Geometric embeddings of metric spaces

By Juha Heinonen

Lectures in the Finnish Graduate School of Mathematics, University

of Jyväskylä, January 2003.

1 Basic Concepts.
2 Gromov-Hausdorff convergence.
3 Some fundamental embeddings.
4 Strong A∞ -deformations.

1
2

Preface
These notes form a slightly expanded version of the lectures that I
gave in the Finnish Graduate School of Mathematics at the University
of Jyväskylä in January 2003. The purpose of this mini-course was to
introduce beginning graduate students to some easily accessible ques-
tions of current interest in metric geometry. Besides additional remarks
and references, the only topics that were not discussed in the course,
but are included here, are the proof of the existence of a Urysohn uni-
versal metric space and the proof of Semmes’s theorem 4.5.
I thank Tero Kilpeläinen for inviting me to give these lectures, Pekka
Koskela for offering to publish the lecture notes in the Jyväskylä Math-
ematics Department Reports series, and Juha Inkeroinen for typing a
preliminary version of this mansucript from my hand written notes.
I am grateful to Bruce Hanson who carefully read the entire manu-
script and made useful suggestions. Finally, I thank the Mathematical
Sciences Research Institute and NSF (Grant DMS 9970427) for their
support.

Berkeley, September 2003

Juha Heinonen
 3

1. Basic Concepts
Let
X = (X, d) = (X, dX )
denote a metric space. Throughout these lectures, we will consider
quite general metric spaces. However, the reader should not think of
anything pathological here (like the discrete metric on some huge set).
We are motivated by problems in analysis and geometry.
For background reading, see [8], [33], [23], [37], [65].
1.1. Examples of metric spaces. (a) Any subset of Rn with the
inherited metric. Many, but by no means all, interesting metric spaces
belong to this group. We can also equip Rn with the p-norms, where
 n  p1

|x|p = |xi |p , 1 ≤ p < ∞,
i=1

and
|x|∞ = max |xi | ,
i=1,...,n
for x = (x1 , . . . , xn ) ∈ R . The corresponding distance or metric is
n

|x − y|p , 1 ≤ p ≤ ∞.

(b) Cones over metric spaces. Let Z be a metric space. The metric
cone C(Z) over Z is the completion of the product Z × (0, ∞) in the
metric that is defined by

d((z, r), (z  , r )) := r2 + r2 − 2r · r cos(dZ (z, z  )) ,
if dZ (z, z  ) ≤ π, and by
d((z, r), (z  , r )) := (r + r ) ,
if dZ (z, z  ) ≥ π. If Z is a closed (=compact, without boundary) smooth
Riemannian manifold, then C(Z) is a smooth Riemannian manifold,
except perhaps at the tip (=the point added in the completion).
The cone Cρ over the circle of length ρ > 0 is flat outside the tip
in that it is locally isometric to a patch in R2 . The tip is a singular
(nonsmooth) point, except when ρ = 2π. The curvature on Cρ can
defined in a certain generalized sense; it is a Dirac measure at the tip
of total mass 2π − ρ.
Read more in [2], [14], [13], [9], [11].

(c) Snowflake spaces

(R, |x − y|ε ) , 0 < ε < 1.
These are indeed metric spaces because the inequality
(a + b)ε ≤ aε + bε
4

is valid whenever a, b ≥ 0 and 0 < ε < 1. The name stems from the
fact that (R, |x − y|ε ), for 12 < ε < 1, admits a bi-Lipschitz embedding
in R2 , and the image resembles the boundary of a snowflake. (See 1.3
for the definition of a bi-Lipschitz embedding.)
More generally we say that a metric space (X, dε ) is a snowflake
version of (X, d) if 0 < ε < 1.
Read more in [5], [72], [23], [37], [65].

(d) Rickman’s rug

(R2 , |x − x | + |y − y  |ε ) , 0 < ε < 1.
This can be thought of as a product of a snowflake and a line. Here
R2 = {(x, y) : x, y ∈ R} .
Metrics like this exhibit an interesting asymmetry. They show up in
problems related to the heat equation, in the theory of pinched cur-
vature Riemannian manifolds, and the theory of quasiconformal and
related mappings.
Read more in [72], [54], [3], [75].

(e) Various classical fractals, such as the Sierpinski carpet or gasket.

Read more in [26], [52], [42], [23], [41].

(f) Sub-Riemannian manifolds and, in particular, Carnot groups.

These are intriguing metric spaces with complicated local geometry,
but many nice global analytic and geometric properties. They arise in
complex analysis, control theory, and partial differential equations.
Read more in [32], [68], [7], [36], [69].

(g) Spaces with generalized curvature bounds. These are metric

spaces where curvature makes sense either as a measure, or as a metric
quantity that can be said to be bounded, or bounded from above or
below. (Compare (b).)
Read more in [2], [21], [56], [55], [14], [13].
1.2. Curves in metric spaces. A (compact) curve in X is a contin-
uous mapping γ : [a, b] → X, where [a, b] ⊂ R is an interval. A curve
is rectifiable, if
n
sup d(γ(ti+1 ), γ(ti )) < ∞ ,
i=0
where the supremum is taken over all sequences
a = t0 ≤ t1 ≤ · · · ≤ tn+1 = b .
A metric space is quasiconvex, if there exists a constant C ≥ 1 with
the property that every pair of points x, y in the space can be joined
by a curve whose length is no more than C · d(x, y).
 5

The Sierpinski carpet and gasket, although fractal objects, are qua-
siconvex. The snowflake space (R, |x − y|ε ), 0 < ε < 1, contains no
nonconstant rectifiable curves. In fact, no snowflake version of a met-
ric space can contain nonconstant rectifiable curves. The only non-
constant rectifiable curves in Rickman’s rug are those with constant
y-coordinate.
By the aid of rectifiable curves, some differential analysis in metric
spaces is possible via reduction to the one variable case.
Read more in [13], [38], [37], [60], [66].

1.3. Embeddings. A map ϕ : X → Y is an embedding if it is a

homeomorphism onto its image. An embedding ϕ is L-bi-Lipschitz,
L ≥ 1, if
1
dX (a, b) ≤ dY (ϕ(a), ϕ(b)) ≤ L dX (a, b)
L
whenever a, b ∈ X. If L = 1, then ϕ is an isometric embedding and X
is isometrically embedded in Y .

1.4. Open problem. Characterize metric spaces that are bi-Lipschitz

embeddable in some Rn .

Very little progress has been made on this problem. See Section 3
for further discussion.
Read more in [48], [63], [37], [49].

1.5. Motivation. Metric spaces arise in nearly all areas of mathemat-

ics. The metric point of view has been useful even in group theory,
where finitely generated groups can be equipped with a word metric.
In analysis, the study of various function spaces has a long history, and
completeness is an important concept there. The study of completeness
of a class of geometric objects is more recent. For example, an impor-
tant problem in Riemannian geometry concerns the degeneration of
metrics on a fixed manifold; singular spaces occur as limits. Similarly,
focusing on graphs of functions, rather than on their analytic expres-
sions, can shed new light on classical problems as well. For example,
the functions x → xn , n = 2, 4, 6, . . . , do not converge to a finite valued
function on R, but their graphs have a nice locally uniform limit as a
subset of R2 .
Read more in [31], [33], [16], [70], [71], [24], [21], [56], [55], [23].

1.6. Kuratowski embedding. Every metric space X embeds isomet-

rically in the Banach space L∞ (X) of bounded functions on X with
sup-norm:
||s||∞ := sup |s(x)| , s : X → R.
x∈X
6

Proof. Fix x0 ∈ X and define

x → sx , sx (a) = d(x, a) − d(a, x0 ) .
Then by the triangle inequality
|sx (a)| ≤ d(x, x0 )
and
|sx (a) − sy (a)| = |d(x, a) − d(y, a)| ≤ d(x, y) ,
and equality occurs if a = x or a = y.

1.7. Remark. If X is bounded, we do not need to subtract the term
d(a, x0 ) in the definition of sx . In this case, the embedding is canonical.
In the unbounded case, the embedding depends upon the choice of the
base point x0 .
The letter s in sx conveys the idea that the elements in L∞ (X) are
bounded sequences s = (sa ) indexed by a ∈ X.
Theorem 1.6 is due to to Kuratowski [43].
1.8. Application. If f, g : X → Y , then the sum f + g is meaningless
in general. However, after embedding Y isometrically,
Y → L∞ (Y ) ,
we have
f + g : X → L∞ (Y ).
For example, we can form the homotopy
ht = (1 − t)f + tg : X → L∞ (Y ) , 0 ≤ t ≤ 1.
Here is a simple application of the idea of enlarging the range from
Y to L∞ (Y ).
1.9. Proposition. Let X be compact and f : X → Y continuous.
Then there exists a sequence (fi ) of Lipschitz continuous maps fi :
X → L∞ (Y ) such that fi → f uniformly as i → ∞.

Recall that a map f : X → Y is Lipschitz, or L-Lipschitz, L ≥ 1, if

dY (f (x), f (y)) ≤ L dX (x, y)
whenever x, y ∈ X.
1.10. Remark. In general, the conclusion of Proposition 1.9 cannot
hold with fi (X) ⊂ Y . For example, the identity map
1
[0, 1] → ([0, 1], |x − y| 2 )
cannot be approximated by Lipschitz maps because the target space
has no nonconstant rectifiable curves.

The proof for Proposition 1.9 requires some concepts that are useful
in other contexts as well. These are discussed next.
 7

1.11. Maximal nets. A maximal ε-net, ε > 0, in a metric space X is

a collection of points
N = N (X, ε) ⊂ X
such that
d(x, x ) ≥ ε
whenever x, x ∈ N are distinct, and that

X= B(x, ε) .
x∈N

Here and later,

B(x, ε) = {y ∈ X : d(x, y) < ε}
is an open ball centered at x with radius ε.
We often call maximal nets simply nets.
The following facts can be proved easily by using Zorn’s lemma:
• Maximal ε-nets exist for every X and ε > 0.
• Given an ε-net N (X, ε) and ε < ε, there exists an ε -net
N (X, ε ) extending N (X, ε), i.e.,
N (X, ε) ⊂ N (X, ε ) .
• Given x0 ∈ X, 0 < R < R , 0 < ε < ε, and an ε-net
N (X, R, ε) := N (B(x0 , R), ε)
of the ball B(x0 , R), then there exists an ε -net
N (X, R , ε ) = N (B(x0 , R ), ε )
extending N (X, R, ε).

1.12. Lipschitz partition of unity. Let

N = N (X, ε) = (xi )
be a countable ε-net in X. Write
Bi = B(xi , ε) , 2Bi = B(xi , 2ε),
and assume that the collection {2Bi } is locally finite in the following
sense: only finitely many balls 2Bj meet a given ball 2Bi . In particular,
if X is compact, such nets exist for every ε > 0.
Define
1
ψi (x) = min{1, dist(x, X\2Bi )}.
ε
Then
• 0 ≤ ψi ≤ 1, ψi |Bi = 1, and ψi |X\2Bi = 0.
• ψi is 1ε -Lipschitz.

• ψ := i ψi ≥ 1 and ψ|2Bi is ε -Lipschitz, where Li is the
Li

number of balls 2Bj meeting 2Bi .

