently short) closed curve γ in a length
space M , then M is a space of curvature not greater than k. The proof of
the latter is simple: it is sufficiently to apply the statement to “triangular”
curves.

9.1.7. Extendibility of geodesics. Recall that a geodesic is a curve which

is locally a shortest path. We assume that all geodesics are naturally
parameterized.
Definition 9.1.24. A geodesic γ : [O, a] → X is extendible beyond the
point γ(a) if γ is a restriction of a geodesic γ̃ : [0, b] → X with b > a.
We will soon see that, loosely speaking, a geodesic in a complete space
of nonpositive curvature is always extendible unless its endpoint looks like
a boundary point of the space. The following simple statement can help to
understand what we mean.
Proposition 9.1.25. Let X be a complete space of nonpositive curvature.
If a geodesic γ : [0, a] is not extendible beyond the point p = γ(a), then the
punctured ball Bε (p) \ {p} is contractible for all sufficiently small ε > 0.
Note that the converse statement is wrong in general: there exists a
space having a boundary point such that all geodesics ending at the point
are extendible beyond it.
Example 9.1.26. Consider the Euclidean cone over the line. The metric
of the cone is locally Euclidean everywhere except the vertex p. Every
punctured ball centered at p is contractible, so p is a (unique) boundary
point of the cone. However every geodesic ending at p is obviously extendible
beyond p.
Exercise 9.1.27. 1. Prove that for a smooth surface in R3 boundary points
are exactly the points that have a contractible punctured neighborhood.
2. Give examples showing that this statement fails for metrics without
upper curvature bound.
Proposition 9.1.28. Let X be a complete metric space and assume that
every geodesic is extendible through all of its points. Then every geodesic γ
is a restriction of an arc-length parameterized geodesic γ : R → X.
318 9. Spaces of Curvature Bounded Above

In other words, every geodesic can be extended in both sides “to infin-
ity”. Note that no curvature restriction is assumed here.

Proof. Actually this fact is a trivial consequence of the simple implication

(i) ⇒ (iii) in the Hopf–Rinow Theorem; this implication does not require
locally compactness of X. Let (a, b) be a maximal open interval (possibly
infinite) to which γ can be extended. We want to prove that (a, b) =
(−∞, ∞). Suppose for instance that b < ∞. Since X is complete, γ can
be defined at b by the implication mentioned above. And the assumptions
of the proposition imply that γ is extendible beyond b. This contradict the
maximality of b. ¤
Definition 9.1.29. If the conditions of Proposition 9.1.28 hold (i.e., if
the space is complete and all geodesics are extendible), we say that X is
geodesically complete.

Since every geodesic is a shortest path as long as it is contained in a

normal ball (see Proposition 9.1.17), we have the following
Corollary 9.1.30. Let X be a geodesically complete space of curvature
bounded above. Then for every point p ∈ X there is a number r > 0 such
that all geodesics starting at p can be extended up to the boundary of the ball
Br (p).
Example 9.1.31. Consider a disc x2 + y 2 < 1 in R2 and add to it the point
(1, 0). This is a “flat” space, and the point (1, 0) is the only point where the
conclusion of the corollary fails.
Example 9.1.32. Consider a bouquet of disjoint segments [0, 1/i], i =
1, 2, . . . , attached to a point 0. This is a compact space of nonpositive
curvature. Ends of the segments are the only points where geodesics are not
extendible. And though all geodesics coming to 0 are extendible beyond 0,
the point 0 has no round neighborhood mentioned in the corollary above.

9.1.8. Spaces of directions and tangent cones. Space of directions.

Let X be a complete locally compact metric space, p ∈ X. Recall (see
Subsection 3.6.6) that the space of directions at p is the completion of the
metric space Σ′p whose points are equivalence classes of the shortest paths
starting at p and the distance between two classes is the angle between
corresponding shortest paths. This space is denoted by Σp . In the case of
either nonpositively or nonnegatively curved space, angles between shortest
paths do exist, so the space of directions is well-defined at every point.
Remark 9.1.33. Even if a space X is geodesically complete, it is possible
that Σ′p is not complete, i.e., some elements of Σp are not represented by
shortest paths. (This is quite surprising in view of Corollary 9.1.30.)
9.1. Definitions and Local Properties 319

For example, consider the (open!) half-plane {(x, y) : y > 0} with

the ray {(0, y) : y ≤ 0} added. Cut the half-plane along rays {(x, y) :
kx ≥ 1, y = xk } for all k ∈ Z \ {0} and glue a copy of the first quadrant
(x, y ≥ 0) into each cut.
The resulting space is homeomorphic to a half-plane with a ray added,
and it is geodesically complete (unlike the bare half-plane or the half-plane
with the ray added). The space of directions at (0, 0) consists of a single
point (corresponding to the vertical ray) and an interval of length π. The
endpoints of this interval are not represented by geodesics.
Exercise 9.1.34. Prove the above statements.

However, if a space X is geodesically complete and locally compact, then

Σp is compact at every point p ∈ X.
Exercise 9.1.35. Prove this.
Hint: Apply Corollary 9.1.30.

Even if X is locally compact and complete (as we usually assume for

simplicity), the direction spaces Σp are not necessarily compact:
Example 9.1.36. The bouquet of segments considered in 9.1.32 is a com-
pact space of nonpositive curvature. Its space of directions at 0 is not
compact—it is a countable discrete space with distances π between different
points.
Example 9.1.37. This example is a space of curvature bounded below (not
above!), so it rather falls under the subject of Chapter 10. Consider the
direct product of countably many round spheres of radii 1/i (i = 1, 2, . . . ).
(Exercise: give a definition of such a product and prove that it is a compact
space of nonnegative curvature.)
Its space of directions at each point is not locally compact—it is the
product of infinite number of copies of S 1 (exercise: formulate and prove
this).

Note that Σp is not always connected. The simplest example is the

line R: its space of directions at any point consists of two points with the
distance π between them. We will need the following
Definition 9.1.38. If every point of a metric space (X, d) has a neighbor-
hood such that restriction of the metric d to this neighborhood is an intrinsic
metric, we say that metric d is locally intrinsic.

The original (angular) metric of a space of directions Σp is not necessarily

intrinsic. One of reasons is that angles never exceed π while shortest paths
320 9. Spaces of Curvature Bounded Above

in a space of directions may be much longer than π. For example, consider

a cone with the total angle 4π around the origin. Its space of directions at
the origin is a circle of length 4π while all distances (angles) are not greater
than π.
So we equip a space of directions not only with its original (angular)
metric but also with the intrinsic metric induced by the original one.
As usual, the intrinsic distances may be infinite. This certainly occurs if
Σp consists of more than one component. Later we will see that in “good”
cases this is the only possible reason for the intrinsic metric to take infinite
values.

Lemma 9.1.39. If X is a complete locally compact space of curvature

bounded above and p ∈ X, then the angular metric of Σp is locally intrinsic.
More precisely, for every point a0 ∈ Σp and every positive r < π/2, the
restriction of the angular metric to the ball Br (a0 ) is intrinsic.

Proof. By Theorem 2.4.16 it is sufficient to prove that for every two points
a, b ∈ Br (a0 ) and every ε > 0 there is an ε-midpoint. Recall that Σp is
the completion of its subset Σ′p consisting of points represented by shortest
paths. Without loss of generality, one can assume that points a and b
belong to Σ′p . So they can be represented by shortest paths α and β,
α(0) = β(0) = p. Take an ε > 0 and choose points x = α(t), y = β(t)
so close to p that 0 ≤ ∡xpy − ∡xpy ≤ ε, where ∡xpy is the angle in a
comparison triangle △xpy. Let z be a midpoint between x and y. Our
assumption guarantees that d(a, b) < π, where d is the distance in Σ′p .
Hence z 6= p for sufficiently small ε. Place comparison triangles △xpz and
△ypz in the k-plane in different half-planes separated by the line pz and,
as usual, compare the resulting quadrilateral with a comparison triangle for
the triangle △xpy. Since |xy| ≤ |x, y|, we obtain ∡xpz + ∡ypz ≤ ∡xpy + ε.
With the obvious equality ∡xpz = ∡ypz, it gives
1 1
∡xpz ≤ ∡xpz ≤ (∡xpy + 2ε), ∡xpz ≤ ∡ypz ≤ (∡xpy + 2ε)
2 2
if points the x and y are sufficiently close to p.
Let c denote the point of Σ′p represented by a shortest path pz. Now the
last two inequalities can be re-written as
1 1
d(a, c) ≤ (d(a, b) + 2ε), d(b, c) ≤ (d(a, b) + 2ε).
2 2
By the triangle inequality for angles we have

d(a, c) + d(b, c) ≥ d(a, b) − ε.

9.1. Definitions and Local Properties 321

These three inequalities immediately imply that |d(a, c) − 12 d(a, b)| ≤ 2ε,
|d(b, c) − 21 d(a, b)| ≤ 2ε. Since ε is arbitrary, this proves the midpoint
property, so the metric of Σ′p is locally intrinsic. ¤

Tangent cone. Denote by Kp the cone over the direction space Σp (see
Subsection 3.6.2 for the definition of the cone over a length space). Every
point w ∈ Kp except its origin o is represented by a pair (ξ, r), where ξ ∈ Σp
and r = |ow|. The cone Kp is called the tangent cone at p. It is clear that
for a smooth surface in R3 all tangent cones are planes.

Remark 9.1.40. There is another approach to the notion of tangent cone,

namely the Gromov–Hausdorff tangent cone that we discussed in Section
8.2. It is defined as the Gromov–Hausdorff limit of pointed spaces (λX, p)
as λ → ∞. (Loosely speaking, one “blows up” our metric space X from a
fixed point p ∈ X and calls the limit space the tangent cone.)
In many cases (but not always!) the two definitions are equivalent. One
of the obstacles to such equivalence is noncompactness of the space of direc-
tions, which brings all difficulties related to Gromov–Hausdorff convergence
of not locally compact spaces. Moreover, the limit in the definition of the
Gromov–Hausdorff tangent cone may fail to exist even if the space of direc-
tions is compact.

Example 9.1.41. Begin with the ray R+ = {x ∈ R : x ≥ 0}, and attach

a segment of length 1/k at every point 1/k, k ∈ N. The space of directions
of the resulting space at 0 is a point, so the corresponding tangent cone is
a ray. On the other hand, the Gromov–Hausdorff tangent cone obviously
does not exist.
Now let us change this example: attach a segment of length x at every
point x ∈ R+ . The resulting space Y is a (not locally compact) space of
nonpositive curvature. The direction space of Y at the point 0 still is a
point. This time the Gromov–Hausdorff tangent cone at 0 does exist (in the
sense of Definition 8.2.1) and is isometric to Y , but it is not a cone over any
metric space.

Remark 9.1.42. In general, a small neighborhood of a point in a space of

curvature bounded above may have much more complicated structure than
the tangent cone. Loosely speaking, the space of directions (and the tangent
cone) may carry much less information than a small neighborhood of a point.
Contrary to this, in spaces with lower curvature bounds the tangent cone
is very similar (at least, homeomorphic) to a small metric ball centered at
a given point. This is the reason why the notion of tangent cone (defined
either in terms of the space of directions or via the Gromov–Hausdorff limit)
is more important for spaces with curvature bounded below.
322 9. Spaces of Curvature Bounded Above

Let X be a space of curvature bounded above, p ∈ X, U = Br (p) a

normal neighborhood of p. We define two maps, the logarithmic map logp
and the exponential map expp . The map logp is a map from U to the tangent
cone at p. Let x ∈ U . Connect p to x by a (unique) shortest path [px] and let
logp x = (ξ, r) ∈ Kp where ξ is the direction of [px] and r = |px|. It is easy
to prove that logp is continuous. We would like to define expp = (logp )−1 .
However logp may be not injective. So we let expp (x) be one (any) point of
the set log−1
p . Even in a small neighborhood V of the origin O ∈ Kp the
map expp may not be defined everywhere. However the set where expp is
defined is dense in V if V is small enough (prove this!).
If X is a space of nonpositive curvature, then logp is a nonexpanding
map. And if X is a space of curvature ≤ k, k > 0, then logp is a Lipschitz
map with a constant 1 + δ where δ goes to zero along with the diameter of
the neighborhood.
Remark 9.1.43. Now let X be a length space (not necessarily of curvature
bounded above), p ∈ X and Kp well defined. Then we can define maps logp
and expp almost as above. However now the situation is somewhat different:
we have no normal neighborhood of p and cannot guarantee uniqueness of
a shortest path [px]. So we choose any of such paths. As a result logp can
be discontinuous (just at points connected with p by two or more shortest
paths having different directions at p). Nevertheless we still can define
expp = (logp )−1 . And if X is a space of curvature bounded below, expp
is continuous and moreover Lipschitz (just like logp in the case of curvature
bounded above).
Theorem 9.1.44. If X is a complete locally compact space of curvature
bounded above, p ∈ X, then the tangent cone Kp is a length space of
nonpositive curvature.

Proof. The idea of the proof is very simple. One takes a triangle in Kp and
considers a sufficiently small homothetic one (with respect to a homothety
centered at the origin). The map expp allows us to associate a triangle in X
to a small triangle in Kp . These two triangles have almost the same lengths
of the respective sides (to the first order), so the curvature comparison
condition translates from the triangle in X to the one in Kp .
To make it more clear, let us assume for now that Kp is locally compact
and that each point of Kp is represented by a shortest path. Let △a1 a2 a3
be a triangle in Kp and a4 the midpoint of the side [a1 a3 ]. Denote by γi ,
i = 1, 2, 3, 4, the shortest paths representing points ai and by aλi the point of
the shortest path γi such that |paλi | = λ|Oai |. Here O is the origin of Kp and
λ is a small positive number. By the very definitions of the angular metric
and the cone’s metric, we have λ−1 |aλi aλj | ∼ |aj aj | as λ → 0 (in the sense that
9.1. Definitions and Local Properties 323

the ratio of the two quantities converges to 1). For a small λ, consider the
comparison triangle △λ for △aλ1 aλ2 aλ3 in the k-plane and the dilated triangle
λ△λ . The latter one has sides that converge to those of △a1 a2 a3 and lies
in the λ−1 k-plane (whose curvature goes to zero as λ → 0). Therefore the
curvature comparison conditions for △aλ1 aλ2 aλ3 and a suitable (close to aλ4 )
point of [aλ1 aλ3 ] turn in the limit (as λ → 0) to the same condition (but with
zero curvature bound!) for △a1 a2 a3 and a4 . This proves the theorem with
our special assumptions.
To get a proof in the general case, use the “generalized” Definition 9.1.2.
The same argument works except that a4 should be an ε-midpoint instead
of the real midpoint (which may not exist), and each point ai should be
replaced by a sufficiently close point represented by a shortest path. ¤

Combining the above theorem with Theorem 4.7.1 about cones over
length spaces yields the following
Corollary 9.1.45. Under the conditions of the theorem, the space of direc-
tions Σp at every point p ∈ X is a space of curvature ≤ 1.

Now we restrict ourselves to geodesically complete spaces (see Definition

9.1.29). We will see now that such spaces enjoy additional nice properties.
Proposition 9.1.46. Let X be a geodesically complete locally compact
length space of curvature bounded above. Then its space of directions at
any point is compact.

Proof. Let p ∈ X and Br (p) be a normal neighborhood of p. Consider

a metric sphere Sr (p), 0 < r < r′ . The shortest paths connecting p with
points of Sr (p) fill in Br (p) and each geodesic starting at p can be extended
to Sr (p) and is a shortest path within Br (p).
The map associating to each point x ∈ Sr (p) the direction at p of the
unique shortest path [px] is continuous and Sr (p) is compact. So the image
of Sr (p) (which is just Σ′p ) is compact as well. ¤
Exercise 9.1.47. Prove that none of the assumptions of the theorem can
be omitted.
Theorem 9.1.48. Let X be a geodesically complete locally compact space
of curvature bounded above. Then at every point the Gromov–Hausdorff
tangent cone exists and is equal to the tangent cone Kp .

Proof. Let us assume for simplicity that X is a space of nonpositive curva-

ture. This allows us to write “nonexpanding map” instead of a cumbersome
map with a Lipschitz constant 1 + δ.
324 9. Spaces of Curvature Bounded Above

Take two positive numbers, R and ε. By Proposition 7.4.12 we have to

find a finite ε-net in BR (O) ⊂ Kp which is a uniform (or Lipschitz, here it
is the same) limit of ε-nets in balls BRλ (p) ⊂ (λX, p) as λ → ∞. Choose

a closed ball B r (p) in a normal neighborhood of p. By the Hopf–Rinow

Theorem 2.5.28 this ball is compact, so there is a finite δ-net in B r (p) for
each δ > 0. Denote by N0 a finite Rr ε-net in B r (p). Now construct ε-nets
N and Nλ in BR λ (p) and B (0), resp. by setting N = R log (N ), where R
R r p 0 r
means homothety in Kp .
For each t, 0 < t ≤ 1, consider a map ht sending every point x ∈ B r (p)
to the point ht (x) of the shortest path [xp] such that |ht (x)p| = t|xp|. The
mapping ht : B r (p) → B tr is distance nonincreasing and surjective (check
R
this). Now put Nλ = λ ◦ ht (N0 ), where λ◦ means re-scaling and t = λr . For
fixed R, λ consider the map
1 λ
Φλ = λ ◦ logp ◦ : BR (p) → BR (0),
λ
where λ : (X, p) 7→ (λX, p) and λ1 : (λX, p) 7→ (X, p) are re-scaling maps.
This map does not increase distances and sends Nλ to N . So the only thing
we still need is to estimate decrease of distances between the points of Nλ
under the map Φλ . The set N0 is finite; thus for every ν > 0, there is a Λ
such that for λ > Λ and a, b ∈ Nλ the angle ∡apb of a comparison triangle
for △apb is not bigger than ∡apb + ν (since homothety does not change
angles).
e
For a, b ∈ Nλ denote ã = Φλ (a), b̃ = Φλ (b). A comparison triangle △pab
and the triangle △Oãb̃ have equal lateral sides, |p̄ā| = |Oã|, |p̄b̄| = |Ob̃|, and
∡āp̄b̄ ≥ ∡ãOb̃ − ν, where we can choose ν → 0 as λ → ∞. This proves that
|ab| − |ãb̃| → 0 uniformly for all couples (a, b) ∈ Nλ as λ → ∞. ¤

9.2. Hadamard Spaces

For a general (possibly positive) curvature bound k, not much is known
about global properties of spaces of curvature ≤ k. Most known results
about such spaces are local. However the assumption that k ≤ 0 implies
deep conclusions about global structure of the space. In particular, if a
space is complete and simply connected, it inherits many nice properties of
Euclidean spaces and, if k < 0, of hyperbolic spaces.
Definition 9.2.1. A complete simply connected space of nonpositive cur-
vature is called a Hadamard space.

An important source of examples of Hadamard spaces is the construction

of a universal covering (see Subsection 3.4.2 for general discussion of cover-
ings). Let X be a complete space of nonpositive curvature (not necessarily
9.2. Hadamard Spaces 325

simply connected). Recall (see Remark 9.1.18) that such a space is locally
simply connected and therefore has a universal covering; i.e., there exist
a simply connected topological space X̃ (the so-called universal covering
space) and a covering map f : X̃ → X.
The metric of X is (canonically) lifted to X̃ so that the covering map
becomes a local isometry. Since being nonnegatively curved is a local
property and f is a local isometry, X̃ is nonpositively curved as long as
X is. And a covering space of a complete space is complete (cf. Exercise
3.4.8). Thus X̃ is a Hadamard space.
One can apply general properties of Hadamard spaces to the universal
covering space X̃ and then derive information about the original space X.
This section and Section 9.3 contain a number of statements whose proofs
work this way.

9.2.1. Cartan–Hadamard Theorem. We already had examples showing

that curvature conditions in spaces of curvature bounded above may fail for
large triangles or hinges. Looking at these examples, one observes that
they also contain geodesic “bi-angles”, i.e., pairs of geodesics (even shortest
paths) connecting the same pairs of points. Later we will show that, indeed,
if there are no such bi-angles (i.e., if every two points are connected by a
unique geodesic), then all curvature conditions hold “in the large” (Theorem
9.2.9).
How does one check that geodesics are unique? In general, this is a hard
question, but for nonpositively curved spaces the answer is easy and is given
by the following fundamental “generalized Cartan–Hadamard Theorem”:

Theorem 9.2.2. Every two points in a Hadamard space are connected by a

unique geodesic. Furthermore, every geodesic segment in a Hadamard space
is a shortest path.

This theorem was proved by E. Cartan for Riemannian manifolds, by

M. Gromov for locally compact Hadamard spaces, and by D. Bishop and
S. Alexander in the general case (see [AB]).
Here we prove it only for locally compact Hadamard spaces. Note that
in this case a shortest path connecting any two points definitely exists. So
in this case it is sufficient to prove that, for every two points, a geodesic
connecting them is unique. For a proof for the general case, see [AB] and
[BH].
Before we begin the proof, it is useful to make some preliminary con-
sideration of geodesics in a space of nonpositive curvature. We assume that
geodesics are parameterized with constant speed.
326 9. Spaces of Curvature Bounded Above

Lemma 9.2.3. Let α and β be two constant-speed geodesics in a nonpos-

itively curved space parameterized by the same interval and contained in a
normal ball. Then the distance d(α(t), β(t)) is a convex function of t.

Proof. Denote d(α(t), β(t)) by δ(t). Since this function is continuous, it

suffices to prove that
t1 + t2 δ(t1 ) + δ(t2 )
δ( )≤ .
2 2
We use the following obvious property of nonpositively curved spaces:
For a triangle △abc in such a space and for midpoints b′ , c′ in its sides
′ ′
ab, ac one has |b′ c′ | ≤ |b c′ |, where b , c′ are midpoints of the corresponding
sides of a comparison triangle abc.
Now let t = (t1 + t2 )/2. Recall that every geodesic segment in a normal
ball is a shortest path. Let p be the midpoint between α(t1 ) and β(t2 ). Then
δ(t) = |α(t)β(t)| ≤ |α(t)p| + |pβ(t)|
1³ ´ 1¡ ¢
≤ |α(t1 )β(t1 )| + |α(t2 )β(t2 )| = δ(t1 ) + δ(t2 ) ,
2 2
and the lemma follows. ¤

The next lemma says that, given a geodesic segment γ connecting points
p and q, for every point q ′ near q there exists a unique geodesic connecting
p and q ′ and passing near γ. Note that there may be other geodesics
connecting p and q ′ but passing “far away” from γ; moreover the constructed
geodesic may be not a shortest path even if γ is. A reader familiar with
Riemannian geometry will notice that this lemma is an counterpart of the
fact that geodesics in a nonpositively curved Riemannian manifold have no
conjugate points.
Lemma 9.2.4. Let X be a complete locally compact space of nonpositive
curvature, γ : [0, 1] → X be a constant-speed geodesic with endpoints p =
γ(0), q = γ(1). Let r > 0 be so small that the convexity radius of γ([0, 1])
(cf. subsection 9.1.3 for the definition) is greater than 10r.
Then for every q ′ ∈ Br (q) there exists a unique constant-speed geodesic
α : [0, 1] → X such that α(0) = p, α(1) = q ′ , and the uniform distance
between γ and α is less than r, i.e., |γ(t)α(t)| < r for all t.
Remark 9.2.5. It is easy to see that the geodesic α depends continuously
on q ′ (prove this as an exercise).

The idea of the proof is simple. Consider a class of “broken lines” (that
is, curves composed of several shortest paths) with a fixed number of vertices
connecting p and q ′ , having short edges and passing in a neighborhood of γ.
9.2. Hadamard Spaces 327

We will show that a broken line of minimal length in this class is the desired
geodesic.

Proof. 1. Divide γ into equal-length intervals by points p0 = p, p1 =

γ(1/N ), p2 = γ(2/N ), . . . , pN = q, where N is so large that the length
of each interval is less than r. Consider the class M of broken lines
a0 a1 a2 · · · aN such that a0 = p, aN = q ′ and |ai pi | ≤ r for all i. By a broken
geodesic a0 a1 a2 · · · aN we mean a concatenation of shortest paths [ai ai+1 ]
each constant speed parameterized by the respective interval [i/N, (i +
1)/N ]. Since the space is locally compact, the set of admissible sequences
(a0 , a1 , . . . , aN ) is compact. Therefore there P is a broken geodesic α =
a0 a1 . . . aN minimizing the length (i.e., the sum i |ai ai+1 |) in the class M.
2. We will show that α is a geodesic; moreover every pair [ai−1 ai ] and
ai ai+1 of adjacent segments form a shortest path. Suppose the contrary, i.e.,
that |ai−1 ai | + |ai ai+1 | > |ai−1 ai+1 | for some i. Let a′i be the midpoint of
a shortest path [ai−1 ai+1 ]. Replacing ai by a′i we obtain a broken geodesic
which is shorter than α. Thus, in order to obtain a contradiction, it suffices
to show that the new broken geodesic belongs to M, i.e., that |a′i pi | ≤ r.
Apply Lemma 9.2.3 to the interval [pi−1 pi+1 ] of α and the shortest path
[ai−1 ai+1 ]. Since pi and a′i are the midpoints of these intervals, it follows that
|a′i pi | ≤ 21 (|ai−1 pi−1 | + |ai+1 pi+1 |) ≤ 12 (r + r) = r. Hence the new (shorter)
broken geodesic belongs to M. This contradiction shows that α is a geodesic.
Moreover, since the distance |α(t)γ(t)| is a convex function on every interval
[i/N, (i + 1)/N ] (by Lemma 9.2.3), we have |α(t)γ(t)| ≤ maxi |ai pi | ≤ r.
3. However α is not parameterized with a constant speed (unless all
distances |ai ai+1 | are equal). We have to show that α reparameterized with
a constant speed remains r-close to γ (with respect to the uniform distance).
The argument from the previous step shows that moving a vertex ai to the
midpoint between ai−1 and ai+1 does not increase the distance between our
broken geodesic and γ. It is easy to see that by doing such moves one can
obtain a sequence of reparameterizations of α converging to a constant-speed
one.
See also a more formal argument in the remark after the proof.
4. It remains to prove the uniqueness of α. Suppose there is another
geodesic β : [0, 1] → X connecting p and q ′ in the r-neighborhood of γ.
Then the (uniform) distance between α and β is less than 2r; hence for all
t ∈ [0, 1] the points α(t) and β(t) belong to some normal ball. By Lemma
9.2.3 it follows that the function t 7→ |α(t)β(t)| is convex and hence has no
local maxima in (0, 1). On the other hand, this function takes zero values
at 0 and 1, so it equals zero everywhere in [0, 1]. This means that α and β
coincide. ¤
328 9. Spaces of Curvature Bounded Above

Remark 9.2.6. Here is another argument replacing steps 2 and 3 in the

above proof. It is simpler but looks “less natural”.
P
Let us minimize over M the sum i |ai ai+1 |2 instead of the bare length.
Then the same midpoint construction (replacing ai by a′i ) shows that a
minimizing broken geodesic a0 a1 . . . aN is a constant-speed geodesic. Indeed,
µ ¶
2 2 |ai−1 ai | + |ai ai+1 | 2
|ai−1 ai | + |ai ai+1 | ≥ 2
2
µ ¶
|ai−1 ai+1 | 2
≥2 = |ai−1 a′i |2 + |a′i ai+1 |2 .
2
The equality here is achieved only if |ai−1 ai | = |ai ai+1 | = 12 |ai−1 ai+1 |, i.e., if
ai is a midpoint between ai−1 and ai+1 . For a minimizing broken geodesic
a0 a1 . . . aN this holds for all i; hence it is a constant-speed geodesic.

Lemma 9.2.4 can be interpreted as follows. Consider the space Xp of all

constant-speed geodesics (parameterized by [0, 1]) emanating from p. The
topology in this space is determined by the uniform distance. Lemma 9.2.4
tells us that for a sufficiently small r the points of the ball Br (p) are in
1-1 correspondence with the r-neighborhood of γ in the space of geodesics.
The correspondence is given by the “exponential” map eg xpp which sends
every geodesic to its endpoint, i.e., eg
xpp (α) = α(1). This map is obviously
continuous. Taking into account that X is locally compact, we see that egxpp
is a local homeomorphism.
Remark 9.2.7. The space Xp may differ from the tangent cone KP and
so eg
xpp may differ from the exponential map expp defined in 9.1.8. They
definitely are different if two geodesics emanating from p have the same
direction at p and none of them is a part of another.

Now turn to the proof of Theorem 9.2.2.

Proof of Theorem 9.2.2. Let X be a Hadamard space. Fix a p ∈ X

and consider the space Xp of geodesics emanating from p and the map
xpp : Xp → X sending every geodesic to its end-point (see above). By
eg
Lemma 9.2.4, eg xpp is a local homeomorphism. Lifting locally the metric
of X to Xp , we equip Xp with a length metric such that eg xpp is a local
isometry. We will show that eg xpp is a covering map. The theorem follows
easily from this. Indeed, since X is simply connected, the covering map
xpp is a homeomorphism; hence for every point q ∈ X the inverse image
eg
xp−1
eg p (q) consists of exactly one geodesic. In other words, there is a unique
geodesic connecting p and q.
To prove that eg
xpp is a covering map, we use the criterion provided by
Theorem 3.4.18: a local isometry from a complete length space to a “good
9.2. Hadamard Spaces 329

enough” length space (more precisely, one with locally unique shortest paths)
is a covering map. Since X is nonnegatively curved, it is “good enough” in
this sense, so it remains to show that Xp (with the length metric lifted from
X) is complete.
By Hopf–Rinow Theorem 2.5.28, it is sufficient to prove that, for some
P ∈ Xp , any constant-speed geodesics Γ : [0, 1) → Xp emanating from P
can be extended to the closed interval [0, 1].
Let P be the constant geodesic (speed 0) resting at p. Then every
geodesic γ : [0, 1) → X with γ(0) = p has a unique lift Γ : [0, 1) → Xp with
Γ(0) = P . (Being a lift means that γ = eg xpp ◦ Γ.) This lift Γ is defined in
a natural way: for every t ∈ [0, 1), the point Γ(t) in the space of geodesics
is nothing but the (reparameterized) restriction γ|[0,t] . The uniqueness of a
lift follows from the fact that eg
xpp is a local homeomorphism.
It is easy to extend such a lift Γ to [0, 1]. Extend γ to [0, 1] (this is
possible since X is complete) and take the lift of the extended curve. In
other words, define Γ(1) = γ|[0,1] . Since every geodesic in Xp is a lift of
some geodesic in X (namely, of its own image under eg xpp ), we have proved
the desired sufficient condition for Xp to be complete. The theorem then
follows as explained above. ¤

A geodesic loop is a geodesic γ : [a, b] → X such that γ(a) = γ(b).

