Fixed Point Theory, 11(2010), No.

1, 45-58
http://www.math.ubbcluj.ro/∼ nodeacj/sfptcj.html

METRIC FIXED POINT THEORY: OLD PROBLEMS

AND NEW DIRECTIONS

W.A. KIRK

Dedicated to Wataru Takahashi on the occasion of his retirement

Department of Mathmatics
University of Iowa
Iowa City, IA 52242 USA
E-mail: [email protected]

Abstract. This is a brief review of some of the things, both past and present, which have motivated
the writer’s interest in metric fixed point theory.
Key Words and Phrases: Nonexpansive mappings, common fixed points, R-trees, product spaces
2000 Mathematics Subject Classification: 47H09, 47H10, 54H25, 05C05.

1. Introduction
Mathematical research often focuses on problems lying on the boundary of an area
of known results, with a view to enlarging that area. This strategy is designed to
assure some success. However there are two other strategies that can be even more
rewarding when successful. One is the study of old and seemingly difficult problems.
Another is the search for genuinely new research directions. In this paper we discuss
how both approaches have have influenced the writer’s career.
This paper is largely expository, and much of the material in Section 3 has already
appeared in [28].

2. Old Problems
In this section we discuss three papers that had a strong impact in stimulating the
writer’s early interest in fixed point theory. The first is a paper by Ralph DeMarr
[14] about commuting continuous mappings on the unit interval, the second is the
fundamental paper of Brodskii and Milman [9] that gave us the concept of ‘normal
structure’, and the third is another paper by Ralph DeMarr [13] about commuting
families of nonexpansive mappings. These papers triggered a huge amount of further
research in the past and to this day leave many related questions open.

The paper was presented at The 9th International Conference on Fixed Point Theory and Its
Applications, July 16-22, 2009, National Changhua University of Education, Changhua, Taiwan
(R.O.C.).

45
46 W.A. KIRK

2.1. Commuting maps. In 1954 Eldon Dyer posed the following question: If two
continuous mappings f, g : [0, 1] → [0, 1] commute, do they necessarily have a common
fixed point? A. L. Shields posed the same question in 1955, as did Lester Dubins
in 1956 and the problem first appeared in the literature as part of a more general
question posed by J. R. Isbell in 1957 [23]. This question immediately attracted a
lot of attention, in part because it had long been known that the answer is yes for
polynomial functions. In 1963 Ralph DeMarr [14] gave a partial positive answer,
and another partial answer to the question was established by Schwartz in 1965, who
showed in [40] that if f has a continuous derivative, then there is a common fixed
point of f and some iterate of g. However the general question resisted a complete
solution for over ten years. In 1967 both W. M. Boyce [7] and H. Huneke [22] gave
counterexamples to the problem. Thus the answer to the original topological question
is negative, but it gave rise to an interesting positive result. (For a more thorough
survey of this topic, see Chapter 7 of [39].)
We now state DeMarr’s result. On the surface DeMarr’s result appears to be a
metric theorem. However an inspection of the proof reveals that it has a strong
order-theoretic component as well. For the convenience of the reader we include the
proof.
Theorem 1 (DeMarr [14]). Let f and g be two commuting continuous mappings of
a closed interval I into itself having respective Lipschitz constants α and β satisfying
β (α − 1) < (α + 1) . Then f and g have a common fixed point.
Proof ([14]). Let N be the set of fixed points of g. Then N is closed and nonempty,
and by commutativity, f : N → N. Assume f does not have a fixed point in N. Let
a = min N and b = max N. Then a < f (a) and f (b) < b. Now pick x0 , x1 ∈ N such
that x0 < f (x0 ) , f (x1 ) < x1 , x0 < x1 and x1 − x0 is minimal. This is possible
because N is compact and f is continuous.
Set y0 = f (x0 ) and y1 = f (x1 ) . We will show that y0 ≥ x1 and y1 ≤ x0 . Suppose
z ∈ N satisfies x0 < z < x1 . Then f (z) < z or f (z) > z, Suppose f (z) < z. Then
x0 < z, f (z) < z, and x0 < z. However z − x0 < x1 − x0 , contradicting minimality
of x1 − x0 . A similar contradiction arises if f (z) > z. Hence no point of N lies in
the open interval (x0 , x1 ) . This in turn implies that g (x) < x for all x ∈ (x0 , x1 )
or x < g (x) for all x ∈ (x0 , x1 ) . Without loss of generality assume x < g (x) for all
x ∈ (x0 , x1 ) .
Since x0 < f (x0 ) and f (x1 ) < x1 , it must be the case that f has a fixed point
in (x0 , x1 ) . Let s be the largest fixed point of f in (x0 , x1 ) , and let t = g (s) . Then
f (t) = t because g maps the fixed point set of f into itself, and since s < t, it must
be the case that x1 < t.
We now assume α > 1. Since f (x1 ) ≤ x0 we now have
t − x0 ≤ f (t) − f (x1 ) ≤ α (t − x1 ) . (1)
Also,
t − x0 = g (s) − g (x0 ) ≤ β (s − x0 ) ; (2)
s − x0 ≤ f (s) − f (x1 ) ≤ α (s − x1 ) . (3)
 METRIC FIXED POINT THEORY 47

