Introduction
The main theorem. We study the large-scale geometry of right-angled Coxeter groups. A right-angled Coxeter group (RACG) W Γ W_{\Gamma} W Γ is determined by a finite simplicial graph Γ = ( V , E ) \Gamma= (V, E) Γ = ( V , E ) . Its generators are the vertices of Γ \Gamma Γ ; each is an involution, and two generators commute precisely when the corresponding vertices are adjacent. A central theme is to determine which features of the coarse geometry of W Γ W_{\Gamma} W Γ are encoded by the combinatorics of Γ \Gamma Γ . For example, W Γ W_{\Gamma} W Γ is hyperbolic if and only if Γ \Gamma Γ has no induced 4-cycle [58 ] ; it is one-ended if and only if Γ \Gamma Γ is connected, not a clique, and has no separating clique [63 , 30 ] . Other graph-theoretic characterizations relate to divergence, thickness, relative hyperbolicity, the Morse boundary, and the structure of quasi-flats [5 , 34 , 11 , 49 , 50 , 25 , 44 , 12 ] . For a broader overview, we refer to the survey [29 ] .
We add a new entry to this dictionary. Our main result characterizes, purely in terms of Γ \Gamma Γ , when one has connectivity of the Gromov boundary of a particular hyperbolic space associated to W Γ W_\Gamma W Γ . The relevant hyperbolic space is called the maximized hyperbolic space , which was introduced by Abbott–Behrstock–Durham in [1 ] .
The geometry of W Γ W_\Gamma W Γ is carried by the Davis complex X = X Γ X=X_\Gamma X = X Γ , the flag cube complex on the Cayley graph of W Γ W_\Gamma W Γ with the standard generating set V ( Γ ) V(\Gamma) V ( Γ ) . It is a CAT(0) cube complex on which W Γ W_\Gamma W Γ acts properly and cocompactly. When Γ \Gamma Γ contains an induced 4-cycle, W Γ W_\Gamma W Γ contains a Z 2 \mathbb{Z}^2 Z 2 subgroup and X X X contains flats, so X X X is far from being hyperbolic. These flats lie inside standard product regions : convex subcomplexes g X Λ gX_\Lambda g X Λ , where g ∈ W Γ g\in W_\Gamma g ∈ W Γ and Λ ⊂ Γ \Lambda\subset\Gamma Λ ⊂ Γ is a nontrivial join subgraph which has a decomposition Λ = Λ 0 ∗ Λ 1 \Lambda=\Lambda_0*\Lambda_1 Λ = Λ 0 ∗ Λ 1 in which every vertex in Λ 0 \Lambda_0 Λ 0 is adjacent to every vertex in Λ 1 \Lambda_1 Λ 1 and neither Λ i \Lambda_i Λ i is a clique. For an RACG, the maximized hyperbolic space is obtained by collapsing standard product regions: X ^ = X ^ Γ \widehat{X}=\widehat{X}_\Gamma X = X Γ is the graph with vertex set X ( 0 ) X^{(0)} X ( 0 ) in which two vertices are connected by an edge when they are adjacent in X X X or lie in a common standard product region (Definition 2.22 ). Coning off these regions removes the non-hyperbolic directions and produces a δ \delta δ -hyperbolic space X ^ \widehat{X} X [12 Corollary 2.16] . Moreover, X ^ \widehat{X} X has infinite diameter precisely when Γ \Gamma Γ is neither a clique nor a nontrivial join; see Lemma 3.4 . If Γ \Gamma Γ is a clique, X Γ X_\Gamma X Γ is bounded; if Γ \Gamma Γ is a nontrivial join, then X Γ X_\Gamma X Γ is a single standard product region, hence X ^ \widehat{X} X has bounded diameter. Thus ∂ X ^ \partial\widehat{X} ∂ X is empty in both cases.
Our main theorem shows that the connectivity of ∂ X ^ \partial\widehat{X} ∂ X can be characterized from the defining graph Γ \Gamma Γ .
Theorem 1.1. Let W Γ W_\Gamma W Γ be a right-angled Coxeter group with defining graph Γ \Gamma Γ , and let X ^ Γ \widehat{X}_\Gamma X Γ be its maximized hyperbolic space. The Gromov boundary ∂ X ^ Γ \partial\widehat{X}_\Gamma ∂ X Γ is nonempty and connected if and only if Γ \Gamma Γ satisfies the following:
(a) Γ \Gamma Γ is connected, is not a clique, and has no separating clique; and
(b) Γ \Gamma Γ has no separating nontrivial join; and
(c) for every vertex v ∈ Γ v\in\Gamma v ∈ Γ , the link of v v v is not contained in a nontrivial join.
Furthermore, when ∂ X ^ Γ \partial\widehat{X}_\Gamma ∂ X Γ is connected, it is linearly connected with respect to a visual metric.
Here, a metric space is linearly connected if there is a constant L ≥ 1 L\ge1 L ≥ 1 such that for any two points x , y x,y x , y , there is a compact connected set containing x , y x,y x , y with diameter at most L d ( x , y ) Ld(x,y) L d ( x , y ) .
The connectedness implication also follows from Cordes–Levcovitz’s theorem on the connectedness of Morse boundaries of Coxeter groups [25 ] , together with the density of the Morse-boundary image in ∂ X ^ Γ \partial\widehat{X}_\Gamma ∂ X Γ . We prove this implication in Corollary 4.8 , within a general framework that also applies to other groups. Our direct geometric argument gives the quantitative conclusion of linear connectedness.
The three conditions have transparent group-theoretic and geometric meanings: (a) says that W Γ W_\Gamma W Γ is one-ended; (b) says that W Γ W_\Gamma W Γ does not split as an amalgam over a special unbounded product subgroup; and (c) says no hyperplane in the Davis complex X Γ X_\Gamma X Γ is contained in a standard product region. The geometry of the Davis complex X X X is closely related to its standard product regions and the maximized hyperbolic space X ^ \widehat{X} X . Given a pair of vertices in X X X , their distance in X X X can be estimated by the sum of the distances of their images in standard product regions under the gate map (a closest point projection) and their images in the maximized hyperbolic space under the identity map. This property is captured by the distance formula, which is a result from [10 ] ; see Theorem 2.24. More generally, right-angled Coxeter groups are hierarchically hyperbolic groups (§ 2.6 ), whose geometries are encoded, by coarse Lipschitz projections, in a collection of hyperbolic spaces with a uniform hyperbolicity constant which are organized by three relations: nesting, orthogonality and transversality . Hierarchically hyperbolic groups/spaces are a generalization of hyperbolic groups in the manner of Masur–Minsky’s description of a mapping class group through the curve graphs of its essential subsurfaces [10 ] . Each HHG/HHS may admit several hierarchically hyperbolic structures. For RACGs, the maximized structure in the sense of Abbott–Behrstock–Durham [1 Theorem 3.7] has X ^ \widehat{X} X as its nested-maximal (top-level) hyperbolic space, and W Γ W_\Gamma W Γ acts on X ^ \widehat{X} X as its largest and universal acylindrical action [1 Theorem A] . Moreover, a quasi-isometry on W Γ W_\Gamma W Γ induces a quasi-isometry on its maximized hyperbolic space X ^ \widehat{X} X [37 Theorem 6.10] , hence a homeomorphism of the Gromov boundary ∂ X ^ \partial\widehat{X} ∂ X . Consequently, the connectivity of ∂ X ^ \partial\widehat{X} ∂ X is a quasi-isometry invariant of W Γ W_\Gamma W Γ , and Theorem 1.1 characterizes this invariant through the combinatorics of Γ \Gamma Γ .
We record some applications of Theorem 1.1 . First, we provide new examples of HHGs with connected Gromov boundary on their nested-maximal hyperbolic spaces; see Figure 1 . The classical examples whose top-level hyperbolic space has connected Gromov boundary include one-ended hyperbolic groups and mapping class groups of surfaces of complexity at least two. A recent work of Choi–Talbott also provides infinitely many examples from random quotients of HHGs [28 ] .
Figure 1. The maximized hyperbolic space X ^ Γ 1 \widehat{X}_{\Gamma_1} X Γ 1 of W Γ 1 W_{\Gamma_1} W Γ 1 has connected Gromov boundary, but the Gromov boundary of X ^ Γ 2 \widehat{X}_{\Gamma_2} X Γ 2 is disconnected since the link of the vertex v v v is contained in an induced 4-cycle in Γ 2 \Gamma_2 Γ 2 .
Second, X ^ \widehat{X} X distinguishes certain right-angled Coxeter groups up to quasi-isometry. The quasi-isometry classification of RACGs, and the question of which RACGs are quasi-isometric or commensurable to RAAGs, have received considerable attention [7 DL24, 21 ] . Because connectivity of ∂ X ^ Γ \partial\widehat{X}_\Gamma ∂ X Γ is a quasi-isometry invariant of W Γ W_\Gamma W Γ , Theorem 1.1 immediately gives the following.
Corollary 1.2. Let W Γ W_\Gamma W Γ and W Λ W_\Lambda W Λ be right-angled Coxeter groups. If Γ \Gamma Γ satisfies conditions (a)–(c) of Theorem 1.1 but Λ \Lambda Λ does not, then W Γ W_\Gamma W Γ and W Λ W_\Lambda W Λ are not quasi-isometric.
Every RAAG A Γ A_\Gamma A Γ is hierarchically hyperbolic. When Γ \Gamma Γ is a connected graph with more than one vertex, A Γ A_\Gamma A Γ admits universal [47 ] and largest [2 ] acylindrical actions on its extension graph. This graph is quasi-isometric to the contact graph which is a quasi-tree [39 ] , so its boundary is totally disconnected; see also Remark 3.7. We therefore obtain the following corollary.
Corollary 1.3. If the defining graph Γ \Gamma Γ satisfies conditions (a)–(c) of Theorem 1.1 , then W Γ W_{\Gamma} W Γ is not quasi-isometric to any right-angled Artin group.
We remark that Corollaries 1.2 and 1.3 also follow from Cordes–Levcovitz’s theorem on the connectedness of Morse boundaries [25 ] .
Third, in the appendix [37 Theorem A.1] , Russell showed that if the nested-maximal hyperbolic space C S CS C S in an HHG G G G with a suitable HHS structure is one-ended, then the group of quasi-isometries of C S CS C S is isomorphic to the group of quasi-isometries of G G G . Combining Russell’s result with Corollary 8.9, we obtain the following application to right-angled Coxeter groups.
