Research Collection
Doctoral Thesis
Spaces with convex geodesic bicombings
Author(s): Descombes, Dominic
Publication Date: 2015
Permanent Link: https://doi.org/10.3929/ethz-a-010584573
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. 23109
Spaces with Convex Geodesic Bicombings
A thesis submitted to attain the degree of
Doctor of Sciences of ETH Zurich
presented by
Dominic Descombes
Master of Science ETH in Mathematics citizen of Lignières, NE and citizen of Italy
accepted on the recommendation of
Prof. Dr. Urs Lang, examiner Prof. Dr. Alexander Lytchak, co-examiner
2015
Life; full of loneliness, and misery, and suffering, and unhappiness — and it’s all over much too quickly. – Woody Allen
Abstract
In the geometry of CAT(0) or Busemann spaces every pair of geodesics, call them α and β, have convex distance; meaning d ◦ (α, β) is a convex function I → R provided the geodesics are parametrized proportional to arc length on the same interval I ⊂ R. Therefore, geodesics ought to be unique and thus even many normed spaces do not belong to these classes. We investigate spaces with non-unique geodesics where there exists a suitable selection of geodesics exposing the said (or a similar) convexity property; this structure will be called a bicombing. A rich class of such spaces arises naturally through the construction of the injective hull for arbitrary metric spaces or more generally as 1- Lipschitz retracts of normed spaces. In particular, the injective hulls of word hyperbolic groups admit bicombings. We are able to generalize a variety of classical results to this broader context of weak non-positive curvature including a flat torus theorem and an asymptotic rank theorem (an equivalence of several notions of embeddings of flats) — results first shown for manifolds of non-positive curvature. New tools are needed in this less restricted setting. We employ a well-known, elegant barycenter construction that will be the key to many proofs.
v
Zusammenfassung
In CAT(0)- oder Busemann-Räumen gilt für beliebige zwei Geodäten, nennen wir diese α und β, dass der Abstand d ◦ (α, β) eine konvexe Funktion I → R ist. Vorausgesetzt wird dabei nur, dass die beiden Geodäten auf demselben Intervall I definiert und pro- portional zur Bogenlänge parametrisiert sind. Geodäten sind daher eindeutig, was für viele interessante Räume eine zu restriktive Forderung darstellt. Wir untersuchen da- her Räume, in welchen wir eine Auswahl an Geodäten treffen können, so dass die obige Konvexitätsbedingung (oder eine ähnliche) zumindest für diese Teilmenge gilt; diese Struktur nennen wir ein bicombing. Diese Klasse umfasst nun auch alle normierten Räume und allgemeiner alle 1-Lipschitz-Retrakte solcher. Weitere Beispiele von Räumen mit bicombings entstehen auf natürliche Weise durch die Konstruktion injektiver Hüllen metrischer Räume und im Speziellen besitzen injektive Hüllen hyperbolischer Gruppen diese Struktur. Wir beweisen zentrale klassische Resultate in diesem allgemeineren Rah- men schwacher, nichtpositiver Krümmung. Darunter einen Satz über flache Tori und ein Theorem über den asymptotischen Rang (eine Äquivalenz über die Einbettbarkeit flacher Räume) — Resultate welche ursprünglich für Mannigfaltigkeiten nichtpositiver Krümmung gezeigt wurden. Manchmal führen die klassischen Beweispfade zum Ziel, häufig aber auch nicht. Unter den neuen Hilfsmitteln, die nun benötigt werden, befindet sich auch eine Schwerpunktkonstruktion, von welcher wir wiederholt und an kritischen Stellen Gebrauch machen.
vii
Acknowledgments
Most of all, I want to express my gratitude to my advisor Urs Lang. Without his patience, gentle guidance, and firm support through these years, I would not have made it. I could always stop by his office to discuss mathematics and always left it with new ideas and renewed motivation for doing my research. I also thank the Swiss National Science Foundation for supporting me financially. And of course, I am also indebted to Alexander Lytchak for agreeing to act as co-examiner. The time I spent at the department of mathematics would not have been so rewarding without the wonderful people of group 1 & 4. Thank you all! And in particular, I am grateful for the time and fun I had with my office mates Dante Bonolis, Maël Pavón and Michael Th. Rassias. Finally, I thank my family for the support during all of my studies and most of all my girlfriend Julia.
ix
Contents
1 Introduction ...... 1
2 Basic concepts ...... 6 2.1 Metric spaces ...... 6 2.2 Bicombings ...... 9 2.3 Group actions ...... 11
3 Improving bicombings ...... 12 3.1 From conical to convex bicombings ...... 12 3.2 Straight curves ...... 15 3.3 Combinatorial dimension ...... 18
4 Boundary at infinity ...... 22
5 Flat planes ...... 28 5.1 Flat strips and half-planes ...... 28 5.2 The Flat Plane Theorem ...... 32
6 Barycenters ...... 36 6.1 The construction ...... 36 6.2 Linear barycenters ...... 42 6.3 Integration ...... 44
7 Semi-simple isometries ...... 45 7.1 Minimal displacement and axes ...... 45 7.2 σ-Axes ...... 48
8 Flat tori ...... 52
9 Asymptotic rank ...... 57
A Injective metric spaces ...... 64 A.1 Basic definitions and injective hulls ...... 64 n A.2 A characterization of injective subsets in `∞ ...... 68
Bibliography ...... 74
xi
Chapter 1
Introduction
The geometry of spaces of global non-positive curvature is largely dominated by the convexity of the distance function. Thus a considerable part of the theory of CAT(0) spaces [Bal, BriH] carries over to Busemann spaces [Bus, Pap] (defined by the property that d ◦ (α, β) is convex for any pair of constant speed geodesics α, β parametrized on the same interval). However, this larger class of spaces has the defect of not being preserved under limit processes. For example, among normed real vector spaces, exactly those with strictly convex norm satisfy the Busemann property, and a sequence of such n norms on R , say, may converge to a non-strictly convex norm. This motivates the study of an even weaker notion of non-positive curvature that dispenses with the uniqueness of geodesics but retains the convexity condition for a suitable selection of geodesics (compare Section 10 in [Kle]). In any normed space, the affine segments t 7→ (1−t)x+ty (t ∈ [0, 1]) provide such a choice. In particular, the relaxed condition carries the potential for simultaneous generalizations of results for non-positively curved and Banach spaces. A selection of geodesics for some metric space (X, d) we call a geodesic bicombing σ and thereby mean a map σ : X × X × [0, 1] → X that singles out a constant speed geodesic σxy := σ(x, y, ·) from x to y (that is, σxy(0) = 0 0 0 x, σxy(1) = y, and d(σxy(t), σxy(t )) = |t − t |d(x, y) for all t, t ∈ [0, 1]) for every pair (x, y) ∈ X × X. Because we work exclusively with bicombings that are composed of geodesics, we suppress the term “geodesic” throughout; a bicombing is thus implicitly understood to be a geodesic bicombing. We are interested in bicombings satisfying one of the following two conditions, each of which implies that σ is continuous and, hence, X is contractible via X ×[0, 1] → X, (x, t) 7→ σox(1−t) for an arbitrary choice of basepoint o ∈ X. We call σ convex if
the function t 7→ d(σxy(t), σx0y0 (t)) is convex on [0, 1] (1.1) for all x, y, x0, y0 ∈ X, and we say that σ is conical if
0 0 d(σxy(t), σx0y0 (t)) ≤ (1 − t) d(x, x ) + t d(y, y ) for all t ∈ [0, 1] (1.2) 2 CHAPTER 1. INTRODUCTION and x, y, x0, y0 ∈ X. Obviously every convex bicombing is conical, but the reverse implica- tion does not hold in general (see Example 3.2). For this, in order to pass condition (1.2) to subsegments, one would need to know in addition that σ is consistent in the sense that σpq(λ) = σxy((1 − λ)s + λt) (1.3) whenever x, y ∈ X, 0 ≤ s ≤ t ≤ 1, p := σxy(s), q := σxy(t), and λ ∈ [0, 1]. Thus, every conical and consistent bicombing is convex. Since we always assume our bicombings to be conical at least (for requiring a space to admit a bicombing is nothing more than it being geodesic), we adopt the convention that bicombing shall mean conical (and geodesic) bicombing from now on. Consequently, a consistent bicombing is automatically convex. Furthermore, we will say that a bicombing σ is reversible if
σxy(t) = σyx(1 − t) for all t ∈ [0, 1] (1.4) and x, y ∈ X. Consistent and reversible bicombings have also been employed in [HitL, FoeL]. We arrive at the following hierarchy of global non-positive (but non-coarse) curvature conditions (A) ⇒ (B) ⇒ (C) ⇒ (D) ⇒ (E) for a geodesic metric space X:
(A) X is a CAT(0) space,
(B) X is a Busemann space,
(C) X admits a consistent bicombing,
(D) X admits a convex bicombing,
(E) X admits a (conical) bicombing.
Clearly, if X is uniquely geodesic, then (E) ⇒ (B). When X is a normed real vec- tor space, (A) holds if and only if the norm is induced by an inner product (compare Proposition II.1.14 in [BriH]), (B) holds if and only if the norm is strictly convex (see Proposition 8.1.6 in [Pap]), and (C) is always satisfied. In the general case, (C) is stable under limit operations, whereas (B) is not (see, for instance, Section 10 in [Kle] again). It is unclear whether (E) ⇒ (D) and (D) ⇒ (C) without further conditions. The goal of Chapter 3 is to establish these implications under suitable assumptions and to address questions of uniqueness. Our first result refers to (E) and (D).
Theorem 3.4. Let X be a proper metric space with a bicombing. Then X also admits a convex bicombing.
The idea of the proof is to resort to a discretized convexity condition and then to gradually decrease the parameter of discreteness by the “cat’s cradle” construction from [AleB]. The passage from (D) to (C) appears to be more subtle. We first observe that the linear segments t 7→ (1 − t)x + ty in an arbitrary normed space X may be characterized as the curves α: [0, 1] → X with the property that t 7→ d(z, α(t)) is convex on [0, 1] for every z ∈ X (see Theorem 3.8). We therefore term curves with this property 3 straight (in a metric space X). Since the geodesics of a convex bicombing are necessarily straight, the above observation shows that normed spaces have no such bicombing other than the linear one. By contrast, there are instances of non-linear straight geodesics in compact convex subspaces of normed spaces as well as multiple convex bicombings in compact metric spaces (see Examples 3.9 and 3.10). Nevertheless, we prove a strong uniqueness property of straight curves (Proposition 3.13) in spaces satisfying a certain finite dimensionality assumption, introduced by Dress in [Dre] and explained further below. This gives the following result regarding items (D) and (C).
Theorem 3.14. Let X be a metric space of finite combinatorial dimension in the sense of Dress, or with the property that every bounded subset has finite combinatorial dimension. Suppose that X possesses a convex bicombing σ. We then have that σ is consistent, reversible, and unique, that is, σ is the only convex bicombing on X.
Our interest in Theorems 3.4 and 3.14 comes from the fact that property (E) holds in particular for all injective metric spaces (equivalently absolute 1-Lipschitz retracts). X is injective if for every isometric inclusion A ⊂ B of metric spaces and every 1-Lipschitz map f : A → X there exists a 1-Lipschitz extension f : B → X of f. Basic examples include the real line, all `∞ spaces, and all metric (R-)trees. Furthermore, by a result of Isbell [Isb1], every metric space X possesses an essentially unique injective hull E(X), which is the smallest injective space into which X embeds isometrically. See Appendix A n for a short survey where we also derive a characterization of injective subsets of `∞. E(X) yields a finite polyhedral complex with `∞ metrics on the cells in case X is finite. Isbell’s explicit construction was rediscovered and further explored by Dress [Dre]. The combinatorial dimension dimcomb(X) of a general metric space X is the supremum of the dimensions of the polyhedral complexes E(Y ) for all finite subspaces Y of X. In case X is already injective, every such E(Y ) embeds isometrically into X, so dimcomb(X) is bounded by the supremum of the topological dimensions of compact subsets of X. Now by a theorem in [Lan], for every word hyperbolic group Γ the injective hull E(Γ) satisfies the prerequisites of the theorem above and, moreover, Γ acts properly and cocompactly on E(Γ).
