On asymmetric distances

Andrea C. G. Mennucci ?
May 28, 2007

Abstract
In this paper we discuss asymmetric metric spaces in an abstract setting, mim-
icking the usual theory of metric spaces. As a typical application, we consider
asymmetric metric spaces generated by functionals in Calculus of Variation.
K EYWORDS : ASYMMETRIC METRIC , GENERAL METRIC , QUASI METRIC , OSTENSIBLE METRIC , F INSLER METRIC ,
PATH METRIC , LENGTH , GEODESIC CURVE , H OPF –R INOW THEOREM

1 Introduction
“Besides, one insists that the distance function be symmetric, that is, d(x, x0 ) = d(x0 , x).
(This unpleasantly limits many applications [· · ·]).” M. Gromov ([12], Intr.)

The main purpose of this paper is to study an “asymmetric metric theory”; this
theory naturally generalizes the metric part of Finsler Geometry, much as symmetric
metric theory generalizes the metric part of Riemannian Geometry.
In order to do this, we will define the “asymmetric metric space” as a set M equipped
with a positive b : M × M → lR+ which satisfy the triangle inequality, but may fail
to be symmetric (see 2.1 for the formal definition). We will then state, (in sec. 2)
“asymmetric definitions” in this space, such as “forward local compactness”, “forward
completeness”, “forward boundedness”, and so on.
The invention of a (possibly) asymmetric distance is quite useful in some applied
fields, and mainly in Calculus of Variation, as we will exemplify in the next section
§1.ii, and discuss more in depth in sec. §5.i. It is then no surprise that this theory was
invented and studied many times in the past.
It was indeed shown in preceding studies that, even if a metric space does not have
the “smooth” character of a Riemannian structure, still many definitions and results can
be reformulated and carried on (1) : see for example in Busemann [5], or in Gromov’s
[12], or in Ambrosio et al [1, 2].
To exemplify this fact, in the next section we will present an example theorem that
generalizes the Hopf–Rinow theorem from finite dimensional Riemannian spaces to
suitable metric spaces; and a further generalization to the asymmetric case.
A different point of view is found in the theory of quasi metric (or ostensible metric)
where much emphasis is given to the topological aspects of the theory; indeed, a quasi-
metric will generate, in general, three different topologies (see §3.ii.1); this brings forth
many different and non-equivalent definitions of “completeness” (and of the other
? Scuola Normale Superiore Piazza dei Cavalieri 7, 56126 Pisa, Italy
(1) and Busemann [4] attributes the first such results in this respect to Cartan, in 1928

1
properties commonly stated in metric theory). In our definition of the “asymmetric
metric space” (M, b), we choose a different topology associated to the space than the
one used in “quasi metric spaces”: this makes it difficult to compare the results in the
two fields. We discuss further this issue in sec. §3.ii.
Our set of hypotheses is more resembling (and slightly generalizes) Busemann
theory of general metric spaces (as was defined in [4]); we will discuss differences and
similarities in section §3.i.
Summarizing, this presentation of the theory of asymmetric metrics is more specific
and more focused on “metric aspects” of what is seen in ostensible metric theory, and at
the same time is less “geometric oriented” than what is seen in Busemann work (indeed,
many of the examples presented are quite non-smooth).

§1.i The Hopf–Rinow theorem

In section I in [12], Gromov models the “metric part” of Riemannian Geometry, as
follows. Consider a metric space (M, d): we can define the length lenγ of a Lipschitz
curve γ : [α, β] → M using the total variation formula (see (2.5)); then we can
define a new metric dg (x, y) as the infimum of lenγ in the class of all Lipschitz curves
connecting x to y. Gromov defines that a metric space is “path-metric” if d = dg ; he
then proves in §1.11 §1.12 in [12] that (2)

Theorem 1.1 (symmetric Hopf-Rinow) Suppose that (M, d) is path-metric and lo-
cally compact; then the two following facts are equivalent

i). (M, d) is complete

ii). closed bounded sets are compact

and they imply that each pair of points can be joined by a geodesic (i.e. a minimum of
lenγ).

The above is the “metric only” counterpart of the theorem of Hopf-Rinow in Rie-
mannian Geometry: indeed, if (M, g) is a finite-dimensional Riemannian manifold, and
d is the associated distance, then (M, d) is path-metric and locally compact.
Since there is a Hopf-Rinow theorem in Finsler Geometry, we would expect that
there would be a corresponding theorem for “asymmetric metric spaces”; and indeed
we can prove this result:

Theorem 1.2 Suppose that the asymmetric metric space (M, b) is path-metric and
forward-locally compact; then the following two facts are equivalent

• forward-bounded closed sets are compact

• (M, b) is forward-complete

and they imply that any two points can be joined by a geodesic.

A first form of this theorem is due to Cohn–Vossen, according to the introduction of

Busemann’s [6], where though it is stated in with stronger hypothesis than the above; so
a more general form is here stated and proved as Theorem 4.33 (a comparison of the
two results follows that theorem).
(2) actually, we have added the easy implication ii=⇒i to the statement in [12]

2
§1.ii asymmetric metric theories in Calculus of Variations
For the sake of this introduction section, let M be a smooth connected differential
manifold.
Consider an integrand F : T M → [0, ∞] that is Borel measurable and such that
v 7→ F (x, v) is convex and positively 1-homogeneous (that is, F (x, lv) = lF (x, v) for
l ≥ 0). Consider the length
Z 1
lenγ = F (γ(s), γ̇(s)) ds (1.3)
0

where γ : [0, 1] → M is a locally Lipschitz curve; this generates a “distance” b(x, y), as
the infimum of the length of curves γ connecting x to y; b is called a Finsler metric, since
is is easily seen to satisfy a triangular inequality; but in general b will be asymmetric.
Consider also the Hamilton-Jacobi equation
n
H(x, Du(x)) = 0, u(a) = 0 (1.4)

where H : T ∗ M → lR is continuous and p 7→ H(x, p) is convex, and the solution

u : M → lR is in the viscosity sense; and a ∈ M .
In the above settings, a broad number of ideas appear naturally; we list them, in
decreasing order of regularity:
• In the field of Finsler Geometry, it is assumed that F is smooth and that v 7→
2 2
F 2 (x, v) is strongly convex for v 6= 0 (that is, the Hessian ∂∂v∂v
F
is positive
definite); many results of Riemannian geometry are generalized to this asymmetric
settings (see [3] for a wide treaty of the subject);
• the part of Finsler Geometry that deals with existence of geodesics can be easily
shown to hold when F ∈ C 2 and v 7→ F 2 (x, v) is strongly convex; so this is
a very common setting in books on Calculus of Variations, Control Theory and
Viscosity Solutions (3) ; it is well known that the Hamilton-Jacobi equation (1.4)
is naturally associated with the distance b(x, y): for example, u(x) = b(a, x) is
the viscosity solution of (1.4) when H is “dual” to F 2 , in the sense of Legendre–
Fenchel (see [16] for details).
• When we try to relax the requirements on F and H further, few results are known
in the literature (see [18] for an exception).
With appropriate hypotheses on H (and in particular on the sets Zx = {p | H(x, p) ≤
0}) [16] establishes the duality between H and F and that F is continuous and
locally bounded (and many other properties). The above work uses some re-
sults that are in section §5.i; for example theorem 5.9 shows that the Hopf–Rinow
theorem implies a purely geometric criteria for the existence of a minimum curve
for problem (1.3).
• Yet even the setting used in section §5.i is not the most general we can envision:
to start this section, we just assumed that F is Borel and F (x, ·) is convex and
1-homogeneous, and this is enough to generate the distance b, so that (M, b) is an
asymmetric metric space. So this study of the abstract properties of asymmetric
metric spaces may be useful in further attempts at generalizing the work that was
done in [16].
(3) for example, in §2.4.2 [9]; or L ∈ C 3 and strongly convex in §1.8 in [10]

3
2 Asymmetric metric spaces
We review some basal concepts and results; those are similar to what is commonly found
in books on metric spaces. In all of this chapter, M will be a generic set.

§2.i Asymmetric metric space

Definition 2.1 b : M × M → [0, ∞) is an asymmetric distance if b satisfies

• b ≥ 0 and b(x, y) = 0 iff x = y

• b(x, y) ≤ b(x, z) + b(z, y) ∀x, y, z ∈ M .

We call the pair (M, b) an asymmetric metric space.

We agree that b defines a topology τ on M , generated by the families of forward

and backward open balls

B − (x, ε)={y | b(y, x) < ε}

def def
B + (x, ε)={y | b(x, y) < ε},

that is, the topology is generated by the symmetric distance

def
d(x, y)=b(x, y) ∨ b(y, x) (2.2)

with balls
B(x, ε)={y | d(x, y) < ε} = B + (x, ε) ∩ B − (x, ε)
def

By the above definition,

Proposition 2.3 xn → x iff d(xn , x) → 0, iff both b(xn , x) → 0 and b(x, xn ) → 0

We will sometimes denote by τ + (resp. τ − ) the topology generated by forward

(resp. backward) balls alone; we remark that those may differ from τ : this issue is
discussed in sec. §3.ii.

Remark 2.4 We may also consider the case of an asymmetric distance b : M × M →

[0, +∞] where there may be points x, y such that b(x, y) = ∞; this is somewhat
uncommon, so we suggest two workarounds:

• by using the triangular inequality, it is easily seen that the equivalence relation
x ∼ y ⇐⇒ d(x, y) < ∞ decomposes the space (M, b) in equivalence classes
that are all open (so two different classes are disconnected); so we may restrict
the study of the space to those equivalence classes;

• defining φ(t) = t/(1 + t) and φ(∞) = 1 and then b̃=φ ◦ b and d˜=φ ◦ d, we may
def def

build a space (M, b̃) that is topologically equivalent, and where the distance does
not assume the value +∞.

Unfortunately those remedies are not really useful in some cases: see 2.17.

4
§2.ii Lipschitz maps
Let (N, δ) be a symmetric metric space, and f : N → M , g : M → N . We could
define that f and g are “Lipschitz w.r.t. b”, or “Lipschitz w.r.t. d”. That is, in the first
case
b(f (x), f (y)) ≤ Cδ(x, y), δ(g(x), g(y)) ≤ Cb(x, y)
for some constant C > 0. In the second case
d(f (x), f (y)) ≤ Cδ(x, y), δ(g(x), g(y)) ≤ Cd(x, y)
for some constant C > 0.
The two concepts are equivalent for f , but not equivalent for g.
As an example, if we provide M 2 with the metric

d2 ((x, y), (x0 , y 0 ))=d(x, x0 ) + d(y, y 0 ) ,

def

then (x, y) 7→ b(x, y) is Lipschitz w.r.t d2 , since

|b(x, y) − b(x0 , y 0 )| ≤ |b(x, y) − b(x, y 0 )| + |b(x, y 0 ) − b(x0 , y 0 )| ≤ d(x, x0 ) + d(y, y 0 )
and this implies that b is continuous.