8

Set
ψi
ϕi = .
ψ
Then 
0 ≤ ϕi ≤ 1 , ϕi |X\2Bi = 0 , and ϕi = 1 .
i

We further claim that ϕi is Li -Lipschitz, where

2Li
Li =
.
ε
To this end, let first x, y ∈ 2Bi . Then
|ψi (x)ψ(y) − ψi (y)ψ(x)|
|ϕi (x) − ϕi (y)| =
ψ(y)ψ(x)
≤ |ψi (x)ψ(y) − ψi (y)ψ(y)| + |ψi (y)ψ(y) − ψi (y)ψ(x)|
1 Li
≤ sup |ψ| d(x, y) + d(x, y)
2Bi ε ε

≤ Li d(x, y) .
Next, let x ∈ 2Bi and y ∈ X\2Bi . Then
1
|ϕi (x) − ϕi (y)| = |ϕi (x)| ≤ |ψi (x)| ≤ d(x, y) .
ε
This proves the claim.

We call a collection of functions {ϕi } as above the Lipschitz partition

of unity associated with the net N .

Proof of Proposition 1.9. The preceding notation and terminology

understood, fix ε > 0 and an ε-net N = (xi ). Define

fε (x) = ϕi (x)f (xi ).
i

Then
fε : X → L∞ (Y )
is Lipschitz, because N is finite. Fix x ∈ X. Let Ix denote the set of
those indices i such that x ∈ 2Bi . We find

|fε (x) − f (x)| = | ϕi (x)(f (xi ) − f (x))|
i∈Ix
≤ max |f (xi ) − f (x)|
i∈Ix
≤ ω(2ε),
where ω is the modulus of continuity of the uniformly continuous map
f . Because ω(ε) → 0 as ε → 0, the proposition follows.
 9

1.13. Remark. The modulus of continuity of a map f : X → Y is the

function
ω = ωf : [0, ∞) → [0, ∞]
defined by
ω(t) = sup{dY (f (x), f (y)) : dX (x, y) ≤ t} .
If X is compact, then the modulus of continuity of every continuous
map is bounded and satisfies ω(t) → 0 as t → 0.

1.14. Doubling spaces. The functions ϕi in the Lipschitz partition

of unity have Lipschitz constant C/ε, with C independent of ε, if the
amount of overlap of the balls 2Bi is uniformly bounded.
We call X doubling if there exists C ≥ 1 such that every ball in X
can be covered by C balls with half the radius.

1.15. Proposition. A metric space X is doubling if and only if there

exist constants C  > 0 and α > 0 such that
C 1
#{εR-net in BR } ≤ α , 0 < ε ≤ ,
ε 2
for every ball BR of radius R in X, where # denotes the cardinality.
The constants C  , α, and the doubling constant C depend only on each
other.
Proof. Exercise.

1.16. Assouad dimension. The infimal α > 0 such that the condition
in Proposition 1.15 holds is called the Assouad dimension of X. It is
always at least as large as the Hausdorff dimension of the space, and
can be strictly larger.

1.17. Examples. (a) Every subspace of a doubling space is a doubling

space (with the same constant).

(b) Rn and all its subsets are doubling with constant C = C(n).
This is a special case of (a) and (c).

(c) If X carries a nontrivial doubling (Borel) measure µ, i.e.,

µ(B(x, 2R)) ≤ Cµ(B(x, R))
for some C ≥ 1 and for every ball B(x, R) in X, then X is doubling.
Conversely, if X is a complete doubling space, then X carries a
nontrivial doubling measure [77], [51].

(d) Infinite dimensional normed spaces are never doubling. In fact, a

complete doubling space is always proper, i.e., closed balls in the space
are compact.
10

1.18. Open problem. Which subsets of Rn carry nontrivial doubling

measures? Difficult and unsolved even for n = 1.

Read more on doubling spaces in [5], [77], [23], [51], [57], [37], [65],
[12], [50].

1.19. Lipschitz extensions. The Kuratowski embedding is useful in

extending Lipschitz maps. Recall first the McShane-Whitney extension
lemma:
If A ⊂ X and f : A → R is L-Lipschitz, then there exists an L-
Lipschitz function F : X → R extending f , i.e., F |A = f .
Proof. The proof is simple. Define
fa (x) = f (a) + L d(x, a) , a ∈ A.
Then fa is L-Lipschitz, fa ≥ f (a), and fa (x) = f (x) if x = a. The
function
F (x) = inf fa (x)
a∈A

is the required extension.

1.20. Remark. For mappings f : A → Y , A ⊂ X, Lipschitz exten-

sions are not always possible simply for topological reasons. These
extensions can also fail to exist for metric reasons. For example, con-
sider the (Lipschitz) identity map
1
{0, 1} → ([0, 1], |x − y| 2 ) ,
and observe that no Lipschitz extension to [0, 1] is possible (cf. 1.10).
On the other hand, if A ⊂ X and f : A → Y is L-Lipschitz, then
there exists an L-Lipschitz map F : X → L∞ (Y ) extending f . (Here
we understand that a Kuratowski embedding, with a base point, has
been fixed for Y ⊂ L∞ (Y ).) This assertion follows from the McShane-
Whitney lemma applied to each component of f . Indeed, recall that
the elements of L∞ (Y ) can be thought of as bounded sequences indexed
by Y . Thus,
f (a) = (fya )y∈Y , |fya − fyb | ≤ ||f (a) − f (b)||∞ ≤ L dX (a, b) ,
and we can extend each of the real-valued functions
a → fya , y ∈ Y.
Get
F (x) = (Fyx )y∈Y , |Fyx1 − Fyx2 | ≤ L dX (x1 , x2 )
for all y ∈ Y , x1 , x2 ∈ X, so that
||F (x1 ) − F (x2 )||∞ ≤ L dX (x1 , x2 ) .
 11

1.21. Example. Let f : X → R be a function such that

|f (x) − f (y)| ≤ (g(x) + g(y)) d(x, y) , x, y ∈ X,
for some g : X → [0, ∞). Then

∞
X= Xi
i=1

where f |Xi is i-Lipschitz. Indeed, define

Xi := g −1 {[0, i/2]}.
If X = Rn , we find by using Rademacher’s theorem and the extension
lemma that every function f as above is approximately differentiable
almost everywhere. These observations lead, for example, to a.e. ap-
proximate differentiability of Sobolev functions.
Read more in [27], [25], [35].

1.22. Fréchet embedding. The Kuratowski embedding has the dis-

advantage that the receiving space L∞ (X) depends on X. Already in
1909, Fréchet [28] 1 observed the following fact:
Every separable metric space X isometrically embeds in the Banach
space l∞ .
Recall that l∞ = L∞ (N) is the space of bounded sequences equipped
with the sup-norm.
Proof. Pick a dense countable subset {x0 , x1 , . . .} of X and consider
x → sx , sxi = d(x, xi ) − d(xi , x0 ).
Note that this is exactly the (unique) extension of the Kuratowski
embedding, with base point x0 ,
{x0 , x1 , . . .} → L∞ ({x0 , x1 , . . .})  l∞ ,
from a dense set to a complete space. The embedding x → sx is
isometric because
|sai − sbi | = |d(a, xi ) − d(b, xi )| ≤ d(a, b) ,
with equality obtained when xi approximates a (or b).

We have thus obtained a universal Banach space l∞ which isomet-
rically contains all separable metric spaces. The isometric embedding
depends on the chosen dense subset, and on the base point x0 , but is
otherwise concretely defined.
Note that the Lipschitz extension lemma 1.19 is valid for functions
valued in l∞ , because the sup-norm is used and so component-wise
extensions are valid.
1The concept of an abstract metric space was first introduced in [28].
12

2. Gromov-Hausdorff convergence
Given X, Y ⊂ Z, we define the Hausdorff distance between X and Y
in Z by
dH (X, Y ) := dZH (X, Y ) := inf{ε > 0 : X ⊂ Y (ε) and Y ⊂ X(ε)} ,
where

A(ε) := {x ∈ Z : dist(z, A) < ε} = B(a, ε)
a∈A

is the ε-neighborhood of a set A ⊂ Z.

Obviously
dH (X, Y ) = dH (Y, X)
and it is easy to see that
dH (X, Y ) ≤ dH (X, W ) + dH (W, Y ) .
Thus, dH behaves like a metric on all subsets of Z, except that it can
take the value ∞, and that dH (X, Y ) = 0 does not necessarily imply
that X = Y . Examples to this effect are easy to give.

2.1. Proposition. We have dH (X, Y ) > 0 for closed subsets X = Y

of Z. In particular, dH determines a metric on the collection of all
compact subsets of a metric space Z.

Proof. Exercise.

Gromov-Hausdorff distance is a generalization of the Hausdorff dis-

tance; it measures how much two abstract metric spaces deviate from
each other. This distance can be defined in several equivalent ways. We
will only consider separable spaces, and employ the Fréchet embedding
to this end.
The Gromov-Hausdorff distance between two separable metric spaces
X, Y is
dGH (X, Y ) := inf d∞
H (ϕ(X), ψ(Y )) ,

where
∞
d∞ l
H := dH

is the Hausdorff distance in l∞ and where the infimum is taken over all
isometric embeddings ϕ : X → l∞ and ψ : Y → l∞ .
Obviously, if X and Y are isometric, then
dGH (X, Y ) = 0.
The converse is not true in general, and it is wise here to restrict to
isometry classes of compact spaces. (Note that compact spaces are
separable.)
 13

2.2. Proposition. The Gromov-Hausdorff distance determines a met-

ric on the isometry classes of all compact metric spaces.

The statement needs some explanation. Among compact metric

spaces we define an equivalence relation, where X ∼ Y if and only
if X is isometric to Y . An isometry class of X is the equivalence class
determined by X under this relation.
Proof. We need to show that dGH (X, Y ) = 0 implies the existence of
a (surjective) isometry ϕ : X → Y . Fix ε > 0 and choose a maximal
ε-net
Nε = {x1 , x2 , . . . , xk } ⊂ X.
We may assume that X, Y ⊂ l∞ and that
d∞ 2
H (X, Y ) < ε .

Write
|a − b| := ||a − b||∞
for a, b ∈ l∞ . For each i = 1, . . . , k choose yi ∈ Y such that
|xi − yi | < ε2 ,
to obtain a map Fε : Nε → Y . We have
|Fε (xi ) − Fε (xj )| = |yi − yj |
≤ |yi − xi | + |xi − xj | + |xj − yj |
≤ 2ε2 + |xi − xj |
≤ (1 + 2ε)|xi − xj |.
Similarly,
|yi − yj | ≥ |xi − xj | − |xi − yi | − |xj − yj |
≥ |xi − xj | − 2ε2
≥ (1 − 2ε)|xi − xj |.
Moreover, if y ∈ Y , then there are x ∈ X, xi ∈ Nε such that
|y − yi | ≤ |y − x| + |x − xi | + |xi − yi | < ε2 + ε + ε2 < 2ε
for all small ε. It follows that, for ε small enough, the map Fε : Nε → Y
is a (1+10ε)-bi-Lipschitz embedding whose image is 2ε-dense in Y , i.e.,
the 2ε-neighborhood of the image contains Y . The standard proof of
the Arzela-Ascoli theorem then gives that Fε → F0 as ε → 0, where F0
is an isometry of a dense subset of X onto a dense subset of Y . (One
should use increasing nets Nε ⊂ Nε , ε < ε, here.) Such an isometry
has an extension X → Y . This completes the proof.