Theorem 3.16. Every word hyperbolic group Γ acts properly and cocompactly by isome- tries on the proper, finite-dimensional metric space E(Γ) admitting a consistent bicomb- ing σ that is furthermore reversible and unique. Hence σ is equivariant with respect to the isometry group of E(Γ), i.e. for every isometry L: E(Γ) → E(Γ) and x, y ∈ E(Γ) we have L ◦ σxy = σL(x)L(y). In Chapter 4 we discuss the asymptotic geometry of complete metric spaces X with a consistent bicombing σ. We define the geometric boundary ∂σX in terms of geodesic rays consistent with σ, and we equip Xσ = X∪∂σX with a natural metrizable topology akin to the cone topology in the case of CAT(0) spaces. If X is locally compact, Xσ is compact. In view of Theorem 3.16, this unifies and generalizes the respective constructions for CAT(0) or Busemann spaces and for hyperbolic groups. By virtue of the bicombing one can then give a rather direct proof of the following result on Xσ. 4 CHAPTER 1. INTRODUCTION
Theorem 4.4. Let X be a complete metric space with a consistent bicombing σ. Then Xσ is an absolute retract; in particular, Xσ is contractible and locally contractible. Moreover, ∂σX is a Z-set in Xσ, that is, for every open set U in Xσ the inclusion U \ ∂σX,→ U is a homotopy equivalence. Note that we do not assume X to be proper or finite dimensional. Bestvina and Mess [BesM] proved the analogous result for the Gromov closure P (Γ) of the (con- tractible) Rips complex of a hyperbolic group Γ. Theorem 3.16 and Theorem 4.4 together thus provide an alternative to their result, which has a number of important applications (see Corollaries 1.3 and 1.4 in [BesM]). The results so far have been published in [DesL1] together with Urs Lang. Then in Chapters 5–9 we carry the analogy with CAT(0) and Busemann spaces fur- ther with regard to existence results for flat subspaces, that is, for subspaces isometric to a normed space. Our first main result is the following generalization of the hyperbolicity criterion for cocompact CAT(0) spaces stated on p. 119 in [Gro]. A detailed proof of Gromov’s result was given in [Bri]. For the case of Busemann spaces, both Theorem 5.5 and Theorem 8.3 below were shown by Bowditch [Bow]. Theorem 5.5 (Flat plane). Let X be a proper metric space with a consistent and re- versible bicombing σ and with cocompact isometry group. Then X is hyperbolic if and only if X does not contain an isometrically embedded normed plane. En route to this theorem we discuss a generalized Flat Strip Theorem. Unlike for Busemann spaces, the σ-convex hull of a pair of parallel σ-lines may be “thick” and the two lines may span different, though pairwise isometric, flat (normed) strips. We also give a criterion for the existence of an embedded normed half-plane. This is then used in Section 5.2 for the proof of the above result. A variant of this theorem for injective metric spaces is also shown. Another well-known result from the theory of spaces of non- positive curvature is the Flat Torus Theorem, originally proven for smooth manifolds in [GroW, LawY]. A detailed account of this result and its applications in the context of CAT(0) spaces is given in Chap. II.7 of [BriH]. We have: Theorem 8.3 (Flat torus). Let X be a proper metric space with a consistent and re- versible bicombing σ. Let Γ be a group acting properly and cocompactly by isometries on X, and suppose that σ is Γ-equivariant. If Γ has a free abelian subgroup A of rank n ≥ 1, then X contains an isometrically embedded n-dimensional normed space on which A acts by translation.
Here, σ being Γ-equivariant means that γ ◦ σxy = σγ(x)γ(y) for all isometries γ ∈ Γ and (x, y) ∈ X × X. In Section 7 we establish basic properties of semi-simple isometries of spaces with bicombings. We employ a barycenter map for finite subsets which was introduced in the context of Busemann spaces in [EsSH]. Section 7.2 addresses the ques- tion whether a hyperbolic (axial) isometry of a metric space with a consistent bicombing σ also possesses an axis that is at the same time a σ-line. This is false in general but holds true for the hyperbolic elements of a group Γ as in Theorem 8.3. As an auxil- iary tool we use a fixed point theorem for non-expansive mappings originally proven for 5
Banach spaces in [GoeK]. Finally, Chapter 8 is dedicated to the proof of Theorem 8.3. Chapters 5–8 with the exception of the chapter about barycenters are a publication in preparation [DesL2] and will be co-authored with Urs Lang. In that paper some of the results (but not the main results) are stated in a somewhat more general setting. Here we decided not to do so since this facilitates the presentation without any drawback for the main results. The final Chapter 9 will revolve around the following theorem which generalizes a result of Kleiner for Busemann spaces; see Proposition 10.22 and the more comprehensive Theorem D in [Kle]. Also compare [Wen].
Theorem 9.1. Let X be a proper metric space with a bicombing σ and cocompact isom- etry group. Suppose there are sequences Rk ∈ (0, ∞), Sk ⊂ X, and a normed vector n 1 n space (R , k · k), so that Rk → ∞, and R Sk converges to the unit ball B ⊂ (R , k · k) in k n the Gromov-Hausdorff topology. Then (R , k · k) can be isometrically embedded in X. Not that the bicombing is not assumed to be consistent nor convex or even reversible. With this we are able to present Kleiner’s Theorem D in our setting. In particular, this then shows that for every proper cocompact metric space X with a bicombing the following is true: X contains a flat (normed) n-plane whenever there is a quasi-isometric n embedding of R into X. Such a result was first shown for manifolds of non-positive curvature in [AndS]. The proof of Proposition 10.22 in [Kle] is performed in the spirit of CAT(0) geometry whereas our generalized statement we prove by a quite different argument. One ingredient is a Riemannian integral defined by means of the barycenter maps. The central steps then are a convolution like operation and a kind of metric differentiation argument. We also devote a whole chapter to introduce the barycenters needed. As a byproduct, we show that a space X is isometric to a convex subset of a normed space if it admits a barycenter satisfying a simple extra property. The final chapter about Theorem 9.1 and the auxiliary results about barycenters form another publication in preparation [Des1]. Chapter 2
Basic concepts
We use this chapter to fix some notation and recall basic definitions and facts concerning metric spaces. For the proofs we omit and for a comprehensive introduction we refer to Part I in [BriH].
2.1 Metric spaces
Let X be a set. By a metric we mean a map d: X × X → R satisfying (i) d(x, y) ≥ 0 for all x, y ∈ X and d(x, y) = 0 if and only if x = y, (ii) d(x, y) = d(y, x) for all x, y ∈ X and (iii) d(x, z) ≤ d(x, y) + d(y, z) for all x, y, z ∈ X (triangle inequality). Then, when endowed with a metric, we call X a metric space; and in case we need to stress which metric is meant, we shall write (X, d). Some notation: For arbitrary x ∈ X; A, B ⊂ X; r ∈ R, we set diam(A) := supa,b∈A d(a, b) the diameter of A, d(x, A) := infa∈A d(x, a), d(A, B) := infa∈A,b∈B d(a, b), and denote by B(x, r) := {y ∈ X | d(x, y) ≤ r} the closed ball around x of radius r and by Ur(A) := {x ∈ X | d(x, A) < r} the open neighborhood of A. A map f : X → Y between metric spaces (X, dX ), (Y, dY ) is an isometry by definition, if it is surjective and dY (f(p), f(q)) = dX (p, q) for all p, q ∈ X. If such an f exists, then X and Y are called isometric. If the requirement of surjectiveness for f is dropped, we call this an isometric embedding of X into Y . The isometry group of a metric space X is the set of all isometries f : X → X with concatenation as group operation. A map f with dY (f(p), f(q)) ≤ c dX (p, q) for all p, q ∈ X is called c-Lipschitz continuous or just Lipschitz in case the value of the constant is irrelevant. A metric space is called proper if every closed ball B(x, r) is compact and this implies completeness. In metric spaces there is the following notion of length for paths (also called curves or sometimes segments) — i.e. for continuous maps α: I → X where I is an interval in R: k X l(α) := sup d(α(ti−1), α(ti)) t0≤t1≤···≤tk i=1 2.1. METRIC SPACES 7 where the supremum runs over all k ∈ N and tagged points t0 ≤ t1 ≤ · · · ≤ tk in I. We adopt the notation N = {1, 2, 3,...} and N0 = {0, 1, 2,...} throughout. When I = [a, b] we say that α is connecting (or joining) the points α(a) and α(b), and (whenever [a, b] ⊂ I) we always have l(α|[a,b]) ≥ d(α(a), α(b)) by the triangle inequality. In case that equality occurs for all a, b provided [a, b] ⊂ I, then α is commonly called a geodesic. But it is useful to adopt a slightly more restricted version of this definition. A continuous curve α: I → X for some interval I ⊂ R is called a geodesic if there is a constant c ∈ R such that d(α(t), α(t0)) = c|t − t0| for all t, t0 ∈ I, (2.1) thus the geodesics are required to be parametrized proportional to arc length (or param- etrized by arc length in case of c = 1). We also refer to this property (2.1) by saying that the geodesics have constant speed (or unit speed if c = 1). Nevertheless, both definitions of geodesics coincide up to monotonic (but not necessarily continuous) reparametrizations. From now on geodesic shall implicitly mean constant speed geodesic when not mentioned otherwise. Note that we allow for c = 0, giving stationary geodesics, and that geodesics are always global geodesics (in contrast to the wording in Riemannian geometry where geodesics are usually only required to be local geodesics). In case we need a local geodesic we will explicitly add the adjective and thereby mean — to be precise — that for a path α: I → X and every t ∈ I there is an ε > 0 such that α|[t−ε,t+ε]∩I is a geodesic. So with our denomination there can be no closed geodesics (i.e. a periodic geodesic) like 2 2 the boundary of an `1 ball in `∞ which is a local geodesic. Throughout this thesis we n n understand `p , for n ∈ N and 1 ≤ p ≤ ∞, to be the Banach space R endowed with the Lp-norm 1 p p p kxkp := (|x1| + ··· + |xn| ) when p < ∞ and kxk = max{|x1|,..., |xn|} for p = ∞. Just for later use we also define the Banach spaces `p(X) for arbitrary index sets X as follows. For a real valued function X f : X → R (the set of all such we denote by R ), the Lp-norm shall be as usual
1 ! p X p kfkp := |f(x)| for p < ∞ and kfk∞ := sup |f(x)| for p = ∞; x∈X x∈X
X quantities which may be infinite. `p(X) then is the subset of R where the said norm is finite. From (2.1) we see that geodesics are just isometric embeddings of intervals where the metric on the interval is rescaled by a factor c. In case I = R and c = 1 we call a geodesic α a line (thus a line is just an isometric embedding of R), and if I = R+ := [0, ∞) and c = 1 we refer to α as a ray and α is said to issue from α(0). Note that there are spaces where no geodesics besides the stationary ones exist. A simple 1 2 example is S := {x | kxk2 = 1} ⊂ R with the restricted Euclidean metric since the arc connecting two distinct points is always longer than their distance. Spaces were there is a geodesic connecting any pair of points are denoted geodesic spaces. If, in addition, for every pair of points there is exactly one geodesic joining them (up to affine reparametrization), the space is called uniquely geodesic. Basic examples of geodesic 8 CHAPTER 2. BASIC CONCEPTS spaces are normed spaces where the linear segment t 7→ (1 − t)x − ty joins x to y realizing the distance. Moreover, they are uniquely geodesic if and only if their norm k · k is strictly convex, i.e. k(1 − t)v + twk < 1 for all t ∈ (0, 1) and distinct vectors v, w n of norm 1. For instance, this is the case in all `p spaces except for p = 1 or p = ∞ (and n ≥ 2) where there are infinitely many geodesics connecting any pair of distinct points. To give an example, for every 1-Lipschitz function f : I → R the paths t 7→ (t, f(t)) and 2 t 7→ (f(t), t) are geodesics in `∞, and in fact every geodesic therein can be described that way. Concerning whether a complete geodesic space is proper or not, we remark that local compactness is sufficient for such spaces to be proper by virtue of the Hopf-Rinow theorem. Nevertheless, we will always write proper instead of locally compact even in these cases. We turn to the notion of Hausdorff and Gromov-Hausdorff distance. Let X be any metric space, then for two non-empty subsets A, B the Hausdorff distance is the real value dH (A, B) := inf{ε | A ⊂ Uε(B),B ⊂ Uε(A)}. Clearly, this is not a metric on the set 2X of all subsets of X; the value can be infinite if no ε with A ⊂ Uε(B),B ⊂ Uε(A) exists and the distance dH (A, B) is zero if and only if the closures A, B coincide. However, triangle inequality holds and dH is a metric on the set of all non-empty, bounded and closed subsets of some metric space X. Often we want to compare spaces which are not subsets of a common ambient space. This leads to the following definition. Let X,Y be two metric spaces, then their Gromov-Hausdorff distance is dGH (X,Y ) := inf dH (ϕ(X), ψ(Y )) ϕ,ψ,Z where the infimum runs over all metric spaces Z with isometric embeddings ϕ: X → Z, ψ : Y → Z. Alternatively, and to circumvent set theoretic anomalies, one may replace Z by the disjoint sum of X,Y and take the Hausdorff distance with respect to all the pseudometrics which restrict to the given metrics on X and Y . Here pseudometric means that we allow distinct points to have zero distance. Now for compact spaces X,Y we have that dGH (X,Y ) = 0 if and only if they are isometric. And a sequence of spaces Xi converges to a limit X in Gromov-Hausdorff distance if any only if there is a sequence of (typically non-continuous) maps fi : X → Xi such that supx,y∈X |d(fi(x), fi(y))−d(x, y)| and dH (fi(X),Xi) converge to zero for i → ∞. (The index set of a sequence is always understood to be N = {1, 2, 3,...}.) Finally, since we use the well-known theorem of Arzelà-Ascoli rather frequently, we state a version adapted to our needs. Recall that a family of maps fi : X → Y from a metric space into another is called equicontinuous if for every ε > 0 there is a δ > 0 such that d(x, y) < δ implies d(fi(x), fi(y)) < ε uniformly for all i and x, y ∈ X. Theorem 2.1 (Arzelà-Ascoli). Let X be a separable metric space and Y a proper metric spaces. Then every sequence fi : X → Y such that
(i) the family fi|C is equicontinuous for every compact set C ⊂ X,
(ii) {fi(x) | i ∈ N} is bounded for every x ∈ X, 2.2. BICOMBINGS 9 then has a subsequence converging uniformly on compact subset to a continuous limit f. Moreover, if X happens to be geodesic, then {fi(x) | i ∈ N} being bounded for one x implies it for every x.
