GAFA, Geom. funct. anal.

Vol. 15 (2005) 534 – 554 c Birkhäuser Verlag, Basel 2005

1016-443X/05/020534-21
DOI 10.1007/s00039-005-0515-x GAFA Geometric And Functional Analysis

ISOPERIMETRIC INEQUALITIES OF
EUCLIDEAN TYPE IN METRIC SPACES

S. Wenger

1 Introduction
The purpose of this paper is to prove an isoperimetric inequality of Eu-
clidean type for complete metric spaces admitting a cone-type inequality.
These include all Banach spaces and all complete, simply-connected metric
spaces of non-positive curvature in the sense of Alexandrov or Busemann.
as a consequence we obtain solutions to the Plateau problem in certain
spaces, such as duals of Banach spaces and complete simply-connected met-
ric spaces of non-positive curvature. The main theorem generalizes results
of Gromov [G] and Ambrosio–Kirchheim [AmK].
1.1 Statement of the main results. The isoperimetric problem of
Euclidean type for a space X and given classes Ik−1 , Ik , and Ik+1 of sur-
faces of dimension k−1, k, and k+1 in X, together with boundary operators
∂ ∂
Ik+1 → Ik → Ik−1 and a volume function M on each class, asks the follow-
ing: Does there exist for every surface T ∈ Ik without boundary, ∂T = 0,
a surface S ∈ Ik+1 with ∂S = T and such that
M(S) ≤ DM(T )(k+1)/k (1)
for a constant D depending only on X and k? A space for which this holds
is said to admit an isoperimetric inequality of Euclidean type for Ik (or in
dimension k). The isoperimetric problem of Euclidean type was resolved
by Federer and Fleming in [FF] for Euclidean space X = R n and in the
class Ik of k-dimensional integral currents, k ∈ {1, . . . , n}. In [G] Gro-
mov extended the result to finite dimensional normed spaces and moreover
to complete Riemannian manifolds admitting a cone-type inequality (for
which the definition will be given below). Gromov worked in the class of
Lipschitz chains, formal finite sums of Lipschitz maps on standard sim-
plices. Recently, Ambrosio and Kirchheim extended in [AmK] the theory
of currents from the Euclidean setting to general metric spaces. The met-
ric integral currents define suitable classes Ik (X) of k-dimensional surfaces
in X. (It is to be noted that there are metric spaces for which Ik (X) only
consists of the trivial current. However, for the spaces considered here,
Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 535

this is not the case (see below)). In [AmK] the isoperimetric inequality of
Euclidean type is proved for dual Banach spaces X admitting an approxi-
mation by finite dimensional subspaces in the following sense: There exists
a sequence of projections Pn : X → Xn onto finite dimensional subspaces
such that Pn (x) weakly∗ -converges to x for every x ∈ X. The authors
then raise the question whether all Banach spaces admit an isoperimetric
inequality of Euclidean type.
In this paper we answer this question affirmatively and, in fact, prove
the Euclidean isoperimetric inequality for a large class of metric spaces
including also many non-linear ones. We will work in the class of metric
integral currents Ik (X) developed in [AmK], the main definitions of which
will be recalled in section 2.2. The intuitive picture of a k-dimensional
integral current T one might have (for the moment) is that of a gener-
alized k-dimensional surface with a multiplicity function and an orienta-
tion of suitably defined tangent spaces (these will be k-dimensional normed
spaces). The orientation defines the boundary ∂T of T which will be a
(k − 1)-dimensional integral current. The volume of T , denoted M(T ) and
called mass of T , is the L1 -norm of the multiplicity function with respect
to a suitably defined ‘Finsler’ volume. (T will be a functional rather than
a set. The set on which T will live is called the support, denoted spt T ).
See section 2.2 and [AmK] for the precise definitions. An integral current
T with ∂T = 0 will be called a cycle.
Definition 1.1. A metric space (X, d) is said to admit a k-dimensional
cone-type inequality (or to admit a cone-type inequality for Ik (X)) if for
every cycle T ∈ Ik (X) with bounded support there exists an S ∈ Ik+1 (X)
satisfying ∂S = T and
M(S) ≤ Ck diam(spt T )M(T )
for a constant Ck depending only on k and X.
The main result can be stated as follows:
Theorem 1.2. Let (X, d) be a complete metric space and k ∈ N. Suppose
that X satisfies a cone-type inequality for Ik (X) and, if k ≥ 2, that X also
satisfies an isoperimetric inequality of Euclidean type for Ik−1 (X). Then
(X, d) admits an isoperimetric inequality of Euclidean type for Ik (X): For
every cycle T ∈ Ik (X) there exists an S ∈ Ik+1 (X) with ∂S = T and such
that
M(S) ≤ Dk [M(T )](k+1)/k
where Dk only depends on k and the constants of the cone-type inequality
in Ik (X) and the isoperimetric inequality in Ik−1 (X).
536 S. WENGER GAFA

We would like to point out that no additional assumptions are made

on T . In particular, T can have unbounded support. However, if T has
bounded (compact) support then there exists an S satisfying the proper-
ties above and which also has bounded (compact) support. The theorem
extends [G, 3.4.C] from the setting of Riemannian manifolds to that of
complete metric spaces. As will be shown in section 2.3 Banach spaces ad-
mit cone-type inequalities in every dimension. This leads to the following
generalization of the result of Ambrosio and Kirchheim and answers the
question raised.
Corollary 1.3. There are universal constants Dk , k ≥ 1, such that
every Banach space E admits an isoperimetric inequality of Euclidean type
for Ik (E) with constant Dk .
In [G, Theorem 4.2.A] Gromov proves an isoperimetric inequality for
Lipschitz cycles in arbitrary Banach spaces. The proof uses the fact that
Lipschitz cycles admit an approximation by Lipschitz cycles with supports
in finite dimensional subspaces. It is not clear whether arbitrary integral
currents can be approximated in a similar way.
The constants Dk in Theorem 1.2 and Corollary 1.3 can be computed
explicitly. However, they are not optimal, not even if one takes X = R n .
Proving optimality of constants has turned out to be a challenging task.
In a major advance Almgren Jr. proved the isoperimetric inequality with
optimal constants for R n in [A], yielding equality in (1) precisely for round
k-dimensional spheres in (k + 1)-dimensional affine subspaces. For domains
in Hadamard manifolds of dimension 3 and 4 the isoperimetric inequality
with optimal Euclidean constants has been proved by Kleiner [Kl1] and
Croke [C], respectively.
The cone-type inequality is for example satisfied in spaces (X, d) ad-
mitting a γ-convex bicombing, for some γ > 0. By this we mean choices,
for all x, y ∈ X, of γd(x, y)-Lipschitz paths cxy : [0, 1] → X joining x to
y and such that the following holds: For any three points u, v, v  ∈ X the
inequality

d cuv (t), cuv
(t) ≤ γd(v, v  )
holds for all t ∈ [0, 1]. Examples of such spaces include all complete simply-
connected metric spaces of non-positive curvature in the sense of Alexan-
drov (called Hadamard spaces in the sequel) and, more generally, spaces
with a convex metric. The definitions will be given in section 2.1. It should
be mentioned here that these spaces contain many rectifiable sets (see [Kl2])
and hence Ik (X) is not trivial.
Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 537