Corollary 1.4. Let W Γ W_{\Gamma} W Γ be a right-angled Coxeter group whose defining graph Γ \Gamma Γ satisfies conditions (a)–(c) in Theorem 1.1 and X ^ Γ \widehat{X}_{\Gamma} X Γ be the maximized hyperbolic space of W Γ W_{\Gamma} W Γ . Then Q I ( W Γ ) QI(W_{\Gamma}) Q I ( W Γ ) is isomorphic to Q I ( X ^ Γ ) QI(\widehat{X}_{\Gamma}) Q I ( X Γ ) .
Fourth, combining the visual decomposition theorem of Mihalik–Tschantz [63 ] with Theorem 10.2 relates algebraic splittings of W Γ W_{\Gamma} W Γ to the topology of ∂ X ^ \partial\widehat{X} ∂ X .
Corollary 1.5. A one-ended right-angled Coxeter group W Γ W_{\Gamma} W Γ splits over a subgroup of an unbounded product subgroup if and only if the Gromov boundary ∂ X ^ \partial\widehat{X} ∂ X is empty or disconnected.
Corollary 1.6. In the class of one-ended right-angled Coxeter groups, splitting over a subgroup of an unbounded product subgroup is a quasi-isometric invariant.
Along the way, we establish several properties that X ^ \widehat{X} X shares with curve graphs. In Section 6 we exhibit two RACGs whose maximized hyperbolic spaces contain dead ends. In § 7 , using the bounded geodesic image theorem, we show that ∂ X ^ \partial\widehat{X} ∂ X is visible, i.e., any two points in X ^ ∪ ∂ X ^ \widehat{X}\cup\partial\widehat{X} X ∪ ∂ X are joined by a geodesic (Lemma 7.4 and Proposition 7.5 ), X ^ \widehat{X} X is geodesic almost-extendable (Proposition 7.7), and when ∂ X ^ \partial\widehat{X} ∂ X is nonempty, ∂ X ^ \partial\widehat{X} ∂ X is non-compact if and only if Γ \Gamma Γ contains an induced 4-cycle (Theorem 7.8 ).
Related work
Boundaries of curve graphs and related complexes. Theorem 1.1 belongs to a broader program relating the topology of boundaries of hyperbolic spaces to the geometry of groups. For mapping class groups, after work of Leininger–Schleimer [52 ] and Leininger–Mj–Schleimer [51 ] , Gabai [36 ] proved that the Gromov boundary of the curve graph of a surface of complexity at least two is path-connected and locally path-connected. Wright sharpened this to linear connectedness [Wri24]. In related settings, Long–Tan showed that the Gromov boundary of the fine curve graph is path connected when the genus of the surface is at least 3 and it is linearly connected under certain choices of the visual metric [53 ] . Bestvina–Chaika–Hensel showed that the Gromov boundary of the free factor complex of a free group of rank at least 18 is path connected and locally path connected [6 ] .
Boundaries of Coxeter and CAT(0) groups. One may instead study the visual (CAT(0)) boundary and the Morse boundary of Coxeter groups. Mihalik–Ruane–Tschantz [61 ] proved that the visual boundary of a one-ended RACG is locally connected when there are no product separators (Definition 2.9 ) and no virtual factor separators (Definition 2.10 ).
Cordes–Levcovitz [25 ] proved that a one-ended RACG with wide-avoidant defining graph has connected and locally connected Morse boundary. These two graph conditions coincide, and together they are equivalent to conditions (a)–(c) of Theorem 1.1 (see Lemma 2.14 for more details). The fan-and-filter construction used in § 8 , an essential ingredient in our proof, builds on [61 , 25 ] .
The Morse boundary ∂ M W Γ \partial_M W_\Gamma ∂ M W Γ admits a natural continuous injection into the Gromov boundary ∂ X ^ \partial\widehat{X} ∂ X of the maximized hyperbolic space; see [1 Theorem 6.6] and [68 Lemma A.6] . For non-virtually-cyclic HHGs containing a Morse element, Abbott–Behrstock–Russell established density of the Morse-boundary image in the HHS boundary [3 Corollary 5.10 and Remark 5.11] .
In § 4 , we record a general density statement for cobounded Morse-recognizing actions on hyperbolic spaces (Proposition 4.2 ). The resulting continuous boundary injection has a dense image, so connectedness of the Morse boundary implies connectedness of the Gromov boundary. We apply the result to the hyperbolic models for finite-rank Coxeter groups constructed in [20 ] . Combined with [25 Theorem B] , this gives an alternative proof of the connectedness implication of Theorem 1.1 .
Splittings and boundary topology. The topology of a group boundary is classically tied to its algebraic splittings. By Stallings’ theorem on the number of ends [71 ] , a finitely generated group splits over a finite subgroup if and only if it has more than one end. Bowditch showed that a one-ended hyperbolic group, which is not virtually cocompact Fuchsian, splits over a two-ended subgroup if and only if its Gromov boundary has a local cut point [17 ] . These were extended to CAT(0) groups by Papasoglu–Swenson [66 , 67 ] , to relatively hyperbolic groups by Groff [38 ] , and for splittings over non-parabolic two-ended subgroups by Haulmark [42 ] in terms of local cut points of the Bowditch boundary. The relation between splittings and connectivity of the visual boundary of CAT(0) groups was studied in [59 , 60 , 43 , 61 ] . Moreover, Papasoglu showed that a finitely presented one-ended group which is not commensurable to a surface group splits over a two-ended subgroup if and only if its Cayley graph is separated by a quasi-line [64 ] .
Outline of the proof
Outline of the proof. For the forward direction, first consider the bounded cases. If Γ \Gamma Γ is a clique or a nontrivial join, then X ^ \widehat{X} X has bounded diameter and hence empty boundary. In the remaining cases, failure of one of conditions (a)–(c) gives a splitting over a finite subgroup or over a subgroup of an unbounded product subgroup. Bass–Serre theory then detects separation of X ^ \widehat{X} X by a bounded set, and the boundary is disconnected. See § 3 for more details.
In the backward direction, we show that under the three conditions (a)–(c) on the defining graph Γ \Gamma Γ , the Gromov boundary ∂ X ^ \partial\widehat{X} ∂ X is linearly connected (Definition 2.21 ), which implies path connectedness and local path connectedness. Fix a point o o o in X ^ \widehat{X} X and take two points a , b a,b a , b in the boundary ∂ X ^ \partial\widehat{X} ∂ X . In Lemma 7.4 , we show that there are geodesic rays α , β \alpha,\beta α , β based at o o o converging to a , b a,b a , b , respectively. Then the key idea is to push paths along α ∪ β \alpha\cup\beta α ∪ β away from o o o . To achieve this, we iteratively use Proposition 8.8 , which states that any two points x x x and y y y on a sphere of radius R R R centered at o o o in X ^ \widehat{X} X can be joined by a path lying in the union of the spheres of radii R R R and R + 1 R+1 R + 1 , with the diameter bounded from above by a uniform constant multiple of d X ^ ( x , y ) d_{\widehat{X}}(x,y) d X ( x , y ) .
The proof of Proposition 8.8 uses the cubical structure of the Davis complex X X X . Identify the vertex sets of X X X and X ^ \widehat{X} X , and choose geodesic segments γ x \gamma_x γ x and γ y \gamma_y γ y from o o o to x x x and y y y . We construct a truncated filter : a van Kampen diagram over W Γ W_\Gamma W Γ whose boundary maps to the union of γ x \gamma_x γ x , γ y \gamma_y γ y , and a path from x x x to y y y in X ( 1 ) X^{(1)} X ( 1 ) (Figure 9 ). The image of the third path in X ^ \widehat{X} X lies in the union of the spheres of radii R R R and R + 1 R+1 R + 1 centered at o o o .
Figure 9. The first two levels of a filter. The edges of the spanning tree T 2 T_2 T 2 are solid, and the deleted top-left edges are dashed.
Mihalik–Ruane–Tschantz introduced filters between geodesic rays to study the visual boundary of RACGs [61 ] ; Cordes–Levcovitz later generalized the construction to study Morse boundaries of Coxeter groups [25 ] . Here we instead construct truncated filters between geodesic segments. A (truncated) filter contains a directed tree whose directed paths map to geodesics in X ( 1 ) X^{(1)} X ( 1 ) . Furthermore, paths in the directed tree after their last intersection with the side paths have uniformly bounded wide subpaths, with a bound depending only on Γ \Gamma Γ . Lemma 5.11 converts this estimate into progress in X ^ \widehat{X} X . Consequently, the truncation requires only finitely many fan attachments.
Open questions
Mapping class groups are quasi-isometrically rigid [14 , 41 ] (for a direct approach using hierarchical hyperbolicity see [12 , 55 ] ). The proof factors through two steps: a quasi-isometry of the mapping class group induces an automorphism of the curve graph, and, by Ivanov’s theorem [46 ] , every such automorphism is realized by a mapping class. A motivation for our work is to find hierarchically hyperbolic groups whose nested-maximal hyperbolic space might play the role of the curve graph in such an argument. The contact graph of the standard hierarchical structure of RACGs is unsuitable: as a quasi-tree it can have an enormous automorphism group [10 Section 7] . The maximized hyperbolic space X ^ \widehat{X} X , when ∂ X ^ \partial\widehat{X} ∂ X is connected, is a more promising candidate, and we expect a quasi-isometric model of it to have a countable automorphism group and to be useful to study the quasi-isometric rigidity problem. This motivates the following question.
Question 1.7. Is W Γ W_\Gamma W Γ quasi-isometrically rigid when Γ \Gamma Γ satisfies conditions (a), (b), and (c)?
In view of Corollary 1.5, we expect that the connectedness of the boundary can be characterized by the group splittings for general hierarchically hyperbolic groups.
Conjecture 1.8. Let G G G be a one-ended hierarchically hyperbolic group with maximized hyperbolic space X ^ \widehat{X} X . Suppose X ^ \widehat{X} X has infinite diameter. Then G G G splits over a subgroup of an unbounded product subgroup if and only if ∂ X ^ \partial\widehat{X} ∂ X is disconnected.
Bonk–Kleiner [13 ] showed that every one-ended hyperbolic group contains a quasi-isometrically embedded hyperbolic plane, and Mackay–Sisto [62 ] extended this to relatively hyperbolic groups with nilpotent peripheral subgroups. We ask an analogous question for HHGs.
Question 1.9. Let G G G be a one-ended hierarchically hyperbolic group with maximized hyperbolic space X ^ \widehat{X} X . Suppose the Gromov boundary of X ^ \widehat{X} X is connected. Does X ^ \widehat{X} X contain a quasi-isometrically embedded hyperbolic plane?
