Research Collection
Doctoral Thesis
The injective hull of hyperbolic groups
Author(s): Moezzi, Arvin
Publication Date: 2010
Permanent Link: https://doi.org/10.3929/ethz-a-006060216
Rights / License: In Copyright - Non-Commercial Use Permitted
This page was generated automatically upon download from the ETH Zurich Research Collection. For more information please consult the Terms of use.
ETH Library DISS. ETH No. 18860
The Injective Hull of Hyperbolic Groups
A dissertation submitted to ETH ZURICH
for the degree of Doctor of Sciences
presented by Arvin Moezzi Dipl. Math. ETH Zurich
born 27.08.1976 citizen of Basel BS, Switzerland
accepted on the recommendation of Prof. Dr. Urs Lang, examiner Prof. Dr. Viktor Schroeder, co-examiner
2010
For my family
Abstract
The construction of the injective hull of a metric space goes back to Isbell in the sixties. In this thesis we investigate this construction in the setting of word metrics of finitely generated groups, hyperbolic groups in particular. There is a natural action of the group on its injective hull. We are mainly interested when this action is properly discontinuous and cocompact; we give a necessary condition. Furthermore, we introduce the concept of Axiom Y and show for a hyperbolic group satisfying Axiom Y that this action is cocompact and properly discontinuous. Moreover, in this case the injective hull can be given explicitly the structure of a polyhedral complex such that each cell in itself is injective and the group acts by cellular isometries on it. We give some simple examples of groups satisfying Axiom Y and a counter example of a Cayley graph of a hyperbolic group that does not satisfy Axiom Y.
Zusammenfassung
Die injektive H¨ulle von metrischen R¨aumen geht zur¨uck auf Isbell in den Sechzigern. Wir betrachten in dieser Arbeit die injektive H¨ulle einer endlich erzeugten Gruppe versehen mit der Wort Metrik - hyperbolische Grup- pen aber im besonderen. Die Gruppe selbst operiert auf nat¨urliche Weise isometrisch auf ihrer injektiven H¨ulle. Es stellt sich die Frage, wann diese Gruppenaktion eigentlich und cocompakt ist. Wir geben eine notwendige Bedingung an die Gruppe. Desweiteren f¨uhren wir die Eigenschaft Ax- iom Y ein und zeigen, dass jede hyperbolische Gruppe, die das Axiom Y erf¨ullt, eigentlich und cocompakt auf ihrer injektiven H¨ulle operiert. In diesem Fall l¨asst sich weiter zeigen, dass die injektive H¨ulle die Struktur eines polyedrischen Komplexes tr¨agt, wobei jede Zelle f¨ur sich wieder injek- tiv ist und die Gruppe zellul¨ar operiert. Wir geben einige einfache Beispiele von Gruppen, die das Axiom Y erf¨ullen, und auch ein Gegenbeispiel eines Cayley Graphen einer hyperbolischen Gruppe, welcher Axiom Y nicht erf¨ullt.
Danksagung
Die vorliegende Doktorarbeit w¨are ohne die Hilfe und Unterst¨utzung an- derer in dieser Form nicht m¨oglich gewesen. Daher m¨ochte ich gerne fol- genden Menschen an dieser Stelle danken; allen voran meinem Doktorvater, Professor Urs Lang. Er war es, der mich in die metrische Geometrie und geometrische Gruppentheorie einf¨uhrte. Dieses faszienerende Teilgebiet der Mathematik w¨are mir ansonsten wohl verschlossen geblieben. Auch nahm er sich stets die n¨otige Zeit f¨ur Fragen und hat die vorliegende Arbeit mit wichti- gen Ideen und Konzepten vorangetrieben; ohne seine grosse Mithilfe w¨are die Doktorarbeit niemals zu dem geworden, was sie ist. Viel Zeit investierte er zudem, die Arbeit mit der von ihm gewohnten Genauigkeit durchzuar- beiten und half mir so, auch die letzten Fehler auszumerzen. Vorallem aber m¨ochte ich ihm daf¨ur danken, dass er mir vertraute und an meine F¨ahigkeiten glaubte - tausend Dank daf¨ur! Ein grosses Dankesch¨on gilt auch dem Koreferenten, Professor Viktor Schroeder, der sich die M¨uhe machte, meine Arbeit zu lesen und f¨ur die m¨undliche Doktorpr¨ufung kurzfristig nach Z¨urich zu reisen. Mein besonderer Dank gilt auch zwei guten Freunden von mir, Johnny Micic und Driton Komani. Beide haben grosse Teile der Arbeit mehrfach korrekturgelesen und haben mich bei der Pr¨ufungsvorbereitung tatkr¨aftig unterst¨utzt. Sehr wichtig f¨ur mich waren auch die zahlreichen mathematis- chen Debatten mit ihnen zu allen m¨oglichen Themen. Diese Diskussionen f¨uhrten mir immer wieder vor Augen was ich am liebsten mache: Mathe- matik. Der lange Weg zum Abschluss der Doktorarbeit w¨are sicherlich unendlich l¨anger gewesen ohne die Freundschaft und die moralische Unterst¨utzung vieler. Nennen m¨ochte ich an dieser Stelle Fabian Roth, Yves Hauser, B´en´edicte Gros, Thilo Schlichenmaier, Stefan Wenger, Michael Anderegg, Driton Komani, Johnny Micic, Leandra Simitovic, Theo Buehler, Beat Steiner, Kathrin Signer, Luzia Huggentobler, Demian Wismer, Kascha, Robin Krom, Ivo K¨ahlin, Lorenz Reichel, Thomas Huber, Ralph Kirchhofer und Jan Oppliger. Vielen Dank, dass ihr f¨ur mich da wart. Und vielen Dank f¨ur eine sehr gelungene Doktorfeier. Danken m¨ochte ich auch Tim Boin f¨urs Korrekturlesen der deutschen Passagen. Zuletzt und ganz besonders m¨ochte ich meiner Familie danken. Ohne ihre Liebe und Unterst¨utzung, die man leider allzu oft als selbstverst¨andlich betrachtet, h¨atte ich dieses Ziel niemals erreicht. Danke f¨ur eure St¨utze und eine Konstante in meinem Leben!