From (1)
αx1 − x0 ≤ (α − 1) t. (4)
From (2) and (4):
αx1 − x0 ≤ (α − 1) [β (s − x0 ) + x0 ] ;
hence αx1 ≤ (α − 1) β (s − x0 ) + αx0 , from which
α (x1 − x0 ) ≤ β (α − 1) (s − x0 ) . (5)
From (3) s − x0 ≤ α [(x1 − x0 ) − (s − x0 )] ; hence
(α + 1) (s − x0 ) ≤ α (x1 − x0 ) . (6)
From (5) and (6)
α + 1 ≤ β (α − 1)
and this contradicts our initial assumption
β (α − 1) < (α + 1) .

The preceding proof assumed α > 1. If α < 1 then f is a contraction mapping
which has a unique fixed point. If g commutes with f then this point is also fixed
under g. In this case no further assumptions on g are needed. The remaining case is
α = 1. In this case f is nonexpansive and its fixed point set is a closed interval which
is invariant under g. Thus if g is merely continuous, then f and g have a common
fixed point.
There seems to be very little known about commuting maps on the unit interval
beyond DeMarr’s result, but two papers might be worth mentioning. In [33] it is
shown that commuting continuous mappings f, g of I → I have a common fixed point
if either of the following two conditions is satisfied: (1) the fixed points of f 2 coincide
with those of f ; (2) there exists an infinite sequence {In } of disjoint open intervals
each of which has as endpoints consecutive points in the set of fixed points of f, and
there is a fixed point c of g, such that for each In there is some integer m such that
f m (c) ∈ Im . Another result involving Lipschitz-type conditions is given in [24].
It is also interesting √
to note that the mappings in Huneke’s counterexample have
Lipschitz constant 3 + 6, so the restriction on the Lipschitz constants in DeMarr’s
result is significant.
The first question we raise in this paper is whether DeMarr’s Theorem is strictly a
result about a real line interval, or do analogs exist in a wider context. We consider
possible extensions below.
2.2. R-trees. In [6] it is shown that if X is a finite tree (in the topological sense), if
fα : X → X is a commuting family of open maps, and if g : X → X is a continuous
map which commutes with each fα , then they have a common fixed point. For
corresponding metric results we look to R-trees.
Definition 1. An R-tree is a metric space M such that for every x and y in M there
is a unique arc between x and y and this arc is isometric to an interval in R (i.e., is
a geodesic segment).
48 W.A. KIRK

Standard examples of R-trees include the ‘radial’ and ‘river’ metrics on R2 . For
the radial metric, consider all rays emanating from the origin in R2 . Define the radial
distance dr between x, y ∈ R2 as follows:
dr (x, y) = d (x, 0) + d (0, y) .
(Here d denotes the usual Euclidean distance and 0 denotes the origin.) For the river
metric ρ, if two points x, y are on the same vertical line, define ρ (x, y) = d (x, y) .
Otherwise define ρ (x, y) = |x2 |+|y2 | + |x1 − y1 | , where x = (x1 , x2 ) and y = (y1 , y2 ) .
More subtle examples of R-trees exist; e.g., the real tree of Dress and Terhalle [15].
The following result follows from a more general 1946 result of G. S. Young (see
[45], Theorem 16). Young explicitly points out in [46] that compactness is not needed.
There is a more constructive metric proof of this result is given in [27].
Theorem 2. A complete geodesically bounded R-tree has the fixed point property for
continuous maps.
Question 1. Let X be a compact or, more generally, a geodesically bounded and
complete R-tree, and let f, g : X → X be commuting and lipschitzian. Then if the
respective Lipschitz constants of f and g are sufficiently near 1, do f and g have a
common fixed point?
A very special case of Question 1 has an affirmative answer.
Theorem 3. Let X be a geodesically bounded complete R-tree, let g be a nonexpansive
mapping of X into itself, and suppose f : X → X is continuous and commutes with
g. Then f and g have a common fixed point.
Proof. By Theorem 2 g has a nonempty fixed point set N, and it is easy to see that N
is closed and convex. Since f commutes with g, f : N → N , and again by Theorem
2 f has a fixed point in N. 
Theorem 3 can be extended a little further. For this result it will be necessary to
discuss another concept.
A metric space X is said to be hyperconvex (Aronszajn and Panitchpakdi [1]) if
every family {B (yα ; rα )}α∈A of closed balls centered at yα ∈ X with radii rα ≥ 0 has
nonempty intersection whenever
d (yα , yβ ) ≤ rα + rβ ∀ α, β ∈ A.
It is known that compact hyperconvex spaces (often called Helly spaces) are con-
tractible and locally contractible; hence they have the fixed point property for contin-
uous mappings (see [36]). It is important to note that if a hyperconvex metric space
X is a subset of a metric space Y then there is always a nonexpansive retraction of
Y onto X. A discussion of this fact and many more facts about hyperconvex spaces
can be found in [16]. We also refer the reader to [5] for a survey that emphasizes
applications of R-trees.
For the remarks that follow we shall need the following fact.
Theorem 4 ([27]). For a metric space X the following are equivalent: (i ) X is a
complete R-tree; (ii ) X is hyperconvex and has unique metric segments.
 METRIC FIXED POINT THEORY 49

