See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/267239341

When Does a Metric Generate Convex Balls?

Article

CITATIONS READS

5 574

1 author:

Timothy Norfolk
University of Akron
20 PUBLICATIONS 204 CITATIONS

SEE PROFILE

All content following this page was uploaded by Timothy Norfolk on 09 April 2015.

The user has requested enhancement of the downloaded file.

 When Does a Metric Generate Convex Balls?
Timothy S. Norfolk
Department of Mathematics and Computer Science,
The University of Akron,
Akron, Ohio 44325.

Abstract
Given a metric imposed on a real linear vector space, we explore the
issue of convexity of the open balls generated by the metric.

1991 Mathematics Subject Classification : Primary 52A07

1 Introduction
The concept of a metric is usually first encountered in beginning courses in
real analysis, presented as an abstraction to more general sets of the concept
of distance in Euclidean space.
This abstraction allows for the discussion of such issues as convergence and
continuity, and extends to the even more general ideas of point-set topology.
When one deals with real linear spaces (vector spaces over the reals), it is
usual to discuss the concept of norm, which abstracts the notion of length,
rather than distance.
However, there are cases in which real linear spaces may be endowed with a
metric, rather than a norm. In such cases, it seems reasonable to ask how
the algebraic structure of the linear space and metric combine to generate a
geometry on the space.
One natural question which arises is the following :
What metric spaces have the property that all open balls in the space are
convex?

1
To my chagrin, this question arose because of an error that I made in the
final examination of a graduate–level analysis course, which I did not notice
until I was constructing the answer key.
Our starting point is the question which I originally posed, and an example
which demonstrates that not all metrics have the property that we seek.
We follow that with a complete characterization of the metrics with this
property, and some special cases, which provide weaker sufficient conditions
that are easier to verify.
We continue by discussing some other convexity notions which appear in the
literature. Namely, the topological concept of local convexity (Willard [2], p.
164), and the functional concept of a (metrically) convex space (Bryant [1]),
and show that both of these notions are independent of the problem under
consideration.
Finally, we discuss two particular constructions. The first shows that there
are metrics in which some of the open balls are convex, and some are not. The
second construction is, in some sense, an inverse problem. Given any convex
set which is symmetric with respect to the origin, we construct a number of
metrics whose open balls are dilations, contractions or translations of this
set.

2 Convex Open Balls in Metric Spaces

As discussed above, the question addressed here appeared on an examination
that I gave in analysis, and led me to the subsequent investigation.
The question posed was to prove the following :

Proposition 2.1 If E is a linear space and ρ is a metric on E, then the

open ball B(x; r) = {y ∈ E : ρ(x, y) < r} is convex.

To see that this is indeed false in general, consider the following :

Proposition 2.2 Consider the linear space E = IR2 , with the metric
q q
ρ(x, y) = ρ((x1 , x2 ), (y1 , y2 )) = |x1 − y1 | + |x2 − y2 | .

Then, B(x; r) is not convex for any x ∈ E, r > 0.

2
 q q
Figure 1: The unit ball B(0; 1) = {y ∈ IR2 : |y1 | + |y2 | < 1}

Proof. Clearly,
√ √ ρ√is a metric on E, which can be proven using the fact that
a + b ≤ a + b, for a, b ≥ 0.
r2
Given any x ∈ E and r > 0, choose < α < r2 , and let y = (x1 + α, x2 ),
2
z = (x1 , x2 + α).
√
Then, ρ(x, y) = α = ρ(x, z) < r, so that y, z ∈ B(x; r).

1 1
 √ 1 1
However, ρ x, y + z = 2α > r, so that y + z ∈ / B(x; r). Q.E.D.
2 2 2 2
In Figure 1, we present the unit ball B(0; 1) in this metric, which is clearly
not a convex set.
Having established that I did indeed make a mistake in this question, it
is natural to explore necessary and sufficient conditions for the validity of
Proposition 2.1, and related problems.
At this point, it is appropriate to note again that, while the questions which
arise are quite natural, and the results obtained in this paper are elementary,
an exhausting search of the literature has failed to find this specific issue
addressed.

3
3 Necessary and Sufficient Conditions
Given that our previous example shows that not all metrics generate convex
open balls, we are in a position to completely characterize those metrics
which do so.