Contents
1 Isbell’s Injective Hull 5 1.1 Preliminaries ...... 5 1.2 isoM-Injectivity ...... 8 1.3 Bicombing and isoperimetric inequality ...... 13 1.4 Isbell’s injective hull ...... 15 1.5 Gromov–Hausdorffconvergence ...... 21 1.6 Compactmetricspaces ...... 26 1.7 Stronglyconvexsubspaces ...... 30 1.8 Normedvectorspaces...... 35 n 1.8.1 Affine, injective subspaces of R∞ ...... 41 n 1.8.2 Injective polyhedrons in R∞ ...... 46 1.9 Finitemetricspaces...... 51 2 Groups and their Injective Hull 57 2.1 Cayleygraphs...... 59 2.2 Conetypes...... 62 2.3 Combablegroups ...... 63 2.4 Hyperbolicgroups...... 69 2.5 AxiomY...... 72 2.6 Injectivehullofhyperbolicgroups ...... 73 3 Examples and a counter-example 83 3.1 Hyperbolicgroups...... 83 3.1.1 ACounterexample ...... 83 3.1.2 Small cancellation groups ...... 86 3.2 Abeliangroups ...... 90 A The Stone–Cechˇ compactification 95 B Group actions on geodesic metric spaces 99 CTheGrowthofaDehnFunction 101
Introduction
A metric space I is said to be isoM-injective if every 1-Lipschitz map, f : A → I, can be extended to B whenever A is isometrically embedded in B. In other words, there exists a 1-Lipschitz map f¯: B → I such that the diagram f A / I ? ¯ f B commutes. isoM-injective metric spaces share nice metric properties; they are contractible and complete geodesic metric spaces, and they all admit an isoperimetric inequality of euclidean type in the class of integral metric currents. One may wonder whether every X can be isometrically embedded into some isoM-injective metric space I and if there is a smallest such space, in the sense that whenever X can be isometrically embedded into some other isoM-injective metric space J, there exists an isometric embedding of I into J such that X / Io J commutes. In this case I is unique up to an isometry, and it is said to be the injective hull of X; we shall write EX for I. In [22] Isbell shows that every metric space X can indeed be embedded isometrically into a smallest isoM-injective metric space EX. He gives an explicit description of EX in terms of some function space on X. We will show in Theorem 1.55 that the assignment X → EX is continuous with respect to the Gromov–Hausdorff distance; that is E E Xn →dGH X ⇒ Xn →dGH X. Furthermore, in Theorem 2.58 we are going to show that Gromov hyperbol- icity for geodesic metric spaces is preserved under this assignment: 4
Theorem. The injective hull of a geodesic hyperbolic metric space is hy- perbolic. However, in this thesis we are mostly interested in the following ques- tion: when does a finitely generated group (Γ,S) act nicely, i.e. properly discontinuous and cocompactly, on EΓS, where ΓS denotes the metric space Γ with the word metric; in particular what is the metric structure of EΓS in this case? We shall give in 2.39 a necessary condition for Γ to act nicely on EΓS: Proposition. Every group that acts properly discontinuous and cocom- pactly on an injective metric space is a combable group. But this condition is by no means sufficient. The finitely generated abelian group Zn, for example, with the standard generating set is a combable group but does not act nicely on EZn; compare page 68. In Section 2.5 we introduce the concept of Axiom Y(δ), a metric property which postulates that for any three points x, z, z′ with d(z, z′) ≤ δ one can find a point y such that d(z,y) ≤ R = R(δ) and
d(x, z) = d(x, y)+ d(y, z) d(x, z′) = d(x, y)+ d(y, z′).
The main result of this thesis, Theorem 2.71, can then be formulated as follows.
Theorem. Suppose (Γ,S) is a δ-hyperbolic group and ΓS satisfies Axiom Y(δ). Then, EΓS is a proper locally finite polyhedral complex with finitely many isometry types of cells, and Γ acts properly discontinuous and cocom- pactly on EΓS by cellular isometries.
The cell structure on EΓS can be given explicitly in terms of some admis- sible graphs on ΓS, and every cell endowed with the induced metric is itself isoM-injective; see Theorem 2.70. Moreover, we shall see in 2.72 that every tangent cone of EΓS is an isoM-injective metric space. For further study, we hope that this property would give some good link conditions on the cell structure of EΓS, allowing us to alter the metric on the cells in such a way that EΓS becomes a CAT(0) space. This would show that every hyperbolic group satisfying Axiom Y(δ) is a CAT(0) group. Finally we will see in Chapter 3 that a subclass of small cancellation groups and all finitely generated abelian groups satisfy Axiom Y(δ) for every δ > 0. We will also give a particular presentation of a δ-hyperbolic group which does not satisfy Axiom Y(δ). Chapter 1
Isbell’s Injective Hull
1.1 Preliminaries
We recall the notion of an injective object and injective hull (cf. [1]). Let C denote an arbitrary category and H a distinguished class of morphisms in C. We call all morphisms from H H-morphisms. An H-morphism i : A → B is called H-essential if for any g : B → C the following holds g ◦ i ∈H⇒ g ∈ H. 1.1 Definition. An object I in C is said to be H-injective if for every pair of objects A, B and every pair of morphisms f : A → I and i : A → B with i ∈ H there exists a morphism f¯ : B → I, such that the following diagram commutes f A / I . (1.2) ? i∈H ¯ f B The extension f¯ need not be unique. If f : A → I is H-essential then I is called the H-injective hull of A. 1.3 Lemma. The H-injective hull is unique up to a not necessarily unique isomorphism. Proof. Let I,I′ be two H-injective hulls of A with H-morphisms i: A → I and i′ : A → I′. From Definition 1.1 we have the following commuting diagram i idI A / I / I (1.4) ? O ? O ? ? idI′ ?? ¯ ′ ? i ? idI i ? ? ? ′ / ′ I ′ I idI 6 Chapter 1. Isbell’s Injective Hull where by definition of H-essential all the maps in the diagram are in H. In particular idI′ ◦ ¯i = idI′ and idI ◦ idI′ = idI . Hence idI′ is an isomorphism and therefore ¯i is also an isomorphism. 1.5 Definition. An H-retract R of I is an H-morphism i: R → I, such that i has a left-inverse r : I → R , i.e. r ◦ i = idR. The left-inverse r is called a retraction of I onto R.
1.6 Definition. Let {Xα}α∈I be a family of objects in C. An object α Xα together with a family of morphisms πα : α Xα → Xα is said toQ be the product of {Xα}α∈I if the following universalQ property is satisfied. For every object Y in C and any family of morphisms fα : Y → Xα there is a unique f : Y → α Xα such that πα ◦ f = fα for all α ∈ I. Q α Xα (1.7) w O πα wwQ ww ww {ww ∃!f Xα dHH HH HH fα HH HH Y
By the universal property the product of a family of objects is unique in C up to an isomorphism. The following fact will be needed later. 1.8 Lemma. The subcategory of H-injective objects is closed under retrac- tions and products. Proof. First of all we show that every H-retract R of an H-injective object I is again H-injective. Consider the commuting diagram
r u R i I O 6> O ~~ f ~~ i◦f ~~i◦f ~~ / A j∈H B where the extension i ◦ f exists since I is H-injective. Define f¯ := r ◦ i ◦ f. Since
f¯◦ j = r ◦ i ◦ f ◦ j = r ◦ i ◦ f ◦ j = r ◦ i ◦ f = idR ◦ f = f, it follows that f¯ is an extension of f in the sense of Definition 1.1. Therefore, R is H-injective. Preliminaries 7
Let α Xα and πα : α Xα → Xα be given as in Definition 1.6 and let all Xα beQ H-injective. WeQ need to show that for any f : A → α Xα and ¯ any H-morphism i : A → B there is an extension f : B → Qα Xα with ¯ f ◦ i = f. Since Xα is H-injective there is an extension πα ◦ f :QB → Xα of πα ◦ f : A → Xα, for every α ∈ I. In other words πα ◦ f ◦ i = πα ◦ f. Due to the universal property of the product α Xα in Definition 1.6 there is a ¯ ¯ unique morphism f : B → α Xα such thatQ πα ◦ f = πα ◦ f. We need to show that f¯ is an extensionQ of f, i.e. f¯◦ i = f. As
πα ◦ (f¯◦ i)= πα ◦ f ◦ i = πα ◦ f =: fα for all α ∈ I and since f is unique with this property, it follows that f¯◦i = f.
α Xα w O bFF πα wwQ F f ww FF ww FF {ww FF Xα f A dH ww H ww H ww πα◦f H ww i H {ww B
1.9 Definition. An absolute H-retract is an object A ∈ C with the property that every H-morphism i: A → B has a left-inverse in C. In other words there is a morphism r : B → A with r ◦ i = idA. 1.10 Lemma. Every H-injective object in C is an absolute H-retract.