§2.iii Length
We induce from b the length lenb γ of a curve γ : [α, β] → M , by using the total
variation (4)
n
X
lenb γ = sup
def

b γ(ti−1 ), γ(ti ) (2.5)
T i=1
where the sup is carried out over all finite subsets T = {t0 , · · · , tn } of [α, β] and
t0 ≤ · · · ≤ tn . If lenb γ < ∞, then γ is called rectifiable.

Remark 2.6 The above length is asymmetric: if γ̂(t)=γ(−t), then in general lenb γ
def

will be different from lenb γ̂; hence there does not exist a measure H on M such that the
length of a curve is the measure of its image, that is
H(image(γ)) = lenb γ ;
indeed image(γ) = image(γ̂) which is incompatible in the general case in which
lenb γ 6= lenb γ̂.
We define bg
bg (x, y) = inf lenb γ (2.7)
where the inf is taken in the class of all continuous curves γ connecting x to y.
Note that
lenb γ ≥ b(γ(α), γ(β)) (2.8)
so bg ≥ b. It is also easy to prove that bg satisfy a triangle inequality: so bg is an
asymmetric distance (possibly taking the value ∞; see 2.4 and 2.17 on this). bg is
called the geodesic (asymmetric) distance induced by b. We may endow M with
def
the distance dg (x, y)= bg (x, y) ∨ bg (y, x); then dg ≥ d, and the topology resulting in
(M, bg ) is finer than the one used in (M, b) (has more open sets and less compact sets);
the two topologies may easily differ, as in this simple example:
(4) according to Busemann [4], this definition goes back to Menger (1928)

5
Example 2.9 Consider

M = {x ∈ lR2 | − 1 ≤ x1 ≤ 1, x2 = 0} ∪
[
{x ∈ lR2 | − 1 ≤ x1 ≤ 1, x2 = (x1 − 1)/n}
n≥1
and b the Euclidean distance (see fig. 1). Then (M, b) is compact but (M, bg ) is not.

Figure 1: example 2.9

Sometimes the above infimum (2.7) is a minimum:

Definition 2.10 (Geodesics) Let’s fix x, y ∈ M such that bg (x, y) < ∞. If there is a
continuous curve γ providing the minimum of the equation (2.7) that defined bg (x, y),
then this curve is called a “(minimal) geodesic” (connecting x to y).

Suppose that γ is rectifiable. We define the running length ` : [α, β] → lR+ of γ to

be the length of γ restricted to [α, t], that is

`(t)=lenb γ|[α,t]
def

(2.11)

Note that:
• if γ is Lipschitz of constant L, then, by direct substitution in (2.5),

`(t) − `(s) ≤ (t − s)L

so ` is Lipschitz. (5)
In this case the derivative d`
dt exists for almost all t, and it is called the metric
derivative of γ (as defined in [1, 2]).
• Supposing that ` is continuous (for simplicity), for t ≥ s,

b γ(s), γ(t) ≤ bg γ(s), γ(t) ≤ `(t) − `(s)

(2.12)

since the length of γ restricted to [s, t] is `(t) − `(s), and γ is indeed a path
connecting γ(s) to γ(t);
• whereas the fact that ` is continuous does not imply that γ is continuous, as shown
in this example:

Example 2.13 Let M = [−1, 1] and let f : M × lR → lR be


|v| if v > 0 or x ≤ 0
f (x, v) = (2.13.?)
|v|/x if v < 0 and x > 0
(5) when b is symmetric, if γ is rectifiable and continuous, then ` is continuous; this seems to be true also for

asymmetric distances, but was not carefully checked

6
From this we induce the asymmetric distance as explained in the introduction §1.ii, to
get 
 |x − y| if x ≤ y or y < x ≤ 0
b(x, y) = − log(y) + log(x) if 0 < y < x (2.13.??)
+∞ if y ≤ 0 < x


then the topology τ + generated by forward balls is the usual euclidean topology on
[−1, 1]; whereas the topology τ − generated by backward balls makes [−1, 0] discon-
nected from (0, 1]; and τ = τ − .
γ : [−1, 1] → M defined by γ(t) = t has lγ (t) = t, but is not continuous.

This is counter-intuitive since it is contrary to what is seen in symmetric metric spaces;

so we will sometimes use this lemma

Lemma 2.14 Suppose that ` is continuous; then γ is continuous if and only if its image
is compact. (Proof follows from lemma 3.7 and (2.12)).

§2.iv Definitions
We add some definitions

• We say that the (asymmetric) metric space (M, b) is a path-metric space, or that
b is intrinsic, if b = bg ;

• and we say that it admits geodesics, or that geodesics exist, if, for all x, y ∈ M ,
there is a geodesic connecting x to y.

• A sequence (xn ) ⊂ M is called forward Cauchy if

∀ε > 0, ∃N ∈ lN such that ∀n, m, m ≥ n ≥ N, b(xn , xm ) < ε (2.15)

We say that (M, b) is forward complete if any forward Cauchy sequence (xn )
converges to a point x (according to the topology τ : cf. 2.3 and §3.ii.1 on the notion of
convergence).
(These definitions agree with those used in Finsler Geometry (see ch. VI of [3]). See 3.4
for a different definition.)

• We say that A ⊂ M is forward bounded if these two equivalent propositions

hold

– ∃x ∈ M, r > 0 s.t. A ⊂ B + (x, r)

– ∀x ∈ M, ∃r > 0 s.t. A ⊂ B + (x, r)

For any forward definition above there is a corresponding backward definition,

obtained by exchanging the first and the second argument of b, or by using the conjugate
distance b defined by
b(x, y) = b(y, x) ; (2.16)
and a corresponding symmetric definition, obtained by substituting d for b.
For the same reason, in this paper we will present mostly the forward versions of
the theorems, since backward results are obtained by substituting b for b.
We nonetheless emphasize these definitions.

7
 • We say that (M, b) is forward-locally compact if ∀x ∃ε > 0 such that
def
D+ (x, y)={y | b(x, y) ≤ ε}

is compact.

• We say that (M, b) is backward-locally compact if ∀x ∃ε > 0 such that

D− (x, y)={y | b(y, x) ≤ ε}

def

is compact.

• We say that (M, b) is symmetrically-locally compact if ∀x ∃ε > 0 such that

{y | d(x, y) ≤ ε} is compact.

• We say that (M, b) is locally compact if ∀x ∃ε > 0 such that both D− (x, y) and
D+ (x, y) are compact.

The following implications hold:

locally compact forward locally compact

backward locally compact symmetrically locally compact

Remark 2.17 In general, even if (M, b) is connected and b < ∞ at all points, it
may be the case that bg (x, y) = ∞ for some points; if we do not like this fact, we
may wish to try the remedies proposed in 2.4; in particular the equivalence relation
x ∼ y ⇐⇒ dg (x, y) < ∞ decomposes the space (M, bg ) in equivalence classes that
coincide with all connected components of (M, bg ) (since bg is itself path-metric, by
4.44).
Unfortunately the second remedy suggested in 2.17 is only a placebo when we are
interested in path-metric spaces and/or in studying geodesics: suppose b is path-metric
def
and we decide to set φ(t) = t/(1 + t) and define b̃=φ ◦ b: then b̃ is not path-metric; so if
we try to substitute b̃ by its generated path-metric distance b̃g we find out, by prop. 4.37,
that b̃g = bg = b, and we are back to square one. So we may sometimes be forced to
address the case when bg (x, y) = ∞ for some points.

3 Comparison with related works

§3.i Comparison with Busemann’s “General metric spaces”
The foundation of the theory of non-symmetric metric spaces was given by Busemann
in his paper Local Metric Geometry [4], where it was christened as “General metric
spaces”.
It should be noted that, although Busemann has presented the foundation of that
theory, most of his work makes the assumption that the distance be symmetric; such is
the case e.g. of the renowned work on The Geometry of Geodesics [5]. The study of
non-symmetric “General metric spaces” was carried on for example in Phadke’s [17]
and Zaustinsky’s [19].
We report now the theory of “General metric spaces” here, in a language that is
more similar to our presentation and to [19], than to [4].

8
 • a “General metric space” is a space satisfying all requisites in 2.1, and moreover

∀x ∈ M, ∀(xn ) ⊂ M, b(xn , x) → 0 iff b(x, xn ) → 0 (3.1)

It is easy to prove that this is equivalent to assuming that τ + = τ − = τ , where the

topology τ + (resp. τ − ) is generated by the families of forward (resp. backward)
balls alone;
• a “General metric space” is finitely compact if any closed set that is contained in
B + (x, r) ∪ B − (x, r) (for a choice of x ∈ M and r > 0) is compact
• a “General metric space” is weakly finitely compact if small closed forward
balls are compact. This is what was here defined as forward local compactness.

We add some comments that highlight the differences between the approach in the
present paper and in [4]:
• the additional hypothesis (3.1) used in defining “General metric spaces” is not
really restrictive; in many applications it does hold (for example, when the space
is locally compact, as in [16]); still it is possible to find examples of path-metric
spaces such as 4.13 and 2.13 where it does not hold (even though this latter comes
from a Calculus of Variations problem of the form discussed in introduction);
• on the other hand, in the aforementioned papers it was (sometimes silently)
assumed that the metric b be intrinsic; (6) that is, no distinction was made between
b and bg ; this is restrictive, since in many simple situations we will consider the
two do differ.
A good reason to consider non-path-metric spaces is as follows: suppose (M 0 , b0 )
is an asymmetric metric space, and (M, b) is a subset of it, with b being the
restriction of b0 ; then it easy to devise examples where (M 0 , b0 ) is path-metric but
(M, b) is not (7) and vice versa. For example, in the forthcoming paper [8] a
family of (symmetric) metric space are studied that are not (and cannot possibly)
be path-metric; the reason being this theorem (adapted from 2.6 in [8]):

Theorem 3.2 Suppose that (M, d) is a complete path-metric space, and that
i : M → E is an isometric immersion in a uniformly convex Banach space E:
then i(M ) is convex, and any two points in M can be joined by a unique minimal
geodesic (unique up to reparametrization).