A quasi-isometrically embedded hyperbolic plane in X ^ \widehat{X} X induces a quasisymmetrically embedded circle in ∂ X ^ \partial\widehat{X} ∂ X . In light of the characterization of quasisymmetric images of circles and arcs by Tukia–Väisälä [72 ] , the following question is a natural step.
Question 1.10. When ∂ X ^ \partial\widehat{X} ∂ X is connected, does it contain a quasicarc?
Organization
Section 2 reviews RACGs, equivalent graph-theoretic conditions, CAT(0) cube complexes, Gromov boundaries, and the maximized HHS structure. Section 3 proves the forward direction of Theorem 1.1 . Section 4 records the density statement for Morse-recognizing actions and its connectedness applications. Section 5 develops the geometry of X ^ Γ \widehat{X}_{\Gamma} X Γ , beginning with taut geodesics, and Section 6 gives two examples of dead ends. Section 7 proves bounded geodesic image and establishes visibility and noncompactness of ∂ X ^ \partial\widehat{X} ∂ X . Section 8 develops filters in X X X and uses them to establish controlled connectivity of spheres in X ^ \widehat{X} X . Section 9 establishes linear connectedness, completing the proof of the main theorem. Finally, Section 10 applies the main theorem to group splittings.
Acknowledgements
I am grateful to my advisor, Jason Behrstock, for introducing me to hierarchically hyperbolic spaces and for his continued support and encouragement throughout this project. I thank Jacob Russell for encouraging me to think about this problem at the Advances in Hierarchical Hyperbolicity workshop, and for his continued interest in this project. His question about the relationship between the Morse boundary and the Gromov boundary, raised during the Buckeye Metric Geometry Workshop II, led to the material in Section 4. I also thank Michael Mihalik for his interest in this work and, in particular, for discussions of Lemma 2.14 and Remark 2.15 during the Vanderbilt Colloquium Series on Groups, Geometry, and Dynamics.
I was partially supported by a CUNY Graduate Center Dissertation Fellowship during this project.
AI disclosure
AI tools were used to proofread the manuscript.
Background
We review the background on right-angled Coxeter groups, CAT(0) cube complexes, Gromov boundaries, and hierarchically hyperbolic structures needed later in the paper.
Right-angled Coxeter groups
Let Γ \Gamma Γ be a simplicial graph with vertex set V ( Γ ) V(\Gamma) V ( Γ ) and edge set E ( Γ ) E(\Gamma) E ( Γ ) . The right-angled Coxeter group defined by Γ \Gamma Γ is
W Γ = ⟨ V ( Γ ) ∣ v 2 = 1 for v ∈ V ( Γ ) , ( v w ) 2 = 1 for ( v , w ) ∈ E ( Γ ) ⟩ . W_{\Gamma}=\langle V(\Gamma)\mid v^{2}=1\text{ for }v\in V(\Gamma),(vw)^{2}=1\text{ for }(v,w)\in E(\Gamma)\rangle. W Γ = ⟨ V ( Γ ) ∣ v 2 = 1 for v ∈ V ( Γ ) , ( v w ) 2 = 1 for ( v , w ) ∈ E ( Γ )⟩ .
We call Γ \Gamma Γ the defining graph of W Γ W_{\Gamma} W Γ and write Σ Γ \Sigma_{\Gamma} Σ Γ for the Cayley graph with respect to V ( Γ ) V(\Gamma) V ( Γ ) . A subgraph Δ ⊂ Γ \Delta\subset\Gamma Δ ⊂ Γ is induced if it contains every edge of Γ \Gamma Γ whose endpoints lie in Δ \Delta Δ . Unless stated otherwise, every subgraph we consider in this article is induced.
The following lemma is usually referred to as the Deletion Condition . We refer the reader to [4 Property 1.4.7] for a proof.
Lemma 2.1 (Deletion Condition). If a word w = w 1 ⋯ w n w=w_{1}\cdots w_{n} w = w 1 ⋯ w n is non-geodesic in Σ Γ \Sigma_{\Gamma} Σ Γ , then there are 0 < i < j ≤ n 0<i<j\le n 0 < i < j ≤ n such that
w = w 1 ⋯ w i − 1 w i + 1 ⋯ w j − 1 w j + 1 ⋯ w n w=w_{1}\cdots w_{i-1}w_{i+1}\cdots w_{j-1}w_{j+1}\cdots w_{n} w = w 1 ⋯ w i − 1 w i + 1 ⋯ w j − 1 w j + 1 ⋯ w n in W Γ W_{\Gamma} W Γ .
A subgroup of W Γ W_{\Gamma} W Γ is special if it is generated by a subset of V ( Γ ) V(\Gamma) V ( Γ ) . The following standard result justifies the terminology.
Theorem 2.2 ([4 Proposition 2.4.1] ). Let W Γ W_{\Gamma} W Γ be a right-angled Coxeter group with defining graph Γ \Gamma Γ . Suppose Δ ⊂ Γ \Delta\subset\Gamma Δ ⊂ Γ is an induced subgraph. Then the special subgroup generated by V ( Δ ) V(\Delta) V ( Δ ) is isomorphic to the right-angled Coxeter group W Δ W_{\Delta} W Δ .
A subgraph Δ \Delta Δ separates Γ \Gamma Γ if Γ ∖ Δ \Gamma\setminus\Delta Γ ∖ Δ is disconnected. A graph Λ \Lambda Λ is a join , written as Λ = Λ 0 ∗ Λ 1 \Lambda= \Lambda_0 \ast\Lambda_1 Λ = Λ 0 ∗ Λ 1 , if Λ 0 \Lambda_0 Λ 0 and Λ 1 \Lambda_1 Λ 1 are nonempty disjoint subgraphs whose union is Λ \Lambda Λ , and every vertex of Λ 0 \Lambda_0 Λ 0 is adjacent to every vertex of Λ 1 \Lambda_1 Λ 1 . A clique is a complete graph.
Definition 2.3 (Nontrivial Join). A join Λ = Λ 0 ∗ Λ 1 \Lambda= \Lambda_0 \ast\Lambda_1 Λ = Λ 0 ∗ Λ 1 is nontrivial if neither Λ 0 \Lambda_0 Λ 0 nor Λ 1 \Lambda_1 Λ 1 is a clique.
Remark 2.4. The right-angled Coxeter group W Λ W_\Lambda W Λ whose defining graph is a nontrivial join decomposes as a direct product W Λ = W Λ 0 × W Λ 1 W_\Lambda= W_{\Lambda_0} \times W_{\Lambda_1} W Λ = W Λ 0 × W Λ 1 , where each factor is infinite. Moreover, the Morse boundary of W Λ W_\Lambda W Λ is empty and no geodesic in W Λ W_\Lambda W Λ is contracting.
Example 2.5. Let Γ \Gamma Γ be a 4-cycle. Its two pairs of opposite vertices give a nontrivial join decomposition. Consequently, W Γ ≅ D ∞ × D ∞ W_\Gamma\cong D_\infty\times D_\infty W Γ ≅ D ∞ × D ∞ , where D ∞ D_\infty D ∞ is the infinite dihedral group, and W Γ W_\Gamma W Γ contains a finite-index copy of Z 2 \mathbb{Z}^2 Z 2 .
We call a subgroup H H H of G G G an unbounded product subgroup if it decomposes as a direct product H = H 0 × H 1 H = H_0 \times H_1 H = H 0 × H 1 and both H 0 , H 1 H_0, H_1 H 0 , H 1 are infinite. If Λ ⊂ Γ \Lambda\subset\Gamma Λ ⊂ Γ is a nontrivial join subgraph, the special subgroup W Λ W_\Lambda W Λ is an unbounded product subgroup of W Γ W_\Gamma W Γ .
Many algebraic and geometric properties of RACGs can be read from their defining graphs. Moussong’s theorem classifies hyperbolic right-angled Coxeter groups.
Theorem 2.6 ([58 ] ). A right-angled Coxeter group W Γ W_\Gamma W Γ is hyperbolic if and only if Γ \Gamma Γ has no induced square.
The number of ends of W Γ W_\Gamma W Γ is detectable in Γ \Gamma Γ , by Stallings’ theorem [71 ] together with work of Mihalik–Tschantz [63 ] or Davis [30 ] .
Theorem 2.7 ([63 Corollary 16] ). W Γ W_\Gamma W Γ is one-ended if and only if Γ \Gamma Γ is connected, is not a clique, and has no separating clique.
We conclude this subsection with a graph-theoretic lemma that will be used repeatedly. For a subgraph Λ ⊂ Γ \Lambda\subset\Gamma Λ ⊂ Γ , its link lk ( Λ ) \operatorname{lk}(\Lambda) lk ( Λ ) is the set of vertices adjacent to every vertex of Λ \Lambda Λ , and its star is st ( Λ ) = V ( Λ ) ∪ lk ( Λ ) \operatorname{st}(\Lambda) = V(\Lambda) \cup\operatorname{lk}(\Lambda) st ( Λ ) = V ( Λ ) ∪ lk ( Λ ) .
Lemma 2.8. Let Λ ⊂ Γ \Lambda\subset\Gamma Λ ⊂ Γ be a nontrivial join and let Σ ⊂ Γ \Sigma\subset\Gamma Σ ⊂ Γ . If Δ = lk ( Σ ) ∩ Λ \Delta= \operatorname{lk}(\Sigma) \cap\Lambda Δ = lk ( Σ ) ∩ Λ is not a clique, then Σ ∪ Δ \Sigma\cup\Delta Σ ∪ Δ is contained in a nontrivial join of Γ \Gamma Γ .