Corollary 1.4. For fixed γ > 0 and k ∈ N, every complete metric space
(X, d) with a γ-convex bicombing admits an isoperimetric inequality of
Euclidean type for Ik (X) with constants Dk depending only on k and γ.
As an application of the above results we will prove the existence of a
solution to the generalized Plateau problem in dual Banach spaces and in
Hadamard spaces.
Theorem 1.5. If E is the dual of a Banach space then for every T ∈ Ik (E)
with compact support and ∂T = 0 there exists an S ∈ Ik+1 (E) with ∂S = T
and  
M(S) = inf M(S  ) : S  ∈ Ik+1 (E), ∂S  = T . (2)
Moreover, every S ∈ Ik+1 (E) which satisfies ∂S = T and (2) has compact
support.
This extends [AmK, Theorem 10.6]. We point out that we do not make
the assumption that the predual of E be separable. In [AmK] there are
examples of non-dual spaces for which the Plateau problem has a solution.
For general Banach spaces, the Plateau problem is unsolved.
Concerning (non-linear) metric spaces we have the following result.
Theorem 1.6. If (X, d) is a Hadamard space then for every T ∈ Ik (X)
with compact support and ∂T = 0 there exists an S ∈ Ik+1 (X) with ∂S = T
and  
M(S) = inf M(S  ) : S  ∈ Ik+1 (X), ∂S  = T . (3)
Moreover, every S ∈ Ik+1 (X) which satisfies ∂S = T and (3) has compact
support.
The two theorems above will follow from a more general theorem which
uses the ultra-completion of metric spaces (section 4). The basic ideas in
the proof are similar to those in [AmK, Theorem 10.6]. The argument using
ultra-completions is due to Urs Lang.
1.2 Outline of the main argument. The proof of Theorem 1.2 is
inspired by Gromov’s argument. However, the methods in [G] rely in several
ways on the bi-Lipschitz embeddability of compact Riemannian manifolds
into Euclidean space. Our approach uses a more intrinsic analysis of k-
dimensional cycles. For the description of our argument it is convenient to
introduce the following terminology: A cycle T ∈ Ik (X) is called round if
diam(spt T ) ≤ EM(T )1/k
for a constant E depending only on k and on the space X. The essential step
in the proof is stated in Proposition 3.1 which claims the following: Under
538 S. WENGER GAFA

the hypotheses of Theorem 1.2 every cycle T ∈ Ik (X) can be decomposed

into the sum T = T1 + · · · + TN + R of round cycles Ti and a cycle R with
the properties that M(Ti ) ≤ (1 + λ)M(T ) and M(R) ≤ (1 − δ)M(T )
for constants 0 < δ, λ < 1 depending only on k and the constant from the
isoperimetric inequality for Ik−1 (X). The construction of such a decomposi-
tion is based on an analysis of the growth of the function r → T (B(y, r)).
Intuitively speaking, T (B(y, r)) is the volume of T lying in the closed ball
B(y, r) with center y and radius r (see section 2.2 for the precise definition
of T ). For almost every y ∈ spt T we have


T  B(y, r) ≥ F r k (4)
for small r > 0 and a constant F depending only on k. Denoting by r0 (y)
the least upper bound of those r satisfying (4) one can prove that, when T
is cut open along the metric sphere with center y and of radius about r0 (y)
only little boundary is created. By closing this boundary with a suitable
isoperimetric filling (Lemma 3.4) one constructs a decomposition of T into
a sum T = T1 + R̃. The cycle T1 , lying essentially in B(y, r0 (y)) is round.
This will easily follow from the definition of r0 (y). By using a simple Vitali-
type covering argument one then shows that enough such round cycles Ti
can be split off in order to leave a rest R satisfying M(R) ≤ (1 − δ)M(T ).
Successive application of Proposition 3.1 will easily establish the proof of
the main theorem.
The paper is structured as follows: In section 2.2 we recall the main
definitions from the theory of currents in metric spaces and state those re-
sults from [AmK] vital for our purposes. Then, following a construction in
[AmK], we prove cone-type inequalities for various classes of metric spaces.
The decomposition of a given cycle is constructed in section 3. This forms
the main part of the paper. Section 3 also contains the proof of the main
theorem. The last section contains a general theorem from which Theo-
rem 1.5 and Theorem 1.6 will be derived.
Acknowledgements. I would like to thank Urs Lang for helpful discus-
sions on the subject, and in particular for the contribution to the Plateau
problem for Hadamard spaces. My thanks also go to Luigi Ambrosio for
comments concerning the final version of the paper.

2 Preliminaries
In section 2.1 we review the definition of a convex bicombing on a met-
ric space and give a list of spaces admitting a convex bicombing. One
Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 539

important class is that of simply-connected metric spaces of non-positive

curvature in the sense of Alexandrov for which the definition will also be
given. Section 2.2 contains the main definitions from the theory of metric
currents developed in [AmK] as well as the results relevant in our con-
text. The purpose of section 2.3 is to construct cone fillings and to prove a
cone-type inequality for metric spaces admitting a convex bicombing. The
preliminary material on ultra-completions needed for solving the Plateau
problem is to be found in section 4.
2.1 Convex bicombings and spaces of non-positive curvature. A
continuous curve c : [0, 1] → X in a metric space (X, d) is called rectifiable
if it has finite length, i.e. if
 N−1 