Then (the symmetric version of) theorem 4.24 was quite useful in establishing
existence of geodesics in [8].
So in most of this paper we will carefully distinguish the two metrics b and bg .
• In most of the cited works (with the exception of Phadke’s [17]) the space is
assumed to be locally compact (or even finitely compact).
According to the notes at end of the introduction of [6], some work was carried
on at that time to extend known results to forward locally compact spaces; but the
results in this paper do not appear in the announced papers.
(6) Note that the definition of “intrinsic metric” was not used at the time of [4]: it was in a sense replaced by

the “Menger convexity” hypothesis (that we will discuss in section §4.i.1)

(7) many examples in this paper are built this way

9
§3.ii Comparison with quasi metric spaces
§3.ii.1 Topology
As already pointed out, the definition 2.1 of the asymmetric metric coincides with
the definition of a quasi metric that is found in the literature, cf. Kelly [14], Reilly,
Subrahmanyam and Vamanamurthy [13] (8) Fletcher and Lindgren [11, (pp 176-181)],
Künzi [15]. The main difference between our theory of asymmetric metric spaces
and quasi metric spaces is in the choice of the associated topology.
Indeed, we have three topologies at hand:
• the topology τ , generated by the families of forward and backward balls, or
equivalently by the metric d defined in (2.2);
• the topology τ + generated by the families of forward balls;
• the topology τ − generated by the families of backward balls;
it may happen that these three topologies are different.
This problem has been studied in [14]: there Kelly introduces the notion of a
bitopological space (M, τ + , τ − ), and extends many definition and theorems, (such as
the Urysohn lemma, the Tietze’s extension theorem, the Baire category theorem (9) ) to
these spaces. Unfortunately Kelly does not include the topology τ in his studies.
In many applications τ = τ + = τ − : see 3.8 and section §5.i. We have chosen to
associate the topology τ to the “asymmetric metric space”. τ is a symmetric kind of
object: as a consequence, we have only one notion of “open set”, of “compact set”, of
“the sequence (xn ) converges to x”, and of “the functions f : N → M and g : M → N
are continuous”. Furthermore
Proposition 3.3 If xn → x (according to τ ) then the sequence (xn ) is both a forward
Cauchy sequence and a backward Cauchy sequence,
in accordance with the symmetric case.

§3.ii.2 Cauchy sequences, and completeness

In papers on quasi metric spaces, the quasi-metric space (M, b) is instead usually
endowed with the topology τ + : this entails a different notion of convergence and
compactness, and poses the problem to find a good (10) definition of “Cauchy sequence”
and “complete space”.
This problem has been studied in [13], where 7 different notions of “Cauchy se-
quence” are presented. (Indeed, the list in [13] includes the three that we defined in §2.iv:
a “forward Cauchy sequence” (resp. backward) is a “left K-Cauchy sequence” (resp. right); a
“symmetrical Cauchy sequence” is a “b-Cauchy sequence”). Combining these 7 definition
with the τ + topology, [13] presents 7 different definitions of “complete space”. Actually,
by combining 7 “Cauchy sequences” with all the above 3 topologies, we may reach a total of 14
(!) different definitions of “complete space” (considering the symmetry of using the b instead of b,
see eq. (2.16)). To our knowledge, no one has taken the daunting task of examining all of them.
One of the notions of “Cauchy sequence” and “complete space” from [13] has been
further studied by Künzi [15]; we present it here.
(8) [13]provides also a wide discussion of the references on quasi metrics
(9) Another version of Baire theorem, using a better definition of completeness, is found in Thm 2 in [13].
(10) What do we mean by “good definition”? We may need that the definition would satisfy a proposition

similar to 3.3; see statement iv) below

10
Definition 3.4 • A sequence (xn ) ⊂ M is a “left b-Cauchy sequence” when
∀ε > 0 ∃x ∈ M and ∃k ∈ lN such that b(x, xm ) < ε whenever m ≥ k.

• (M, b) is a “left b-sequentially complete space” if any left b-Cauchy sequence

converges to a point, according to the topology τ + .

It is easy to prove that

i). if xn → x according to τ + then the sequence (xn ) is a “left b-Cauchy sequence”.

ii). Any “forward Cauchy sequence” (as defined in (2.15)) is a “left b-Cauchy se-
quence” (11) .

iii). If τ = τ + , then any “left b-sequentially complete space” is a “forward complete

metric space” as defined in this paper.
In case τ 6= τ + , the implication may not hold.

iv). Whereas, if xn → x according to τ + then the sequence (xn ) may fail to be

either a “forward Cauchy sequence” or a “backward Cauchy sequence”. (12) Cf.
example 4.13.(iv) here.

For those reasons, it is not easy to compare the results and examples in the above
papers, with the result and examples here presented.
From here on, we return to our setting of “asymmetric metric spaces”: we will
always use the topology τ .

§3.iii Comparison with symmetric case

There are subtle but important differences between asymmetric and symmetric distances.
We start with a striking remark.

Remark 3.5 Let ε > 0 and x, y, z be such that b(x, y) < ε and b(x, z) < ε. This does
not imply, in general, that b(y, z) < 2ε
z
x
?
y

This has consequences such as

i). In example 4.13.(iv) there exists a sequence such that b(xn , x) → 0 but xn 6→ x.
(Note that this example cannot be found in Busemann’s “General metric space”,
due to property (3.1)).

ii). Fix a sequence (xn ) ⊂ M . Suppose that ∀ε > 0 there exists a converging
sequence (yn ) such that b(yn , xn ) < ε. If b is symmetric and (M, b) is complete,
then (xn ) converges. If b is asymmetric and (M, b) is complete, then there is a
counter-example in 4.13.(vi).
(11) just choose n = N = k and x = xn in the definition of left b-Cauchy sequence
(12) Indeed, Kelly [14] had encountered this problem, which was a motivation of [13]

11
 iii). In example 4.14.(iv) there is a ball B + (a, r) and points ai 6∈ B + (a, r) such that
B + (a, r) ⊂ i B + (ai , r/2): it is then difficult to find a useful definition of a
S
“precompact set”. (13)

Further examples are in §4.i.2 and §4.ii.3.

To circumvent the above problems, we will use often the following propositions.

Proposition 3.6 If (xn ) ⊂ M is a sequence such that b(x, xn ) → 0, and M is forward-

locally compact, then xn → x.

The above may be proved by using this Lemma (that is similar to (2.3) and (2.6) in
Zaustinsky’s [19])

Lemma 3.7 (modulus of symmetrization) Let C ⊂ M be a compact set: then there

exists a continuous monotonic non decreasing function ω : lR+ → lR+ , with ω(0) = 0,
such that
∀x, y ∈ C, b(x, y) ≤ ω(b(y, x))

Proof. Define
f (r) = sup b(y, x)
x,y∈C, b(x,y)≤r

and then f is monotone. Since C is compact, then f < ∞ and limr→0 f (r) = 0:
otherwise we may find ε > 0 and xn , yn s.t. b(xn , yn ) → 0 while b(yn , xn ) > ε:
extracting converging subsequences, we obtain a contradiction.
R 2r
From f we can define an ω as required, for example ω(r) = 1r r f (s)ds (note
that ω ≥ f ).

This lemma may also be used as follows

Corollary 3.8 If (M, b) is locally compact then ∀x ∈ M, ε > 0 ∃r > 0 s.t.

B + (x, r) ⊂ B − (x, ε), B − (x, r) ⊂ B + (x, ε)

and then τ = τ + = τ − .

Therefore, in locally compact spaces, asymmetry is not important for topological

questions (in the sense explained in §3.ii.1); but we still care for asymmetry in metric
questions (such as completeness). Cf. example 4.14 and sec. §5.i.

4 Study of asymmetric metric spaces

§4.i Properties
Proposition 4.1 For any A ⊂ M , the following properties are equivalent

• supx,y∈A b(x, y) < ∞

• A is forward bounded and backward bounded

• A is symmetrically bounded
(13) An useful definition of “precompact set” is found in [15], where, though, a different completeness

hypothesis is used (see 3.4 here).

12
But note that in example 4.13.(v) there exists a set A ⊂ M that is forward bounded but
not backward bounded (and vice versa in 4.14.(i)).
Similarly
Proposition 4.2 Let (xn ) ⊂ M be a sequence; then the following are equivalent
• (xn ) is forward Cauchy and backward Cauchy
• (xn ) is symmetrically Cauchy
From that we obtain: if (M, b) is either forward or backward complete, then it is
symmetrically complete.
See examples 4.14, 4.15.
Proof. Suppose that (xn ) is symmetrically Cauchy: then ∀ε > 0 ∃N such that ∀n, m >
N , d(xn , xm ) < ε: then b(xn , xm ) < ε.
Suppose that (xn ) is forward Cauchy and is backward Cauchy: ∀ε > 0 ∃N 00 such
that ∀n > m > N 00 , b(xn , xm ) < ε, and ∃N 0 such that ∀n > m > N 0 , b(xm , xn ) < ε:
then we let N = N 0 ∨ N 00 , and ∀n, m > N , d(xn , xm ) < ε.
Suppose that (M, b) is forward complete; let (xn ) be symmetrically Cauchy: then it
is forward Cauchy, and then, since (M, b) is forward complete, there is an x such that
xn → x. Similarly if (M, b) is backward Cauchy.

Proposition 4.3 Suppose that (xn ) is forward Cauchy and there exist subsequence nk
and a point x such that limk xnk = x: then limn xn = x.
Proof. fix ε > 0; since (xn ) is forward Cauchy, ∃N such that ∀m, m0 with m0 ≥ m ≥
N , b(xm , xm0 ) ≤ ε; let H be such that nH ≥ N and ∀k ≥ H, d(x, xnk ) ≤ ε; for
n ≥ nH ,

b(x, xn ) ≤ b(x, xnH ) + b(xnH , xn ) ≤ d(x, xnH ) + b(xnH , xn ) ≤ 2ε

at the same time, choosing a large h such that nh ≥ n,

b(xn , x) ≤ b(xn , xnh ) + b(xnh , x) ≤ b(xn , xnh ) + d(xnh , x) ≤ 2ε

so in conclusion d(xn , x) ≤ 2ε

§4.i.1 Mid-point properties

In path metric spaces, for any two points x, y, there is always a curve joining them
with a quasi optimal distance (that is, very near to b(x, y)): this is used in the following
proposition, in many different but equivalent fashions, to find approximate intermediate
points z at a prescribed distance b from x and y.