Proof. Let I be H-injective and let i : I → B be an H-morphism. Then there exists a morphism idI : B → I such that
idI I / I ? i∈H idI B commutes. Therefore, r := idI is a left-inverse of i. Note that in general an absolute H-retract need not be H-injective. A counter-example is given next.
1.11 Example. Let C be the category consisting of {Z3, Z3 × Z5, Z3 ⋊ Z2} and of all group homomorphisms between them, and let H be the subclass 8 Chapter 1. Isbell’s Injective Hull
∼ of all injective homomorphisms in C; note that Z3 ⋊ Z2 = S3. Since 15 = |Z3 × Z5| > |Z3|, |Z3 ⋊ Z2| there is no H-morphism from Z3 × Z5 into Z3 or Z3 ⋊ Z2. On the other hand every H-morphism from Z3 × Z5 into itself is an isomorphism and hence invertible. It follows that Z3 × Z5 is an absolute H- retract. But on the other hand there is no group homomorphism ¯i: Z3⋊Z2 → Z3 × Z5 such that i Z3 / Z3 × Z5 8 qqq j∈H qqq qq ¯i qq Z3 ⋊ Z2 commutes, j(z) := (z, 0) ∈ Z3 ⋊ Z2 and i(z) := (z, 0) ∈ Z3 × Z5. Otherwise Z2 < Z3 ⋊ Z2 would be a normal subgroup, but it isn’t. Hence, Z3 × Z5 is not H-injective in C. However if a category has enough H-injective objects then the two no- tions, H-injectivity and absolute H-retract, are in fact equivalent. This is the subject of the next lemma. We say that a category C has enough H- injectives if every object in C can be embedded into an H-injective object by means of an H-morphism. 1.12 Lemma. Let C be a category that has enough H-injectives. Then, an object I ∈ C is H-injective if and only if it is an absolute H-retract in C. Proof. One implication directly follows from Lemma 1.10. We only need to show that every absolute H-retract R is H-injective. Let i: R → I be an H- morphism into an H-injective object I ∈ C. This exists since C has enough H-injectives. Then by assumption there is r : I → R with r◦i = idR. In other words R is an H-retract of I. Therefore, R is H-injective by Lemma 1.8.
1.2 isoM-Injectivity
Throughout this thesis Metr1 will denote the category of metric spaces and 1-Lipschitz maps, i.e. objects are metric spaces (X,dX ) and morphisms are maps f : (X,dX ) → (Y,dY ) such that
′ ′ dY (f(x), f(x )) ≤ dX (x, x ) for all x, x′ ∈ X. isoM will denote the class of all isometric embeddings, which we shall use as the distinguished class of morphisms H from Section 1.1 for the category Metr1. For convenience we shall write injective instead of isoM-injective unless otherwise stated. isoM-Injectivity 9
Injective objects in Metr1 can be characterized by the ball intersection property as Proposition 1.13 will show. This property is crucial in under- standing the idea of Isbell’s construction, [22], of the injective hull of a metric space X, which we shall introduce in Section 1.4.
1.13 Proposition (Ball Intersection Property). Let X =(X,d) be a metric space. The following three statements are equivalent:
(i) X is an injective object in Metr1. (ii) Whenever f : A → X is a 1-Lipschitz map defined on a subset A of a metric space Y , then there exists a 1-Lipschitz extension f¯: Y → X of f.
(iii) Whenever {B(xi,ri)}i∈I is a nonempty family of closed balls in X with the property that ri +rj ≥ d(xi, xj) for all pairs of indices i, j ∈ I, then
the intersection i∈I B(xi,ri) is nonempty. Proof. The equivalenceT of (i) and (ii) is just a reformulation of Definition 1.1 where C is Metr1 and H is isoM. We first show (ii) ⇒ (iii). Let A denote the subset {xi}i∈I ⊂ X endowed with the metric inherited from X. Let Y = A ∐{y} be the extension of A by one more point with distance d(y, xi) := ri. By assumption d(y, xi)+ d(y, xj)= ri + rj ≥ d(xi, xj) for all i, j ∈ I, thus d is a metric on Y . Because X satisfies condition (ii) the inclusion map i : A ֒→ X can be extended to a 1-Lipschitz map ¯i : Y → X. Since
d (¯i(y), xi) ≤ d(y, xi)= ri ⇒ ¯i(y) ∈ B(xi,ri), ∀i ∈ I we see that i∈I B(xi,ri) = ∅. We proveT (iii) ⇒ (ii). Using Zorn’s lemma one can see that in order to prove (ii) it is enough to verify that every 1-Lipschitz map f : A → X can be extended by one more point y ∈ Y . In other words the domain of f¯ is the metric subspace Y¯ := A ∐{y} ⊂ Y where A is isometrically embedded. Therefore, we only need to define f¯(y) in a proper way such that f¯ remains 1-Lipschitz. Let ri := dY (xi,y). Since
ri + rj = dY (xi,y)+ dY (xj,y) ≥ dY (xi, xj) ≥ dX (f(xi), f(xj)) and condition (iii) holds, we know that the intersection F := ¯ i∈I B (f(xi),ri) is nonempty. So choose a point x ∈ F and set f(y) := x. TBy construction we then have ¯ dX f(y), f(xi) = dX (x, f(xi)) ≤ ri = dY (y, xi) and therefore f¯ is 1-Lipschitz. 10 Chapter 1. Isbell’s Injective Hull
1.14 Remark. There is also an intrinsic characterization of isoM-injective objects, among metric spaces with finite Assouad-Nagata dimension, in the much “larger” category M of all metric spaces with arbitrary Lipschitz maps as morphisms. See [24] for more details. In the literature, injective metric spaces are also known as hyperconvex spaces (cf. [2]) or tight spans (cf. [19], [20], [17], [11], [10]). As a first example we will show that the real line is injective. Since every geodesic line in a metric space X is the isometric image of the real line, it follows that every geodesic in X is injective. 1.15 Corollary. The real line R with its usual metric induced by the eu- clidean norm, d(x, y)= |x − y|, is injective in Metr1.
Proof. Let {B(xi,ri)}i∈I be a nonempty family of closed balls in R such that xi − xj ≤|xi − xj| ≤ ri + rj. Hence
∀i, j ∈ I : xi − ri ≤ xj + rj ⇒ ∃x ∈ R : sup(xi − ri) ≤ x ≤ inf(xj + rj). i∈I j∈I
Therefore, xi − ri ≤ x ≤ xi + ri for all i ∈ I, which implies x ∈
i∈I B(xi,ri) = ∅. By the ball intersection property 1.13(iii) R must be injective.T We will list some properties of injective metric spaces in general. Recall that a metric space X is said to be geodesic if every two points in X can be joined by a geodesic segment. 1.16 Proposition. An injective metric space X is complete and geodesic. -Proof. Let X¯ denote the metric completion of X and i : X ֒→ X¯ the inclu ¯ sion map. Since X is injective there is an extension r := idX : X¯ → X of the identity map idX : X → X with r ◦ i = idX. Let (xn) be a Cauchy sequence in X and setx ¯n := i(xn). Because X¯ is complete (¯xn) converges to a point x¯ ∈ X¯. Therefore, we get that
d r(¯x), xn = d r(¯x),r ◦ i(xn) ≤ d x,¯ i(xn) = d(¯x, x¯n) → 0. Hence X is complete. On the other hand let x, y ∈ X be two arbitrary points and let A := {0,D} ⊂ [0,D] =: ID ⊂ R with D := d(x, y). Define a map f : A → X by setting f(0) := x and f(D) := y. The map f is an isometric embedding because d f(0), f(D) = d(x, y) = D = |0 − D|. In particular f is 1- ¯ Lipschitz. Therefore, there is an extension f : ID → X of f. Let s < t ∈ ID. D = d f¯(0), f¯(D) ≤ d f¯(0), f¯(s) + d f¯(s), f¯(t) + d f¯(t), f¯(D) (f¯ is 1-Lipschitz) ≤ s +(t − s)+( D − t)= D. isoM-Injectivity 11
Hence we have equality everywhere and d f¯(s), f¯(t) = t − s = |s − t|. 1.17 Proposition. Let (X,d) be an injective metric space. Every closed ball B(x, r) in X endowed with the inherited metric d is injective.