Most often in this thesis the family of functions we are considering will have a uniform Lipschitz constant on every compact set, this clearly renders them equicontinuous on these set.
2.2 Bicombings
Recall the definition and various properties of bicombings defined on page 1–2 in the introduction. Also recall that since a bicombing σ is a selection of one geodesic for every (ordered) pair of points, a space possessing such a σ is necessarily geodesic. Occasionally, when readability demands it, we will write [x, y] instead of σxy for the geodesics of a bicombing. And, when using this notation, we do not explicitly distinguish between [x, y] as a map and its image (or trace) as long as no confusion arises. The most basic of weak non-positive curvature conditions we are considering (and which every bicombing has by convention) is the conical property (1.2). Moreover, let us mention here that there is no loss of generality in assuming a space X with (conical) bicombing σ to be complete. As for two points x, y in the completion X, we may (and must) set
σxy := limk→∞ σxkyk given two sequences xk → x, yk → y with xk, yk ∈ X. For every t ∈ [0, 1], the sequence σxkyk (t) is then Cauchy and the convergence σxkyk → σxy is uniform and the limit independent of the chosen sequences by (1.2). Furthermore, and since we very often work with proper spaces, let us note:
Proposition 2.2. If a proper metric space X admits a (possibly non-reversible) bicomb- ing σ, then X admits a reversible bicombing σ˜.
Proof. First we claim that the existence of a bicombing is equivalent to the existence of a midpoint assignment (x, y) 7→ x#y (i.e. x#y is a point such that d(x, x#y) = 1 d(x#y, y) = 2 d(x, y)) with the additional (conical) property 1 1 d(x#y, x0#y0) ≤ d(x, x0) + d(y, y0); (2.2) 2 2 and this equivalence holds for any complete space X (which every proper space is). Obviously (x, y) 7→ σxy(1/2) is a valid midpoint assignment with (2.2) given a bicombing σ. To define a bicombing from a midpoint assignment, set σxy(0) := x, σxy(1) := y, k k and Q0 := {0, 1}. Then for every k ≥ 1 let Qk := {m/2 | m odd , 0 < m < 2 }, and −k −k −k −k inductively define σxy(t) at t ∈ Qk to be σxy(t−2 )#σxy(t+2 ). See that t−2 , t+2 lie in ∪0≤i≤k−1Qi, and by induction it is easy to verify (1.2) for t ∈ Q := ∪k∈N0 Qk as well 0 0 0 as d(σxy(t), σxy(t )) = |t − t |d(x, y) for all t, t ∈ Q. So the maps σxy : Q → X defined so far are Lipschitz continuous for every pair (x, y), hence extend uniquely to geodesics σxy : [0, 1] → X (preserving (1.2)) as X is complete. 10 CHAPTER 2. BASIC CONCEPTS
In view of this it suffices to show that, starting from #, we can construct a symmetric 2 midpoint assignment , i.e. one with x y = y x for all x, y ∈ X. For a pair (x, y) ∈ X we define the sequences xi, yi by
x0 = x, y0 = y and xi = xi−1#yi−1, yi = yi−1#xi−1 for all i ≥ 1.
We have d(xi+1, yi+1) ≤ d(xi, yi) by applying (2.2) to the definition of the sequences as well as d(x , y ) d(x , x ) = d(x , y ) = i i = d(x , y ) = d(y , y ) for all j > i. i j j i 2 i j j i The latter results by induction over j and the fact that a, b ∈ M(x, y) implies a#b, 1 b#a ∈ M(x, y) where M(x, y) := {z | d(x, z) = d(z, y) = 2 d(x, y)} denotes the set of all midpoints between x and y. Consequently, if the monotone sequence d(xi, yi) would not converge to zero, then d(xi, xj) ≥ ε for all i, j and some ε > 0 thereby contradicting compactness. So both sequences converge to a common limit and we set x y = limi→∞ xi = limi→∞ yi = y x. The verification that obeys (2.2) follows when taking the limit i → ∞ of either of the inequalities 1 1 d(x , x0 ) ≤ d(x, x0) + d(y, y0), i i 2 2 1 1 d(y , y0) ≤ d(x, x0) + d(y, y0), i i 2 2 0 0 0 0 where xi, yi are the sequences in the construction of x y . And these inequalities are easily shown by mutual induction. Given a bicombing σ, we call a set C σ-convex if the image of every σ-geodesic (but not necessarily every geodesic) connecting two points of C is contained in C. Note that in the absence of reversibility for σ this requires both geodesics σxy, σyx to run entirely in C whenever x, y ∈ C. For any set D ⊂ X, we define the (closed) σ-convex hull to be the intersection of all closed σ-convex sets containing D. If we denote this hull by D, we have diam(D) = diam(D). This is a direct consequence of the following equivalent construction of the σ-convex S hull. Given D, let Dk be defined inductively as D0 = D and Dk = 2 σxy([0, 1]) (x,y)∈Dk−1 for k ≥ 1. Clearly Dk−1 ⊂ Dk for all k and the set ∪k∈N0 Dk is σ-convex. Now D is the closure of that set (which is still σ-convex as the bicombing is continuous). Since diam(Dk−1) = diam(Dk) from the conical inequality, taking the σ-convex hull does not increase the diameter. Also, observe that for a non-consistent or non-reversible bicombing the σ-convex hull of {x, y} may not just be σxy([0, 1]). In fact, the only bicombings for which the trace of every geodesic σxy is σ-convex are precisely the ones that are consistent and reversible. For arbitrary bicombings, examples of σ-convex sets include balls and neighborhoods Ur(A) of σ-convex sets A as well as M(x, y), the set of midpoints for any pair x, y.A midpoint for x, y is a point z with d(x, z) = d(z, y) = 1 2 d(x, y), and every geodesic α from α(0) = x to α(1) = y has to pass M(x, y) exactly at α(1/2) and only then. 2.3. GROUP ACTIONS 11
2.3 Group actions
Here we collect the definitions about group actions we need in the sequel. By an action by isometries of a group Γ on a metric space X we mean a group homomorphism ϕ:Γ → Isom(X) from Γ into the group of isometries on X. We do not distinguish (notation- wise) between γ as an element of Γ and the isometry ϕ(γ) assigned to it. Therefore we simply write γ(x) instead of (ϕ(γ))(x) for all x ∈ X. Equivalently to the definition above, we may define an action by isometries as a map Γ × X → X, (γ, x) → γ(x) such that γ(·): X → X is an isometry for all γ ∈ Γ and γ(δ(x)) = (γδ)(x), e(x) = x for all x ∈ X and γ, δ ∈ Γ where e ∈ Γ denotes the identity element. A proper action is one where {γ | γ(C) ∩ C 6= ∅} is finite for every compact subset C ⊂ X. We will use this mainly in the context of proper metric spaces X and there we can replace compact by bounded and have that {γ | γ(B(x, r)) ∩ B(x, r) 6= ∅} is finite for all x ∈ X, r ∈ R. An action by Γ on X is called cocompact if there exists a compact set C ⊂ X such that ΓC = X where ΓC = {γ(c) | γ ∈ Γ, c ∈ C}, thus the orbit of that compact set covers the whole space. This last property is especially useful in combination with the Arzelà-Ascoli Theorem 2.1. Let fi : X → Y be a sequence of maps and assume some group Γ acts cocompactly by isometries on the proper target space Y . If fi fulfills the requirements of Theorem 2.1 for a geodesic X but with the exception of part (ii) ({fi(x) | i ∈ N} is bounded for every x ∈ X), then there is a sequence γi ∈ Γ such that property (ii) holds for the sequence γi ◦ fi. Whenever the canonical action of the full isometry group Isom(X) on X is cocompact, we simply call X a cocompact metric space. Finally, given an action by Γ on a space with bicombing σ, we then say that σ is Γ-equivariant if
γ ◦ σxy = σγ(x)γ(y) for all x, y and γ ∈ Γ. (2.3)
If we want to express that (2.3) holds for a single isometry γ, we say that σ is γ- equivariant. Chapter 3
Improving bicombings
First we discuss some simple properties and examples of (conical) bicombings as defined in (1.2). Lemma 3.1. Let X¯ be a metric space with a bicombing σ¯. If π : X¯ → X is a 1-Lipschitz ¯ retraction onto some subspace X of X, then σ := π ◦ σ¯|X×X×[0,1] defines a bicombing on X. Furthermore, if σ¯ is reversible so is σ.
Proof. Note that since π is a 1-Lipschitz retraction, σxy is indeed a geodesic for every pair (x, y) ∈ X2. Furthermore, since π is 1-Lipschitz and σ¯ is conical, we have
0 0 d(σxy(t), σx0y0 (t)) ≤ d(¯σxy(t), σ¯x0y0 (t)) ≤ (1 − t) d(x, x ) + t d(y, y ) for all t ∈ [0, 1] and x, y, x0, y0 ∈ X, so σ is conical as well. The last statement follows from σxy(t) = π(¯σxy(t)) = π(¯σyx(1 − t)) = σyx(1 − t). As mentioned earlier (also see Theorem A.5), a direct consequence of this lemma is that all injective metric spaces (or absolute 1-Lipschitz retracts) admit bicombings. An equally simple application gives an example of a bicombing that is not convex. 2 Example 3.2. Consider the set X of all points (u, v) ∈ `∞ with |u| ≤ 2 and b(u) := 2 |u| − 1 ≤ v ≤ |b(u)| endowed with the metric induced by the `∞ norm. Note that this is a geodesic metric space (even injective by Lemma A.8). With respect to this metric, the vertical retraction π from the triangle X¯ := {(u, v) | b(u) ≤ v ≤ 1} onto X that maps (u, v) to (u, min{v, |b(u)|}) is 1-Lipschitz. The linear bicombing σ¯ on X¯, defined by σ¯xy(t) := (1 − t)x + ty, is convex. By Lemma 3.1, σ := π ◦ σ¯|X×X×[0,1] defines a bicombing on X. For x = (−2, 1) and y = (2, 1), we have σxy(1/4) = (−1, 0), σxy(3/4) = (1, 0), and σxy(1/2) = (0, 1). Hence, for z = (0, −1), the function t 7→ kσxy(t) − σzz(t)k∞ = kσxy(t) − zk∞ is clearly not convex.