For a mapping g : X → X and each δ > 0, let

Fδ (g) = {x ∈ X : d (x, g (x)) ≤ δ} .
Theorem 5. Let X be a complete R-tree and let g be commuting nonexpansive map-
ping of X into itself, and suppose f : X → X is a continuous mapping which com-
mutes with g. If Fδ (g) is geodesically bounded for some δ > 0, then f and g have a
common fixed point.
Proof. By Theorem 4 X is hyperconvex, and it is known that approximate fixed point
sets of nonexpansive mappings in hyperconvex spaces are themselves hyperconvex
(Sine[42]). Thus Fδ (g) is hyperconvex. Also, again by Theorem 4, a hyperconvex
subspace of a complete R-tree is also a complete R-tree. Thus Fδ (g) is a geodesically
bounded complete R-tree. Since g : Fδ (g) → Fδ (g) , g has a nonempty fixed point
set N, which must be convex and geodesically bounded (since it lies in Fδ (g)). Since
f : N → N the conclusion follows from Theorem 2. 
It should be noted that the assumption that Fδ (g) is geodesically bounded is
strictly weaker than the assumption that X is geodesically bounded.
In [17] it is proved (Theorem 4.3) that if X is a geodesically bounded R-tree
then every commuting family of nonexpansive mappings of X → X has a nonempty
common fixed point set. The above approach yields the following (slight) extension
of that result.
Theorem 6. Let X be a complete R-tree and let F be a commuting family of non-
expansive mappings of X → X. If Fδ (g) is bounded for some g ∈ F, then F has a
nonempty common fixed point set.
Proof. Proceed as in the preceding proof and conclude that f : N → N for each
f ∈ F. The conclusion now follows from Theorem 4.3 of [17] applied to N . 
If the answer to Question 1 is negative then another question arises, namely to what
extent does DeMarr’s result character a real line interval among compact R-trees?
Question 2. If X is a compact R-tree which is not an interval, do there always
exist two continuous maps f, g : X → X, each having Lipschitz constant arbitrarily
near one, which fail to have a common fixed point?
2.3. Isometries and the Brodskii-Milman paper. We now turn to the original
paper of Brodskii and Milman [9] in which the concept of ‘normal structure’ was
introduced. This paper too was motivated by a common fixed point result.
Let K be a convex bounded set in a Banach space X. A point p0 in K is said to
be diametral if supp∈K kp − p0 k = d, the diameter of K.
Definition 2. The set K is said to have normal structure if every convex subset of
K which contains more than one point contains at least one nondiametral point.
A sequence {xn } of points of X is called a diametral sequence if the distance
dn from xn+1 to the convex hull of {xi : i ≤ n} tends to the diam{xi : 1 ≤ i < ∞}.
It was first shown in [9] that K has normal structure if and only if it contains no
diametral sequence. This fact of course implies that all compact convex sets have
50 W.A. KIRK

normal structure. However the main result in [9] deals explicitly with common fixed
points of isometries.
Assume now that K has normal structure and is compact in some topology in
which spheres are closed (e.g., the weak or weak∗ topologies). Then the following
construction yields a uniquely determined point, called the (Brodskii-Milman) center
of K. Let Kε be the intersection of all closed spheres of radius d−ε centered at points
of K; normal structure shows that there exists ε > 0 such that Kε 6= ∅. Compactness
shows that, if ε0 = sup {ε : Kε 6= ∅}, then Kε0 6= ∅. Denote Kε0 = K 1 . Transfinite
induction, using this step for nonlimit ordinals and intersections for limit ordinals,
yields a transfinite sequence K 1 ⊃ · · · ⊃ K θ ⊃ · · ·; normal structure assures that these
inclusions are proper until a θ is reached for which K θ is a singleton, the center of K.
This name is justified by the fact that the center of a convex bounded K for which
the above construction can be carried through is a fixed point of every isometry of
K onto itself. It is also shown that if K is compact in the norm topology, the center
is a fixed point of every nonexpansive mapping of K onto itself. This is of course a
direct consequence of a classical result of Freudenthal and Hurewicz [18]. (See the
next section.)
In his review [12] of the Brodskii-Milman paper, M. M. Day states that they prove
that the center of K is fixed under every isometry of K into itself. However it
seems likely that the mapping is meant to be surjective, especially in view of the
proof outlined above and the fact that the corresponding result about nonexpansive
mappings is clearly false if the mapping is not surjective.
Now let X be a Banach space and K a nonempty bounded subset of X. For x ∈ X,
set
rx (K) = sup {kx − yk : y ∈ K}
and denote the diameter of K by diam K := sup {ky − zk : y, z ∈ K} . As usual,
B (x; r) denotes the closed ball centered at x with radius r ≥ 0. co (S) denotes the
closure of the convex hull of S. The number rK := inf {rx (K) : x ∈ K} is called
the Chebyshev radius of K and the set C (K) := {x ∈ K : rx (K) = rK } is called the
Chebyshev center of K. It is well known that if K is weakly compact and convex
then C (K) is a nonempty closed convex subset of K, and if diam K > 0 then K has
normal structure if and only if C (K) is a proper subset of K. Indeed, the following
is a direct consequence of the definition of normal structure.
Lemma 1. Suppose K is a weakly compact convex subset of a Banach space with
diam K > 0, and suppose K has normal structure. Then diam C (K) = rK < diam
K.
Another fact is immediate:
Proposition 1. If T : K → K is a surjective isometry, then T (C (K)) = C (K) .
This brings us to another classical open question.
Question 3. If T is not surjective in the previous proposition, does one still have
T (C (K)) ⊂ C (K)?
Notice that if the answer to this question is affirmative, then the Brodskii-Milman
center of K is a fixed point of every isometry of K into itself.
 METRIC FIXED POINT THEORY 51