Closed (or periodic) geodesics are particular case of geodesic loops.

Corollary 9.2.8. Let X be a complete locally compact space of nonpositive

curvature and p ∈ X. Then every element of the fundamental group π1 (X, p)
contains exactly one geodesic loop.

Proof. Consider a universal covering map f : X̃ → X and lift the metric of

X to X̃. Then X̃ is a Hadamard space. Fix a point x ∈ f −1 (p). For every
closed curve γ in X with endpoints at p there is a unique lift γ̃ in X̃ starting
at X; furthermore γ̃ is a geodesic if and only if γ is. Two curves represent the
same element of the fundamental group if and only if the endpoints of their
lifts coincide. So the set of curves representing a given element of π1 (X, p)
corresponds to the set of curves in X̃ connecting x and some fixed point
y ∈ f −1 (p). The latter contains exactly one geodesic by Cartan–Hadamard
Theorem; hence π1 (X, p) contains exactly one geodesic loop. ¤

9.2.2. Globalization.

Theorem 9.2.9. Let X be a Hadamard space of curvature ≤ k. Then

the curvature conditions hold for all triangles in X; i.e., X is a space of
curvature ≤ k “in the large”.
 330 9. Spaces of Curvature Bounded Above

Remark 9.2.10. In fact, the condition that X is a Hadamard space can be

replaced by the conclusion of the Hopf–Rinow Theorem: every two points
are connected by a unique shortest path and the shortest paths depend
continuously on the endpoints.

Proof of the theorem. We will prove that the angle condition holds for
all triangles. (This implies that all other curvature conditions hold globally,
because the “global” versions of the definitions of bounded curvature are
equivalent just like the local ones.)
1. The key observation is the following: if △abc is a triangle in X, d is
a point in its side [ac], and the angle condition holds for the triangles △abd
and △cbd, then it holds for △abc.
Indeed, place comparison triangles △abd and △cbd for △abd and △cbd
in the k-plane in different half-planes with respect to their common side bd.
From the angle condition for △abd and △cbd we have ∡bda ≤ ∡bda and
∡bdc ≤ ∡bdc; hence ∡bda + ∡bdc ≥ ∡bda + ∡bdc ≥ π. Therefore the angle
at d of the quadrilateral abcd is not less than π. Using Alexandrov’s Lemma
4.3.3 we obtain that ∡bad ≤ ∡bad ≤ b1 a1 c1 , where △b1 a1 c1 is a comparison
triangle for △bac. Similarly ∡bcd ≤ ∡b1 c1 a1 . Finally,
∡abc ≤ ∡abd + ∡cbd ≤ ∡abd + ∡cbd = ∡abc ≤ ∡a1 b1 c1
(the last inequality follows from the fact that |ac| ≤ |ad| + |cd| = |a1 c1 |).
2. We say that a triangle △abc is slim if its edges [ab] and [ac] are very
close to each other; more precisely, the (uniform) distance between those
sides is less than r/10 where r is the convexity radius of [ab] ∪ [ac] (so every
r-ball centered in a point of [ab] ∪ [ac] is a normal region).
c
PSfrag replacements
a
ba
c b

Figure 9.2: Cutting a slim triangle.

Let us “cut” such a slim triangle △abc into small triangles as shown in
Figure 9.2. Each small triangle is contained in a normal ball and therefore
satisfies the angle condition. Then the angle condition for △abc follows by
induction on the number of small triangles from the fact proved in the first
step.
3. Thus the angle condition is satisfied for all slim triangles. Now
consider an arbitrary triangle △abc. Since shortest paths in X are unique
9.2. Hadamard Spaces 331

and depend continuously on their endpoints (recall that we suppose M to be

locally compact), we have a continuous family of shortest paths connecting
a to the points of the side [b, c]. Then one can partition [bc] into small
intervals [bi bi+1 ], i = 0, 1, . . . , N − 1, b0 = b, bN = c, so that each triangle
△abi bi+1 is slim. These slim triangles satisfy the angle condition. Then the
result of the first step implies by induction on N that △abc satisfies the
angle condition. ¤
Remark 9.2.11. Conversely, if a complete space has nonpositive curvature
“in the large”, then it is simply connected and hence is a Hadamard space.
Indeed, suppose that such a space X is not simply connected. Fix a
point a ∈ X and find a shortest nontrivial geodesic loop γ with endpoint
at a. Then divide γ by points b and c into three intervals with equal lengths.
These intervals are shortest path; hence we have a triangle △abc whose sides
form a single geodesic. This triangle violates the angle condition.

9.2.3. Convex functions.

Definition 9.2.12. Let X be a length space. A function f : X → R is said
to be convex if its restriction to any constant-speed geodesic γ (that is, the
function t 7→ f (γ(t)) ) is convex.

If the function f is continuous and the metric of X is strictly intrinsic,

convexity of f is equivalent to the following: for any points x, y, z ∈ X such
that z is a midpoint between x and y, one has f (z) ≤ 21 (f (x)+f (y)). Indeed,
this midpoint criterion is equivalent to requiring that f is convex along any
shortest path, and the latter implies convexity along any geodesic (because
being a convex function is a local property).
In Hadamard spaces, many natural functions are convex. First, the
globalization theorem allows us to “globalize” Lemma 9.2.3:
Proposition 9.2.13. Let X be a Hadamard space. Then for any two
constant-speed geodesics α and β the function δ(t) = d(α(t), β(t)) is convex.
Corollary 9.2.14. If X is a Hadamard space, then for every p ∈ X the
function x 7→ |px| is convex.

Proof. Let β be a constant geodesic staying at p, and apply Proposition

9.2.13. ¤
Corollary 9.2.15. Let X be a Hadamard space and Y ⊂ X be a convex set.
Then the distance from Y (that is, the function x 7→ dist(x, Y )) is convex.

Proof. We assume that X is locally compact. Let x, y ∈ X, z be a midpoint

between x and y, x′ , y ′ be points nearest to x and y in the closure of Y ,
and z ′ be a midpoint between x′ and y ′ . Since Y is convex, z ′ belongs to
332 9. Spaces of Curvature Bounded Above

the closure of Y . Then dist(z, Y ) ≤ |zz ′ | ≤ 12 (|xx′ | + |yy ′ |). Here the second
inequality follows from Proposition 9.2.13. ¤
Corollary 9.2.16. Let X be a Hadamard space and i : X → X an isometry
of X. Then the “displacement function” δ(x) = d(x, i(x)) is convex.

Proof. The i-image of a geodesc γ is a geodesic i ◦ γ. And Proposition

9.2.13 says that the restriction of δ to γ, i.e., the function d(γ(t), i ◦ γ(t)) is
convex. ¤

The notion of λ-convex functions was introduced in Section 4.4,

Example 4.4.4. Recall that λ-convex functions are in a sense “no less convex”
than a quadratic function t 7→ λt2 . Roughly speaking, these are functions
whose second derivatives (in some generalized sense) are bounded below
by 2λ. More precisely:
Definition 9.2.17. Let X be a length space and λ > 0. A function
f : X → R is λ-convex if for any unit-speed geodesic γ in X the function
t 7→ f (γ(t)) − λt2 is convex.
A function is said to be strongly convex if it is λ-convex for some λ > 0.

Here is a reformulation of the definition in terms of midpoints.

Proposition 9.2.18. A continuous function f : X → R is λ-convex if and
only if, for any x, y ∈ X and z a midpoint between x and y,
f (x) + f (y) λ
f (z) ≤ − |xy|2 .
2 4
Proof. The proof is a straightforward computation. Let γ be a unit-speed
geodesic, and denote g(t) = f (γ(t)) − λt2 . Convexity of g is equivalent to
requiring that g(t1 ) + g(t2 ) ≥ 2g( 12 (t1 + t2 )) for all t1 , t2 . If t = 21 (t1 + t2 ),
x = γ(t1 ), y = γ(t2 ), z = γ(t), this inequality can be rewritten as
f (x) − λt21 + f (y) − λt22 ≥ 2f (z) − 2λt2 ,
or, equivalently,
λ
f (x) + f (y) − 2f (z) ≥ λ(t21 + t22 − 2( t1 +t2 2 t1 −t2 2
2 ) ) = 2λ( 2 ) = |xy|2
2
which is just the inequality given in the formulation. ¤

The next theorem characterizes Hadamard spaces in terms of convexity

of distance functions. In fact, this theorem is nothing but a (globalized) re-
formulation of the definition of nonpositive curvature (compare with Section
4.4).
9.2. Hadamard Spaces 333

We use the following notation: for a metric space X and a point p ∈ X,

dp denotes the distance function of p, that is, the function defined by
dp (x) = |px|.

Theorem 9.2.19. For every point p in a Hadamard space, the function d2p
is 1-convex.
Conversely, if X is a complete space with strictly intrinsic metric and
d2p is 1-convex for every p ∈ X, then X is a Hadamard space.

Proof. In the Euclidean plane, the function d2p has the form t 7→ t2 + const
along a unit-speed straight line. Since in a Hadamard space the distance
function is “more convex” than in the plane (this is the distance condition,
our first definition of nonpositive curvature), d2p is “more convex” than t2 .
Here is the same argument filled in with formulas. By Proposition 9.2.18,
it is sufficient to verify that

|px|2 + |py|2 |xy|2

|pz|2 ≤ −
2 4
if z is a midpoint between x and y. In the Euclidean plane this inequality
turns to equality. (This is the formula for the length of a triangle’s median
in terms of sides.) By the distance condition, the median [pz] of △pxy is not
longer than the respective median of a comparison triangle, and the desired
inequality follows.
Conversely, if d2p is 1-convex for every point p, the above inequality for the
median holds for all triangles. This means that X has nonpositive curvature
“in the large”. This implies that X is a Hadamard space, see Remark 9.2.11
after the proof of Globalization theorem. ¤

One of the most useful properties of strongly convex functions is that

they always attain minima.

Proposition 9.2.20. Let X be a complete space with a strictly intrinsic

metric, and f : X → R a continuous strongly convex function bounded from
below. Then f has a unique minimum point.

Proof. The uniqueness part is easy. If f (x) = f (y) = min f and x 6= y, let
z be a midpoint between x and y, and apply Proposition 9.2.18. This yields
f (z) < 21 (f (x) + f (y)) = min f , contradiction.
Now let us prove that a point of minimum exists. Let m = inf f and
{xi }∞
i=1 be a minimizing sequence for f , i.e., f (xi ) → m as i → ∞. We are
going to show that {xi } is a Cauchy sequence.
334 9. Spaces of Curvature Bounded Above

Fix an ε > 0 and let i0 be so large that f (xi ) < m + ε for all i > i0 .
Then for all i, j > i0 and z ∈ X we have
f (xi ) + f (xj )
f (z) ≥ m > − ε.
2
Taking a midpoint between xi and xj for z and subtracting the inequality
from
p Proposition 9.2.18, we obtain that λ|xi xj |2 /4 < ε. Thus |xi xj | <
4ε/λ for all i, j > i0 . Since λ is fixed and ε is arbitrarily small, it follows
that {xi } is a Cauchy sequence.
Now define x = lim xi . Then f (x) = lim f (xi ) = m, so x is the desired
minimum point. ¤

Many geometric constructions in Hadamard spaces are based on mini-

mum points of convex functions. Below are some examples.
Definition 9.2.21. Let {pi }ni=1 be a collection of points in a metric space
and {mi }ni=1 a collection of positive numbers. A barycenter (or a center of
mass) of points {pi } with masses {mi } is the minimum point of the function
P
mi d2pi where dpi (x) = |pi x|.

In a Hadamard space, the functions d2pi are strongly convex (Theorem

P
9.2.19), and so is the function mi d2pi . Therefore a barycenter exists and
is unique.
Exercise 9.2.22. Prove that the barycenter depends continuously on the
points {pi } and masses {mi }, and is a Lipschitz function of {pi } if the masses
mi are fixed.
Definition 9.2.23. Let X be a metric space and Y ⊂ X a bounded set.
A (closed) ball of minimal radius among the balls containing Y is called a
circumscribed ball of Y , its center a circumcenter of Y , and its radius the
circumradius.
Proposition 9.2.24. In a Hadamard space, every bounded set has a unique
circumscribed ball.

Proof. A circumcenter of Y is a minimum point of a function RY defined

by RY (x) = sup{|xy| : y ∈ Y }. Let us minimize the function RY2 instead.
This function is the supremum of functions d2y over y ∈ Y . A supremum of
1-convex functions is obviously 1-convex. So RY2 is 1-convex and therefore
a minimum point of RY2 exists and unique. ¤
Exercise 9.2.25. Consider the space M(X) of all bounded subsets of X
equipped with Hausdorff distance. Prove that, if X is a Hadamard space,
then the circumcenter is continuous as a function on M(X).
9.2. Hadamard Spaces 335

9.2.4. Parallel rays and lines. Recall that a line in a length space X is
a (unit-speed) geodesic γ : R → X such that any closed subinterval of γ
is a shortest path. A ray is a geodesic γ : R+ → X possessing the same
property. In Hadamard spaces, every geodesic segment is a shortest path,
so lines and rays are just complete geodesics of infinite length.
As in Euclidean and hyperbolic spaces, one can introduce a notion of
parallel rays and lines in a Hadamard space.
Definition 9.2.26. Two rays or two lines are said to be parallel if the
uniform distance between them is finite. Parallel rays are called asymptotic
as well.

It is obvious that being parallel is an equivalence relation.

Exercise 9.2.27. Prove that two rays or lines are parallel if and only if the
Hausdorff distance between them is finite.

In the Euclidean plane two rays or lines are parallel if and only if they
are parallel in ordinary (“scholar”) sense. It is easy to see that if two lines
in the Lobachevsky plane are parallel, then they coincide (but parallel rays
starting at points a, b in the Lobachevsky plane do exist for every pair of
points a, b). We will see soon that in the general case the situation is very
similar.
Proposition 9.2.28. Let X be a Hadamard space and p ∈ X. Then for
every ray γ in X there exists a unique ray starting at p and parallel to γ.

Proof. We prove this fact only for locally compact spaces. Let γt denote the
shortest path connecting p to γ(t). Since the distance between geodesics is a
convex function (Proposition 9.2.13), this shortest path is contained in the
(closed) r-neighborhood of γ where r = |pγ(0)|. Since the space is locally
compact, one can choose a sequence {ti }, ti → ∞, such that {γti } converges
to some ray α : R+ → X. This ray stays within the r-neighborhood of γ
and hence is parallel to γ.
To prove the uniqueness, suppose that there are two rays α and β starting
at p and parallel to γ. Since being parallel is an equivalence relation, the
rays α and β are parallel, so the function t 7→ |α(t)β(t)| is bounded. On the
other hand, this function is convex and equals zero at t = 0; therefore it is
zero for all t. This means that α = β. ¤

With the notion of parallel rays, one defines the ideal boundary of a
Hadamard space similarly to the ideal boundary of the Lobachevsky plane.
Namely, the points of the ideal boundary are equivalence classes of parallel
rays. The ideal boundary of a Hadamard space X is denoted by X(∞). It is
common to say that a ray “connects” its initial point with the point of the
336 9. Spaces of Curvature Bounded Above

ideal boundary that it represents. In this language, the above proposition

tells us that every point of X and every point of X(∞) are connected by a
unique ray.
The ideal boundary and various structures on it provide a useful lan-
guage for describing asymptotic properties of a Hadamard space. In this
book we do not get further into this important subject ( see [BGS], [E],
[BH] for some further material and references).
Now we turn to parallel lines. It turns out that parallel lines do not exist
unless a space contains some flat subsets; in particular, there are no parallel
lines in Hadamard spaces of strictly negative curvature.
Theorem 9.2.29. Let X be a Hadamard space, and γ1 and γ2 be two parallel
lines. Then either γ1 and γ2 coincide, or they span a convex flat strip.
Therefore if X has curvature ≤ k for a k < 0, then there are no parallel
lines in X except coinciding ones.

By a flat strip spanned by γ1 and γ2 we mean a convex subset isometric

to a strip between two parallel lines in R2 via an isometry which maps the
boundary lines to γ1 and γ2 .
For the proof we need the following simple lemma.
Lemma 9.2.30. Let Q = abcd be a quadrangle in Hadamard space and the
sum of the angles of Q no less than 2π. Then this sum of angles equals 2π,
and Q is a boundary of a flat convex region isometric to a planar quadrangle.

Proof. Draw the “diagonal” [ac] in the quadrangle Q. The angle condition
for nonpositive curvature implies that the sums of angles of △abc and △adc
are not greater than π. On the other hand, the triangle inequality for angles
implies that the sum of all six angles in △abc and △adc is not less than the
sum of angles of Q and hence is no less than 2π. Therefore all mentioned
inequalities turn to equalities; in particular, the sum of angles of Q equals 2π,
and in both triangles △abc and △acd the sum of angles equals π. Therefore
(by Proposition 9.1.19) each of these triangles bounds a totally geodesic flat
surface (a “solid” triangle). Denote these surfaces by T1 and T2 , respectively.
We are going to show that T1 ∪ T2 is a desired quadrangle.
By a similar argument, the triangles △abd and △cbd bound flat solid
triangles T3 and T4 . To finish the proof, it is sufficient to show that
T1 ∪ T2 = T3 ∪ T4 , or, equivalently, that the shortest path [ac] lies in the
surface T3 ∪ T4 . Observe that both surfaces T1 ∪ T2 and T3 ∪ T4 are arcwise-
isometric to a quadrangle abcd whose sides and angles equal the respective
sides and angles of Q. Then the length of [ac] equals the diagonal ac, and
this diagonal corresponds to a curve of the same length in T1 ∪ T2 . Thus
there is a curve in T3 ∪ T4 whose length equals |ac|. This means that this
9.2. Hadamard Spaces 337

curve is a shortest path and, since shortest paths in a Hadamard space are
unique, coincides with [ac]. ¤

Proof of the theorem. Let α and β be two parallel lines. By Proposition

9.2.13, the function t 7→ |α(t)β(t)| is convex on R. Since this function is
bounded, it is a constant. Note that this remains true if one shifts the
parameterization of α or β by a change of variable t 7→ t + const. Let us
choose a parameterizations so that the value of |α(t)β(t)| is the minimum
possible. To do this, let β(0) be a point nearest to α(0) in the line β. Then
for every t ∈ R we have

|α(t)β(t)| = |α(0)β(0)| = min |α(0)β(c)|

c∈R
= min |α(t)β(t + c)| = min |α(t − c)β(t)|,
c∈R c∈R

so the point β(t) is nearest to α(t) in β, and the point α(t) is nearest to β(t)
in α. Then by the first variation formula (Theorem 4.5.6) all four angles
that a shortest path [α(t)β(t)] forms with half-lines of α and β are no less
than π/2. Hence the quadrangle α(0)β(0)β(t)α(t) satisfies the assumption
of Lemma 9.2.30, so its angles equal π/2 and it bounds a flat surface. The
union of such rectangles over all t ∈ R is the desired flat strip. ¤

Now we can prove a splitting theorem for Hadamard spaces.

Theorem 9.2.31. Let X be a Hadamard space, γ a line in X, and Y ⊂ X

be a union of lines parallel to γ. Then Y is isometric to a product Z × R
for some metric space Z.
In particular, if every point of X belongs to a line parallel to γ, then X
is isometric to a product Z × R where Z is a Hadamard space.

Proof. If α and β are two parallel lines, x ∈ α, and y ∈ β is a nearest point

to x in β, we call y a projection of x to β.
Let us choose a parameterization of every line α parallel to γ so that
α(0) is a projection of γ(0) to α. We first prove that, for every two lines α
and β parallel to γ, the point β(0) is the projection of α(0) to β. Suppose
this is not the case. Let p = α(0), and let q = β(t0 ) be a projection of p to β.
We may assume that t0 > 0; otherwise change the orientation of the lines.
By Theorem p 9.2.29 the lines α and β bound a flat strip of width |pq|; hence
|β(t)p| = |β(t)q|2 + |pq|2 for all t ∈ R. It follows that |β(t)p| − |β(t)q| → 0
as t → ±∞ (because |β(t)q| → ∞ and |pq| is fixed). Then

|β(t)p| − t = |β(t)p| − |β(t)β(0)| = |β(t)p| − |β(t)q| + t0 −−−−→ t0 .

t→+∞
338 9. Spaces of Curvature Bounded Above

Therefore |β(t)p| > t + t0 /2 for all large enough t. We are going arrive
at a contradiction by showing that there is a path from β(t) to p (passing
through a point of γ) whose length is less than t + t0 /2.
By Theorem 9.2.29 the lines γ and α bound some flat strip S1 , and
lines γ and β bound some flat strip S2 . Consider the union S = S1 ∪ S2
of these strips. It is arcwise-isometric to a planar strip whose width is the
sum of widths of S1 and S2 (more precisely, S is the image of this planar
strip under an arcwise isometry). Furthermore, the segment between points
corresponding to p = α(0) and β(0) in the planar strip is orthogonal to its
boundary lines. Reasoning as above, we conclude that
t − |β(t)p|s = |β(t)β(0)|s − |β(t)p|s −−−−→ 0
t→+∞

where | · · |s is the intrinsic metric of S. Therefore |β(t)p|s < t0 /2 for all

large enough t. This contradicts the inequality |β(t)p| > t0 /2 that we had
above, because the intrinsic metric of S is not less than the metric of X.
This contradiction proves that β(0) is the projection of α(0) to β; in other
words, the relation of being a projection is transitive. Now it is easy to finish
the proof. Let Z be the set of all projections of γ(0) to the lines of which Y
is composed. There is a natural map from Z × R to Y which maps a pair
(z, t) to the point γz (t) where γz is the line parallel to γ passing through z
(recall that z = γz (0) by our choice of parameterizations). Theorem 9.2.29,
along with the fact that points of Z are projections of one another to their
respective lines, implies that this map is an isometry. ¤

9.3. Fundamental Group

of a Nonpositively Curved Space
In this section we prove some classical results about fundamental groups of
nonpositively curved spaces. If X is a compact space of strictly negative
curvature (that is, of curvature ≤ k for a k < 0), the results will actually
strengthen the fact that these groups are hyperbolic. For instance, let us
mention one corollary in advance: for X as above every nontrivial abelian
subgroup of π1 (X) is isomorphic to Z.

Let us recall the basic facts from Subsection 3.4.2 about fundamental
groups and universal coverings. Let X be a length space and f : X̃ → X
be the universal covering map. Then the fundamental group π1 (X) acts on
X̃ by isometries; more precisely, π1 (X) is isomorphic to the group of deck
transformations, that is, of maps from X̃ to itself commuting with f .
In other words, the isometry group Iso(X̃) contains a subgroup iso-
morphic to π1 (X); moreover this subgroup acts freely (i.e., every nontrivial
9.3. Fundamental Group of a Nonpositively Curved Space 339

element of this subgroup is a map without fixed points) and totally discon-
tinuously.
If X is a complete nonpositively curved space, then X̃ is a Hadamard
space. Thus, instead of fundamental groups of such spaces X, one can study
groups acting by isometries on a Hadamard space. In this case the action is
assumed to be free and totally discontinuous.

Theorem 9.3.1. Let X be a Hadamard space. Then every finite subgroup

Γ of Iso(X) has a fixed point; i.e., there exists an q ∈ X such that γ(q) = q
for all γ ∈ Γ.

Proof. For every point p ∈ X its orbit Γp = {γ(p) : γ ∈ Γ} is finite. By

Proposition 9.2.24, there is a unique circumcenter q of the orbit. For every
γ ∈ Γ the map x → γ(x) sends any orbit of Γ onto itself. Therefore γ(q) = q
for every γ ∈ Γ. ¤

Since deck transformations have no fixed points, the theorem yields the
following

Corollary 9.3.2. Let X be a complete space of nonpositive curvature. Then

the fundamental group π1 (X) does not contain nontrivial finite subgroups.
In other words, every element of π1 (X) except the identity generates an
infinite subgroup.

Theorem 9.3.3 (Preissmann). Let X be a compact space of curvature ≤ k

where k < 0. Then every two commuting elements of π1 (X, p) belong to a
cyclic subgroup.

Proof. Suppose that a0 , b0 ∈ π1 (X, p) and a0 b0 = b0 a0 If one of elements

a0 , b0 is trivial or a0 = b0 , then there is nothing to prove. So suppose that
this is not the case. Consider the free homotopy class of the element a0
considered as maps S 1 → X. There is a shortest representative in this
class. (The existence of a shortest loop is proved similarly to the existence
of shortest paths in a compact space; see Subsection 2.5.2.) Such a path α
is a closed (periodic) geodesic (check this). Pick a point q in α and consider
α as a loop with the vertex q. There is an isomorphism of π1 (X, p) onto
π1 (X, q) sending a0 to [α] = a. Denote by b the image of b0 under this
isomorphism. Obviously ab = ba.
Let f : X̃ → X be a universal covering map. A lift of α is the complete
geodesic α̃. Let us identify π1 (X, q) with the deck transformations group Γ.
The element a of Γ sends the geodesic α̃ into itself and acts in α̃ as a group
Z of shifts. (This element cannot change orientation of α̃; otherwise there
would be a fixed point in α̃.) The isometry b sends α̃ to a geodesic β̃.
340 9. Spaces of Curvature Bounded Above

First of all we want to show that β̃ = α̃. Suppose it is not the case.
Take a point x of the geodesic α̃, and consider the quadrangle Q with the
vertices x, x1 = ax, x2 = bx and x3 = abx = bax.
Since a, b are isometries, we have equalities

∡x1 xx2 + ∡xx2 x3 = π, ∡xx1 x3 + ∡x1 x3 x2 = π.

Now it follows from Lemma 9.2.30 that Q is isometric to a planar quadrangle.

This contradicts the assumption that X has strictly negative curvature.
Therefore β̃ = α̃.
Now consider restrictions of the actions of a, b to α̃. These restrictions
may be identified with numbers ā, b̄ ∈ R1 . The group generated by these
numbers, that is, the lattice {mā + nb̄}m,n∈Z , cannot have accumulation
points, i.e., is discrete.
It is a well-known fact of the number theory that in this case there is
a number c̄ such that c̄ = pā + q b̄ and ā = rc̄, b̄ = sc̄ where p, q, r, s are
integers, rp + sq = 1. (To prove this, observe that the set of positive sums
pa + qb assumes its minimum which is just the desired c.)
Let p, q be as above, and consider the element c = ap bq ∈ Γ. We show
that a = cr , b = cs . Indeed, ac−r keeps α̃ fixed and so ac−r is the identity
transformation. The proof of the equality b = cs is the same. ¤

Combining this theorem with the previous corollary, we obtain

Corollary 9.3.4. Let X be a compact space of curvature ≤ k where k < 0.

Then every nontrivial abelian subgroup of π1 (X) is isomorphic to Z.

Generalizations. The above theorem fails if one drops the assumption

about strictly negative curvature. Tori and direct products T n × M , where
M is a Hadamard space, are good counter-examples. In fact, every such
counter-example contains a subspace isometric to a flat torus. To get an idea
why this happens, recall the considerations of parallel lines in Subsection
9.2.4, in particular Theorem 9.2.29. Flat tori in a nonpositively curved
space correspond to abelian subgroups in the fundamental group. More
precisely, the following theorem holds.

Theorem 9.3.5. Let X be a compact space of nonpositive curvature. If

π1 (X) contains an abelian subgroup G of rank k > 1, then X contains a
convex subset isometric to a k-dimensional flat torus T .
Moreover, this torus T can be chosen so that the inclusion map T ֒→ X
induces an injective homomorphism of the fundamental groups and G is the
image of this homomorphism.
9.4. Example: Semi-dispersing Billiards 341

This theorem is a particular case of more general results on the action

of properly discontinuous isometry groups on a Hadamard space (see [BH]
for details and further references).

9.3.1. Final remarks. This chapter is only an introduction to the upper

curvature bounded world. The reader can find more results in several
recent books and papers. First of all, there is [BH] where, besides many
other things, the reader can find a detailed exposition of the fundamental
group properties (compare with 9.3) and of the geometry of the boundary
at infinity—the areas which are hardly touched in our textbook (see also
[BGS], [Gro3]). Also polyhedral complexes of curvature bounded above
(spaces are built of simplexes in a space form of curvature k) are considered
there in detail. Since 2-dimensional polyhedral spaces are already a very
interesting case, we mention here also the paper [BB], where one can find
appropriate references and some generalization and unsolved problems.
Local structure and some asymptotic properties of Alexandrov spaces
of curvature bounded above were studied recently in [Kl]. This study is
based on the notion of geometric dimension dimG . It is defined inductively
as follows: dimG (X) = 0 if X is a discrete space and dimG (X) = 1 +
supp∈X dimG (Σp ) in other cases. B. Kleiner proved that geometric dimension
is equal to topological dimension dimT op in the following sense: dimG (X) =
sup dimT op (K) where supremum is taken over all compact sets K ⊂ X.
If dimG (X) < ∞, then X has nice properties; for instance, for every
ε > 0 there exists a (1 + ε)-bi-Lipschitz map of an open (nonempty) set
U ⊂ Rn → X where n = dimG (X).
For Riemannian manifolds, some aspects of Hadamard spaces are more
developed than in the general case (see [BGS], [E]).