Proposition 4.4 Suppose that the (asymmetric) metric space (M, b) is path-metric:
i). let ρ > 0 and
def
S + (a, ρ)={y | b(a, y) = ρ}
def
D+ (a, ρ)={y | b(a, y) ≤ ρ}
(which are closed, since b is continuous) then

B + (a, ρ) = D+ (a, ρ), S̊ + (a, ρ) = ∅ (4.4.?)

13
 ii). ∀θ ∈ (0, 1), ∀x, y ∈ M

∀ε > 0 ∃z ∈ M , such that

b(x, z) < θb(x, y) + ε, b(z, y) < (1 − θ)b(x, y) + ε (4.4.??)

iii). let x, y ∈ M and ε > 0, ε ≤ b(x, y) then

inf b(z, y) = b(x, y) − ε (4.4.)

z∈S + (x,ε)

inf b(z, y) = b(x, y) − ε . (4.4.)

z∈D + (x,ε)

Note that, if there exists a minimum point z ∈ D+ (x, ε) to (4.4.), then b(x, z) = ε,
i.e. z ∈ S + (x, ε). (14)
In the general case of a space (M, b), all the above applies to the metric space
(M, bg ) (that is path-metric, by 4.44).
On the other hand,
Proposition 4.5 if M is either forward or backward complete, and any one of the
following properties holds:
• ∃θ ∈ (0, 1) s.t. ∀x, y, property (4.4.??) holds,
• ∃θ ∈ (0, 1) s.t. ∀x, y, when ε = θb(x, y), property (4.4.) holds,
• ∃θ ∈ (0, 1) s.t. ∀x, y, when ε = θb(x, y), property (4.4.) holds,
then (M, b) is path-metric.
By using Zorn’s Lemma, the above may be strengthened to this statement
Proposition 4.6 if M is symmetrically complete, and any one of the following proper-
ties holds:
• ∀x, y, ∃θ ∈ (0, 1) s.t. property (4.4.??) holds,
• ∀x, y, there exists ε > 0, ε < b(x, y) s.t. property (4.4.) holds,
• ∀x, y, there exists ε > 0, ε < b(x, y) s.t. property (4.4.) holds,
then (M, b) is path-metric.
Regarding the completeness hypothesis, see example 4.11.
All of the above results related the existence of approximate intermediate points
to the fact that (M, b) be path-metric. If M admits geodesics, then, for any two
points x, y, there is always a curve joining them with optimal distance bg (x, y) (but not
necessarily equal to b(x, y)): this is again used to find exact intermediate points z at a
prescribed distance bg from x and y; this is at the base of the definition of “Menger
convexity”; unfortunately, there are, in the literature, two different definitions:
• M is “Menger (weakly) convex” if given two different points x, y ∈ M , a third
(different from x, y) point z exists such that bg (x, z) + bg (z, y) = bg (y, z);
(14) Indeed, b(x, z) ≥ b(x, y) − b(z, y) by triangular inequality, and b(x, y) − b(z, y) = ε by (4.4.);

while b(x, z) ≤ ε since z ∈ D+ (x, ε)

14
 • M is “Menger (strongly) convex” if given any two points x, y ∈ M , and any
0 ≤ λ ≤ bg (y, z), a third point z exists such that bg (x, z) = λ, bg (z, y) =
λ − bg (y, z);

the weaker definition being used for example in Busemann [4] in proposition (1.16), the
stronger in more modern treaties.

Example 4.7 The subset M = lR3 \ {(x1 , x2 , x3 ), x1 = x2 = 0} obtained by deleting

a line from 3-space, endowed with the Euclidean distance b(x, y) = |x − y|, is a simple
example of a space that is path metric, it is Menger weakly convex, but not Menger
strongly convex. It is also locally compact but not finitely compact (as defined in pages
8 and 9) and not complete.

Other similar examples are in 4.11.

It is easy to see that, if b admits geodesics then it is Menger convex (both in weaker
and stronger sense). Vice versa,

Proposition 4.8 if M is either forward or backward complete, and it is Menger strongly

convex, then (M, b) admits geodesics.
If M is symmetrically complete, and it is Menger weakly convex, then (M, b) admits
geodesics.

In the case when Menger convexity is intended in the weak form, the proof needs Zorn’s
Lemma. The above result is similar to proposition (1.16) in [4], (where though the
hypothesis was “M is Menger (weakly) convex and finitely compact”).

Remark 4.9 We may add an intermediate statement: if M is either forward or back-

ward complete, and

∃θ ∈ (0, 1) , ∀x, y ∈ M , ∃z ∈ M , such that

bg (x, z) = θbg (x, y), bg (z, y) = (1 − θ)bg (x, y) . (4.9.?)

then (M, b) admits geodesics.

The above property (4.9.?) is stronger than the weak Menger convexity, and weaker than
the stronger. (15) Still, it is a condition that entails convexity, as seen in this example

Example 4.10 Let b(x, y) = |x − y|, M ⊂ lRn . If M is closed in lRn and (4.9.?) holds,
then M is convex (in the usual Euclidean sense).

§4.i.2 Examples
We consider several examples, for a better understanding of the above properties and
definitions.
This subset of lR2 complements the already seen example 4.7

Example 4.11 The set M = (lR × Q) ∪ ({0} × lR), equipped with b(x, y) = |x − y|,
admits approximate intermediate points, in the various forms seen in proposition 4.4;
and it is Menger weakly convex; and M is Lip-arc-connected; but (M, b) is not complete;
and (M, b) is not path-metric, since bg (x, y) = |x1 | + |y1 | + |x2 − y2 | when x2 6= y2 .
(15) Indeed “Menger weak convexity” may be equivalently stated as: ∀x, y ∈ M ∃θ ∈ (0, 1) s.t. bg (x, z) =

θbg (x, y), bg (z, y) = (1 − θ)bg (x, y) .

15
Example 4.12 Consider M ⊂ lRn to be an open set, and b to be the Euclidean distance;
then

• (M, b) is locally compact and locally path-metric; whereas

• (M, b) admits geodesics iff M is convex,

• and (M, b) is complete iff M = lRn ;

• if moreover M is equal to the interior of the closure of M in lRn , then (M, b) is

path-metric iff M is convex.

Example 4.13 Let M be the disjoint union of segments

def
[
M= ({n} × [0, 1])/ ∼
n∈Z

with ∼ identifying all 0s into a class [0], and 1s into [1], i.e. (n, 0) ∼ (m, 0) and
(n, 1) ∼ (m, 1). Let ln = {n} × (0, 1). Let b be defined on segments as

def |x − y|(1 + en ) if x ≥ y and x, y ∈ ln
b(x, y)=
|x − y|(1 + e−n ) if x ≤ y and x, y ∈ ln

and
b([0], [1]) = b([1], [0]) = 1
and extend b (geodesically) using the rule

b(x, y) = inf b(x, y), b(x, [0]) + b([0], y), b(x, [1]) + b([1], y)

We highlight the following properties

i). (M, b) is a complete path-metric space.

ii). (M, b) is not locally compact.

iii). There is no geodesic connecting [0] to [1]

iv). Let xn = (−n, 1/n) then

b(xn , [0]) = (1 + e−n )/n → 0

but
b([0], xn ) = (1 + en )/n → ∞

16
 forw back
4 4
fbr(a,r) bbr(a,r)
a a
3 fbl(a,r) 3 bbl(a,r)

2 2

1 1

0 0

-1 -1

-2 -2

-3 -3

-4 -4
-4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 4

Figure 2: forward and backward balls, left and right extrema (M is vertical, r = 1/2, a
in abscissa)

v). Moreover the set A = {[0], xn | n} is backward bounded but not forward
bounded.

vi). Moreover ∀ε > 0 we can define


xn if nε < 2
yn =
[0] if nε ≥ 2

and then b(xn , yn ) < ε and yn → [0]; but nonetheless xn 6→ [0].

Example 4.14 Let M = lR and let

ey − ex

if x < y
b(x, y) = (4.14.?)
e−y − e−x if x > y
then b generates on lR the usual topology, and (M, b) is locally compact and path-metric
(this may be proved using 5.6 since b comes from a Finsler structure: see in sec. 5.3.1 in
[16]). The balls are the open intervals

B + (a, r) = {y | b(a, y) < r} = − log(r + e−a ), log(r + ea )

B − (a, r) = {x | b(x, a) < r} = log(ea − r), − log(e−a − r)

where log(z) = −∞ if z ≤ 0. (see fig. 2 and 3)

i). We immediately note that B − (0, 1) = lR, that is, M is backward bounded (but is
not forward bounded).

17
 forw back
4 4
fbr(a,r) bbr(a,r)
a a
3 fbl(a,r) 3 bbl(a,r)

2 2

1 1

0 0

-1 -1

-2 -2

-3 -3

-4 -4
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Figure 3: forward and backward balls, left and right extrema (M is vertical, a = 1/2, r
in abscissa)

ii). Let xn = n: then for m < n, b(xn , xm ) = e−m − e−n < e−m so this sequence
is backward-Cauchy: then M is not backward complete.

iii). (M, b) is forward complete (and then is symmetrically complete, by 4.2). Proof:
Suppose that (xn ) is an increasing forward-Cauchy sequence: then for m > n > N ,
b(xn , xm ) = exm − exn < ε that implies xm ≤ log(ε + exn ), so xm has a limit
xm → x (16) ; while if it were a decreasing forward-Cauchy sequence for m > n > N ,
b(xn , xm ) = e−xm − e−xn < ε that implies xm ≥ − log(ε + e−xn ), and again
xm has a limit. Suppose that xn is a generic forward-Cauchy sequence: from any
subsequence xnk of xn we may extract a monotonic sub-sub-sequence (xnk h )h : this
would be convergent to a point x; this point does not depend on the choice of subsequence:
indeed, b(xn , xnk h ) →h b(xn , x) < ε for n large.

iv). Consider the ball B + (0, ρ) of extrema (−R, R) with R(ρ) = log(ρ + 1), and
then the two balls

B + (R, r) = −log(r+e−R ), log(r+eR ) , B + (−R, r) = −log(r+eR ), log(r+e−R )

then if r ≥ ρ/(ρ + 1),

B + (0, ρ) ⊂ B + (R, r) ∪ B + (−R, r)