T.-C. Lim, et al., take up Question 3 in [31] and show that the answer is affir-
mative if the space X is uniformly convex. In this case the Chebyshev center of K
is a singleton and coincides with the Brodskii-Milman center of K. But the general
question remains open.

2.4. Commuting nonexpansive mappings. This brings us to the third influen-

tial paper, another 1963 paper by Ralph DeMarr [13], and its opening paragraph is
prophetic. We quote:
Kakutani [25] and Markov [32] have shown that if a commutative
family of continuous linear transformations of a linear topological
space into itself leaves some nonempty compact convex subset in-
variant, then the family has a common fixed point in this invariant
subset. The question naturally arises as to whether this is true if
one considers a commutative family of continuous (not necessarily
linear) transformations. We shall show that it is true in a rather spe-
cial, but non-trivial case, thus giving hope that further investigation
of the general question will yield positive results.
DeMarr then proceeded to state his result: Every commutative family of nonexpan-
sive mappings which map a compact convex subset of a Banach space into itself has a
common fixed point. It is interesting to note that Lemma 1 of DeMarr’s paper states
precisely that if a nonempty compact convex subset of a Banach space has positive
diameter, then it must contain a nondiametral point, a fact already apparent from
the results of Brodskii and Milman.
DeMarr’s paper motivated the Belluce-Kirk paper [4] in which it is proved that
every finite family of commuting nonexpansive mappings defined on a weakly compact
convex set K has a common fixed point if K has normal structure. Since DeMarr’s
result holds for infinite families this raised the obvious question of whether the result
of [4] could be extended to infinite families, a question subsequently settled in the
affirmative by T.-C. Lim [30], and also by R. E. Bruck in his definitive study [10] of
nonexpansive retracts.

2.5. Surjective isometries. Another metric question may be of interest. The result
of Freudenthal and Hurewicz mentioned Section 2.3 states that a surjective nonexpan-
sive self-mapping of a compact metric space is necessarily an isometry. It is interesting
to note that there are noncompact spaces for which this assertion also holds.

Example 1. Let {en } be the standard unit basis in `2 , and for each n ≥ 1 let
Ln = ten : 0 ≤ t ≤ 1 − n1 . Take X = ∪Ln . Then the only surjective nonexpansive


mapping of X onto X is the identity.

Example 2. Consider R2 with the radial metric ρ. Let {xn } be a sequence of distinct
points on the unit sphere of R2 . Let yn = xn for 1 ≤ n ≤ 10, and for n > 10 let yn
be the point on the segment [0, xn ] such that ρ (0, yn ) = 1 − n1 . Now let Ln = [0, yn ]
and take X = ∪Ln . In this case every surjective nonexpansive mapping of X onto X
is an isometry, but there exist nontrivial surjective nonexpansive mappings.
52 W.A. KIRK

Question 4. Is it possible to classify metric spaces for which surjective nonex-

pansive self-mappings are always isometries? Do such spaces share any additional
properties with compactness?