Proof. Let X′ := B(x, r). To distinguish balls in X′ and X we write B′ for balls in X′ instead of B. In other words B′(z,s)= B(z,s) ∩ X′. We have to verify the ball intersection property 1.13(iii) for X′. Let ′ ′ {B (xi,ri)}i∈I be a nonempty family of closed balls in X with d(xi, xj) ≤ ri + rj and let B := {B(xi,ri)}i∈I ∪ B(x, r). The family B satisfies the conditions of 1.13 (iii) because d(x, xi) ≤ r ≤ r + ri and d(xi, xj) ≤ ri + rj. Therefore, since X is injective and due to Proposition 1.13, the intersection of B is non-empty. And we get
′ B (xi,ri)= B(xi,ri) ∩ B(x, r) = B(xi,ri) ∩ B(x, r) = ∅. \i∈I \i∈I \i∈I Hence B(x, r) is injective.
1.18 Lemma. Let X be a proper metric space, that is every closed, bounded subset of X is compact, and let {An}n∈N be an exhausting, increasing family of subsets of X, i.e.
An ⊂ An+1 and An = X. n[∈N
Then X is injective if all the An’s are injective.
Proof. Let B := {B(xi,ri)}i∈I be a family of closed balls in X as in 1.13 (iii). By Proposition 1.13 we need to verify that the intersection of B is non-empty. n Let B := {B(xi,ri)∩An | xi ∈ An, i ∈ I}. Since X = n An and An ⊂ An+1 for every i ∈ I there is an Ni ∈ N such that xi ∈ ASn for all n ≥ Ni. In n particular B(xi,ri) ∩ An ∈ B for all n ≥ Ni. Let j ∈ I be arbitrary but fixed. W.l.o.g. we may assume that xj ∈ An for all n, otherwise we take ′ n An := An ∪ ANj . As An is injective, the intersection of the family B is non- empty. Let an ∈ B∈Bn B. Since an ∈ B(xj,rj) for all n and X is proper, there is a convergentT subsequence ank → a. Note that ank ∈ B(xi,ri) for all nk ≥ Ni. Therefore, a ∈ B(xi,ri) for all i ∈ I and the claim follows.
1.19 Corollary. A proper metric space X is injective if and only if every closed ball in X is injective.
Proof. This is a direct consequence of Proposition 1.17 and Lemma 1.18. 12 Chapter 1. Isbell’s Injective Hull
1.20 Definition. Let (Xi, xi) i∈I be a family of pointed metric spaces. The ℓ∞-product is the metric space given by
(Xi, xi) := f ∈ Xi sup dXi (fi, xi) < ∞ , i∈I YI YI c ∞ endowed with the metric d (f,g) := supi∈I dXi (fi,gi). In the case where ∞ (Xi, xi)=(R, 0) we just write l (I) for I (R, 0). This is the Banach space of all bounded functions on the set I. Qb
The metric space (Xi, xi) depends on the choice of the family {xi}i∈I . Note that from the categoricalQb point of view satisfies the universal prop- erty of the product in the category of pointed metric spaces with 1-Lipschitz Qb maps. The following proposition shows that injectivity is preserved under prod- ucts from Definition 1.20.
∞ ∞ 1.21 Proposition. The ℓ -product I (Xi, xi),d is injective in Metr1 whenever the X ’s are. In particular the Banach space ℓ∞(I) is injective. i Qb
Proof. Every closed ball in (Xi, xi) of radius r and with center {mi}i∈I is the product of the balls B(m ,r) ⊂ X . The claim follows from the ball Qb i i intersection property 1.13 in Xi. The real line R endowed with the euclidean metric is injective by 1.15. Therefore, the Banach space ℓ∞(I) is injective.
For our purposes ℓ∞(I) is an important class of examples of injective met- ric spaces. We shall show that every metric space X can be isometrically embedded into ℓ∞(I) for I = X. This is Kuratowski’s embedding theo- rem, which we prove next. In particular if X is injective the Banach space structure of ℓ∞(X) leads to good properties of X itself; see Section 1.3.
1.22 Theorem (Kuratowski Embedding Theorem). Every metric space (X,d) can be isometrically embedded into the Banach space ℓ∞(X) by means ∞ of ix0 : X ֒→ ℓ (X) given by
ix0 (x) := d(x, ) − d(x0, ), where x0 ∈ X.
The Kuratowski embedding ix0 is not canonical since it depends on the choice of the base point x0. Bicombing and isoperimetric inequality 13
Proof. We have to show that ix0 (x) is a bounded function on X.
sup | ix0 (x) (y)| = sup |d(x, y) − d(x0,y)| ≤ d(x, x0) < ∞. y∈X y∈X ∞ Hence ix0 (x) ∈ ℓ (X). Furthermore
∞ ′ ′ ′ d ix0 (x), ix0 (x ) = sup |d(x, y) − d(x ,y)| ≤ d(x, x ). y∈X ′ ′ ′ ′ ′ On the other hand supy∈X |d(x, y)−d(x ,y)|≥|d(x, x )−d(x , x )| = d(x, x ).
This implies that ix0 is an isometric embedding.
Theorem 1.22 says that the category Metr1 has enough injectives, i.e. every metric space can be isometrically embedded into an injective one.
1.23 Definition. A metric space X is called an absolute 1-Lipschitz retract, or 1-ALR, if X is an absolute isoM-retract in Metr1 that is if every iso- metric embedding of X into another metric space Y is a 1-Lipschitz retract of Y .
1.24 Proposition. X is injective in Metr1 if and only if X is a 1-ALR. Proof. This is follows directly from Lemma 1.12.
1.3 Bicombing and isoperimetric inequality
In the following let I = [0, 1] ⊂ R. A bicombing on a metric space X is a choice of a, not necessarily continuous, path between any pair of points of X. The particular choice depends on the problem one is dealing with. In this thesis, for example, we use bicombings by quasi-geodesics (cf. Defini- tion 2.24). Another important class are convex bicombings, see Definition 1.25 be- low. Metric spaces admitting a convex bicombing share some properties with normed vector spaces. They are for instance 1-Lipschitz contractible and a metric version of the Theorem of Krein–Mil’man holds (see [6]). Examples of such metric spaces are normed spaces, CAT(0) spaces (cf. [5]) and in fact all injective metric spaces. The latter is shown in Lemma 1.27.
1.25 Definition. A convex bicombing on a metric space X is a map
c : X × X → C(I,X) with the following properties. 14 Chapter 1. Isbell’s Injective Hull
(i) The path cxy := c(x, y): I → X is a geodesic segment from x to y parametrized with constant speed v = d(x, y), i.e.
′ ′ ′ d cxy(t),cxy(t ) = |t − t | d(x, y), ∀t, t ∈ I. (1.26) (ii) For every four points x, x′,y,y′ ∈ X the function D : I → R,
D(t) := d cxy(t),cx′y′ (t) , is a convex function on I.