Proof. Write Λ = Λ 0 ∗ Λ 1 \Lambda= \Lambda_0 \ast\Lambda_1 Λ = Λ 0 ∗ Λ 1 , where neither factor is a clique, and put Δ i = Δ ∩ Λ i \Delta_i = \Delta\cap\Lambda_i Δ i = Δ ∩ Λ i . At least one Δ i \Delta_i Δ i is not a clique; otherwise Δ = Δ 0 ∗ Δ 1 \Delta= \Delta_0 \ast\Delta_1 Δ = Δ 0 ∗ Δ 1 would be a clique. After interchanging the factors, assume that Δ 0 \Delta_0 Δ 0 is not a clique. Because Δ ⊂ lk ( Σ ) \Delta\subset\operatorname{lk}(\Sigma) Δ ⊂ lk ( Σ ) , every vertex of Δ 0 \Delta_0 Δ 0 is adjacent to every vertex of Σ ∪ Δ 1 \Sigma\cup\Delta_1 Σ ∪ Δ 1 . Hence
Δ 0 ∗ ( Σ ∪ Δ 1 ) \Delta_0 \ast(\Sigma\cup\Delta_1) Δ 0 ∗ ( Σ ∪ Δ 1 )
is a join containing Σ ∪ Δ \Sigma\cup\Delta Σ ∪ Δ . If its second factor is not a clique, this join is nontrivial. Otherwise choose a nonclique subgraph Λ 1 ′ ⊂ Λ 1 \Lambda'_1 \subset\Lambda_1 Λ 1 ′ ⊂ Λ 1 containing Δ 1 \Delta_1 Δ 1 . Then
Δ 0 ∗ ( Σ ∪ Λ 1 ′ ) \Delta_0 \ast(\Sigma\cup\Lambda'_1) Δ 0 ∗ ( Σ ∪ Λ 1 ′ )
is a nontrivial join containing Σ ∪ Δ \Sigma\cup\Delta Σ ∪ Δ . □
Equivalent graph-theoretic conditions
We compare the graph-theoretic hypotheses of our main theorem with those used by Mihalik–Ruane–Tschantz and Cordes–Levcovitz. For one-ended right-angled Coxeter groups, we show that conditions (b) and (c) are equivalent both to the absence of product separators and virtual factor separators and to wide-avoidance. These equivalences will be used in Section 4 to apply results on Morse boundaries, and in Section 8 to apply the fan-and-filter construction.
Definition 2.9 ([61 Definition 5.1] ). A product separator for Γ \Gamma Γ is an induced subgraph Λ \Lambda Λ such that Λ \Lambda Λ separates Γ \Gamma Γ and the special subgroup W Λ W_{\Lambda} W Λ decomposes as W Λ = W Λ 0 × W Λ 1 W_{\Lambda}=W_{\Lambda_0}\times W_{\Lambda_1} W Λ = W Λ 0 × W Λ 1 , where both W Λ 0 W_{\Lambda_0} W Λ 0 and W Λ 1 W_{\Lambda_1} W Λ 1 are infinite.
Definition 2.10 ([61 Definition 3.1] ). A virtual factor separator for Γ \Gamma Γ is a triple ( C , C 1 , K ) (C,C_1,K) ( C , C 1 , K ) of induced subgraphs such that
C 1 ⊂ C C_1\subset C C 1 ⊂ C and C C C separates Γ \Gamma Γ ;
W C 1 W_{C_1} W C 1 has finite index in W C W_C W C ;
K ⊂ lk ( C 1 ) K\subset\operatorname{lk}(C_1) K ⊂ lk ( C 1 ) and W K W_K W K is infinite.
We record a theorem about the visual boundary of W Γ W_{\Gamma} W Γ by Mihalik–Ruane–Tschantz.
Theorem 2.11 ([61 ] ). Let W Γ W_{\Gamma} W Γ be a one-ended right-angled Coxeter group. If the defining graph Γ \Gamma Γ has no product separator and no virtual factor separator, then the visual boundary of W Γ W_{\Gamma} W Γ is locally connected.
We note that, in the right-angled case, the notion of a wide subgraph used in [25 Definition 3.1 and Remark 3.4.1] agrees with that of an induced nontrivial join.
Definition 2.12 ([25 Definition 3.3] ). We say the graph Γ \Gamma Γ is wide-avoidant if given any nontrivial join subgraph Δ ⊂ Γ \Delta\subset\Gamma Δ ⊂ Γ and any pair of vertices s , t ∈ Γ s,t\in\Gamma s , t ∈ Γ , there exists a path α ⊂ Γ \alpha\subset\Gamma α ⊂ Γ from s s s to t t t such that α ∩ Δ ⊂ { s , t } \alpha\cap\Delta\subset\{s,t\} α ∩ Δ ⊂ { s , t } .
We record a theorem about the Morse boundary of W Γ W_{\Gamma} W Γ by Cordes–Levcovitz.
Theorem 2.13 ([25 Theorem B] ). Let W Γ W_{\Gamma} W Γ be a one-ended right-angled Coxeter group. Then the Morse boundary ∂ M W Γ \partial_M W_{\Gamma} ∂ M W Γ is connected and locally connected if and only if Γ \Gamma Γ is wide-avoidant.
The following observation is due to Michael Mihalik.
Lemma 2.14. Let Γ \Gamma Γ be the defining graph of a one-ended RACG W Γ W_{\Gamma} W Γ . Then the following are equivalent:
(i) Γ \Gamma Γ has neither product separator nor virtual factor separator;
(ii) Γ \Gamma Γ is wide-avoidant;
(iii) Γ \Gamma Γ has no separating nontrivial join and for each vertex v ∈ Γ v\in\Gamma v ∈ Γ , the link of v v v is not contained in a nontrivial join.
Proof. For (i)⇔ \Leftrightarrow ⇔ (iii), observe first that a special subgroup is infinite exactly when its defining subgraph is not a clique. Thus product separators are precisely separating nontrivial joins. When no product separator is present, [61 Lemma 5.3] identifies virtual factor separators with vertices whose links are contained in nontrivial joins. This proves the equivalence.
We prove (ii)⇔ \Leftrightarrow ⇔ (iii) directly. A separating nontrivial join Δ \Delta Δ prevents vertices in different components of Γ ∖ Δ \Gamma\setminus\Delta Γ ∖ Δ from being joined while avoiding Δ \Delta Δ . Likewise, if lk ( v ) ⊂ Δ \operatorname{lk}(v)\subset\Delta lk ( v ) ⊂ Δ for a nontrivial join Δ \Delta Δ , then every path from v v v to a vertex outside Δ ∪ { v } \Delta\cup\{v\} Δ ∪ { v } meets Δ \Delta Δ immediately. Hence failure of either condition in (iii) implies failure of wide-avoidance.
Conversely, assume (iii) and let Δ \Delta Δ be a nontrivial join. Then Γ ∖ Δ \Gamma\setminus\Delta Γ ∖ Δ is connected. Given s , t ∈ V ( Γ ) s,t\in V(\Gamma) s , t ∈ V ( Γ ) , let s ′ = s s'=s s ′ = s if s ∉ Δ s\notin\Delta s ∈ / Δ . If s ∈ Δ s\in\Delta s ∈ Δ , condition (iii) gives a vertex s ′ ∈ lk ( s ) ∖ Δ s'\in\operatorname{lk}(s)\setminus\Delta s ′ ∈ lk ( s ) ∖ Δ . Define t ′ t' t ′ similarly. Join s ′ s' s ′ to t ′ t' t ′ by a path in Γ ∖ Δ \Gamma\setminus\Delta Γ ∖ Δ and, when necessary, append the edges to s s s and t t t , respectively. The resulting path meets Δ \Delta Δ only at those endpoints that already lie in Δ \Delta Δ . Thus Γ \Gamma Γ is wide-avoidant. □ \square □
Remark 2.15. The hypothesis excluding product separators is essential in the equivalence above. Without it, the absence of virtual factor separators is not equivalent to condition (c). See examples observed by Mihalik in Figure 2 and Figure 3 .
Figure 2. No virtual factor separator but the link of v 0 v_0 v 0 is contained in a nontrivial join.
Figure 3. No link of a vertex is contained in a nontrivial join but { a 1 , a 2 , a 3 , a 4 , b 1 , b 2 } \{a_1,a_2,a_3,a_4,b_1,b_2\} { a 1 , a 2 , a 3 , a 4 , b 1 , b 2 } spans a virtual factor separator.
CAT(0) cube complexes
The geometry of a Coxeter group is encoded by its Davis complex ; see [30 ] for the construction and its properties. We write X = X Γ X=X_{\Gamma} X = X Γ for the Davis complex of W Γ W_{\Gamma} W Γ . If W Δ W_{\Delta} W Δ is a special subgroup of W Γ W_{\Gamma} W Γ , then X Δ X_{\Delta} X Δ embeds isometrically as a convex subcomplex of X Γ X_{\Gamma} X Γ . Throughout the paper, Davis complexes carry their combinatorial metrics rather than their CAT(0) metrics.
For right-angled Coxeter groups, the 1-skeleton X ( 1 ) X^{(1)} X ( 1 ) is precisely the Cayley graph Σ Γ \Sigma_{\Gamma} Σ Γ . After choosing a basepoint o ∈ X o\in X o ∈ X , we label its vertices by elements of W Γ W_{\Gamma} W Γ and its edges by generators in V ( Γ ) V(\Gamma) V ( Γ ) . If α \alpha α is an edge path in X X X , let α ‾ = s 1 ⋯ s n \overline{\alpha}=s_1\cdots s_n α = s 1 ⋯ s n be the word obtained by reading its edge labels in order, and set
Label ( α ) = Label ( α ‾ ) = { s ∈ V ( Γ ) : s occurs in α ‾ } . \operatorname{Label}(\alpha)=\operatorname{Label}(\overline{\alpha})=\{s\in V(\Gamma):s\text{ occurs in }\overline{\alpha}\}. Label ( α ) = Label ( α ) = { s ∈ V ( Γ ) : s occurs in α } .
All edges dual to a given hyperplane of X X X have the same label v ∈ V ( Γ ) v \in V(\Gamma) v ∈ V ( Γ ) ; we say that such a hyperplane is of type v v v . Let e v e_v e v be the edge labeled v v v incident to o o o , and let H v H_v H v be its dual hyperplane. Then
Stab W Γ ( H v ) = ⟨ st ( v ) ⟩ . \operatorname{Stab}_{W_\Gamma}(H_v)=\langle\operatorname{st}(v)\rangle. Stab W Γ ( H v ) = ⟨ st ( v )⟩ .
More generally, a hyperplane g H v gH_v g H v of type v v v has stabilizer g ⟨ st ( v ) ⟩ g − 1 g\langle\operatorname{st}(v)\rangle g^{-1} g ⟨ st ( v )⟩ g − 1 .
Each hyperplane H H H separates X X X into two halfspaces, denoted by H ← \overleftarrow{H} H and H → \overrightarrow{H} H . Its carrier N ( H ) \mathcal{N}(H) N ( H ) is the union of the closed cubes that meet H H H . There is a cubical isometric embedding H × [ − 1 2 , 1 2 ] ↪ X H \times[-\frac{1}{2},\frac{1}{2}] \hookrightarrow X H × [ − 2 1 , 2 1 ] ↪ X ; the images of H × { ± 1 2 } H \times\{\pm\frac{1}{2}\} H × { ± 2 1 } are the two combinatorial hyperplanes , denoted by H ± H^{\pm} H ± .