length(c) := sup d c(ti ), c(ti+1 ) : 0 = t1 < · · · < tN = 1 < ∞ .
i=1
An isometric embedding c : [0, d(x, y)] → X satisfying c(0) = x and
c(d(x, y)) = y is called a geodesic segment from x to y. The space X is
called geodesic if every two points x and y can be connected by a geodesic
segment.
Definition 2.1. A γ-convex bicombing on a metric space (X, d) is a choice
of curves cxy : [0, 1] → X joining x to y, for each two points x, y ∈ X, such
that the following two conditions hold:
(i) For any points u, v ∈ X the curve cuv is Lipschitz with Lipschitz
constant γd(u, v).
(ii) For any three points u, v, v  ∈ X and for t ∈ [0, 1] we have


d cuv (t), cuv
(t) ≤ γd(v, v  ) .
If a metric space has a γ-convex bicombing for some γ > 0 then it
is said to admit a convex bicombing. Let ϕ : X → Y be a bi-Lipschitz
homeomorphism between two metric spaces X and Y . Then X admits a
convex bicombing if and only if Y does. In this sense, the above definition
is bi-Lipschitz invariant.
Example. For a normed space E the straight lines cxy (t) := ty + (1 − t)x
where t ∈ [0, 1] and x, y ∈ E define a 1-convex bicombing.
Complete simply-connected Alexandrov spaces of non-positive curva-
ture (called Hadamard spaces) form another important class of spaces ad-
mitting a convex bicombing. They are defined as follows: Let (X, d) be a
metric space. A geodesic triangle in X consists of three points x, y, z ∈ X
and of a choice of three geodesic segments joining them, denoted
540 S. WENGER GAFA

by [x, y], [y, z], [z, x]. A comparison triangle for the geodesic triangle
∆([x, y], [y, z], [z, x]) is a triangle ∆([x, y], [y, z], [z, x]) in R 2 with vertices
x, y, z ∈ R 2 satisfying d(x, y) = |x − y|, d(y, z) = |y − z|, and d(z, x) =
|z − x|. A point p ∈ [x, y] is called a comparison point for p ∈ [x, y] if
d(x, p) = |x − p|. Comparison points for points on the other geodesic sides
of the triangle are defined similarly.
Definition 2.2. A geodesic metric space (X, d) is called CAT(0)-space
if for every geodesic triangle ∆ and its comparison triangle ∆, and for all
points p, q ∈ ∆ the comparison points p, q ∈ ∆ satisfy
d(p, q) ≤ |p − q| .
A complete CAT(0)-space is called Hadamard space. We refer to [BH]
for an account on these spaces.
Example. Every complete simply-connected Riemannian manifold of non-
positive sectional curvature is a Hadamard space.
CAT(0)-spaces clearly admit a convex bicombing. Indeed, every two
points can be joined by a unique geodesic and for points u, v, v  the repara-
metrizations cuv and cuv
(to [0, 1]) of the geodesics joining u to v and,
respectively, u to v  satisfy


d cuv (t), cuv
(t) ≤ td(v, v  ) for all t ∈ [0, 1] . (5)
A uniquely geodesic metric space X for which (5) holds for all points
u, v, v  ∈ X is said to have a convex metric. Such a space clearly admits a
convex bicombing.
2.2 Currents in metric spaces. The general reference for this section
is [AmK]. Let (X, d) be a complete metric space and let D k (X) denote
the set of (k + 1)-tuples (f, π1 , . . . , πk ) of Lipschitz functions on X with f
bounded. The Lipschitz constant of a Lipschitz function f on X will be
denoted by Lip(f ).
Definition 2.3. A k-dimensional metric current T on X is a multi-linear
functional on D k (X) satisfying the following properties:
(i) If πij converges pointwise to πi as j → ∞ and if supi,j Lip(πij ) < ∞
then
T (f, π1j , . . . , πkj ) −→ T (f, π1 , . . . , πk ) .
k
(ii) If {x ∈ X : f (x) = 0} is contained in the union i=1 Bi of Borel sets
Bi and if πi is constant on Bi then
T (f, π1 , . . . , πk ) = 0 .
Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 541

(iii) There exists a finite Borel measure µ on X such that

k
T (f, π1 , . . . , πk ) ≤ Lip(πi ) |f |dµ (6)
i=1 X

for all (f, π1 , . . . , πk ) ∈ D k (X).

The space of k-dimensional metric currents on X is denoted by Mk (X)
and the minimal Borel measure µ satisfying (6) is called mass of T and
written as T . We also call mass of T the number T (X) which we
denote by M(T ). The support of T is, by definition, the closed set spt T
of points x ∈ X such that T (B(x, r)) > 0 for all r > 0. Here, B(x, r)
denotes the closed ball B(x, r) := {y ∈ X : d(y, x) ≤ r}.
Remark 2.4. As is done in [AmK] we will also assume here that the cardi-
nality of any set is an Ulam number. This is consistent with the standard
ZFC set theory. We then have that spt T is separable and furthermore that
T  is concentrated on a σ-compact set, i.e. T (X\C) = 0 for a σ-compact
set C ⊂ X (see [AmK]).
The restriction of T ∈ Mk (X) to a Borel set A ⊂ X is given by
(T A)(f, π1 , . . . , πk ) := T (f χA , π1 , . . . , πk ) .
This expression is well defined since T can be extended to a functional on
tuples for which the first argument lies in L∞ (X, T ).
The boundary of T ∈ Mk (X) is the functional
∂T (f, π1 , . . . , πk−1 ) := T (1, f, π1 , . . . , πk−1 ) .
It is clear that ∂T satisfies conditions (i) and (ii) in the above definition. If
∂T also has finite mass (condition (iii)) then T is called a normal current.
The respective space is denoted by Nk (X).
The push-forward of T ∈ Mk (X) under a Lipschitz map ϕ from X to
another complete metric space Y is given by
ϕ# T (g, τ1 , . . . , τk ) := T (g ◦ ϕ, τ1 ◦ ϕ, . . . , τk ◦ ϕ)
for (g, τ1 , . . . , τk ) ∈ D k (Y ). This defines a k-dimensional current on Y , as
is easily verified.
In this paper we will mainly be concerned with integer rectifiable and
integral currents. For notational purposes we first repeat some well-known
definitions. The Hausdorff k-dimensional measure of A ⊂ X is defined
to be
 ∞  ∞ 
diam(Bi ) k
H (A) := lim inf
k
ωk :B⊂ Bi , diam(Bi ) < δ ,
δ0 2
i=1 i=1
542 S. WENGER GAFA

where ωk denotes the Lebesgue measure of the unit ball in R k . The k-

dimensional lower density Θ∗k (µ, x) of a finite Borel measure µ at a point
x is given by the formula
µ(B(x, r))
Θ∗k (µ, x) := lim inf .
r0 ωk r k
An Hk -measurable set A ⊂ X is said to be countably Hk -rectifiable if there
exist countably many Lipschitz maps fi : Bi → X from subsets Bi ⊂ R k
such that   
Hk A\ fi (Bi ) = 0 .