1.27 Lemma. Every injective metric space X admits a convex bicombing. Proof. The idea of the proof is to first embed X isometrically into the Banach space l∞(X) and then push forward the natural convex bicombing on ℓ∞(X), as a normed vector space, onto X using a retraction. Here are the details. We can suppose that X is a subspace of ℓ∞(X). Otherwise we take a base ∞ point x0 ∈ X and embed X into ℓ (X) using the Kuratowski embedding ix0 from Theorem 1.22. Since X is injective there is by Proposition 1.24 a 1-Lipschitz retraction r : ℓ∞(X) → X. Define the bicombing on X by
cxy(t) := r (1 − t) x + ty , for all x, y ∈ X. We have to check that cxy is a geodesic and that 1.25(ii) holds for cxy. Let s, t ∈ I with t>s. It follows that
d(x, y) ≤ d x, cxy(s) + d cxy(s),cxy(t) + d cxy(t),y
(r is 1-lipschitz) ≤ s x − y ∞+(t − s) x − y ∞+ (1 − t) x − y ∞
= x − y ∞ = d(x, y). Therefore, we have equality and in particular
d cxy(s),cxy(t) = |t − s| x − y ∞, for all s, t ∈ I. Hence cxy is geodesic parametrized with constant speed v = x − y ∞. Furthermore
′ ′ d cxy(t),cx′y′ (t) = r (1 − t) x + ty − r (1 − t) x + ty ∞ ′ ′ (r is 1-Lipschitz) ≤ (1 − t) x + ty − (1 − t) x + ty ∞ ′ ′ = (1 − t)(x − x)+ t( y − y) ∞ ′ ′ ≤ (1 − t) x − x ∞ + t y − y ∞ = (1 − t) d(x, x′)+ td(y,y′), which proves that c is a convex bicombing. Isbell’s injective hull 15
1.28 Corollary. Every injective metric space X is contractible.
Proof. Choose a base pointx ¯ ∈ X. By Lemma 1.27 X admits a convex bicombing c: X × X → C(I,X). Let H(x, t) := cxx¯ (t). H is a contraction of X ontox ¯. We have
′ ′ ′ d H(x, t), H(x , t ) = d cxx¯ (t),cxx¯ ′ (t ) ′ ≤ d cxx¯ (t),cxx¯ ′ (t) + d cxx¯ ′ (t),cxx¯ ′ (t ) (c is a convex bicombing) ≤ td (x, x′)+ |t − t′| d(¯x, x′).
Hence H: X × I → X is a continuous map. Furthermore we have by defini- tion that H(x, 0)=x ¯ and H(x, 1) = x.
In [26] S. Wenger has shown that every complete metric space admitting a convex bicombing has an isoperimetric inequality of euclidean type in the setting of integral metric currents. It would lead us too far astray to give the precise definitions and state the theorem here. But the interested reader is encouraged to read [26] for the details and proofs. Since every injective metric space admits a convex bicombing and is complete we get the following corollary.
1.29 Corollary. Every injective metric space X admits an isoperimetric inequality of euclidean type in the class of integral metric currents.
Proof. See [26].
1.4 Isbell’s injective hull
The aim of this section is to show that every metric space X has an isoM- injective hull in Metr1. The injective hull, denoted by EX, can be described explicitly in terms of a function space. This construction goes back to Isbell [22]. We are going to recall the details and give the proofs.
1.30 Definition. Let (E, ) be the poset of all functions g : X → R satis- fying g(x)+ g(y) ≥ d(x, y) ∀x, y ∈ X, (1.31) where f g :⇔ f(x) ≥ g(x) for all x ∈ X. We call a function f : X → R extremal if f is a minimal element of the poset (E, ). In other words whenever a function g : X → R is given, satisfying condition (1.31) and with f g, we have g = f. 16 Chapter 1. Isbell’s Injective Hull
1.32 Remark. Let f ′ : X → R be a function that satisfies condition (1.31) ′ and let Ef ′ := {g ∈ E | f g} ⊂ E. Every linearly ordered subset of the poset Ef ′ has a minimum in Ef ′ - take the infimum of the subset for instance. Therefore, by Zorn’s lemma we obtain a minimal function in Ef ′ , which is by definition an extremal function on X. Hence for every function f ′ ∈ E there is an extremal function f with f ′ f. 1.33 Proposition. Every extremal function is non-negative and 1-Lipschitz. Proof. Let f be extremal. By definition we have 2f(x) = f(x)+ f(x) ≥ d(x, x) = 0. Hence f is non-negative. Furthermore let x, y ∈ X and define g : X → R by setting g(y) := f(x)+ d(x, y) and g(x′) := f(x′) for all x′ ∈ X \{y}. Note that either f g or g f, since g and f coincide everywhere except for y. Moreover g satisfies condition (1.31) because g(x′)+ g(y) = f(x′)+ f(x)+ d(x, y) ≥ d(x′, x)+ d(x, y) ≥ d(x′,y). Hence by the minimality of f we obtain that f g and in particular f(y) ≤ g(y)= f(x)+ d(x, y). Reversing the roles of x and y we get |f(x) − f(y)| ≤ d(x, y) for all x, y ∈ X. (1.34) Therefore, f is 1-Lipschitz. In the following we give an alternative characterization of extremal func- tions on X. 1.35 Proposition. A function f : X → R is extremal if and only if f sat- isfies condition (1.31) and for every x ∈ X inf f(x)+ f(y) − d(x, y) =0. (1.36) y∈X
Proof. Let f be an extremal function and suppose that there exists an x ∈ X such that f(x)+ f(y) − d(x, y) ≥ 2ε > 0 for all y ∈ X. Define g : X → R by setting g(x) := f(x) − ε ≥ 0 and g(y) := f(y) for y ∈ X \{x}. The function g satisfies condition (1.31), because by assumption g(x)+ g(y) = f(x) − ε + f(y) ≥ d(x, y) for all y ∈ X \{x}. But g f, which contradicts the minimality of f. On the other hand let f be a function satisfying condition (1.31) and (1.36). By Remark 1.32 there is an extremal function g with f g. Suppose that g = f, i.e. g(x)+ ε ≤ f(x) for some x ∈ X and some ε> 0. Since g is extremal the above implies that 0 = inf f(x)+ f(y) − d(x, y) ≥ inf g(x)+ ε + g(y) − d(x, y) = ε> 0. y∈X y∈X Isbell’s injective hull 17
The following two estimates will be needed later in this section.
1.37 Lemma. Let f1, f2 be two extremal functions on X. Then the following two inequalities hold for all x, y ∈ X:
|d(x, y) − f1(x)| ≤ f1(y) (1.38) and |f1(x) − f2(x)| ≤ f1(y)+ f2(y). (1.39) Proof. This is a direct consequence of (1.34) and (1.31). 1.40 Definition (Injective Hull). Let X be a metric space. Then EX de- notes the set of all extremal functions on X equipped with the metric
d¯(f1, f2) := sup |f1(x) − f2(x)|. x∈X d¯ is well-defined since the supremum is finite by inequality (1.39). ¯ We shall show that (EX, d) is the isoM-injective hull of (X,d) in Metr1. This is done in three steps. First of all we construct in Lemma 1.41 an isometric embedding i: X ֒→ EX. Then we verify that EX is injective and that the embedding i is isoM-essential, see Theorem 1.45. 1.41 Lemma. Every metric space X can be isometrically embedded into EX by z ∈ X → iz ∈ EX where iz : X → R denotes the distance function
iz(y) := d(z,y).