3. New Directions
There has been a trend over the years to seek applications of metric fixed point
theory in settings where the underlying algebraic linear structure of a Banach space
is not present, at least in any explicit sense. While this trend has been somewhat
sporadic, several developments stand out.
First there was the era of the early to mid-eighties. This brought the discovery
that ideas of nonlinear analysis could be applied to hyperbolic geometry. The pi-
oneering paper [21] by Goebel, Sekowski, and Stachura appeared in 1980. Notable
in this development is the 1984 book Uniform convexity, hyperbolic geometry, and
nonexpansive mappings by Goebel and Reich [20]. At about the same time there was
the realization that there is a beautiful synergism between hyperconvex metric spaces
and nonexpansive mappings. This came to fruition with a 1986 paper of J.-B. Bail-
lion [2]. The early eighties also brought the realization that certain iteration processes
(Krasnoselskii, Ishikawa) could be carried out in a so-called hyperbolic setting (see be-
low) ([26], [19]). Later came the discovery that complete R-trees (often called metric
trees) provided prime examples of hyperconvex metric spaces with unique topologi-
cal properties (Kirk [27]). Also it became clear that many of the standard ideas of
nonlinear analysis could be extended to the class of so-called CAT(0) spaces [28]. We
discuss some of these developments in more detail below. First, however, we fix some
terminology.
HYPERBOLIC SPACES. The meaning of the term ‘hyperbolic metric’ has been
inconsistent. Here we adopt the terminology of Kohlenbach [29]. A hyperbolic space
is a triple (X, ρ, W ), where (X, ρ) is a metric space and W : X × X × [0, 1] → X
satisfies
(W1) ρ (z, W (x, y, λ)) ≤ (1 − λ) ρ (z, x) + λρ (z, y) ,



(W2) ρ W (x, y, λ) , W x, y, λ̄ = λ − λ̄ ρ (x, y) ,
(W3) W (x, y, λ) = W (y, x, 1 − λ) ,
(W4) ρ (W (x, z, λ) , W (y, w, λ)) ≤ (1 − λ) ρ (x, y) + λρ (z, w) .
If only axiom (W1) is assumed this structure is a convex metric space in the
sense of Takahashi [43]. If (W1)-(W3) are assumed the notion is equivalent to spaces
called of hyperbolic type in [19]. Axiom (W4) is used for example in [38]. However
Kohlenbach’s definition is less restrictive than that given in [38] in that it does not
require the existence of metric lines. Hence it includes all CAT(0) spaces, whereas
the definition in [38] includes only those CAT(0) space which have the unique geodesic
extension property.

3.1. CAT(0) spaces and graph theory. (Much of the material in this section is
taken from the author’s article [28].)
 METRIC FIXED POINT THEORY 53

Theorem 7 (Kirk, 1981; cf. Penot, 1979). Let (X, d) be a nonempty bounded metric
space and suppose the admissible sets (ball intersections) in X are countably compact
and normal. Then every nonexpansive mapping T : X → X has a fixed point.
The following statement is found in M. van de Vel’s book ”Theory of Convex
Structures” [44].
[Problem 6.23.3] Observe that an EP function G → G is nonex-
pansive. If G is a median graph, then the diametric convex sets
are precisely the graphic cubes. However, the preceding discussion
shows that the median convexity of G will rarely include metric disks.
Nevertheless, the similarities between the Invariant Cube Theorem
and [Theorem 7] are too detailed to be pure coincidences. Is there a
common generalization of both results?
A few years later Victor Chepoi responded to the above question in [11].
A Euclidean cell is a convex polytope in some Euclidean space. By a piecewise
Euclidean (PE) cell complex is meant a space H formed by gluing together Euclidean
cells via isometries of their faces, together with the subdivision of H into cells. It is
assumed that the intersection of two cells is either empty or a single face of each of
the cells. The following can be found in the book by Bridson and Haefliger.
Proposition 2. Let H be a simply connected PE complex with only finitely many
isometry types of cells and endowed with the intrinsic metric. Then H satisfies CAT(0)
if and only if given any two geodesic segments α and β in H, then the function f :
[0, 1] → H given by f (t) = d (α (t) , β (t)) is convex.
Every median graph G gives rise to an abstract cubical complex K (G) consisting
of all cubes of G, i.e., subgraphs of G isomorphic to cubes of any dimension. The
geometric realization |K (G)| is called a median complex. In his paper Chepoi proves
that CAT(0) cube complexes coincide with the cubical cell complexes arising from
median graphs. In [11] Chepoi goes on to state:
It seems that the CAT(0) property of median complexes gives a par-
tial explanation to [van de Vel’s problem]. Namely, assume addi-
tionally that f : G → G is a cell-to-cell map. Then f induces a
continuous map f : |K (G)| → |K(G)| by extending f affinely over
the geometric cubes. As a result, we obtain a nonexpansive map of
a CAT(0) space |K(G)| . From [Proposition 2] and the result of Kirk
[Theorem 7], f has a fixed point p. Consider the smallest cube C of
G such that p belongs to the relative interior of |C| . Clearly, C is an
invariant cube of the original map f. This remark can be extended to
obtain fixed simplexes of edge-preserving maps of underlying graphs
of CAT(0) simplicial complexes, but does not extend to arbitrary
edge-preserving maps of CAT(0) polysimplicial complexes (cell com-
plexes whose cells are products of simplexes).
These two quotes raised a number of questions, among them: What are edge
preserving maps? What are median graphs. What does Invariant Cube Theorem
say? What are CAT(0) spaces?
54 W.A. KIRK