Definition 2.5. A current T ∈ Mk (X) with k ≥ 1 is said to be rectifiable

if
(i) T  is concentrated on a countably Hk -rectifiable set and
(ii) T  vanishes on Hk -negligible sets.
T is called integer rectifiable if, in addition, the following property holds:
(iii) For any Lipschitz map ϕ : X → R k and any open set U ⊂ X there
exists θ ∈ L1 (R k , Z) such that
 
∂πi
ϕ# (T U )(f, π1 , . . . , πk ) = θf det ∂x j
dLk
Rk

for all (f, π1 , . . . , πk ) ∈ D k (R k ).

A 0-dimensional (integer) rectifiable current is a T ∈ M0 (X) of the
form ∞

T (f ) = θi f (xi ) , f Lipschitz and bounded ,
i=1
for suitable θi ∈ R (or θi ∈ Z) and xi ∈ X.
The space of rectifiable currents is denoted by Rk (X), that of integer
rectifiable currents by Ik (X). Endowed with the mass norm Mk (X) is
a Banach space, Rk (X) a closed subspace, and Ik (X) a closed additive
subgroup. This follows directly from the definitions. Integer rectifiable
normal currents are called integral currents. The respective space is denoted
by Ik (X). As the mass of a k-dimensional normal current vanishes on Hk -
negligible sets [AmK, Theorem 3.9] it is easily verified that the push-forward
of an integral current under a lipschitz map is again an integral current. In
the following, an element T ∈ Ik (X) with zero boundary ∂T = 0 will be
called a cycle. An element S ∈ Ik+1 (X) satisfying ∂S = T is said to be a
filling of T .
Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 543

The characteristic set ST of a rectifiable current T ∈ Rk (X) is defined

by  
ST := x ∈ X : Θk (T , x) > 0 . (7)
It can be shown that ST is countably Hk -rectifiable and that T  is
concentrated on ST . In the next theorem the function λ : ST → (0, ∞)
denotes the area factor on the (weak) tangent spaces to ST as defined in
[AmK]. We do not provide a definition here since for our purposes it is
enough to know that λ is Hk -integrable and bounded from below by k−k/2
(see [AmK, Lemma 9.2]).
Theorem 2.6 [AmK, Theorem 9.5]. If T ∈ Rk (X) then there exists a
Hk -integrable function θ : ST → (0, ∞) such that
T (A) = λθdHk for A ⊂ X Borel,
A∩ST
that is, T  = λθdHk ST . Moreover, if T is an integral current then θ
takes values in N := {1, 2, . . . } only.
The following slicing theorem (proved in [AmK, Theorems 5.6 and 5.7])
is, besides Theorem 2.6, the only result from the theory of metric currents
needed in the proof of the main result.
Theorem 2.7. Let be T ∈ Nk (X) and a Lipschitz function on X. Then
there exists for almost every r ∈ R a normal current T, , r ∈ Nk−1 (X)
with the following properties:
(i) T, , r = ∂(T { ≤ r}) − (∂T ) { ≤ r};
(ii)  T, , r  and ∂ T, , r  are concentrated on −1 ({r});
(iii) M( T, , r ) ≤ Lip( ) dr d
M(T { ≤ r}).
Moreover, if T ∈ Ik (X) then T, , r ∈ Ik−1 (X) for almost all r ∈ R.
2.3 Cone constructions and cone-type inequalities. The following
cone construction is a slightly modified version of the one given in [AmK].
Let (X, d) be a complete metric space and T ∈ Nk (X) and endow
[0, 1] × X with the Euclidean product metric. Given a Lipschitz function f
on [0, 1] × X and t ∈ [0, 1] we define the function ft : X → R by ft (x) :=
f (t, x). To every T ∈ Nk (X) and every t ∈ [0, 1] we associate the normal
k-current on [0, 1] × X given by the formula


[t] × T (f, π1 , . . . , πk ) := T (ft , π1t , . . . , πkt ) ,
The product of a normal current with the interval [0, 1] is defined as follows.
Definition 2.8. For a normal current T ∈ Nk (X) the functional [0, 1] × T
on D k+1 ([0, 1] × X) is given by
544 S. WENGER GAFA


[0, 1] × T (f, π1 , . . . , πk+1 )

k+1 1 
i+1 ∂πit
:= (−1) T ft , π1t , . . . , πi−1t , πi+1t , . . . , πk+1t dt
0 ∂t
i=1
for (f, π1 , . . . , πk+1 ) ∈ D k+1 ([0, 1] × X).
We now have the following result whose proof is analogous to that of
[AmK, Proposition 10.2 and Theorem 10.4].
Theorem 2.9. For every T ∈ Nk (X) with bounded support the func-
tional [0, 1] × T is a (k + 1)-dimensional normal current on [0, 1] × X with
boundary


∂ [0, 1] × T = [1] × T − [0] × T − [0, 1] × ∂T .
Moreover, if T ∈ Ik (X) then [0, 1] × T ∈ Ik+1 ([0, 1] × X).
Proposition 2.10. If (X, d) is a complete metric space admitting a γ-
convex bicombing then every cycle T ∈ Ik (X), k ≥ 1, with bounded support
has a filling S ∈ Ik+1 (X) satisfying
M(S) ≤ (k + 1)γ k+1 diam(spt T )M(T ) .
Proof. We fix x0 ∈ spt T and define a locally Lipschitz map ϕ : [0, 1]×X → X
by ϕ(t, x) := cx0 x (t). Then, for fixed x ∈ spt T , the map t → ϕ(t, x) is
γ diam(spt T )-Lipschitz, whereas for fixed t ∈ [0, 1] the map x → ϕ(t, x)
is γ-Lipschitz. Theorem 2.9 implies that ϕ# ([0, 1] × T ) ∈ Ik+1 (X) and
furthermore





∂ϕ# [0, 1] × T = ϕ# ∂([0, 1] × T ) = ϕ# [1] × T − ϕ# [0] × T = T .
To obtain the estimate on mass we compute for (f, π1 , . . . , πk+1 ) ∈ D k+1 (X)
that
ϕ# ([0, 1] × T )(f, π1 , . . . , πk+1 )

k+1 1 
∂πi ◦ϕt
≤ T f ◦ ϕt , π1 ◦ ϕt ,...,πi−1 ◦ ϕt , πi+1 ◦ ϕt ,...,πk+1 ◦ ϕt dt
0 ∂t
i=1

k+1 1
∂(πi ◦ ϕt )
≤ Lip(πj ◦ ϕt ) f ◦ ϕt dT dt
0 j=i X ∂t
i=1
k+1 1
≤ (k + 1)γ k+1 diam(spt T ) Lip(πj ) f ◦ ϕ(t, x) dT (x)dt .
j=1 0 X

From this it follows that

 

ϕ# [0, 1]×T  ≤ (k + 1)γ k+1 diam(spt T )ϕ# L1 ×T  ,
and this concludes the proof.

Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 545

3 Partial Decomposition and Proof of the Main Result

The aim of this section is to prove the proposition below which forms the
crucial step when decomposing a cycle into the sum of round cycles. This
result will be used to prove the Theorem 1.2.
Proposition 3.1. Let (X, d) be a complete metric space and k ≥ 1
an integer. If k ≥ 2 then suppose furthermore that X has a Euclidean
isoperimetric inequality for Ik−1 (X) with a constant C > 0. There then
exist constants E > 0 and 0 < δ, λ < 1 depending only on k and C with
the following
 property: Every cycle T ∈ Ik (X) admits a decomposition
T = N T
i=1 i + R into a sum of integral cycles satisfying:
(i) diam(spt Ti ) ≤ EM(Ti )1/k ;
N ≤ (1 − δ)M(T );
(ii) M(R)
(iii) i=1 M(Ti ) ≤ (1 + λ)M(T ).
We first state some preparatory lemmas. The first will be employed to
obtain the estimate in (ii) for the cycle R.
Lemma 3.2. Let (Y, d) be a metric space, µ a finite Borel measure on Y ,
and F > 0, k ∈ N. For y ∈ Y define
 
r0 (y) := max r ≥ 0 : µ(B(y, r)) ≥ F r k .
If r0 (y) > 0 for µ-almost every y ∈ Y then there exist points y1 , . . . , yN ∈ Y
satisfying
(i) r0 (yi ) > 0;
(ii) B(yi , 2r0 (yi )) ∩ B(yj , 2r0 (yj )) = ∅ if i = j;
N
(iii) i=1 µ(B(yi , r0 (yi ))) ≥ αµ(Y );
for a constant α > 0 depending only on k.
The proof, provided for completeness, is analogous to that of the simple
version of the Vitali covering lemma.
Proof. Set Y1 := Y and r1 := sup{r0 (y) : y ∈ Y1 } and choose y1 ∈ Y1 such
that r0 (y1 ) > 23 r1 . If y1 , . . . , yj are chosen, we define

j


Yj+1 := Y B yi , 5r0 (yi )
i=1

and rj+1 := sup{r0 (y) : y ∈ Yj+1 }. If µ(Yj+1 ) > 0 then we choose
yj+1 ∈ Yj+1 such that r0 (yj+1 ) > 23 rj+1  . This procedure yields (possibly fi-

nite) sequences yi ∈ Yi and r1 ≥ r2 ≥ . . . and we claim that



B(yi , 2r0 (yi )) ∩ B(yj , 2r0 (yj )) = ∅ if i = j. This is immediate since

d(yi , yj ) ≥ 5r0 (yi ) > 2r0 (yi ) + 2rj ≥ 2r0 (yi ) + 2r0 (yj )
546 S. WENGER GAFA

for i < j. If we have µ(YN ) = 0 for some N then it follows that


N

 N
 k 
N


5 k
µ B(yi , r0 (yi )) = F 5r0 (yi ) > µ B(yi , 5r0 (yi )) ≥ µ(Y ) .
i=1 i=1 i=1

all N∞∈ N then it follows

 that rN → 0
If, on the other hand, µ(YN ) > 0 for 

as N → ∞ and furthermore that µ Y \ i=1 B(yi , 5r0 (yi )) = 0. The proof

then follows as in the finite case.

The study of the growth of the function r → T (B(x, r)) will play
a predominant role in the proof of Proposition 3.1. In this context the
following easy fact will be helpful.
Lemma 3.3. Fix C̄ > 0, k ≥ 2, 0 ≤ r0 < r1 < ∞, and suppose
β : [r0 , r1 ] → (0, ∞) is non-decreasing and satisfies
rk0
(i) β(r0 ) = C̄ k−1 kk
;

(ii) β(r) ≤ C̄[β (r)]k/(k−1) for a.e. r ∈ (r0 , r1 ).
Then it follows that
rk
β(r) ≥ for all r ∈ [r0 , r1 ] .
C̄ k−1 kk
Proof. By rearranging (ii) we obtain
β  (t) 1
k−1
≥ k−1
β(t) C̄ k
k

and integration from r0 to r yields the claimed estimate.

The next statement is concerned with the support of fillings. It will be

used to prove the roundness of the cycles Ti . The Ti will be constructed by
restricting T to a ball B(yi , r) and filling in the boundary ∂(T B(yi , r))
by a filling satisfying the isoperimetric inequality. Lemma 3.4 ensures that
we can choose a filling whose support stays near its boundary.
Lemma 3.4. Let (X, d) be a complete metric space and k ≥ 2. Suppose
that X admits a Euclidean isoperimetric inequality for Ik−1 (X) with a
constant C > 0. Then there exists for every cycle T ∈ Ik−1 (X) and every
ε > 0 an S ∈ Ik (X) satisfying ∂S = T and
 
M(S) ≤ inf M(S  ) : S  ∈ Ik (X), ∂S  = T + ε
and furthermore

 rk
S B(x, r) ≥
(3C)k−1 kk
for all x ∈ spt S and 0 ≤ r ≤ dist(x, spt T ).
Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 547

In particular, there exists for every cycle T ∈ Ik−1 (X) an S ∈ Ik (X)

with ∂S = T and
M(S) ≤ C[M(T )]k/(k−1)
and such that

spt S ⊂ B spt T, 3CkM(T )1/(k−1) .
Here, B(A, ) denotes the -neighborhood of the set A. The proof of the
lemma is essentially contained in the proof of [AmK, Theorem 10.6].
Proof. Let M denote the complete metric space consisting of all fill-
ings S ∈ Ik (X) of T and endowed with the metric given by dM (S, S  ) :=
M(S − S  ). By the Ekeland variational principle [E], for every ε ∈ (0, 1/2),
there exists an S ∈ M satisfying
 