Furthermore, for all f ∈ EX the following identity holds:
d¯ f, iz = f(z). (1.42)
Proof. Let z ∈ X. By the triangle inequality it follows that iz(x)+ iz(y)= d(z, x)+ d(z,y) ≥ d(x, y). Therefore, iz satisfies condition (1.31). In order to show that iz is extremal we need to verify that iz is minimal. Using Remark 1.32 we get an extremal function f ∈ EX with the property that f iz. Then f(z) = 0 because 0 ≤ f(z) ≤ iz(z) = d(z, z) = 0. Hence iz(x)= d(z, x) ≤ f(x)+ f(z)= f(x) ≤ iz(x) for all x ∈ X. Since we have
d¯(iy,iz) d(y, z)= |d(y,y) − d(y, z)| ≤ sup |d(x, y) − d(x, z)| ≤ d(y, z), xz∈X }| { 18 Chapter 1. Isbell’s Injective Hull it follows that i is an isometric embedding of X into EX. On the other hand we get by (1.38)
d¯(iz,f) f(z)= |d(z, z) − f(z)| ≤ sup |d(x, z) − f(x)| ≤ f(z) xz∈X }| { for all f ∈ EX.
The following lemma will be needed for the main theorem of this section, Theorem 1.45, to prove that i: X ֒→ EX is in fact the injective hull of X in Metr1.
Lemma. Let i: X ֒→ EX be the isometric embedding from 1.43 Lemma 1.41. Then the following two statements are true.
(i) If s: EX → R is an extremal function on EX then s ◦ i: X → R is an extremal function on X.
(ii) If F : EX → EX is a 1-Lipschitz map that leaves i(X) pointwise fixed then F is the identity on EX.
Proof. (i) Since s satisfies (1.31) and i is an isometric embedding, we get that s ◦ i(x)+ s ◦ i(y) = s(ix)+ s(iy) ≥ d¯(ix, iy) = d(x, y) for all x, y ∈ X. Thus s ◦ i fulfills condition (1.31). It remains to show that s ◦ i is minimal. Let g : X → R be a function on X with g ≤ s ◦ i and satisfying (1.31), and let x ∈ X. Define t: EX → R by setting t(ix) := g(x) and t(f) := s(f) for all f ∈ EX \{ix}. If suffices to prove that
t(ix)+ t(f)= g(x)+ s(f) > f(x) − ǫ (1.44) for all f ∈ EX \{ix} and ǫ> 0. Because by (1.42) we have f(x) = d¯(f, ix) and therefore (1.44) implies that t satisfies (1.31). By the minimality of s it follows that s ◦ i(x) ≤ t(ix) = g(x). Since x was arbitrary this proves that s ◦ i is minimal. Let f ∈ EX \{ix} and ǫ> 0. By Proposition 1.35 there is a y ∈ X such that f(x)+ f(y) g(x)+ s(iy) ≥ g(x)+ g(y) ≥ d(x, y) > f(x)+ f(y) − ǫ. On the other hand by (1.38) and (1.42), we get s(f) − s(iy) ≥ −d¯(iy, f)= −f(y). Isbell’s injective hull 19 Adding these two estimates we obtain g(x)+ s(f) > f(x) − ǫ. (ii) Let f ∈ EX and g := F (f) ∈ EX. Since F : EX → EX is a 1-Lipschitz map and because of (1.42) we get that g(x)= d¯(ix,g)= d¯ F (ix), F (f) ≤ d¯(ix, f)= f(x) for all x ∈ X. By minimality of f it follows F (f)= g = f. Hence F is the identity. Theorem. The metric space (EX, d¯) is injective and i: X ֒→ EX is 1.45 an isoM-essential map. Therefore, EX is the isoM-injective hull of X in Metr1. Proof. To prove the injectivity of EX we verify the ball intersection prop- erty 1.13(iii). Let B(fi,ri) i∈I be a nonempty family of closed balls in (EX, d¯) such that ri + rj ≥ d¯(fi, fj) for all i, j ∈ I. Let r : EX → R be the function given by r(f) := inf ri + d¯(fi, f) . i∈I In particular for f = fi we have that r(fi) = ri. Note that r satisfies ′ ¯ ¯ ¯ ′ condition (1.31) since r(f)+r(f ) ≥ infi,j∈I d(fi, fj)+d(fi, f)+d(fj, f ) ≥ d¯(f, f ′) for all f, f ′ ∈ EX. Therefore, by Remark 1.32 there is an extremal function s: EX → R with s ≤ r. Using Lemma 1.43 we conclude that s ◦ i is an extremal function on X. Let x ∈ X. By (1.42) and (1.38) it follows ¯ |f(x) − s ◦ i(x)| = |d(ix, f) − s(ix)| ≤ s(f) ≤ r(f). ¯ E Hence d(f,s ◦ i) = supx∈X |f(x) − s ◦ i(x)| ≤ r(f) for all f ∈ X. In other words s ◦ i is a point of intersection of the balls B(fi,ri) i∈I since s ◦ i ∈ B f,r(f) ⊂ B(fi,ri). f∈\EX \i∈I On the other hand let H and G be two maps given as in the following commutative diagram i X / EX D U) DD DD H ¯i G DD D! Y with G ∈ isoM. In order to prove that i is isoM-essential we need to verify that H is an isometric embedding. Because EX is injective and G ∈ isoM 20 Chapter 1. Isbell’s Injective Hull there is an extension ¯i: Y → EX of the map i. Let F := ¯i ◦ H : EX → EX. Then for all x ∈ X we have F i(x) = (¯i ◦ H) ◦ i (x)= ¯i ◦ (H ◦ i)(x)= ¯i ◦ G(x)= i(x). Therefore, F fixes i(X) pointwise and by Lemma 1.43 (ii). it follows that ¯i ◦ H = F = idEX . Now H is an isometric embedding because d H(f),H(f ′) ≥ d¯ ¯i(H(f)),¯i(H(f ′)) = d¯(f, f ′) ≥ d H(f),H(f ′) for all f, f ′ ∈ EX. The following corollary says that EX is the smallest injective metric space in which X can be isometrically embedded. 1.46 Corollary. Let j : X → Y be an isometric embedding of X into some injective metric space Y . Then every extension ¯j : EX → Y of j i.e. ¯j◦i = j, where i: X → EX, is an isometric embedding. Proof. Note that an extension ¯j : EX → Y of j : X → Y always exists by the ,injectivity of Y . Since ¯j ◦ i = j ∈ isoM and i: X ֒→ EX is isoM-essential cf. 1.45, it follows that ¯j ∈ isoM. 1.47 Lemma. Let j : X → Y and h: Y → EX be isometric embeddings such that h ◦ j = i, where i: X → EX. Then EY and EX are isometric. Proof. Consider the commuting diagram i X / EX (1.48) D O bE DD EE h¯ DD h EE j DD EE D! ? E Y / EY iY where iY : Y → EY is the isometric embedding from Lemma 1.41 of Y into its injective hull and h¯ is an extension of h. Note that h¯ exists since EX is injective and iY ∈ isoM; it is an isometric embedding due to Corollary 1.46. ¯ ¯ ¯ Moreover, i(X) ⊂ h(EY ) because i = h◦iY ◦j. We claim that h(EY )= EX. Let r : EX → h¯(EY ) ⊂ EX be a 1-Lipschitz retraction of EX onto h¯(EY ), which by Lemma 1.10 exists since the latter is injective. Then r fixes the subset i(X) ⊂ h¯(EY ) pointwise. Therefore, r is the identity on EX by 1.43; in particular EX = r(EX)= h¯(EY ). Gromov–Hausdorff convergence 21 1.49 Remark. The isometric embedding h: Y → EX from 1.47 is uniquely determined by j and the condition h ◦ j = i since h(y)(x)= d¯ h(y), i(x) = d¯ h(y), h j(x) = dY y, j(x) . ¯−1 ¯−1 It follows in particular that h = iY because h ◦ h = iY and iY ◦ h = iY ; compare diagram (1.48). 