3.1 From conical to convex bicombings
We now define a relaxed notion of convexity that will be useful for the proof of Theo- rem 3.4. We say that a bicombing σ on a metric space X is 1/k-discretely convex if, for 3.1. FROM CONICAL TO CONVEX BICOMBINGS 13 all x, y, x0, y0 ∈ X, the convexity condition t − s s − r d(σ (s), σ 0 0 (s)) ≤ d(σ (r), σ 0 0 (r)) + d(σ (t), σ 0 0 (t)) xy x y t − r xy x y t − r xy x y holds whenever the three numbers r < s < t belong to [0, 1] ∩ (1/k)Z. To check this condition, it clearly suffices to verify the “local” inequality
2d(σxy(s), σx0y0 (s)) ≤ d(σxy(s − 1/k), σx0y0 (s − 1/k))
+ d(σxy(s + 1/k), σx0y0 (s + 1/k)) for all s ∈ (0, 1) ∩ (1/k)Z. Note that every (conical) bicombing is 1/2-discretely convex. Proposition 3.3. Suppose that X is a complete metric space with a bicombing σ that is 1/k-discretely convex for some integer k ≥ 2. Then X also admits a bicombing that is 1/(2k − 1)-discretely convex.
Proof. Set l := 2k − 1. To construct the desired new bicombing σ˜, fix x and y and define the sequences pi and qi recursively as q0 := σxy(k/l) and
pi := σxqi−1 (1 − 1/k), qi := σpiy(1/k), for i ≥ 1.
p σ˜xy q
p4 q3
y x
q1 = σp1y(1/k)
p1 = σxq0 (1 − 1/k)
q0 = σxy(k/l)
Since σ is conical, we get the inequalities
d(pi, pi+1) ≤ (1 − 1/k) d(qi−1, qi),
d(qi, qi+1) ≤ (1 − 1/k) d(pi, pi+1).
It follows that pi and qi are Cauchy sequences, so pi → p and qi → q. Then σxqi → σxq and σpiy → σpy uniformly, again because σ is conical. Note that p = σxq(1 − 1/k) and q = σpy(1/k); we can thus define σ˜xy(s) for s ∈ [0, 1] ∩ (1/l)Z so that ( σxq (l/k)s if s ≤ k/l, σ˜xy(s) = (3.1) σpy (l/k)s − (k − 1)/k if s ≥ (k − 1)/l. 14 CHAPTER 3. IMPROVING BICOMBINGS
0 To declare σ˜xy on all of [0, 1], we connect any pair of consecutive points x :=σ ˜xy(s) 0 and y :=σ ˜xy(s + 1/l), where s ∈ [0, 1) ∩ (1/l)Z, by the geodesic t 7→ σx0y0 (l(t − s)) for t ∈ [s, s + 1/l]. 0 0 0 0 Now if p , q and σ˜x0y0 result from the same construction for two points x and y , we want to show that for s ∈ (0, 1) ∩ (1/l)Z we have
2d(˜σxy(s), σ˜x0y0 (s)) ≤ d(˜σxy(s − 1/l), σ˜x0y0 (s − 1/l))
+ d(˜σxy(s + 1/l), σ˜x0y0 (s + 1/l)). In view of (3.1) this corresponds to the inequality
2d(α(t), β(t)) ≤ d(α(t − 1/k), β(t − 1/k)) + d(α(t + 1/k), β(t + 1/k)) where α = σxq, β = σx0q0 , t = (l/k)s if s ≤ (k − 1)/l; and α = σpy, β = σp0y0 , t = (l/k)s − (k − 1)/n if s ≥ k/l. However, these inequalities hold since σ is 1/k- discretely convex. This shows that σ˜ is 1/l-discretely convex. Now it follows easily from the construction that σ˜ is also conical.
From this result, we now obtain Theorem 3.4 by an application of the Arzelà-Ascoli Theorem 2.1, which requires X to be proper. We do not know whether the implication (E) ⇒ (D) holds in general without this assumption. Theorem 3.4. Let X be a proper metric space with a bicombing. Then X also admits a convex bicombing. Proof. Starting from the given bicombing σ1 := σ, we construct, by means of Proposi- i i tion 3.3, a sequence of bicombings σ on X such that σ is 1/ki-discretely convex, where i k1 = 2 and ki+1 = 2ki −1. This collection of maps σ is equicontinuous on every bounded set, and for every fixed (x, y, t) in the separable domain X × X × [0, 1], the sequence i i(j) σxy(t) remains in a compact subset of X. One may thus extract a subsequence σ that converges uniformly on every compact set to a map σ¯, which is clearly a bicombing. Convexity t − s s − r d(¯σ (s), σ¯ 0 0 (s)) ≤ d(¯σ (r), σ¯ 0 0 (r)) + d(¯σ (t), σ¯ 0 0 (t)), xy x y t − r xy x y t − r xy x y
i(j) i(j) where r < s < t, follows from the corresponding inequality for σxy , σx0y0 and rl < sl < tl by choosing these numbers in [0, 1] ∩ (1/ki(j))Z such that rl → r, sl → s, and tl → t. The following observation regarding the above construction will be used in Exam- ple 3.10. Remark 3.5. Let σ be a bicombing on the proper metric space X, and suppose that for some pair of points x, y the consistency condition holds, i.e. σx0y0 (λ) = σxy((1−λ)s+λt) 0 0 for all 0 ≤ s ≤ t ≤ 1 and λ ∈ [0, 1], where x := σxy(s) and y := σxy(t). Then it is easily seen that σxy as well as all positively oriented subsegments are unaltered by the procedure in the above proof. In other words, the resulting convex bicombing σ¯ satisfies 0 0 σ¯x0y0 = σx0y0 for all x , y as above, in particular σ¯xy = σxy. 3.2. STRAIGHT CURVES 15
3.2 Straight curves
Whereas the preceding section dealt with the existence of convex bicombings, we now turn to the question of uniqueness. First we consider a property every geodesic from a convex bicombing necessarily shares. Let X be a metric space. We call a curve α:[a, b] → X straight (or a straight segment) if for every z ∈ X the function dz ◦ α is convex, where dz = d(z, ·). In particular, for fixed s, t ∈ [a, b], taking z := α(s) one gets the inequality
d(α(s), α((1 − λ)s + λt)) ≤ λ d(α(s), α(t)) for all λ ∈ [0, 1], whereas for z := α(t) one obtains
d(α((1 − λ)s + λt), α(t)) ≤ (1 − λ) d(α(s), α(t)) for all λ ∈ [0, 1].
By taking the sum of these two inequalities one sees that straight curves are geodesics (of constant speed). The terminology is further justified by Theorem 3.8 below. In the proof of this result as well as in Example 3.10 we use Isbell’s injective hull construction [Isb1]. For convenience we review this in Appendix A. We denote the injective hull of a metric space X by E(X) and given a straight curve in X, we first observe that this property persists when we pass from X to E(X).
Lemma 3.6. Let α:[a, b] → X be a straight curve in some metric space X. Regarding X as a subspace of its injective hull E(X) we then have that α is also straight in E(X).
Proof. By (A.2), the distance from an element f ∈ E(X) to a point x ∈ X ⊂ E(X) equals f(x). So we need to show that the function f ◦ α is convex. Given x := α(s), y := α(t), and q := α((1 − λ)s + λt), where λ ∈ [0, 1], let ε > 0 and choose (using the extremity of f, see (A.1)) a point p ∈ X such that f(p) + f(q) ≤ d(p, q) + ε. Since α is straight in X, we have
d(p, q) ≤ (1 − λ) d(p, x) + λ d(p, y).
Furthermore, d(p, x) ≤ f(p) + f(x) and d(p, y) ≤ f(p) + f(y). Combining these inequal- ities we get f(p) + f(q) ≤ f(p) + (1 − λ)f(x) + λf(y) + ε. Since ε was arbitrary, this gives f(q) ≤ (1 − λ)f(x) + λf(y), as desired.
Now let X = V be a normed real vector space. Isbell [Isb2] and Rao [Rao] (see also [CiaD]) showed that then the injective hull E(V ) has a Banach space structure with respect to which the isometric embedding e: V → E(V ) is linear. Since E(V ) is injective, collections of balls in E(V ) have the, so-called, binary intersection property, so the Banach space E(V ) is also injective in the linear category by [Nac]. Then a theorem of Nachbin, Goodner, and Kelley [Kel] implies that E(V ) is isometrically isomorphic to the space C(M) of continuous functions, with the supremum norm, on some extremally 16 CHAPTER 3. IMPROVING BICOMBINGS disconnected compact Hausdorff space M. Summarizing, we may thus view V as a linear subspace of C(M), where M is such that E(V ) =∼ C(M). This fact will be used below. As usual, we call a curve α:[a, b] → V in a vector space V linear if it is of the form t 7→ p + tv for some p, v ∈ V . This is obviously a local property.
Proposition 3.7. Let M be a compact Hausdorff space, and let α:[a, b] → U be a curve in an open subset U of C(M). Then α is straight (in U with the induced metric) if and only if it is linear.
Proof. Clearly every linear curve is straight. For the other direction, we assume that α: [0, 1] → U is a straight curve from 0 to y, where kyk∞ = l, and the closed 2l- neighborhood of α([0, 1]) is contained in U. We have to show that the two functions α(λ) and λy agree for every λ ∈ (0, 1). So fix λ ∈ (0, 1) as well as m ∈ M. For an arbitrary ε > 0, choose an open neighborhood B of m such that
|α(λ)(m0) − α(λ)(m)| < ε/2, |y(m0) − y(m)| < ε/2 for all m0 ∈ B. By Urysohn’s lemma there is a non-negative function ϕ ∈ C(M) vanishing on M \ B and kϕk∞ = 2l. Put z± := α(λ) ± ϕ and note that these are elements of U. Since α is straight,
2l = kz± − α(λ)k∞ ≤ (1 − λ)kz±k∞ + λkz± − yk∞. (3.2)
0 Now if f ∈ C(M) is a function satisfying kfk∞ ≤ l as well as f(m ) < f(m) + ε for all 0 m ∈ B, then (f + ϕ)|B < 2l + f(m) + ε, (f + ϕ)|M\B ≤ l, −(f + ϕ) ≤ −f ≤ l, and l ≤ 2l + f(m), hence kf + ϕk∞ ≤ 2l + f(m) + ε.
Taking f = α(λ) and f = α(λ) − y we get kz+k∞ ≤ 2l + α(λ)(m) + ε and kz+ − yk∞ ≤ 2l + α(λ)(m) − y(m) + ε, respectively. (Note that kfk∞ ≤ l since α is a geodesic from 0 to y.) Together with (3.2), this shows that
α(λ)(m) − λy(m) + ε ≥ 0.
Similarly, for f = −α(λ) and f = y − α(λ) we obtain kz−k∞ ≤ 2l − α(λ)(m) + ε and kz− − yk∞ ≤ 2l + y(m) − α(λ)(m) + ε, respectively, which gives
λy(m) − α(λ)(m) + ε ≥ 0.
Letting ε → 0 we conclude that α(λ)(m) = λy(m), ending the proof as λ and m were arbitrary.
Combining these results we obtain the desired characterization of linear geodesics in normed spaces.
Theorem 3.8. A straight curve in an arbitrary normed space V is linear. Hence the bicombing of linear geodesics is the only convex bicombing on V . 3.2. STRAIGHT CURVES 17
Proof. As mentioned above, we can regard the normed space V as being a linear subspace of its injective hull E(V ), and we have E(V ) =∼ C(M) for some compact Hausdorff space M. Now straight curves in V are straight in C(M) by Lemma 3.6 and linear by Proposition 3.7.
As we learned after developing the above result, on every normed space there is in fact no other (conical) bicombing than the linear one; see [GaeM]. One may ask whether the equivalence between straight and linear holds more generally for linearly convex subsets of normed spaces. The answer turns out to be negative.
Example 3.9. Consider the Banach space `∞([0, 1]) of bounded real valued functions on [0, 1], equipped with the supremum norm. The subset
C := {f ∈ `∞([0, 1]) | f(0) + f(1) = 1, f is convex, and f ∈ ∆1([0, 1])} is compact and linearly convex (see Appendix A for the definition of ∆1). Hence there is the convex bicombing of linear geodesics in C. But there are more straight segments: the curve α: [0, 1] → C, α(t) = dt, is non-linear, and by (A.2) we have t 7→ kf − α(t)k∞ = f(t) for every f ∈ C, which is a convex function by definition of C.