For 0-cubes x , y ∈ X x,y \in X x , y ∈ X , the distance d ( x , y ) d(x,y) d ( x , y ) equals the number of hyperplanes separating them. The convex hull of a subcomplex Y ⊂ X Y \subset X Y ⊂ X is the intersection of all convex subcomplexes containing Y Y Y . By Sageev’s construction [69 ] , the convex hull of two 0-cubes is cubically isometric to the cube complex dual to the wallspace of hyperplanes separating them.
A graph is median if every triple of vertices a , b , c a,b,c a , b , c has a unique vertex m ( a , b , c ) m(a,b,c) m ( a , b , c ) lying on a geodesic between each pair. The 1-skeleton of a CAT(0) cube complex is a median graph [23 ] . For 0-cubes x , y ∈ X x,y \in X x , y ∈ X , the median interval between them is the 1-skeleton of their convex hull; we denote it by [ x , y ] [x,y] [ x , y ] .
If Y ⊂ X Y \subset X Y ⊂ X is convex, the closest-point projection g Y : X → Y \mathfrak{g}_Y:X \to Y g Y : X → Y is well defined and we call it the gate map . Thus g Y ( x ) \mathfrak{g}_Y(x) g Y ( x ) is the unique 0-cube y ∈ Y y \in Y y ∈ Y minimizing the distance d ( x , z ) d(x,z) d ( x , z ) over all 0-cubes z ∈ Y z \in Y z ∈ Y . Let W ( Y ) \mathcal{W}(Y) W ( Y ) denote the set of hyperplanes crossing Y Y Y .
The following is a standard application of the gate map, which is usually referred to as the bridge lemma introduced in [5 ] . See [FFH24, Lemma 3.23] for a complete proof.
Lemma 2.16 (Bridge Lemma). Let P , Q P,Q P , Q be convex subcomplexes of a CAT(0) cube complex X X X . Let g P : X → P \mathfrak{g}_P:X \to P g P : X → P and g Q : X → Q \mathfrak{g}_Q:X \to Q g Q : X → Q be the gate maps. Then the following hold:
g P ( Q ) \mathfrak{g}_P(Q) g P ( Q ) and g Q ( P ) \mathfrak{g}_Q(P) g Q ( P ) are convex subcomplexes.
W ( g P ( Q ) ) = W ( g Q ( P ) ) = W ( P ) ∩ W ( Q ) \mathcal{W}(\mathfrak{g}_P(Q))=\mathcal{W}(\mathfrak{g}_Q(P))=\mathcal{W}(P)\cap\mathcal{W}(Q) W ( g P ( Q )) = W ( g Q ( P )) = W ( P ) ∩ W ( Q ) .
The map g P : g Q ( P ) → g P ( Q ) \mathfrak{g}_P:\mathfrak{g}_Q(P)\to\mathfrak{g}_P(Q) g P : g Q ( P ) → g P ( Q ) is an isomorphism of CAT(0) cube complexes.
The cubical convex hull [ g P ( Q ) , g Q ( P ) ] [\mathfrak{g}_P(Q),\mathfrak{g}_Q(P)] [ g P ( Q ) , g Q ( P )] is a CAT(0) cube complex isomorphic to g P ( Q ) × [ x , g P ( Q ) ] \mathfrak{g}_P(Q)\times[x,\mathfrak{g}_P(Q)] g P ( Q ) × [ x , g P ( Q )] for a vertex x ∈ g Q ( P ) x\in\mathfrak{g}_Q(P) x ∈ g Q ( P ) .
If P ∩ Q ≠ ∅ P\cap Q\ne\varnothing P ∩ Q = ∅ , then g P ( Q ) = g Q ( P ) = P ∩ Q \mathfrak{g}_P(Q)=\mathfrak{g}_Q(P)=P\cap Q g P ( Q ) = g Q ( P ) = P ∩ Q and W ( P ∩ Q ) = W ( P ) ∩ W ( Q ) \mathcal{W}(P\cap Q)=\mathcal{W}(P)\cap\mathcal{W}(Q) W ( P ∩ Q ) = W ( P ) ∩ W ( Q ) .
The next lemma is standard in Coxeter group theory; see [45 ] . We include a cubical proof. It says that the intersection of the median interval [ o , x ] [o,x] [ o , x ] with the 1-neighborhood of x x x spans a cube.
Lemma 2.17. Let α \alpha α be a geodesic from o o o to x x x in X X X . Suppose e e e and f f f are 1-cubes containing x x x such that α e \alpha e α e and α f \alpha f α f are not geodesic. Then e e e and f f f span a 2-cube.
Proof. Let H e , H f H_e,H_f H e , H f be the hyperplanes dual to e , f e,f e , f , respectively. Suppose H e H_e H e and H f H_f H f do not cross. H e H_e H e separates X X X into two halfspaces H e ← \overleftarrow{H_e} H e and H e → \overrightarrow{H_e} H e . Suppose x ∈ H e → x\in\overrightarrow{H_e} x ∈ H e . Since H f H_f H f does not cross H e H_e H e , H f ⊂ H e → H_f\subset\overrightarrow{H_e} H f ⊂ H e . Since α e \alpha e α e is not geodesic, H e H_e H e separates o o o and x x x . In particular, o ∈ H e ← o\in\overleftarrow{H_e} o ∈ H e . Therefore, H f H_f H f does not separate o o o and x x x , which implies α f \alpha f α f is geodesic. Contradiction. Hence H e H_e H e must cross H f H_f H f and the 1-cubes e , f e,f e , f span a 2-cube. □ \square □
Remark 2.18. Let α \alpha α be a geodesic from x x x to y y y in X X X and α ‾ \overline{\alpha} α be the word in W Γ W_\Gamma W Γ corresponding to α \alpha α . Denote by B ( α ) = B ( α ‾ ) B(\alpha)=B(\overline{\alpha}) B ( α ) = B ( α ) the set of vertices v ∈ Γ v\in\Gamma v ∈ Γ such that α ‾ v \overline{\alpha}v α v is non-geodesic. By Lemma 2.17, B ( α ) B(\alpha) B ( α ) spans a clique in Γ \Gamma Γ .
Let J ⊂ Γ J \subset\Gamma J ⊂ Γ be a nontrivial join. It is maximal if it is not properly contained in another nontrivial join. For a maximal nontrivial join J J J and g ∈ W Γ g \in W_\Gamma g ∈ W Γ , we call g X J ⊂ X Γ gX_J \subset X_\Gamma g X J ⊂ X Γ a standard product region .
Gromov-hyperbolic spaces and their boundaries
Let ( X , d ) (X,d) ( X , d ) be a metric space and fix o ∈ X o \in X o ∈ X . The Gromov product based at o o o is
⟨ x , y ⟩ o = 1 2 ( d ( o , x ) + d ( o , y ) − d ( x , y ) ) \langle x,y\rangle_o = \frac{1}{2}\bigl(d(o,x)+d(o,y)-d(x,y)\bigr) ⟨ x , y ⟩ o = 2 1 ( d ( o , x ) + d ( o , y ) − d ( x , y ) )
for any x , y ∈ X x,y \in X x , y ∈ X .
A metric space is Gromov hyperbolic , or δ \delta δ -hyperbolic, if there is δ ≥ 0 \delta\ge0 δ ≥ 0 such that
⟨ x , y ⟩ o ≥ min { ⟨ x , z ⟩ o , ⟨ y , z ⟩ o } − δ \langle x,y\rangle_o \ge\min\{\langle x,z\rangle_o,\langle y,z\rangle_o\}-\delta ⟨ x , y ⟩ o ≥ min {⟨ x , z ⟩ o , ⟨ y , z ⟩ o } − δ
for all x , y , z , o ∈ X x,y,z,o \in X x , y , z , o ∈ X .
When X X X is geodesic, this is equivalent, up to changing the hyperbolicity constant, to requiring every geodesic triangle to be thin. Moreover, ⟨ x , y ⟩ o \langle x,y\rangle_o ⟨ x , y ⟩ o coarsely measures the distance from o o o to a geodesic from x x x to y y y , with additive error depending only on δ \delta δ .
A sequence ( x n ) n ≥ 0 (x_n)_{n\ge0} ( x n ) n ≥ 0 in X X X converges at infinity if
lim inf n , m → ∞ ⟨ x n , x m ⟩ o = ∞ . \liminf_{n,m\to\infty}\langle x_n,x_m\rangle_o=\infty. n , m → ∞ lim inf ⟨ x n , x m ⟩ o = ∞.
This condition is independent of the basepoint. Two sequences ( x n ) (x_n) ( x n ) and ( y n ) (y_n) ( y n ) converging at infinity are equivalent if
lim inf n , m → ∞ ⟨ x n , y m ⟩ o = ∞ . \liminf_{n,m\to\infty}\langle x_n,y_m\rangle_o=\infty. n , m → ∞ lim inf ⟨ x n , y m ⟩ o = ∞.
This is an equivalence relation; write [ ( x n ) ] [(x_n)] [( x n )] for the class of ( x n ) (x_n) ( x n ) .
Definition 2.19 (Gromov boundary). Let ( X , d ) (X,d) ( X , d ) be a Gromov-hyperbolic space. Its Gromov boundary is
∂ X : = { [ ( x n ) ] ∣ ( x n ) is a sequence converging at infinity in X } . \partial X:=\{[(x_n)]\mid(x_n)\text{ is a sequence converging at infinity in }X\}. ∂ X := {[( x n )] ∣ ( x n ) is a sequence converging at infinity in X } .
If X X X is proper, meaning that every closed ball is compact, then every boundary point is represented by a geodesic ray [9 III.H, Lemma 3.13] . For a nonproper hyperbolic space, every boundary point is still represented by a quasi-geodesic ray; see, for example, [53 Proposition 2.3] .
The Gromov product can be extended to the boundary points. For x ∈ X x \in X x ∈ X and ζ ∈ ∂ X \zeta\in\partial X ζ ∈ ∂ X ,
⟨ x , ζ ⟩ o = inf { lim inf n → ∞ ⟨ x , x n ⟩ o : ( x n ) ∈ ζ } . \langle x,\zeta\rangle_o=\inf\{\liminf_{n\to\infty}\langle x,x_n\rangle_o:(x_n)\in\zeta\}. ⟨ x , ζ ⟩ o = inf { n → ∞ lim inf ⟨ x , x n ⟩ o : ( x n ) ∈ ζ } .