1.50 Corollary. If i(X) ⊂ Y ⊂ EX then EY is isometric to EX. Proof. The claim follows directly from Lemma 1.47 with h: Y ֒→ EX the inclusion map and j = i: X → Y ⊂ EX. 1.5 Gromov–Hausdorff convergence In Section 1.4 we saw that every metric space X has an injective hull EX and that the latter can be given explicitly. In this section we are going to show that the assignment X → EX is continuous with respect to the Gromov–Hausdorff pseudo-distance dGH in the sense that E E Xn →dGH X ⇒ Xn →dGH X. (1.51) We give the definition of the Gromov–Hausdorff pseudo-distance dGH (X,Y ) of two metric spaces X and Y . There are several equivalent definitions, cf. [5], but the following one is appropriate for our purposes. Let NZ (X, R) := {z ∈ Z | ∃x ∈ X : d(x, z) ≤ R}. (1.52) Define the Gromov–Hausdorff pseudo-distance between X and Y to be Z dGH (X,Y ) := inf{ dH(X,Y ) | X,Y ⊂ Z isometrically embedded}, (1.53) Z where Z dH (X,Y ) := inf{R> 0 |X ⊂ NZ (Y, R) and Y ⊂ NZ(X, R)} denotes the Hausdorff distance between X and Y as subspaces of Z. However we can assume that the metric space Z in the definition (1.53) is injective. Otherwise we can embed Z isometrically into an injective metric space by Theorem 1.22. In order to prove (1.51) we need the following technical lemma. 22 Chapter 1. Isbell’s Injective Hull Y 1.54 Lemma. Let X ⊂ Y be two metric spaces such that dH (X,Y ) ≤ ε, and let iX : X ֒→ EX and iY : Y ֒→ EY be defined as in Lemma 1.41. Then for every isometric embedding i : EX → EY with i ◦ iX = iY |X we have EY E E dH i( X), Y ≤ 4ε. Note that the condition i ◦ iX = iY |X is satisfied whenever i is an extension . of the map iY |X : X ֒→ EY Proof. Let f ∈ EY . We shall show that there exists a gf ∈ i(EX) with d¯EY (f,gf ) ≤ 4ε. Y By assumption dH (X,Y ) ≤ ε. Therefore for each y ∈ Y there is a xy ∈ X with d(xy,y) ≤ ε. Consider the restriction f|X : X → R. Then f|X satisfies condition (1.31) since X ⊂ Y . By Remark 1.32 there is an extremal function g ∈ EX with g f|X . We show by contradiction that |g(x) − f|X (x)| ≤ 2ε must hold for all x ∈ X. Suppose that g(x) < f|X (x) − 2ε for some point x ∈ X. Define f ′ : Y → R by setting f ′(x) := g(x)+2ε and f ′(y) := f(y) otherwise. Then the function f ′ fulfills condition (1.31) since f ′(x)+ f ′(y) = g(x)+2ε + f(y) (f is 1-lipschitz and d(xy,y) ≤ ε) ≥ g(x)+ ε + f(xy) (f(xy) ≥ g(xy) and g ∈ EX) ≥ d(x, xy)+ ε (d(xy,y) ≤ ε) ≥ d(x, y). But this is a contradiction to the minimality of f because f ′(x) = g(x)+ 2ε < f(x). Therefore we have that |f(x) − g(x)| ≤ 2ε for all x ∈ X. Let gf := i(g) ∈ EY . Then gf (x)= d¯EY (gf , iY (x)) = d¯EY (i(g), i ◦ iX (x)) = d¯EX (g, iX(x)) = g(x) for all x ∈ X. Hence d¯EY (f,gf ) = sup |f(y) − gf (y)| y∈Y (f and gf are 1-Lipschitz) ≤ sup |f(xy) − gf (xy)| +2ε y∈Y ≤ sup |f(x) − gf (x)| +2ε x∈X = sup |f(x) − g(x)| +2ε ≤ 4ε. x∈X Gromov–Hausdorff convergence 23 1.55 Theorem. Let X and Y be two metric spaces. Then, dGH (EX, EY ) ≤ 8dGH (X,Y ). In particular the assignment X → EX is continuous with respect to dGH in the sense of (1.51). Proof. Let R>dGH (X,Y ). Then by the definition of dGH there is a metric Z space Z with X,Y ⊂ Z and such that dH (X,Y ) ≤ R, i.e. X ⊂ NZ (Y, R) and Y ⊂ NZ (X, R). In particular Y ⊂ NZ(X, R) ⊂ NZ (Y, 2R). Therefore we get that Z Z dH Y, NZ(X, R) ≤ 2R, dH X, NZ (X, R) ≤ R. Let N := NZ (X, R) and let IX := iN |X : EX → EN denote the extension of .iN |X : X ֒→ EN where iN : N ֒→ EN is the embedding given by Lemma 1.41 Compare the commutative diagram. ⊆ iN X / N / EN (1.56) m6 m m iX m m mI :=i | m m X N X EX By Corollary 1.46 it follows that IX is an isometric embedding. Since Z dH (X, N) ≤ R we get by Lemma 1.54 that EN E E dH IX ( X), N ≤ 4R. In the same way we define IY := iN |Y : EY → EN and conclude that EN E E dH IY ( Y ), N ≤ 8R. Therefore dGH (EX, EY ) ≤ 8R whenever dGH (X,Y ) < R and the claim follows. 1.57 Remark. The assignment X → EX is not an isometry. For example n n n n n consider Z ⊂ R1 where R1 := (R ,ℓ1) and Z is endowed with the induced n n 1 ℓ1-metric. Then it is easy to see that dGH (Z , R1 ) ≥ 2 > 0. But on the E n E n E n ∼ E n other hand dGH ( Z , R1 ) = 0 since Z = R1 ; see Example 1.99. 1.58 Corollary. Let Y be injective and X an arbitrary metric space with dGH (Y,X)=0. Then iX (X) ⊂ EX is dense, where iX : X ֒→ EX is the isometric embedding from Lemma 1.41. If in addition X is complete, then X is injective. 24 Chapter 1. Isbell’s Injective Hull B Proof. First of all note that A ⊂ B is dense if and only if dH (A, B) = 0. This follows easily from the definition of the Hausdorff distance, see (1.53). EX E Thus it is enough to verify that dH iX (X), X = 0. ∼ Since dGH (Y,X) = 0 and Y is injective, i.e. Y= EY , Theorem 1.55 gives us the following estimate 0 ≤ dGH (Y, EX)= dGH (EY, EX) ≤ 8dGH (Y,X)=0. Hence dGH (Y, EX) = 0. We then obtain by the triangle inequality that dGH (X, EX) = 0, because 0 ≤ dGH (X, EX) ≤ dGH(X,Y )+ dGH (Y, EX)=0. In other words, for every ε> 0 there is an injective metric space Z = Z(ε) in E Z E which X and X are isometrically embedded and such that dH (X, X) ≤ ε. W.l.o.g. we may assume that X, EX ⊂ Z. In order to distinguish the isometric image of EX in Z and EX itself, we shall write EZ X for the former. We have EZX ⊂ NZ(X,ε) and X ⊂ NZ (EZX,ε), thus X ⊂ NZ (EZX,ε) ⊂ NZ (X, 2ε). Let N := NZ (EZ X,ε). W.l.o.g. we may fur- ther assume that EZX ⊂ EN ⊂ Z. Otherwise we can embed EN isometrically into Z, using Corollary 1.46, and get N ⊂ EN ⊂ Z. Notice that the inclusion map i : EZX ֒→ EN obviously Z E E satisfies the conditions of Lemma 1.54 and so dH ( Z X, N) ≤ 4ε, since Z E E E dH ( Z X, N) ≤ ε. In particular we get N ⊂ NZ ( ZX, 