Let us mention here that we do not know whether the non-linear straight curve α belongs to a convex bicombing, thus whether there is a bicombing σ on C with σα(0)α(1) = α. It may very well be that there is no convex bicombing on C other than the linear one. In contrast to this we conclude this section with an example of a compact metric space that admits at least two distinct convex bicombings.
Example 3.10. Consider two geodesics α, β : [0, 1] → `∞([0, 1]) from d0 to d1: the linear geodesic α(s) = (1 − s)d0 + sd1, and the Kuratowski embedding β(t) = dt of [0, 1]. Let B ⊂ `∞([0, 1]) be the bigon composed of these two geodesic segments. By (A.2) we have
kα(s) − β(t)k∞ = α(s)(t) = s + t − 2st. (3.3)
Since this last term is symmetric in s and t, there is an isometric involution ι: B → B that interchanges α and β. Furthermore, as `∞([0, 1]) is injective, we can embed the injective hull E(B) of B into `∞([0, 1]) (and identify it with its image), so that B ⊂ E(B) ⊂ `∞([0, 1]). Now E(B) is not linearly convex in `∞([0, 1]), but by retracting the linear geodesics with endpoints in E(B) to E(B) we obtain a bicombing σ on E(B)
(compare Lemma 3.1). Note that since α was already linear, we have σd0d1 = α. Note further that E(B) is compact, because B is compact. Theorem 3.4 then yields a convex bicombing σ¯ which, by Remark 3.5, still satisfies σ¯d0d1 = α. Finally, the involution ι extends uniquely to an isometry I of E(B) (see Proposition A.4 which is Proposition 3.7 in [Lan]). Mapping σ¯ by I we get a convex bicombing τ¯ on E(B) with τ¯d0d1 = β, distinct from σ¯. 18 CHAPTER 3. IMPROVING BICOMBINGS
3.3 Combinatorial dimension
The example just described contrasts with Theorem 3.14, which we will prove in this section. First we discuss the structure of injective hulls of finite metric spaces and the notion of combinatorial dimension. X ∼ n Let X be a finite metric space, with |X| = n ≥ 1, say. The set ∆(X) ⊂ R = R is an unbounded polyhedral domain, determined by the finitely many linear inequalities f(x) + f(y) ≥ d(x, y) for x, y ∈ X (in particular f ≥ 0). As X is finite, a function f ∈ ∆(X) is extremal if and only if for every x ∈ X there exists a point y ∈ X such that f(x) + f(y) = d(x, y). So the injective hull is a polyhedral subcomplex of ∂∆(X) of dimension at most n/2. (It is not difficult to see that E(X) consists precisely of the bounded faces of ∂∆(X).) For n ≤ 5, the various possible combinatorial types of E(X) are depicted in Section 1 of [Dre] (where ∆(X) and E(X) are denoted PX and TX , respectively). To describe the structure of E(X) further, one may assign to every f ∈ E(X) the undirected graph with vertex set X and edge set
A(f) = {x, y} | x, y ∈ X, f(x) + f(y) = d(x, y) .
Note that this graph has no isolated vertices (because f is extremal), but may be dis- connected, and there is a loop {x, x} if and only if f(x) = 0, which occurs if and only if f = dx (by (A.2)). Call a set A of unordered pairs of (possibly equal) points in X admis- sible if there exists an f ∈ E(X) with A(f) = A, and denote by A (X) the collection of admissible sets. The family of polyhedral faces of E(X) is then given by {P (A)}A∈A (X), where P (A) = {f ∈ ∆(X) | A ⊂ A(f)}, and where P (A0) is a face of P (A) if and only if A ⊂ A0. We define the rank rk(A) of A as the dimension of P (A). This number can be read off as follows. If f, g are two elements of P (A), then f(x) + f(y) = d(x, y) = g(x) + g(y) for {x, y} ∈ A, so f(y) − g(y) = −(f(x) − g(x)). Thus the difference f − g has alternating sign along all edge paths in the graph (X,A). It follows that there is either no or exactly one degree of freedom for the values of f ∈ P (A) on every connected component of (X,A), depending on whether or not the component contains a cycle of odd length. We call such components (viewed as subsets of X) odd or even A-components, respectively. The rank rk(A) = dim(P (A)) is then precisely the number of even A-components of X. (Here we have adopted the notation from [Lan], whereas [Dre] uses Kf = {(x, y) ∈ X × X | f(x) + f(y) = d(x, y)} in place of A(f).) Now let again X be a general metric space. We recall that the combinatorial dimen- sion of X, introduced by Dress, is the possibly infinite number
dimcomb(X) = sup{dim(E(Y )) | Y ⊂ X, |Y | < ∞}, see Theorem 90 on p. 380 in [Dre]. This theorem contains in particular a characterization of the inequality dimcomb(X) ≤ n in terms of a 2(n + 1)-point inequality, which may be rephrased as follows. 3.3. COMBINATORIAL DIMENSION 19
Theorem 3.11 (Dress). Let X be a metric space, and let n ≥ 1 be an integer. The inequality dimcomb(X) ≤ n holds if and only if for every set Z ⊂ X with |Z| = 2(n + 1) and every fixed point free involution i: Z → Z there exists a fixed point free bijection j : Z → Z distinct from i such that
X d(z, i(z)) ≤ X d(z, j(z)). (3.4) z∈Z z∈Z
The case n = 1 corresponds to the much simpler fact that dimcomb(X) ≤ 1 if and only if X is 0-hyperbolic in the sense of Gromov [Gro] or tree-like in the terminology of [Dre], that is, for every quadruple of points x, x0, y, y0 ∈ X,
d(x, x0) + d(y, y0) ≤ max{d(x, y) + d(x0, y0), d(x, y0) + d(x0, y)}.
Theorem 3.11 follows from more general considerations in Section (5.3) of [Dre]. We will not use this result in the sequel. By the results of [Lan], every proper metric space with integer valued metric that is discretely geodesic and δ-hyperbolic has finite combinatorial dimension. By contrast, the unit circle S1 in the Euclidean plane with either the induced 1 (chordal) or the induced inner metric satisfies dimcomb(S ) = ∞, as is seen by looking at the constant extremal function f = diam(S1)/2, restricted to the vertices of a regular 2n-gon. Similarly, the metric bigon B constructed in Example 3.10 has dimcomb(B) = ∞: consider the function f defined by f(α(s)) = f(β(1 − s)) = kα(s) − β(1 − s)k∞/2 for s ∈ [0, 1]. Among the finite dimensional normed spaces, only those with a polyhedral norm have finite combinatorial dimension, equal to the number of pairs of opposite facets of the unit ball. The following proposition is the key observation for the proof of Theorem 3.14.
Proposition 3.12. Let X be a metric space of finite combinatorial dimension. Then for every pair of points x0, y0 ∈ X there exists a δ > 0 such that
d(x0, y0) + d(x, y) ≤ d(x, y0) + d(x0, y) for all pairs of points x ∈ B(x0, δ) and y ∈ B(y0, δ).
Proof. When x0 = y0, the triangle inequality in X gives the result. Now assume that x0 6= y0. Denote by F the collection of all real valued functions with finite support spt(f) ⊂ X such that f ∈ E(spt(f)), x0, y0 ∈ spt(f), and {x0, y0} ∈ A(f), that is,
f(x0) + f(y0) = d(x0, y0).
Since dimcomb(X) < ∞, there exist an integer n and a function f ∈ F such that rk(A(g)) ≤ n = rk(A(f)) for all g ∈ F . Since x0 6= y0, we have n ≥ 1, thus f > 0. By restricting f to a smaller set if necessary, we can arrange that A(f) is the collection of n disjoint pairs {x0, y0}, {x1, y1},..., {xn−1, yn−1}. There exists a δ > 0 such that for all p ∈ {x0, y0} and q ∈ {x1, y1, . . . , xn−1, yn−1} we have d(p, q) > δ, f(p) ≥ δ, and
f(p) + f(q) ≥ d(p, q) + 2δ. (3.5) 20 CHAPTER 3. IMPROVING BICOMBINGS
Note that d(x0, y0) = f(x0) + f(y0) ≥ 2δ. Let x ∈ B(x0, δ) and y ∈ B(y0, δ). If x = x0 or y = y0 or x = y, the desired inequality holds. So assume that x0 6= x 6= y 6= y0. Then x, y, x0, y0, . . . , xn−1, yn−1 are pairwise distinct. Put a := d(x, y0) − f(y0) and b := d(x0, y) − f(x0). We have
a ≥ d(x0, y0) − d(x, x0) − f(y0) = f(x0) − d(x, x0) ≥ f(x0) − δ, (3.6) b ≥ d(x0, y0) − d(y, y0) − f(x0) = f(y0) − d(y, y0) ≥ f(y0) − δ.
Since f(x0), f(y0) ≥ δ, this gives in particular a, b ≥ 0 and hence
a + f(x0) ≥ δ ≥ d(x, x0), (3.7) b + f(y0) ≥ δ ≥ d(y, y0).
Furthermore, for every q ∈ {x1, y1, . . . , xn−1, yn−1}, combining (3.6) and (3.5) we obtain
a + f(q) ≥ f(x0) + f(q) − δ ≥ d(x0, q) + δ ≥ d(x, q), (3.8) b + f(q) ≥ f(y0) + f(q) − δ ≥ d(y0, q) + δ ≥ d(y, q).
Now, in the case that a + b < d(x, y), we could define a function g with finite support by putting g(q) := f(q) for q ∈ {x0, y0, . . . , xn−1, yn−1} and by choosing g(x) > a and g(y) > b such that g(x) + g(y) = d(x, y). In view of (3.7) and (3.8), this function would satisfy A(g) = {x0, y0},..., {xn−1, yn−1}, {x, y} , so that g ∈ F and rk(A(g)) = n + 1, in contradiction to the maximality of n. We conclude that d(x, y) ≤ a + b = d(x, y0) + d(x0, y) − d(x0, y0).
Resuming the discussion of straight curves, we can now prove the following.
Proposition 3.13. Suppose that X is a metric space of finite combinatorial dimension, and α, β : [0, 1] → X are two straight curves. Then the function s 7→ d(α(s), β(s)) is convex on [0, 1]. In particular, every pair of points in X is joined by at most one straight segment, up to reparametrization.
Proof. For s, t ∈ [0, 1], put h(s, t) := d(α(s), β(t)). Fix s0 ∈ (0, 1). Then it follows from Proposition 3.12 that there exists an ε > 0 such that [s0 − ε, s0 + ε] ⊂ [0, 1] and, for all s, t ∈ [s0 − ε, s0 + ε],
h(s0, s0) + h(s, t) ≤ h(s, s0) + h(s0, t). (3.9)
Now suppose that s0 − ε ≤ s < s0 < t ≤ s0 + ε, and let λ ∈ (0, 1) be such that s0 = (1 − λ)s + λt. Since h(s, ·) and h(·, t) are convex functions on [0, 1], we have
h(s, s0) ≤ (1 − λ) h(s, s) + λ h(s, t), (3.10) h(s0, t) ≤ (1 − λ) h(s, t) + λ h(t, t). 3.3. COMBINATORIAL DIMENSION 21
Combining (3.9) and (3.10) we conclude that
h(s0, s0) ≤ (1 − λ) h(s, s) + λ h(t, t).
Note that this holds whenever s0 − ε ≤ s < s0 < t ≤ s0 + ε and s0 = (1 − λ)s + λt, where ε > 0 depends on s0. Since s 7→ h(s, s) = d(α(s), β(s)) is continuous on [0, 1], it follows easily that this function is convex.
Now the next theorem is an immediate corollary of Proposition 3.13. Note that a bicombing σ can be restricted to any ball because balls are σ-convex.
Theorem 3.14. Let X be a metric space of finite combinatorial dimension in the sense of Dress, or with the property that every bounded subset has finite combinatorial di- mension. Suppose that X possesses a convex bicombing σ. We then have that σ is consistent, reversible, and unique, that is, σ is the only convex bicombing on X. Since straight segments are unique, σ is also equivariant with respect to the isometry group.
Recall that equivariant with respect to the isometry group means that for every isometry γ : X → X and all x, y ∈ X we have γ ◦ σxy = σγ(x)γ(y). The assumptions of the theorem are, in particular, satisfied for proper injective metric spaces of finite combinatorial dimension.