The following fundamental observation is due to Gromov [33].
14

2.3. Gromov’s compactness theorem (GCT). Let M = M(C, N )

be a family of metric spaces X such that
• diam(X) ≤ C ,
• #(every ε-net in X) ≤ N (ε),
where C is a finite constant and N : (0, ∞) → (0, ∞) is a positive
function. Then every sequence (Xi ) ⊂ M contains a subsequence that
converges in the Gromov-Hausdorff distance to a metric space X∞ . If,
in addition, each Xi is compact, then X∞ is compact.

We also have the following result (partially a corollary to 2.3):

2.4. Theorem. Let C denote the collection of isometry classes of com-

pact metric spaces. Then (C, dGH ) is a contractible, complete, and sep-
arable metric space.
Proof. C is contractible because, given X = (X, d), the spaces
Xt := (X, t d)
converge in the Gromov-Hausdorff distance to a one point space as
t → 0.
Completeness follows from Gromov’s compactness theorem, because
it easily follows that for a Cauchy sequence (Xi ) ⊂ C,
• supi diam(Xi ) < ∞
• supi #(every ε-net in Xi ) < ∞ , ε > 0.
Finally, to prove separability, observe first that (the isometry classes
of) finite spaces are dense in C. On the other hand, given a finite
set X = {x1 , x2 , . . . , xN }, the isometry classes of metrics on X are
described by a subset of the set of symmetric N × N -matrices, thus by
2
a subset of RN . This is a separable space.

Proof of the GCT. Let (Xi ) ⊂ M. We will pass to several subse-

quences in the course of the proof; these will all be denoted by (Xi ).
For each i and each j = 1, 2, . . . choose increasing maximal 2−j -nets

Ni,j := N Xi , 2−j ⊂ Ni,j+1 .
We assume, as we may, that there are infinitely many nets Ni,1 = ∅.
Moreover, we may assume that
n1 := #Ni,1 ≤ N (2−1 )
for all i.
Let Y1 = {1, 2, . . . , n1 } and choose bijections
ψi : Y1 → Ni,1 ⊂ Xi .
Pull back the metric dXi to Y1 , to obtain a metric
δi (p, q) := dXi (ψi (p), ψi (q)) , p, q ∈ Y1 .
 15

Write
Yi,1 := (Y1 , δi ).
We have
δi : (Y1 × Y1 )\diagonal → [2−1 , C],
so we can find a subsequence of metrics (hence spaces) δi , such that
δi → e1 : (Y1 × Y1 )\diagonal → [2−1 , C] ,
where e1 is a metric on Y1 .
Next, by passing to a further subsequence, we may assume that
n2 := #Ni,2 ≤ N (2−2 )
for all i. Extend the bijections ψi to bijections
ψi : Y2 → Ni,2 ⊂ Xi ,
where Y2 = {1, 2, . . . , n1 , n1 + 1, . . . , n2 }. As above, get a convergent
sequence of metrics δi on Y2 that converges to a metric e2 on Y2 , with
e2 : (Y2 × Y2 )\diagonal → [2−2 , C].
Continuing in this manner, always passing to a subsequence when nec-
essary, we get metrics
ej : (Yj × Yj )\diagonal → [2−j , C],
where
Yj = {1, 2, . . . , nj } , nj := #Ni,j ,
for each i. In sum, the metric spaces
Zj := (Yj , ej )
have the following property: for each j, there exists Xj from the original
sequence, a 2−j -net Nj in Xj , #Nj = nj , and a bijection
ψj : Zj → Nj
satisfying
|dXj (ψj (p), ψj (q)) − ej (p, q)| < 10−j
for p, q ∈ Zj . (Instead of 10−j , one can choose, in advance, any positive
function of j here.) In particular,
1 dXj (ψj (p), ψj (q)) 1
1 ≤ ≤1+ .
1+ j
ej (p, q) j
The proof will be completed by two lemmas. It is easy to see that
without loss of generality we can assume nj → ∞.

2.5. Lemma. There exists a metric d on N such that Zj → (N, d) in

dGH .
16

2.6. Lemma. If X, Y are two bounded and separable metric spaces and
if Ψ : X → Y is a (1 + ε)-bi-Lipschitz bijection, then
dGH (X, Y ) ≤ ε max{diam(X) , diam(Y )}.

Assume the lemmas. Then

C
dGH (Zj , Nj ) ≤ .
j
Because also
dGH (Nj , Xj ) ≤ 2−j
we have that
dGH (Xj , (N, d)) ≤ dGH (Xj , Nj ) + dGH (Nj , Zj ) + dGH (Zj , (N, d)) → 0
as j → ∞.
We also remark here that if, in the GCT, all spaces are compact, then
the completion of the limit space (N, d) is compact, too, because it is
easily seen to be totally bounded. (In fact, every Gromov-Hausdorff
limit of finite spaces is totally bounded.)
So it suffices to prove 2.5 and 2.6.
Proof of Lemma 2.5. Note first that
Zj ⊂ Zj+1
isometrically (ej+1 extends ej ), so the definition for d on N is obvious.
Because each Zj is (1 + 1/j)-bi-Lipschitz to a maximal 2−j -net Nj
in Xj , with Nj ⊂ Nj+1 , it is clear that for
p ∈ N\Zj , p ∈ Zk , k > j,
we have
d(p, Zj ) ≤ d(p, pk−1 ) + d(pk−1 , pk−2 ) + · · · + d(pj+1 , pj )
for pl ∈ Zl , which is less than

∞
2 2−k .
k=j

This expression → 0 as j → ∞, and the lemma follows.

Proof of Lemma 2.6. We will define a metric d, an extension of both
dX and dY , in the disjoint union Z = X  Y such that the Hausdorff
distance between X and Y in Z is less than C ε, where
C = max{diam(X), diam(Y )}.
This suffices.
To such end, define
d(x, y) := inf {dX (x, a) + dY (y, Ψ(a))} + C ε
a∈X
 17

for x ∈ X and y ∈ Y . To prove that this is a metric, we need to

consider two cases.
Case 1. x ∈ X, and y, y  ∈ Y .
Then
d(x, y) ≤ dX (x, a) + dY (y, Ψ(a)) + C ε
≤ dX (x, a) + dY (y, y  ) + dY (y  , Ψ(a)) + C ε
for all a ∈ X, so
d(x, y) ≤ d(x, y  ) + d(y  , y).
Case 2. x ∈ X, and y, y  ∈ Y .
Then
d(y, y  ) ≤ dY (y, Ψ(a)) + dY (Ψ(a), Ψ(a )) + dY (Ψ(a ), y  )
≤ dY (y, Ψ(a)) + dX (a, a ) + C ε + dY (Ψ(a ), y  )
≤ dY (y, Ψ(a)) + dX (a, x) + dX (x, a )
+ C ε + dY (Ψ(a ), y  ) + C ε .
By taking the infimum over all a, a ∈ X, we get
d(y, y  ) ≤ d(y, x) + d(x, y  )
as required.
Next, fix y ∈ Y and take x = Ψ−1 (y). Then
d(x, y) ≤ dX (Ψ−1 (y), a) + dY (y, Ψ(a)) + C ε
for each a ∈ X. In particular, we can take a = Ψ−1 (y) so that d(x, y) ≤
C ε. By symmetry,
dGH (X, Y ) ≤ C ε.
The lemma follows.

2.7. Concrete description of Gromov-Hausdorff convergence.
Let X, Y be separable and bounded spaces, and suppose that there exist
maximal ε-nets N (X, ε) and N (Y, ε) that are bi-Lipschitz equivalent
with constant (1 + ε). Then
dGH (X, Y ) ≤ ε max{diam(X), diam(Y )} + 2ε .
In fact, the proof of Lemma 2.6 shows that there is a metric in the
disjoint union Z = X  Y so that the Hausdorff distance between X, Y
in Z satisfies the preceding inequality. More precisely, the first part of
the proof of Lemma 2.6 goes unchanged, just take the infimum over all
a ∈ N (X, ε). To compute the Hausdorff distance, we argue
dGH (X, Y ) ≤ dGH (X, N (X, ε))
+ ε max{diam(X), diam(Y )} + dGH (N (Y, ε), Y ) .

The proof of the following result (stated for simplicity for compact
spaces) is a straightforward exercise.
18

2.8. Proposition. A sequence of compact spaces (Xi ) converges in

dGH if and only if there are isometries Ψi : Xi → l∞ such that

d∞
H (Ψi (Xi ), Z) → 0 , i → ∞,

for some compact subset Z ⊂ l∞ .

2.9. GCT for noncompact spaces. Recall that a metric space is

said to be proper if its closed balls are compact.
Let ((Xi , xi )) be a sequence of pointed proper spaces. We say that
((Xi , xi )) Gromov-Hausdorff converges to a (pointed proper) space
(Z, z) if
B(xi , R) → B(z, R)
in the Gromov-Hausdorff distance for each R > 0.
Here and later

B(x, R) = {y ∈ X : d(x, y) ≤ R}

denotes a closed ball.

The GCT together with a standard diagonalization argument gives
the following important result.

2.10. Theorem. A sequence of proper, uniformly doubling metric spaces

has a Gromov-Hausdorff convergent subsequent.

By a uniformly doubling sequence we mean a sequence of spaces that

are all doubling with the same constant (cf. 1.14).
In particular, we have the following corollary.

2.11. Theorem. Let (X, d) be a locally compact doubling space, and

let x0 ∈ X. Then every sequence

(εi ), εi → 0, contains a subsequence
−1
(εij ) such that the sequence X, εij d, x0 Gromov-Hausdorff converges
to some proper, doubling metric space (W, w).

2.12. Tangent spaces. A pointed space W = (W, w) as in 2.11 is

called a tangent space of X at x0 .
Tangent spaces reflect the local (infinitesimal) structure of X. The
concept of a tangent space is used in powerful ”blow-up” or rescaling
methods. For example, the infinitesimal behavior of functions often
translates into the global behavior of rescaled limit functions on tangent
spaces. In general, a space X as in 2.11 has many tangent spaces at a
given point x0 ∈ X.
Read more in [23], [15], [40], [41], [17], [18], [19].
 19

3. Some fundamental embeddings

We have seen that every separable metric space isometrically embeds
in l∞ . This Fréchet embedding has two shortcomings. First, if we want
to embed a given metric space “a point at a time”, it is not clear that
the Fréchet embedding helps here. For example, a given finite set
{x1 , x2 , . . . , xn } ⊂ X
embeds via
xj → (d(xj , x1 ) − d(x1 , x0 ), . . . , d(xj , xn ) − d(xn , x0 ), 0, . . .) ⊂ l∞ ,
while
{x1 , x2 , . . . , xn , xn+1 } ⊂ X
embeds via
xj → (d(xj , x1 ) − d(x1 , x0 ), . . . , d(xj , xn+1 ) − d(xn+1 , x0 ), 0, . . .) ⊂ l∞ ,
which is in general different.
Second, l∞ is not separable itself; it would be desirable to have a
separable Banach space as a universal target.
This section addresses the two shortcomings.
3.1. Urysohn universal space. A metric space (U, e) is called Urysohn
universal if it is separable and complete, and has the following property:
given any finite subspace X ⊂ U and any one point metric extension
of X,
(X ∗ , d∗ ) = (X ∪ {x∗ }, d∗ ) , d∗ |X × X = e ,
then there is a point Ψ(x∗ ) ∈ U such that
e (Ψ(x∗ ), x) = d∗ (x∗ , x)
for all x ∈ X.
The following property of Urysohn universal spaces is clear from the
definitions.
3.2. Proposition. Let U be Urysohn universal, and let
X = {x1 , x2 , . . . , xn , . . .}
be a countable metric space. Then every isometric embedding
Ψn : {x1 , x2 , . . . , xn } → U
can be extended to an isometric embedding
Ψn+1 : {x1 , x2 , . . . , xn , xn+1 } → U.
In particular, every separable metric space embeds isometrically in U .