3.2. Some answers. A graph is an ordered pair (V, E) where V is a set and E is a
binary relation on V (E ⊆ V × V ) . Elements of E are called edges. We are concerned
here with (undirected) graphs that have a “loop” at every vertex (i.e., (a, a) ∈ E for
each a ∈ V ) and no “multiple” edges. Such graphs are called reflexive. In this case
E ⊆ V × V corresponds to a reflexive (and symmetric) binary relation on V.
Given a graph G = (V, E) , a path of G is a sequence a0 , a1 , · · ·, an−1 , · · · with
(ai+1 , ai ) ∈ E for each i = 0, 1, 2, · · ·. A cycle is a finite path (a0 , a1 · ··, an−1 ) with
(a0 , an−1 ) ∈ E. A graph is connected if there is a finite path joining any two of its
vertices. A finite path (a0 , a1 , · · ·, an−1 ) is said to have length n. Finally, a tree is a
connected graph with no cycles.
For a graph G = (V, E) a map f : V → V is edge-preserving if (a, b) ∈ E ⇒
(f (a) , f (b)) ∈ E. For such a mapping we simply write f : G → G. There is a
standard way of metrizing connected graphs; let each edge have length one and take
distance d (a, b) between two vertices a and b to be the length of the shortest path
joining them. With this metric edge preserving mappings become precisely the non-
expansive mappings. Keep in mind that in a reflexive graph an edge-preserving map
may collapse edges between distinct points since loops are allowed.
The classical Fixed Edge Theorem in graph theory due to Nowakowski and Rival
[34] asserts that an edge preserving mapping defined on a connected graph that has
no cycles or infinite paths always leaves some edge of the graph fixed.
The Invariant Cube Theorem (due to H.-J. Bandelt and M. van de Vel [3]) ap-
peared in 1987 . It involves median graphs. A median graph is a connected graph G
(with standard metric d) that has the following property: For each triple of points
a1 , a2 , a3 ∈ G there is a unique point x ∈ G such that

d (ai , x) + d (x, aj ) = d (ai , aj ) for all i 6= j ∈ {1, 2, 3} .

In this case x is called the median of a1 , a2 , a3 and is written x = m (a1 , a2 , a3 ) .

Thus each three points of G have a unique common between point. In particular a
connected graph which has no cycles (to which the Fixed Edge Theorem applies) is a
median graph.
The ternary operation m : G3 → G has the following properties, of which the first
two are obvious; the third less so:
(M 1) m (a, b, c) = m (b, a, c) = m (c, b, a) ,
(M 2) m (a, a, b) = a,
(M 3) m (m (a, b, c) , u, v) = m (a, m (b, u, v) , m (c, u, v)) .

Theorem 8 (Invariant Cube Theorem ). Let G be a finite median graph and let
f : G → G be an edge preserving function. Then there is a graphic cube of G that is
mapped isomorphically onto itself by f. If the number of vertices of G is odd and f is
an automorphism, then f has a fixed point.

3.3. The approximate fixed point property. A subset K of a metric space is said
to have the approximate fixed point property (for nonexpansive mappings) if given any
nonexpansive f : K → K, inf {d (x, f (x)) : x ∈ K} = 0. To characterize this concept
we need some more definitions.
 METRIC FIXED POINT THEORY 55

Definition 3. A geodesic metric space X is said to have the geodesic extension

property if for every local geodesic c : [a, b] → X, with a 6= b, there exists ε > 0 and a
local geodesic c0 : [a, b + ε] → X such that c0 |[a,b] = c.
Definition 4 (Shafrir, 1990). Let X be a metric space. A curve γ : X → [0, ∞) is
said to be directional ( with constant b) if there is b ≥ 0 such that
t − s − b ≤ d (γ (s) , γ (t)) ≤ t − s
for all t ≥ s ≥ 0. A subset of X is said to be directionally bounded if it does not
contain a directional curve.
Lemma 2. If X is a CAT (0) space, then X has the geodesic extension property
if and only if every non-constant geodesic c : [a, b] → X can be extended to a line
c : R → X.
If a complete CAT(0) space is homeomorphic to a finite dimensional manifold, then
it always has the geodesic extension property ([8]).
In [41] it is proved that a closed convex subset of a complete hyperbolic metric
space with the geodesic extension property has the approximate fixed point property
for nonexpansive mappings if and only if it is directionally bounded. As an immediate
corollary, a closed convex subset of a complete CAT(0) space with the geodesic exten-
sion property has the approximate fixed point property if and only if it is directionally
bounded. However in this case a stronger assertion holds.
Theorem 9 ([28]). If a complete CAT(0) space is geodesically bounded then it is
directionally bounded.
In view of Shafrir’s result we now have the following.
Corollary 1. A closed convex subset of a complete CAT(0) space with the geodesic
extension property has the approximate fixed point property for nonexpansive mappings
if and only if it does not contain a geodesic ray.
Proof of Theorem 9. ([28])Suppose K is a closed convex set in X and suppose K
contains a directional curve γ. We show that this implies K contains a geodesic ray.
Let xn = γ (n) , n = 0, 1, 2, · · ·, and fix an arbitrary ρ > b, where b is the directional
constant associated with γ. For each n ≥ ρ, let yn be the point of geodesic segment
[x0 , xn ] with distance ρ from x0 . Now suppose m > n ≥ ρ, and let α be the comparison
angle ∠ ¯ x̄ (x̄n , x̄m )in R2 . By the law of cosines
0