Theorem 3.15. Let X be a proper injective metric space of finite combinatorial dimen- sion. Then X possesses a unique convex bicombing, which furthermore is consistent, reversible, and equivariant with respect to the whole isometry group.
Proof. Such spaces admit a bicombing by the remark after Lemma 3.1 and a convex bicombing by Theorem 3.4.
As addressed before, by a theorem of [Lan] we know that if Γ is a Gromov hyperbolic group, endowed with the word metric with respect to some finite generating set, then the injective hull E(Γ) is proper, has finite combinatorial dimension, and Γ acts properly and cocompactly on E(Γ) by isometries. Therefore the above theorem is applicable, and we obtain the last result of this section mentioned in the introduction.
Theorem 3.16. Every word hyperbolic group Γ acts properly and cocompactly by isome- tries on the proper, finite-dimensional metric space E(Γ) admitting a consistent bicomb- ing that is furthermore reversible, and unique, hence equivariant with respect to the isometry group of X. Chapter 4
Boundary at infinity
We now consider a complete metric space X with a consistent bicombing. Note that completeness is no restriction as any bicombing may be extended to the completion of the underlying space. We define the geometric boundary and the closure of X by means of geodesic rays that are consistent with the given bicombing, and we equip the closure with a simple explicit metric. (Some general references for the analogous constructions in the case of CAT(0) spaces or Gromov hyperbolic spaces are [Bal, BriH, BuyS, Gro].) Then we prove Theorem 4.4. Recall that by a ray in X we mean an isometric embedding of R+ := [0, ∞). Two rays ξ, η in X are asymptotic if the function t 7→ d(ξ(t), η(t)) is bounded or, equivalently, the Hausdorff distance between the images of ξ and η is finite. In the presence of a consistent bicombing σ on X we call a ray ξ : R+ → X a σ-ray if
ξ((1 − λ)s + λt) = σxy(λ) (4.1) whenever 0 ≤ s ≤ t, x := ξ(s), y := ξ(t), and λ ∈ [0, 1]. It follows that for any two 0 0 0 0 σ-rays ξ, η the map t 7→ d(ξ(a + a t), η(b + b t)) is convex on R+ for all a, a , b, b ≥ 0. The geometric boundary ∂σX is the set of equivalence classes of mutually asymptotic σ-rays in X, and we write
Xσ := X ∪ ∂σX.
For a unified treatment of the two parts of Xσ it is convenient to associate with every pair (o, x) ∈ X × X the eventually constant curve %ox : R+ → X, ( σox(t/d(o, x)) if 0 ≤ t < d(o, x), %ox(t) := (4.2) x if t ≥ d(o, x).
As a preliminary remark we note that for any basepoint o ∈ X and r ∈ R+, the radial retraction ϕr : X → B(o, r) defined by ϕr(x) := %ox(r) satisfies
2r d(ϕ (x), ϕ (y)) ≤ d(x, y) (4.3) r r d(o, x) 23
0 whenever d(o, x) ≥ d(o, y) and d(o, x) > r. To see this, let s := r/d(o, x) and y := σoy(s). 0 We have ϕr(x) = σox(s), and since σ is conical, d(ϕr(x), y ) ≤ s d(x, y). Further- 0 0 0 0 more, d(y , ϕr(y)) = d(o, ϕr(y)) − d(o, y ) ≤ d(o, ϕr(x)) − d(o, y ) ≤ d(ϕr(x), y ) and so 0 0 d(ϕr(x), ϕr(y)) ≤ d(ϕr(x), y )+d(y , ϕr(y)) ≤ 2s d(x, y). In particular, ϕr is 2-Lipschitz. 2 The constant 2 is optimal, as one can see by looking at the space `∞ with 0 as the basepoint, the map ϕ1, and the points (1, 1), (1 + ε, 1 − ε). In order to prove that for every pair (o, x¯) ∈ X × ∂σX there is a σ-ray issuing from o and representing the class x¯, we shall need the following estimate (compare Lemma II.8.3 in [BriH] for the case of CAT(0) spaces).
Lemma 4.1. Let X be a metric space with a consistent bicombing σ, and let o, p ∈ X. Then for any σ-ray ξ with ξ(0) = p we have
2t d(o, p) d(% (t),% (t)) ≤ ox oy T − d(o, p) whenever T > 2d(o, p), x, y ∈ ξ([T, ∞)), and 0 ≤ t ≤ T − 2 d(o, p).
0 Proof. Assume d(o, x) ≤ d(o, y) and let s := d(p, x)/d(p, y) and y := σoy(s). Since ξ is a σ-ray, we have x = σpy(s). As σ is conical,
d(x, y0) ≤ (1 − s) d(p, o) ≤ d(o, p). (4.4)
Now d(o, x) ≥ d(p, x) − d(o, p) ≥ T − d(o, p) and so d(o, y0) ≥ d(o, x) − d(x, y0) ≥
o %oy(t)
%ox(t) ∂B(o, t)
0 y = σoy(s) p ξ T x y
0 T −2 d(o, p). Hence, for 0 ≤ t ≤ T −2 d(o, p), we have ϕt(x) = %ox(t) and ϕt(y ) = %oy(t), and (4.3) gives the result.
The next result now follows by a standard procedure.
Proposition 4.2. Let X be a complete metric space with a consistent bicombing σ. Then for every pair (o, x¯) ∈ X × ∂σX there is a unique σ-ray %ox¯ with %ox¯(0) = o that represents the class x¯. Furthermore, if r ≥ 0 and x = %ox¯(r), then %xx¯(t) = %ox¯(r + t) for all t ∈ R+.
Proof. Let ξ be a σ-ray in the class x¯, and let p := ξ(0). For n = 1, 2,... , put %n := %oξ(n). It follows from the preceding lemma that for every fixed t ≥ 0 the sequence %n(t) is Cauchy. In the limit one obtains a σ-ray %ox¯ issuing from o. As in (4.4), we have 24 CHAPTER 4. BOUNDARY AT INFINITY d(ξ(t),%n[t d(o, ξ(n))/n]) ≤ d(o, p) for 0 ≤ t ≤ n, hence d(ξ(t),%ox¯(t)) ≤ d(o, p) for all 0 t ≥ 0. In particular, %ox¯ is asymptotic to ξ. Finally, if % is another σ-ray issuing 0 from o and asymptotic to ξ, then t 7→ d(%oξ(t),% (t)) is a non-negative, bounded, convex 0 function on R+ that vanishes at 0, so % = %oξ. From this uniqueness property, the last assertion of the proposition is clear.
A natural topology on Xσ = X ∪ ∂σX may be described in different ways. First we fix a basepoint o ∈ X and consider the set
Rσ,o := {%ox¯ | x¯ ∈ Xσ} = {%ox | x ∈ X} ∪ {%ox¯ | x¯ ∈ ∂σX} of generalized σ-rays based at o, given by (4.2) and Proposition 4.2. We equip Rσ,o with the topology of uniform convergence on compact subsets of R+. Clearly Rσ,o is compact if and only if X is proper, as a consequence of the Arzelà-Ascoli theorem. The topology of Rσ,o agrees, under canonical identification, with the cone topology on Xσ, a basis of which is given by the sets
Uo(¯x, t, ε) := {y¯ ∈ Xσ | d(%ox¯(t),%oy¯(t)) < ε} (4.5)
∂σX x¯
%ox¯(t)
ε Uo(¯x, t, ε)
o
Figure 4.1: Uo(¯x, t, ε) for an x¯ ∈ ∂σX for x¯ ∈ Xσ and t, ε > 0. Note that since ϕr is 2-Lipschitz, we have
d(%ox¯(r),%oy¯(r)) ≤ 2 d(%ox¯(t),%oy¯(t)) for all r ∈ [0, t]. (4.6)
Note also that Uo(x, t, ε) is just the open ball U(x, ε) in case t ≥ d(o, x) + ε. It follows readily from the following lemma that this topology on Xσ is independent of the choice of basepoint o.
Lemma 4.3. Given o ∈ X, x¯ ∈ ∂σX, ε, t > 0, and p ∈ X, there exists T > 0 such that Up(¯x, ε/4,T ) ⊂ Uo(¯x, ε, t). 25
Proof. It follows from Lemma 4.1 and the construction of the ray %ox¯ in Proposition 4.2 that if T is chosen sufficiently large, depending on d(o, p) and ε, t, and if x := %px¯(T ), then d(%ox(t),%ox¯(t)) ≤ ε/4.
Likewise, for any point y¯ ∈ Xσ, if y := %py¯(T ) and d(p, y) is large enough, then
d(%oy(t),%oy¯(t)) ≤ ε/4.
Now if y¯ ∈ Up(¯x, T, ε/4), that is, d(x, y) < ε/4, then
d(%ox(t),%oy(t)) ≤ 2 d(x, y) < ε/2 by (4.3), provided d(o, x), d(o, y) > t. We conclude that d(%ox¯(t),%oy¯(t)) < ε and thus y¯ ∈ Uo(¯x, t, ε) for sufficiently large T .
Next we equip Xσ with the metric defined by Z ∞ −s Do(¯x, y¯) := d(%ox¯(s),%oy¯(s)) e ds (4.7) 0
(compare Section 8.3.B in [Gro]). We have Do(o, x¯) ≤ 1, with equality if and only if x¯ ∈ ∂σX. For d(%ox¯(t),%oy¯(t)) = a, (4.6) yields Z ∞ Z t Z ∞ a −s −s −s e ds ≤ Do(¯x, y¯) ≤ 2a e ds + 2s e ds, t 2 0 t and it follows easily from these estimates that the metric (4.7) induces the cone topology. Observe also that if d(o, x) = R and y = ϕr(x) for some 0 ≤ r ≤ R, then Z R Z ∞ −s −s −r −R Do(x, y) = (s − r) e ds + (R − r) e ds = e − e . r R In particular, for any σ-ray ξ issuing from o, the curve λ 7→ ξ(− log(1 − λ)), λ ∈ [0, 1), is a unit speed geodesic with respect to Do. Accordingly, for λ ∈ [0, 1], we define the radial retraction ψλ : Xσ → BDo (o, λ) = {y¯| Do(o, y¯) ≤ λ} such that ψ1 = id and ψλ(¯x) := %ox¯(t) for λ < 1 and t := − log(1 − λ). In this latter case, the generalized ray %ox with endpoint x := ψλ(¯x) agrees with ϕt ◦ %ox¯, and since ϕt is 2-Lipschitz we obtain Z ∞ −s Do(ψλ(¯x), ψλ(¯y)) = d(ϕt(%ox¯(s)), ϕt(%oy¯(s))) e ds ≤ 2 Do(¯x, y¯) (4.8) 0 for all x,¯ y¯ ∈ Xσ. Thus ψλ is 2-Lipschitz with respect to Do. Finally, we note that if x¯ ∈ Xσ and λ, µ ∈ [0, 1], then clearly
Do(ψλ(¯x), ψµ(¯x)) ≤ |λ − µ|, (4.9) with equality when Do(o, x¯) ≥ max{λ, µ}. Now we turn to the result mentioned in the introduction. Recall that a metrizable space is an absolute retract if it is a retract of every metrizable space containing it as a closed subspace. 26 CHAPTER 4. BOUNDARY AT INFINITY
Theorem 4.4. Let X be a complete metric space with a consistent bicombing σ. Then Xσ is an absolute retract; in particular, Xσ is contractible and locally contractible. Moreover, ∂σX is a Z-set in Xσ, that is, for every open set U in Xσ the inclusion U \ ∂σX,→ U is a homotopy equivalence.