M(S) ≤ inf M(S  ) : S  ∈ Ik+1 (X) , ∂S  = T + ε
(and thus the isoperimetric inequality for ε > 0 small enough) and such
that the function
S  ∈ M → M(S  ) + εM(S  − S)
is minimal at S  = S. Let be x ∈ spt S\ spt T and set x (y) := d(x, y).
Then, for almost every 0 < r < d(x, spt T ) the slice S, x , r exists, has zero
boundary, and belongs to Ik−1 (X). For an isoperimetric filling Sr ∈ Ik (X)
of S, x , r the integral current S B c (x, r) + Sr has boundary T and thus,
comparison with S yields



M S B c (x, r) + Sr + εM S B(x, r) − Sr ≥ M(S) .
Here, B c (x, r) denotes the complement of the ball B(x, r). Together with
the isoperimetric inequality, the above estimate implies that

 1+ε
k/(k−1)
M S B(x, r) ≤ M(Sr ) ≤ 3CM S, x , r
1−ε
for almost every r ∈ (0, dist(x, spt T )). Setting β(r) := S(B(x, r)) and
using the slicing theorem we obtain the inequality
 k/(k−1)

β(r) ≤ 3C β  (r) for a.e. r ∈ 0, dist(x, spt T )
which, after applying Lemma 3.3, yields

 rk
S B(x, r) ≥ for all 0 ≤ r < dist(x, spt T ) .
(3C)k−1 kk
This proves the lemma.

We are now ready to prove the proposition.

548 S. WENGER GAFA

Proof of Proposition 3.1. First of all, let F be given by F := λk−1 /C k−1 kk

with λ ≤ 1/6 small enough such that F < ωk /kk/2 . Let T ∈ Ik (X) be a
cycle and define for y ∈ X
 
r0 (y) := max r ≥ 0 : T (B(y, r)) ≥ F r k .
By [K, Theorem 9] we have
Hk (ST ∩ B(y, r))
lim =1
r0 ωk r k
for Hk -almost all y ∈ ST , the set ST being defined as in (7). Together with
Theorem 2.6 this implies that the set Y of points y ∈ ST satisfying r0 (y) > 0
has full T -measure. By Lemma 3.2 there exist points y1 , . . . , yN ∈ Y with
r0 (yi ) > 0, and such that the balls B(yi , 2r0 (yi )) are pairwise disjoint and
satisfy
N


T  B(yi , r0 (yi )) ≥ αT (Y ) = αM(T ) (8)
i=1
for a constant α > 0 depending only on k. Fix i ∈ {1, . . . , N } and set
r0 := r0 (yi ) and β(r) := T (B(yi , r)). It is clear that β is non-decreasing,
that β(r0 ) = F r0k , and that β(r) < F r k for all r > r0 . Denote furthermore
by the function (x) := d(yi , x). By Theorem 2.7 the slice T, , r =
∂(T B(yi , r)) exists for almost all r, is an element of Ik−1 (X), and satisfies


M T, , r ≤ β  (r) for a.e. r . (9)
We now consider one dimensional and higher dimensional cycles sepa-
rately: If k = 1 it follows from the fact that F = 1 and from the definition
of β that there exists a measurable set Ω ⊂ [r0 , 2r0 ) of positive measure
and such that β  (r) < 1 for all r ∈ Ω. Since, for r ∈ Ω, the slice T, , r
is a 0-dimensional integral current, M( T, , r ) is an integer number and
hence, by (9), the integral current Ti := T B(yi , r) has zero boundary.
Applying this to each i ∈ {1, . . . , N } one easily obtains a decomposition
T = T1 + · · · + TN + R satisfying all the properties stated in the proposition
(with λ = 0, δ ≥ α, and E < 4).
If k ≥ 2 then Lemma 3.3 and the definitions of F and r0 imply the
existence of Ω ⊂ [r0 , 4r0 /3] of positive measure such that
 k/(k−1)
C β  (r) < λβ(r) for all r ∈ Ω . (10)
By Theorem 2.7 we can assume without loss of generality that the slice
T, , r exists for every r ∈ Ω and is an element of Ik−1 (X). Choose an
r ∈ Ω arbitrarily and a filling S ∈ Ik (X) of T, , r as in Lemma 3.4.
Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 549

Together with (9) and (10) this implies that

M(S) ≤ λβ(r) , (11)
and, furthermore, since λ ≤ 1/6, that the support of S lies in the ball with
center yi and with a radius r̄ satisfying

4   1 4 3Ck(λF )1/k
r̄ ≤ r0 + 3Ck M( T, , r ) k−1
≤ 1+ r0 ≤ 2r0 .
3 3 C 1/k
Clearly, Ti := T B(yi , r) − S defines an integral cycle which satisfies
(1 − λ)β(r) ≤ M(Ti ) ≤ (1 + λ)β(r) .
Since Ti has support in B(yi , 2r0 (yi )) it follows that
4  1/k 4
diam(spt Ti ) ≤ 4r0 (yi ) = 1/k β(r0 ) ≤ M(Ti )1/k
F [F (1 − λ)]1/k
and hence Ti fulfills condition (i) with E := 4/[F (1 − λ)]1/k .
Since our construction of Ti leaves T B c (yi , 2r0 (yi )) unaffected (by
the fact that the balls B(yj , 2r0 (yj )) are pairwise disjoint) we can apply
the above construction to everyi ∈ {1, . . . , N } to obtain round cycles
T1 , . . . , TN . Setting R := T − N i=1 Ti this yields a decomposition T =
T1 + · · · + TN + R satisfying the claimed properties. Indeed, we have

N 
N
M(Ti ) ≤ (1 + λ) T (Bi ) ≤ (1 + λ)M(T )
i=1 i=1
where Bi is the ball chosen individually for every i as above. The estimate
for M(R) is also obvious since, by (8) and (11), we have
   

M(R) ≤ T  X\ Bi + λ T (Bi ) ≤ 1 − α(1 − λ) M(T ) .

This completes the proof of the proposition with δ := α(1 − λ).

The isoperimetric inequality now easily follows from Proposition 3.1.