9.4. Example: Semi-dispersing Billiards

Everybody has seen a billiard table with several balls moving and colliding
in it. Here we demonstrate how the theory of Alexandrov spaces can be
applied to a certain class of problems in the theory of semi-dispersing billiard
systems.
Apparently, one of the motivations to study semi-dispersing billiard
systems comes from gas models in statistical physics. For instance, the
hard ball model is a system of round balls moving freely and colliding
elastically in empty space or in a box. Physical considerations naturally lead
to several mathematical problems regarding the dynamics of such systems.
For instance, one of the central problems of this kind asks for upper bounds
on the number of collisions that could occur in a billiard system (in a given
time). A series of basic problems of this type in their “physical” version go
342 9. Spaces of Curvature Bounded Above

back to Boltzmann, while their rigorous mathematical study was initiated

by Ya. Sinai.
As a (leading) example, consider a system of N round balls moving freely
and colliding elastically in empty space R3 (or in a box). Every ball moves
along a straight line with constant speed until two balls collide, and then
the new velocities of the two balls are determined by the (conservation)
laws of classical mechanics. (Explicit formulas for the velocities after a
collision in terms of the masses, radii, and the initial velocities of the balls
are usually discussed even in high-school physics classes.) To simplify our
considerations, we consider only situations where at most two balls collide
simultaneously.
A position of a collection of N balls can be represented by a point in
R3N . Namely if ai ∈ R3 is the center of the i-th ball and (xi , yi , zi ) are its
Cartesian coordinates, the corresponding point in R3N is

(a1 , a2 , . . . , aN ) = (x1 , y1 , z1 , x2 , y2 , z2 , . . . , xN , yN , zN ).

Not every point in R3N represents a valid configuration of balls. We

have to exclude positions where some of the balls overlap. The ith and
jth ball intersect if |ai − aj | < ri + rj where ri and rj are the radii of
3n
S This inequality defines a cylinder Cij ⊂ R . The complement
the balls.
3N
R \ i6=j Cij is the configuration space of our system. Its points correspond
to valid positions of the system of balls.
It is natural to consider the configuration space R3N with a Euclidean
structure given by the kinetic energy of the system:

N
X
K((v1 , v2 , . . . , vN ), (v1 , v2 , . . . , vN )) = mi hvi , vi i ,
i=1

where h, i is the usual scalar product for each of the “three-dimensional”

coordinates, and mi is the mass of the ith ball. The evolution of the system
of balls traces a path in the configuration space. It is easy to verify that
the point representing the configuration of balls moves straight and at a
constant speed until it hits one of the cylinders Cij (this event corresponds
to a collision in the system of balls), and then it continues following the
standard law of billiard collision: the angle of reflection is equal to the angle
of incidence. In other words, when a trajectory hits a cylinder Cij , the
projection of the velocity vector onto (the plane tangent to) the cylinder
does not change, and the normal component of the velocity vector changes
its sign. Of course, both the projection and the normal component are
considered with respect to K.
9.4. Example: Semi-dispersing Billiards 343

Thus the dynamics of a system of several balls is (equivalent to) a

particular case of a semi-dispersing billiard system, which is defined as
follows.
Definition 9.4.1. Let M be a complete nonpositively curved Riemannian
manifold with nonzero injectivity radius, and {Bi }i∈I be a finite (or at least
locally-finite) collection of convex subsets of M with smooth boundaries
Wi = ∂Bi (which areS smooth convex hypersurfaces). The (closure of)
the complement M \ Bi is called a semi-dispersing billiard table. The
hypersurfaces Wi (as well as the convex sets Bi ) will be referred to as walls
of the table.
A billiard trajectory is aSunit-speed piecewise smooth curve (“brokenS
line”) in the closure of M \ Bi with break point at walls (i.e., in Wi ),
such that the smooth intervals are Riemannian geodesics, and at every break
point the left and right velocities have equal projections onto the tangent
plane to the wall. In other words, it is a trajectory of a point that moves
along a geodesic until it hits one of the walls Wi , and then it gets reflected
according to the standard law of ellastic collision. For simplicity, we exclude
the trajectories that ever experience a collision with more than one wall
simultaneously.

An informal idea that a semi-dispersing billiard system is somewhat sim-

ilar to a geodesic flow on a negatively curved manifold has been around for
quite a while. Perhaps it was first explicitly mentioned by V. Arnold. In
the early sixties V. Arnold “speculated” that “such systems can be consid-
ered as the limit case of geodesic flows on negatively curved manifolds (the
curvature being concentrated on the collisions hypersurface)”. Indeed, it
is nowadays well known that a large portion of the results in the smooth
theory of (semi-)hyperbolic systems can be generalized (with appropriate
modifications) to (semi-)dispersing billiards. In spite of this, the construc-
tion suggested by Arnold has never been used until recently. It also caused
several serious objections; in particular, A. Katok pointed out that such ap-
proximations by geodesic flows on manifolds necessarily produce geodesics
that “bend around” collision hypersurfaces and therefore have no analogs in
the billiard system.
To study the billiard flow for a fixed time and in a small neighborhood
of a fixed point, one can use doubling by taking two copies of M , removing
the interiors of the walls and then gluing the copies along the boundaries
of the walls. One can approximate the singular metric resulting from this
procedure by smooth metrics (analogously one substitutes hard collisions by
a very steep repelling potential). The geodesic flows of the resulting metrics
will naturally converge to the billiard flow on a fixed time interval in a
small neighborhood of each point. Even though this construction does not
344 9. Spaces of Curvature Bounded Above

seem very useful, it already can deliver certain information. For instance,
the Liouville theorem (invariance of the Liouville measure) for billiard flows
follows immediately from the Liouville theorem for geodesic flows.
Let us illustrate Arnold’s suggestion by a simple example of the billiard
table in the complement of a disc in a two-torus (or the Euclidean plane).
Starting with two copies of the torus with (open) discs removed, and gluing
them along the boundary circles of the discs, one obtains a Riemannian
manifold (a surface of genus 2) with a metric singularity along the gluing
circle. This manifold is flat everywhere except at this circle. One can think
of this circle as carrying singular negative curvature, and indeed this is a
nonpositively curved length space. Smoothing this metric by changing it in
an (arbitrarily small) collar around the circle of gluing, one can obtain a
smooth nonpositively curved Riemannian metric, which is flat everywhere
except in this collar. To every segment of a billiard trajectory, one can
(canonically) assign a geodesic in this metric. Collisions with the disc would
correspond to intersections with the circle of gluing, where the geodesic
leaves one copy of the torus and goes to the other one.
Exercise 9.4.2. Repeat the same construction in dimension three: consider
two copies of R3 with a unit ball removed, and glue them along the boundary
spheres. Show that this is a space of curvature bounded above by 1, and it
is not a nonpositively curved space!
Hint: In this space, the gluing locus (which is a sphere) is a totally
geodesic subspace.
Show that the presence of positive curvature in this example persists un-
der smoothenings that do not change the metric outside some neighborhood
of the gluing locus.

This exercise shows that there are serious problems with using this
construction in higher dimensions. Moreover, even in dimension two many
geodesics do not correspond to billiard trajectories. They can be described
as coming from “fake” trajectories hitting the disc at zero angle, following
an arc of its boundary circle (possibly even making several rounds around
it) and then leaving it along a tangent line. Dynamically, such geodesics
carry “the main portion of entropy” and they cannot be disregarded. On
the other hand, it is difficult to tell actual trajectories from the fake ones
when analyzing the geodesic flow on this surface.
Main construction. The purpose of this section is to give an informal
and elementary account of how Arnold’s idea can nonetheless be formalized
(and the difficulties mentioned above can be partially avoided) by using
Alexandrov spaces. Several important open problems in the area were
recently solved by this method.
9.4. Example: Semi-dispersing Billiards 345

Our discussion is concentrated around the idea of gluing several copies

of M together and then developing billiard trajectories into this new space.
This idea is very old and its simplest versions arise even in elementary high-
school mathematical puzzles. For instance, if the billiard table is a square,
one can consider a tiling of Euclidean plane by such squares, and billiard
trajectories turn into straight lines.
We are going to represent billiard trajectories by geodesics in a nonposi-
tively curved space. To demonstrate the power of this method, we will show
how metric geometry allows to solve the problem of estimating the number
of collisions. Let us mention that, unfortunately, a construction that would
allow us to represent all billiard trajectories as geodesics in one compact
space is unknown in dimensions higher than three. Attempts to do this lead
to a striking open question: Is it possible to glue finitely many copies of a
regular 4-simplex to obtain a (boundary-less) nonpositive pseudo-manifold?
We introduce a construction that represents trajectories from a certain
combinatorial class, where by a combinatorial class of (a segment of) a
billiard trajectory we mean a sequence of walls that it hits.
Fix such a sequence of walls K = {Wni , i = 1, 2, . . . , N }. Consider a
sequence {Mi , i = 0, 1, . . . , N } of isometric copies of M . For each i, glue Mi
and Mi+1 along Bni . Since each Bni is a convex set, the resulting space MK
has the same upper curvature bound as M due to Reshetnyak’s Theorem
9.1.21.
There is an obvious projection MK → M , and M can be isometrically
embedded into MK by identifying it with one of Mi ’s (regarded as subsets of
MK ). Thus every curve in M can be lifted to MK in many ways. A billiard
trajectory whose combinatorial class is K admits a canonical lifting to MK :
we lift its segment till the first collision to M0 ⊂ MK , the next segment
between collisions to M1 ⊂ MK and so on. It is called the development of a
trajectory. It is easy to see that a development of a trajectory is a geodesic
in MK .
Note that, in addition to several copies of the billiard table, MK contains
other redundant parts formed by identified copies of Bi ’s. For example, if
we study a billiard in a curved triangle with concave walls, Bi ’s are not the
boundary curves. Instead, we choose as Bi ’s some convex ovals bounded
by extensions of these walls. (One may think of a billiard in the compact
component of the complement to three discs.) In this case, these additional
parts look like “fins” attached to our space. In case of the billiard in the
complement of a disc in a two-torus (see discussion above), the difference
is that we do not remove the disc when we glue together two copies of the
torus. Now a geodesic cannot follow an arc of the disc boundary, as the
latter can be shortened by pushing inside the disc. Still, there are “fake”
346 9. Spaces of Curvature Bounded Above

geodesics, which go through the disc. However, there are fewer of them than
before and it is easier to separate them.
It might seem more natural to glue along the boundaries of Wni rather
than along the whole Bni . For instance, one would do so thinking of this
gluing as “reflecting in a mirror” or by analogy with the usual development
of a polygonal billiard. However, gluing along the boundaries will not give
us a nonpositively curved space in any dimension higher than 2 (see Exercise
9.4.2).
One may wonder how the interiors of Bi ’s may play any role here, as they
are “behind the walls” and billiard trajectories never get there. For instance,
instead of convex walls in a manifold without boundary, one could begin with
a manifold with several “convex” boundary components (with a nonnegative
definite second fundamental forms with respect to the inner normal). Even
for one boundary component, there are examples where it is impossible to
“fill in” the boundary by a nonpositively curved manifold. Moreover, our
main dynamical result does fail for an example of this sort. Thus, it is
indeed important that the walls are not only locally convex surfaces, and we
essentially use the fact that they are filled by convex bodies.

Estimating the number of collisions. Now we are turning to the main

application of our construction, estimating the number of collisions in a hard
ball gas model. For the hard ball system, one asks whether the number of
collisions that may occur in this system can be estimated from above by
a bound depending only on the number of balls and their masses. If we
consider the balls moving in unbounded Euclidean space, we count the total
number of collisions in infinite time. For a system of balls in a box, we mean
the number of collisions in unit time (for a fixed value of kinetic energy).
We will consider a general case of a semi-dispersing billiard system. In this
case, it is clear that an additional assumption is needed. Indeed, already
for a two-dimensional billiard table bounded by several concave walls, a
trajectory may experience an arbitrarily large number of collisions (in time
one) in a neighborhood of a vertex if two boundary curves are tangent to
each other. Thus, we want to assume that (in a certain sense) our billiard
table is nondegenerate.
For simplicity, let us introduce the following nondegeneracy assumption,
which rules out various degenerations of the arrangements of hyperplanes
tangent to walls (it can be essentially weakened for noncompact billiard
tables):

• there exists a number C such that, for all sufficiently small ε, if a

point is ε-close to all sets from some sub-collection I of the Bi ’s,
9.4. Example: Semi-dispersing Billiards 347

then it is Cε-close to the intersection of the Bi ’s from this sub-

collection.
Exercise 9.4.3. Verify this condition for a system of several balls in R3 .
Hint: This condition means that for every configuration such that, for
any pair of indices from a certain set of pairs I, the corresponding balls are
close, there is a nearby configuration for which the balls from each of the
pairs intersect. This can be shown by using the following procedure: Choose
a ball and move all the balls simultaneously and with equal velocities along
the segments connecting their centers with the center of the chosen ball.
For a system of balls in a jar with concave walls the nondegeneracy
condition is satisfied except for some special sets of radii, when it is possible
to “squeeze the balls tightly between the walls.” Actually, it is known that
in those situations the system may have arbitrarily many collisions locally.
Now the main local result reads as follows:
Theorem 9.4.4. If a semi-dispersing billiard table satisfies the nondegener-
acy assumption, then there exists a finite number P such that every point p
in the billiard table possesses a neighborhood U (p) such that every trajectory
segment contained in U (p) experiences no more than P collisions.
Passing to estimating the global number of collisions (for infinite time)
we want to stay away from situations such as a particle infinitely bouncing
between two disjoint walls:
Theorem 9.4.5. If a semi-dispersing billiard table satisfies the
T nondegen-
eracy assumption, M is simply connected and the intersection Bi of Bi ’s
is nonempty, then there exists a finite number P such that every trajectory
experiences no more than P collisions.
Of course, for a hard ball gas model, the cylinders Cij have nonempty
intersection: for instance, the origin (0, 0, . . . , 0) ∈ R3N belongs to all
Cij (note that this point obviously does not correspond to any physically
realizable configuration of balls). One can check that the maximal number
of collisions that may occur in a system of N hard elastic balls (of arbitrary
masses and radii) never exceeds
µ ¶2N 4
2 mmax
400N ,
mmin
where mmax and mmin are, correspondingly, the maximal and the minimal
masses in the system.
Exercise 9.4.6. A trivial question: How many collisions can occur in a
system of two balls of equal masses in R3 ? A tricky question: How many
collisions can occur in a system of three balls of equal masses in R3 ?
348 9. Spaces of Curvature Bounded Above

Answer: four—and it is not that easy even to give an example...

Sketch of Proof for Theorem 9.4.4 for two walls: To outline the
idea of the proofs of uniform estimates on the number of collisions we restrict
ourselves to the simplest nontrivial case where the result was unknown: two
walls W1 and W2 bounding two convex sets B1 and B2 . Thus we avoid
inessential combinatorial complications and cumbersome indices. The reader
can understand where the difficulty lies by thinking of a particle shot “almost
parallel to the intersection line of W1 and W2 ” and bouncing between them
experiencing “almost tangential collisions”. The problem is to show that
the number of such collisions can be estimated by a bound independent of
how closely the particle follows the intersection line.
Let us assume that M is simply connected; otherwise, one can pass to its
universal cover. Consider a billiard trajectory γ connecting two points x and
y and choose any point z ∈ B1 ∩ B2 . Denote by K = {W1 , W2 , W1 , W2 , . . . }
the combinatorial class of γ, and consider the development γ̃ of γ in MK .
This is a geodesic between two points x′ and y ′ . By the Cartan–Hadamard
Theorem 9.2.2 every geodesic in a complete simply connected nonpositively
curved space is the shortest path between its endpoints. Note that z
canonically lifts to MK since all copies of z in different copies of M got
identified. Denoting this lift by z ′ , we see that |zx| = |z ′ x′ | and |zy| = |z ′ y ′ |.
Thus we conclude that the lengths of γ between x and y is less that |xz|+|zy|
for all z ∈ B1 ∩ B2 . In other words, any path in M connecting x and y and
visiting the intersection B1 ∩ B2 is longer than the segment of γ between x
and y.
The following argument is the core of the proof. It shows that if a
trajectory made too many collisions then it can be modified into a shorter
curve with the same endpoints and passing through the intersection B1 ∩B2 .
This contradicts the previous assertion and thus gives a bound on the
number of collisions.
Assume that γ is contained in a neighborhood U (p) and it collided
with W1 at points a1 , a2 , . . . , aN alternating with collisions with W2 at
b1 , b2 , . . . , bN . Let zi be the point in B1 ∩ B2 closest to bi and hi be the
distance from bi to the shortest geodesic [ai ai+1 ]. By the nondegeneracy
assumption, |zi bi | ≤ C · dist(bi , B1 ) ≤ Chi . Thus the distance Hi from zi to
the shortest geodesic [ai ai+1 ] is at most (C + 1)hi .
Plugging this inequality between the heights of the triangles △ai bi ai+1
and △ai zi ai+1 into a routine argument which develops these triangles on
both the Euclidean plane and k-plane, one concludes that Di ≤ C1 · di ,
where di = |ai bi | + |bi ai+1 | − |ai ai+1 |, Di = |ai zi | + |zi ai+1 | − |ai ai+1 |. Here
k is the infimum of the sectional curvature in U (p), and a constant C1 can
be chosen depending on C alone provided that U (p) is sufficiently small.
9.4. Example: Semi-dispersing Billiards 349

Let dj be the smallest of di ’s. Let us modify the trajectory γ into a

curve with the same endpoints: substitute its pieces ai bi ai+1 by the shortest
segments [ai ai+1 ] for all i’s excluding i = j. This new curve is shorter than
γ by at least (N − 1)dj . Let us make a final modification by replacing the
piece aj bj aj+1 by aj zj aj+1 . It makes the path longer by Dj , which is at
most C1 dj . Hence, N ≤ C1 + 1 because otherwise we would have a curve
with the same endpoints as γ, passing through zj ∈ B1 ∩ B2 and shorter
than γ. This proves the local bound on the number of collisions. ¤
Let us pass to the proof of global estimates, where geometry works in
its full power. Again, to stay away from combinatorial complications, we
restrict ourselves to the case of two walls W1 and W2 .

Sketch of Proof for Theorem 9.4.5 for two walls. Consider a trajec-
tory γ making N collisions with the (only possible) sequence of walls
K = {1, 2, 1, . . . , 2, 1}. Reasoning by contradiction, assume that N > 3P +1,
where P is the local bound on the number of collisions (whose existence is
guaranteed by Theorem 9.4.4). Again consider the space MK , but now we
will “close it up” by gluing M0 ∈ MK and MN ∈ MK along the copies of
B1 . Denote the resulting space by M̃ . We cannot use Reshetnyak’s Theo-
rem 9.1.21 directly to conclude that M̃ is a nonpositively curved space any
more, since we identify points in the same space and we do not glue two
spaces along a convex set.
We recall that a space has nonpositive curvature if and only if every
point possesses a neighborhood such that, for every triangle contained in
the neighborhood, its angles are no bigger than the corresponding angles
of the comparison triangle in the Euclidean plane. However, using the
correspondence between geodesics and billiard trajectories, one can conclude
(reasoning exactly as in the proof of the local estimates on the number of
collisions) that each side of a small triangle cannot intersect interiors of
more than P copies of the billiard table. Since N > 3P + 1, for every
small triangle for which we want to verify the angle comparison property,
we can undo one of the gluings without tearing the sides of the triangle. This
ungluing may only increase triangle’s angles, and now we find ourselves in
a nonpositively curved space (which is actually just MK ), and thus we get
the desired comparison for the angles of the triangle.
To conclude the proof, it remains to notice that the development of γ
in M̃ is a geodesic connecting two points in the same copy of B1 . This
is a contradiction since every geodesic in a simply connected nonpositively
curved space is the only shortest path between its endpoints; on the other
hand, there is a shortest path between the same points going inside this
copy of B1 . ¤
 Chapter 10

Spaces of Curvature
Bounded Below

This is an introduction to the theory of length spaces with lower curvature

bounds. We consider only complete length spaces (in fact, local complete-
ness is sufficient in all local statements) of curvature ≥ k for some k ∈ R.
Throughout the chapter, the term “Alexandrov space” is reserved for (con-
nected) spaces from this class.
In the case of a positive curvature bound k it is convenient to exclude
some exceptional one-dimensional spaces. Namely, we do√ not count the line
R, the half-line R+ , segments
√ of length greater that π/ k, and circles of
length greater than 2π/ k as Alexandrov spaces of curvature ≥ k. As we
will see in Section 10.4, excluding these spaces is equivalent
√ to requiring
that the diameter of a space is not greater than π/ k. Thus the term
“Alexandrov space of curvature ≥ k” in this chapter means a (connected)
complete length space of curvature ≥ k which is not isometric to one of
the above listed exceptions if k > 0. A developed theory for such spaces
was created during 1980–90s. Here we present only some basic results from
the foundations of this theory, and a few “global” results that generalize
classical theorems of Riemannian geometry. More advanced exposition of
results and techniques can be found in [BGP], [Pl]; for other references and
applications see [GP].
There are two main points making the theory very different from the
case of curvature bounded above. First, Toponogov’s Globalization Theorem
(Theorem 10.3.1) says that in any Alexandrov space of curvature ≥ k (for
any k, and without any topological assumptions like simple connectedness),

351
352 10. Spaces of Curvature Bounded Below

the triangle comparison conditions hold “in the large”. This yields a number
of results about global geometry of Alexandrov spaces (recall from Chapter 9
that one can say much more about a Hadamard space than about a general
space of curvature bounded above). Furthermore, Toponogov’s Theorem
implies that the class of Alexandrov spaces of curvature ≥ k is closed with
respect to Gromov–Hausdorff convergence and this allows us to study this
class “as a whole” (recall the discussion in the beginning of Chapter 7).
Second, (finite-dimensional) Alexandrov spaces have nice local structure.
By dimension we mean Hausdorff dimension (cf. Subsection 1.7.4) but in
fact all known notions of dimension are equivalent for an Alexandrov space
(in particular, Hausdorff dimension equals topological dimension). Some
of the local properties of finite-dimensional Alexandrov spaces are worth
mentioning right now. Hausdorff dimension of such a space is always an
integer; such a space is a manifold (in fact, almost a Riemannian one)
everywhere except a tiny set of singular points; the space of directions of an
n-dimensional Alexandrov space is an (n − 1)-dimensional Alexandrov space
of curvature ≥ 1 and in fact is isometric to S n−1 at almost every point.
Though these formulations are nice and clear, known proofs are long
and technical. We do not get into these details until the end of the chapter
(Sections 10.8 and 10.9). Some knowledge of the local structure is, however,
required in earlier sections; in such cases we refer to the appropriate results
from the last sections.

10.1. One More Definition

In Chapter 4 we gave several equivalent definitions of a space of nonnegative
curvature. Replacing R2 by the k-plane, one obtains definitions of a space of
curvature ≥ k where k is an arbitrary real number (cf. Section 4.6). Again,
these definitions are equivalent and the proof of this equivalence is almost
the same as in the case k = 0.
Here we add one more equivalent formulation that has no counterpart
for spaces of curvature bounded above. Its main feature is logical simplicity;
it has the form “for every four points (in a small neighborhood) the distances
between them satisfy certain inequalities”. Recall that the definitions we had
so far involve collections of points satisfying additional restrictions (e.g., one
is a midpoint for two others), or even angles.
The new equivalent formulation is given in Proposition 10.1.1. For
convenience we introduce the following notation for comparison angles,
generalizing it from Chapter 3.
Notation. Let a, b and c be three different points in a length space. We
e
denote by ∡abc (comparison angle) the angle at b of a comparison triangle
10.1. One More Definition 353

abc in the k-plane (assuming k is fixed). In case of ambiguity we add k as a

e k abc.
subscript: ∡

Note that ∡e k abc is a (continuous) function of the three distances |ab|,

|ac| and |bc|. It is well-defined if |ab| + |ac| + |bc| < 2Rk where Rk is the
diameter of the k-plane.
Proposition 10.1.1. A locally compact length space X is a space of curva-
ture not less than k if and only if every point x ∈ X has a neighborhood U
such that for any collection of four different points a, b, c, d ∈ U the following
condition is satisfied:
(10.1) e k bac + ∡
∡ e k cad + ∡
e k dab ≤ 2π.

We call the inequality (10.1) the quadruple condition (a quadruple is a

collection of four different points) for a quadruple (a; b, c, d). Note that the
condition is symmetric in {b, c, d} but a is in a special position. Proposition
10.1.1 allows us to use the quadruple condition as yet another definition of
a space of curvature ≥ k. Note that this new definition does not rely on
shortest paths, so it can be used unmodified for not strictly intrinsic spaces.

Proof of Proposition 10.1.1. First suppose that (10.1) holds for every
quadruple. To verify the triangle condition (from Definition 4.1.9) for a
triangle △abc and d ∈ [ac], apply the quadruple condition to (d; a, b, c).
e
Since ∡adc e
= π, it follows that ∡bdc e
+ ∡bda ≤ π. Then Lemma 4.3.3
implies the desired inequality |db| ≥ |db|.
Now let X be a space of curvature ≥ k. Consider a quadruple (a; b, c, d)
and a point a′ in a shortest path [ab]. Then
e ′ d + ∡da
∡ba e ′ c + ∡ca
e ′ b ≤ ∡ba′ d + ∡da′ c + ∡ca′ b
≤ (∡ba′ d + ∡da′ a) + (∡aa′ c + ∡ca′ b) = 2π.
We used the angle condition (Definition 4.1.15), the triangle inequality for
angles, and the fact that the sum of adjacent angles equals π (Lemma 4.3.7).
Now let a′ converge to a; then (10.1) follows by continuity of comparison
angles. ¤

We leave as exercises several simple facts about spaces of curvature ≥ k.

Exercise 10.1.2. Prove that geodesics in a space of curvature ≥ k do not
branch. Namely, if two geodesics have a common interval, then they are
subintervals of one geodesic.
Exercise 10.1.3. Show that the quadruple condition implies the following
fact: for any three geodesics emanating from one point, the sum of three
angles between these geodesics is not greater than 2π.
354 10. Spaces of Curvature Bounded Below

Exercise 10.1.4. Prove that, if two geodesics in a space of curvature ≥ k

start at one point and form a zero angle at that point, then one of them is
a subinterval of the other.
Hint: Use the angle condition to obtain a local statement; then refer to
Exercise 10.1.2.
Exercise 10.1.5. Prove that, if shortest paths in a space of curvature ≥ k
have two points in common, then these points are their endpoints.

10.2. Constructions and Examples

There are many constructions of Alexandrov spaces of curvature bounded
below. A number of simple ones were presented in Chapter 4. Here we add
two new examples: quotients by isometry groups and convex surfaces. In
the second example, the proof of the fact that the space has nonnegative
curvature is not trivial and relies on results that we will prove later.

10.2.1. Products and cones.

Example 10.2.1 (products). Let X and Y be Alexandrov spaces of cur-
vature ≥ k, k ≤ 0. Then the direct product X × Y is a space of curvature
≥ k. The proof is straightforward.
Note that a product of spaces of curvature ≥ k is not a space of curvature
≥ k if k > 0 unless one of the multiplied spaces is a single point. See
comment just after 9.1.5.
Indeed, consider two arbitrary shortest paths in spaces X and Y . These
shortest paths (as subspaces of X and Y ) are isometric to intervals of R;
hence their product in X × X is a convex set isometric to a region in R2 .
Therefore X × Y cannot have strictly positive curvature.
Example 10.2.2 (cones). Euclidean and spherical cones were already con-
sidered in Subsections 3.6.2, 3.6.3. Here we give a uniform description for
these constructions and define the similar notion of a hyperbolic cone. In
the context of spaces with lower curvature bounds, it does not make sense
to consider cones over spaces of diameter greater than π.
Let k ∈ R and X be a metric space with diam(X) ≤ π. The k-cone over
X, denoted by Conk (X), consists of the origin √ (or apex) o and pairs (x, r)
where x ∈ X and r > 0 (in addition, r ≤ π/ k if k > 0). The distance from
(x, r) to the origin equals r, and then the distance between a1 = (x1 , r1 ) and
a2 = (x2 , r2 ) is defined so that
∡e k a1 oa2 = |x1 x2 |.
In other words, |a1 a2 | equals the side |a1 a2 | of a triangle a1 oa2 in the k-plane
such that |oai | = ri for i = 1, 2, and ∡a1 oa2 = |x1 x2 |.
10.2. Constructions and Examples 355

√
If k > 0, all pairs (x, π/ k) should be identified (because the distance
between them is zero). These pairs represent a point o′ ∈ Conk (X) √ which
could be taken as the origin instead of o (the map (x, r) 7→ (x, π/ k − r)
is an isometry of Conk (X)). In this case the k-cone is also called the k-
spherical cone or the k-suspension. The origins o and o′ are referred to
as its poles. The standard spherical suspension defined in Subsection 3.6.3
corresponds to k = 1.
If k < 0, k-cones are also called (−k)-hyperbolic cones. A hyperbolic
cone is a k-cone for k = −1. If k = 0, k-cones are ordinary (“Euclidean”)
cones.
By the very definition, Conk (S 1 ) is isometric to the k-plane. Similarly,
Conk (S n ) is the standard (n+1)-dimensional
√ space form of curvature k, i.e.,
an n-dimensional sphere of radius 1/ k, a rescaled hyperbolic space, or Rn ,
depending on the sign of k.
The primary use of cones in this chapter is to consider cones over spaces
of directions. Note that a space X can be recovered as the space of directions
of the cone Conk (X) at the origin.
The results of Section 4.7 about curvature bounds of cones can be
simplified for Alexandrov spaces in view of Toponogov’s Theorem 10.3.1
(saying that the triangle comparison conditions hold in the large) and
Corollary 10.4.2 (saying that the perimeter of every triangle in a space of
curvature ≥ 1 is no greater than 2π).
Theorem 10.2.3. Let X be a complete metric space and k ∈ R.
1. If X is an Alexandrov space of curvature ≥ 1, then Conk (X) is an
Alexandrov space of curvature ≥ k.
2. If Conk (X) is an Alexandrov space of curvature ≥ k and, in addition,
|xy| + |yz| + |xz| ≤ 2π for all x, y, z ∈ X, then X is an Alexandrov space
of curvature ≥ 1 or a space consisting of two points at distance π from each
other.
The proof of this theorem is in no way different from that of Theorem
4.7.1. Note that in the second part of the theorem we do not assume in
advance that X is a length space. Due to the other assumption, this property
follows from the fact that Conk (X) is a length space. Nor do we assume
that X is connected—unlike the case of curvature bounded above, a cone
over X cannot have a lower curvature bound if X is not connected (except
the trivial case of a two-point space).