The extension property for finite subspaces, stipulated in the defi-

nition for Urysohn universal spaces, is stronger than merely receiving
isometrically every separable metric space. In fact, we have the follow-
ing description of Urysohn universal spaces.
20

3.3. Proposition. Let U be a separable and complete metric space that

contains an isometric image of every separable metric space. Then U is
Urysohn universal if and only if U has the following transitivity prop-
erty: every isometry between finite subsets of U extends to an isometry
of U onto itself.

The proof of Proposition 3.3 is left to the reader.

The next theorem, due to Urysohn2 [73], guarantees that Urysohn
universal spaces exist.
3.4. Theorem. A Urysohn universal space exists. Moreover, up to
isometry there is only one such space.
Proof. We prove the existence of a Urysohn universal space. The second
assertion is left to the reader.
Given an arbitrary metric space X = (X, d), denote by E(X) the
collection of functions f : X → R that satisfy both
|f (x) − f (y)| ≤ d(x, y) ≤ f (x) + f (y)
whenever x, y ∈ X, and
f (x) = inf{d(x, y) + f (y) : y ∈ Y }
for all x ∈ X, where Y ⊂ X is a finite set (allowed to depend on f ).
A set Y as above is called a support of f . For example, if x0 ∈ X,
then the function
fx0 (x) := d(x0 , x)
belongs to E(X) with support Y = {x0 }.
Alternatively, E(X) can be described as the collection of all those 1-
Lipschitz functions f : X → R that are McShane-Whitney extensions
of 1-Lipschitz functions (see 1.19) defined on finite subsets (supports)
of X and satisfying the condition
d(x, y) ≤ f (x) + f (y)
for x and y in the support.
We equip E(X) with the sup-metric,
dE (f, g) := sup{|f (x) − g(x)| : x ∈ X} .
It follows from the definitions that dE is indeed a metric. Moreover,
we have that X embeds isometrically in (E(X), dE ) via
x0 → fx0 , fx0 (x) = d(x0 , x) .
Note that this embedding is canonical, cf. 1.7.
It is easy to check that
dE (f, x0 ) = dE (f, fx0 ) = f (x0 )
2Pavel
Samuilovich Urysohn (1898-1924) drowned in Brittany at the age of 26
while swimming. The cited article [73] was published posthumously.
 21

for each x0 ∈ X and f ∈ E(X), where we have identified x0 ∈ X and

the function fx0 .
If we denote by En (X) the subspace of E(X) consisting of those
functions in E(X) that have a support of cardinality at most n, then
X ⊂ E1 (X) ⊂ E2 (X) ⊂ . . . ,
and

∞
E(X) = En (X) .
n=1
Finally, starting with an arbitrary separable metric space X, we
define by induction
X0 := X , Xn+1 := E(Xn ) , n = 0, 1, 2, . . . .
We then claim that the metric completion X ∞ of the space

∞
X∞ := Xn
n=0

is Urysohn universal. Note that the metric d in X∞ is unambiguously

defined, as each Xn sits inside Xn+1 canonically and isometrically.
To prove the claim, we first observe that X ∞ is complete by defini-
tion; it is also separable because functions that take rational values in
their supports are dense in every E(X). Next, let {x1 , . . . , xn } be a
finite subset of X ∞ , and let d∗ be a metric in a set {x1 , . . . , xn , x∗ } satis-
fying d∗ (xi , xj ) = d(xi , xj ). Let f : X ∞ → R be the McShane-Whitney
extension of the function
xi → d∗ (xi , x∗ ) .
Next, fix 5 > 0 and choose points
x1 , . . . , xn ∈ Xm
such that
d(xi , xi ) < 5 .
Then the McShane-Whitney extension f : Xm → R of the function
xi → f (xi )
belongs to
E(Xm ) = Xm +1 ⊂ X ∞ .
It is easy to see from the definitions that y := (f ), 5 > 0, forms a
Cauchy sequence in X∞ , so that y ∈ X ∞ . Moreover,
d(y, xi ) = lim d(f , xi ) = f (xi ) = d∗ (xi , x∗ ) ,
→0

as required. This completes the proof of the theorem.

22

3.5. Remark. It is hard to find a proof of Urysohn’s theorem in mod-

ern sources. There is a proof in [33, pp. 78–79], but crucial details are
missing. (See however [29].)
Read more in [33], [29], [76].

The following theorem due to Banach [6] in 1932 provides a concrete

separable target for all separable spaces.

3.6. Theorem. Every separable metric space embeds isometrically in

C[0, 1].

Here C[0, 1] is the (separable) Banach space of continuous real-valued

functions on the closed unit interval equipped with the sup-norm.

3.7. Remark. The Banach space C[0, 1] cannot be Urysohn universal,

for it cannot satisfy the transitivity requirement given in Proposition
3.3. This is easily ascertained, by using the fact that every isometric
bijection between Banach spaces is an affine map [8, p. 341].

The proof of Theorem 3.6 requires the following auxiliary result,

interesting in its own right.

3.8. Theorem. Given a compact metric space X, there is a continuous

surjection from the Cantor set C to X.

In the statement, the Cantor set C is the standard 13 -Cantor set.

Alternatively, one can think of C as an arbitrary compact, perfect, and
totally disconnected space; then the Cantor set refers to the unique
such space, up to homeomorphism.
Read more in [23], [47, Chapter 2, Section 6], [44, Chapter XV,
Section 8].

Let us assume Theorem 3.8 for the moment.

Proof of Theorem 3.6. Let X be a separable metric space. By

the Fréchet embedding, we can assume that X ⊂ l∞ . Moreover, by
considering the completion of the span of a countable dense set in X,
we can assume that X itself is a separable Banach space. (Indeed, the
finite sums of the form

q x , q ∈ Q , x ∈ D,
where D ⊂ X is countable and dense, are dense in the span.)
By the Banach-Alaoglu theorem, the closed unit ball U of the dual
space X ∗ , equipped with the weak∗ -topology, is a compact metrizable
space [79, p. 32]. The separability of X is used to infer the last assertion
about metrizability.
 23

Recall that
U = {Λ : X → R linear with |Λ(x)| ≤ ||x||X for all x ∈ X},
and that Λi → Λ in the weak∗ -topology if and only if Λi (x) → Λ(x) for
all x ∈ X.
Next, let
h:C→U
be a surjective continuous map guaranteed by Theorem 3.8. We define
the map
X → C(C) , x(t) = h(t), x for t ∈ C ,
where x is identified with its image, and a bracket notation Λ, x =
Λ(x) is used for the dual action. Indeed, because X ⊂ X ∗∗ , we have
h x
C−
→U −
→R
so that the map X → C(C) is simply x → x ◦ h.
We observe
|x(t) − x (t)| = |h(t), x − x | ≤ ||x − x ||X
for each t ∈ C. On the other hand, given x, x ∈ X, x = x , there
exists, by the Hahn-Banach theorem, an element Λ ∈ U such that
Λ, x − x  = ||x − x ||X .
By surjectivity, Λ = h(t) for some t ∈ C, and we conclude that
||x − x ||C[0,1] = ||x − x ||X .
It follows that the embedding X → C(C) is isometric. By using the
linear extension in the complementary intervals of C, we further have
an isometric embedding
C(C) → C[0, 1] .
The proof is complete.

Proof of Theorem 3.8. Let X be a compact, and let
N1 = N (X, 2−1 ) = {x1 , x2 , . . . , xn1 }
be a maximal 2−1 -net in X. By repeating points, if necessary, we may
assume that
n1 = 2m1
for some integer m1 . Here and below we consider finite point sets on
X that are like maximal nets, but a point in the set may be counted
more than once; we call these sets nets as well.
Consider the m1 th stage in the Cantor construction. Thus, we have
n1 = 2m1 intervals of length 3−m1 ,
I11 , I21 , . . . , In11 .
Define
h1 : {I11 , I21 , . . . , In11 } → N1 , Ii1 → xi .
24

Next, extend N1 to a 2−2 -net N2 . We group the points in N2 into n1

disjoint groups,

N21 , N22 , . . . , N2n1 , N2i ⊂ B xi , 2−1 ,
each containing
n2 = 2m2
elements (counted with multiplicity). Then we subdivide each interval
Ii1 from the previous construction into n2 = 2m2 intervals of length
3−m1 −m2 ,
Ii21 , Ii22 , . . . , Ii2n2 ,
and define
h2 : {Ii21 , Ii22 , . . . , Ii2n2 } → N2 ,
where Iij2 maps to the corresponding point (depending on j) in N2i .
Continuing in this manner, we get (locally constant) maps
hk : Ck → Nk ,
where Ck is the (m1 + · · · + mk )th stage in the Cantor construction,
Nk is a maximal 2−k -net in X (with points counted with multiplicity).
Moreover hk has the property that
|hk (x) − hk+l (x)| ≤ 2 · 2−k .
In particular,
h(x) = lim hk (x)
 k→∞
exists for each x ∈ C = Ck . This map is clearly continuous (by con-
struction, it is a uniform limit of continuous functions) and surjective.
This proves the theorem.

3.9. Embeddings in finite dimensional spaces. Thus far we have
only studied embeddings of metric spaces in infinite dimensional (Ba-
nach) spaces. The case of finite dimensional targets is much harder,
even when we consider bi-Lipschitz embeddings in place of isometries.
It is a wide open problem to decide which metric spaces can be em-
bedded bi-Lipschitzly in some Euclidean space. We next discuss some
examples along these lines.
3.10. Theorem. Every closed (=compact without boundary) Riemann-
ian manifold embeds smoothly and bi-Lipschitzly in some Euclidean
space.
Proof. Cover X n by finitely many balls B1 , B2 , . . . , BN of radius ε > 0
such that for each i there exist 2-bi-Lipschitz diffeomorphisms
ϕi : 5Bi → ϕi (5Bi ) ⊂ Rn .
(See [39, Chapter 1, Section 6].) Without loss of generality we assume
that
|ϕi (x)| ≥ 1
 25

for all i and x ∈ 5Bi .

Let ui ∈ C0∞ (2Bi ) be such that 0 ≤ ui ≤ 1 and ui |Bi = 1, and let
vi ∈ C0∞ (5Bi ) be such that 0 ≤ vi ≤ 1 and vi |4Bi = 1. Then set

N 
N
ϕ(x) := ϕi (x) ui (x) + ϕi (x) vi (x) ,
i=1 i=1

where we think
ϕ : X → RnN × RnN = R2nN .
Obviously ϕ is smooth, and hence Lipschitz. We will show that ϕ is
bi-Lipschitz. To this end, an equivalent norm

2nN
|a| = |aj | , a = (a1 , a2 , . . . , a2nN ) ,
j=1

can be used in R2nN .