Also for ζ , η ∈ ∂ X \zeta,\eta\in\partial X ζ , η ∈ ∂ X , we have
⟨ ζ , η ⟩ o = inf { lim inf n , m → ∞ ⟨ x n , y m ⟩ o : ( x n ) ∈ ζ , ( y m ) ∈ η } . \langle\zeta,\eta\rangle_o=\inf\{\liminf_{n,m\to\infty}\langle x_n,y_m\rangle_o:(x_n)\in\zeta,(y_m)\in\eta\}. ⟨ ζ , η ⟩ o = inf { n , m → ∞ lim inf ⟨ x n , y m ⟩ o : ( x n ) ∈ ζ , ( y m ) ∈ η } .
The sets
U ( ζ , r ) = { η ∈ ∂ X ∣ ⟨ ζ , η ⟩ o > r } U(\zeta,r)=\{\eta\in\partial X\mid\langle\zeta,\eta\rangle_o>r\} U ( ζ , r ) = { η ∈ ∂ X ∣ ⟨ ζ , η ⟩ o > r }
for r > 0 r>0 r > 0 form a neighborhood basis at ζ \zeta ζ and define the topology on ∂ X \partial X ∂ X .
The boundary admits a class of mutually quasisymmetric visual metrics; see [Vä05, DSU17] for definitions in this generality. These metrics relate boundary distance to the Gromov product as follows.
Proposition 2.20 ([Vä05, Proposition 5.16], [33 Proposition 3.6.8] ). If b > 1 b>1 b > 1 is sufficiently close to 1 1 1 , then for each o ∈ X o\in X o ∈ X , a visual metric d ∞ d_{\infty} d ∞ on ∂ X \partial X ∂ X satisfies the following inequality:
b − ⟨ ζ , η ⟩ o / 4 ≤ d ∞ ( ζ , η ) ≤ b − ⟨ ζ , η ⟩ o b^{-\langle\zeta,\eta\rangle_o/4}\le d_{\infty}(\zeta,\eta)\le b^{-\langle\zeta,\eta\rangle_o} b − ⟨ ζ , η ⟩ o /4 ≤ d ∞ ( ζ , η ) ≤ b − ⟨ ζ , η ⟩ o for any ζ , η ∈ ∂ X \zeta,\eta\in\partial X ζ , η ∈ ∂ X . Furthermore, this metric is compatible with the topology on ∂ X \partial X ∂ X .
A ( K , C ) (K,C) ( K , C ) -quasi-isometry of a δ \delta δ -hyperbolic space extends to a quasisymmetry of its boundary, with distortion controlled by K K K , C C C , and δ \delta δ [Vä05, Theorem 5.35].
Linear connectedness
We recall the definition of linear connectedness.
Definition 2.21 (Linear connectedness). A metric space ( X , d ) (X,d) ( X , d ) is L L L -linearly connected, for L ≥ 1 L\ge1 L ≥ 1 , if every x , y ∈ X x,y\in X x , y ∈ X lie in a compact connected set Y ⊂ X Y\subset X Y ⊂ X with diam ( Y ) ≤ L d ( x , y ) \operatorname{diam}(Y)\le Ld(x,y) diam ( Y ) ≤ L d ( x , y ) .
A metric space is linearly connected if it is L L L -linearly connected for some L L L .
After increasing L L L by an arbitrarily small amount, one may require Y Y Y to be an arc; see the introduction of [54 ] . A linearly connected metric space is path-connected and locally path-connected, and linear connectedness is preserved by quasisymmetries.
A hierarchically hyperbolic structure on RACGs
Behrstock–Hagen–Sisto introduced hierarchically hyperbolic spaces and groups (HHSs and HHGs) in [10 ] ; we refer the reader there for the full definition. Special groups in the sense of Haglund–Wise, including right-angled Artin and Coxeter groups, are hierarchically hyperbolic. An HHG may admit several HHS structures. Here we use the maximized structure on RACGs introduced in [1 ] .
Let X = X Γ X=X_{\Gamma} X = X Γ be the Davis complex. Recall that an induced nontrivial join J ⊂ Γ J\subset\Gamma J ⊂ Γ determines an unbounded product subgroup W J W_J W J and an isometrically embedded subcomplex X J ⊂ X Γ X_J\subset X_{\Gamma} X J ⊂ X Γ . When J J J is maximal, each translate g X J gX_J g X J is a standard product region.
We now define the maximized hyperbolic space X ^ \widehat{X} X , the nested-maximal domain in this HHS structure. We give only the RACG-specific definition; the general construction appears in [1 Section 3.3] .
Definition 2.22 (Maximized hyperbolic space). Let X = X Γ X=X_{\Gamma} X = X Γ be the Davis complex of W Γ W_{\Gamma} W Γ . The maximized hyperbolic space X ^ = X ^ Γ \widehat{X}=\widehat{X}_{\Gamma} X = X Γ is the graph with vertex set X ( 0 ) X^{(0)} X ( 0 ) in which distinct vertices x , y x,y x , y are adjacent if they are joined by an edge of X X X or lie in a common standard product region.
Give every edge of X ^ \widehat{X} X length one and denote the resulting path metric by d X ^ d_{\widehat{X}} d X . We identify the vertex sets of X X X and X ^ \widehat{X} X throughout.
If W Γ W_{\Gamma} W Γ is hyperbolic, equivalently if Γ \Gamma Γ contains no induced 4 4 4 -cycle, then X ^ = X ( 1 ) \widehat{X}=X^{(1)} X = X ( 1 ) . In general, X ^ \widehat{X} X has infinite diameter if and only if Γ \Gamma Γ is neither a clique nor a nontrivial join; see Lemma 3.4 .
By [1 Theorem 3.7] , X ^ \widehat{X} X is δ \delta δ -hyperbolic. Applying the HHS quasi-flats theorem iteratively shows that every quasi-isometry of X X X induces a quasi-isometry of X ^ \widehat{X} X ; see [37 Theorem 6.10] .
The rest of this subsection describes the HHS structure used only in § 7 ; it may be skipped on a first reading.
In a CAT(0) cube complex, two hyperplanes contact if no hyperplane separates them, equivalently if their carriers intersect. The contact graph C X \mathcal{C}X C X has one vertex for each hyperplane of X X X and an edge for each contacting pair. Hagen proved that C X \mathcal{C}X C X is a quasi-tree [39 ] , hence hyperbolic. There is a coarse Lipschitz map
c : X ⟶ 2 C X , x ⟼ { H : x ∈ N ( H ) } . c:X\longrightarrow2^{\mathcal{C}X},\qquad x\longmapsto\{H:x\in\mathcal{N}(H)\}. c : X ⟶ 2 C X , x ⟼ { H : x ∈ N ( H )} .
Suppose Y Y Y is a convex subcomplex of X X X . Its contact graph C Y \mathcal{C}Y C Y is an induced subgraph of C X \mathcal{C}X C X : every hyperplane of Y Y Y is the intersection of Y Y Y with a hyperplane of X X X , and two such hyperplanes contact in Y Y Y if and only if they contact at a cube of Y Y Y , by the Helly property.
Let Y \mathfrak{Y} Y be a collection of convex subcomplexes of X X X . The coned-off contact graph C ^ X \widehat{\mathcal{C}}X C X of X X X with respect to Y \mathfrak{Y} Y is a graph obtained from the contact graph C X \mathcal{C}X C X by adding an edge between a pair of vertices in an induced subgraph C Y \mathcal{C}Y C Y for every Y ∈ Y Y\in\mathfrak{Y} Y ∈ Y .
Two convex subcomplexes Y , Y ′ ⊂ X Y,Y'\subset X Y , Y ′ ⊂ X are parallel if a hyperplane crosses Y Y Y if and only if it crosses Y ′ Y' Y ′ .
Now we describe the maximized hierarchically hyperbolic structure on W Γ W_{\Gamma} W Γ . Let X = X Γ X=X_{\Gamma} X = X Γ be the Davis complex. Let L \mathcal{L} L be the collection of induced subgraphs Λ \Lambda Λ of Γ \Gamma Γ such that neither Λ \Lambda Λ nor lk ( Λ ) \operatorname{lk}(\Lambda) lk ( Λ ) is a clique. Note that if Λ ∈ L \Lambda\in\mathcal{L} Λ ∈ L , then lk ( Λ ) ∈ L \operatorname{lk}(\Lambda)\in\mathcal{L} lk ( Λ ) ∈ L . Then we collect the parallelism classes of all convex subcomplexes of X X X of the form g X Λ gX_{\Lambda} g X Λ where g ∈ W Γ g\in W_{\Gamma} g ∈ W Γ and Λ ∈ L \Lambda\in\mathcal{L} Λ ∈ L . Note that two subcomplexes g X Λ gX_{\Lambda} g X Λ and h X Δ hX_{\Delta} h X Δ are parallel only if Λ = Δ \Lambda=\Delta Λ = Δ and g − 1 h ∈ W Λ ∗ lk ( Λ ) g^{-1}h\in W_{\Lambda*\operatorname{lk}(\Lambda)} g − 1 h ∈ W Λ ∗ lk ( Λ ) .