Theorem 3.1 Proposition 2.1 holds if and only if

ρ(x, ty + (1 − t)z) ≤ max{ρ(x, y), ρ(x, z)} ,
for all x, y, z ∈ E and 0 ≤ t ≤ 1.

Proof. On one hand, suppose that the indicated condition holds.

That is, given any x ∈ E, with y, z ∈ B(x; r) and 0 ≤ t ≤ 1, we have
ρ(x, ty + (1 − t)z) ≤ max{ρ(x, y), ρ(x, z)} < r ,
so that ty + (1 − t)z ∈ B(x; r), the desired convexity condition.
On the other hand, if there exist x, y, z ∈ E and 0 ≤ t ≤ 1 such that
ρ(x, ty + (1 − t)z) > max{ρ(x, y), ρ(x, z)} ,
we may choose
max{ρ(x, y), ρ(x, z)} < r < ρ(x, ty + (1 − t)z) ,
from which y, z ∈ B(x; r), but ty + (1 − t)z ∈
/ B(x; r), meaning that the
latter set is not convex. Q.E.D.
Note that, if ρ satisfies this condition, x ∈ E, and P is any convex polytope
in E, then the function f : P → [0, ∞) defined by
f (y) = ρ(x, y)
attains its maximum value at one of the vertices of P .
However, this optimization condition does not categorize the metrics that
we seek, since it can be satisfied by functions which are not metrics, such as
f (x, y) = (x1 − y1 )2 + (x2 − y2 )2 on IR2 .
While Theorem 3.1 gives a complete description of the metrics which generate
convex balls, the functional condition given is, in practice, difficult to verify.
Consequently, it is appropriate at this point to consider some special cases.

4
Corollary 3.2 If any of the following conditions holds on a metric ρ on a
linear space E, then Proposition 2.1 holds. That is, the open balls of the
metric are convex.

1. ρ is the discrete metric. That is, ρ(x, y) = 1 if x 6= y.

2. For each fixed x ∈ E, ρ is a convex function on E.

That is, for all y, z ∈ E and 0 ≤ t ≤ 1 :

ρ(x, ty + (1 − t)z) ≤ tρ(x, y) + (1 − t)ρ(x, z)

3. ρ satisfies the following conditions :

(a) ρ(x + y, x + z) = ρ(y, z) for all x, y, z ∈ E.

(ρ is translation invariant).
and
(b) ρ(tx, ty) ≤ tρ(x, y) for all x, y ∈ E and 0 ≤ t ≤ 1.

4. There exists a norm ||.|| on E such that ρ(x, y) = ||x−y|| for x, y ∈ E.

Proof.

1. Suppose that ρ is the discrete metric. Then,

B(x; r) = {x} for any x ∈ E and 0 ≤ r ≤ 1
and
B(x; r) = E for any x ∈ E and 1 < r.
In either case, B(x; r) is clearly convex.

2. Suppose that the indicated condition holds.

Then, for any x, y, z ∈ E and 0 ≤ t ≤ 1, we have

ρ(x, ty + (1 − t)z) ≤ tρ(x, y) + (1 − t)ρ(x, z) ≤ max{ρ(x, y), ρ(x, z)} ,

which, by Theorem 3.1 means that all open balls are convex.

5
 3. Suppose that the given pair of conditions holds.
We will show that this implies the conditions of 2.
For any x, y, z ∈ E and 0 ≤ t ≤ 1, we have

ρ(x, ty + (1 − t)z) = ρ(0, t(y − x) + (1 − t)(z − x))

≤ ρ(0, t(y − x)) + ρ(t(y − x), t(y − x) + (1 − t)(z − x))
≤ tρ(0, y − x) + ρ(0, (1 − t)(z − x))
≤ tρ(x, y) + (1 − t)ρ(0, z − x)
= tρ(x, y) + (1 − t)ρ(x, z) ,

the desired result.

4. Suppose that ρ is derived from a norm. Then, it clearly satisfies the

conditions of 3, since it is translation invariant, and also satisfies the
stronger condition ρ(tx, ty) = tρ(x, y) for all x, y ∈ E and 0 ≤ t.
Q.E.D.