Proof of Theorem 1.2. Let T ∈ Ik (X) be a cycle. Successive appli-
cation of Proposition 3.1 yields (possibly finite) sequences of cycles (Ti ),
(Rn ) ⊂ Ik (X) and an increasing sequence (Nn ) ⊂ N with the following
properties:
 n
• T = N i=1 Ti + Rn ;
• diam(spt Ti ) ≤ EM(Ti )1/k ;
• M(Rn ) ≤ (1 − δ)n M(T );
∞  ∞  1+λ
• i=1 M(Ti ) ≤ (1 + λ) i=0 (1 − δ) M(T ) = δ M(T ).
i
550 S. WENGER GAFA

The isoperimetric filling of T is then constructed as follows. We first

fill each Ti with an Si ∈ Ik+1 (X) from the cone inequality, i.e. one with
∂Si = Ti and such that
M(Si ) ≤ Ck diam(spt Ti )M(Ti ) ≤ Ck EM(Ti )(k+1)/k . (12)
∞ n
Nn
The finiteness of i=1 M(Ti ) implies that the sequence S := i=1 Si is a
Cauchy-sequence with respect to the mass norm because
∞  ∞  k+1
k+1 k
n→∞
M(S n+q
−S ) ≤ Ck E
n
M(Ti ) k ≤ Ck E M(Ti ) −→ 0 .
i=Nn +1 i=Nn +1
Since Ik+1 (X) is a Banach space the sequence Sn ∈ Ik+1 (X) ⊂ Ik+1 (X)
converges to a limit current S ∈ Ik+1 (X). As T − ∂S n = Rn converges to
0 it follows that ∂S = T and, in particular, that S ∈ Ik+1 (X). Finally, S
is an isoperimetric filling of T . Indeed, we have
   k+1
k+1 1+λ k k+1
M(S) ≤ M(Si ) ≤ Ck E M(Ti ) k ≤ Ck E M(T ) k ,
δ
which completes the proof.

We note that if T has bounded support then there exists an S ∈ Ik+1 (X)
as in the theorem and which, in addition, has bounded support. This follows
directly from the remark after Lemma 3.4. Furthermore, if T has compact
support then it is easy to prove, using Lemma 3.4, that there exists such
an S with compact support. For this see also the second part of the proof
of Theorem 1.6.

4 The Plateau Problem

Here we will solve the Plateau problem in generalized form and from this
we will then derive Theorem 1.5 and Theorem 1.6.
Definition 4.1. A non-principal ultra-filter on N is a finitely additive
probability measure ω on N (together with the σ-algebra of all subsets)
such that ω takes values in {0, 1} only and ω(A) = 0 whenever A ⊂ N is
finite.
Using Zorn’s lemma it is not too hard to establish the existence of non-
principal ultra-filters on N. It is also easy to prove the following fact. If
(Y, τ ) is a compact topological Hausdorff space then for every sequence
(yn )n∈N ⊂ Y there exists a unique point y ∈ Y such that


ω {n ∈ N : yn ∈ U } = 1
Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 551

for every U ∈ τ containing y. We will denote the point y by limω yn .

Now let (X, d) be a metric space and fix a non-principal ultra-filter ω
on N. We call a sequence (xn )n∈N ⊂ X bounded if diam{xn } < ∞. An
equivalence relation on the set of bounded sequences is given by
(xn ) ∼ (xn ) if and only if limω d(xn , xn ) = 0 .
Definition 4.2. The ultra-completion (X)ω of X is the set of equivalence
classes of bounded sequences (xn )n∈N together with the metric given by


dω (xn ), (xn ) := limω d(xn , xn )
where (xn ), (xn ) are bounded sequences in X.
The space X can be isometrically embedded into (X)ω by the map
ι : X → (X)ω assigning to x the constant sequence (x)n∈N . We are now in
a position to state the main theorem of this section.
Theorem 4.3. Let (X, d) be a complete metric space, k ∈ N, and
suppose X has an isoperimetric inequality of Euclidean type for Ik (X).
Let furthermore ω be a non-principal ultra-filter on N. Then there exists
for every T ∈ Ik (X) with ∂T = 0 and compact support an S ∈ Ik+1 ((X)ω )
such that ∂S = ι# T and
 
M(S) ≤ inf M(S  ) : S  ∈ Ik+1 (X), ∂S  = T . (13)
Remark 4.4. It should be noted that S is a (k + 1)-chain in the ultra-
completion of X whereas the infimum in (13) is taken over all (k + 1)-
chains S  in X. In particular, this theorem does not claim the existence of
a solution to the Plateau problem in the ultra-completion of X.
The ideas of proof are similar to that of the weak∗ -compactness theorem
for duals of separable Banach spaces in [AmK, Theorem 6.6]. The idea of
using the ultra-completion of a metric space to attack the Plateau problem
in the case of a Hadamard space is due to Urs Lang.
Proof. Let T ∈ Ik (X) be a cycle with compact support. We then claim
that there exists a sequence Si ∈ Ik+1 (X) with the properties listed below:
(i) ∂Si = T for every i ∈ N;
(ii) M(Si ) −→ inf{M(S  ) : S  ∈ Ik+1 , ∂S  = T } with i → ∞;
(iii) The sequence Zi := spt Si is equi-bounded and equi-compact.
By Lemma 3.4 we find a sequence (Si ) ⊂ Ik+1 (X) satisfying properties
(i) and (ii) and such that for x ∈ spt Si \ spt T we have

 r k+1
Si  B(x, r) ≥ whenever 0 ≤ r ≤ dist(x, spt T ) . (14)
(3C)k (k + 1)k+1
552 S. WENGER GAFA

We prove that the sequence of sets Zi is equi-bounded and equi-compact. In

order to do so fix > 0 and cover spt T by finitely many balls B(ym , ), m =
1, . . . , n. We then choose points x1 , x2 , . . . in Zi \ ∪ B(ym , 2 ) inductively
such that the balls B(xm , /2) are pairwise disjoint. By (14) there exist
only finitely many such points, say x1 , . . . , xN . (Note that N can be chosen
independent of i.) It follows that