10.2.2. Quotients. Recall (see Section 3.3) that if a group Γ acts on a

metric space (X, d) by isometries and the orbits o(p) = {γp : γ ∈ Γ} are
closed, then we equip the quotient Q = X/Γ with the strongest topology for
356 10. Spaces of Curvature Bounded Below

which the canonical projection π : X → Q, π(p) = o(p) is continuous. And

we equip Q with the metric ρ such that
(10.2) ρ(o(p), o(q)) = inf{d(p, r)|r ∈ o(q)}.

Proposition 10.2.4. If (X, d) is a length space of curvature ≥ k and a

group Γ acts on X by isometries with closed orbits, then Q = X/Γ also is a
space of curvature ≥ k.

Proof. To simplify the considerations, we assume that X is locally compact.

Let p0 ∈ Q and let p ∈ π −1 (p0 ) where π : X → Q is the projection map.
Let r > 0 be such that the ball U = Br (p) is a normal region in the sense
that the quadruple condition (cf. Proposition 10.1.1) is satisfied for every
for quadruple contained in U . We will show that the same is true in the
ball U0 = Br/2 (p0 ) in the quotient space. Let (a0 ; b0 , c0 , d0 ) be a quadruple
in U0 . Since X is locally compact and the orbit π −1 (a0 ) is a closed set, there
is a point a ∈ π −1 (a0 ) which is nearest to p, i.e., |pa| = dist(p, π −1 (a0 )). The
definition (10.2) of the quotient metric implies that |pa| = |p0 a0 |. Then one
can similarly find points b, c, d ∈ U such that |ab| = |a0 b0 |, |ac| = |a0 c0 |
and |ad| = |a0 d0 |. Note that a, b, c, d ∈ U . By (10.2), we have |bc| ≥ |b0 c0 |,
e
|cd| ≥ |c0 d0 | and |db| ≥ |d0 b0 |; hence ∡bac ≥ ∡be 0 a0 c0 , ∡cad
e e 0 a0 d0 and
≥ ∡c
e e
∡bad ≥ ∡b0 a0 d0 . Therefore
e 0 a0 c0 + ∡c
∡b e 0 a0 d0 + ∡b
e 0 a0 d0 ≤ ∡bac
e e
+ ∡cad e
0 + ∡bad ≤ 2π

and the proposition follows. ¤

This proposition is commonly used to construct examples of spaces with

curvature bounded below. One of the simplest examples of this kind is the
following. Let Z2 = {e, γ} act on R3 by symmetries: γ(x) = −x for all
x ∈ R3 . Then Q2 = R3 /Z2 is a space of nonnegative curvature.

Exercise 10.2.5. Check that Q2 is not a manifold. Prove that Q2 is

isometric to the cone K(P2 ) over the projective space P2 equipped with
the canonical metric of constant curvature 1.

Similarly, if a group G acts on the standard sphere S n ⊂ Rn+1 by

isometries, then Q = S n /G is a space of curvature ≥ 1. Note that Q is a
Riemannian manifold if G acts freely (i.e., without fixed points); otherwise Q
may have metric or even topological singularities. As an example, consider
the action of G = Z2 = {e, γ} on S n defined by
γ(x1 , . . . , xn , xn+1 ) = (−x1 , . . . , −xn , xn+1 ).
Here the quotient space is the spherical cone over Pn−1 .
10.2. Constructions and Examples 357

10.2.3. Convex surfaces in R3 . Metrics of curvature bounded below

arose initially in papers by A. D. Alexandrov as nonnegatively curved metrics
on two-dimensional surfaces. Alexandrov proved that the class of two-
dimensional nonnegatively curved length spaces essentially coincide with
the class of convex surfaces in R3 . In this section we prove one part of
this coincidence, namely, that every convex surface in R3 (with its intrinsic
metric) is a space of nonnegative curvature.
The other part is a remarkable theorem stating that every length space
of nonnegative curvature homeomorphic to the 2-sphere is isometric to the
boundary of a convex body in R3 or, as a degenerate case, to a twice
covered planar convex region, i.e., two copies of a region glued together
along the boundaries. From this theorem it easily follows that every point
of a two-dimensional manifold with a metric of nonnegative curvature has
a neighborhood isometric to a region on the boundary of a convex body in
R3 .
Similar results can be proved for metrics of curvature ≥ k and surfaces in
spaces of constant curvature k. The restriction that the space is a manifold
is not essential because one can prove that every two-dimensional space of
curvature bounded below is a manifold (possibly with a boundary).

Theorem 10.2.6. The intrinsic metric of a convex surface in R3 is a metric

of nonnegative curvature.

By a convex surface C we mean the boundary of a convex body X. It

makes sense to add “degenerate surfaces” to make the class closed. The
theorem is trivial for convex sets of dimensions 0 and 1. If dim X = 2, we
consider not the boundary of X but the twice covered region X, i.e. two
copies of X glued together along their boundaries (convex curves). It is not
difficult to prove the theorem in this case. The case of unbounded X is
reduced to the compact case due to the local nature of the theorem.

Proof of Theorem 10.2.6. Now we assume that X is a convex body, that

is, a compact convex set with nonempty interior. Let C = ∂X. The idea
of the proof is very simple: we approximate X by convex polyhedra and
prove that the intrinsic metrics of their boundaries converge to the intrinsic
metric of C. It is easy to verify that the intrinsic metric of a convex
polyhedral surface is nonnegatively curved. Later in Section 10.3 we will
show that a Gromov–Hausdorff limit of nonnegatively curved spaces is itself
nonnegatively curved (Proposition 10.7.1). Applying this fact finishes the
proof. Now we pass to formal details.

Lemma 10.2.7. If convex bodies Xi ⊂ R3 (i → ∞) converge in the

Hausdorff metric to a convex body X ⊂ R3 , then the intrinsic metrics of
358 10. Spaces of Curvature Bounded Below

Ci = ∂Xi converge uniformly (cf. Definition 7.1.5) to the intrinsic metric

of C = ∂X.

Proof. We denote the intrinsic metrics of C and Ci by d and di respectively.

We may assume that the origin 0 belongs to the interior of X and hence X
contains a ball Br (0) for some r > 0. Since X is convex, every ray emanating
from the origin intersects C = ∂X exactly once. Hence the central projection
from R3 \ {0} to C (which maps every ray to its point of intersection with
C) is well-defined. We will show that this central projection restricted to
Ci is a homeomorphism from Ci to C (for all sufficiently large i) and its
distortion goes to zero as i → ∞.
First consider arbitrary convex bodies X and X ′ in R3 such that Br (0) ⊂
X ⊂ X ′ . Let C = ∂X, C ′ = ∂X ′ , d and d′ denote the intrinsic metrics of C
and C ′ , and let δ be the maximum distance from a point in C ′ to its central
projection in C. We are going to estimate how much the central projection
from C ′ to C may enlarge the intrinsic distances. The key observation is
the following “Busemann–Feller Lemma”:
(i) For every point a′ ∈ R3 there is a unique nearest point in X, i.e., a
point a ∈ X such that |a − a′ | = dist(a′ , X).
(ii) If a, b ∈ X are nearest points in X for a′ , b′ ∈ R3 , then |a − b| ≤
|a′ − b′ |.
To prove (i), suppose that a1 and a2 are two nearest points for a′ in X.
Then the midpoint a = 12 (a1 + a2 ) belongs to X because X is convex. But
|a′ − a| < 21 (|a′ − a1 | + |a′ − a2 |); hence a1 and a2 are not nearest points.
To prove (ii), observe that the angles ∡abb′ and ∡baa′ are not less than
π/2. Indeed, if ∡abb′ < π/2, then the segment [ab] ⊂ X contains a point b1
with |b′ − b1 | < |b′ − b| (for example, pick b1 close to b). Since the angles
∡abb′ ≥ π/2 and ∡baa′ ≥ π/2, the projection of the segment [ab] to the line
a′ b′ contains the segment [a′ b′ ]. Hence [ab] is not longer than [a′ b′ ].
The Busemann–Feller Lemma can be reformulated as follows: the
nearest-point map from R3 to X is well-defined and nonexpanding. We
call this map the orthogonal projection to X. Since it is nonexpanding, it
does not increase lengths of curves. Therefore if a′ , b′ are two points in C ′
and a0 , b0 are their orthogonal projections to X, then d(a0 , b0 ) ≤ d′ (a′ , b′ ).
(Note that a0 and b0 belong to C because X ⊂ X ′ .)
Now let a and b be the central projections of a′ and b′ in C. Consider
the straight segment [aa′ ] and project it onto C by means of orthogonal
projection. The resulting curve γ connects a to a0 in C, and L(γ) ≤
L([aa′ ]) ≤ δ. Hence d(a, a0 ) ≤ δ, and similarly d(b, b0 ) ≤ δ. It follows
that d(a, b) ≤ d′ (a′ , b′ ) + 2δ.
10.2. Constructions and Examples 359

Now return to the original problem. We leave as an exercise to the reader

the following fact about convex bodies: the Hausdorff convergence of {Xi }
to X implies that there is a sequence {εi } of positive numbers such that
εi → 0 and (1 − εi )X ⊂ Xi ⊂ (1 + εi )X for all large enough i. It follows that
the central projection to Xi is well-defined. Applying the above estimate to
the central projection from Xi to (1 − εi )X, we obtain that
(1 − εi )d(a, b) ≤ di (ai , bi ) + 2εi D
where D = diam(X), ai and bi are central projections of a and b in Xi .
Applying the same estimate to the central projection from (1 + εi )X to Xi ,
we obtain that
d(ai , bi ) ≤ (1 + εi )d(a, b) + 2εi D.
These two inequalities imply that the distortion of the central projection is
not greater than εi (diam(C, d) + 2D).
Since diam(C, d) < ∞ (for example, diam(C, d) ≤ πD because the
orthogonal projection from the sphere of radius D to C is nonexpanding and
surjective), it follows that the distortions of central projections converge to
zero. Hence (Ci , di ) converge uniformly to (C, d). ¤

It is not difficult to approximate X by convex polyhedra Pi . For example,

let Pi be a convex hull of a finite εi -net in X where εi → 0. As we have
seen in Chapter 4 (Theorem 4.2.14), the boundary of a convex polyhedron
is a nonnegatively curved space. By the above lemma, C with its intrinsic
metric is a uniform limit of these polyhedral spaces. Now the theorem
follows from Proposition 10.7.1 that a lower curvature bound is preserved
under Gromov–Hausdorff convergence. ¤

Generalizations. Theorem 10.2.6 holds for convex hypersurfaces in Rn for

all n ≥ 3. The proof is almost the same. (However, the converse theorem
that every nonnegatively curved space is locally isometric to a convex surface
is no longer true in higher dimensions.)
Moreover, one can replace Rn by a spherical or hyperbolic space of
curvature k; in this case convex hypersurfaces have curvature ≥ k. A
more general (and not that easy) result that a convex hyper-surface in any
Riemannian manifold of curvature ≥ k is itself a space of curvature ≥ k, is
proved by S.Buyalo.1 It is still an open problem whether the boundary of
a convex set in an Alexandrov space of curvature ≥ k is itself a space of
curvature ≥ k.

1S. Buyalo, Shortest paths on convex hypersurfaces of a Riemannian manifold, in Studies in

Topology, Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 66 (1976) 114-132
(Russian)
360 10. Spaces of Curvature Bounded Below

10.3. Toponogov’s Theorem

Toponogov’s Theorem is the “globalization theorem” for spaces of curvature
bounded below. A space of curvature bounded below is defined in local
terms; namely, certain triangle comparison inequalities are required to hold
in a neighborhood of every point. Toponogov’s Theorem says that this local
definition implies that the same inequalities hold “in the large”, i.e., for all
triangles (cf. Definition 4.6.6 in section 4.6.2). Unlike the case of curvature
bounded above (Theorem 9.2.9), Toponogov’s Theorem works for curvature
bounds of any sign and, more importantly, does not require topological
conditions like simple connectedness.
In the two-dimensional case this theorem was proved by A. Alexandrov.
V. Toponogov proved it for Riemannian manifolds of any dimension. The
general case is due to G. Perelman.

Theorem 10.3.1 (Toponogov’s Theorem). Let X be a complete length space

of curvature ≥ k. Then X has curvature ≥ k in the large.

Remark 10.3.2. Recall that “in the large” means that the curvature
comparison conditions are satisfied for all triangles for which comparison
triangles exist and are unique. The latter requirement on triangles is
essential only if k > 0 and is approximately
√ equivalent to the statement
that the perimeter is not greater than 2π/ k. In Subsection 4.1.15 we will
√ of curvature ≥ k cannot contain a
show that, in fact, an Alexandrov space
triangle of perimeter greater than 2π/ k.

The proof of Theorem 10.3.1 is quite complicated and may be considered

optional. To simplify the exposition, we prove the theorem only for locally
compact spaces and leave to the reader most of the technical details arising
in the case k > 0. Another proof of the theorem can be found in [BGP].

Proof of Theorem 10.3.1. The global versions of the definitions are equiv-
alent just like the local ones. We will show that the “angle condition” 4.1.15
is satisfied for every triangle. Note that angles between shortest paths do
exist and the sum of adjacent angles is equal to π because these properties
are local.
We present a proof for the case k ≤ 0 and then explain how to modify
the argument for the case k > 0.
Step 1. Suppose that the theorem is false. Then there exists a triangle
△pqr satisfying the following:
e
(i) The angle condition fails for its angle at q, i.e., ∡pqr < ∡pqr.
10.3. Toponogov’s Theorem 361

(ii) The angle condition is satisfied for every triangle of perimeter no

greater than 0.99 R whose vertices are contained in the ball B100R (q), where
R = per(△pqr) = |pq| + |pr| + |qr|.
To prove the existence of such a triangle, suppose the contrary and
construct a sequence of triangles {△pi qi ri }∞i=1 as follows. Let p1 q1 r1 be
e
an arbitrary triangle such that ∡p1 q1 r1 < ∡p1 q1 r1 . Then for i = 1, 2, . . .
define Ri = per(△pi qi ri ) and let △pi+1 qi+1 ri+1 be a triangle of perimeter
no greater than 0.99 Ri , with vertices in B100Ri (qi ), and such that
e i+1 qi+1 ri+1 .
∡pi+1 qi+1 ri+1 < ∡p
¡ ¢i
Then Ri+1P ≤ 0.99 Ri ≤ 0.99 R1 . Since |qi qi+1 | ≤ 100Ri , it follows that
∞
the series i=1 |qi qi+1 | converges; hence {qi } is a Cauchy sequence. Let
q = limi→∞ qi and U be a normal neighborhood of q. Then △pi qi ri ⊂ U for
a large enough i; hence △pi qi ri must satisfy the angle condition. This is a
contradiction.
Step 2. Let △pqr satisfy the conditions (i) and (ii) from Step 1. Then
there exists a triangle △abc with vertices in BR (q), of perimeter not greater
than R, which is “thin” in the sense that |ac| < 0.01 R, and such that the
angle condition fails for ∡bac.
To prove this, divide the side [qr] of △pqr into small intervals by
points c0 = q, c1 , . . . , cn = r so that each triangle △ci pci+1 is “thin”
(|ci ci+1 | < 0.01 R). Note that the perimeters of the triangles △ci pci+1 are
not greater than per(△pqr) ≤ R. We will show that the angle condition fails
for ∡pci ci+1 or ∡pci+1 ci in one of these triangles. Suppose the contrary.
Then it follows by induction that the angle condition is satisfied for the
angles at q and ci in △qpci for i = 1, . . . , n. Indeed, assuming this for △qpci
(i < n), place comparison triangles △qpci and △ci pci+1 in different half-
planes with respect to their common side pci in the k-plane. Then Lemma
4.3.3 implies that
e
∡pqc e
i+1 ≤ ∡pqci = ∡pqci ≤ ∡pqci = ∡pqci+1
e i+1 q ≤ ∡pci+1 q. This completes the induction step. Now
and similarly ∡pc
let i = n; then the statement we just proved implies that the angle condition
is satisfied for ∡pqr, contrary to our assumptions.
Step 3. Let △abc be a “thin” triangle constructed in Step 2 and such
that the angle condition fails just at c, and let d be the midpoint of [bc].
Then |bd| = |cd| ≤ 14 R, |ad| ≤ 0.26 R and per(△adc) ≤ 0.52 R; hence the
angle condition holds for the angles of △adc. Therefore by Lemma 4.3.3
e
(compare with Step 2), it follows that ∡adb < ∡adb. Choose an ε > 0 such
that
(10.3) e
∡adb ≤ ∡adb − ε = ∡adb − ε
 362 10. Spaces of Curvature Bounded Below

where △adb is a comparison triangle.

Let us take a point x close to d and place comparison triangles for △adb,
△axb, △axd, and △dxb as shown in Figure 10.3. We want a point x to be
such that the comparison triangles △abx3 , △dbx2 and △adx1 have disjoint
interiors and ∡x1 dx2 ≥ ε. To obtain such an x we do the following: take
a1 and b1 close to d in the shortest paths [ad] and [bd], and let x be the
midpoint of a shortest path [a1 b1 ].
Observe that the angle condition is satisfied for triangles △aa1 x, △a1 dx,
△bb1 x, △b1 dx and △a1 db1 because their perimeters are smaller than 0.99 R.
Using this and the implied monotonicity of angles, we obtain
e 1 dx1 + ∡b
∡adx1 + ∡bdx2 ≤ ∡a e 1 dx2 ≤ ∡a
e 1 db1 ≤ ∡adb ≤ ∡adb − ε.

Hence the triangles △adx1 , and △bdx2 in Figure 10.3 have no common
interior points and
∡x1 dx2 = ∡adb − ∡adx1 − ∡bdx2 ≥ ε.
Since |ax3 | = |ax1 | and |bx3 | = |bx2 |, it then follows that △abx3 has no
common interior points with △adx1 , and △bdx2 .

d
PSfrag replacements
x1
x1 gaga x2
x2
x3 al be
a al be
b x3
d
α
β
γ γa b

Figure 10.1: Toponogov’s theorem, Step 3.

Step 4. Let α = ∡ax1 x3 = ∡ax3 x1 , β = ∡bx2 x3 = ∡bx3 x2 and

γ = ∡dx1 x2 = ∡dx2 x1 (see Figure 10.3). Now, applying the Gauss–Bonnet
formula to the triangle △x1 x2 x3 and taking into account that 2α ≤ π,
2β ≤ π, 2γ + ε ≤ 2γ + ∡x1 dx2 ≤ π (here we use our assumption that k ≤ 0),
we obtain that
2π = ω(△x1 x2 x3 ) + 2(α + β + γ) + (∡ax1 d + ∡dx2 b + ∡ax3 b) − 3π
≤ (3π − ε) + (∡axd + ∡dxb + ∡axb) + (−∡axb + ∡ax3 b) − 3π
10.3. Toponogov’s Theorem 363

where ω is the integral curvature, ω(E) = k · Area(E) ≤ 0. Since ∡axd +

∡dxb + ∡axb ≤ 2π, it follows that

e
∡axb − ∡axb = ∡ax3 b − ∡axb > ε.
ε
On the other hand, |ax| + |bx| ≤ |ad| + |db| − 10 |dx| if |dx| is small enough.
Step 5. Consider the set S of all triangles △ayb (with a and b fixed)
such that
e
(i) ∡ayb − ∡ayb ≥ ε, and
(ii) max{|ay|, |by|} ≤ 0.26 R.
We supposed that this set was nonempty. Since our space is locally
compact, there is a triangle △abd ∈ S with

|ad| + |bd| = min{|ay| + |yb| : △ayb ∈ S}.

However Step 4 shows that one can find a point x (close to d) such that
△axb ∈ S and |ax| + |xb| < |ad| + |db|. This contradiction proves the
theorem for the case k ≤ 0.
Step 6. If k > 0, the same scheme works with certain modifications.
First, one has to check carefully that the involved comparison triangles are
well-defined. Second, in the inequalities that follow from the Gauss–Bonnet
formula in Step 4, one cannot drop the integral curvature of the triangles
△x1 x2 x3 and dx1 x2 . As a result, one loses the condition (i) in Step 5 when
△adb is replaced by △axb. To work this around, replace the condition (i)
in Step 5 by a more complicated one, namely,

e 1
(10.4) (∡ayb − ∡ayb) − c(|ay| + |yb|) ≥ ε,
2
where c is a sufficiently small positive constant. Then the loss in the first
term of (10.4) is compensated by the gain in the second one. The details
are left to the reader as an exercise .
The reader who has checked all these details might have noticed that √
the argument works only for triangles of perimeter strictly less than 2π/ k.
√ out the case of a triangle △abc with perimeter
So we have to work √ |ab| +
|bc| + |ac| = 2π/ k, and prove that ∡abc = π if max{|ab|, |bc|} < π/ k. Fix
a positive ε < k. Observe that a space of curvature ≥ k is also a space of
curvature ≥ k − ε. Applying the already proven part of the theorem to X as
a space of curvature ≥ k − ε, we obtain that ∡abc ≥ ∡ e k−ε abc. As ε → 0, a
comparison triangle for △abc in the (k − ε)-plane converges to a comparison
e k−ε abc → ∡
triangle in the k-plane. Therefore ∡ e k abc = π as ε → 0, implying
that ∡abc ≥ π. ¤
364 10. Spaces of Curvature Bounded Below

10.4. Curvature and Diameter

This section can be viewed as an addendum to Toponogov’s Theorem in the
case of positive curvature bound k. The conclusion of the theorem reads:
if a triangle is “not too large” so that it has a well-defined (i.e., unique
up to an isometry) comparison triangle in the k-plane, then it possesses
the comparison properties.
√ “Not too large” here means that the perimeter
is not greater
√ than 2π/ k. (One√should also exclude triangles △abc with
|ac| = π/ k and |ab| + |bc| = π/ k for which a comparison triangle is not
unique. Or, alternatively, choose comparison triangles such that the side
corresponding to [ac] passes through the point corresponding to b.)
A question
√ arises: what about triangles whose perimeters are greater
than 2π/ k ? In this section we give an easy answer: such triangles simply
do not exist. Recall that we have excluded some spaces from the class of
Alexandrov spaces√ of curvature ≥ k. These exceptions are: √
circles of length
greater than 2π/ k, line segments of length greater than π/ k, the half-line
R+ , and the line R.
Theorem 10.4.1. Let X be√ an Alexandrov space of curvature ≥ k where
k > 0. Then diam(X) ≤ π/ k.

Proof. Suppose
√ the contrary, and let a, b ∈ X√ be two points such that
|ab| > π/ k. We may assume that |ab| = (π+ε)/ k where√0 < ε < π/4. Let
c be the midpoint of a shortest path [ab] and U be an (ε/3 k)-neighborhood
of c.
First we show that U contains a point which does not belong to [ab].
Indeed, suppose the contrary. For every point x ∈ X consider a shortest
path γ connecting c to x. Our assumption implies that this shortest path
coincides with a subinterval of [ab] in a neighborhood of c. Since geodesics do
not branch (Exercise 10.1.2), it follows that x belongs to the unique geodesic
containing [ac]. Therefore X is covered by two minimal geodesics (that
is, geodesics whose intervals are shortest paths) starting at c and passing
through a and b respectively. Depending on whether these geodesics are
finite or infinite and whether their endpoints coincide, we conclude that X
is one of the one-dimensional exceptional spaces.
Choose√an x ∈ U \ [ab], and
√ let y be a nearest point to x in [ab]. Then
|ay| > π/2 k and |by| > π/2 k. The first variation formula implies (see
Corollary 4.5.7) that ∡xya = ∡xyb = π/2. Consider a comparison√triangle
△xya for △xya in the k-plane which is the sphere of radius 1/ k. √By
Toponogov’s Theorem we have ∡xya ≤ ∡xya = π/2. Since |ya| > π/2 k,
it follows that |xa| < |ya|. (To see this, place a at the north pole of the
sphere; then the inequality ∡xya ≤ π/2 implies that x lies above the great
circle passing through y orthogonally to the meridian [ax]. Since y is strictly
10.4. Curvature and Diameter 365

below the equator, y is the lowest point of this great circle; hence y is strictly
lower than x, i.e., |xa| < |ya|.) Thus |xa| < |ya| and similarly |xb| < |yb|.
Hence |ya| + |yb| > |xa| + |xb| ≥ |ab|, contrary to the fact y belongs to a
shortest path [ab]. ¤

Corollary 10.4.2. Let X be an Alexandrov space of curvature ≥ k√ where

k > 0. Then every triangle in X has perimeter no greater than 2π/ k.

Proof. We first prove the theorem √ under an additional assumption that

diam(X) is strictly less than π/ k. Suppose √ the contrary and let x, y, z ∈ X
be such that |xy| + |yz| + |xz| > 2π/ k. Fix shortest paths [xy] and
[xz]. By continuity, there exist√ points y ′ ∈ [xy] and z ′ ∈ [xz] such
that |xy ′ | + |xz√ ′ | + |y ′ z ′ | = 2π/ k. Consider a triangle △xy ′ z ′ . Since
√
diam(X) < π/ k, its sides are shorter than π/ k; hence a comparison
triangle △xy ′ z ′ for △xy √
′ z ′ in the k-plane is well-defined. Since the perimeter

of △xy ′ z ′ equals 2π/ k, its angles equal π; therefore ∡xy ′ z ′ = ∡xz ′ y ′ = π

by Toponogov’s Theorem. Hence ∡zz ′ y ′ = ∡yy ′ z ′ = 0, and this implies that
the shortest path [y ′ z ′ ] passes through y and z. Then the perimeter of △xyz
is less or equal to that of √ △xy ′ z ′ . This contradiction proves the statement
in the case diam(X) < π/ k.
Now we pass to the general √ case. Fix
√ a positive ε < k. By Theorem
10.4.1 we have diam(X) ≤ π/ k < π/ k − ε. Furthermore, since X is a
space of curvature ≥ k, it is also a space of curvature ≥ k − ε. Therefore
we can apply the above arguments with k replaced by k −√ε and conclude
that the perimeter of every triangle is no greater than 2π/ k − ε. Since ε
is arbitrary, the desired statement follows. ¤

Exercise 10.4.3. Let X be a compact space of curvature ≥ 1 and diam(X) =

π. Prove that X is isometric to a suspension (i.e., spherical cone; cf. Sub-
section 10.2.1) over a compact space of curvature ≥ 1.
Hint: Let p, q ∈ X be such that |pq| = π. Define

Y = {x ∈ X : |px| = |qx| = π/2}.

Then prove the following facts. First, for every x, y ∈ Y , shortest paths [px]
and [qx] are unique and ∡pxy = ∡qxy = ∡pyx = ∡qyx = π/2. Then the
triangles △pxy and △qxy (as unions of shortest paths) are isometric to their
comparison triangles in the sphere. Then Y is convex in X and therefore Y
is a length space of curvature ≥ 1. Finally, every point x ∈ X belongs to
some shortest path connecting p and q, and therefore to the union [py] ∪ [qy]
for some y ∈ Y . These facts imply that X is the spherical cone over Y .
Another proof is based on the splitting theorem from the next section;
see Remark 10.5.8.
366 10. Spaces of Curvature Bounded Below

Exercise 10.4.4 (Berger’s theorem). Let X be an n-dimensional Riemann-

ian manifold of curvature ≥ 1 and diam(X) = π. Prove that X is isometric
to the standard sphere S n .
Exercise 10.4.5. The radius rad(X) of a compact metric space X is the
minimal number r > 0 such that X = B r (p) for some p ∈ X. Prove that
1
1. 2 diam(X) ≤ rad(X) ≤ diam(X) for every metric space X.
2. If X is an n-dimensional Alexandrov space of curvature ≥ 1 and
rad(X) = π, then X is isometric to S n .
Hint: Using the statement of Exercise 10.4.3, prove that X contains
subsets isometric to S 1 , S 2 , . . . .