Let us assume first that d(x, y) > 3 ε. Then there exists i such that
ui (x) = 1 and ui (y) = 0. Thus,
|ϕ(x) − ϕ(y)| ≥ |ϕi (x) ui (x) − ϕi (y) ui (y)|
= |ϕi (x)|
d(x, y)
≥1≥ .
diam X
On the other hand, if d(x, y) ≤ 3 ε, then there exists i such that vi (x) =
1 = vi (y). Thus,
|ϕ(x) − ϕ(y)| ≥ |ϕi (x) − ϕi (y)| ≥ 12 d(x, y).
The theorem follows.

3.11. Remark. According to a celebrated and deep theorem of Nash
[53], every smooth Riemannian manifold, whether compact or not, em-
beds isometrically in some finite dimensional Euclidean space. Here,
however, the isometry has to be interpreted intrinsically; that is, the
image has to be considered as a submanifold of the surrounding Eu-
clidean space.
Read more in [34].

Partition of unity type arguments can be used more generally. The

following theorem is due to Aharoni [1]; the proof is due to Assouad
[4].
3.12. Theorem. There exists an absolute constant K ≤ 12 such that
every separable metric space embeds K-bi-Lipschitzly in the (separable)
Banach space c0 .

Recall that c0 is the Banach subspace of l∞ consisting of all bounded

sequences (xi ) with xi → 0 as i → ∞.
26

Proof. Fix x0 ∈ X, and write B(r) = B(x0 , r). We claim that there
exists a doubly infinite sequence of subsets
Qk,j ⊂ X , k ∈ Z , j ≥ 0 ,
satisfying the following:

(i) j≥0 Qk,j = X\B( 32 · 2−k ),
(ii) diam Qk,j ≤ 2 · 2−k ,
(iii) Given k ∈ Z and x ∈ X, B(x, 2−k−1 ) meets only finitely many
Qk,j .
Thus, the Qk,j ’s form a sort of locally finite tesselation of X\B( 32 · 2−k )
by sets with diameter not exeeding 2 · 2−k .
Assume, for a moment, the existence of the Qk,j ’s and define
ϕk,j (x) = max{12 · (2−k−1 − dist(x, Qk,j )) , 0) } .
Then
• ϕk,j is 12-Lipschitz.
• 0 ≤ ϕk,j ≤ 12 · 2−k−1 and ϕk,j |B(2−k−1 ) = 0.
• Given x ∈ X and ε > 0, then ϕk,j (x) > ε only for finitely many
k, j.
Now we index the standard basis of c0 by Z × {0, 1, 2, . . .}, to have
(ek,j ), and define

ϕ(x) := ϕk,j (x)ek,j , ϕ : X → c0 .
k,j

Clearly ϕ is 12-Lipschitz. We will show that ϕ is 12-bi-Lipschitz. Fix

x = y and choose k ∈ Z such that
3 · 2−k < d(x, y) ≤ 3 · 2 · 2−k .
We may assume that x ∈ B( 32 · 2−k ), so that x ∈ Qk,j for some j. Thus
ϕk,j (x) = 12 · 2−k−1 .
On the other hand,
d(y, Qk,j ) ≥ d(y, x) − diam Qk,j ≥ 3 · 2−k − 2 · 2−k = 2−k
so that
ϕk,j (y) = 0 .
Thus
12 · 2−k−1
|ϕ(x) − ϕ(y)| ≥ 12 · 2−k−1 ≥ d(x, y) = d(x, y) ,
3 · 2 · 2−k
which is more than sufficient.
It suffices to construct the sets Qk,j . Fix a countable dense set
{y0 , y1 , . . .} in X. Set

Qk,0 = B(y0 , 2−k )\B 32 · 2−k ,
Qk,j = B(yj , 2−k )\(B( 32 · 2−k ) ∪ Qk,0 ∪ · · · ∪ Qk,j−1 ) .
 27

Only property (iii) requires an argument: Fix k ∈ Z, x ∈ X, and con-

sider B(x, 2−k−1 ). Arbitrarily close to x there is yi , so that B(yi , 2−k ) ⊃
B(x, 2−k−1 ) and so that for no index j after i can Qk,j meet B(x, 2−k−1 ).
The proof is now complete.

3.13. Remark. An easy sharpening of the above argument shows that
one can take any K > 4 in Theorem 3.12. It is not known what the best
constant is. Aharoni [1] shows that for every K-bi-Lipschitz embedding
ϕ : l 1 → c0
one must have K ≥ 2, where l1 is the (separable) Banach space of
absolutely summable sequences.
3.14. Assouad embedding. As mentioned earlier, there is no known
characterization of metric spaces that admit bi-Lipschitz embedding
in some Euclidean space. The following theorem of Assouad [5], in
contrast, gives a beautiful characterization for doubling spaces in terms
of embeddings.
3.15. Theorem. A metric space (X, d) is doubling if and only if the
snowflake space (X, dε ) admits a bi-Lipschitz embedding in some Eu-
clidean space for every 0 < ε < 1.
3.16. Remarks. (a) There are examples (none of which are trivial)
of doubling spaces that do not embed bi-Lipschitzly in any Euclidean
space. Thus Theorem 3.15 is, in a sense, sharp.
Read more in [61], [62], [65], [48], [45], [46].

(b) Theorem 3.15 is sharp also in the sense that isometric embeddings
of doubling snowflake spaces in some Rn do not always exist. Consider,
for example, the space
1
X = ([0, 1], |t − s| 2 ).
Then there is an isometric embedding in a Hilbert space,
X → L2 [0, 1] , t → χ[0,t] ,
where χ· denotes the characteristic function of a set, but no such em-
bedding can be found in some Rn .
To see this, fix N ≥ 1. Consider N orthogonal vectors in X ⊂ L2 ,
f1 = χ[0, 1 ] , . . . , fi = χ[ i , i+1 ] , . . . , fN = χ[ N −1 ,1] .
N N N N

We observe that
√
1 2
||fi ||L2 = √ , ||fi − fj ||L2 = √ ,
N N
for i = j. If we now assume that the linear span of X inside L2 (after
an appropriate isometric embedding) has finite dimension n, then upon
rescaling the preceding norm expressions, we find points x1 , . . . , xN in
28
√
the unit sphere of Rn with pairwise distance at least 2. This is an
obvious contradiction for N  n.

(c) Assouad [5] has shown that ([0, 1], |t − s|ε ), 0 < ε < 1, embeds
bi-Lipschitzly in
R[ ε ]+1 .
1

This is sharp by a result of Väisälä [74]: the image of [0, 1] in Rn under

such a map must have Hausdorff dimension < n.

(d) One can embed every snowflaked Rn , (Rn , |x − y| ) with 0 < 5 <
1, in the Hilbert space l2 of square summable sequences. Indeed, by
setting √
x → fx , fx (t) := e −1x·t − 1 ,
we obtain an element fx of the Hilbert space L2 (Rn ; µ), where
dµ(t) := c(5, n) |t|−2−n dt , t ∈ Rn ,
and it is straightforward to check that
||fx − fy ||2L2 (Rn ;µ) = |x − y|2
under an appropriate choice for the constant c(5, n).
This fact was pointed out to me by Eero Saksman. It also follows
from more general, similar and rather elementary facts about positive
definite kernel functions on metric spaces given in [8, Chapter 8]. Defi-
nite functions on metric spaces were introduced long ago by Schoenberg
and von Neumann 3[58], [78], who studied isometric embeddability of
metric spaces in Hilbert spaces via Fourier analysis.

(e) The study of definite functions as in [8, Chapter 8] also leads to

the proof that every snowflaked (separable) Hilbert space (l2 , |x − y|2 )
embeds isometrically in l2 . One is led to wonder which Banach spaces
have the property that they isometrically or bi-Lipschitzly contain
snowflaked versions of themselves. Besides l2 , the space C[0, 1] has
this property (isometrically) by Theorem 3.6, and c0 (bi-Lipschitzly)
by Theorem 3.12. I do not know if this question has been studied in
the literature.
Proof of Theorem 3.15. Let N0 ⊂ X be a maximal 1-net. Then by
the doubling condition
#N0 ∩ B(x, 12) ≤ M
for all x ∈ X and for some M independent of x. We claim that there
exists a “coloring map”
κ : N0 → {e1 , . . . , eM } = standard basis of RM
3I thank Piotr Haj/lasz for these references.
 29

with
κ(a) = κ(b) if d(a, b) < 12 .
To see this, enumerate N0 = {a0 , a1 , . . .} and suppose that a map
κi : {a1 , . . . , ai } → {e1 , . . . eM }
has been defined so that the claim holds for κi . Because
#{a1 , . . . , ai } ∩ B(ai+1 , 12) ≤ M − 1 ,
we can assign a value κi+1 (ai+1 ) to extend κi so that the claim holds
for κi+1 .
We now define a map ϕ0 : X → RM ,

ϕ0 (x) = max{2 − d(x, ai ), 0} κ(ai ).
ai ∈N0

(Compare the proof of Theorem 3.12.)

Then
1
· 8 < d(x, y) ≤ 8
2
implies
|ϕ0 (x) − ϕ0 (y)| ≥ 1 ,
while
|ϕ0 (x) − ϕ0 (y)| ≤ C min{d(x, y), 1}
for every x, y ∈ X, with C ≥ 1 independent of x and y.
Now we also have that
#Nj ∩ B(x, 2−j · 12) ≤ M
for all x ∈ X, and j ∈ Z, where Nj is a maximal 2−j -net and we
similarly obtain maps
ϕ j : X → RM
with the property
|ϕj (x) − ϕj (y)| ≥ 1
if
2−j−1 · 8 < d(x, y) ≤ 2−j · 8 .
(This follows directly by applying the construction of ϕ0 to the metric
space (X, 2j · d).) Moreover,
|ϕj (x) − ϕj (y)| ≤ C min{2j d(x, y), 1} .
Next, consider R2N with the standard basis { ej } cyclically extended
to all j ∈ Z. We claim that the map

ϕ(x) := 2−εj ϕj (x) ⊗ ej
j∈Z

is bi-Lipschitz (X, dε ) → RM ⊗ R2N , provided N is large enough (de-

pending on the given data only).
Here we have normalized (as we may) the maps ϕj so that ϕj (x0 ) = 0
for a fixed base point x0 ∈ X.
30

First note that the above series converges, because


|ϕ(x)| = |ϕ(x) − ϕ(x0 )| ≤ C 2−εj min{2j d(x, x0 ), 1} < ∞ .
j∈Z

Next, fix x, y ∈ X and choose k ∈ Z such that

2−k−1 · 8 < d(x, y) ≤ 2−k · 8 .
Then
 
|ϕ(x) − ϕ(y)| ≤ 2−εj |ϕj (x) − ϕj (y)| + 2−εj |ϕj (x) − ϕj (y)|
j≥k+1 j≤k
−εk
≤ C · (2 +2 k(1−ε)
d(x, y))
≤ C · d(x, y) . ε

Moreover,

|ϕ(x) − ϕ(y)| ≥ | 2−εj (ϕj (x) − ϕj (y)) ⊗ ej |
−N +k<j≤N +k

− 2−εj |ϕj (x) − ϕj (y)|
j>N +k

− 2−εj |ϕj (x) − ϕj (y)|
j≤−N +k

≥ 2−εk |ϕk (x) − ϕk (y)| − c · 2−ε(N +k)

− c · 2−ε(−N +k) 2−N +k d(x, y)
1
≥ 2−εk |ϕk (x) − ϕk (y)| − · 2−εk
2
−εk
≥ c2 ≥ c · d(x, y) ,
ε

provided N is large enough. This completes the proof.