Proof. To start, we wish to show that Xσ is contractible and locally contractible. The map (¯x, λ) 7→ ψ1−λ(¯x) is continuous on Xσ ×[0, 1] by (4.8) and (4.9) and contracts Xσ to o, so Xσ is contractible. As for the local property, we prove that in fact each of the sets Uo(¯x, t, ε) for x¯ ∈ Xσ and t, ε > 0 (see (4.5)) is contractible. The same map as above, but −t restricted to Uo(¯x, t, ε) × [0, e ], contracts Uo(¯x, t, ε) to the subset U(%ox¯(t), ε) ∩ B(o, t), which as an intersection of balls is σ-convex and hence itself contractible. To prove that ∂σX is a Z-set in Xσ, let an open set U in Xσ be given, and assume that U 6= ∅, Xσ. We want to find a homotopy H : U × [0, 1] → U from the identity on U to a map into U \ ∂σX such that the restriction of H to (U \ ∂σX) × [0, 1] takes values in U \ ∂σX. To this end, we note that the function h: U → R defined by 1 h(¯x) := D (o, x¯) − inf{D (¯x, y¯) | y¯ ∈ X \ U} o 2 o σ is Lipschitz continuous with respect to Do and satisfies h(¯x) < Do(o, x¯) for all x¯ ∈ U because U is open. It is then easy to see that H(¯x, λ) := ψmax{1−λ,h(¯x)}(¯x) serves the purpose. It remains to show that Xσ is an AR (absolute retract). In case X has finite topo- logical dimension, we have
dim(Xσ) = dim(X) < ∞.
Indeed, for every ε ∈ (0, 1), any open covering of the Do-ball BDo (o, 1 − ε) with mesh −1 ≤ ε gives rise, via ψ1−ε, to an open covering of Xσ with mesh ≤ 3ε and the same mul- tiplicity (see, for instance, [BuyS] for the definitions). It is then a standard result that contractible and locally contractible metrizable spaces of finite dimension are absolute retracts (see [Dug1]). However, finite dimension is not needed. Every metric space X with a bicombing is strictly equiconnected, as defined in [Him], so X is an AR by Theorem 4 in that paper (see also [Dug2]). Now, by Corollary 6.6.7 in [Sak] (a result attributed to O. Hanner and S. Lefschetz), Xσ is an ANR (absolute neighborhood re- tract) as it contains X as a homotopy dense subset; alternatively, one can give a short direct argument along the lines of Lemma 1.4 in [Tor]. Thus Xσ is a contractible ANR or, equivalently, an AR (see Corollary 6.2.9 in [Sak]).
As a concluding remark, we first note that ∂σX may be strictly smaller than the set ∂X of asymptote classes of (general) rays in X. For instance, any 1-Lipschitz function 2 f : R+ → R determines a ray t 7→ (t, f(t)) in `∞, thus there are plenty of such rays that are not asymptotic to linear (σ-)rays. Presumably, if X admits different consistent bicombings σ, then ∂σX also depends on the choice of σ. On the other hand, ∂σX agrees with ∂X if, for instance, X has roughly unique geodesics, that is, if there exists a constant δ ≥ 0 such that any two geodesics α, β : [0, 1] → X connecting the same points 27 are within uniform distance at most δ from each other. Indeed, a simple modification of Lemma 4.1 and Proposition 4.2 first shows that for every ray ξ in X there is a σ-ray asymptotic to ξ and issuing from the same point. As in the proof of Lemma 4.3, one can then conclude that distinct bicombings yield homeomorphic boundaries. Clearly every Gromov hyperbolic geodesic metric space X has roughly unique geodesics, and if X is proper, it is well-known that there is also a natural bijection between ∂X and the boundary ∂∞X defined in terms of sequences converging to infinity. This latter fact fails in general for non-proper spaces but remains true, in the presence of a consistent bicombing σ, when X is complete. In fact, for any complete Gromov hyperbolic metric space X with such a bicombing, ∂σX is homeomorphic to ∂∞X with the topology induced by any visual metric (however, the above Do rarely induces a visual metric on the boundary). Chapter 5
Flat planes
0 0 Two lines ξ, ξ in X are called parallel if sups∈R d(ξ(s), ξ (s)) < ∞. And, as before, two rays η, η0 in X are asymptotic if sup d(η(s), η0(s)) < ∞. We use the definition of s∈R+ σ-ray from the previous chapter. As for lines, ξ : R → X will be called a σ-line if its trace is σ-convex; equivalently,
[ξ(s), ξ(t)](λ) = ξ((1 − λ)s + λt) (5.1) for all s, t ∈ R and λ ∈ [0, 1]. Here we used the notation [x, y](t) for σxy(t), and by [x, y] we shall denote the map as well as its trace and will continue to do so without further comment. For two σ-lines ξ, ξ0, the function s 7→ d(ξ(s), ξ0(s)) is convex and non- negative, hence constant in case ξ, ξ0 are parallel. Note that there is a subtle difference between the otherwise identical definitions of σ-ray and σ-line. For the condition (4.1) we restrict to s ≤ t, whereas equation (5.1) must hold for all s, t ∈ R. But since we work with reversible bicombings throughout this chapter, the two properties are equivalent.
5.1 Flat strips and half-planes
Proposition 5.1 (Flat strip). Let X be a metric space with a consistent and reversible 0 bicombing σ. Suppose that ξ, ξ : R → X are two parallel σ-lines with disjoint images. Then the map 0 f : R × [0, 1] → X, f(s, t) = [ξ(s), ξ (s)](t) is an isometric embedding with respect to the metric on R × [0, 1] induced by some norm 2 on R .
If X is CAT(0), the norm is Euclidean. If X is a Busemann space, the norm is strictly convex. If X has finite combinatorial dimension, the norm is polyhedral.
0 0 Proof. For r ∈ R, put h(r) := d(ξ(0), ξ (r)). We have d(ξ(R), ξ (R + r)) = h(r) for every R ∈ R since the left hand side is a convex non-negative bounded function of R, hence 5.1. FLAT STRIPS AND HALF-PLANES 29 constant. We claim that for every pair of points p = (s, t) and p0 = (s + ∆s, t + ∆t) in R × [0, 1] we have (|∆t| h(∆s/∆t) if ∆t 6= 0, d(f(p), f(p0)) = (5.2) |∆s| if ∆t = 0.
There is no loss of generality in assuming ∆t ≥ 0. Suppose first that ∆t > 0, and put r := ∆s/∆t. Let q := (s − tr, 0) and q0 := (s + (1 − t)r, 1) denote the points where the 0 line through p, p intersects R × {0} and R × {1}. Then d(f(q), f(q0)) = d(ξ(s − tr), ξ0(s − tr + r)) = h(r). (5.3) We put η := [ξ(s), ξ0(s)] and η0 := [ξ(s + ∆s), ξ0(s + ∆s)]. By convexity, we get
f(q0) ξ0
f(p0) = η0(t + ∆t)
η(1 − ∆t) ∆t η(t) = f(p) ∆s η0(∆t)
ξ f(q) = ξ(s − tr) that d(f(q), f(p)) = d(ξ(s − tr), η(t)) ≤ t d(ξ(s − r), η(1)) = t h(r). (5.4) Likewise, we have d(f(p0), f(q0)) ≤ (1 − t − ∆t) h(r) (5.5) as well as d(η(0), η0(∆t)) ≤ ∆t h(r) and d(η(1 − ∆t), η0(1)) ≤ ∆t h(r). Hence, by the convexity of λ 7→ d(η(λ), η0(λ + ∆t)) on [0, 1 − ∆t], also d(f(p), f(p0)) = d(η(t), η0(t + ∆t)) ≤ ∆t h(r). (5.6) From (5.3)–(5.6) and the triangle inequality we see that all inequalities derived so far are in fact equalities. In view of (5.6), this shows in particular the first part of (5.2).
The second case follows by continuity from the first, since |r| − h(0) ≤ h(r) ≤ h(0) + |r| for all r ∈ R and hence lim |∆t| h(∆s/∆t) = |∆s|. ∆t→0 0 Now, to conclude the proof, note that h(r) > 0 for all r ∈ R, as ξ and ξ have disjoint 2 images. It then follows readily from (5.2) that there is a norm k · k on R such that 0 0 0 d(f(p), f(p )) = kp − pk for all p, p ∈ R × [0, 1]. Note that the triangle inequality for k · k is just inherited from X. 30 CHAPTER 5. FLAT PLANES
The following example shows that, in general, if we replace ξ by s 7→ ξ(s + a) for some a 6= 0, we may get a different strip in X. Example 5.2. Define piecewise affine functions g, h: R × [0, 1] → R such that 1 t if s ≤ 0, 2 1 g(s, t) = 2 |s − t| if 0 ≤ s ≤ 1, 1 2 (1 − t) if s ≥ 1, 1 2 and h(s, t) = 2 − g(s, 1 − t). Note that g = h outside of (0, 1) , whereas the graphs of g 2 3 and h over [0, 1] bound a simplex Σ in R . Consider the space 3 X := {(s, t, u) ∈ `∞ | g(s, t) ≤ u ≤ h(s, t)} 3 equipped with the metric induced from `∞. X is injective by Lemma A.8 and therefore
ξ0(0)
ξ0(1)
ξ(1) ξ(0)
Figure 5.1: the simplex Σ in X
0 possesses a unique consistent and reversible bicombing σ. The geodesics ξ, ξ : R → X given by ξ(s) := (s, 0, g(s, 0)) and ξ0(s) := (s, 1, g(s, 1)) are two (parallel) σ-lines. It is then not difficult to see (compare the reasoning in Example 8.4) that the strip formed by the segments [ξ(s), ξ0(s + 1)] corresponds to the graph of g, whereas the segments [ξ(s + 1), ξ0(s)] trace out the graph of h. We also see that in Proposition 5.1, for fixed t ∈ (0, 1), the lines s 7→ f(s, t) need not be σ-lines in general: Clearly the lines s 7→ [ξ(s), ξ0(s + 1)](1/2) and s 7→ [ξ(s + 1), ξ0(s)](1/2) in the above example cannot both be σ-lines, and the strip there is never σ-convex as the hull always equals the whole space. For a proper X, however, one may 0 always obtain an embedding f : R × [0, 1] → X such that f(·, 0) = ξ, f(·, 1) = ξ , and s 7→ f(s, t) is a σ-line parallel to ξ and ξ0 for every fixed t ∈ (0, 1). Every accumulation point (w.r.t. compact-open topology) of the family of embeddings fa, a ∈ N, defined 0 through fa(s + ta, t) = [ξ(s), ξ (s + a)](t) for all (s, t) ∈ R × [0, 1], yields such an f. We now proceed to an existence result for embedded flat half-planes, which will be instrumental in the proof of Theorem 5.5. We need the following analogue of the Tits cone in the case of CAT(0) spaces. Let ∂σX denote the set of equivalence classes of mutually asymptotic σ-rays like we defined them at the beginning of Chapter 4. For (a, ξ), (b, η) ∈ R+ × ∂σX, we put 1 d∞((a, ξ), (b, η)) := lim d(ξ(aλ), η(bλ)). λ→∞ λ 5.1. FLAT STRIPS AND HALF-PLANES 31
Note that the limit exists by convexity, and |a − b| ≤ d∞((a, ξ), (b, η)) ≤ a + b. This defines a pseudometric d∞ on R+ × ∂σX, and the respective quotient metric space (shrinking {0}×∂σX to a single point) is a metric cone over ∂σX (compare [Bal], p. 38). In particular, for a > 0, d∞((a, ξ), (a, η)) = a d∞((1, ξ), (1, η)), and this is zero if and only ξ and η are asymptotic. The following result should now be compared with Proposition II.4.2 in [Bal] and Proposition II.9.8 and Corollary II.9.9 in [BriH].