4 Other Convexity Concepts

While there are at least two other concepts involving convexity which appear
in the literature, we show in this section that these concepts are independent
of the central question of this work.
While there is a class of metric spaces called (metrically) convex spaces, the
open balls in such spaces are not necessarily convex, since the definition
(given below) is clearly independent of the geometry of the space.

Definition 4.1 (cf. Bryant [1]) A metric space (X, ρ) is (metrically) convex
if and only if, given distinct points x, y ∈ X, there exists z ∈ X, different
from both x and y, such that

ρ(x, z) + ρ(z, y) = ρ(x, y) .

On one hand, by considering again the discrete metric, it is clear that a

metric on a linear space which generates convex open balls is not necessarily
convex.

6
On the other hand, the metric of Proposition 2.2 satisfies the Intermediate
Value Theorem, and so is a convex metric, but generates no open convex
balls.
We may thus deduce that the definition of convex metric is independent of
the convexity of the open balls of the metric.
One can also consider the topology that is induced by the metric, which leads
one to the concept of a locally convex space (Willard [2], p.164), in which each
point of the space has a base of convex sets.
This standard definition leads immediately to the following result :

Proposition 4.2 If ρ is a metric on the linear space E such that all open
balls are convex, then the induced linear topological space is locally convex.

On the other hand, if the metric ρ induces a locally convex topology on the
linear space E, it is not necessarily true that all open balls are convex. In
fact, it can be the case that no open balls in a locally convex metric space
are themselves convex.
To see this, consider the following :

Example 4.3 Let E = IR2 , and ρ be the metric of Proposition 2.2. Then,
none of the open balls is convex, but the metric space is locally convex.

In Proposition 2.2, it was shown that no open balls are convex in this case.
However, it is clear that each open ball contains the interior of a disk (a
convex set) and that each such disk can be expressed as a union of non-
convex open balls. Hence, each such disk is an open set, and forms a base
for the metric topology, which means that the topology is locally convex.
Just as in the above, this demonstrates that local convexity is independent
of the convexity of the open balls of the metric.

5 Two Constructions
To conclude, we provide two constructions. The first demonstrates that there
exist metrics on linear spaces which have some open balls which are convex,
and some which are not. The second shows that, given a convex set in a real

7
linear space which is symmetric with respect to the origin, it is possible to
construct a metric whose open balls are dilations, contractions or translations
of this convex set.
We recall the following result, which appears in many introductory texts,
frequently as an exercise :

Theorem 5.1 If ρ1 and ρ2 are metrics on a set E, then the following is also
a metric on E :

ρ0 (x, y) = max{ρ1 (x, y), ρ2 (x, y)} .

Using this result, it is possible to construct metric spaces in which some of

the open balls are convex, and others are not. To illustrate, consider :

Example 5.2 Let E = IR2 , ρ1 be the metric of Proposition 2.2, and ρ2 be

the Euclidean qmetric. Thatqis, q
ρ1 (x, y) = |x1 − y1 |+ |x2 − y2 | and ρ2 (x, y) = (x1 − y1 )2 + (x2 − y2 )2 .
Then, using the metric ρ0 =√max{ρ1 , ρ2 }, the corresponding
√ open ball B 0 (x; r)
is not convex for 0 < r < 8, and is convex if 8 ≤ r.

To see this, it suffices to make two observations :

If B1 (x; r), B2 (x; r) and B 0 (x; r) denote the open balls centred at x with
radius r in the metrics ρ1 , ρ2 and ρ0 respectively, then

B 0 (x; r) = B1 (x; r) ∩ B2 (x; r) (5.1)

and, via some simple analysis,

B1 (x; r/81/4 ) ⊂ B2 (x; r2 /81/2 ) ⊂ B1 (x; r) (5.2)

for any 0 < r. Moreover, the inclusions in (5.2) are sharp.

√
Now, if 8 ≤ r, then, by (5.1) and (5.2), we have B 0 (x; r) = B2 (x; r), which
is clearly convex, since it is the interior of a disk. On the other hand, if
0 < r ≤ 1, then B 0 (x; r) = B1 (x;
√ r), 0which is not convex, as demonstrated in
Proposition 2.2. For 1 < r < 8, B (x; r) is a region whose boundary has 8
arcs, 4 of which are portions of the boundary of B1 (x, r), and hence cannot
be convex, by our previous observations.