Proof: indeed the right extrema of B + (−R, r) is positive when log(r + e−R ) ≥ 0 that is
(r + e− log(ρ+1) ) ≥ 1 that is (r + 1/(ρ + 1)) ≥ 1

(16) x → x in the topological sense, i.e. d(xm , x) → 0

m

18
Example 4.15 Consider two copies (M, b) of the above space, join them at the origin,
reverse the metric on one: the resulting space M̃ , b̃ is symmetrically complete, but is
neither forward nor backward complete.
More precisely: let M+ = lR × {+}, M− = lR × {−},

M̃ = M+ ∪ M− / ∼

where (0, +) ∼ (0, −). Let


 b(x, y) x, y ∈ M+
b̃(x, y) = b(y, x) x, y ∈ M−
b̃(x, [0]) + b̃([0], y) otherwise


§4.ii Geodesics
Lemma 4.16 (change of variable) Let γ : [α, β] → M , ξ : [α1 , β1 ] → M be linked
by
γ =ξ◦ϕ
where ϕ : [α, β] → [α1 , β1 ] is monotone non decreasing, and continuous, and ϕ(α) =
α1 , ϕ(β) = β1 . Then
lenb γ = lenb ξ

Proof. For any partition T = {ti } of [α, β] we associate the partition S = {si = ϕ(ti )}
of [α1 , β1 ]: in this case,
n
X n

X

b γ(ti−1 ), γ(ti ) = b ξ(si−1 ), ξ(si ) (4.17)
i=1 i=1

Similarly, for any partition S = {si } of [α1 , β1 ] we choose ti ∈ ϕ−1 ({si }), and
associate the partition T = {ti } of [α, β]: then again (4.17) is verified regardless of the
choice of ti : indeed if t0i ∈ ϕ−1 ({si }) then γ(ti ) = γ(t0i ).

Remark 4.18 The above lemma holds also when ϕ is not bijective: this is very powerful,
since it simplifies the next lemma, and leads to a very short and simple proof of 4.24
and of 4.29.

The following is a long–known but very powerful result (that may be found in
section I in [4]; see also in Theorem 4.2.1 in [1], that, for the symmetric case, also
discusses the metric derivative issue):
Lemma 4.19 (reparametrization to arc parameter) For any rectifiable curve
γ : [α, β] → M of length L such that `γ is continuous, there exists an unique Lipschitz
curve ξ : [0, L] → M such that

γ(t) = ξ(`γ (t)), `ξ (t) = t, t ∈ [0, L] (4.19.?)

where `γ is the running length of γ, and `ξ of ξ (see eq. (2.11)).

ξ is the reparametrization to arc parameter of γ.
Proof. γ(t) = ξ(`γ (t)) uniquely defines ξ: indeed, if `γ (t) = `γ (t0 ) then γ(t) = γ(t0 )
by (2.12). Let ˆl = `γ (t̂). If we restrict γ to [a, t̂] and ξ to [0, ˆl], we can write γ = ξ ◦ `γ .
By the previous lemma, `γ (t̂) = `ξ (ˆl), that is ˆl = `ξ (ˆl).

19
 In general we will say that a curve ξ is parametrized by arc parameter when `ξ (t) =
t. Combining (2.12) and (4.19.?), we can state that, for any curve ξ parametrized by
arc parameter, for 0 ≤ s < t ≤ L,
bg ξ(s), ξ(t) ≤ t − s


(4.20)
This leads to the following definition (due to Busemann [5], [6]):
Definition 4.21 (Segments) A curve γ : I → M (where I ⊂ lR is an interval) is a
segment if
∀s, t ∈ I, s < t , t − s = bg (γ(s), γ(t)) (4.21.?)
The “time parameter” is arbitrary: if γ(·) is a segment, then γ(· + t) is a segment as
well.
Unfortunately, a segment may fail to be continuous: this was seen in example 2.13
where the curve γ is a segment but is not continuous. The above example cannot be
found in Busemann’s general metric spaces, since it contradicts the extra hypothesis
(3.1). That example is one of the reasons why we stressed the necessity that curves be
continuous in the definition 2.10 of what a geodesic is.
Segments enjoy some nice properties:
• by 4.43, a segment is always parametrized by arc parameter, and the length of a
segment defined on I = [α, β] is β − α ;
• then a continuous segment γ is a minimal geodesic connecting γ(s) to γ(t),
∀s, t ∈ I, s < t.
Vice versa,
Lemma 4.22 (almost segment) Suppose γ : [0, L] → M is a curve parametrized by
arc parameter, connecting γ(0) = x to γ(L) = y; and suppose that
lenγ ≤ bg (x, y) + ε
for ε ≥ 0; then for any 0 ≤ s < t ≤ L,
t − s − ε ≤ bg γ(s), γ(t) ≤ t − s


(4.22.?)
(that is an approximate version of (4.21.?)); the rightmost inequality is (4.20); whereas
bg (x, y) ≤ bg x, γ(s) + bg γ(s), γ(t) + bg γ(s), y ≤

≤ s + bg γ(s), γ(t) + L − t .

again by (4.20).
So (choosing ε = 0) we obtain that a minimal geodesic is a continuous segment.
Segments can be joined, as follows:
Lemma 4.23 (joining) Suppose that γ : [0, L] → M , γ : [0, L0 ] → M , are geodesic
segments. Suppose that z = γ 0 (0) = γ(L): then the segments may be joined to form a
continuous curve ξ : [0, L + L0 ] → M , by defining

γ(t) for t ∈ [0, L]
ξ(t) =
γ 0 (t − L) for t ∈ [L, L + L0 ].
Then ξ is a geodesic segment if and only if
bg ξ(0), ξ(L + L0 ) = L + L0 .

20
§4.ii.1 Existence of geodesics
The following results rephrase (5),(6),(7) in section I in [6], but without assuming that
(M, b) be path–metric (by properly distinguishing b by bg ):

Theorem 4.24 Fix x. Let ρ > 0. Suppose

def
D+g (x, ρ)={y | bg (x, y) ≤ ρ}

is compact (in the (M, b) topology (17) ). Then for any y such that bg (x, y) ≤ ρ, there is
a geodesic connecting x to y.

Note that it is not sufficient to assume that D+ (x, ρ) is compact (instead of D+g ):
see in example 4.34. It is instead enough to require that D+ (x, ρ) is compact for all
ρ > 0 (as noted also in remark 4.3.3. in [1]).
In the same spirit we state those corollaries

Corollary 4.25 • Suppose D+g (x, ρ) and

D−g (x, ρ)={y | bg (y, x) ≤ ρ}

def

are compact, and bg (z, x) + bg (x, y) ≤ ρ and bg (y, x) + bg (x, z) ≤ ρ, then z, y

may be joined by a geodesic.

• As a consequence, if D+g (x, 2ρ) and D−g (x, 2ρ) are compact, and z, y ∈
D−g (x, ρ) ∩ D+g (x, ρ), then z, y may be joined by a geodesic.

We now prove the theorem; this needs some extra steps w.r.t. the proof in General
Metric Spaces (where (3.1) is assumed to hold).

Proof. Fix x, y ∈ M , ρ > 0, as above. We will write D+g instead of D+g (x, ρ) for
brevity. By lemma 3.7, let ω be the modulus of symmetrization of D+g (x, ρ); let
def
ω̃(r)= sup{r, ω(r)}: then

d(z, y) ≤ ω̃ b(z, y) (4.26)
for any z, y ∈ D+g .
Let L = bg (x, y); suppose y 6= x (otherwise γ ≡ x is the geodesic). Let γn :
def
[0, Ln ] → M be a sequence of continuous paths from x to y such that Ln =lenγn → L.
By using the above lemma 4.19, we assume without loss of generality that γn are
parametrized by arc parameter; by (4.26) above,


d γn (t), γn (s) ≤ ω̃(|t − s|) (4.27)

for all s, t ∈ [0, Ln ]. To confront different paths, we extend so that γn (t) = γn (Ln ) for
t > Ln .
Suppose also that L = bg (x, y) < ρ for simplicity; then definitively Ln ≤ ρ: by
(2.12), we know that all of γn is contained in D+g . Combining this argument and
(4.27) we can apply the Ascoli-Arzelà theorem: we know that there is a γ : [0, L] → M
(that again satisfies (4.27)) such that, up to a subsequence, there is uniform convergence
of γn → γ; this uniform convergence is w.r.t the distance d(x, y) = b(x, y) ∨ b(y, x).
(17) We recall that if D ⊂ M is compact in (M, bg ) then it is compact in (M, b); the opposite is not true, as

shown by example 2.9.

21
The functional γ 7→ lenb γ is lower semicontinuous w.r.t. uniform convergence, since it
is the supremum of the continuous functionals
n
X

b γ(ti−1 ), γ(ti )
i=1

so we conclude that γ is a geodesic connecting x to y.

When L = bg (x, y) = ρ, if γn is frequently wholly contained in D+g , all works as
above; otherwise the proof is obtained by a slight change in the above argument. Let tn
be the last of the times such that γn ([0, t]) ⊂ D+g , that is,
def
tn = inf{s | γn (s) 6∈ D+g (x, ρ)} = inf{s | bg (x, γn (s)) > ρ}

then tn ≥ ρ, since γn is continuous and by arc parameter. Define


γn (t) t < tn
γ̃n (t) = ; (4.28)
γn (tn ) t ≥ tn

since γ̃n are wholly contained in D+g , we can apply the above reasoning to say that
there is a continuous γ̃ such that γ̃n → γ̃ uniformly. To conclude the proof we need to
prove that γ̃(L) = y: since L = ρ ≤ tn ≤ Ln , then tn → L; moreover

b(γn (tn ), y) ≤ bg (γn (tn ), y) = Ln − ρ

so by (4.26) again, γn (tn ) → y: the sequence γ̃n is uniformly equicontinuous, this

implies that γ̃(L) = y.
Those results immediately imply what is nowadays known as Busemann’s theorem
Theorem 4.29 Suppose that (M, b) is compact and Lipschitz-arc-connected: then it
admits geodesics.

§4.ii.2 Hopf-Rinow Theorem

We now restate and prove theorem 1.2. To this end we prove some preliminary lemma.
def
Lemma 4.30 (radius of compactness) Let D+ (a, ρ)={y | b(a, y) ≤ ρ}.
We define the forward radius of compactness R : M → lR+ ∪ {∞} as
def
R(x)= sup{ρ ≥ 0 | D+ (x, ρ) is compact } .