4. Strong A∞ -deformations
In this section, we consider metric spaces that result when the stan-
dard Euclidean metric on Rn is deformed by a weight. For smooth
positive weights this procedure is called a conformal deformation; in
our case, the weight is allowed to have singularities in a controlled
manner. We follow David and Semmes [22], [59].

4.1. Metric doubling measures. Let µ be a nontrivial doubling

measure in Rn (see 1.17 (c)). For x, y ∈ Rn set
dµ (x, y) := µ(Bx,y )1/n ,
where
Bx,y := B(x, |x − y|) ∪ B(y, |x − y|) .
 31

Then dµ satisfies all the axioms for a metric on Rn except perhaps the
triangle inequality. We call µ a metric doubling measure if there exists
a constant C ≥ 1 such that

N
dµ (x, y) ≤ C inf dµ (xi , xi+1 ) ,
i=0

where the infimum is taken over all sequences of points

x = x0 , x1 , . . . , xN +1 = y
in Rn . If µ is a metric doubling measure, then the above infimum
determines a metric δµ on Rn that satisfies
δµ (x, y) ≤ dµ (x, y) ≤ C δµ (x, y)
whenever x, y ∈ Rn .
It is easy to see that every doubling measure on R is a metric doubling
measure. For n ≥ 2, metric doubling measures possess subtle geometric
and analytic features; in particular, not every doubling measure is a
metric doubling measure.
Read more in [63], [61], [62].

4.2. Strong A∞ weights. The following theorem was proved by David

and Semmes in [22].

4.3. Theorem. Every metric doubling measure µ on Rn , n ≥ 2, is of

the form dµ = w dx, where w is an A∞ weight.

We omit the proof. The idea behind this result goes back to Gehring’s
fundamental work [30].
A nonnegative locally integrable function w on Rn is called a weight.
A weight w is an A∞ weight if there exist a constant C ≥ 1 and a
positive number 5 > 0 such that

1  1/(1+) 1

1+
w dx ≤ C w dx
|B| B |B| B
for every ball B ⊂ Rn . Here and hereafter,

|E| := dx
E

denotes the Lebesgue measure of a measurable set.

The A∞ weights that arise as densities of metric doubling measures
are called strong A∞ weights in the literature. There are many inter-
esting open questions about strong A∞ -weights.
Read more in [64], [37], [10].
32

4.4. A1 -weights. A weight w on Rn is called an A1 weight if there

exists a constant C ≥ 1 such that

1
w dx ≤ C essinfB w
|B| B
for every ball B ⊂ Rn . It is not difficult to see that every A1 weight is
a strong A∞ weight. Let δw denote the corresponding metric. Thus,
δw (x, y) ≤ w(Bx,y )1/n ≤ C δw (x, y) ,
for some constant C ≥ 1 independent of x and y. Here and hereafter,
we write 
w(E) := w dx ,
E
if w is a weight on Rn .
We will prove the following theorem of Semmes [59].
4.5. Theorem. Let w be an A1 weight on Rn , n ≥ 2. Then the metric
space (Rn , δw ) admits a bi-Lipschitz embedding in some RN .

It is not known how low the receiving dimension N in Theorem 4.5

can be.
4.6. Open problem. Let w be an A1 weight on Rn , n ≥ 2. Is (Rn , δw )
bi-Lipschitz homeomorphic to Rn ?

It even seems to be an open problem if one can choose N such that

it depends only on n and on the constant in the A1 condition. See [59,
p. 228].
4.7. Remark. (a) As explained in [22], [59], [10], an affirmative answer
to the question in 4.6 would imply that with every A1 weight w on Rn
there is associated a quasiconformal homeomorphism f : Rn → Rn
such that
C −1 Jf (x) ≤ w(x) ≤ C Jf (x)
for almost every x ∈ Rn , for some constant C ≥ 1. Here Jf denotes
the Jacobian determinant of the derivative matrix of f .
A theorem of Coifman and Rochberg [20], [68, p. 214] asserts that
for every Radon measure ν on Rn , and for every 0 < δ < 1, the function
w := (M ν)δ
is an A1 weight provided it is finite almost everywhere, where M ν
denotes the Hardy-Littlewood maximal function of the measure ν. In
fact, all A1 weights essentially arise in this manner.
It follows from the Coifman-Rochberg theorem that given any set of
Lebesgue measure zero in Rn , there is an A1 weight that takes the value
∞ (in an appropriate sense) on that set. Thus, an affirmative answer
to the above question would imply that given any set of measure zero
 33

in Rn , there exists a quasiconformal homeomorphism f : Rn → Rn

with Jacobian determinant infinite (in an appropriate sense) on that
set. Theorem 4.5 implies that a mapping with such Jacobian always
can be found, if we allow for a larger class of mappings with range in
some higher dimensional Euclidean space.

(b) The metric space (Rn , δw ), and all its bi-Lipschitz images, pos-
sesses many good geometric properties, if w is a strong A∞ weight.
For example, it is easy to see that w determines an Ahlfors n-regular
measure on the space, which means by definition that
C −1 rn ≤ w(Bw (r)) ≤ Crn
for every metric ball Bw (r) of radius r > 0 in (Rn , δw ), for some con-
stant C ≥ 1. In particular, (Rn , δw ) has Hausdorff dimension n and
the n-dimensional Hausdorff measure satisfies an Ahlfors n-regularity
condition as well.

(c) There exist strong A∞ weights w on Rn for each n ≥ 2 such that

the corresponding space (Rn , δw ) is not bi-Lipschitz embeddable in any
Hilbert space, or even in any uniformly convex Banach space. See [46],
[61].

Proof of Theorem 4.5. We give a reasonably detailed proof, but

leave some straightforward checking for the reader. I would like to
point out that Semmes proves in [59] a more general theorem, where
the conclusion of Theorem 4.5 is established under the hypothesis that
w satisfies a stability property in terms of taking powers. A1 weights
have such a stability property. We forgo the general case here.
We now fix an A1 weight w in Rn , n ≥ 2, and let C ≥ 1 denote a
generic positive constant that only depends on n and on w.
The A1 condition readily implies that w satisfies

∗ 1
w(x) ≤ w (x) := sup w dx ≤ C w(x)
B,x∈B |B| B

for almost every x. Here we use the uncentered Hardy-Littlewood

maximal function in the middle (see [67, Ch. I]). This minor smoothing
of w is required only to have the sets {w∗ > λ} open for each λ ∈ R.
We could also work with convolutions w of w, and obtain estimates
independent of 5. It is somewhat more elementary to use the maximal
function, especially because we only use the definition and some of its
direct consequences.
Next, we fix a large real number L > 0, to be determined later so as
to depend only on n and w. Then set
Ωj := {x ∈ Rn : w∗ (x) > Ljn } , j ∈ Z,
34

and define
j0 := inf{j : Ωj+1 = Rn } .
It may happen that j0 = −∞.
As in the proof of Assouad’s Theorem 3.15, the desired embedding
will be defined rather explicitely by using locally defined building blocks
that are maps to some Euclidean space. The A1 condition is used to
control the resulting infinite sum. For the building blocks, we employ
a Whitney decomposition in the open sets Ωj .
We divide each Ωj , j > j0 , into countably many pairwise essentially
disjoint cubes {Qj,k } such that
C −1 dist(Qj,k , Rn \ Ωj ) ≤ diam(Qj,k ) ≤ 5 dist(Qj,k , Rn \ Ωj ) ,
1
where 0 < 5 < 100 is a small constant to be chosen at a later stage.
Such a decomposition can be found by appropriately subdividing the
cubes in the standard Whitney decomposition (as constructed in [67,
p. 16], for example).
If j0 > −∞, then define

w(B) 1/n
M0 := inf ,
|B|
where the infimum is taken over all balls B such that
B ∩ (Rn \ Ωj0 +1 ) = ∅ .
Next, for j > j0 , set

w(Q ) 1/n
j,k
Mj,k := .
|Qj,k |
Note that Mj,k is roughly the distance distortion between the metric
δw and the Euclidean metric at the scale of the cube Qj,k .
Fix Lipschitz bump functions θj,k , with Lipschitz constant not ex-
ceeding 2(diam(Qj,k ))−1 , such that
0 ≤ θj,k ≤ 1 , θj,k |2Qj,k = 1 , θj,k |Rn \ 3Qj,k = 0 .
By setting 
ϕj,k := θj,k ( θj,l )−1 ,
l
we obtain a Lipschitz partition of unity for Ωj ,

ϕj,k ≡ 1
k

with the Lipschitz condition (cf. 1.12)

|ϕj,k (x) − ϕj,k (y)| ≤ C (diam(Qj,k ))−1 |x − y| , x, y ∈ Rn .
Further auxiliary mappings are defined as follows. Denote by qj,k the
center point of the cube Qj,k . If j0 > −∞, then define
f0 := M0 x , if x ∈ Rn \ Ωj0 +1 ,
 35

and 
f0 (x) := M0 ϕj0 +1,k (x) qj0 +1,k , if x ∈ Ωj0 +1 .
k
If j0 = −∞, then define f0 ≡ 0. For j > j0 , set
hj,k (x) := Mj,k (x − qj,k ) , if x ∈ 3Qj,k \ Ωj+1 ,
and

hj,k (x) := Mj,k ϕj+1,l (x) (qj+1,l − qj,k ) , if x ∈ 3Qj,k ∩ Ωj+1 ,
l

and then
fj,k := θj,k hj,k .
We understand that fj,k is everywhere defined; it vanishes outside 3Qj,k .
Next we record some estimates for the given functions. Define
dj (x) := dist(x, Rn \ Ωj ) .
In the proof of the ensuing proposition, we use the estimate
M0 ≤ Lj0+1 , if j0 > −∞ ,
which easily follows from the definitions.
4.8. Proposition.
(1) |f0 (x) − M0 x| ≤ C M0 5 dj0 +1 (x) , if j0 > −∞ ,
(2) |f0 (x) − f0 (y)| ≤ C Lj0 +1 |x − y| ,
(3) |fj,k (x)| ≤ C w(Qj,k )1/n ,
(4) |fj,k (x) − fj,k (y)| ≤ C Lj+2 |x − y|
for every x, y ∈ Rn and j, k.
Proof. To prove (1) and (2), we may assume that x ∈ Ωj0 +1 . Then

|f0 (x) − M0 x| = M0 | ϕj0 +1,k (x) (qj0 +1,k − x)|
k
≤ C M0 5 dj0 +1 (x) ,

which gives (1).

To prove (2), we consider first the case when
1
|x − y| > dj +1 (x) .
10 0
Then, by using (1),
|f0 (x) − f0 (y)| ≤ |f0 (x) − M0 x| + M0 |x − y| + |f0 (y) − M0 y|
≤ C M0 |x − y| .
36

Thus (2) follows in this case. We also have that


|f0 (x) − f0 (y)| ≤ M0 | (ϕj0 +1,k (x) − ϕj0 +1,k (y))(qj0 +1,k − x)| ;
k

and because only a fixed number of cubes 3Qj0 +1,k meet x, we easily
see that the last quantity is less than
C M0 |x − y| ,
provided
1
|x − y| ≤ dj +1 (x) .
10 0
Thus (2) follows in all cases.
Inequality (3) follows from the definitions and from the fact that
Mj,k ≤ C w(Qj,k )1/n (diam(Qj,k ))−1 .