N 
n
Zi ⊂ B(xm , 2 ) ∪ B(ym , 2 ) =: U .
m=1 m=1
This shows the equi-compactness of the Zi . To show equi-boundedness
it is enough to note that, by minimality of Si , there exists no connected
component of U with empty intersection with spt T . This establishes the
existence of a sequence with the properties listed above.
By Gromov’s compactness theorem there now exists a compact metric
space (Z, dZ ) and (after extracting a subsequence) isometric embeddings
ϕi : Zi → Z such that the subsets ϕ(Zi ) form a Cauchy sequence with re-
spect to the Hausdorff metric on Z. Since Z is compact we may assume
by Arzelà–Ascoli theorem that (after extracting a further subsequence) the
ϕi |spt T converge uniformly to an isometric embedding ϕ̃ : spt T → Z. By
the closure and compactness theorems for currents ([AmK, Theorems 5.2
and 8.5]) we may furthermore assume that ϕi# Si converges weakly to a
current Ŝ ∈ Ik+1 (Z). It follows easily that
∞ 

spt Ŝ ⊂ spt(ϕi# Si ) ⊂ limH ϕi (Zi )
m=1 i≥m

where limH denotes the Hausdorff limit for sequences of compact sets in Z.
We define a map ψ : Y := limH ϕi (Zi ) → (X)ω as follows: For y ∈ Y there
exists a sequence zi ∈ Zi such that ϕ(zi ) → y. We set ψ(y) := (zi )i∈N .
As is easily seen, the map ψ is well defined and an isometric embedding.
Furthermore, we have for S := ψ# Ŝ that
 
M(S) ≤ lim inf M(ϕi# Si ) = inf M(S  ) : S  ∈ Ik+1 (X), ∂S  = T .
It remains to show that ∂S = ι# T . Clearly,
∂ϕi# Si = ϕi# T → ϕ̃# T
and hence ∂ Ŝ = ϕ̃# T . Since, furthermore, ψ ◦ ϕ̃ = ι|spt T we obtain
∂S = ∂ψ# Ŝ = (ψ ◦ ϕ̃)# T = ι# T ,
concluding the proof.

Vol. 15, 2005 ISOPERIMETRIC INEQUALITIES IN METRIC SPACES 553

From the above theorem we now derive Theorems 1.5 and 1.6. Of course,
it suffices to prove the existence of a 1-Lipschitz retraction from the ultra-
completion to the space itself. We first deal with the case of Hadamard
spaces.
Proof of Theorem 1.6. If X is a Hadamard space then it follows easily
that the ultra-completion (X)ω is again a Hadamard space and that ι(X)
is a closed convex subspace of (X)ω . Hence, by Proposition II.2.4 of [BH]
there is a 1-Lipschitz retraction ϕ : (X)ω → ι(X). If T and S are as in the
above theorem then clearly (ι−1 ◦ ϕ)# (S) is a minimal filling of T .
To prove the second statement of the theorem let S ∈ Ik+1 (X) satisfy
∂S = T and (3). It suffices to show that

 r k+1
S B(x, r) ≥ k (15)
C (k + 1)k+1
for every x ∈ spt S and 0 ≤ r ≤ dist(x, spt T ), see the proof of Theo-
rem 4.3. For this, fix x ∈ spt S and define β(r) := S(B(x, r)). Using the
isoperimetric inequality and the slicing theorem it follows that
 k+1/k

β(r) ≤ C β  (r) for a.e. r ∈ 0, dist(x, spt T ) ,
from which (15) follows by applying Lemma 3.3.

Proof of Theorem 1.5. As above, we show that there is a 1-Lipschitz

retraction from (E)ω onto E. We define a retraction ϕ : (E)ω → E by
ϕ((xn )n ) := limω xn . Note that limω xn exists uniquely by the weak∗ -
compactness of balls in E. We must check that ϕ is 1-Lipschitz. For
this, let (xn ) and (yn ) be bounded sequences in E and fix ε > 0. We set
x := limω xn and y := limω yn and choose an element z in the predual F of
E with z = 1 and such that
x(z) − y(z) ≥ x − y − ε
4 .
By the definition of limω and the definition of the weak∗ -topology on E we
obtain
ω(A1 ∩ A2 ∩ A3 ) = 1 ,
where A1 := {m : |xm (z) − x(z)| ≤ ε/4}, A2 := {m : |ym (z) − y(z)| ≤ ε/4},
and A3 := {m : |xm −ym −dω ((xn ), (yn ))| ≤ ε/4}. For an m ∈ A1 ∩A2 ∩A3
we thus conclude


x − y ≤ xm (z) − ym (z) + 3ε 3ε
4 ≤ xm − ym  + 4 ≤ dω (xn ), (yn ) + ε .
Since ε > 0 was chosen arbitrarily this shows that ϕ is indeed 1-Lipschitz.
Finally, the second statement of the theorem follows exactly as in the
proof of Theorem 1.6.

554 S. WENGER GAFA

Remark 4.5. For a general Banach space E there need not exist a 1-
Lipschitz retraction from (E)ω onto E. Indeed, an example is given by
the space c0 of real sequences tending to zero together with the supremum
norm. (However, given a cycle in c0 with compact support K one easily
constructs a compact subset in c0 containing K and which is a 1-Lipschitz
retract of ∞ . The same methods as above then show that the Plateau
problem has a solution for compactly supported cycles in c0 .)

References
[A] F.J. Almgren Jr., Optimal isoperimetric inequalities, Indiana Univ.
Math. J. 35 (1986), 451–547.
[AmK] L. Ambrosio, B. Kirchheim, Currents in metric spaces, Acta Math.
185:1 (2000), 1–80.
[BH] M.R. Bridson, A. Haefliger, Metric Spaces of Non-Positive Curvature,
Grundlehren der mathematischen Wissenschaften 319, Springer, 1999.
[C] C. Croke, A sharp four dimensional isoperimetric inequality, Comment.
Math. Helv. 59 (1984), 187–192.
[E] I. Ekeland, On the variational principle, J. Math. Anal. Appl. 47 (1974),
324–353.
[FF] H. Federer, W.H. Fleming, Normal and integral currents, Ann. Math.
72 (1960), 458–520.
[G] M. Gromov, Filling Riemannian manifolds, J. Diff. Geom. 18 (1983),
1–147.
[K] B. Kirchheim, Rectifiable metric spaces: local structure and regularity
of the Hausdorff measure, Proc. Am. Math. Soc. 121:1 (1994), 113–123.
[Kl1] B. Kleiner, An isoperimetric comparison theorem, Invent. Math. 108
(1992), 37–47.
[Kl2] B. Kleiner, The local structure of length spaces with curvature bounded
above, Math. Z. 231:2 (1999), 409–456.

Stefan Wenger, Department of Mathematics, ETH Zentrum, CH – 8092 Zürich,

Switzerland [email protected]

Received: August 2003

Revision: June 2004
Accepted: August 2004