Then, for any ρ < R(x), D+ (x, ρ) is compact.

Either R ≡ ∞, or R < ∞ and R is 1-Lipschitz w.r.t. d: indeed, for any x, y ∈ M ,
for any ρ < R(x), we have that D+ (y, ρ − b(x, y)) is compact, since

D+ (y, ρ − b(x, y)) ⊂ D+ (x, ρ) .

This implies that R(y) ≥ ρ − b(x, y), and then R(y) ≥ R(x) − b(x, y); if R is
finite, the above entails d(x, y) ≥ |R(x) − R(y)|, since x, y are arbitrary.
In general (even when R(x) < ∞) it is possible to find examples where D+ (x, R(x))
is compact, and examples where it is not. In path–metric and forward-locally compact
spaces, instead, we can precisely describe the behaviour of R(x) as follows

22
Lemma 4.31 Suppose that (M, b) is path–metric and forward-locally compact. Choose
def
x ∈ M, ρ > 0 such that R(x) < ∞ and D+ (x, ρ) is compact; let δ = miny∈D+ (x,ρ) R(y)
(note that δ > 0); then R(x) = ρ + δ.

Proof. Since (M, b) is forward-locally compact, then δ > 0. Choose 0 < t < s < δ;
we want to prove that [
def
V= D+ (x, t)
y∈D + (x,ρ)

is compact; indeed if yn ∈ V , then yn ∈ D+ (xn , t), for a choice of xn ∈ D+ (x, ρ);

up to a subsequence, xn → x ∈ D+ (x, ρ), so that for n large, b(x, xn ) ≤ s − t hence
b(x, yn ) ≤ s, that is, yn is contained in the compact set D+ (x, s), so we can extract a
converging subsequence.
Assuming that (M, b) is path–metric, then V ⊃ D+ (x, ρ + t) (by using (4.4.),
with ε = ρ), so V ⊃ D+ (x, ρ + t) (by (4.4.?)) and then D+ (x, ρ + t) is compact; and
also V = D+ (x, ρ + t), by existence of geodesics 4.24. This implies that R(x) ≥ ρ + t,
and we conclude by arbitrariety of t that R(x) ≥ ρ + δ. The opposite inequality is easily
inferred from the previous lemma, where it was proved that R(y) ≥ R(x) − b(x, y),
that implies δ ≥ R(x) − ρ.

A corollary of the above lemma is that (in the above hypothesis) D+ (x, R(x)) is not
compact; so the above lemma is the quantitative version of the argument shown in (8) in
section I in [6].

Lemma 4.32 Suppose that (M, b) is path–metric, for simplicity (18) . Suppose that
0 < R(x) < ∞, let ρ = R(x); then for any y with b(x, y) = ρ there is a continuous
geodesic segment γ : [0, ρ) → M such that γ(0) = x and

b(γ(t), y) = ρ − t ∀t ∈ [0, ρ]. (4.32.?)

(or, equivalently, γ is segment on [0, ρ] but it is continuous only on [0, ρ)).

Proof. Let z1 = x, and ρn = (1 − 2−n )ρ; we would like to define zn+1 as a minimum
point for the problem
min −n b(z, y) ; (4.32.??)
+ z∈D (zn ,2 ρ)

iteratively, using 4.4, we may prove those facts: b(x, zn ) ≤ ρn−1 and

D+ (zn , 2−n ρ) ⊂ D+ (x, ρn ) (4.32.)

where the r.h.s is compact: so the above problem has a minimum zn+1 that satisfy
b(zn , zn+1 ) = 2−n ρ, b(x, zn+1 ) ≤ ρn and

b(zn+1 , y) = 2−n ρ . (4.32.)

Using (4.32.) and 4.24, there exists a geodesic segment γn connecting zn to zn+1
(of length 2−n ρ). By using (4.32.??),(4.32.) and the triangle inequality for h > n

b(zn , zh ) + 2−h+1 ρ = b(zn , zh ) + b(zh , y) ≥ b(zn , y) = 2−n+1 ρ

(18) It is possible to state a more general lemma without assuming that (M, b) is path–metric; but the added

complexity is not useful in the following theorem

23
but on the other side
h−1
X h−1
X
b(zn , zh ) ≤ b(zj , zj+1 ) = ρ2−j
j=n j=n

we obtain that
b(zn , zh ) = (2−n+1 − 2−h+1 )ρ
so by the joining lemma 4.23 we can join all the segments γn to build the required
segment γ. By (4.32.), we obtain (4.32.?).

The example 2.13 shows that we cannot expect that γ(t) be continuous at ρ in general:
indeed the “identity curve” γ : [0, 1] → [0, 1] is a segment but is not continuous at 0.
We now use the above lemmas to prove the asymmetric Hopf–Rinow theorem (in a
more detailed version than what was seen in the introduction):
Theorem 4.33 (Hopf-Rinow) Suppose that (M, b) is path-metric and forward-locally
compact; then the following are equivalent
i). forward-bounded and closed sets are compact,
ii). (M, b) is forward complete,
iii). any continuous geodesic segment γ : I → M defined on I = [α, β) may be
completed to a continuous geodesic segment on [α, β],
and all imply that (M, b) admits geodesics.
In general, the implications i=⇒ii=⇒iii hold for any asymmetric metric space.
Proof. • Suppose that forward-bounded closed sets are compact. If xn is a forward-
Cauchy sequence, then there exists N s.t. b(xN , xm ) ≤ 1 for m > N , that is,
xm ∈ D+ (xN , 1) that is compact; then we can extract a converging subsequence,
and use lemma 3.7 to obtain the result.
• Suppose that (M, b) is forward complete, let γ be the segment; then as tn ↑ β,
the sequence γ(tn ) is forward–Cauchy, so there is an unique limit limn γ(tn ) that
we use to define γ(β).
• Suppose that any segment γ : I → M defined on I = [α, β) may be completed:
we will prove that forward–bounded closed sets are compact (that is, that the
radius of compactness R(x) ≡ ∞).
We proceed by contradiction, and suppose that there is a x s.t. ρ = R(x) < ∞:
we will show that D+ (x, ρ) is compact, contradicting 4.31. To this end let
{yn } ⊂ D+ (x, ρ). Let Ln = b(x, yn ).
If lim inf n Ln < ρ, we can extract a subsequence nk s.t. Lnk ≤ t < ρ, that is,
{ynk } ⊂ D+ (x, t) that is compact: so we can extract a converging subsequence.
Suppose now that limn Ln = ρ. For any yn s.t. Ln < ρ, we use 4.24 to
obtain the minimal geodesic segments γn : [0, Ln ] → M connecting x to yn ;
we extend γn constantly to [Ln , ρ] (as was done in eqn. (4.28)). When instead
Ln = ρ we use 4.32 to define γn : [0, ρ) → M , and define γn (ρ) = yn . We
use Ascoli–Arzelà theorem and a diagonal argument to find a subsequence nk
and a curve γ : [0, ρ) → M such that for each 0 < s < ρ, limk γnk (t) = γ(t)

24
 uniformly for t ∈ [0, s]. Since the length functional is lower-semicontinuous,
we obtain that γ is a continuous geodesic segment on [0, ρ); by hypothesis, it can
be completed to a continuous segment on [0, ρ]. Let y = γ(ρ) Fix now ε > 0;
since γ is continuous, there is a t s.t. b(y, γ(s)) < ε for all s, t ≤ s ≤ ρ; fix
s = t ∨ ρ − ε; we apply the triangular inequality to prove that
b(y, yn ) ≤ b(y, γ(s)) + b(γ(s), γn (s)) + b(γn (s), yn ) ;
this is less than 3ε, definitively in n; so b(y, yn ) → 0, and then yn → y (by
prop. 3.6).
• Existence of geodesics is guaranteed by theorem 4.24.

We remark that the proof of the above equivalence cannot simply follow from the
proof for metric spaces §1.11 in [12], since that proof uses the property (ii) in sec. §3.iii;
neither it does follow from the proof in Finsler Geometry (see section VI of [3]), since
the latter uses the exponential map.
Hence we devised a different proof, that combines the idea of the diagonalization of
a sequence of subsequences, as in [12], and the idea of repeated application of (4.4.)
as in [3]; it is similar to the proof for general metric spaces in §I in [6], but it does
not use the hypothesis (3.1), and it exploits the special ingredients of the radius of
compactness Lemma and of 4.3.

§4.ii.3 More examples

We show some examples to highlight the impact of the hypotheses in the above theorems
4.33 (a.k.a. 1.2) and 4.24.
We tweak the example 2.9 a bit, to build this:
Example 4.34 Consider
M = {x ∈ lR2 | − 1 ≤ x1 ≤ 1, x1 6= 0, x2 = 0} ∪
{x ∈ lR2 | − 1 ≤ x1 ≤ 1, 1/4 ≥ x2 > 0} ∪
[
{x ∈ lR2 | − 1 ≤ x1 ≤ 1, x2 = (x1 − 1)/n}
n
and b the Euclidean distance (see fig. 4 on the next page). Let A = (−1, 0), B = (1, 0).
(M, b) is locally compact, but is not path-metric and is not complete; (M, b) does not
admit geodesics locally around A, that, the points A and xn = (−1, −2/n) do not
admit a minimal connecting geodesic.
The disc {y | b(B, y) ≤ 1/2} is compact for b; but the discs {y | bg (B, y) ≤ ε} are
never compact for bg .

Example 4.35 Let

[
{x ∈ lR2 | x21 + n2 x22 = 1, x2 ≥ 0}
def
M=
n

and b be the geodesic distance induced on M by the Euclidean distance; loosely

speaking, M is the disjoint union of countable segments ln , with length ∼ 2 + 1/n, and
with common end points.
Then (M, b) is path-metric, is complete, and is bounded, and it locally admits
geodesics; but is not locally compact (and then it is not compact), and there is no
geodesic curve connecting (−1, 0) to (1, 0).

25
 A B

Figure 4: example 4.34

§4.iii Complements
§4.iii.1 Generating (path) metrics
For any asymmetric metric b, we may build many other asymmetric metrics b̃, as follows

Proposition 4.36 Let ϕ : lR+ → lR+ be continuous and concave, ϕ(x) = 0 only for
def
x = 0; let b̃=ϕ ◦ b: then b̃ is an asymmetric distance.
def
The above family b̃=ϕ ◦ b is quite large; hence the following question arises: what
happens if we induce b̃g from b̃: can we build a large family of path metrics on M , by
this simple procedure? The answer is no.