8
 Figure 2: The open ball B 0 (0; 5/2)

To clarify, Figure 2 shows the ball B 0 (0; 5/2).

As claimed above, we conclude by constructing a metric from a particular
convex set.
In a general space, quantities such as boundary, boundedness, closed set and
the interior of a set require that a metric be imposed a priori. However, in
the case of a linear space, we may implicitly define these quantities by means
of the underlying algebraic structure.
Let E be a real linear space, and C be a subset of E which satisfies the
following conditions :
i) C is convex.
ii) For each x ∈ E, x 6= 0, there exists a unique T = T (x) > 0 such that
L(x) := {t ∈ IR : tx ∈ C} = [−T, T ] .

That is, each line through the origin intersects C in a line segment which is
symmetric about 0. Examples of such sets in IRn include closed spheres and
closed polytopes symmetric with respect to 0.
Define the set B in E via
B = {T (x)x : x 6= 0} .

9
B is the boundary of C, in the topology that we generate, and the set C −B =
{x ∈ C : x ∈ / B} its interior.
Note that, for any t ≥ 0, the sets tC = {tx : x ∈ C} and t(C − B) are also
convex sets.
Furthermore, let f : (0, ∞) → (0, ∞) be a function which satisfies the
following conditions :
i) f (t) is non-decreasing.
ii) For all t1 , t2 ≥ 0, f (t1 + t2 ) ≤ f (t1 ) + f (t2 ).
Examples of such functions include f (t) = tp for 0 ≤ p ≤ 1, and piecewise-
defined functions such as
(
1
3
if 0 < t < 1
f (t) =
1 if 1 ≤ t

Using the properties of the set B, and the functions T and f , we may define
a metric ρ on E via
(
0 if x = y
ρ(x, y) =
f (1/T (x − y)) if x 6= y

To verify that this is a metric, we check the conditions of the definition :

1. ρ(x, y) ≥ 0 for all x, y ∈ E , ρ(x, y) = 0 if and only if x = y.

This follows from the definition of ρ, using the facts that T and f are
strictly positive functions.

2. ρ(x, y) = ρ(y, x) for all x, y ∈ E.

This follows from the symmetry condition on C, which implies that
T (−x) = T (x).

3. ρ(x, z) ≤ ρ(x, y) + ρ(y, z) for all x, y, z ∈ E.

This condition clearly holds if any two of x, y and z are equal. If not,
we set a = 1/T (x − y) and b = 1/T (y − z), from which ρ(x, y) = f (a),

10
 ρ(y, z) = f (b), and there exist u, v ∈ B such that x − y = au and
y − z = bv. Hence,
x−z = (x − y) + (y − z)
= au
 + bv 
a b
= (a + b) a+b u + a+b v

a b
By the convexity of C, it follows that the vector w = u+ v∈
a+b a+b
C, whence T (w) ≥ 1, and so 1/T (w) ≤ 1.
Thus, x − z = (a + b)w ∈ (a + b)C, from which

ρ(x, z) ≤ f (a + b) ≤ f (a) + f (b) = ρ(x, y) + ρ(y, z) ,

the desired inequality.

To illustrate, consider :

Example 5.3 Let E be the real linear space IR2 , C be the closed hexagon √
with vertices {(1, 0), (0, 1), (−1, 1), (−1, 0), (0, −1), (1, −1)} and f (t) = t.
Then, the set B is the boundary of this polygon. Given the point √ x = (2, 0.5),
we have 0.4x ∈ B, so that T (x) = 0.4 and hence ρ(0, x) = 2.5. This is
illustrated in figure 3.

Having thus defined the metric ρ, we may make the following observations :
i) By the definition, ρ is translation invariant.
ii) With the choice f (t) = t, ρ generates a norm on E.
iii) For any r > 0 and x ∈ E, the open ball B(x, f (r)) is precisely the set
r(C − B), which, as noted above, is convex.

References
[1] V. W. Bryant. The convexity of the subset space of a metric space.
Compositio Math., 22 : 383–385, 1970.

[2] S. Willard. General Topology. Addison-Wesley Publishing, 1970.

11
Figure 3: The convex polygon and point (2, 0.5)
View publication stats

12