4ε). On the other hand EZ X ⊂ NZ (X,ε) and hence EN ⊂ NZ (X, 5ε). Moreover, because X ⊂ N ⊂ EN ⊂ Z, there exists by corollary 1.46 an isometric embedding j : EX → EN such that j ◦ iX = idX . We conclude that for all ε> 0 X ⊂ j(EX) ⊂ EN ⊂ NZ (X, 5ε), or rather, since j ◦ iX = idX , iX (X) ⊂ EX ⊂ NEX iX (X), 5ε . EX E In other words dH iX (X), X = 0. As a Corollary we get the following. The completion of the Gromov– Hausdorff limit of a sequence of injective metric spaces is injective. Gromov–Hausdorff convergence 25 1.59 Corollary. Let {Xn}n∈N be a sequence of injective metric spaces con- verging to a metric space X with respect to the Gromov–Hausdorff distance, i.e. Xn →dGH X. Then iX (X) lies dense in EX. If in addition X is complete, then X is injective. E E Proof. By Theorem 1.55 we know that Xn →dGH X. Since all Xn are in- E ∼ E E jective, i.e. Xn = Xn, we get that Xn →dGH X. Therefore dGH (X, X)= 0, because 0 ≤ dGH (X, EX) ≤ dGH (X,Xn)+ dGH (Xn, EX) → 0. The claim then follows using Corollary 1.58. 1.60 Remark. Recall that a sequence of pointed metric spaces (Xn, xn) is said to converge to (X, x) in the pointed Gromov–Hausdorff topology, and we write Xn →pGH X, if for every r > 0 there is a sequence of numbers rn → r such that B(xn,rn) →dGH B(x, r) (cf. [7] for a better definition; however the definition above is sufficient for our purposes). One may think that the assignment X → EX is also continuous with respect to this topology. But this is not true in general; take for instance 2 X = R2, the plane with the euclidean norm. Then the sequence of closed 2 balls Xn := B(0, n) ⊂ R2, n ∈ N, converges to X with respect to the pointed E 2 Gromov–Hausdorff topology. However R2 is an infinite dimensional Banach space and therefore not proper, see Theorem 1.103, whereas the completion of limpGH EXn must be a proper metric space due to Proposition 1.66 below and [5, I.5.44]. Anyhow the following is true: 1.61 Corollary. Let Xn be a family of injective metric spaces, such that (Xn, xn) converges to a proper metric space (X, x) in the pointed Gromov– Hausdorff topology for some xn ∈ Xn and x ∈ X. Then X is injective. Proof. This is an immediate consequence of Corollary 1.19, Theorem 1.55 and the definition of pointed Gromov–Hausdorff convergence from Re- mark 1.60. 1.62 Remark. If X is a proper metric space and Xn ⊂ X a family of injective subspaces such that B(x, R) ∩ Xn = ∅, n ∈ N, for some x ∈ X and R > 0; let xn ∈ B(x, R) ∩ Xn. Then by an application of [7, 7.3.8], 26 Chapter 1. Isbell’s Injective Hull Theorem of Blaschke, there is a subsequence of (Xn, xn) converging to a proper subspace (Y,y) of X in the Gromov–Hausdorff topology. Therefore, Y is injective by Corollary 1.61. 1.63 Corollary. Suppose X is a proper metric space and An ⊂ X, n ∈ N, an increasing family of injective subspaces of X such that An = X. n[∈N Then X is injective. Proof. By Remark 1.62 we know that there is a subsequence of (An) con- verging to a proper injective metric subspace Y ⊂ X. Since n∈N An ⊂ Y and n∈N An = X it follows that Y = X. Hence, X is injective.S S 1.6 Compact metric spaces For the sake of completeness we recall [22] regarding the injective hull of a compact metric space C. We shall show in Proposition 1.66 that the injective hull EC is compact. Furthermore, we are interested in the following question. What is the smallest closed subset EC ⊂ C, such that EEC and ∼ EC are isometric, i.e. EEC = EC? Is EC unique with this property? The uniqueness of EC may surprise, since it would follow that besides EC there is no other isometric copy of EC in C, compare Lemma 1.77. 1.64 Lemma. Let X be a metric space. Then diam(X) = diam(EX). Proof. Recall that diam(X) := supx,y∈X d(x, y). Since X can be isometri- cally embedded into EX we have diam(X) ≤ diam(EX). If diam(X) = ∞ then it follows that diam(EX) = ∞ and the claim follows. Therefore, as- sume D := diam(X) < ∞. Let f ∈ EX and x ∈ X, and let ε > 0. By Proposition 1.35 there is an y ∈ X such that f(x)+ f(y) ≤ d(x, y)+ ε. In particular f(x) ≤ f(x)+ f(y) ≤ d(x, y)+ ε ≤ D + ε, or rather f(x) ≤ D since ε is arbitrary. Let g ∈ EX. We may further assume w.l.o.g. that g(x) ≤ f(x). Then |f(x) − g(x)| = f(x) − g(x) ≤ f(x) ≤ D. Therefore, ¯ E d(f,g) = supx∈X |f(x) − g(x)| ≤ D. It follows diam( X) ≤ diam(X). This proves the assertion. 1.65 Corollary. The injective hull of a bounded metric space is bounded. Proof. This follows directly from Lemma 1.64. Compact metric spaces 27 1.66 Proposition. Let C be a compact metric space. Then EC is compact. 1 Proof. Let Cn be a maximal n -separated net in C. Because C is compact, Cn is finite. By Lemma 1.65 the injective hull ECn is a bounded and closed ∞ ∼ |Cn| subset of ℓ (Cn) = R . Thus ECn is compact for all n ∈ N. C 1 On the other hand by construction we have that dH (Cn,C) ≤ n . Hence the sequence of subsets Cn ⊂ C converges to C in the Gromov–Hausdorff topol- E E ogy. Therefore by Proposition 1.55 we obtain that Cn →dGH C. Since every complete Gromov–Hausdorff limit of a sequence of compact metric spaces is compact [5, 5.40], it follows that EC is compact. Proposition 1.66 can also be proved using Arzel`a-Ascoli (cf. [22]), since extremal functions are 1-Lipschitz and as such uniformly continuous. Before we can answer the first question, we need first to settle some notations and show a simple observation about extremal functions. 1.67 Definition. Let X be a metric space. By the defect of an extremal function f ∈ EX at x, y ∈ X, we mean the number ∆(f; x, y) := f(x)+ f(y)−d(x, y) ≥ 0. The defect ∆ : EX ×X ×X → R is a continuous function. Furthermore, we say that y lies between x andx ¯, when d(x, x¯)= d(x, y)+ d(y, x¯) or equivalently, when ∆(iy; x, x¯) = 0, where iy is the distance function defined in Lemma 1.41. We write x