10.5. Splitting Theorem

Recall that a geodesic γ : (−∞, ∞) → X is said to be a straight line (or,
briefly, a line) if every one of its segments is a shortest path between its
endpoints. Here is a remarkable splitting theorem.
Theorem 10.5.1 (Splitting theorem). If a locally compact Alexandrov space
X of nonnegative curvature contains a line, then X is isometric to a direct
metric product R × Y for some nonnegatively curved Alexandrov space Y .
This theorem was proved by V. Toponogov for Riemannian manifolds
and generalized by A. Milka for general spaces of nonnegative curvature.2
Remark 10.5.2. For Riemannian manifolds, a stronger theorem holds: one
can assume nonnegativity of Ricci curvatures instead of nonnegativity of
sectional curvatures (see [CG]).
In the course of the proof of the theorem, X always denotes an Alexan-
drov space of nonnegative curvature. We introduce the following definition
of parallel lines (used in this section only and having no relation to parallel
lines considered in Chapter 9).
Definition 10.5.3. Two straight lines γ, γ1 in X are said to be parallel
if there are two parallel straight lines γ, γ 1 in R2 and an isometry of the
metric space γ ∪ γ1 onto the metric space γ ∪ γ 1 .
Here we mean an isometry with respect to the restrictions of the metrics
of X and R2 to these sets.
We say that the lines γ and γ 1 in R2 correspond to the lines γ and γ1
in X. We fix an isometry from γ ∪ γ1 to γ ∪ γ 1 and refer to the image of a
point a ∈ γ ∪ γ1 under this isometry as the point corresponding to a.
2A. D. Milka, Metric structure of some class of spaces containing straight lines, Ukrain.
Geometrical. Sbornik, vyp. 4, 1967, Kharkov, pp. 43-48 (in Russian). In dim 2 this theorem was
proved by Cohn-Vossen much earlier.
10.5. Splitting Theorem 367

Lemma 10.5.4. Let γ be a line in X and p ∈ X. Then for every two points
x, y ∈ γ the angles at x and y of a triangle △pxy are equal to the respective
e
angles of its comparison triangle △pxy.

Proof. Suppose that ∡pxye ≤ ∡pxy − ε, ε > 0. Take two points zi ∈ γ

such that x is located between z1 and z2 , and y is between z1 and x. Place
e
comparison triangles △pxz e
1 and △pxz2 in different half-planes with respect
to their common side px. And place a comparison triangle △pxy in the
e
same half-plane where the triangle △pxz 1 is placed. The rectangles pz 1 xz 2
and pyxz 2 are convex. Due to the angle monotonicity definition we have
e
∡pxz e e
1 ≤ ∡pxy ≤ ∡pxy − ε and, therefore, the angle ∡z1 xz2 < π − ε. This
is impossible if z1 and z2 are sufficiently far from x because |z1 p| + |pz2 | ≥
|z1 x| + |xz2 |. ¤

By monotonicity of angles the lemma immediately implies the following

Corollary 10.5.5. 1. For points p′ ∈ [px] and y ′ ∈ [yx] and points p′ ∈ [px]
and y ′ ∈ [yx] such that |p′ x| = |p′ x|, |y ′ x| = |y ′ x| the equality |p′ y ′ | = |p′ y ′ |
holds.
2. For a line γ and a point p there is an unique nearest point to p in γ.

This nearest point is called the projection of p to γ. (Note however that

a shortest path [pp′ ] may not be unique.)
Lemma 10.5.6. 1. For every line γ and every point p, there is no more
than one line γ1 parallel to γ and passing through p.
2. The relation of being parallel is transitive.

Proof. Suppose that there are two lines, γ1 and γ2 , passing through p
and parallel to γ. Consider three corresponding parallel lines γ 1 , γ, and
γ 2 in the plane R2 . We assume that γ lies between the two other lines.
Let points a ∈ γ 1 , b ∈ γ 2 correspond to points a ∈ γ1 , b ∈ γ2 such that
|pa| = |pb| and let the segment [ab] intersect the line γ at point c. Then
|ab| ≤ |ac| + cb| = |ac| + |cb| = |ab|. Applying Lemma 10.5.4 to the point a
and the line γ2 , one sees that |ab| → ∞ as |pa| = |pb| → ∞. The latter is
impossible because |ab| is a constant.
The second assertion of the lemma is proved similarly. The only differ-
ence is that one should consider four parallel lines in the plane. ¤

Proof of the theorem. Step 1. Let γ be a straight line. First of all, for
every point p ∈ X \ γ, we construct a straight line γp passing through p and
parallel to γ. To do this, denote ci = γ(ti ) and consider shortest paths [pci ]
where ti → ∞ as i → ∞. There is a subsequence of these paths converging
to a geodesic ray γp+ : [0, ∞) → X. Similarly, taking points di = γ(−ti ),
368 10. Spaces of Curvature Bounded Below

ti → ∞, one can find a geodesic ray γp− : [0, −∞)) → X. We are going to
show that these two rays together form a straight line γp parallel to γ.
The comparison angles ∡c e i pdi obviously approach π as i → ∞. Since
e i pdi , this implies that angles ∡ci pdi converge to π as well. Hence
∡ci pdi ≥ ∡c
the angle between γp+ and γp− at p equals π.
Now let us prove that γp is a line. Fix a t > 0 and let b+ −
i and bi be
the points in the shortest paths [pci ] and [pdi ] at the distance t from p.
These points converge to the points b+ = γp (t), b− = γp (−t). Consider a
+ −
comparison triangle △ci adi and points bi , bi in its sides at distances t of
+ −
a. It is obvious that |b+ b− | = limi→∞ |b+ −
i bi | ≥ limi→∞ |bi bi | = 2t. This
means that each segment of γp is a shortest path; hence γp is a line.
Now we prove that γ and γp are parallel. Consider two lines γ, γp at the
distance |pp′ | in a plane R2 where p′ is the projection of p to γ. Let p ∈ γ p
and p′ be the projection of p to γ. Then the limit process used above and
Corollary 10.5.5 imply that for every two points a ∈ γ, b ∈ γa and points
a ∈ γ, b ∈ γa , lying in the corresponding rays and such that |pa| = |pa|,
|p′ b| = |p′ b|, the equality |ab| = |ab| holds. This means that lines γ and γp
are parallel.
Note that we have proved that for every sequence {ti }, ti → +∞,
shortest paths [pγ(ti )] converge to a half-line of a geodesic parallel to γ
if they converge at all. Since such a geodesic is unique (Lemma 10.5.6),
it follows that these shortest paths converge to this half-line γp+ for every
sequence {ti }, ti → +∞. Moreover the angles between these paths and γp+
go to zero.
Step 2. Applying Lemma 10.5.6, we see now that X is split into parallel
lines, and we have the relation “one is the projection of the other” between
points. We are going to show that this relation is symmetric and transitive.
To prove that this relation is symmetric, consider two parallel lines γ1
and γ2 and points a1 ∈ γ1 , a2 ∈ γ2 such that a2 is the projection of a1 to γ2 .
Consider triangles a1 a2 γ(t) as |t| → ∞. Since ∡a1 a2 γ(t) = ∡a e 1 a2 γ2 (t) =
π/2 and |a2 γ2 (t)| → ∞, we have ∡a2 a1 γ2 (t) = ∡a e 2 a1 γ2 (t) → π/2. Therefore
[a1 a2 ] is orthogonal to γ1 , or, equivalently, a1 is the projection of a2 to γ1 .
To prove that the projection relation is transitive, consider three parallel
lines γ1 , γ2 , γ3 and let a2 and a3 be the projections of a point a1 ∈ γ1 to γ2
and γ3 , respectively. Then
|a2 γ1 (t)| − |a3 γ1 (t)| → 0, t → +∞,
because |a2 γ1 (t)|−|a1 γ1 (t)| → 0 and |a3 γ1 (t)|−|a1 γ1 (t)| → 0. It follows that
e 2 a3 γ(t) → π/2, and therefore [a2 a3 ] is orthogonal to γ + = lim[a3 , γ1 (t)].
∡a 3
Hence a3 is the projection of a2 to γ3 .
10.6. Dimension and Volume 369

Step 3. Now it is not difficult to finish the proof of the theorem. For a
point p ∈ γ denote by N (p) the set of projections of p to all lines parallel
to γ. This set is convex in X and hence is a length space of nonnegative
curvature. Every two such sets N (p) and N (q) are canonically isometric via
an isometry defined as follows: if x ∈ N (p), then the image of x is N (q) ∩ γx ,
where γx is the line parallel to γ and passing through x.
Fixing p = γ(0), consider the set N (p)×R1 and the map f : N (p)×R1 →
M such that f (x, t) is the point of N (γ(t)) which corresponds to x under
the canonical isometry between N (p) and N (γ(t)). It is clear that f is an
isometry. ¤

Remark 10.5.7. There is a different (and ideologically somewhat better)

exposition of the proof of the theorem based on convexity properties of Buse-
mann functions and horoballs (see 5.3.3). In Hadamard spaces (outward)
equidistants (r-neighborhoods) of convex sets are convex, and hence so are
all horoballs. Contrary to that, in spaces of nonnegative curvature inward
equidistants preserve convexity; in particular, the reader can prove as an ex-
ercise that complements to horoballs are convex. This allows us to construct
a family of convex sets exhausting X. In the case of manifolds, one can use
smoothness properties of horoballs. Dividing the line (from the condition
of the theorem) into two rays and studying their Busemann functions and
corresponding families of horoballs easily gives a proof of a Riemannian case
of the theorem. A similar argument allows us to prove the Cheeger–Gromoll
Soul Theorem. For Alexandrov spaces, where even defining boundaries of
convex sets requires a hard Stratification Theorem, this approach encoun-
ters technical difficulties. This is the reason why we use a modification of
the proof (following an argument due to Milka) avoiding such difficulties.

Remark 10.5.8. The Splitting Theorem implies that an Alexandrov space

X of curvature ≥ 1 and with diameter π is a spherical cone (see Exercise
10.4.3). Indeed, let Y be the cone over X; then Y contains a line (corre-
sponding to a pair of points in X at the distance π from each other). Then
Y is a direct product of R and some nonnegatively curved space Z. Since
X is a space of directions of Y (X = Σo (Y ) where o is the origin), it follows
that X is a spherical cone over the space of directions Σo (Z).

10.6. Dimension and Volume

10.6.1. Dimensional homogeneity.

Theorem 10.6.1. Let X be a complete locally compact space of curvature

bounded below. Then all open subsets of X have the same Hausdorff dimen-
sion, namely dimH (U ) = dimH (X) for every open set U ⊂ X.
370 10. Spaces of Curvature Bounded Below

The property expressed in the theorem is referred to as dimensional

homogeneity of X. We do not exclude a priori noninteger and infinite values
for dimension.

Proof. Every open set U contains a ball Br (p) for some p ∈ X and r > 0,
and dimH (Br (p)) ≤ dimH (U ) ≤ dimH (X) because Br (p) ⊂ U ⊂ X. It
is sufficient to prove that dimH (Br (p)) = dimH (X). We will show that
dimH (Br (p)) = dimH (BR (p)) for every R > r. The theorem follows from
this, because X is can be represented as a union of countably many balls of
the form BR (p); for example, take R = r + 1, r + 2, . . . . If the dimensions
of these balls equal dimH (Br (p)), it follows that dimH (X) = dimH (Br (p));
cf. Proposition 1.7.19.
Denote A = Br (p) and B = BR (p). The inequality dimH (A) ≤ dimH (B)
is trivial because A ⊂ B. To prove that dimH (B) ≤ dimH (A), we construct
a map f : B → A such that for some c > 0 one has |f (x)f (y)| ≥ c|xy| for
all x, y ∈ A. If such a map f exists, its inverse f −1 : f (B) → B is Lipschitz;
therefore by Proposition 1.7.19 we have
dimH (B) ≤ dimH (f (B)) ≤ dimH (A).
This map f is used so many times in this chapter that it deserves a special
description:
Lemma 10.6.2. Let X be a complete locally compact space of curvature
≥ k, p ∈ X, and 0 < λ < 1. For every x ∈ X, let f (x) be a point in some
(arbitrarily chosen) shortest path [px] such that |pf (x)| = λ|px|. Then
(1) If k = 0, then |f (x)f (y)| ≥ λ · |xy| for all x, y ∈ X.
(2) If k < 0, then for every R > 0 there is a positive number c(k, λ, R)
such that |f (x)f (y)| ≥ c(k, λ, R) · |xy| for all x, y ∈ BR (p).
sinh(−kλR)
In fact, one can take c(k, λ, R) = sinh(−kR) .

Proof. Follows immediately from monotonicity of angles. The number

c(k, λ, R) is a coefficient for which the inequality holds for the similar map
in an R-ball in the k-plane. (The map in the k-plane is a diffeomorphism
and hence is bi-Lipschitz in every ball, so a desired constant exists. A
straightforward computation yields c(k, λ, R) = sinh(−kλR)
sinh(−kR) .) ¤

We refer to the map f from the lemma as a λ-homothety centered at p.

To finish the proof of the theorem, set λ = r/R and let f be a λ-
homothety centered at p. Then f (B) = f (BR (p)) ⊂ Br (p) = p and
|f (x)f (y)| ≥ c(k, λ, R) · |xy|
for all x, y ∈ BR (p). As explained above, the theorem follows from this. ¤
10.6. Dimension and Volume 371

Exercise 10.6.3. Prove that, if X is a locally compact (but not necessarily

complete) space of curvature bounded below, then all precompact open sets
in X have the same Hausdorff dimension.
Remark 10.6.4. In fact, the theorem remains valid without the local
compactness assumption. This can be proved using results of Section 10.8;
see Exercise 10.8.22 there.
Corollary 10.6.5. Let X be a complete locally compact space of curvature
bounded below. Then dimH (U ) = dimH (X) for every open set U ⊂ X.

Proof. Since X is complete and locally compact, all balls in X are pre-
compact. By Theorem 10.6.1 all these balls have the same Hausdorff
dimension. Since X can be covered by a countable collection of balls (e.g.,
by balls of integer radii centered at a fixed point), dimH (X) equals the
dimension of these balls; cf. Proposition 1.7.19. ¤

10.6.2. Gromov–Bishop inequality. It is easy to see from the proof

of dimensional homogeneity that the Hausdorff measure of a larger ball
BR (p) can be estimated from above in terms of the measure of a smaller
ball Br (p). In other words, a lower curvature bound imposes an upper
bound for the growth rate of volumes of balls. In fact, the volumes of balls
in an Alexandrov space grow not faster than the volumes of balls in the
comparison space of the respective dimension. This fact is known as the
Gromov–Bishop inequality.
Notation. Recall that a space form is a simply connected complete space of
constant curvature, i.e., a sphere, a Euclidean space, or a hyperbolic space.
For an integer n ≥ 2, we denote the n-dimensional space form of curvature
k by Mkn . As a special convention, we set Mk1 = R if k ≤ 0 and Mk1 = √1k S 1
√
(that is, the circle of length 2π/ k) if k > 0. We fix an n ≥ 1 and denote
by Vrk and Srk the volume of the r-ball and the (n − 1)-dimensional area of
the r-sphere in Mkn .
Theorem 10.6.6 (Gromov–Bishop inequality). Let X be a locally compact
Alexandrov space of curvature ≥ k and n be a positive integer. Then for
every p ∈ X the ratio
µn (Br (p))
Vrk
is nonincreasing in r, where µn (Br (p)) is the n-dimensional Hausdorff
measure of a ball of radius r > 0 centered at p. In other words, if R ≥ r > 0,
then
µn (BR (p)) µn (Br (p))
k
≤ .
VR Vrk
372 10. Spaces of Curvature Bounded Below

We give a complete proof only for k = 0 (and this proof is essentially

the same as that of dimensional homogeneity). In the general case, we give
a proof based on the formula
Z r
(10.5) µn (Br (p)) = µn−1 (St (p)) dt
0

where St (p) is the sphere of radius t centered at p. If X is a Riemannian

manifold, (10.5) is a simple particular case of the so-called coarea formula
(Theorem 3.2.11 in [Fe]). For general Alexandrov spaces of curvature
bounded below formula (10.5) is still valid but its proof requires hard
technical work.
An alternative approach is to modify the proof below so that it uses
finite sums instead of the integral formula (10.5). Namely, one can split
Br (p) into thin layers by a collection of spheres of appropriate radii and
replace the inequalities involving (n − 1)-dimensional volumes of spheres by
similar ones with n-volumes of these layers. Then, replacing the integral
in (10.5) by the sum of volumes of the layers, one can obtain the desired
inequality with any given precision. This modification is straightforward but
the resulting argument is full of irrelevant technical details about choosing
the radii, estimating correction terms and so on. So we present a more
clear argument with the formula (10.5), leaving a more elementary (but
technically more involved) method as exercise to the reader.
We begin however with a simple proof in the case k = 0.

Proof of Theorem 10.6.6 for k = 0. Let f : X → X be a (r/R)-homo-

thety centered at p (cf. Lemma 10.6.2). Clearly f maps BR (p) to Br (p).
Lemma 10.6.2 implies that f is injective and its inverse f −1 is Lipschitz
with Lipschitz constant R/r. Therefore

VR0
µn (BR (p)) ≤ (R/r)n µn (Br (p)) = µn (Br (p))
Vr0

and the theorem follows. ¤

Note that for k 6= 0 the same argument proves an inequality of the form
µn (BR (p)) ≤ C(r, R, k)µn (Br (p)) for some (nonoptimal) constant C(r, R, k).
This is sufficient for many purposes.
To prove the theorem in full generality, we need the following lemma
which is similar to the Gromov–Bishop inequality but involves spheres
instead of balls.
10.6. Dimension and Volume 373

Lemma 10.6.7. Let St (p) denote the sphere of radius t centered at p ∈ X.

Then
µn−1 (Sr (p)) Sk
(10.6) ≥ rk
µn−1 (SR (p)) SR
if R ≥ r > 0.

Proof. We will prove the inequality for k = 1. By means of re-scaling, the

statement then follows for every k > 0. To obtain a proof for k < 0, simply
replace sines in the formulas by hyperbolic sines.
Consider an (r/R)-homothety f (see Lemma 10.6.2) restricted to the
sphere SR (p), and let f0 be the analog of f in the 1-plane. We consider f0
as a map from the R-sphere in the 1-plane to the r-sphere in the 1-plane.
The derivative of f0 multiplies lengths of all tangent vectors by sin r/ sin R.
Hence f0 is locally Lipschitz with local Lipschitz constant sin r/ sin R and
f0−1 is locally Lipschitz with local Lipschitz constant sin R/ sin r. Therefore
for every ε > 0 there is a δ > 0 such that
|f0 (x)f0 (y)| sin r
≥ (1 − ε)
|xy| sin R
whenever |xy| < δ. By triangle comparison the same statement holds for f .
Let us partition SR (p) into finitely or countably many sets Di of diameter
less than δ. The above inequality implies that
µ ¶
sin r n−1
µn−1 (f (Di )) ≥ (1 − ε) ;
sin R
hence
X µ ¶n−1
sin r
µn−1 (Sr (p)) ≥ µn−1 (f (Di )) ≥ (1 − ε) µn−1 (SR (p))
sin R
Srk
= (1 − ε) k
µn−1 (SR (p)).
SR
Since ε is arbitrary, the lemma follows. ¤

Proof of Theorem 10.6.6. It is a matter of simple computation to derive

the Gromov–Bishop inequality from (10.6) and (10.5). Define V (t) =
µn (Bt (p)) and S(t) = µn−1 (St (p)). Integrating the inequality (10.6) in r
over [0, R] yields
V (R) Vk
≥ Rk ,
S(R) SR
or, equivalently (with R replaced by t),
d S(t) Sk d
log V (t) = ≤ tk = log Vtk .
dt V (t) Vt dt
374 10. Spaces of Curvature Bounded Below

Integrating this in t over [r, R] yields the assertion of the theorem.

To justify the above computations with derivatives, observe that (10.6)
implies that the function t 7→ S(t) is a product of the continuous function
t 7→ Stk and a monotone function. Hence for all t except at most countably
many values, S(t) is continuous at t and equals dV /dt. ¤

10.6.3. Bishop inequality. Suppose that X is an n-dimensional Rie-

mannian manifold of curvature ≥ k. Fix a p ∈ X. Then volumes of small
balls in X centered at p are almost Euclidean; hence
µn (Br (p))
→ 1, r → 0.
Vrk
Then the Gromov–Bishop inequality implies that µn (BR (p)) ≤ VRk for every
R > 0. This fact (known as the Bishop inequality) can be generalized to
all Alexandrov spaces.
Theorem 10.6.8 (Bishop inequality). Let n be a positive integer and X an
n-dimensional Alexandrov space of curvature ≥ k. Then for every p ∈ X
and every r > 0
µn (Br (p)) ≤ Vrk ,
where Vrk is volume of a ball of radius r in the space form Mkn (see the
notations in Subsection 10.6.2).

If k > 0, the Bishop inequality and the upper bound for the diameter
(Theorem 10.4.1) imply that the volume of
√ X is no greater than the volume
of the n-dimensional sphere of radius 1/ k. In other words, the following
statement holds.
Corollary 10.6.9. If X is an n-dimensional Alexandrov space of curvature
≥ k where k > 0, then
µn (S n )
µn (X) ≤ .
k n/2
Theorem 10.6.8 immediately follows from the next proposition. The
proposition says that the ball Br (p) is “not greater” than the r-ball in Mkn
in the sense that there is a noncontracting map from one to the other.
Proposition 10.6.10. Let X be an n-dimensional Alexandrov space of
curvature ≥ k, and p ∈ X. Then there exists a map from f : X → Mkn
such that |f (x)f (y)| ≥ |xy| for all x, y ∈ X (i.e., f is noncontracting) and
|f (p)f (x)| = |px| for all x ∈ X.

Proof. We argue by induction in n using the following fact (Theorem 10.8.6)

proved later in this chapter: if n ≥ 2 and X is an n-dimensional Alexandrov
10.6. Dimension and Volume 375

space, then the space of directions Σp (X) at every point p ∈ X is an (n − 1)-

dimensional Alexandrov space of curvature ≥ 1; if X is one-dimensional,
then Σp (X) consists of one or two points. This fact and the inductive
hypothesis imply that there exists a noncontracting map from Σp (X) to
M1n−1 = S n−1 .
Denote by Kpk (X) the k-cone (cf. Subsection 10.2.1) over the space of
directions Σp (X). The following lemma follows immediately from angle
comparison.
Lemma 10.6.11. Let X be an Alexandrov space of curvature ≥ k and
p ∈ X. Consider the map logp : X → Kpk (X) which maps p to the origin
of the cone and every point x ∈ X \ {p} to the pair (ξ, |px|) where ξ is the
direction of some (arbitrarily chosen) shortest path [px]. This map logp is
noncontracting.

By inductive hypothesis, we have a noncontracting map from Σp (X)

to S n−1 . It naturally extends to a noncontacting map F : Kpk (X) → Mkn
preserving the distance from the origin (note that Mkn is the k-cone over
S n−1 ). Then the map F ◦ logp is a desired noncontracting map from X
to Mkn . ¤
Exercise 10.6.12. Prove that the inequality in Theorem 10.6.8 turns into
equality if and only if Br (p) is isometric to the r-ball in Mnk . In particular,
if the equality takes place for all r, then X is isometric to Mkn .
Remark 10.6.13. The Gromov–Bishop and Bishop inequalities are also
valid for Riemannian manifolds with a lower bound for Ricci curvature.

10.6.4. Appendix: Rough dimension. Here we introduce one more

notion which simplifies considerations in many cases. It is especially useful
in the case when a space is not supposed to be locally compact and finite-
dimensional. Let X be a metric space. For every ε > 0, denote by βX (ε)
the maximal cardinality of an ε-separated subset of X, that is, a collection
of points pi ∈ X such that |pi pj | ≥ ε for all i 6= j. The case βX (ε) = ∞ is
not excluded.
Definition 10.6.14. The rough d-dimensional volume Vold X is
lim sup εd βX (ε).
ε→0

Rough dimension dimr is defined by the equality

dimr (X) = sup{d : Vold X = ∞}.
Exercise 10.6.15. Prove that
1. dimH X ≤ dimr X, where dimH denotes Hausdorff dimension.
376 10. Spaces of Curvature Bounded Below

Hint: this is obvious.

2. For every map f : X → Y such that |f (x)f (y)| ≥ c|xy| for some
constant c, one has dimr f (X) ≥ dimr X.
3. dimr (X) = inf{d : Vold (X) = 0};
4. If f : X → Y is a Lipschitz map, then dimr f (X) ≤ dimr X. In
particular, dimr f (X) = dimr X if f is bi-Lipschitz.

10.7. Gromov–Hausdorff Limits

In this section we consider the class of Alexandrov spaces as a subset of
the Gromov–Hausdorff “space of metric spaces”. First of all, Toponogov’s
Theorem makes it possible to prove that Alexandrov spaces form a closed
set (note that the similar statement about spaces of curvature ≤ k is false).
Proposition 10.7.1. A Gromov–Hausdorff limit of Alexandrov spaces of
curvature ≥ k is itself a space of curvature ≥ k. The same is true for
Gromov–Hausdorff limits of pointed spaces (in the sense of Definition 8.1.1).

Proof. By Theorem 7.5.1, a limit of length spaces is a length space. So it

suffices to verify the quadruple condition (cf. Proposition 10.1.1) for a limit
space. Let Xn −→ GH X where {Xn } is a sequence of spaces of curvature ≥ k.
Then for every quadruple (a; b, c, d) in X there is a sequence of quadruples
(an ; bn , cn , dn ) in Xn such that |an bn | → |ab|, |an cn | → |ac|, and so on.
Then the quadruple condition for (a; b, c, d) follows by continuity from that
for (an ; bn , cn , dn ). ¤

The main result of this section is the following Gromov’s Compactness

Theorem. For every k ∈ R, D > 0 and an integer n ≥ 0 we define
M(n, k) = {X : X is an AS of curvature ≥ k and dimH (X) ≤ n},
M(n, k, D) = {X ∈ M(n, k) : diam(X) ≤ D}.
Theorem 10.7.2 (Gromov’s compactness theorem). The class M(n, k, D),
regarded with the Gromov–Hausdorff metric, is compact. In particular,
M(n, k) is compact if k > 0.

An alternative formulation is: M(n, k) is boundedly compact, i.e., all its

closed bounded subsets are compact.
Some ingredients of the theorem will be proved later. Namely, in Section
10.8 we will show that
• All finite-dimensional Alexandrov spaces are locally compact (The-
orem 10.8.1). Therefore all bounded Alexandrov spaces are com-
pact, so M(n, k) is indeed a subset of the Gromov–Hausdorff metric
space.
10.7. Gromov–Hausdorff Limits 377

• The dimension of a finite-dimensional Alexandrov space is always

an integer (Theorem 10.8.2).
• A Gromov–Hausdorff limit of Alexandrov spaces of dimension ≤ n
has dimension ≤ n as well (Corollary 10.8.25). Since the diameter
is a continuous function, it follows that M(n, k) is a closed set in
the Gromov–Hausdorff space.
Given all this, it is sufficient to prove the following
Proposition 10.7.3. For every n, k, D, the class of all Alexandrov spaces
X of curvature ≥ k with dimH (X) = n and diam(X) ≤ D is precompact.

Proof. According to Theorem 7.4.15, it suffices to prove that for every ε > 0
there is an N = N (ε, n, k, D) > 0 such that a space from this class cannot
contain an ε-separated set of more than N points.
Let X be an n-dimensional Alexandrov space of curvature ≥ k with
diam(X) ≤ D. By Proposition 10.6.10, there is a noncontracting map from
X to the ball of radius D in the space form Mkn . Let N be the maximal
possible cardinality of an ε-separated set in the D-ball in Mkn . Then X
cannot contain an ε-separated set of more than N points, because the image
of an ε-separated set under a noncontracting map is ε-separated. ¤
Remark 10.7.4. There is another proof which appears simpler when all
necessary components from the next sections are substituted. Instead of
Proposition 10.6.10, one needs the Gromov–Bishop inequality (a simple
version with nonsharp constants is sufficient), and the following fact implied
by the Gromov–Bishop inequality and Theorem 10.8.3: the volume µn of an
n-dimensional bounded Alexandrov space is positive and finite.
Now let X be as above. Then by the Gromov–Bishop inequality we have
µn (Bε (p)) ≥ c(n, k, D)µn (BD (p)) = µn (X)
for every p ∈ X (note that B D (p) = X since diam(X) ≤ D). Therefore X
cannot contain more than N = c(n, k, D)−1 disjoint ε-balls, or, equivalently,
cannot contain a (2ε)-separated set of more than N points. The proposition
follows.
Remark 10.7.5. The method described in the previous remark also proves
the following statement: the class of Riemannian n-manifolds of Ricci curva-
ture ≥ k and of diameter no greater than D is precompact in the Gromov–
Hausdorff space (because the Gromov–Bishop inequality works with lower
bounds for the Ricci curvature). However, no reasonable description is
known for the closure of this class in the Gromov–Hausdorff space.
Remark 10.7.6. A limit of n-dimensional spaces may have dimension
strictly less than n. For example, for every compact nonnegatively curved
378 10. Spaces of Curvature Bounded Below

space X rescaled spaces {λX} are nonnegatively curved and converge to

a point as λ → 0. It is possible to make the dimension drop in the limit
keeping both lower and upper curvature bounds. For example, “thin” flat
tori S 1 × (λS 1 ) converge to the circle as λ → 0.
Sequences of Alexandrov spaces whose limits have lower dimension are
called collapsing sequences.

10.8. Local Properties

Now we pass to local properties of finite-dimensional Alexandrov spaces
(recall that by dimension we mean the Hausdorff dimension). This section
is very technical. For the first reading we recommend skipping long proofs
in subsection 10.8.2.

10.8.1. Formulations. Below are some of the formulations that we prove

(or explain how to prove) in this and the next sections. The first statement
in the list is the local compactness (the property that we assumed in so
many places to simplify technical details in proofs).
Theorem 10.8.1. All finite-dimensional Alexandrov spaces are locally com-
pact.
Theorem 10.8.2. The Hausdorff dimension of any Alexandrov space is an
integer or infinity.
Theorem 10.8.3. An n-dimensional Alexandrov space contains an open
dense set which is an n-dimensional manifold. Moreover, there is an open
dense set such that every its point has a neighborhood bi-Lipschitz homeo-
morphic to an open region in Rn .