Estimate (4) requires a little more work. We need the following

lemma, which is a rather straightforward consequence of the A1 con-
dition and Jensen’s inequality. We leave the proof to the reader (who
may also consult [59, Lemma 7.14]).

4.9. Lemma. Let Q, Q be two cubes such that Q ⊂ Ωj , that

Q ∩ (Q \ Ωj+1 ) = ∅ ,
and that Q ⊂ 100Q. Then
 
1 1
|  log w dx − log w dx| ≤ n log L + C .
|Q | Q |Q| Q

Now to prove (4), we observe that analogously to (1) and (2), we

have the following estimates:
|hj,k (x) − Mj,k (x − qj,k )| ≤ C Mj,k 5 dj+1 (x) ,
and
|hj,k (x) − hj,k (y)| ≤ C Lj+2 |x − y| .
The proof of the first inequality if analogous to the proof of (1), and,
armed with Lemma 4.9, the proof of the second inequality is analogous
to the proof of (2).
Therefore, by using
|fj,k (x)−fj,k (y)| ≤ |θj,k (x)||hj,k (x)−hj,k (y)|+|hj,k (y)||θj,k (x)−θj,k (y)| ,
we only need the estimate
|hj,k (y)| ≤ C w(Qj,k )1/n ≤ C Lj (diam(Qj,k )) ,
which follows from the definitions. The proposition is proved.

 37

We consider an integer M = m1 · m2 , where the integers m1 and m2

will be selected momentarily (and separately).
Arguing as in the proof of Theorem 3.15, we can find a map
κ : {Qj,k } → {1, 2, . . . , M }
such that
κ(Qj,k ) = κ(Qj  ,k )
implies both that j ≡ j  modm1 , and that k = k  if also j = j  and
1
dist(3Qj,k , 3Qj  ,k ) < dist(Qj,k ∪ Qj  ,k , Rn \ Ωj ) .
2
The integer m2 is chosen here large enough so that the second of these
implications is possible. Then define, for i = 1, 2, . . . , M ,

fi := fj,k
{(j,k):κ(Qj,k )=i}

and 
gi := gj,k ,
{(j,k):κ(Qj,k )=i}

where
gj,k := Lj dist(x, Rn \ 2Qj,k ) .
We claim that
F := (f0 , f1 , . . . , fM , g1 , . . . , gM ) , F : (Rn , δw ) → R2M +1 ,
is a bi-Lipschitz embedding.
We first show that F is Lipschitz. By Proposition 4.8, estimate (2),
f0 is Lipschitz. To handle the remaining components, it is convenient to
consider a fixed pair (fi , gi ). Given an integer j1 > j0 , we use estimate
(4) in Proposition 4.8 and observe that each term in the sum
 
(|fj,k (x) − fj,k (y)| + |gj,k (x) − gj,k (y)|)
j0 <j≤j1 k

is bounded by
C Lj+2 |x − y| .
Because both fj,k and gj,k are supported in the cube 3Qj,k , we also find
that only a bounded number of terms are nonzero, for each given j.
This implies that the sum has a bound
C Lj1 +2 |x − y|
for large enough L.
Next, we infer from estimate (3) in Proposition 4.8, and from the
definitions, that the sum

(|fj,k (x)| + |gj,k (x)|)
j1 <j k
38

has a bound 
C w(Qj,k )1/n ,
j1 <j

where Qj,k ⊂ Ωj is a cube such that x ∈ 3Qj,k . Now the definition for
Ωj implies that
dist(x, Rn \ Ωj+1 ) ≤ C L−1 dist(x, Rn \ Ωj ) ,
which in turn gives that
w(Qj+1,k ) ≤ C L−α w(Qj,k )
if x ∈ Qj+1,k ⊂ Qj,k , where α > 0 depends only on n and w. (This
latter claim follows from the doubling property of w.) Therefore,

(|fj,k (x)| + |gj,k (x)|) ≤ C w(Qj1 +1,k )1/n ,
j1 <j k

where either x ∈ Qj1 +1,k or x ∈

/ Ωj1 +1 .
To prove the Lipschitz condition for the pair (fi , gi ), we use the
preceding two bounds as follows. Fix x, y ∈ Rn and let j1 be the
largest integer j for which B(x, 10|x − y|) ⊂ Ωj . By splitting the sums
involving fj,k and gj,k as above, with this choice of j1 , we obtain that
|fi (x) − fi (y)| + |gi (x) − gi (y)| ≤ C Lj1 +2 |x − y| + C δw (x, y) .
On the other hand,
Lj ≤ C (w(Bx,y ))1/n |x − y|−1
in this case, whence the Lipschitz condition follows.
Notice that for the Lipschitz condition, we did not need the coloring
map κ which was used to split the fj,k ’s into different generations. This
splitting is used next in the proof of the lower estimate
|F (x) − F (y)| ≥ C −1 δw (x, y) , x, y ∈ Rn .
To that end, fix x, y, and define j1 as above. In particular, we have
that
1
|x − y| > dj1 +1 (x) .
10
If j1 = j0 > −∞, then
|F (x) − F (y)| ≥ |f0 (x) − f0 (y)|
≥ M0 |x − y| − |f0 (x) − M0 x| − |f0 (y) − M0 y|


≥ M0 |x − y| − C 5 (dj0 +1 (x) + dj0 +1 (y))
1
≥ M0 |x − y| ,
2
 39

by estimate (1) in Proposition 4.8, provided 5 is chosen small enough.

It is also easy to see, by employing Lemma 4.9, that
M0 |x − y| ≥ C δw (x, y) ,
which gives the claim if j1 = j0 .
Assume next that j1 > j0 . Fix a cube Qj1 ,k1 that contains x, and
let i = κ(Qj1 ,k1 ) be the color of the cube. We need to distinguish two
cases depending upon whether y is in 2Qj1 ,k1 or not. Suppose first that
y ∈ 2Qj1 ,k1 . In this case,
|F (x) − F (y)| ≥ |fi (x) − fi (y)|

≥ |fj1 ,k1 (x) − fj1 ,k1 (y)| − |fj,k (x) − fj,k (y)| ,
(j,k)∈A

where
A = {(j, k) : (j, k) = (j1 , k1 ) , κ(Qj,k ) = i} .
To estimate the first term on the right hand side, we observe that
θj1 ,k1 (x) = θj1 ,k1 (y) = 1, which gives
|fj1 ,k1 (x) − fj1 ,k1 (y)| = |hj1 ,k1 (x) − hj1 ,k1 (y)|
≥ Mj1 ,k1 |x − y| − |hj1 ,k1 (x) − Mj1 ,k1 (x − qj1 ,k1 )|
− |hj1 ,k1 (y) − Mj1 ,k1 (y − qj1 ,k1 )|


≥ Mj1 ,k1 |x − y| − C 5 (dj1 +1 (x) + dj1 +1 (y))
1
≥
Mj ,k |x − y| .
2 1 1
It is again easy to see from the basic (doubling) properties of w, by
employing Lemma 4.9, that
Mj1 ,k1 |x − y| ≥ C −1 L−1 δw (x, y) ,
and it suffices to show that the sum

|fj,k (x) − fj,k (y)|
(j,k)∈A

is small compared to C −1 L−1 δw (x, y). To do so, observe first that the
indices of the type (j1 , k) do not contribute to the sum, which therefore
splits into two sums
  
:= |fj,k (x) − fj,k (y)|
1 m1 +j1 ≤j k

and   
:= |fj,k (x) − fj,k (y)| .
2 j0 <j≤j1 −m1 k
40

By now it is straightforward to verify that

 

−α m1 −1 −m1
+ ≤ C (C L ) +L δw (x, y) ,
1 2
which means that a large enough choice for m1 gives
|F (x) − F (y)| ≥ |fj1 ,k1 (x) − fj1 ,k1 (y)| ≥ C −1 L−1 δw (x, y) .
Finally, we assume that y ∈
/ 2Qj1 ,k1 . This time we estimate
|F (x) − F (y)| ≥ |gi (x) − gi (y)|

≥ |gj1 ,k1 (x) − gj1 ,k1 (y)| − |gj,k (x) − gj,k (y)| ,
(j,k)∈A

where A is as before. Arguing as above, and using the definition for

the functions gj,k , it is easily ascertained that
|gj1 ,k1 (x) − gj1 ,k1 (y)| ≥ C −1 δw (x, y) ,
and that the sum 
|gj,k (x) − gj,k (y)|
(j,k)∈A
is small in comparison.
This completes the proof of Theorem 4.5.

References
[1] Aharoni, I. Every separable metric space is Lipschitz equivalent to a subset
of c+
0 . Israel J. Math. 19 (1974), 284–291.
[2] Aleksandrov, A. D., and Zalgaller, V. A. Intrinsic geometry of sur-
faces. Translated from the Russian by J. M. Danskin. Translations of Math-
ematical Monographs, Vol. 15. American Mathematical Society, Providence,
R.I., 1967.
[3] Alestalo, P., and Väisälä, J. Quasisymmetric embeddings of products of
cells into the Euclidean space. Ann. Acad. Sci. Fenn. Ser. A I Math. 19 (1994),
375–392.
[4] Assouad, P. Remarques sur un article de Israel Aharoni sur les prolongements
lipschitziens dans c0 (Israel J. Math. 19 (1974), 284–291). Israel J. Math. 31,
1 (1978), 97–100.
[5] Assouad, P. Plongements lipschitziens dans Rn . Bull. Soc. Math. France 111
(1983), 429–448.
[6] Banach, S. Théorie des óperations linéaires. Chelsea Publishing Co., New
York, 1955.
[7] Bellaı̈che, A. The tangent space in sub-Riemannian geometry. In Sub-
Riemannian geometry. Birkhäuser, Basel, 1996, pp. 1–78.
[8] Benyamini, Y., and Lindenstrauss, J. Geometric Nonlinear Functional
Analysis, Volume I, vol. 48 of Colloquium Publications. Amer. Math. Soc.,
2000.
[9] Bonk, M., and Eremenko, A. Uniformly hyperbolic surfaces. Indiana Univ.
Math. J. 49, 1 (2000), 61–80.
[10] Bonk, M., Heinonen, J., and Saksman, E. The quasiconformal jacobian
problem. Proceeding of the 2001 Ahlfors-Bers Colloquium (to appear).
 41

[11] Bonk, M., and Lang, U. Bi-lipschitz parameterization of surfaces. Math.