Proposition 5.3 (Flat half-plane). Let X be a metric space with a consistent and reversible bicombing σ. Suppose that ξ : R → X is a σ-line and for every s ∈ R, ηs : R+ → X is the σ-ray with ηs(0) = ξ(s) asymptotic to η := η0. Then, for all a, b > 0, the function s 7→ d(ξ(s + a), ηs(b)) is non-decreasing on R with limit
lim d(ξ(s + a), ηs(b)) = d∞((a, ξ), (b, η)). (5.7) s→∞
Furthermore, if for every a ∈ R the function s 7→ d(ξ(s + a), ηs(1)) is constant on R and non-zero, the map f : R × R+ → X, f(s, t) := ηs(t), is an isometric embedding with respect to the metric on R × R+ induced by some norm 2 on R . Proof. Let a, b > 0. First we show that for all 0 < r ≤ λ ≤ r + 1, 1 d(ξ(ar + a), η (b)) ≥ d(ξ(aλ), η(bλ)). (5.8) ar λ
Since ηar and the ray t 7→ η(br + t) are asymptotic, we have
d(ηar(b), η(br + b)) ≤ d(ηar(0), η(br)) = d(ξ(ar), η(br)). (5.9)
It follows that
d(ξ(ar + a), ηar(b)) ≥ d(ξ(ar + a), η(br + b)) − d(ξ(ar), η(br)) r + 1 r ≥ − d(ξ(aλ), η(bλ)), λ λ which is (5.8). Putting λ = 1 we get d(ξ(ra + a), ηra(b)) ≥ d(ξ(a), η0(b)). Likewise, for all s ∈ R and 0 < r ≤ 1,
d(ξ(s + ra + a), ηs+ra(b)) ≥ d(ξ(s + a), ηs(b)), so s 7→ d(ξ(s + a), ηs(b)) is non-decreasing. Furthermore, for all s ∈ R and λ ≥ 1, we have 1 d(ξ(s + a), η (b)) ≤ d(ξ(s + aλ), η (bλ)). s λ s Together with (5.8), this gives (5.7). 32 CHAPTER 5. FLAT PLANES
For the second part of the proposition we have that s 7→ d(ξ(s+a), ηs(1)) is constant for every a ∈ R and that these values h(a) := d(ξ(a), η0(1)) are all positive. We first claim that d(ξ(s + ta), ηs(t)) = t h(a) (5.10) for all t ≥ 0. The left hand side is convex as a function of t, thus it suffices to show this equality for 0 ≤ t ∈ Z. For t = 0, 1, (5.10) clearly holds. Consequently, by convexity, d(ξ(s + ta), ηs(t)) ≥ t h(a) for all t > 1. The reverse inequality for 1 < t ∈ Z follows by the triangle inequality since
d(ηs+ka(t − k), ηs+ka−a(t − k + 1)) ≤ d(ηs+ka(0), ηs+ka−a(1)) = h(a) for k = t, t − 1,..., 1 (compare (5.9)). Next, we claim that for every pair of points 0 p = (s, t) and p = (s + ∆s, t + ∆t) in R × R+ we have
(|∆t| h(−∆s/∆t) if ∆t 6= 0, d(f(p), f(p0)) = |∆s| if ∆t = 0, similarly as in the proof of Proposition 5.1. To show this, suppose without loss of generality that ∆t ≥ 0. Let first ∆t > 0, and put a := ∆s/∆t and q := (s − ta, 0). Then (5.10) yields
0 d(f(p), f(p )) = d(ηs(t), ηs+∆s(t + ∆t))
≤ d(ηs(0), ηs+∆s(∆t)) = ∆t h(−a) as well as d(f(q), f(p)) = t h(−a) and d(f(q), f(p0)) = (t + ∆t) h(−a). This gives d(f(p), f(p0)) = ∆t h(−a), as claimed. The rest of the proof follows as in Proposi- tion 5.1.
5.2 The Flat Plane Theorem
We now turn to Theorem 5.5. Recall that a metric space X is δ-hyperbolic, for some constant δ ≥ 0, if for every quadruple (w, x, y, z) ∈ X4,
d(w, y) + d(x, z) ≤ max{d(w, x) + d(y, z), d(w, z) + d(x, y)} + 2δ.
If such a δ exists, X is said to be hyperbolic. As is well known, for a geodesic metric space this is equivalent to saying that geodesic triangles are slim, in an appropriate sense. It also suffices to consider triangles whose sides are given by a fixed bicombing.
Lemma 5.4. Let X be a metric space with a map that selects for every pair of points x, y ∈ X a geodesic segment [x, y] = [y, x] ⊂ X connecting them. If for every triple 3 δ (x, y, z) ∈ X the segment [x, z] is contained in the closed 2 -neighborhood of [x, y]∪[y, z], then X is δ-hyperbolic. 5.2. THE FLAT PLANE THEOREM 33
4 δ Proof. Let (w, x, y, z) ∈ X . The union of the closed 2 -neighborhoods of [x, y] and [y, z] δ covers [x, z], and also the union of the closed 2 -neighborhoods of [z, w] and [w, x] covers [x, z]. It follows that there is either a pair of points x0 ∈ [x, y] and z0 ∈ [z, w] with d(x0, z0) ≤ δ or a pair of points y0 ∈ [y, z] and w0 ∈ [w, x] with d(y0, w0) ≤ δ. In the first case,
d(w, y) + d(x, z) ≤ d(w, z0) + δ + d(x0, y) + d(x, x0) + δ + d(z0, z) = d(w, z) + d(x, y) + 2δ.
Similarly, in the second case, d(w, y) + d(x, z) ≤ d(w, x) + d(y, z) + 2δ.
In particular, a non-hyperbolic X with a bicombing σ contains a sequence of fatter and fatter σ-triangles. The following argument then uses a ruled surface construction together with the cocompact isometric action to produce a collection of mutually asymp- totic rays as in Proposition 5.3. This differs from the strategy in [Bow] and is inspired by the proof for CAT(0) spaces in [Bri, BriH], although we make no use of angles.
Theorem 5.5 (Flat Plane). Let X be a proper metric space with a consistent and reversible bicombing σ and with cocompact isometry group. Then X is hyperbolic if and only if X does not contain an isometrically embedded normed plane.
Proof. If X contains an isometrically embedded normed plane, then clearly X cannot be hyperbolic. Suppose now that X is not hyperbolic. By Lemma 5.4 there are sequences 1 2 3 1 3 of points yk, yk, yk ∈ X and pk ∈ [yk, yk] such that
1 2 2 3 B(pk, k) ∩ ([yk, yk] ∪ [yk, yk]) = ∅ (5.11)
i i i+1 for all integers k ≥ 1. Put rk(·) := d(pk, ·). For i = 1, 2, let xk be a point in [yk, yk ] with i i minimal distance to pk, and let ξk : [0, rk(xk)] → X be a unit speed parametrization of the i i segment [pk, xk] from pk to xk. Then, for every pair (i, j) ∈ {(1, 1), (1, 2), (2, 2), (2, 3)}, we define the “ruled surface”
i,j i j ∆k : [0, rk(xk)] × [0, rk(yk)] → X
i,j i i i,j so that ∆k (·, 0) = ξk and, for each s ∈ [0, rk(xk)], ∆k (s, ·) is a constant speed parame- i j i j trization of the segment [ξk(s), yk] from ξk(s) to yk. Thus
i j i,j i,j 0 d(ξk(s), yk) 0 0 d(∆k (s, t), ∆k (s, t )) = j |t − t | ≤ 2|t − t |, rk(yk)
i j i j j i because d(ξk(s), yk) ≤ rk(xk) + rk(yk) ≤ 2rk(yk) by the choice of xk. Note also that by convexity, i,j i,j 0 i i 0 0 d(∆k (s, t), ∆k (s , t)) ≤ d(ξk(s), ξk(s )) = |s − s |. (5.12) 34 CHAPTER 5. FLAT PLANES
i,j 2 It follows that each ∆k is 2-Lipschitz, where here and below we equip R with the 0 i 0 l1-metric. Furthermore, putting s := rk(xk), we notice that for 0 ≤ r ≤ s ≤ s and j 0 ≤ t ≤ rk(yk), i i,j i,j 0 i,j i,j 0 d(ξk(r), ∆k (s, t)) ≥ rk(∆k (s , t)) − r − d(∆k (s, t), ∆k (s , t)) i 0 ≥ rk(xk) − r − (s − s) = s − r (5.13)
i by the triangle inequality, the choice of xk, and (5.12). Now we choose a sequence of isometries γk of X so that γk(pk) ∈ K for all k and i j for some fixed compact set K. By (5.11), rk(xk), rk(yk) > k. Since X is proper, we i,j can extract a sequence k(l) so that each of the four sequences γk(l) ◦ ∆k(l) converges uniformly on compact sets, as l → ∞, to a 2-Lipschitz map
i,j f : R+ × R+ → X i i,j j i,j with boundary rays ξ := f (·, 0) and η := f (0, ·). Furthermore, for every s ∈ R+, i,j i,j j i,j ηs := f (s, ·) is a ray asymptotic to η , so f is in fact 1-Lipschitz. From the construction we also have that d(η1(t), η3(t)) = 2t for all t ≥ 0, in particular η1, η3 are non-asymptotic. Hence, there is at least one pair (i, j) such that ξi, ηj are non- i,j i j i,j asymptotic. We put f := f , ξ := ξ , η := η , and ηs := ηs for some such pair. We claim that for all a ∈ R and b > 0, the limit
L(a, b) := lim d(ξ(s + a), ηs(b)) s→∞ exists and is strictly positive. Clearly L(0, b) = b. If a>0, then L(a, b) = d∞((a, ξ), (b, η)) > 0 by the first part of Proposition 5.3 and since ξ, η are non-asymptotic. If a < 0, the same result still shows that s 7→ d(ξ(s + a), ηs(b)) is non-increasing on [|a|, ∞), so the limit exists, and L(a, b) ≥ |a| as a consequence of (5.13). Next, for every integer l ≥ 1 we define the 1-Lipschitz map
fl :[−l, ∞) × R+, fl(s, t) := f(l + s, t) = ηl+s(t).
Then we choose isometries γ¯l of X so that (¯γl ◦ fl)(0, 0) ∈ K for all l and for some fixed compact set K. As above, there exists a subsequence l(m) such that the sequence γ¯l(m) ◦ fl(m) converges uniformly on compact sets to a 1-Lipschitz map ¯ f : R × R+ → X ¯ ¯ ¯ with boundary line ξ := f(·, 0) and mutually asymptotic rays η¯s := f(s, ·) for s ∈ R. For every a ∈ R and b > 0, we now have that
d(ξ¯(s + a), η¯s(b)) = L(a, b) > 0 for all s ∈ R. Hence, by Proposition 5.3, f¯ is an isometric embedding with respect to 2 some norm on R . Using once more that X is cocompact, we then conclude that X contains an isometrically embedded normed plane. 5.2. THE FLAT PLANE THEOREM 35
It is clear that if X is a CAT(0) or a Busemann space, then this property is inherited by any isometrically embedded normed plane, thus the corresponding norm must be Euclidean or strictly convex, respectively. We briefly discuss another variant of Theo- rem 5.5, which happens to have a very short proof, without reference to bicombings. Recall the notion of injective metric space from Appendix A. Given any metric space Q = {w, x, y, z} of cardinality four, suppose that c := d(w, y)+d(x, z) is not less than the maximum of a := d(w, x) + d(y, z) and b := d(w, z) + d(x, y). The injective hull (or the tight span [Dre]) of Q is isometric to the (possibly degenerate) rectangle [0, (c − a)/2] × 2 [0, (c − b)/2] in `1 with four segments of appropriate lengths attached at the corners, where the terminal points of these segments correspond to Q (see Fig. A1 on p. 336 in [Dre]). Now the δ-hyperbolicity of Q means precisely that the width (the minimum of the two side lengths) of this `1-rectangle is not bigger than δ. This has the following easy consequence.
Theorem 5.6. A proper, cocompact injective metric space X is hyperbolic if and only 2 2 if X does not contain an isometric copy of `1 or, equivalently, of `∞. Proof. Suppose that X is not hyperbolic. Then, by the above discussion, for arbitrarily large δ > 0 there exists a quadruple Q ⊂ X whose injective hull contains an isometric 2 copy of [0, δ] × [0, δ] ⊂ `1. Since X is injective, this `1-square embeds isometrically into X by the respective property of the injective hull. From a sequence of such squares with side lengths tending to infinity we obtain an isometric embedding of the entire `1-plane, using the fact that X is proper and cocompact. Chapter 6
Barycenters
Here we develop the tool that forms a crucial component of many proofs to come. In [EsSH], Es-Sahib and Heinich introduced an elegant barycenter construction for Buse- mann spaces, which was reviewed and partly improved in a recent paper by Navas [Nav].
6.1 The construction
The construction and proofs translate almost verbatim to spaces with reversible bicomb- ings. As such spaces may lack unique midpoints, the only modification required is to set bar2(x, y) := σxy(1/2). For finite subsets, the result is as follows.
Theorem 6.1. For every complete space X with reversible bicombing σ and every n ∈ N n there is a barycenter map barn : X → X with the following properties:
(i) barn(x1, . . . , xn) lies in the σ-convex hull of {x1, . . . , xn},
(ii) barn is permutation invariant, i.e. for any permutation π ∈ Sn we have barn(x1, . . . , xn) = barn(xπ(1), . . . , xπ(n)),
(iii) γ(barn(x1, . . . , xn)) = barn(γ(x1), . . . , γ(xn)) for every isometry γ of X provided σ is γ-equivariant,