These three theorems are proved in the end of Subsection 10.8.2; see
Corollaries 10.8.20, 10.8.21 and 10.8.23.
Theorem 10.8.3 is the first step in the study of the local topological
structure of an Alexandrov space. According to it, a space is divided into the
set of topologically regular points (those having Euclidean neighborhoods),
and a nowhere dense set of topological singularities. To obtain more detailed
information, one can study the set of singular points; some results about it
are discussed later in Section 10.10.
Another way to obtain a classification of points into good and bad ones
is based on tangent cones. Namely, a point is said to be regular if the
tangent cone at it is Euclidean, i.e., isometric to Rn , or, equivalently, the
space of directions is isometric to the standard sphere S n−1 .
Theorem 10.8.4. Let X be an n-dimensional Alexandrov space and p ∈ X.
Then the following three assertions are equivalent.
10.8. Local Properties 379

(1) p is regular.
(2) For every ε > 0 there is a neighborhood U of p at a Lipschitz
distance less than ε from some open set in Rn . (See Section 7.2 for
the definition of Lipschitz distance.)
(3) The Gromov–Hausdorff tangent cone of X at p is isometric to Rn .
In other words, the dilated balls 1ε Bε (p) converge in the Gromov–
Hausdorff sense to the unit ball of Rn as ε → 0.

This theorem is composed of Theorem 10.9.3 and Theorem 10.9.16. In

fact, we prove a more general statement: if the space of directions Σp (X) is
close to S n−1 in the Gromov–Hausdorff metric, then a neighborhood of p is
close to an open set in Rn in the Lipschitz metric.

Theorem 10.8.5. Let X be an n-dimensional Alexandrov space. Then

the set of regular points is dense in X, and moreover is an intersection of
countably many open sets.

This theorem is proved as Corollary 10.9.13. In fact, the set of nonregular

points has Hausdorff dimension no greater than n − 1, but we do not prove
this. See Section 10.10 for more detailed formulations and discussion.
Spaces of directions and tangent cones are considered in Section 10.9.
Among other things, we prove the following theorem (see Corollary 10.9.6).

Theorem 10.8.6. Let X be an n-dimensional Alexandrov space with n ≥ 2.

Then for every point p ∈ X the space of directions Σp (X) is an Alexandrov
space of curvature ≥ 1 and of dimension n − 1. A space of directions of a
one-dimensional Alexandrov space consists of one or two points.

This theorem sounds very similar to the respective fact (Theorem 9.1.44)
about spaces of curvature bounded above. However the proof for the case
of curvature bounded below is not that easy. Even the fact that the
metric of Σp (X) is intrinsic appears difficult to prove. We prove Theorem
10.8.6 using Gromov–Hausdorff convergence. The core facts are: Σp (X) is
compact, and the tangent cone Kp (X) is also the Gromov–Hausdorff tangent
cone of X at p. Then the desired properties of Σp (X) follow from the
respective properties of Kp (X), which in turn follow from general properties
of Gromov–Hausdorff limits.

Remark 10.8.7. A common technique when using a space of directions

is to carry out an induction on dimension. This is the reason why we do
not take our usual approach “prove only for locally compact spaces” in this
section. Even if we did, we could not avoid dealing with possibly not locally
compact spaces when passing down to the space of directions.
380 10. Spaces of Curvature Bounded Below

Note that compactness of the space of directions does not follow from
(local) compactness of a space itself, if one does not assume that the di-
mension is finite. For example, consider a direct metric product of infinitely
many spaces Xi where Xi is a sphere of radius 1/i and i ranges over all
positive integers. This product is a compact length space of nonnegative
curvature; however its space of direction at any point is not compact (it
contains an infinite (π/2)-separated set and in fact is isometric to a sphere
in an infinite-dimensional Hilbert space).

Exercise 10.8.8. Give a definition of a direct metric product of a (good

enough) countable collection of metric spaces, and prove the statements
made in this remark.

We fix a number k and let X denote an Alexandrov space of curvature

≥ k. Since the formulations do not depend on k, we assume that k ≤ 0 (to
make sure that comparison triangles always exist). Since our considerations
are local, we may assume k is arbitrarily close to 0 by means of re-scaling.
For simplicity we assume that k = 0. Making the proofs work for a negative
k is a matter of introducing negligibly small correction terms in the formulas.

10.8.2. Strainers. The notion of a strainer is a useful technical tool which

somewhat resembles the notion of an orthogonal frame in Rn , and is used
for the same purpose—to introduce a (local) coordinate system. To see
how one may come up with a definition of a strainer, consider a point
p ∈ X such that Σp (X) is isometric to S n−1 . The sphere contains n pairs of
points such that the (angular) distance between points in each pair equals π,
and the distances between points from different pairs all equal π/2. These
points correspond to n curves γ1 , . . . , γn in X passing through p and having
mutually orthogonal directions at p, similarly to the coordinate axes in Rn .
Then one can introduce coordinates in a neighborhood of p as follows:
the coordinates of a point q equal the distance |pq| multiplied by the cosines
of the angles between a shortest path [pq] and the “coordinate axes”. This
gives the standard coordinates if X is the Euclidean space, but in the general
case these “coordinates” may even fail to be continuous (e.g., if a shortest
path [pq] is not unique). It is clear how to get rid of discontinuity problems
once and forever: fix a point ai in each γi and use comparison angles
e
(∡qpa i ) instead of angles. (In fact, this does not solve all the problems,
only some of them. The coordinates that we actually introduce later are
slightly different.) After this, “coordinate axes” and even their directions
are no longer needed, and one formulates the following
10.8. Local Properties 381

Definition 10.8.9. A point p ∈ X is an (m, ε)-strained point if there are

m pairs of points (ai , bi ) in X such that
e i pbi > π − ε,
∡a
e i paj > π − 10ε,
∡a 2
e i pbj > π − 10ε,
∡a 2
e i pbj > π − 10ε
∡b 2

for all i, j ∈ {1, . . . , m}, i 6= j. The collection {(ai , bi )} itself is called an

(m, ε)-strainer for p.

The constant 10 in the definition is not very important. It is introduced

only to simplify the formulation of one of the subsequent lemmas.

Remark 10.8.10. Recall that ∡a e i pbi + ∡a

e i paj + ∡a
e j pbi ≤ 2π by the
e
quadruple condition. It follows that the angles ∡ai paj and ∡a e j pbi (and,
e i pbj ) are not only greater than π − 10ε, they are also less than
similarly, ∡b 2
π π
2 + 11ε; i.e., they are actually close to 2 .

One does not have to send ε to zero. For our purposes it is sufficient to fix
1
once and forever a small ε, say, ε < ε0 = 100m . If ε satisfies this inequality,
we omit it and write simply “m-strainer” and “m-strained point”. Obviously
the set of (m, ε)-strained points is open for any fixed m and ε.
Definition 10.8.11. Let X be an Alexandrov space. The strainer number
of X is the supremum of numbers m such that there exists an m-strainer
in X.
A strainer number at a point x ∈ X is the supremum of numbers m such
that every neighborhood of x contains an m-strained point.

We will show in this section that the strainer number equals the Haus-
dorff dimension of the space (in particular, the latter is an integer or infinity).
Right now, observe that X admits an (1, ε)-strainer for every ε > 0 unless
X is a single point. To prove this, pick any two points a, b and let p be an
almost midpoint (more precisely, a δ-midpoint with δ depending on |ab| and
ε); then (a, b) is the desired (1, ε)-strainer for p.
The following proposition tells us that the notion of a strained point is
local, and moreover can be formulated in terms of the space of directions.
Proposition 10.8.12. 1. If {(ai , bi )} is an (m, ε)-strainer for p ∈ X, and
a′i ∈ [ai p], b′i ∈ [bi p] for i = 1, . . . , m, then {(a′i , b′i )} is an (m, ε)-strainer for
p as well. Even if shortest paths do not exist, the same is true for points a′i
and b′i taken in suitable almost shortest paths. In particular, there exists an
(m, ε)-strainer {(a′i , b′i )} with arbitrarily small distances |pa′i | and |pb′i |.
382 10. Spaces of Curvature Bounded Below

2. If {(ai , bi )} is an (m, ε)-strainer for p and shortest paths [pai ] and

[pbi ] exist for all i, then the angles between these shortest paths satisfy the
inequalities
∡ai pbi > π − ε,
∡ai paj > π2 − 10ε,
∡ai pbj > π2 − 10ε,
∡bi pbj > π2 − 10ε
for all i, j ∈ {1, . . . , m}, i 6= j.
3. Conversely, if the inequalities from the second statement hold for
some points p, {ai } and {bi } (i = 1, . . . , m), then p is an (m, ε)-strained
point.

Proof. The first statement follows from the monotonicity of angles. The
second one is obtained as the limit of the first one as a′i and b′i converge to p
along the respective shortest paths. To prove the third statement, observe
that {(a′i , b′i )} is an (m, ε)-strainer if a′i and b′i are sufficiently close to p in
the respective geodesics [pai ] and [pbi ]. ¤

The following two lemmas are simple but important technical facts that
are used everywhere in this section. The first one is an analog of the fact
that the sum of adjacent angles equals π, but with a strainer and comparison
angles instead of a geodesic and angles between directions.
Lemma 10.8.13. Let (a, b) be a (1, ε)-strainer for p, and q ∈ X be such
that
ε
|pq| < min{|pa|, |pb|}.
4
Then
e
|∡apq e
+ ∡bpq − π| < ε.
In particular, if shortest paths [pa], [pb] and [pq] exist, the angles between
them are close to the comparison angles:
e
0 < ∡apq − ∡apq < 2ε,
e
0 < ∡bpq − ∡bpq < 2ε.

Proof. By the quadruple condition for (p; a, b, q) we have

e
∡apq e
+ ∡bpq e
≤ 2π − ∡apb < π + ε.
e
It remains to prove that ∡apq e
+ ∡bpq > π − ε. Suppose the contrary, and
place the comparison triangles △apq and △bpq in different half-planes with
respect to [pq]. Then
e
∡apb = ∡apq + ∡bpq < π − ε < ∡apb;
 10.8. Local Properties 383

e
hence |ab| > |ab|. The inequality |pq| < 4ε |pa| implies ∡paq < arcsin 4ε < 2ε ;
then
e e ε e
∡aqp = π − ∡paq − ∡apq > π − − ∡apq,
2
e
and similarly ∡bqp > π − ε/2 − ∡bpq. Thus
∡aqp + ∡bqp > 2π − ε − (∡apqe e
+ ∡bpq) >π
(i.e., the angle at q of the quadrangle apbq is greater than π). Since
|ab| > |ab|, it follows that
e
∡aqb > ∡aqb = 2π − ∡aqp − ∡bqp,
e
or, equivalently, ∡aqb e
+ ∡aqp e
+ ∡bqp > 2π, contrary to the quadruple
condition for (q; a, b, p).
The second statement of the lemma (about angles between shortest
paths) follows from the first one. Indeed, the quadruple condition implies
that ∡apq + ∡bpq ≤ 2π − ∡apb < π + ε. Therefore
e
(∡apq − ∡apq) e
+ (∡bpq − ∡bpq) < (π + ε) − (π − ε) = 2ε.
e
Since both terms ∡apq − ∡apq e
and ∡bpq − ∡bpq are positive, it follows that
they are bounded above by 2ε. ¤
Lemma 10.8.14. Let p, a1 , a2 , b1 , b2 ∈ X be such that (a1 , b1 ) and (a2 , b2 )
are (1, ε)-strainers for p,
ε
|a2 b2 | < min{|pa1 |, |pb1 |}
4
and ¯ ¯
¯|a1 a2 | − |a1 b2 |¯ < ε|a2 b2 |.
Then {(a1 , b1 ), (a2 , b2 )} is a (2, ε)-strainer for p.
PSfrag replacements
a1 a2
a2 b1
p
b1 a1
b2
p b2

Figure 10.2: Lemma 10.8.14.

Proof. It is sufficient to prove that the comparison angles ∡a e 1 pa2 and

e 1 pb2 are greater than − 8ε. Indeed, if this is true, then the quadruple
∡a π
2
condition for (p; a1 , a2 , b2 ) implies that these two angles are less than π2 + 9ε
(compare with Remark 10.8.10 after the definition of a strainer). Note that
|pa2 | < |a2 b2 |, |pb2 | < |a2 b2 |, so Lemma 10.8.13 is applicable to the 1-strainer
384 10. Spaces of Curvature Bounded Below

(a1 , b1 ) for p and points q = a2 , q = b2 . WWW Then Lemma 10.8.13 implies

that the angles ∡a e 2 pb1 and ∡b
e 1 pb2 are greater than π − 10ε.
2
Suppose the contrary, for example; let ∡a e 1 pa2 < π − 8ε. Then
2
e 1 pa2 > π + 7ε
e 2 pb1 > π − ε − ∡a
∡a
2
by Lemma 10.8.13,
e 1 pb2 < 2π − ∡a
∡b e 1 pa2 < π − 6ε
e 2 pb2 − ∡b
2
by the quadruple condition for (p; a2 , b1 , b2 ), and
e 1 pb2 > π − ε − ∡b
∡a e 1 pb2 > π + 5ε
2
by Lemma 10.8.13. We are going to show that the difference |a1 b2 | − |a1 a2 |
is too large to satisfy the last condition of the lemma.
Here and later in this section we need the following elementary fact: if
△xyz is a triangle in R2 , then
(10.7) |xz| > |xy| + |yz| · sin(∡xyz − π2 ).
To prove this inequality, observe that the right-hand side of the inequality
equals the distance between x and the projection of z in the line xy.
e 1 a2 < ε ; therefore
Since |pa2 | < |a2 b2 | < ε |pa1 |, we have ∡pa
4 2
e 1 pa2 − ε > π + 7ε.
e 1 a2 p > π − ∡a
∡a
2 2
Applying (10.7) to the comparison triangle for △a1 a2 p, we obtain that
e 1 pb2 > π + 5ε
|a1 p| > |a1 a2 | + sin(7ε)|a2 p|. Similarly the inequality ∡a 2
implies that |a1 b2 | > |a1 p| + sin(5ε)|pb2 |. Thus
|a1 b2 | > |a1 a2 | + sin(5ε)(|a2 p| + |pb2 |) > |a1 a2 | + ε|a2 b2 |,
which contradicts the last condition of the lemma. ¤

Now we introduce local coordinates in a neighborhood of a strained

point. We use distances to the points ai as coordinate functions. More
formally, if {(ai , bi )} is an m-strainer for p, we define the map f : U → Rm ,
where U is a neighborhood of p, by
f (x) = (|xa1 |, . . . , |xam |).
The functions x 7→ |xai | (and the map f itself) are referred to as distance
coordinates associated with the strainer {(ai , bi )}. Obviously f is a Lipschitz
map because its coordinate functions are.
The neighborhood U is supposed to be sufficiently small. First, the set
of x ∈ X such that {(ai , bi )} is an m-strainer for x is open (recall that an
1
m-strainer is an (m, ε)-strainer with ε = 100m ). So we may assume that U
10.8. Local Properties 385

is contained in this set. Second, we want the diameter of U be significantly

smaller than min{|pai |, |pbi |}.
If X = Rn and m = n, the level sets of the functions x 7→ |xai |
are spheres intersecting almost orthogonally near p. Therefore, for each
collection of radii (r1 , . . . , rn ) near (|px1 |, . . . , |pxn |) the respective spheres
have a unique intersection point near p. The same behavior occurs for a
maximal strainer in any Alexandrov space. This statement consist of two
parts: that f is an open map (i.e., the image of every open set is open), and
that f is an injective map whose inverse map is Lipschitz. We begin with
the first part (which holds for any strainer, not only a maximal one).
1
Proposition 10.8.15. Let p ∈ X be an (m, ε)-strained point with ε = 100m .
Then there is a neighborhood U of p such that the distance coordinates form
an open map from U to Rm .

Proof. Let {(ai , bi )} be an (m, ε)-strainer for p and f : U → Rm be the

associated distance coordinates. We may assume that |qai | > 1 and |qbi | > 1
for all q ∈ U , i = 1, . . . , m. This can be achieved by means of re-scaling and
choosing U sufficiently small. Furthermore, we assume that U is so small
that {(ai , bi )} is an (m, ε)-strainer for every point q ∈ U . Let us show that
f (U ) contains a neighborhood of f (p) in Rm . Then the same arguments
applied to an arbitrary q ∈ U instead of p will prove that f (U ) contains a
neighborhood of f (q) and therefore f is an open map.
Let y = (y1 , . . . , ym ) ∈ Rm be a point in a small neighborhood of
f (p). We have to find a point x ∈ U such that f (x) = y. Such a
point is obtained by means of consecutive approximations. The idea is the
following. Take x = p as the initial approximation. Then, at each step,
pick an i0 ∈ {1, . . . , m} and move x along [xai0 ] or [xbi0 ] until the distance
|xai0 | takes the desired value (i.e., becomes equal yi0 ). Since we move x
almost orthogonally to the other paths [xai ], the distances |xai | (i 6= i0 )
remain almost unchanged. Therefore the point f (x) gets closer to y. On
the other hand, the distance between the old and new positions of our point
x approximately equals the change in the i0 th coordinate of f (x) or, which
is almost the same, the distance between the old and new position of f (x).
Since f (x) approaches y quite rapidly, the path that it goes along cannot
be long; therefore x also does not run away and the process converges. The
formal details follow.
Set x0 = p and consider a sequence x0 , x1 , . . . of points where¯ xn+1 is¯
chosen recursively depending on xn as follows. PDefine δi = δi,n = ¯yi −|xn ai |¯
for i = 1, . . . , m and ∆n = ky − f (xn )k1 = m i=1 |δi,n |. It will follow by
induction that ∆n < ∆0 , and we may assume that ∆0 < ε/10. Suppose that
xn ∈ U (this also has to be verified later by induction). Define δ = maxi |δi |
386 10. Spaces of Curvature Bounded Below

1
and choose i0 for which |δi0 | = δ. Clearly m ∆n ≤ δ ≤ ∆n . Now let
xn+1 be a point in the union of shortest paths [xn ai0 ] ∪ [xn bi0 ] such that
|xn+1 ai0 | = yi0 ; i.e., the i0 -th coordinate of f (xn+1 ) takes the desired value.
(If shortest paths do not exist, use “almost shortest” paths instead.) We are
going to show that other coordinates of f (xn+1 ) are only slightly different
from those of f (xn ).
First, observe that |xn xn+1 | < 2δ. If |xn ai0 | > |xn+1 ai0 |, then xn+1 ∈
[xn ai0 ] and we have |xn xn+1 | = |xn ai0 | − |xn+1 ai0 | = δ. Otherwise xn+1 ∈
e i xn xn+1 ≥ ∡a
[xn bi0 ]; then ∡a e i xn bi > π − ε by monotonicity of angles,
0 0 0
and the inequality |xn xn+1 | < 2δ follows from (10.7) and the relation
|xn+1 ai0 | = |xn ai0 | + δ.
Recall that {(ai , bi )} is an (m, ε)-strainer for xn as long as xn ∈ U ,
and by Proposition 10.8.12 it remains an (m, ε)-strainer if one replaces ai0
or bi0 by xn+1 (depending on whether xn belongs to [xn ai0 ] or [xn bi0 ]).
Hence |∡ae i xn xn+1 − π | < 11ε for every i 6= i0 (cf. Remark 10.8.10). Since
2
|ai xn | > 1 and |xn xn+1 | < 2δ, we have ∡x e n ai xn+1 < 4δ < ε; hence
e i xn+1 xn − | < 12ε if i 6= i0 . Then (10.7) implies that
|∡a π
2
¯ ¯
¯|ai xn | − |ai xn+1 |¯ < |xn xn+1 | · sin(12ε) < 24δε
and therefore |δi,n+1 | < |δi,n | + 24εδ for all i 6= i0 . Thus
∆n+1 < ∆n − δ + 24(m − 1)εδ < ∆n − 21 δ < (1 − 1
2m )∆n
1
(here we use that ε ≤ 100m ). Note that we have verified one of our inductive
1 n
assumptions: ∆n+1 < ∆n < ∆0 . Moreover, ∆n < (1 − 2m ) ∆0 , so {∆n } is
a vanishing sequence and f (xn ) → y as n → ∞, provided that xn ∈ U for
all n.
It remains to prove that the sequence {xn } does not leave U and is a
Cauchy sequence. Recall that
1 n
|xn xn+1 | < 2∆n < 2∆0 (1 − 2m )
while xn ∈ U ; therefore
n
X
|pxn+1 | = |x0 xn+1 | ≤ |xj xj+1 | < ∆0 · C(m)
j=0
P
where C(m) = ∞ 1 j
j=0 (1 − 2m ) < ∞. We may choose our target y ∈ R
m

so close to f (p) that ∆0 = ky − f (p)k1 < diam(U )/C(m); then |pxn+1 | <
diam(U
P∞ ), and by induction the whole sequence {xn } is contained in U . Since
n=1 n xn+1 | < diam(U ) < ∞, the sequence {xn } is Cauchy one and its
|x
limit belongs in U . This limit is the desired point x such that f (x) = y. ¤
Corollary 10.8.16. The strainer number of an Alexandrov space is not
greater than its Hausdorff dimension.
10.8. Local Properties 387

Proof. If there is an m-strainer in X, then by Proposition 10.8.15 there

exists a Lipschitz map from a subset of X onto an open set V ⊂ Rm . Then
dimH (X) ≥ dimH (V ) = m because Lipschitz maps do not increase the
Hausdorff dimension (Proposition 1.7.19). ¤

Recall that an m-strainer is defined as an (m, ε)-strainer where ε is an

1
explicit small number, namely, ε = 100m . It might sound more natural if we
required existence of (m, ε)-strainers with arbitrarily small ε. Fortunately,
this stronger requirement follows automatically: once any m-strainer is
found, its “quality” can be improved to whatever value is needed.

Proposition 10.8.17 (improving a strainer). If p is an (m, ε)-strained point

1
with ε ≤ 100m , then for any ε′ > 0 any neighborhood of p contains an (m, ε′ )-
strained point.

Proof. Let {(ai , bi )} be an (m, ε)-strainer for p. Let us move these points
(including p) so as to obtain an (m, ε′ )-strainer. We will abuse notation and
denote the new points by the same letters p, ai , bi . The improvement of the
strainer is done in two stages. First, the angles ∡a e i pbi are made very close
π
to π. Then other angles are made very close to 2 .
To make an angle ∡a e i pbi close to π, do the following. Consider the
0 0
distance coordinates f : U → Rm associated with {(ai , bi )} where U is a
very small neighborhood of p. Since f (U ) is an open set in Rm , one can find
two points x = (x1 , . . . , xm ) and y = (y1 , . . . , ym ) in f (U ) such that xi = yi
for all i 6= i0 , and xi0 6= yi0 . Choose a and b in U such that x = f (a) and
y = f (b); then |ai a| = |ai b| for all i 6= i0 . Replace ai0 by a, bi0 by b, and
p by a midpoint p′ between a and b (or an almost midpoint p′ such that
e ′ b > π − ε′ ). Lemma 10.8.14 applied to points p′ , ai , bi , a, b implies that
∡ap
the new collection of points is an (m, ε)-strainer for the new point p = p′ .
Applying this construction for i0 = 1, . . . , m in turn, one obtains an
e i pbi > π − ε′ . (One may wonder why the relation,
(m, ε)-strainer in which ∡a
e 1 pb1 > π − ε′ , does not break down while we fix other angles. The
say, ∡a
answer is easy: the subsequent neighborhoods in which we choose points are
much smaller than the first one.)
Now apply the same procedure one more time. Since all (ai , bi ) are
now (1, ε′ )-strainers, Lemma 10.8.14 implies that the resulting collection of
points is an (m, ε′ )-strainer. ¤

Theorem 10.8.18. If p ∈ X is an m-strained point and m equals the

local strainer number at p, then p has a neighborhood which is bi-Lipschitz
homeomorphic to an open region in Rn .
388 10. Spaces of Curvature Bounded Below

A bi-Lipschitz homeomorphism is provided by distance coordinates asso-

ciated with any m-strainer. Moreover the Lipschitz constants of this map
and its inverse are not greater than 500m.

Proof. Let {(ai , bi )} be an m-strainer for a p ∈ X. Assume that there are no

(m + 1)-strained points in a small neighborhood U of p, and let f : U → Rm
be the map composed from the distance coordinates. We already know that
f is an open map (Proposition 10.8.15).
To demonstrate the idea of the proof, let us first show that f is injective
(and therefore is a local homeomorphism). Suppose it is not; i.e., there are
points a, b ∈ U such that |ai a| = |ai b| for i = 1, . . . , m. Then Lemma 10.8.14
implies that an almost midpoint between a and b is (m + 1)-strained by the
original points {ai } and {bi } and the pair (a, b) taken as (am+1 , bm+1 ).
To prove that f is bi-Lipschitz, we have to turn this argument to an
effective estimate for |f (a)f (b)| from
¯ below in ¯terms of |ab|. By Lemma
10.8.14, for at least one i we have ¯|ai a| − |ai b|¯ ≥ 4ε |ab| where ε = 100m
1
;
otherwise we have an (m + 1)-strainer¯ for an almost midpoint between a
and b. Then |f (a)f (b)| ≥ ||ai a| − |ai b|¯ ≥ 500m
1
|ab|; hence the dilatation of
−1
f is not greater than 500m.
The dilatation of f itself is not greater than m because the coordinate
functions are nonexpanding. ¤
Remark 10.8.19. The Lipschitz constant 100m in the theorem is a poor
achievement. A more delicate argument shows that the dilatations of f and
f −1 are (arbitrarily) close to 1 provided that p is an (m, ε)-strained point
with a sufficiently small ε.

Now we are in position to prove some of the theorems formulated in the

beginning of this section.
Corollary 10.8.20. All finite-dimensional Alexandrov spaces are locally
compact.

Proof. Let X be an Alexandrov space and m be its strainer number. We

know that m ≤ dimH (X) < ∞. Let p ∈ X be an m-strained point; then
by Theorem 10.8.18 some neighborhood U of X is homeomorphic to Rm .
In particular, U is locally compact, so there is an r > 0 such that the ball
Br (p) is precompact.
Let us prove that for every R > 0 the ball BR (p) is precompact. Suppose
the contrary; then there is an ε > 0 such that BR (p) contains an infinite
ε-separated set S. Recall the construction from Lemma 10.6.2: if the metric
is strictly intrinsic, one defines a homothety map f : BR (p) → Br (p) such
that |f (x)f (y)| ≥ Rr |xy|. Namely, the image f (x) of an x ∈ BR (p) is a
10.8. Local Properties 389

point in a shortest path [px] such that |pf (x)| = Rr |px|. The metric of X
is not strictly intrinsic a priori, but one can modify this construction using
“almost shortest” paths instead of shortest ones. This way an inequality
r
|f (x)f (y)| > 2R |xy| can be ensured for points x, y ranging over any given
finite set, in particular, any finite subset S ′ ⊂ S. Thus the ball Br (p)
r
contains an 2R -separated net f (S ′ ) of an arbitrarily large number of points
and therefore it is not precompact. This contradiction shows that every ball
in X is precompact; hence X is locally compact. ¤

Corollary 10.8.21. The Hausdorff dimension of a locally compact Alexan-

drov space (in particular, of any finite-dimensional one) equals its strainer
number.
Therefore the Hausdorff dimension of any Alexandrov space is an integer
or infinity.

Proof. Let X be an Alexandrov space and m be its strainer number.

We already know (Corollary 10.8.16) that m ≤ dimH (X). If m = ∞,
there is nothing else to prove. Otherwise consider a neighborhood U of
an m-strained point which is bi-Lipschitz homeomorphic to Rm . Since a
bi-Lipschitz homeomorphism preserves the Hausdorff dimension, we have
dimH (U ) = m. By dimensional homogeneity (Theorem 10.6.1) it then
follows that dimH (X) = dimH (U ) = m. ¤

Exercise 10.8.22. Prove the statement of the above corollary without the
local compactness assumption.
Hint: The only case not covered yet is the following: dimH (X) = ∞ but
m < ∞. Since m < ∞, there is at least one precompact open set (namely, a
neighborhood of an m-strained point). Show that the proof of dimensional
homogeneity (namely the construction of a homothety) works in this case.

Corollary 10.8.23. Let X be an n-dimensional Alexandrov space. Then

the set of n-strained points is an open dense set in X. Furthermore, this set
is an n-dimensional manifold.

Proof. The fact that the set of n-strained points is an n-manifold follows
immediately from Theorem 10.8.18. The first statement of the corollary is
equivalent to saying that the local strainer number at every point p ∈ X
equals n. Let m be the local strainer number at p. Then some neighborhood
of U of p is bi-Lipschitz homeomorphic to Rm and hence dimH (U ) = m.
Then m = dimH (X) = n by dimensional homogeneity. ¤

Combining Corollary 10.8.23 and Proposition 10.8.17 about improving

a strainer, we obtain the following
390 10. Spaces of Curvature Bounded Below

Corollary 10.8.24. Let X be an n-dimensional Alexandrov space. Then

for every ε > 0 the set of (n, ε)-strained points is an open dense set in X.

Finally, we prove the estimate for the dimension of a Gromov–Hausdorff

limit which was used in the proof of Compactness Theorem 10.7.2 and has
numerous other applications.
Corollary 10.8.25. A Gromov–Hausdorff limit of compact Alexandrov
spaces of curvature ≥ k and of dimension no greater than n is also an
Alexandrov space of curvature ≥ k and dimension no greater than n.
The same is true for Gromov–Hausdorff limits of pointed spaces.