There is a set of domains, S \mathcal{S} S , which index a collection of δ \delta δ -hyperbolic spaces to which X X X projects. Each domain U ∈ S U\in\mathcal{S} U ∈ S indexes either X ^ \widehat{X} X or the parallelism class [ g X Λ ] [gX_{\Lambda}] [ g X Λ ] or a dummy domain (see [1 Appendix A] ). Every domain is associated with a δ \delta δ -hyperbolic space. The hyperbolic space associated with X ^ \widehat{X} X is itself. The hyperbolic space associated with [ g X Λ ] [gX_{\Lambda}] [ g X Λ ] is the coned-off contact graph C ^ g X Λ \widehat{\mathcal{C}}gX_{\Lambda} C g X Λ with respect to contact graphs of h X Δ hX_{\Delta} h X Δ where Δ ∈ L \Delta\in\mathcal{L} Δ ∈ L , Δ ⊊ Λ \Delta\subsetneq\Lambda Δ ⊊ Λ and h X Δ hX_{\Delta} h X Δ is parallel to a subcomplex of g X Λ gX_{\Lambda} g X Λ . The hyperbolic space associated with a dummy domain is a singleton. We denote by d U d_U d U the metric in the hyperbolic space associated with the domain U ∈ S U\in\mathcal{S} U ∈ S . Moreover, for each domain U ∈ S U\in\mathcal{S} U ∈ S there is a natural coarsely well-defined Lipschitz projection, denoted by π U \pi_U π U , from X X X to the hyperbolic space associated with U U U . When U = X ^ U=\widehat{X} U = X , π U \pi_U π U is the inclusion map from X X X to X ^ \widehat{X} X . When U U U is a dummy-domain, π U \pi_U π U is the trivial surjection. When U = [ g X Λ ] U=[gX_{\Lambda}] U = [ g X Λ ] , π U : X → C ^ g X Λ \pi_U:X\to\widehat{\mathcal{C}}gX_{\Lambda} π U : X → C g X Λ is defined as the composition
X ⟶ g X Λ ⟶ C g X Λ ⟶ C ^ g X Λ X\longrightarrow gX_{\Lambda}\longrightarrow\mathcal{C}gX_{\Lambda}\longrightarrow\widehat{\mathcal{C}}gX_{\Lambda} X ⟶ g X Λ ⟶ C g X Λ ⟶ C g X Λ
of the gate map g g X Λ : X → g X Λ g_{gX_{\Lambda}}:X\to gX_{\Lambda} g g X Λ : X → g X Λ , the coarse Lipschitz map c : g X Λ → C g X Λ c:gX_{\Lambda}\to\mathcal{C}gX_{\Lambda} c : g X Λ → C g X Λ and the inclusion of C g X Λ \mathcal{C}gX_{\Lambda} C g X Λ into the coned-off space C ^ g X Λ \widehat{\mathcal{C}}gX_{\Lambda} C g X Λ . We say the projection π U \pi_U π U is coarsely well-defined because the image π U ( x ) \pi_U(x) π U ( x ) is actually a subset of C ^ g X Λ \widehat{\mathcal{C}}gX_{\Lambda} C g X Λ (coarsely) with uniformly bounded diameter (well-defined).
For two non-dummy domains U , V ∈ S U,V\in\mathcal{S} U , V ∈ S , we say U U U is nested in V V V , written U ⊑ V U\sqsubseteq V U ⊑ V , if either V = X ^ V=\widehat{X} V = X , or U = [ g X Λ 0 ] U=[gX_{\Lambda_0}] U = [ g X Λ 0 ] and V = [ h X Λ 1 ] V=[hX_{\Lambda_1}] V = [ h X Λ 1 ] for some g , h ∈ W Γ g,h\in W_{\Gamma} g , h ∈ W Γ with Λ 0 ⊂ Λ 1 \Lambda_0\subset\Lambda_1 Λ 0 ⊂ Λ 1 and g − 1 h ∈ W Λ 1 g^{-1}h\in W_{\Lambda_1} g − 1 h ∈ W Λ 1 . The relation ⊑ \sqsubseteq ⊑ is a partial order on S \mathcal{S} S , with maximal element X ^ \widehat{X} X . For each non-dummy domain U = [ g X Λ ] ≠ X ^ U=[gX_{\Lambda}]\ne\widehat{X} U = [ g X Λ ] = X , fix a vertex ρ X ^ U ∈ g X Λ ⊂ X ^ \rho_{\widehat{X}}^U\in gX_{\Lambda}\subset\widehat{X} ρ X U ∈ g X Λ ⊂ X . Two non-dummy domains U = [ g X Λ 0 ] U=[gX_{\Lambda_0}] U = [ g X Λ 0 ] and V = [ h X Λ 1 ] V=[hX_{\Lambda_1}] V = [ h X Λ 1 ] are orthogonal , written U ⊥ V U\perp V U ⊥ V , if Λ 0 ⊂ lk ( Λ 1 ) \Lambda_0\subset\operatorname{lk}(\Lambda_1) Λ 0 ⊂ lk ( Λ 1 ) and g − 1 h ∈ W Λ 1 g^{-1}h\in W_{\Lambda_1} g − 1 h ∈ W Λ 1 . Their union then lies in an unbounded product region. Two domains are transverse if they are neither nested nor orthogonal.
Now we recall the distance formula for X X X from [10 Theorem 9.1] . It will be used in § 7 .
We recall some necessary notation. For any a , b ∈ R a,b\in\mathbb{R} a , b ∈ R and K ≥ 1 , C ≥ 0 K\ge1,C\ge0 K ≥ 1 , C ≥ 0 , we say a ≍ K , C b a\asymp_{K,C}b a ≍ K , C b if a K − C ≤ b ≤ a ⋅ K + C \frac{a}{K}-C\le b\le a\cdot K+C K a − C ≤ b ≤ a ⋅ K + C . For any a , s ∈ R a,s\in\mathbb{R} a , s ∈ R ,
{ { a } } s = { a if a ≥ s , 0 otherwise . \{\!\{a\}\!\}_s=
\begin{cases}
a & \text{if }a\ge s,\\
0 & \text{otherwise}.
\end{cases} { { a } } s = { a 0 if a ≥ s , otherwise .
Theorem 2.24 ([10 Theorem 9.1] ). (Distance Formula ) There is a constant s 0 ≥ 0 s_0\ge0 s 0 ≥ 0 such that for every s ≥ s 0 s\ge s_0 s ≥ s 0 , there are K ≥ 1 , C ≥ 0 K\ge1,C\ge0 K ≥ 1 , C ≥ 0 such that for all 0 0 0 -cubes x , y ∈ X ( 0 ) x,y\in X^{(0)} x , y ∈ X ( 0 )
d ( x , y ) ≍ K , C ∑ U ∈ S { { d U ( π U ( x ) , π U ( y ) ) } } s . d(x,y)\asymp_{K,C}\sum_{U\in\mathcal{S}}\{\!\{d_U(\pi_U(x),\pi_U(y))\}\!\}_s. d ( x , y ) ≍ K , C U ∈ S ∑ { { d U ( π U ( x ) , π U ( y ))} } s .
Remark 2.25. As a consequence of the distance formula, for each s ≥ s 0 s\ge s_0 s ≥ s 0 , only finitely many domains U ∈ S U\in\mathcal{S} U ∈ S satisfy d U ( π U ( x ) , π U ( y ) ) ≥ s d_U(\pi_U(x),\pi_U(y))\ge s d U ( π U ( x ) , π U ( y )) ≥ s .
Disconnectedness of the Gromov boundary
In this section, we identify exactly when X ^ Γ \widehat{X}_{\Gamma} X Γ has an empty or disconnected Gromov boundary. Lemma 3.4 shows that the boundary is empty precisely when Γ \Gamma Γ is a clique or a nontrivial join. Proposition 3.6 shows that if the boundary is nonempty but one of conditions (a)–(c) of Theorem 1.1 fails, then X ^ Γ \widehat{X}_{\Gamma} X Γ is multi-ended : the complement of some closed ball has at least two unbounded components. Finally, in Proposition 3.8 we show that when X ^ Γ \widehat{X}_{\Gamma} X Γ is multi-ended, it has a disconnected Gromov boundary.
We first use contracting geodesics in X Γ X_{\Gamma} X Γ to characterize when X ^ Γ \widehat{X}_{\Gamma} X Γ has infinite diameter and a nonempty Gromov boundary. We recall the definition of a contracting subset from [1 Definition 4.2] .
Definition 3.1 (Contracting ). A subset γ \gamma γ in a metric space X X X is said to be D D D -contracting if there exists a map π γ : X → γ ⊂ X \pi_\gamma:X\to\gamma\subset X π γ : X → γ ⊂ X and constants A , D > 0 A,D>0 A , D > 0 satisfying the following conditions:
for any x ∈ γ x\in\gamma x ∈ γ , we have diam X ( x , π γ ( x ) ) < D \operatorname{diam}_X(x,\pi_\gamma(x))<D diam X ( x , π γ ( x )) < D ;
if x , y ∈ X x,y\in X x , y ∈ X with d X ( x , y ) < 1 d_X(x,y)<1 d X ( x , y ) < 1 , then diam X ( π γ ( x ) , π γ ( y ) ) < D \operatorname{diam}_X(\pi_\gamma(x),\pi_\gamma(y))<D diam X ( π γ ( x ) , π γ ( y )) < D ;
for all x ∈ X x\in X x ∈ X , if we set R = A ⋅ d ( x , γ ) R=A\cdot d(x,\gamma) R = A ⋅ d ( x , γ ) , then diam X ( π γ ( B R ( x ) ) ) ≤ D \operatorname{diam}_X(\pi_\gamma(B_R(x)))\le D diam X ( π γ ( B R ( x ))) ≤ D .
In this article, we only consider the contracting property of geodesics in a CAT(0) cube complex and the map π γ \pi_\gamma π γ can be considered as the gate map.
In the following lemma, we describe when W Γ W_{\Gamma} W Γ has contracting elements.
Lemma 3.2 (Existence of contracting elements ). Let Λ ⊂ Γ \Lambda\subset\Gamma Λ ⊂ Γ be an induced subgraph of Γ \Gamma Γ . Suppose Λ \Lambda Λ is not a clique and is not contained in a nontrivial join . Then the special subgroup W Λ W_{\Lambda} W Λ contains a contracting element.
Proof. Let C C C be the set of cone vertices of Λ \Lambda Λ ; that is, a vertex v ∈ V ( Γ ) v\in V(\Gamma) v ∈ V ( Γ ) is contained in C C C if and only if v ∈ V ( Λ ) v\in V(\Lambda) v ∈ V ( Λ ) and for every w ∈ V ( Λ ) ∖ { v } w\in V(\Lambda)\setminus\{v\} w ∈ V ( Λ ) ∖ { v } , w w w is adjacent to v v v . Let Λ ′ = Λ ∖ C \Lambda'=\Lambda\setminus C Λ ′ = Λ ∖ C .