Ann. (to appear).
[12] Buckley, S. M., Hanson, B., and MacManus, P. Doubling for general
sets. Math. Scand. 88, 2 (2001), 229–245.
[13] Burago, D., Burago, Y., and Ivanov, S. A course in metric geometry,
vol. 33 of Graduate Studies in Mathematics. American Mathematical Society,
Providence, RI, 2001.
[14] Burago, Y., Gromov, M., and Perel
man, G. A. D. Aleksandrov spaces
with curvatures bounded below. Uspekhi Mat. Nauk 47, 2(284) (1992), 3–51,
222.
[15] Cheeger, J. Differentiability of Lipschitz functions on metric measure spaces.
Geom. Funct. Anal. 9 (1999), 428–517.
[16] Cheeger, J. Degeneration of Riemannian Metrics under Ricci Curvature
Bounds. Scuola Normale Superiore, 2001. Lezioni Fermiane, Pisa 2001.
[17] Cheeger, J., and Colding, T. H. On the structure of spaces with Ricci
curvature bounded below. I. J. Differential Geom. 46, 3 (1997), 406–480.
[18] Cheeger, J., and Colding, T. H. On the structure of spaces with Ricci
curvature bounded below. II. J. Differential Geom. 54, 1 (2000), 13–35.
[19] Cheeger, J., and Colding, T. H. On the structure of spaces with Ricci
curvature bounded below. III. J. Differential Geom. 54, 1 (2000), 37–74.
[20] Coifman, R. R., and Rochberg, R. Another characterization of BMO.
Proc. Amer. Math. Soc. 79, 2 (1980), 249–254.
[21] Colding, T. H. Spaces with Ricci curvature bounds. In Proceedings of the In-
ternational Congress of Mathematicians, Vol. II (Berlin, 1998) (1998), no. Ex-
tra Vol. II, pp. 299–308 (electronic).
[22] David, G., and Semmes, S. Strong A∞ weights, Sobolev inequalities
and quasiconformal mappings. In Analysis and partial differential equations,
vol. 122 of Lecture Notes in Pure and Appl. Math. Marcel Dekker, 1990,
pp. 101–111.
[23] David, G., and Semmes, S. Fractured fractals and broken dreams: self-
similar geometry through metric and measure, vol. 7 of Oxford Lecture Series
in Mathematics and its Applications. Clarendon Press Oxford University Press,
1997.
[24] Dorronsoro, J. R. A characterization of potential spaces. Proc. Amer. Math.
Soc. 95, 1 (1985), 21–31.
[25] Evans, L. C., and Gariepy, R. F. Measure Theory and Fine Properties
of Functions. Studies in Advanced Mathematics. CRC Press, Boca Raton,
Florida, 1992.
[26] Falconer, K. J. Fractal geometry. Mathematical Foundations and Applica-
tions. John Wiley and Sons Ltd., Chichester, 1990.
[27] Federer, H. Geometric Measure Theory, vol. 153 of Die Grundlehren der
mathematischen Wissenschaften. Springer-Verlag, New York, 1969.
[28] Fréchet, M. Les dimensions d’un ensemble abstrait. Math. Ann. 68 (1909–
10), 145–168.
[29] Gao, S., and Kechris, A. S. On the classification of Polish metric spaces
up to isometry. Mem. Amer. Math. Soc. 161, 766 (2003), viii+78.
[30] Gehring, F. W. The Lp -integrability of the partial derivatives of a quasicon-
formal mapping. Acta Math. 130 (1973), 265–277.
[31] Gromov, M. Hyperbolic groups. In Essays in Group Theory, S. Gersten,
Editor. MSRI Publications, Springer-Verlag, 1987, pp. 75–265.
42

[32] Gromov, M. Carnot-Carathéodory spaces seen from within. In Sub-

Riemannian Geometry, vol. 144 of Progress in Mathematics. Birkhäuser, Basel,
1996, pp. 79–323.
[33] Gromov, M. Metric structures for Riemannian and non-Riemannian spaces,
vol. 152 of Progress in Mathematics. Birkhäuser Boston Inc., Boston, MA,
1999. Based on the 1981 French original, With appendices by M. Katz, P.
Pansu and S. Semmes, Translated from the French by Sean Michael Bates.
[34] Gromov, M. L., and Rohlin, V. A. Imbeddings and immersions in Rie-
mannian geometry. Uspehi Mat. Nauk 25, 5 (155) (1970), 3–62.
[35] Haj4lasz, P. Sobolev spaces on an arbitrary metric space. Potential Anal. 5
(1996), 403–415.
[36] Heinonen, J. Calculus on Carnot groups. In Fall School in Analysis
(Jyväskylä, 1994), vol. 68. Ber. Univ. Jyväskylä Math. Inst., Jyväskylä, 1995,
pp. 1–31.
[37] Heinonen, J. Lectures on analysis on metric spaces. Springer-Verlag, New
York, 2001.
[38] Heinonen, J., and Koskela, P. Quasiconformal maps in metric spaces with
controlled geometry. Acta Math. 181 (1998), 1–61.
[39] Helgason, S. Differential geometry, Lie groups, and symmetric spaces, vol. 34
of Graduate Studies in Mathematics. American Mathematical Society, Provi-
dence, RI, 2001. Corrected reprint of the 1978 original.
[40] Keith, S. A differentiable structure for metric measure spaces. Advances in
Mathematics (to appear).
[41] Keith, S., and Laakso, T. J. Conformal assouad dimension and modulus.
Preprint (2003).
[42] Kigami, J. Analysis on fractals, vol. 143 of Cambridge Tracts in Mathematics.
Cambridge University Press, Cambridge, 2001.
[43] Kuratowski, C. Quelques problèmes concernant les espaces métriques non-
séparables. Fund. Math. 25 (1935), 534–545.
[44] Kuratowski, K. Introduction to set theory and topology, english ed. Perga-
mon Press, Oxford, 1972. Containing a Supplement on ‘Elements of algebraic
topology’ by Ryszard Engelking, First edition translated from the Polish by
Leo F. Boron, International Series of Monographs in Pure and Applied Math-
ematics, Vol. 101.
[45] Laakso, T. Ahlfors Q-regular spaces with arbitrary Q admitting weak
Poincaré inequalities. Geom. Funct. Anal. 10 (2000), 111–123.
[46] Laakso, T. J. Plane with A∞ -weighted metric not bi-Lipschitz embeddable
to RN . Bull. London Math. Soc. 34, 6 (2002), 667–676.
[47] Lacey, H. E. The isometric theory of classical Banach spaces. Springer-
Verlag, New York, 1974. Die Grundlehren der mathematischen Wissenschaften,
Band 208.
[48] Lang, U., and Plaut, C. Bilipschitz embeddings of metric spaces into space
forms. Geom. Dedicata 87, 1-3 (2001), 285–307.
[49] Luosto, K. Ultrametric spaces bi-Lipschitz embeddable in Rn . Fund. Math.
150 (1996), 25–42.
[50] Luukkainen, J. Assouad dimension: antifractal metrization, porous sets, and
homogeneous measures. J. Korean Math. Soc. 35 (1998), 23–76.
[51] Luukkainen, J., and Saksman, E. Every complete doubling metric space
carries a doubling measure. Proc. Amer. Math. Soc. 126 (1998), 531–534.
[52] Mattila, P. Geometry of sets and measures in Euclidean spaces, vol. 44
of Cambridge Studies in Advanced Mathematics. Cambridge University Press,
Cambridge, 1995.
 43

[53] Nash, J. The imbedding problem for Riemannian manifolds. Ann. of Math.
(2) 63 (1956), 20–63.
[54] Pansu, P. Dimension conforme et sphère à l’infini des variétés à courbure
négative. Ann. Acad. Sci. Fenn. Ser. A I Math. 14 (1989), 177–212.
[55] Perelman, G. Spaces with curvature bounded below. In Proceedings of the
International Congress of Mathematicians, Vol. 1, 2 (Zürich, 1994) (Basel,
1995), Birkhäuser, pp. 517–525.
[56] Petersen, P. Riemannian geometry, vol. 171 of Graduate Texts in Mathe-
matics. Springer-Verlag, New York, 1998.
[57] Saksman, E. Remarks on the nonexistence of doubling measures. Ann. Acad.
Sci. Fenn. Ser. A I Math. 24, 1 (1999), 155–163.
[58] Schoenberg, I. J. Metric spaces and positive definite functions. Trans. Amer.
Math. Soc. 44, 3 (1938), 522–536.
[59] Semmes, S. Bi-Lipschitz mappings and strong A∞ weights. Ann. Acad. Sci.
Fenn. Ser. A I Math. 18 (1993), 211–248.
[60] Semmes, S. Finding curves on general spaces through quantitative topology,
with applications to Sobolev and Poincaré inequalities. Selecta Math. 2 (1996),
155–295.
[61] Semmes, S. Good metric spaces without good parameterizations. Rev. Mat.
Iberoamericana 12 (1996), 187–275.
[62] Semmes, S. On the nonexistence of bi-Lipschitz parameterizations and geo-
metric problems about A∞ -weights. Rev. Mat. Iberoamericana 12 (1996), 337–
410.
[63] Semmes, S. Bilipschitz embeddings of metric spaces into Euclidean spaces.
Publ. Mat. 43, 2 (1999), 571–653.
[64] Semmes, S. Metric spaces and mappings seen at many scales (appendix).
Birkhäuser Boston Inc., Boston, MA, 1999. Based on the 1981 French orig-
inal, Translated from the French by Sean Michael Bates.
[65] Semmes, S. Some novel types of fractal geometry. Oxford Mathematical Mono-
graphs. The Clarendon Press Oxford University Press, New York, 2001.
[66] Shanmugalingam, N. Newtonian spaces: an extension of Sobolev spaces to
metric measure spaces. Rev. Mat. Iberoamericana 16, 2 (2000), 243–279.
[67] Stein, E. M. Singular integrals and differentiability properties of functions.
Princeton University Press, Princeton, N.J., 1970. Princeton Mathematical
Series, No. 30.
[68] Stein, E. M. Harmonic analysis: real-variable methods, orthogonality, and
oscillatory integrals. Princeton Univ. Press, Princeton, NJ, 1993. With the
assistance of Timothy S. Murphy, Monographs in Harmonic Analysis, III.
[69] Strichartz, R. S. Sub-Riemannian geometry. J. Diff. Geom. 24 (1986), 221–
263.
[70] Toro, T. Surfaces with generalized second fundamental form in L2 are Lips-
chitz manifolds. J. Differential Geom. 39, 1 (1994), 65–101.
[71] Toro, T. Geometric conditions and existence of bi-Lipschitz parameteriza-
tions. Duke Math. J. 77, 1 (1995), 193–227.
[72] Tukia, P. A quasiconformal group not isomorphic to a Möbius group. Ann.
Acad. Sci. Fenn. Ser. A I Math. 6 (1981), 149–160.
[73] Urysohn, P. Sur un espace métrique universel. Bull. Sc. Math. 2e série 51
(1927), 43–64.
[74] Väisälä, J. Porous sets and quasisymmetric maps. Trans. Amer. Math. Soc.
299 (1987), 525–533.
[75] Väisälä, J. Quasiconformal maps of cylindrical domains. Acta Math. 162
(1989), 201–225.
44

[76] Vershik, A. M. The universal Uryson space, Gromov’s metric triples, and
random metrics on the series of natural numbers. Uspekhi Mat. Nauk 53, 5(323)
(1998), 57–64.
[77] Vol´berg, A. L., and Konyagin, S. V. On measures with the doubling
condition. Izv. Akad. Nauk SSSR Ser. Mat. 51 (1987), 666–675. English trans-
lation: Math. USSR-Izv., 30:629–638, 1988.
[78] von Neumann, J., and Schoenberg, I. J. Fourier integrals and metric
geometry. Trans. Amer. Math. Soc. 50 (1941), 226–251.
[79] Wojtaszczyk, P. Banach spaces for analysts. Cambridge University Press,
Cambridge, 1991.

Department of Mathematics, University of Michigan, Ann Arbor,

MI 48109, USA ([email protected])

MSRI, 17 Gauss Way, Berkeley, CA 94720, USA