Proof. The proofs for compact and pointed spaces are similar. Let {Xi } be
a sequence of compact Alexandrov spaces of curvature ≥ k, dimH (Xi ) ≤ n
for all i, Xi −→
GH X as i → ∞. By Proposition 10.7.1, X is a (compact)
space of curvature ≥ k. Suppose that dimH (X) > n. Since dimH (X) equals
the strainer number of X (Corollary 10.8.21), there is an (n + 1)-strained
point p ∈ X. Fix an (n + 1)-strainer {(aj , bj )}n+1
j=1 for p. As soon as Xi gets
sufficiently close to X, one can find points p , a′j , b′j whose distances from
′

one another are almost equal to the respective distances between points p,
aj , bj in X. Then the comparison angles (∡) e involving these points in Xi are
almost equal to the respective comparison angles in X, because the angles ∡ e
are continuous functions of distances. In particular, the inequalities from the
definition of a strained point are satisfied for p′ , a′j , b′j . Thus Xi contains an
(n + 1)-strainer, and therefore dimH (Xi ) ≥ n + 1, for all large enough i. ¤

10.9. Spaces of Directions and Tangent Cones

10.9.1. Basic properties.
Proposition 10.9.1. Let X be a finite-dimensional Alexandrov space. Then
for every p ∈ X the space of directions Σp (X) is compact.

Proof. The proof is based on the following

Lemma 10.9.2. There is a constant C = C(n, k) satisfying the following. If
X is an n-dimensional Alexandrov space of curvature ≥ k, p ∈ X, 0 < r < 1
and 0 < ε < 1, then the ball Br (p) cannot contain an (εr)-separated set
consisting of more than C/εn points.

Proof of the lemma. This lemma is nothing but a refinement of Corollary

10.8.20 about local compactness. All we need is to plug in explicit estimates
for the number of points in separated sets.
The statement is trivial for balls in Rn . It then follows for neighborhoods
of an n-strained point p, with a constant C written in terms of Lipschitz
10.9. Spaces of Directions and Tangent Cones 391

constants given in Theorem 10.8.18. Then, with a homothety map from

Lemma 10.6.2, the statement is extended to all balls. (One needs, however,
an upper bound for the radius to make this work for k < 0.) The details
follow.
Let p be an n-strained point and r be so small that U = Br (p) admits
a bi-Lipschitz map f : U → Rn with dil f ≤ c1 and dil f −1 ≤ c1 where
c1 = 500n (cf. Theorem 10.8.18). Then every (εr)-separated set in Br (p)
corresponds to an (εr/c1 )-separated set in a ball of radius c1 r in Rn . Such
a set cannot contain more than C/εn points where C = c2 · c21 , c2 is the
respective constant for Rn .
Almost the same constant C works for every ball centered at any point.
Indeed, consider an (εr)-separated set S ⊂ Br (p) where p ∈ X is an arbitrary
point and 0 < r < 1. Find an n-strained point q near p and a small
ball U = Bδ (q) satisfying Theorem 10.8.18. Consider a (δ/2r)-homothety
f : B2r (q) → Bδ (q). We have |f (x)f (y)| ≥ (δ/2r) · |xy| for all x, y; hence
f (S) is an (εδ/2)-separated set in Bδ (q). Therefore the cardinality of f (S),
and that of S, are no greater than C(ε/2)n .
These formulas should be slightly modified if k < 0. It suffices to
multiply the constant by sinhm (−k) which corresponds to the correction
term sinh(−k) arising in the lower bound for |f (x)f (y)| in the unit ball of
the k-plane. ¤

Now continue the proof of the proposition. We will show that for every
ε > 0 the space of directions cannot contain an ε-separated set of more
than C/εn points where C is some constant (almost the same one as in the
lemma). Let S = {xi }N i=1 be an ε-separated set in Σp (X). We may assume
that the points xi are representable by shortest paths γi emanating from p.
e i (t)pγj (t) → ∡(γi , γj ) = |xi xj | as t → 0.
By the definition of angle, ∡γ
Therefore one can find such a small r > 0 that ∡γ e i (t)pγj (t) > 1 |xi xj | > ε/2,
2
and hence |γi γj | > 2r sin(ε/4) > rε/4, for all i, j (i 6= j). In other words, we
obtain an (rε/4)-separated set in an r-ball. Applying Lemma 10.9.2 yields
a desired bound for the number of points in the set. Since ε is arbitrary, it
follows that Σp (X) is compact. ¤

Recall that the tangent cone Kp (X) over the space of directions Σp (X)
consists of the origin o and points represented by (ξ, r) where x ∈ Σp (X)
and r > 0. The distance from (ξ, r) to the origin equals r, and the distance
e 0 a1 oa2 = ∡(ξ1 , ξ2 ).
from a1 = (ξ1 , r1 ) and a2 = (ξ2 , r2 ) is defined so that ∡
Theorem 10.9.3. Let X be a finite-dimensional Alexandrov space and
p ∈ X. Then the Gromov–Hausdorff tangent cone of X at p exists and
is isometric to Kp (X).
392 10. Spaces of Curvature Bounded Below

Proof. Let B denote the unit ball centered at the origin o of the cone
Kp (X). We have to show that the rescaled balls 1r Br (p) converge to B in
the Gromov–Hausdorff metric as r → 0. Fix an ε > 0. By one of the
criteria of Gromov–Hausdorff convergence (namely, by Proposition 7.4.11),
1
¯ ε-nets {xi }¯ in B and {yi } in r Br (p), where r is small
it suffices to find finite
enough, such that ¯|xi xj | − |yi yj |¯ < ε for all i, j.
Pick a finite ε-net {xi } in B whose elements are representable by shortest
paths. Let γi denote a shortest path representing xi , parameterized with the
constant speed |oxi |. Assume that r is so small that the point pi = γi (r) is
defined for every i, and
¯ ¯
¯ |pi pj | ¯
¯ ¯
¯ r − |xi xj |¯ < ε

for all i, j. (This inequality for a small r is an immediate consequence of the

definition of the tangent cone.) Now let {yi } be the points corresponding
1
¯ {pi } in the¯ rescaled ball r Br (X). The above inequality means that
to
¯|yi yj | − |xi xj |¯ < ε for all i, j.
It remains to show that {yi } is an ε-net in 1r Br (X), or, equivalently, in
the unit ball of the rescaled space X ′ = 1r X. To do this, recall that the
“logarithm” map logp : X ′ → Kp (X ′ ) = Kp (X) from Lemma 10.6.11 is
noncontracting if X is of curvature ≥ 0. Since {xi } is an ε-net in the unit
ball of Kp (X), for every point y in the unit ball of X ′ there is an i such
that | logp (y)xi | < ε. Since logp is noncontracting and we may assume that
xi = logp (yi ), it follows that |yyi | ≤ ε. Therefore {yi } is an ε-net.
If the curvature bound k is nonzero, the same argument works with
minor corrections. Since the rescaled space X ′ = 1r X has curvature ≥ kr2
(which goes to zero as r → 0), for a sufficiently small r the logarithm map
satisfies | logp (x) logp (y)| ≥ 21 |xy| within the unit ball. Then it follows that
{yi } is a (2ε)-net. ¤

Remark 10.9.4. Note that the above argument not only proves that Kp (X)
is the Gromov–Hausdorff limit of rescaled balls, but also gives explicit small-
distortion maps from the balls to the cone. Namely the distortion of the
“logarithm” map goes to zero with the radius of the ball. Reformulating
this in terms of angles instead of distances, one gets the following useful
fact: for every ε > 0 there is an r > 0 such that |∡xpy e − ∡xpy| < ε
whenever |px| and |py| are less than r.

Corollary 10.9.5. Let X be an n-dimensional Alexandrov space and

p ∈ X. Then the tangent cone Kp (X) is an n-dimensional Alexandrov
space of nonnegative curvature.
10.9. Spaces of Directions and Tangent Cones 393

Proof. Kp (X) is a length space and has nonnegative curvature as a Gromov–

Hausdorff limit (by Proposition 10.7.1). More precisely, if X has curvature
≥ k, then λ−1 X has curvature ≥ λ2 k, and this curvature bound goes to 0
as λ → 0. Then the Gromov–Hausdorff limit Kp (X) has curvature ≥ −ε for
every ε > 0, and hence it is nonnegatively curved.
It remains to prove that dimH (Kp (X)) = dimH (X). The inequality
dimH (Kp (X)) ≥ dimH (X) follows from the fact that there is a noncon-
tracting map from X to Kp (X), namely, the “logarithm” map (see Lemma
10.6.11). The inverse inequality follows from Theorem 10.9.3 and Corollary
10.8.25 about the dimension of a Gromov–Hausdorff limit. ¤
Corollary 10.9.6. Let X be an n-dimensional Alexandrov space and p ∈ X.
Then the space of directions Σp (X) is an (n − 1)-dimensional Alexandrov
space of curvature ≥ 1, provided that n ≥ 2. If n = 1, then Σp (X) consists
of one or two points.

Proof. Since the cone Kp (X) is nonnegatively curved, Theorem 10.2.3

implies that Σp (X) is a length space of curvature ≥ 1 or a pair of points.
(The additional requirement that no triangle in Σp (X) has perimeter greater
than 2π follows from Exercise 10.1.3.) Since dimH Kp (X) = n, one has
dimH Σp (X) = n − 1. This seemingly obvious implication is left as an
exercise to the reader. ¤

10.9.2. Regular points.

Definition 10.9.7. Let X be an n-dimensional Alexandrov space. A point
p ∈ X is said to be regular if the space of directions Σp (X) is isometric to
the standard sphere S n−1 .
Equivalently, p is regular if the tangent cone Kp (X) is isometric to Rn .
Example 10.9.8. If X is a convex polyhedral surface in R3 , then all points
but vertices are regular. This is because the intrinsic metric of a polyhedral
surface is flat everywhere except at vertices.
Example 10.9.9. If X is the doubly-covered disc (that is, two copies of the
disc glued along the boundary), then all points of X are regular.
The following lemma is an interesting global fact about Alexandrov
spaces of curvature ≥ 1. It tells that one can recognize the standard
sphere just by pointing out a finite collection of points with certain distances
between them.
Lemma 10.9.10. Let Y be an Alexandrov space of curvature ≥ 1 and
of dimension no greater than n − 1 (n ≥ 2). Suppose that there are
n pairs of points {(xi , yi )}ni=1 in Y such that |xi yi | = π for all i, and
|xi xj | = |xi yj | = |yi yj | = π2 if i 6= j. Then Y is isometric to S n−1 .
394 10. Spaces of Curvature Bounded Below

We refer to a collection {(xi , yi )} satisfying the assumptions of the lemma

as an orthogonal collection. We mainly use this term in the case when Y is
a space of directions. In this case, the lemma tells us that a point in an n-
dimensional Alexandrov space is regular if and only if there is an orthogonal
collection of n pairs of directions at this point.

Proof. We prove the lemma via an induction on n. Since diam(Y ) = π,

Y is a suspension (spherical cone) over an Alexandrov space Z of curvature
≥ 1 (see Exercise 10.4.3). Moreover one can take the pair (xn , yn ) as the
poles of the suspension; then
π
Z = {x ∈ Y : |xxn | = |xyn | = }.
2
In particular, Z contains the points xi and yi for all i = 1, . . . , n − 1. Thus Z
is isometric to S n−2 by the inductive assumption and therefore Y is isometric
to S n−1 . ¤
Exercise 10.9.11. Let Y be an Alexandrov space of curvature ≥ 1 con-
taining n pairs of points {(xi , yi )} such that |xi yi | = π for all i and the
determinant of the n × n matrix (cos |xi xj |) is nonzero. Prove that Y con-
tains a subset isometric to the standard sphere S n−1 , and that Y is isometric
to S n−1 if dimH (Y ) = n − 1.
Corollary 10.9.12. A point p in an n-dimensional Alexandrov space is
regular if and only if p is (n, ε)-strained for every ε > 0.

Proof. Suppose that p is regular. Then there exist n pairs of directions

{(ξi , ηi )}∞
i=1 in Σp (X) with ∡(ξi , ηi ) = π for all i and ∡(ξi , ξj ) = ∡(ξi , ηj ) =
∡(ηi , ηj ) = π/2 for i 6= j. Consider shortest paths whose directions are close
to ξi and ηi and take points close to p on these shortest path. These points
form an (n, ε)-strainer (see the third part of Proposition 10.8.12).
Now suppose that p is (n, ε)-strained for every ε > 0. Then for every
ε > 0 the space of directions Σp (X) contains n pairs of points (ξi , ηi ) such
that |ξi ηi | > π−ε, and other distances between these points are (11ε)-close to
π
2 (see the second part of Proposition 10.8.12). Since Σp (X) is compact, one
can extract a converging sequence of such pairs corresponding to a sequence
εi → 0. The limit is a collection of points in Σp satisfying the conditions of
Lemma 10.9.10; hence Σp (X) is isometric to S n−1 . ¤
Corollary 10.9.13. The set of regular points in an Alexandrov space is
dense and moreover is an intersection of a countable collection of open dense
sets.

Proof. Let X be an n-dimensional Alexandrov space. For every positive

integer i consider the set Si of (n, 1/i)-strained points. By Corollary 10.8.24,
10.9. Spaces of Directions and Tangent Cones 395

Si is an open dense set. By Corollary 10.9.12, the set of regular points is

the intersection of all Si . Now Baire’s Theorem 1.5.13 implies that this set
is dense in X. ¤

Our next goal is to prove that a small neighborhood of a regular point is

Lipschitz-close to a Euclidean region. We begin with a refinement of Lemma
10.9.10. It tells us that there is a certain stability in Lemma 10.9.10; namely,
if a space of curvature ≥ 1 contains an “almost orthogonal” collection of
directions, then it is close to the sphere in the Gromov–Hausdorff topology.
This proof illustrates the important technique of using Gromov’s Com-
pactness Theorem. This technique allows us to derive “stability” statements
from “rigidity” statements almost automatically in many cases,; see for ex-
ample Exercise 10.9.15 below.

Lemma 10.9.14. For every integer n ≥ 2 and every ε > 0 there is δ > 0
such that the following holds.
Let Y be an Alexandrov space of curvature ≥ 1 and of dimension no
greater than n − 1. Suppose that Y contains n pairs of points {(xi , yi )}ni=1
such that |xi yi | > π−δ for all i, and the distances |xi xj |, |xi yj |, |yi yj | (i 6= j)
are δ-close to π/2. Then dGH (Y, S n−1 ) < ε.

Proof. Suppose the contrary. Then there is a sequence of spaces {Ym }∞m=1
satisfying the assumptions of the lemma for δ = δm where δm → 0 and
such that dGH (Ym , S n−1 ) ≥ ε. Due to the Compactness Theorem 10.7.2
we may assume that {Ym } converges in the Gromov–Hausdorff topology to
an Alexandrov space Y of curvature ≥ 1 and of dimension no greater than
n − 1. We are going to show that Y is isometric to S n−1 , contrary to our
assumption.
Recall (Theorem 7.3.25) that there is a correspondence between Ym
and Y with distortion no greater than 2dGH (Ym , Y ). Thus an “almost
orthogonal” collection of points in Ym corresponds to an “almost orthogonal”
collection in Y (with δm replaced by δm + dGH (Ym , Y ) which goes to zero
anyway). Extracting a converging subsequence of these collections in Y ,
one obtains an orthogonal collection. By Lemma 10.9.10, it follows that
Y ≃ S n−1 . ¤

Here are a few exercises on similar use of the Compactness Theorem.

Exercise 10.9.15. Prove that for every n ∈ N and ε > 0 there is a δ > 0
such that for any n-dimensional Alexandrov space X of curvature ≥ 1 the
following holds.
396 10. Spaces of Curvature Bounded Below

1. If diam(X) > π − δ, then X is ε-close in the Gromov–Hausdorff

metric to a spherical cone over some Alexandrov space of curvature ≥ 1 and
dimension at most n − 1. (Compare with Exercise 10.4.3.)
2. If the radius of X (see Exercise 10.4.5) is greater than π − δ, then X
is ε-close to S n , i.e., dGH (X, S n ) < ε.
3. If µn (X) > µn (S n ) − δ, then X is ε-close to S n .
Hint: Use the Bishop inequality to reduce this statement to the previous
one.
Theorem 10.9.16. For every integer n ≥ 1 and every ε > 0 there is a δ > 0
such that every (n, δ)-strained point in any n-dimensional Alexandrov space
has a neighborhood which is ε-close with respect to the Lipschitz distance (cf.
Section 7.2) to an open region in Rn .
We give two slightly different proofs demonstrating different techniques.

Proof. We will show that, in a sufficiently small neighborhood U , the map

f : U → Rn of distance coordinates associated with an (n, δ)-strainer
{(ai , bi )}, and the inverse map f −1 have Lipschitz constants not greater
than 1 + ε.
We may assume that {(ai , bi )} is an (n, δ)-strainer for every point of U .
We want to show that for every two points x, y ∈ U the ratio |f (x)f (y)|/|xy|
is close to 1. Considering a comparison triangle for △ai xy, one sees that
e i xy + o(|xy|).
fi (y) − fi (x) = |ai y| − |ai x| = |xy| · cos ∡a
Then
X X
|f (y) − f (x)|2 = |fi (y) − fi (x)|2 = |xy|2 · e i xy + o(|xy|2 ).
cos2 ∡a
To prove that the ratio |f (x)−f (y)|/|xy| is close to 1, it suffices to show that
P e i xy is close to 1. Fix shortest paths connecting x to y, ai and bi ,
cos2 ∡a
and denote by ξ, ξi and ηi the respective elements of the space of directions
Σx (X). By the last statement of Lemma 10.8.13, the angles |ξξi | = ∡ai xy
between these shortest paths are almost equal to the comparison angles
e i xy. Therefore it is sufficient to prove that P cos2 |ξξi | is close to 1.
∡a
Note that this sum equals 1 if the space of directions Σx (X) is the
standard sphere S n−1 and {(ξi , ηi )} is an orthogonal collection (in the sense
of Lemma 10.9.10). What we know is that {(ξi , ηi )} is “almost orthogonal”
(up to a term of order δ), and hence by Lemma 10.9.14 the space of directions
Σx (X) is close to S n−1 in the Gromov–Hausdorff metric. Then by Theorem
7.3.25 there is a correspondence with a small distortion between Σx (X) and
S n−1 . Let ξ, ξ i and η i be points in the sphere corresponding to ξ, ξi and ηi .
Then X X
cos2 |ξξi | ≈ cos2 |ξξ i | ≈ 1
10.9. Spaces of Directions and Tangent Cones 397

(the second approximate equality follows from the fact that in the sphere
there is a real orthogonal collection near {(ξ i , η i )}. ¤

Another proof. One can prove the theorem using regular points instead of
Lemma 10.9.14.
P Namely, the above argument shows that the approximate
equality cos2 |ξξi | ≈ 1 holds if Σx (X) is isometric to S n−1 , i.e., x is a
regular point. Then the inequalities (1 + ε)−1 |xy| ≤ |f (x)f (y)| ≤ (1 + ε)|xy|
follow for a regular point x. Since regular points are dense, these inequalities
follow for all x, y. ¤

Note that the second proof makes it possible to write an explicit bound
for δ in terms of ε, while the first one only proves that a δ exists (see how
Lemma 10.9.14 was proved).
Sometimes the following reformulation of Theorem 10.9.16 is useful.

Corollary 10.9.17. For every integer n ≥ 1 and every δ > 0 there is a

ε > 0 such that the following holds. If X is an n-dimensional Alexandrov
space and p ∈ X is a point such that dGH (Σp (X), S n−1 ) < ε, then p has
a neighborhood which is δ-close with respect to the Lipschitz distance to an
open region in Rn .

Remark 10.9.18. Points p ∈ X with dGH (Σp (X), S n−1 ) < ε are referred to
as ε-regular points. The set Xε of ε-regular points features some interesting
properties. Obviously every point of Xε is (n, 2ε)-strained and hence has a
Euclidean neighborhood. For the same reason, the interior of Xε is dense in
X (see Corollary 10.8.23).
Furthermore, though it sounds a little surprising, the set Xε is convex
in the sense that a shortest path connecting two points of Xε does not leave
Xε . This follows from the fact that the tangent cone is constant along any
shortest path except its endpoints; in other words, if x and y belong to the
interior of some shortest path, then Kx (X) is isometric to Ky (X) [Pet].

Corollary 10.9.19. Let X be an n-dimensional Alexandrov space and

p ∈ X be a regular point. Then p has neighborhoods arbitrarily close to
regions in Rn in the sense of Lipschitz distance.

Exercise 10.9.20. Let n ∈ N , k ∈ R and r > 0 be fixed. Prove that there is

a δ > 0 such that the following holds. If X is an n-dimensional Alexandrov
space of curvature ≥ k, p ∈ X and µn (Br (p)) > Vk,rn − δ, where V n is the
k,r
volume of the r-ball in the n-dimensional space form of curvature k, then p
has a neighborhood homeomorphic to Rn .
n − δ,
In particular, if the volume of every r-ball in X is no less than Vk,r
then X is a manifold.
398 10. Spaces of Curvature Bounded Below

Hint: Using a modified argument from the proof of the Bishop inequality,
show that the (n−1)-dimensional measure of Σp (X) is close to that of S n−1 .
Then apply the last part of Exercise 10.9.15.

10.10. Further Information

10.10.1. Stratification. One says that a collection {Xi }N i=1 of subsets of
a topological space X forms a (finite) stratification of X into topological
manifolds if
S
(1) The sets Xi are disjoint, and Ni=1 Xi = X.
(2) Every set Xi is a topological manifold (without boundary).
(3) dim X1 > dim X2 > · · · > dim XN .
S
(4) For every k = 1, . . . , N , the set Xk+ = N i=k Xi is closed in X.
Equivalently, the closure of Xk is contained in Xk+ for all k.
The sets Xi are called strata. Each stratum is not necessarily connected.

Theorem 10.10.1. Every finite-dimensional Alexandrov space X admits a

stratification into topological manifolds.
Moreover, there is a canonical stratification such that every component
C of every stratum is topologically homogeneous in the following sense:
every two points p, q ∈ C have neighborhoods in X which are pointed
homeomorphic to each other (pointed means that the homeomorphism maps
p to q).

This stratification can be described as follows: the stratum X1 is the set

of points possessing a Euclidean neighborhood in X; then X2 is the set of
points in X \ X1 possessing a Euclidean neighborhood of maximal possible
dimension, and so on. The matter is that this procedure exhausts the whole
space. For example, a disc is stratified into its interior and its boundary;
the cone over RP2 is stratified into the complement of the origin (which is
a three-dimensional flat space) and the origin itself.
Corollary 10.9.6 allows us to give an inductive definition of the boundary
of an Alexandrov space. First, the boundary of a one-dimensional space is
its topological boundary (recall that all one-dimensional Alexandrov spaces
are segments, rays, lines and circles). Then, assuming that the notion of
boundary is already defined in dimensions 1, . . . , n − 1, define the boundary
of an n-dimensional Alexandrov space X as the set of points p ∈ X such
that the space of directions Σp (X) has nonempty boundary.
Using the above stratification theorem, it is possible to prove that the
boundary of an Alexandrov space possesses various nice properties; for
10.10. Further Information 399

example, it is always a closed subset of codimension one. In fact, the

boundary coincides with the closure of the stratum of codimension one.
It can be proved that there is no stratum of codimension two. So if an
Alexandrov space has no boundary the set of topologically singular points
is very thin: it has codimension at least three.
The following theorem is a local counterpart of the Theorem 10.10.1.
Theorem 10.10.2. Every point p of a finite-dimensional Alexandrov space
has a neighborhood pointed homeomorphic to the tangent cone at p.

Though Theorem 10.10.2 looks much simpler than Theorem 10.10.1,

there is no known “simple” proof for it. Both theorems can be proved
simultaneously by induction in dimension, and the proof of the second
theorem in dimension n involves the first theorem in dimensions less than n.
The proof of both theorems (and many other strong results as well)
is contained in the manuscript “A.D.Alexandrov’s spaces with curvatures
bounded from below, II” by G. Perelman. Unfortunately this paper is not
published.
Actually Theorem 10.10.1 can be formulated in a stronger form. Namely,
there is a canonical stratification into so-called primitive extremal subsets.
This is proved in [Per1].
We do not define here what is an extremal set. Let us only mention that
the boundary of an Alexandrov space is always an extremal set. Another
example of an extremal set is a point whose space of directions has diameter
less than π.
The proofs of the above mentioned theorems are complicated and are
based on study of analytical properties of distance functions and some
nonelementary facts from topology. These topics are beyond the scope of
this book.
In dimension 2, the results about topological structure of Alexandrov
spaces take very simple form:
Corollary 10.10.3. Every two-dimensional Alexandrov space is a topolog-
ical manifold, possibly with boundary.
Exercise 10.10.4. 1. Prove Corollary 10.10.3 directly (without the strati-
fication theorem).
2. Let X be a three-dimensional Alexandrov space having no boundary.
Prove that topologically singular points of X are isolated and, for every such
point p, the space of directions Σp (X) is homeomorphic to RP2 .

The reader can observe that a classification of topological singularities

in dimension n is roughly equivalent to a classification of topological types
400 10. Spaces of Curvature Bounded Below

of Alexandrov spaces of curvature ≥ 1 in dimension n − 1. This appears

easy for n ≤ 3 (see the exercise above) but in general to obtain such a
classification seems hopeless.
Generalizing Theorem 10.10.2 above, a very strong stability theorem is
proved in the same manuscript by G. Perelman.
Theorem 10.10.5 (Stability theorem). For every k ∈ R and for every
compact Alexandrov space X n of curvature ≥ k, there exists an ε > 0 such
that every compact Alexandrov space Y n of curvature ≥ k and such that
dGH (X n , Y n ) < ε
is homeomorphic to X n (here dGH is the Gromov–Hausdorff distance and
upper indices denote the dimension of spaces).

In particular, the theorem implies that an Alexandrov space having

topological singularities cannot be approximated by Riemannian manifolds
of the same dimension and with the same curvature bound.

10.10.2. Convexity arguments. Let X be a locally compact length

space. Recall that a function f : X → R is said to be convex if its re-
striction to every geodesic is a convex function in one variable.
Theorem 10.10.6. Let X be a finite-dimensional Alexandrov space of
nonnegative curvature with nonempty boundary. Then the distance function
d(·, ∂X) is convex. Moreover it is strictly convex if X has curvature ≥ k > 0.

Clearly a convex subset of an Alexandrov space is an Alexandrov space

with the same curvature bound. In particular, such subsets have well-defined
boundary. Jointly with the theorem above, this allows one to extend the
proof of the famous Soul Theorem by J. Cheeger and D. Gromoll [CG1] to
Alexandrov spaces.
Theorem 10.10.7. Let X n be a noncompact finite-dimensional Alexandrov
space of nonnegative curvature. Then there is a convex compact subset
S ⊂ X n without boundary and such that S is a deformation retract of X n .

Note that if X is a Riemannian manifold, a convex closed subset without

boundary is a totally geodesic submanifold.
Remark 10.10.8. In the case of Riemannian manifolds, the Soul Theorem
can be formulated in a stronger form: X is homeomorphic (in fact, diffeo-
morphic) to the total space of the normal bundle of S in X. For Alexandrov
spaces a similar strengthening is false as the following example shows:
Example 10.10.9. (G. Perelman) The quotient of the complex space Ck
under the equivalence x ∼ λx, λ ∈ C, |λ| = 1, is the cone over the complex
projective space CPk ; denote this quotient by K0 (CPk ).
10.10. Further Information 401

The natural projection (z1 , . . . , z3 ) → (z1 , z2 ) of C3 onto C2 induces the

projection π : K0 (CP2 ) → K0 (CP1 ). Let us equip CPk with the canonical
metric having sectional curvatures between 1 and 4, and equip the cones with
corresponding cone metrics. Let B 0 (1) be the unit ball in CP1 centered at the
origin. Set X 5 = π −1 (B 0 (1)). X 5 is a convex subset of CP2 , and therefore it
is a complete noncompact Alexandrov space of nonnegative curvature and
5
with nonempty boundary. It is possible to prove that the doubling X of X 5
is a space of the same type but without boundary. Since 0 is a topologically
5
singular point of X 5 , the space X has two singular points. One can show
5
that the doubling S of B 0 (1) is a soul of X . Since S is homeomorphic
to a sphere S 3 , a fiber bundle over S is a manifold and hence cannot be
5
homeomorphic to X .

10.10.3. Hausdorff measure. Recall that the m-dimensional Hausdorff

measure µm (X) of m-dimensional compact Alexandrov space X is always
finite and positive. And as we already mentioned after Theorem 10.8.5, the
set of nonregular points has zero measure; in particular, the set of topological
singularities has zero measure. Moreover, for every sufficiently small ε > 0
1
(say, ε < 100m ) one has µm (X \ Xε ) = 0, where Xε is the set of ε-regular
points.
Let {Xi } be a sequence of compact metric spaces, Xi −→ GH X as i → ∞.
Then one can choose a sequence of numbers νi such that νi → 0 as i → ∞
and Xi is a νi -approximation for X. By Corollary 7.3.28 there exist maps
fi : X → Xi such that dis fi < 2νi and fi (X) is 2νi -net in Xi . One says that
measures ωi defined on Xi weakly converge to a measure ω on X if for every
closed set E ∈ X such that ω(∂E) = 0 one has ωi (U2νi (fi−1 (E))) → ω(E)
as i → ∞.
Theorem 10.10.10. Let m-dimensional compact Alexandrov spaces X and
Xi , i = 1, 2, . . . , have the same curvature bound and Xi −→
GH X as i → ∞.
Then m-dimensional Hausdorff measures of Xi converge weakly to that of
X (for any choice of fi and νi in the definition of weak convergence).

This theorem has an obvious

Corollary 10.10.11. Let Xi be an m-dimensional compact Alexandrov
space with the same curvature bound and converging to a compact space X.
Then µm (Xi ) → 0 if and only if dimH X < m.