2 2 2
d (x0 , xn ) + d (x0 , xm ) − d (xn , xm )
cos (α) = .
2d (x0 , xn ) d (x0 , xm )
Using the inequalities
n − b ≤ d (x0 , xn ) ≤ n;
m − b ≤ d (x0 , xm ) ≤ m;
m − n ≥ d (xn , xm ) ,
56 W.A. KIRK

we have
2 2 2
(n − b) + (m − b) − (m − n)
cos (α) ≥
" 2nm
# " #
2 2 2
n (n − b) m (m − b) (m − n)
= + −
2m n2 2n m2 2nm
2
n m (m − n) b b b2
= + − − − +
2m 2n 2nm  n m nm
1 1 b
=1−b + − .
n m nm
Thus cos (α) → 1 as m, n → ∞; hence α → 0. If ȳn , ȳm are the respective points of
the comparison triangle ∆ (x̄0 , x̄n , x̄m ) corresponding to yn , ym , then by the CAT(0)
inequality d (yn , ym ) ≤ d (ȳn , ȳm ) . The fact that α → 0 as m, n → ∞ implies that
{ȳn } , hence {yn } , is a Cauchy sequence. Since ρ > b is arbitrary it now follows that
the sequence {[x0 , xn ]} of geodesic segments converges to a geodesic ray issuing from
x0 . 
Question 5. Is the geodesic extension property needed for the characterization of
Corollary 2.1?
Other questions remain open as well.
Question 6. Does Corollary 2.1 extend to hyperconvex metric spaces?
Question 7. Is it the case that a closed convex subset of a complete CAT (0) space
has the fixed point property if and only if it is bounded (as in the Hilbert space case,
cf. Ray, [37])?

References
[1] N. Aronszajn and P. Panitchpakdi, Extensions of uniformly continuous transformations and
hyperconvex metric spaces, Pacific J. Math., 6(1956), 405-439.
[2] J.-B. Baillon, Nonexpansive mappings and hyperconvex spaces, in Fixed Point Theory and its
Applications (R. F. Brown, ed.), Contemporary Mathematics, 72, Amer. Math. Soc., Providence,
RI, 1988, 11-19.
[3] H.-J. Bandelt and M. van de Vel, A fixed cube theorem for median graphs, Discrete Math.,
62(1987), 129-137.
[4] L. P. Belluce and W. A. Kirk, Fixed point theorems for families of contraction mappings, Pacific
J. Math., 18(1966),213-217.
[5] M. Bestvina, R-trees in topology, geometry, and group theory, in Handbook of geometric topol-
ogy, 55-91, North-Holland, Amsterdam, 2002.
[6] S.A. Bogatyi and O.D. Frolkina, A common fixed point of commuting mappings of a tree (Rus-
sian) Vestnik Moskov. Univ. Ser. I Mat. Mekh. 2002, no. 6, 3-10, 69; translation in Moscow
Univ. Math. Bull., 57 (2002), no. 6, 1-8 (2003).
[7] W.M. Boyce, Commuting functions with no common fixed point, Trans. Amer. Math. Soc.,
137(1969), 77-92.
[8] M. Bridson and A. Haefliger, Metric spaces of non-positive curvature, Springer-Verlag, Berlin,
Heidelberg, New York, 1999.
[9] M.S. Brodskii and D.P. Milman, On the center of a convex set (Russian), Dokl. Akad. Nauk.
SSSR, 59(1948), 837-840.
[10] R.E. Bruck, A common fixed point theorem for a commuting family of nonexpansive mappings,
Pacific J. Math., 53(1974), 59-71.
 METRIC FIXED POINT THEORY 57