Proposition 4.37 Set everything as in the previous proposition. If moreover the deriva-
tive of ϕ exists and is finite at 0, then

ϕ0 (0)lenb γ = lenb̃ γ

and
bg (x, y)ϕ0 (0) = b̃g (x, y) .

Proof. Let ε > 0. In the definition (2.5) of lenb γ it is not restrictive to use only subsets
T of [α, β] such that b(γ(ti ), γ(ti+i )) ≤ ε ∀i ∈ {1, . . . , n − 1}.
Let now δ > 0; then there exists a ε > 0 such that

x(a − δ) ≤ φ(x) ≤ x(a + δ) ∀x ∈ [0, ε]

with a = ϕ0 (0); we obtain that

(a − δ)lenb γ ≤ lenb̃ γ ≤ (a + δ)lenb γ

hence the conclusion

If ϕ0 (0) = ∞, wild things may happen: see for example 1.4.b in [12].
It is also possible to prove that we cannot generate other path metrics by iterating
the operation (2.7): indeed bg = (bg )g ; see 4.44.

§4.iii.2 Length structure

We hereby define a length structure len : C → lR+ , where C is a family of curves
γ : [α, β] → M (and α, β may vary); we say that (19)
(19) the following conditions are not independent: some of them imply some others; we do not detail

26
 • C connects points: ∀x, y ∈ M there is at least one γ ∈ C, γ : [α, β] → M such
that γ(α) = x, γ(β) = y
def
• C is reversible: if γ ∈ C and γ̂(t)=γ(−t) then γ̂ ∈ C
• len is independent of reparametrization: if γ ∈ C then, ∀h > 0, k ∈ lR, its linear
reparametrization γ 0 (s) = γ(hs + k) is in C, and lenγ = lenγ 0 ,
• len is monotonic: if γ ∈ C then its restriction γ 0 = γ|[c,d] is in C, and lenγ 0 ≤
lenγ
• len is additive: if γ 0 , γ 00 ∈ C, γ 0 : [α0 , β 0 ] → M , γ : [α00 , β 00 ] → M , and
β 0 = α00 , then we may join the two curves and obtain γ: then γ is in C, and

lenγ = lenγ 0 + lenγ 00

• len is run-continuous: that is, the running length t 7→ len(γ|[α,t] ) is continuous

for all γ ∈ C
• lenγ = 0 when γ is constant (and constants are in C)
We then define b(x, y) on M to be the infimum of this length lenγ in the class of all
ξ ∈ C, ξ : [0, 1] → M with given extrema ξ(0) = x, ξ(1) = y.
Proposition 4.38 b is an asymmetric semidistance: b satisfies the triangle inequality,
since len is independent of reparametrization and additive.
b ≥ 0, and b(x, x) = 0 since len is zero on constant curves.
b may fail to be an asymmetric distance, since we cannot be sure that b(x, y) =
0 =⇒ x = y
We also induce lenb γ from b using the total variation, as in (2.5).

Lemma 4.39 Suppose len and C satisfy all the hypotheses above. If len is lower semi
continuous w.r.t. uniform convergence, then lenγ = lenb γ for any γ ∈ C

The proof for the asymmetric case is not different from the symmetric case (cf. 1.6 [12]).
We provide here a detailed proof, for convenience of the reader.
Proof. Fix γ ∈ C. We insert len into the definition (2.5) of lenb γ:
n
X
lenb γ = sup inf lenξi (4.40)
T ξi
i=1

where the inf is done on ξi : [0, 1] → M, ξi ∈ C, connecting γ(ti−1 ) to γ(ti ); equiv-

alently (since len is independent of reparametrization) we may choose ξi : [ti−1 , ti ] →
M ; so we can join all the ξi and write

lenb γ = sup inf lenξ (4.41)

T ξ∈ΞT

where the sup is done on all choices of T finite subset of [0, 1] and the inf is done
on the class ΞT of ξ ∈ C, ξ : [0, 1] → M , such that ξ(t) = γ(t) for t ∈ T .
Since we could choose ξ = γ, we immediately obtain that

lenb γ ≤ lenγ (4.42)

27
 Fix a T : by (4.41),
inf lenξ ≤ lenb γ
ξ∈ΞT

Let η > 0. Define ΞT,η as the class of ξ in ΞT such that lenξ ≤ lenb γ + η and, for
any interval [ti , ti+1 ] defined by T ,

len(ξ|[ti ,tt+1 ] ) ≤ len(γ|[ti ,tt+1 ] )

Then
inf lenξ = inf lenξ
ξ∈ΞT ξ∈ΞT ,η

We drop η for simplicity.

Since len is run-continuous, then ∀δ > 0, there is a ε > 0 such that, for any interval
(s, t) with |t − s| < ε,
len(γ|[s,t] ) < ε, len(γ̂|[−t,−s] ) < ε
def
where γ̂(t)=γ(−t); and then b(γ(t), γ(s)) < δ, b(γ(s), γ(t)) < δ, by our definition of
b. We define the density |T | as
def
|T |= sup(ti+1 − ti )
i

when T = {t1 < t2 < . . . tn }.

Consider any T with density |T | ≤ ε; for any point s ∈ [0, 1], there ∃t, t0 ∈ T , with
|t − t| ≤ ε and t ≤ s ≤ t0 :
0

b(ξ(s), γ(s)) ≤ b(ξ(s), ξ(t)) + b(γ(t), γ(s)) ≤ len(ξ|[t,s] ) + δ ≤

≤ len(ξ|[t,t0 ] ) + δ ≤ len(γ|[t,t0 ] ) + δ ≤ 2δ
and similarly b(γ(s), ξ(s)) ≤ 2δ
We can order the sets T by inclusion, so that the class of such sets is a direct class;
moreover T ⊂ T 0 implies
inf lenξ ≤ inf lenξ
ξ∈ΞT ξ∈ΞT 0

so we can write
sup inf lenξ = lim inf lenξ
T ξ∈ΞT T →∞ ξ∈ΞT

As T → ∞, we have |T | → 0, and then ξ → γ uniformly; since lenγ is l.s.c by

hypothesis, then
lenb γ = lim inf lenξ ≥ lenγ
T ΞT
b
and then lenγ = len γ.
Corollary 4.43 Suppose we are given an asymmetric space (M, b); let C be the class
of rectifiable curves with continuous run-length, and len=lenb : this isg a length structure
def

satisfying all above hypotheses. len induces b which induces lenb γ. lenb γ is lower
b g

semi continuous since it is the supremum of the continuous functionals

n
X

b γ(ti−1 ), γ(ti )
i=1
By the lemma
g
lenb γ = lenb γ

28
 This immediately entails the result:

Corollary 4.44 The operation b 7→ bg is idempotent, that is, bg is path-metric, idem est
bg = (bg )g .

5 Applications
§5.i Asymmetric metric manifold
Suppose now that M is a differential manifold (20)
, with an atlas A and a topology τ M .

Definition 5.1 (A-Lip) We will say that a curve ξ : [α, β] → M is locally Lipschitz
w.r.t. A, if for any local chart ϕ : U → lRn in the atlas A, ϕ ◦ ξ is locally Lipschitz for
t ∈ ξ −1 (U ).

Hypotheses 5.2 Suppose that we are given a positive function F : T M → lR+ ;

suppose
• F is lower-semi-continuous,
• v 7→ F (x, v) is positively 1-homogeneous, that is,

F (x, λv) = F (x, v)λ ∀λ ≥ 0

• F (x, v) = 0 iff v = 0.

We define the length lenL γ of a locally Lipschitz curve ξ : [0, 1] → M as

Z 1
L
len γ = ˙
F (ξ(s), ξ(s)) ds (5.3)
0

Remark 5.4 (The regular Finsler case) If we would suppose that

• F is continuous, and C ∞ on the slit tangent bundle T M \ 0
• v 7→ F 2 (x, v) is strongly convex (21)
for v 6= 0, ∀x,
then this section would be exactly what is found in section 6.2 [3]; we will instead use
less regular assumptions.

As in §4.iii.2, we then define the asymmetric distance b(x, y) on M to be the

infimum of this length lenL γ in the class of all locally Lipschitz ξ with given extrema
ξ(0) = x, ξ(1) = y.
In the above hypotheses 5.2, b is an asymmetric distance; indeed, b is a semidistance
by 4.38, and the relation b(x, y) =⇒ x = y can be proved by the methods in the
following Lemma. This metric is also called a Finsler metric, and is naturally associated
to Hamilton-Jacobi equations; see [18] and references therein.
There may be a confusion here, since we have two topologies at hand: the topology
τ M of the differential manifold M , and the topology τ b induced by b; but we may
use this lemma, that simply restates the lemma 6.2.1 in [3] (with fewer regularity
assumptions on F )
(20) more precisely, let M be a finite dimensional connected smooth differential manifold, without boundary
(21) that is, setting f (v) = F 2 (x, v) then in local coordinates the Hessian of f is positive definite for v 6= 0

29
Lemma 5.5 Assume 5.2 and let F be locally bounded from above. Then for any
z ∈ M there exists a compact neighborhood U of z, associated with local coordinates
ϕ : U → W ⊂ lRn , such that the push forward of b in local coordinates is bounded
from above and below by the Euclidean norm: that is, ∃c0 > c > 0

c0 |x̂ − ŷ| ≥ b(x, y) ≥ c|x̂ − ŷ| (5.5.?)

where ŷ = ϕ(y) and x̂ = ϕ(x).

Proof. Let z ∈ M and Z be a neighborhood of z, with Z compact w.r.t. τ A , associated

to local coordinates ϕ : Z → W ⊂ lRn , with W convex bounded.
Let F̂ be the push-forward of F ; for any x, y ∈ Z, we may join ŷ = ϕ(y) to
x̂ = ϕ(x) with a segment, and then b(x, y) ≤ c0 |ϕ(x) − ϕ(y)| where c0 is the upper
bound of F̂ (w, v) for w ∈ W and |v| ≤ 1.
In the above setting, let U ⊂⊂ Z, let c be the minimum of F̂ (w, v) for w ∈ ϕ(Z)
and |v| = 1 (this minimum exists and is positive, since F is l.s.c); and let ε be the
minimum of |ϕ(x) − ϕ(z)| for x ∈ ∂U, z ∈ ∂Z; let ξˆ=ϕ ◦ ξ when ξ ∈ Z.
def

Then, for any ξ connecting x to y,

Z 1 Z t
˙ ˙ˆ
lenL ξ = F (ξ(s), ξ(s)) ds ≥ c|ξ(s)| ds ≥ c(ε ∧ |x̂ − ŷ|)
0 0

where t = 1 if ξ never exits from Z, otherwise it is the first time s such that
ξ(s) ∈ ∂Z: then b(x, y) ≥ c(ε ∧ |x̂ − ŷ|), and we choose a smaller U .