10.10.4. Collapse. The situation described in the previous corollary is

called a collapse. More precisely, one says that a sequence {Xi } of m-
dimensional Alexandrov spaces with the same curvature bound collapses to
a compact space Y if Xi −→
GH Y and µm Xi → 0 as i → ∞ (or, equivalently,
dim Y < m).
402 10. Spaces of Curvature Bounded Below

It is clear that in this case Y is an Alexandrov space with the same

curvature bound. Let us restrict ourselves to a partial case when the spaces
Xi are manifolds homeomorphic to one another or, equivalently, that we
have a sequence of metrics on a fixed manifold.
Then, if the limit space Y is a Riemannian manifold or at least X
has only weak singularities, then Xi (for all sufficiently large i) admits a
structure of locally trivial fibration over Y (see [Yam]).
If a collapse happens for a sequence of manifolds having in addition
a common upper curvature bound, then the fibers of this fibration have a
very special topological type; namely, they are so-called nil-manifolds (see
[CFG]).
Another interesting question concerning collapse is whether a given
Alexandrov space can be approximated by Riemannian manifolds (with some
fixed lower curvature bound). Recall that due to Stability Theorem 10.10.5,
an Alexandrov space X having a topological singularity cannot be approx-
imated by Riemannian manifolds of the same dimension. This does not
exclude approximation by higher dimensional manifolds, i.e., with collapse.
There is an example of an Alexandrov space of nonnegative curvature which
can not be approximated (even with collapse) by nonnegatively curved Rie-
mannian manifolds; this however does not exclude approximation by man-
ifolds having a bit smaller curvature bounds. No affirmative results of this
sort are known, and the problem looks very hard.

10.10.5. Quasigeodesics. Geodesic lines in a smooth surface or a Rie-

mannian manifold can be characterized in two ways: they are locally short-
est paths, and they are straightest paths. The latter means that geodes-
ics are curves of zero geodesic curvature. Geodesics in a length space are
defined as locally shortest paths. Unfortunately geodesics in Alexandrov
spaces lack some useful properties of Riemannian geodesics. First, there
may be points such that no geodesic passes through them (for instance,
consider the apex of a sharp 2-cone). Even at a regular point, there may be
directions not representable by shortest paths (emanating from the point).
And not every geodesic can be extended beyond its endpoint even if this
point is regular. These circumstances stimulate consideration of a wider
class of curves, so-called quasigeodesics. Roughly speaking, quasigeodesics
are locally straightest paths in Alexandrov spaces.
To give a formal definition, recall the study of distance functions in
Subsection 4.1.2. Given a point p and a geodesic γ in a length space X, we
considered the distance function g(t) = d(γ(t), p) and compared it with the
similar distance function gk (with the same values at endpoints) defined by
a point and a geodesic in the k-plane. The definition of a space of curvature
10.10. Further Information 403

≥ k says that for every sufficiently short segment of γ and every point p
sufficiently close to it, the function g must be no less concave than gk .
The same comparison property may hold for curves γ that are not
geodesics. Curves possessing this property are called k-quasigeodesics. A
quasigeodesic is a k-quasigeodesic in an Alexandrov space of curvature ≥ k.
Quasigeodesics have many nice properties; in particular all the above
mentioned missing properties of geodesics do hold for quasigeodesics.3
In addition, the class of quasigeodesics is closed under noncollapsing
Gromov–Hausdorff convergence.
Let us mention one more property. If X is a space of curvature ≥ 1,
two points a, b ∈ X are called antipodes if |ax| + |xb| ≤ π for all x ∈ X. If
two quasigeodesics have a common endpoint p and their directions at p are
antipodes, then the concatenation of these paths (taken with the opposite
orientation) is a quasigeodesic too. This allows one to build quasigeodesics
“from pieces”.
A geodesic in an extremal set (in particular, a geodesic in the boundary)
is a quasigeodesic in the ambient Alexandrov space [PP].
There are several important applications of quasigeodesics. One of them
is the proof of the following gluing theorem (see [Pet1] and the preprint
mentioned above).
Theorem 10.10.12. Let X and Y be Alexandrov spaces of the same dimen-
sions and of curvature ≥ k. Suppose that their boundaries are isometric.
Then the space Z obtained by gluing X and Y along an isometry of their
boundaries is a space of curvature ≥ k.

10.10.6. Analysis on Alexandrov spaces. Recall that almost every

point of an Alexandrov space X has a bi-Lipschitz Euclidean neighborhood.
So one can consider “atlases” of bi-Lipschitz charts covering X up to a set of
zero measure. In such a chart, a Riemannian metric tensor can be defined at
almost every point (using the fact that a Lipschitz function is differentiable
almost everywhere, and that the tangent cone of X is Euclidean at almost
every point). The coefficients of such a metric tensor are a priori only
measurable functions, and their relation to metric properties of X are not
clear. One may ask if it is possible to choose local coordinates so that (some
of) the machinery of differential geometry works. For example, one may want
to define Christoffel symbols (as functions, or measures, or whatever), use
Riemannian formulas for first and second variation, and so on. To make this
analysis work, one has to find coordinates in which the metric coefficients

3See the preprint by G. Perelman and A. Petrunin, Quasigeodesics and gradient curves in
Alexandrov spaces.
404 10. Spaces of Curvature Bounded Below

are as regular as possible (at least, differentiable almost everywhere), and/or

utilize some generalized interpretation of differential formulas.
There are some affirmative results in this direction. Let us mention [OS]
and [Ot] among already published papers; several more papers are to appear
soon. (This field is far from being complete and is changing rapidly.)

10.10.7. Curvature bounded from both sides. One may wonder what
can be said about a space of curvature bounded both above and below. The
answer is simple though it required hard work to prove it: loosely speaking,
such a space is a Riemannian manifold whose metric tensor has slightly
lower smoothness than usually considered in Riemannian geometry. More
precisely, the following theorem holds.
Theorem 10.10.13 (I. Nikolaev, see [BN]). Let (X, d) be a length space
with both side upper and lower curvature bounds (of curvature ≥ k and ≤ K).
If (X, d) has no boundary, then X is a manifold and possesses a C 3 -smooth
atlas such that in its charts the metric d can be defined by a Riemannian
metric tensor whose coefficients gij are in the class Wp2 for every p > 1.
If X has a boundary, then the interior of X is an almost Riemannian
manifold in the same sense as above; however the boundary may be non-
smooth.
If k = K, then (X, d) is usual space of constant curvature.

Recall that Wp2 means the space of functions whose second Sobolev
derivatives belong to Lp . (In particular gij ∈ C 1,α for all α < 1.)
The theorem allows us to define formal Christoffel symbols and curva-
ture tensor via derivatives of gij . These formal values are defined almost
everywhere; nevertheless they have a reasonable geometric meaning. For
instance, parallel transport defined geometrically coincides (for almost all
paths) with one given by Christoffel symbols. And formal sectional curva-
tures have geometric meaning as well.
Also I. Nikolaev proved that such a space can be approximated (in
the sense of uniform convergence) by Riemannian manifolds of the same
dimension and with curvature bounds converging to that of (X, d).
Bibliography

[AB] S. Alexander, R. Bishop, The Hadamard-Cartan theorem in locally convex metric

spaces, Enseign. Math. (2) 36 (1990), no. 3-4, 306-320
[AB1] S. Alexander, R. Bishop, Warped products of Hadamard spaces, Manuscripta Math.,
96 (1998), 487-505
[Al] A. D. Alexandrov, Intrinsic Geometry of Convex Surfaces (in Russian). German
translation: Die innere Geometrie der konvexen Flächen, Akad. Verl., Berlin, 1955.
[Bab] I.K.Babenko, Forte souplesse intersystolique de variétés fermées et de polyèdres,
Annales de l’Ecole Norm. Super. (to appear).
[Bal] W. Ballmann, Lectures on spaces of nonpositive curvature, DMV Seminar 25,
Birkhäuser (1995), viii+112.
[Ban] V. Bangert Minimal geodesics, Erg. Theory and Dyn. Syst. 10 (1990), 263–286.
[BC] R. Bishop and R. Crittenden, Geometry of Manifolds, Academic Press, New York –
London, 1964.
[BGS] W. Ballmann, M. Gromov, and V. Schroeder, Manifolds of nonpositive curvature,
Progress in Mathematics 61, Birkhäuser, Boston (1985), vi+263.
[BH] M. R. Bridson and A. Haffliger Metric spaces of non-positive curvature, in Ser. A
Series of Comprehensive Stadies in Mathematics, vol. 319, Springer-Verlag, Berlin,
1999.
[BN] V. Berestovskii and I. Nikolaev, Multidimensional generalized Riemannian spaces,
in Geometry IV. Non-regular Riemannian geometry. Encyclopaedia of Mathematical
Sciences, Springer-Verlag, Berlin, 1993, 165–244.
[BM] M. Bonk, Quasi-geodesic segments and Gromov hyperbolic spaces, Geom. Dedicata
62 (1996), no. 3, 281–298.
[BP] R. Benedetti and C. Petronio, Lectures on Hyperbolic Geometry, Springer, 1992
[BR] A. Belläche and J.-J.Risler (editors), Sub-Riemannian Geometry, Progress in Math-
ematics, vol. 144, Birkhäuser, 1996.
[Bur] D. Burago, Periodic metrics, Advances in Soviet Math. 9 (1992), 205-210.
[BI] D. Burago and S. Ivanov, On asymptotic volume of tori, Geom. Func. Anal. 5 (1995),
800–808

405
406 Bibliography

[BB] Yu. Burago, S. Buyalo, Metrics of upper bounded curvature on 2-polyhedra. II, St.
Petersburg Math. J. 10 (1999), 619–650.
[BGP] Yu. Burago, M. Gromov, G. Perelman, A. D. Alexandrov spaces with curvatures
bounded below, Russian Math. Surveys 47 (1992), 1–58.
[BZ] Yu. Burago and V. Zalgaller, Geometric Inequalities, A Series of Comprehensive
Studies in Mathematics 285, Springer-Verlag, Berlin, 1988.
[Car] M. P. do Carmo, Riemannian Geometry, Birkhäuser, Boston, 1992.
[CE] J. Cheeger and D. Ebin, Comparison Theorems in Riemannian Geometry, North
Holland, New York, 1975.
[CFG] J. Cheeger, K. Fukaya, and M. Gromov Nilpotent structures and invariant metrics
on collapsed manifolds, J. Amer. Math. Soc. 5 , no. 2,(1992) 327–372.
[CG1] J. Cheeger and D. Gromoll, On the structure of complete manifolds of non-negative
curvature, Ann. of Math. 96 (1972), 413-443.
[CG] J. Cheeger and D. Gromoll, The splitting theorem for manifolds of nonnegative Ricci
curvature, J.Differential Geometry 6 (1971-72), 119-128.
[Cha] I. Chavel, Riemannian Geometry — a modern introduction, Cambridge Univ. Press,
Cambridge, 1993.
[Ch] , C. H. Chen Warped products of metric spaces of curvature bounded from above,
Trans. Amer. Math. Soc. 351 (1999),4727-4740.
[E] P. B. Eberlein, Geometry of Nonpositively Curved Manifolds, The University of
Chicago Press, Chicago and London, 1996.
[Fe] H. Federer, Geometric Measure Theory, Springer-Verlag, Berlin, 1969.
[GHL] S. Gallot, D. Hulin, and J. Lafontaine, Riemannian geometry , Springer-Verleg,
Berlin, 1990, xiv+284.
[Gro1] M. Gromov, Filling Riemannian manifolds, J. Differential Geometry 18 (1983),
1–147.
[Gro2] M. Gromov, Systols and intersystolic Inequalities, Actes de la table rande de
géométrie differentielle en l’honeur de Marcel Berger. Collection SMF 1 (1996), 291–
362.
[Gro3] M. Gromov, Metric Structures for Riemannian and Non-Riemannian Spaces,
Progress in Mathematics 152, Birkhäuser, Boston, 1999.
[Grove] K. Grove, Metric differential geometry, Lecture Notes in Math. 1263 (1987), 171–
227.
[GP] K. Grove and P. Petersen (Editors), Comparison geometry, Papers from the Special
Year in Differential Geometry held in Berkeley, CA. 1993–94. Research Institute
Publications, 30 (1997), Cambridge University Press, Cambridge, x+262 pp.
[HM] S. Hartman and J. Mikusiński,, The theory of Lebesgue measure and integration,
International series of monographs on Pure and Applied Mathematics, vol. 15, 1961.
[HS] J. Hass and P. Scott, Bounded 3-manifolds admits negatively curved metric with
concave boundary, J. Differential Geom. 40 (1994), no. 3, 449-459
[J] J. Jost, Nonpositive curvature: Geometric and analytic Aspects, Lectures in Mathe-
matics ETH Birkhäuser, 1997.
[Ka] H. Karcher, Riemannian comparison constructions, in Global Differential Geometry
(S. S. Chern ed.), MAA Studies in Mathematics 27 (1989), 170–222.
[Kl] B. Kleiner, The local structure of length spaces with curvature bounded above, Math.
Z. 231 (1999), no. 3, 409–456.
Bibliography 407

[KS] M. Katz and A. Suciu, Systolic freedom of loopspace, (2000), 1-13, preprint.
[Mas] W. S. Massey, Algebraic topology: an introduction, Graduate Texts in Mathematics,
vil. 56. Springer, New York, 1987
[Mi] J. Milnor, Topology from the differentiable viewpoint, Univ. Press of Virginia, 1965.
[Mo] F. Morgan, Geometric measure theory: a beginner’s guide, Academic Press, 1988.
[Mun] Munkres, J. R. Topology: a first course, Prentice-Hall, Inc., Englewood Cliffs, N.J.,
1975. xvi+413
[Mun1] Munkres, J. R. Analysis on manifolds, Addison-Wesley Publishing Company,
Advanced Book Program, Redwood City, CA, 1991, xiv+366 pp.
[OS] Y. Otsu and T. Shioya The Riemannian structure of Alexandrov spaces J. Differential
Geom. 39 (1994), no. 3, 629–658.
[Ot] Y. Otsu, Differential geometric aspects of Alexandrov spaces, in Comparison Geome-
try , Math. Sci. Res. Inst. Publ 30 Cambridge Univ. Press, Cambridge (1997), 135–148.
[Per1] G. Perelman, The Beginning of the Morse Theory on Alexandrov spaces, St. Pe-
tersburg Math. Journ. 5 (1994), no. 1, 205-214.
[Pet] A. Petrunin, Parallel transportation for Alexandrov spaces with curvature bounded
below, Geom. Funct. Analysis 8 (1998), no. 1, 123-148.
[Pet1] A. Petrunin, Applications of quasigeodesics and gradiant curves, in Comparison
Geometry, Math. Sci. Res. Inst. Publ 30, Cambridge Univ. Press, Cambridge (1997),
203-219
[Pl] C. Plaut, Metric spaces of curvature ≥ k, in “A handbook of Geometric Topology”
(to appear)
[PP] G. Perelman and A. Petrunin, Extremal subset in Alexandrov spaces and a generalised
Liberman theorem, St. Petersburg Math. Journ. 5 (1994), no. 1, 215-227.
[PP1] G. Perelman and A. Petrunin, Quasigeodesics and gradient curves in Alexandrov
spaces, Preprint (1994), 1-23.
[Resh] Yu. G. Reshetnyak, Nonexpending maps into spaces of curvature not greater K,
Siberian. matem. j. 9, no. 4 (1968), 918-927 (in Russian).
[Sp] M. Spivak, Calculus on manifolds. A modern approach to classical theorems of
advanced calculus, W. A. Benjamin, Inc., New York-Amsterdam 1965, xii+144 pp.
[St] L.Stoyanov, An estimate from above of the number of periodic orbites for semi-
dispersedbilliards Comm. Math. Phys. 24 (1989), no. 2, 217-227.
[Yam] T. Yamaguchi, A convergence theorem in the geometry of Alexandrov spaces, Actes
de la Table Ronde de Geometrie Differentielle (Luminy, 1992), 601–642, Semin. Congr.,
1, Soc. Math. France, Paris, 1996.
Index

(−k)-hyperbolic cone, 355 totally discontinuous, 83

(ε, δ)-approximation, 262 admissible curve, 178, 179
(m, ε)-strained point, 381 Alexandrov space, 101
(m, ε)-strainer, 381 of curvature ≥ k, 351
C-quasi-geodesic, 290 Alexandrov’s Lemma, 115
F -concave function, 120 angle, 96
F -convex function, 120 condition, 108
G-invariant metric, 279 triangle inequality, 98
Z2 -periodic metric, 301 upper, 98
δ-hyperbolicity, 285 angular metric, 31
ε-approximation, 262 arcwise isometry, 85
ε-isometry, 258 Arzela–Ascoli Theorem, 47
ε-midpoint, 41, 42 asymptotic
ε-net, 13 cone
ε-regular point, 397 Gromov–Hausdorff, 276
ε-separated, 14 of a convex set, 244
λ-convex function, 332 line (geodesic), 169
λ-homothety, 370 lines, 168
σ-additivity, 18 norm, 300
σ-algebra, 17 volume, 304
Borel, 18 auxiliary Euclidean structure, 145
k-plane, 127, 308
k-spherical cone, 355
Baire’s Theorem, 13
k-suspension, 355
ball, metrical, 7
m-strained point, 381
Ball-Box Theorem, 185
m-strainer, 381
barycenter, 334
bending, 145
abelian group Besikovitch inequality, 202, 205
rank, 298 Bieberbach inequality, 22
torsion, 298 billiard
absolute geometry, 158 nondegeneracy assumption, 346
action table (semi-dispersing), 343
co-bounded, 280 trajectory, 343
co-compact, 280 Bishop inequality, 374
free, 83 Blaschke Theorem, 253
of a group, 75 Borel

409
410 Index

σ-algebra, 18 Gromov–Hausdorff, 254

measure, 18 Lipschitz, 251
boundary of Alexandrov space, 398 uniform, 248
boundedly compactness, 14 uniform of curves, 47
bouquet of length spaces, 110 convex
Busemann function, 331
boundary, 172 set, 89
compactification, 171 surface, 357
function, 170 convexity radius, 312
of a set, 312
Cantor diagonal process, 47 coordinate
Carnot–Carathéodory space, 179 line, 138
Cartan–Hadamard Theorem, 325 system, 138
Cartan-Alexandrov-Toponogov Comparison degenerate, 139
Theorem, 240 normal, 153
CAT-condition, 106 vector, 138
Cauchy sequence, 10 correspondence, 256
Cayley graph, 73 covariant
center of mass, 334 derivative, 212, 215
change of variable formula, 195 differentiation, 214
Chow–Rashevsky Theorem, 183, 184 covering, 79
circumcenter, 334 base, 79
circumradius, 334 deck transformation, 82
circumscribed ball, 334 group of, 81
co-bounded action, 280 number of sheets, 79
co-compact action, 280 regular, 81
comass, 200 universal, 80
commuting vector fields, 216 coverings equivalent, 80
compact space, 13 curvature, 155
compactness, 14 Gaussian, 145, 224, 233
comparison of the sphere, 163
angle, 352 operator, 224
function, 103 principal, 145, 220
space, 154 sectional, 224
triangle, 107, 308 tensor, 224
complete length structure, 30 vector, 223
completion, 12 curve
concatenation, 26 admissible, 178
cone, 90, 275 rectifiable, 34
(−k)-hyperbolic, 355 simple, 45
k-spherical, 355 unparameterized, 44
asymptotic, 244, 276
hyperbolic, 355 deck transformation, 82
metric, 91 degenerate coordinate system, 139
projection, 93 degree
spherical, 95 for oriented manifolds, 204
tangent, 321 modulo 2, 203
configuration space, 342 for manifolds with boundary, 203
conformal map, 168 of a vertex, 71
conformally flat, 32 derivative, 138
conjugate point, 233, 238 diameter, 11
constant curvature region, 235 of k-plane, 308
constraint, 179 dilatation, 9, 249
nonholonomic, 179 dimension, 17
control, 180 geometric, 341
function, 180 Hausdorff, 22
convergence of a polyhedral space, 70
Index 411

rough, 375 strongly convex, 332

dimensional homogeneity, 370 functional
direction, 100 semi-continuous, 35
principal, 220 fundamental form
directional space, 100 first, 220
disc model of hyperbolic plane, 161 second, 220
displacement function, 332
distance, 1 Gauss Lemma, 148, 153
condition, 103 Gauss-Bonnet formula, 172, 226
coordinates, 384 Gaussian curvature, 145, 233
function geodesic, 51, 150, 222
1-dimensional, 102 equation, 222
Gromov–Hausdorff, 254 minimal, 52
Hausdorff, 252 geodesically complete space, 318
Lipschitz, 250 gluing, 60, 62, 67
minimizers, 48 graph, 68
distortion Cayley, 73
of a correspondence, 257 Gromov hyperbolic space, 290
of a map, 249 Gromov’s Compactness Theorem, 376
distribution, 178 Gromov–Bishop inequality, 371
dotted line, 42 Gromov–Hausdorff
dyadic rational, 43 convergence, 254, 260
of pointed spaces, 272
edge distance, 254
of a graph, 70 space, 259
of a polyhedral space, 68 topology, 260
equidistant, 231 group
equivalent curves, 137 action, 75
evenly covered set, 79 action by isometries, 280
exponential map, 322 finitely generated, 73
extremal subset, 399 isometry, 74
extrinsic properties, 145 of covering, 81
of deck transformations, 82
face of a polyhedral space, 68
Fifth Euclid’s Postulate, 157 Hadamard space, 324
Finsler Hausdorff
length structure, 33 dimension, 22
structure, 198 distance, 252
first fundamental form, 220 measure, 19
first variation hinge, 114
formula, 121 Holmes–Thompson volume, 200
formula, 123 homogeneous, 149
of length, 222, 223 fully, 166
fixed point, 11 homogeneous space, 75
flat Hopf–Rinow–Cohn-Vossen Theorem, 52
space, 75 horoball, 170
torus, 65 horocycle, 170
form hyperbolic
quadratic, 5 distance, 160, 162
form bilinear, 5 plane, 149, 154, 159
free action, 83 rigid motions, 159
Frobenius Theorem, 216 hyperbolic cone, 355
fully homogeneous, 166 hyperparallel, 169
function
λ-convex, 332 ideal
comparison, 103 boundary, 168
convex, 331 of Hadamard space, 335
412 Index

triangle, 169 induced, 34

induced metric of disjoint union, 63
intrinsic metric, 36 of a curve, 34
length structure, 31 of a tangent vector, 141
inequality of curve, 33
Besikovitch, 202, 205 of path, 26
Bishop, 374 space, 28, 29
injectivity radius, 152 polyhedral, 33
integral curve, 138 quasi-geodesically stable, 290
interval, 26 structure, 26
intrinsic complete, 30
geometry, 144 conformal, 139
metric, 4 Finsler, 139
intrinsic metric Finslerian, 140
induced, 36 induced, 31, 34
inversion, 159, 177 Riemannian, 32
involution, 177 semi-continuous, 39
isometric Levi-Civita Lemma, 216
embedding, 86 Lie bracket, 182, 215
pointed spaces, 273 lift, 81
spaces, 74 of a metric, 79
isometric spaces, 2 line, 335
isometry, 2, 74 straight, 366
arcwise, 85 lines
surjective, 86 corresponding, 366
group, 74 parallel, 366
local, 78 link, 94
pointed, 273 Lipschitz
isotropic, 155, 166 convergence, 250
distance, 249, 250
Jacobi map, 250
equation, 228 Lobachevsky plane, 159
vector field, 228 local
Jacobian, 194 isometry, 78
signed, 194 locality of intrinsic metric, 60
locally
kinetic energy, 342 compact topological space, 49
Klein bottle, 177 convex set, 90
intrinsic metric, 319
Lagrangian, 140 path connected, 30
Lebesgue logarithmic map, 322
integration, 55, 194
lemma, 15 map
measure, 18 bi-Lipschitz, 9
number, 15 covering, 79
left-invariant metric, 280 distance-preserving, 2
on the fundamental group, 281 exponential, 322
lemma Lipschitz, 9
Alexandrov, 115 locally Lipschitz, 9
Busemann–Feller, 358 logarithmic, 322
covering homotopy, 81 nonexpanding, 78
Gauss, 148, 153 nonexpending, 9
Lebesgue, 15 uniformly continuous, 15
Levi-Civita, 216 maximal semi-metric, 66
Morse, 175 measure, 17
length, 26 Borel, 18
first variation, 222, 223 Hausdorff, 19
Index 413

Lebesgue, 18 vector field, 220

metric, 1 parameterization, 44
G-invariant, 83 by arc-length, 46
angular, 31 constant speed, 46
bi-Lipschitz, 9 natural, 45
coefficients, 141 unit speed, 46
dilated, 4 partition, 34
direct product, 88 path, 26
graph, 70 never-locally-constant, 45
induced, 38 simple, 45
intrinsic, 4, 29 Poincaré model, 159
length, 29 in the disk, 161
locally intrinsic, 319 in the upper half-plane, 159
maximal, 65 point
quotient, 76 ε-regular, 397
rescaled, 4 regular, 378, 393
space, 1 pointed
complete, 10 isometry, 273
pointed, 272 metric space, 272
strictly intrinsic, 30, 41, 42 polar set, 200
word, 74 polyhedral
metric bouquet, 110 space, 67, 69
midpoint, 41 dimensionally homogeneous, 70
minimal geodesic, 52 finite, 70
Minkowski space, 5 locally finite, 70
monotonicity condition, 114 principal
Morse Lemma, 175, 290 curvature, 145, 220
direction, 220
natural parameterization, 45 product metric, 88, 89
net, 278 projection, 337
nondegeneracy assumption, 346
nonnegatively curved space, 103, 107, 108, quadruple condition, 353
114 quasi-geodesic, 290
nonpositively curved space, 103, 107, 108, in hyperbolic geometry, 175
114 quasi-geodesically stable, 290
norm, 4 quasi-isometric
Euclidean, 6 groups, 282
normal spaces, 277
ball, 312 quasi-isometry, 278
coordinates, 152 quasigeodesic, 402
neighborhood, 312 quotient, 355
region, 107, 116, 312 metric space, 62
normed space, 5 space, 2
number of sheets, 79
Rademacher Theorem, 196
orbit, 76, 280 radius of a space, 366
metric, 280 Rauch Comparison Theorem, 237
orthogonal ray, 335
collection, 394 ray boundary, 172
projection to a convex body, 358 rectifiable curve, 34
orthonormal regular
basis, 6 covering, 81
frame, 6 point, 202, 378, 393
polygon, 176
parallel value, 202
rays, lines, 335 reparameterization, 27
transport, 220, 221 Reshetnyak’s Theorem, 316
414 Index

Riemannian spherical cone, 95

area, 147 splitting theorem, 366
length structure, 32 stability theorem, 400
manifold, 143 stable norm, 300, 301
metric, 141 strictly convex, 304
isometry, 142 strainer, 380
volume, 193 number, 381
coordinate formula, 196 stratification, 398
rotation vector, 302 stratum, 398
rough strictly intrinsic
dimension, 375 length structure, 30
volume, 375 metric, 30, 41, 42
strongly convex function, 332
Sard–Brown Theorem, 202 sub-Riemannian metric, 179
scalar product, 6 subcone, 288
second fundamental form, 220 finite, 288
semi-continuity of angles, 118 submanifold totally geodesic, 90
semi-continuous functional, 35 surface totally geodesic, 90
semi-metric, 2 suspension, 95
maximal, 66 symplectic volume, 200
quotient, 62
separated net, 278 tangent
set cone
ε-separated, 14 Gromov–Hausdorff, 275
Borel, 18 of a convex set, 243
compact, 13 vector, 137
convex, 89 tangent cone, 321
evenly covered, 79 Gromov–Hausdorff, 321
locally convex, 90 tessellation, 176
measurable, 17 theorem
nowhere dense, 13 Arzela–Ascoli, 47
polar, 200 Baire, 13
shortest path, 26, 30, 47, 48, 52 Blaschke, 253
simplex, 68 Cartan–Hadamard, 325
slim triangle, 330 Cartan-Alexandrov-Toponogov, 240
Soul Theorem, 400 Chow–Rashevsky, 183, 184
space Frobenus, 216
δ-hyperbolic, 285 Gromov’s compactness, 376
compact, 13 Hopf–Rinow–Cohn-Vossen, 52
flat, 75 Rademacher, 196
form, 312, 371 Rauch, 237
geodesically complete, 318 Reshetnyak, 316
homogeneous, 75 Sard–Brown, 202
length, 29 Toponogov, 360
locally Toponogov’s Theorem, 360
arcwise connected, 80 totally bounded, 13
simply connected in the large, 80 totally geodesic, 90
locally isometric to, 75 triangle
of curvature ≥ k, 128, 353 condition, 107
of curvature ≤ k, 128, 308, 309 ideal, 169
of curvature bounded above, 308 inequality for angles, 98
of directions, 98, 100, 318 slim, 330
semi-locally simply connected, 80
speed of a curve, 55 uniform
sphere, 7 convergence
spherical of functions, 248
geometry, 31, 149 of metrics, 248
Index 415

of metrics spaces, 249

on compact sets, 272
distance, 86
uniformly totally bounded class, 263
universal covering, 80
unparameterized curve, 44
upper angle, 98

variation of a curve, 211

vector field, 138
commuting, 216
coordinate, 138
parallel, 220
vertex, 67
of a graph, 70
Vitali’s Covering Theorem, 21
volume
Finsler, 193
Holmes–Thompson, 200
of Lipschitz map, 197
Riemannian, 193
symplectic, 200
Voronoi region, 177

walls of billiard table, 343

weak convergence, 401
word metric, 74, 244