[11] V. Chepoi, Graphs of some CAT(0) complexes, Advances in Appl. Math., 24(2000), 125-179.
[12] M.M. Day, Review of: On the center of a convex set (Russian), Dokl. Akad. Nauk. SSSR,
59(1948), 837-840, Mathematical Reviews MR0024073 (9,448f), Amer. Math. Soc., Providence,
1948.
[13] R. DeMarr, Common fixed points for commuting contraction mappings, Pacific J. Math.,
13(1963), 1139-1141.
[14] R. DeMarr, A common fixed point theorem for commuting mappings, Amer. Math. Monthly,
70(1963), 535–537.
[15] A. Dress and W. Terhalle, The real tree, Adv. Math., 120(1996), no. 2, 283-301.
[16] R. Espı́nola and M.A. Khamsi, Introduction to hyperconvex spaces, in Handbook of metric fixed
point theory, 391-435, Kluwer Acad. Publ., Dordrecht, 2001.
[17] R. Espı́nola and W.A. Kirk, Fixed point theorems in R-trees with applications to graph theory,
Topology Appl., 153(2006), 1046-1055.
[18] H. Freudenthal and W. Hurewicz, Dehnungen, Verkürzungen, Isometrien, Fund. Math.,
72(1936), 120-122.
[19] K. Goebel and W.A. Kirk, Iteration processes for nonexpansive mappings. Topological methods
in nonlinear functional analysis (Toronto, Ont., 1982), 115–123, Contemp. Math., 21, Amer.
Math. Soc., Providence, RI, 1983.
[20] K. Goebel, S. Reich, Uniform convexity, hyperbolic geometry, and nonexpansive mappings,
Monographs and Textbooks in Pure and Applied Mathematics, 83, Marcel Dekker, Inc., New
York, 1984. ix+170 pp.
[21] K. Goebel, T. Sekowski, A. Stachura, Uniform convexity of the hyperbolic metric and fixed
points of holomorphic mappings in the Hilbert ball, Nonlinear Anal., 4(1980), no. 5, 1011-1021.
[22] J.P. Huneke, Two commuting continuous functions from the closed unit interval onto the closed
unit interval without a common fixed point, Topological Dynamics (Symposium, Colorado State
Univ., Ft. Collins, Colo., 1967) pp. 291-298 Benjamin, New York, 1968.
[23] J.R. Isbell, Commuting mappings of trees, Research problem #7, Bull. Amer. Math. Soc.,
63(1957), 419.
[24] G. Jungck, Commuting mappings and common fixed points, Amer. Math. Monthly, 73(1966),
735-738.
[25] S. Kakutani, Two fixed-point theorems concerning bicompact convex sets, Proc. Imp. Acad.
Tokyo, 14(1938), 242-245.
[26] W.A. Kirk, Krasnoselskii’s iteration process in hyperbolic space, Numer. Funct. Anal. Optimiz.,
4(1981-82), 371-381.
[27] W.A. Kirk, Hyperconvexity of R-trees, Fund. Math., 156(1998), 67-72.
[28] W.A. Kirk, Geodesic geometry and fixed point theory. Seminar of Mathematical Analysis
(Malaga/Seville, 2002/2003), 195-225, Colecc. Abierta, 64, Univ. Sevilla Secr. Publ., Seville,
2003.
[29] U. Kohlenbach, Some logical metatheorems with applications in functional analysis, Trans.
Amer. Math. Soc., 357(2005), 89-128.
[30] T.-C. Lim, A fixed point theorem for families on nonexpansive mappings, Pacific J. Math.,
53(1974), 487-493.
[31] T.-C. Lim, P.-K. Lin, C. Petalas and T. Vidalis, Fixed points of isometries on weakly compact
convex sets, J. Math. Anal. Appl., 282(2003), 1-7.
[32] A. Markov, Quelques Theoremes sur les ensembles Abeliens, Dokl. Akad. Nauk. SSSR, 10(1936),
311-314.
[33] J.E. Maxfield and W.J. Mourant, Common fixed points of commuting continuous functions on
the unit interval, Nederl. Akad. Wetensch. Proc. Ser. A, 68 - Indag. Math. 27(1965), 668-670.
[34] R. Nowakowski and I. Rival, Fixed-edge theorem for graphs with loops, J. Graph Theory, 3(1979),
339-350.
[35] J.P. Penot, A fixed point theorem without convexity, Bull. Soc. Math. France, 60(1979), 129-152.
[36] A. Quilliot, On the Helly property working as a compactness criterion on graphs, J. Comb.
Theory, Series A, 40(1985), 186-193.
58 W.A. KIRK

[37] W.O. Ray, The fixed point property and unbounded sets in Hilbert space, Trans. Amer. Math.
Soc., 258(1980), 531-537.
[38] S. Reich and I. Shafrir, Nonexpansive iterations in hyperbolic spaces, Nonlinear Anal., 15(1990),
537-558.
[39] I.A. Rus, Fixed Point Structure Theory, Cluj University Press, 2006.
[40] A.J. Schwartz, Common periodic points of commuting functions, Mich. J. Math., 12(1965),
353-355.
[41] I. Shafrir, The aproximate fixed point property in Banach and hyperbolic spaces, Israel J. Math.,
71(1990), 211-223.
[42] R. Sine, Hyperconvexity of approximate fixed points, Nonlinear Anal., 13(1989), 863-869.
[43] W. Takahashi, A convexity in metric space and nonexpansive mappings I, Kodai Math. Sem.
Rep., 22(1970), 142-149.
[44] M. van de Vel, Theory of Convex Structures, Elsevier, Amsterdam, 1993.
[45] G.S. Young, The introduction of local connectivity by a change of topology, Amer. J. Math.,
68(1946), 479-494.
[46] G.S. Young, Fixed point theorems for arcwise connected continua, Proc. Amer. Math. Soc.,
11(1960), 880-884.

Received: 31.12.2009; Accepted:10.02.2010.