Corollary 5.6 Under the same hypotheses, τ b = τ M , i.e. b generates the same topology
that the atlas of M induces. Moreover both coincide with the topology generated by the
forward balls; or respectively, by the backward balls. (22)
This topology is locally compact: then
(23)
• xn → x iff b(xn , x) → 0 iff b(x, xn ) → 0 ; and
• (M, b) locally admits geodesics (by 4.24).
By (5.5.?), a curve ξ : [0, 1] → M is locally Lipschitz w.r.t A, as defined in 5.1, iff it
is locally Lipschitz w.r.t b, as defined in §2.ii.

Proposition 5.7 i). (M, b) is path-metric, that is, b = bg

ii). suppose 5.2, and v 7→ F (x, v) is convex; then for any Lipschitz γ, lenL γ coincides
with the lenb γ defined in (2.5).

Proof. i) is a consequence of 4.4. (24) By prop. 4.4, we need to check that ∀x, y ∈
M, ∀ε > 0 ∃z ∈ M such that b(x, z) < b(x, y)/2 + ε and b(z, y) < b(x, y)/2 + ε. Fix
x, y. Let ε > 0 and ξ connecting x to y such that lenL ξ < b(x, y) + 2ε; let
Z t
f (t) = ˙
F (ξ(s), ξ(s)) ds
0
(22) cf. 3.8 , or sec. 6.2C in [3] for the regular case
(23) cf. 6.2.5 in [3] for the regular case
(24) If v 7→ F (x, v) is convex, then this is an immediate consequence of the second point. Alternatively, we

have a proof that does not need that L(x, ·) is convex: indeed the operation of computing b implicitly replaces
F by its relaxed l.s.c.

30
then f is continuous, so let t0 be the time when f (t0 ) = f (1)/2, and z = ξ(t0 ); by
reparametrizing, ξ 0 (s) = ξ(st0 ), ξ 00 (s) = ξ(1 − (1 − s)(1 − t0 )); then since b(x, z) <
lenL ξ 0 = f (1)/2 < b(x, y)/2 + ε, and b(z, y) < lenL ξ 00 = f (1)/2 < b(y, z)/2 + ε.
ii). The proof is similar to the proof of lemma 4.39. We set C to be the class of
Lipschitz curves, and len=lenL . Fix γ, let L=lenL γ. If L is small, then the curve is
def def

contained in local coordinates. Indeed, choose, as per 5.5, U a compact neighborhood

of γ(0), associated to local coordinates ϕ : U → W ⊂ lRn . We suppose for the moment
that L < R, with B + (γ(0), R) ⊂ U : then γ(t) ∈ U ∀t ∈ [0, 1]. We fix η = (R − L)/2
in the proof of the aforementioned lemma 4.39: then any curve ξ ∈ ΞT,η is wholly
contained in U : indeed by (4.42), lenb ξ ≤ lenL ξ ≤ lenb γ + ε ≤ L + ε < R. As
T → ∞, ξ → γ uniformly.
By (5.5.?), the uniform convergence happens also in local coordinates. By well-
known theory (cf. thm. 2.3.3 in [7]), lenL γ is l.s.c, and we conclude as in the Lemma 5.5.
If we don’t suppose that γ(t) ∈ U ∀t ∈ [0, 1]: we fix a very dense partition S T̂ , so
that we can associate to any point ti ∈ T̂ a local chart ϕi : Ui → lRn , so that i Ui
covers the image of γ, and more precisely γ([ti , ti+1 ]) ⊂ Ui . We repeat the above
argument in any Ui .

Lemma 5.8 (Reparametrization) For any γ : [α, β] → M Lipschitz, define l :

[α, β] → [0, L]
Z t
l(t) = F (γ(s), γ̇(s)) ds
a
(where L = lenγ).
Then there is a unique ξ such that γ = ξ ◦ l, and
˙
F (ξ(s), ξ(s)) =1 ˜ ∈ [0, L]
∀s (5.8.?)
Rt
Proof. We apply 4.19, and just note that `ξ (t) = t means that ˙ =t
F (ξ, ξ)
0

We can then state this version of the Hopf-Rinow theorem

Theorem 5.9 Assume that the function F : T M → lR+
• is locally bounded from above
• v 7→ F (x, v) is positively 1-homogeneous, and convex,
and 5.2 holds.
Then the asymmetric metric space (M, b) is path-metric, locally compact, and
locally admits geodesics (that we can always reparametrize so that they satisfy (5.8.?)).
If (M, b) satisfies any (=all) of the conditions i,ii,iii in the Hopf–Rinow theorem 4.33,
then for any x, y ∈ M , there is a locally Lipschitz curve ξ connecting them that
minimizes (5.3), satisfying (5.8.?).

Further applications of these ideas are in [16]; in particular, in the appendix of [16],
2
we consider a Lagrangian L : T M → lR+ (that is defined as L(x, v) = F (x, v) ) and
∗ +
its Legendre-Fenchel dual Hamiltonian H : T M → lR ; we show that, if H ∈ C 1,1
and p 7→ H(x, p) is positively 2-homogeneous and strictly convex, we can define an
exponential map and add it to the statement of the Hopf-Rinow theorem, to obtain a
statement exactly as in ch. VI of [3].

31
Acknowledgements
The author thanks Prof. S. Mitter, who has reviewed and corrected various versions of
this paper, and suggested many improvements; and Prof. L. Ambrosio for the discus-
sions.

Contents
1 Introduction 1
§1.i The Hopf–Rinow theorem . . . . . . . . . . . . . . . . . . . . . . . 2
§1.ii asymmetric metric theories in Calculus of Variations . . . . . . . . 3

2 Asymmetric metric spaces 4

§2.i Asymmetric metric space . . . . . . . . . . . . . . . . . . . . . . . 4
§2.ii Lipschitz maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
§2.iii Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
§2.iv Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Comparison with related works 8

§3.i Comparison with Busemann’s “General metric spaces” . . . . . . . 8
§3.ii Comparison with quasi metric spaces . . . . . . . . . . . . . . . . . 10
§3.ii.1 Topology . . . . . . . . . . . . . . . . . . . . . . . . . . 10
§3.ii.2 Cauchy sequences, and completeness . . . . . . . . . . . . 10
§3.iii Comparison with symmetric case . . . . . . . . . . . . . . . . . . . . 11

4 Study of asymmetric metric spaces 12

§4.i Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
§4.i.1 Mid-point properties . . . . . . . . . . . . . . . . . . . . 13
§4.i.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 15
§4.ii Geodesics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
§4.ii.1 Existence of geodesics . . . . . . . . . . . . . . . . . . . . 21
§4.ii.2 Hopf-Rinow Theorem . . . . . . . . . . . . . . . . . . . . 22
§4.ii.3 More examples . . . . . . . . . . . . . . . . . . . . . . . 25
§4.iii Complements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
§4.iii.1 Generating (path) metrics . . . . . . . . . . . . . . . . . . 26
§4.iii.2 Length structure . . . . . . . . . . . . . . . . . . . . . . . 26

5 Applications 29
§5.i Asymmetric metric manifold . . . . . . . . . . . . . . . . . . . . . 29

References
[1] L. Ambrosio and P. Tilli. Selected topics in ”analysis in metric spaces”. appunti.
edizioni Scuola Normale Superiore, Pisa, 2000.

[2] Luigi Ambrosio, Nicola Gigli, and Giuseppe Savarè. Gradient flows in metric
spaces and in the Wasserstein space of probability measures. Birkhüser, 2004.

[3] D. Bao, S. S. Chern, and Z. Shen. An introduction to Riemann-Finsler geometry.

(October 1, 1999 version).

32
 [4] H. Busemann. Local metric geometry. Trans. Amer. Math. Soc., 56:200–274,
1944.

[5] H. Busemann. The geometry of geodesics, volume 6 of Pure and applied mathe-
matics. Academic Press (New York), 1955.

[6] H. Busemann. Recent syntetic differential geometry, volume 54 of Ergebnisse der

Mathematik und ihrer Grenzgebiete. Springer Verlag, 1970.

[7] G. Buttazzo. Semicontinuity, relaxation and integral representation in the calculus

of variation. Number 207 in scientific & technical. Longman, 1989.

[8] A. Duci and A. Mennucci. Banach-like metrics and metrics of compact sets. in
preparation, 2007.

[9] A. Fathi. Weak KAM theorem in Lagrangian dynamics. (Preliminary version, 2

Feb 2001).

[10] W. H. Fleming and H. M. Soner. Controlled Markov processes and viscosity

solutions. Springer–Verlag, 1993.

[11] P. Fletcher and W.F.Lindgren. Quasi-uniform spaces, volume 77 of Lecture notes

in pure and applied mathematics. Marcel Dekker, 1982.

[12] M. Gromov. Metric Structures for Riemannian and Non-Riemannian Spaces.

Birkhäuser, 1999.

[13] M.K.Vamanamurthy I.L.Reilly, P.V.Subrahmanyam. Cauchy sequences in quasi-

pseudo-metric spaces. Monat.Math., 93:127–140, 1982.

[14] J. C. Kelly. Bitopological spaces. Proc. London Math. Soc., 13(3):71–89, 1963.

[15] H. P. Künzi. complete quasi-pseudo-metric spaces. Acta Math. Hung., 59(1-

2):121–146, 1992.

[16] A. C. G. Mennucci. Regularity and variationality of solutions to Hamilton-Jacobi

equations. part ii: variationality, existence, uniqueness. 1st (draft) version, on-
line since 2005 at http://cvgmt.sns.it/papers/men05/, 2nd version
ibidem, 2006.

[17] B. B. Phadke. Nonsymmetric weakly complete g-spaces. Fundamenta Mathemati-

cae, 1974.

[18] Antonio Siconolfi. Metric character of Hamilton-Jacobi equations. Transactions

of the american mathematical society, 355(5):1987–2009, 2003. electronically
published on January 8, 2003.

[19] E. M. Zaustinsky. Spaces with non-symmetric distances. Number 34 in Mem.

Amer. Math. Soc. AMS, 1959.

33