It is clear that C C C spans a clique. Since Λ \Lambda Λ is not a clique, Λ ′ \Lambda' Λ ′ is nonempty. Moreover, Λ ′ \Lambda' Λ ′ has no cone vertex since a vertex adjacent to all vertices of Λ ′ \Lambda' Λ ′ is also adjacent to all vertices of C C C , and would be a cone vertex of Λ \Lambda Λ . Now we claim that Λ ′ \Lambda' Λ ′ is not contained in a nontrivial join of Γ \Gamma Γ . Suppose not. That is, there is a nontrivial join subgraph Δ = Δ 0 ∗ Δ 1 \Delta= \Delta_0 * \Delta_1 Δ = Δ 0 ∗ Δ 1 of Γ \Gamma Γ containing Λ ′ \Lambda' Λ ′ , where neither Δ 0 \Delta_0 Δ 0 nor Δ 1 \Delta_1 Δ 1 is a clique. Since Λ ′ \Lambda' Λ ′ is not a clique, at least one of Λ ′ ∩ Δ 0 \Lambda' \cap\Delta_0 Λ ′ ∩ Δ 0 or Λ ′ ∩ Δ 1 \Lambda' \cap\Delta_1 Λ ′ ∩ Δ 1 is not a clique. Without loss of generality, we assume that Λ ′ ∩ Δ 0 \Lambda' \cap\Delta_0 Λ ′ ∩ Δ 0 is not a clique. Every vertex of Λ ′ ∩ Δ 0 \Lambda' \cap\Delta_0 Λ ′ ∩ Δ 0 is adjacent to every vertex of Δ 1 ∪ C \Delta_1 \cup C Δ 1 ∪ C . Hence, Λ \Lambda Λ is contained in a nontrivial join ( Λ ′ ∩ Δ 0 ) ∗ ( Δ 1 ∪ C ) (\Lambda' \cap\Delta_0) * (\Delta_1 \cup C) ( Λ ′ ∩ Δ 0 ) ∗ ( Δ 1 ∪ C ) . Contradiction.
Assume the vertex set V ( Λ ′ ) = { v 1 , … , v n } V(\Lambda') = \{v_1, \ldots, v_n\} V ( Λ ′ ) = { v 1 , … , v n } . Let
g = v 1 ⋯ v n ∈ W Λ ′ g = v_1 \cdots v_n \in W_{\Lambda'} g = v 1 ⋯ v n ∈ W Λ ′
be the Coxeter element of W Λ ′ W_{\Lambda'} W Λ ′ . For every m ∈ N m \in\mathbb{N} m ∈ N , g m g^m g m is geodesic. Indeed, if g m g^m g m is not geodesic for some m m m , then the Deletion Condition (see Lemma 2.1) would provide two occurrences of some letter v i v_i v i such that v i v_i v i commutes with every letter in between. For an innermost such pair, this would imply that v i v_i v i commutes with every vertex in V ( Λ ′ ) ∖ { v i } V(\Lambda') \setminus\{v_i\} V ( Λ ′ ) ∖ { v i } , and hence v i v_i v i would be a cone vertex of Λ \Lambda Λ . Contradiction.
The axis α ⊂ X Γ \alpha\subset X_\Gamma α ⊂ X Γ labeled by
⋯ ( v 1 ⋯ v n ) ( v 1 ⋯ v n ) ( v 1 ⋯ v n ) ⋯ \cdots(v_1 \cdots v_n)(v_1 \cdots v_n)(v_1 \cdots v_n)\cdots ⋯ ( v 1 ⋯ v n ) ( v 1 ⋯ v n ) ( v 1 ⋯ v n ) ⋯
is an axis for g g g . It is clear that every subpath of α \alpha α of length at least n n n contains all labels in V ( Λ ′ ) V(\Lambda') V ( Λ ′ ) , which is not contained in a nontrivial join. Hence, every subpath of α \alpha α with length at least n n n is not contained in a standard product region. By [27 Theorem 4.2] , α ⊂ X Γ \alpha\subset X_\Gamma α ⊂ X Γ is contracting, and hence g ∈ W Γ g \in W_\Gamma g ∈ W Γ is a contracting element. □ \square □
Using the preceding remark, we characterize when the Gromov boundary ∂ X ^ Γ \partial\widehat X_\Gamma ∂ X Γ of the maximized hyperbolic space is nonempty.
Lemma 3.4 (Nonempty boundary ∂ X ^ Γ \partial\widehat X_\Gamma ∂ X Γ ). The maximized hyperbolic space X ^ Γ \widehat X_\Gamma X Γ has infinite diameter and a nonempty Gromov boundary if and only if Γ \Gamma Γ is neither a clique nor a nontrivial join. Whenever nonempty, ∂ X ^ Γ \partial\widehat X_\Gamma ∂ X Γ contains at least two points.
Proof. If Γ \Gamma Γ is a clique, then both X Γ X_\Gamma X Γ and X ^ Γ \widehat X_\Gamma X Γ are bounded. If Γ \Gamma Γ is a nontrivial join, then X Γ X_\Gamma X Γ is a product of two unbounded subspaces and is itself a standard product region, so diam ( X ^ Γ ) = 1 \operatorname{diam}(\widehat X_\Gamma) = 1 diam ( X Γ ) = 1 . Thus the boundary is empty in either case.
Conversely, suppose Γ \Gamma Γ is neither a clique nor a nontrivial join. By Lemma 3.2 , there is a contracting geodesic α \alpha α in X Γ X_\Gamma X Γ . By [1 Theorem D] , the image of α \alpha α in X ^ Γ \widehat X_\Gamma X Γ under the natural identity map is a quasi-geodesic. Thus, X ^ Γ \widehat X_\Gamma X Γ has infinite diameter. Moreover, ∂ X ^ Γ \partial\widehat X_\Gamma ∂ X Γ is nonempty and contains at least two points. □ \square □
For the rest of the section, assume that Γ \Gamma Γ is neither a clique nor a nontrivial join. A subspace A A A of a metric space X X X coarsely separates X X X if, for some K ∈ N K \in\mathbb{N} K ∈ N , the complement X ∖ N K ( A ) X \setminus N_K(A) X ∖ N K ( A ) has at least two components, which are not contained in a finite neighborhood of A A A . The following is a classical result from Bass-Serre theory. We refer the reader to [70 Lemma 1.8] and [65 Lemma 2.2] for a proof.
Lemma 3.5. If a finitely generated group G G G splits over a finitely generated subgroup H H H , then H H H coarsely separates the Cayley graph of $G.
If Δ \Delta Δ separates Γ \Gamma Γ , then W Γ W_{\Gamma} W Γ splits over W Δ W_{\Delta} W Δ and X Δ X_{\Delta} X Δ coarsely separates X Γ X_{\Gamma} X Γ . When Δ \Delta Δ is a clique or lies in a nontrivial join, its image in X ^ Γ \widehat{X}_{\Gamma} X Γ has bounded diameter and also coarsely separates X ^ Γ \widehat{X}_{\Gamma} X Γ . Adapting the argument of [65 Lemma 2.2] , we moreover find quasi-geodesic rays in two distinct unbounded complementary components.
Proposition 3.6. Let W Γ W_{\Gamma} W Γ be the right-angled Coxeter group with defining graph Γ \Gamma Γ . Suppose Γ \Gamma Γ is not a clique, not a nontrivial join, and one of the conditions (a), (b) and (c) from Theorem 1.1 fails. Then there is a constant K ′ ≥ 0 K' \ge0 K ′ ≥ 0 and a vertex o ∈ X ^ Γ o \in\widehat{X}_{\Gamma} o ∈ X Γ such that X ^ Γ ∖ B K ′ ( o ) \widehat{X}_{\Gamma} \setminus B_{K'}(o) X Γ ∖ B K ′ ( o ) has at least two unbounded components. Moreover, at least two unbounded components contain quasi-geodesic rays.
Proof. If condition (a) fails, then Γ \Gamma Γ is disconnected or has a separating clique (the clique case is excluded by hypothesis). If condition (b) fails, then Γ \Gamma Γ has a separating nontrivial join. Finally, suppose condition (c) fails, so that lk ( v ) ⊂ J \operatorname{lk}(v) \subset J lk ( v ) ⊂ J for some vertex v v v and some nontrivial join J J J . Because Γ \Gamma Γ itself is not a nontrivial join, Γ ∖ st ( v ) \Gamma\setminus\operatorname{st}(v) Γ ∖ st ( v ) is nonempty. Thus lk ( v ) \operatorname{lk}(v) lk ( v ) separates v v v from Γ ∖ st ( v ) \Gamma\setminus\operatorname{st}(v) Γ ∖ st ( v ) . In every case there is a separating subgraph Δ ⊂ Γ \Delta\subset\Gamma Δ ⊂ Γ that is either a clique or is contained in a nontrivial join. When Γ \Gamma Γ is disconnected, we take Δ \Delta Δ to be the empty clique.
Let Γ 0 \Gamma_0 Γ 0 be the union of Δ \Delta Δ and a component of Γ ∖ Δ \Gamma\setminus\Delta Γ ∖ Δ . Let Γ 1 = ( Γ ∖ Γ 0 ) ∪ Δ \Gamma_1 = (\Gamma\setminus\Gamma_0) \cup\Delta Γ 1 = ( Γ ∖ Γ 0 ) ∪ Δ . Hence Γ \Gamma Γ is the union Γ = Γ 0 ⊔ Δ Γ 1 \Gamma= \Gamma_0 \sqcup_{\Delta} \Gamma_1 Γ = Γ 0 ⊔ Δ Γ 1 , and the right-angled Coxeter group W Γ W_{\Gamma} W Γ splits as W Γ ≅ W Γ 0 ∗ W Δ W Γ 1 W_{\Gamma} \cong W_{\Gamma_0} *_{W_{\Delta}} W_{\Gamma_1} W Γ ≅ W Γ 0 ∗ W Δ W Γ 1 .
Let X X X be the 1-skeleton of the Davis complex X Γ X_{\Gamma} X Γ . We identify X X X with Cayley graph of W Γ W_{\Gamma} W Γ with respect to the standard generating set V ( Γ ) V(\Gamma) V ( Γ ) and let X Δ ⊂ X X_{\Delta} \subset X X Δ ⊂ X be the subgraph spanned by the vertex set W Δ W_{\Delta} W Δ . We fix a vertex o ∈ X Δ ⊂ X o \in X_{\Delta} \subset X o ∈ X Δ ⊂ X . Let T T T be the Bass-Serre tree of W Γ W_{\Gamma} W Γ for the splitting W Γ ≅ W Γ 0 ∗ W Δ W Γ 1 W_{\Gamma} \cong W_{\Gamma_0} *_{W_{\Delta}} W_{\Gamma_1} W Γ ≅ W Γ 0 ∗ W Δ W Γ 1 . Note that the vertex set of T T T is { g W Γ i ∣ g ∈ W Γ , i = 0 , 1 } \{gW_{\Gamma_i} \mid g \in W_{\Gamma}, i = 0,1\} { g W Γ i ∣ g ∈ W Γ , i = 0 , 1 } , the edge set of T T T is { g W Δ ∣ g ∈ W Γ } \{gW_{\Delta} \mid g \in W_{\Gamma}\} { g W Δ ∣ g ∈ W Γ } , and g W Δ gW_{\Delta} g W Δ connects g W Γ 0 gW_{\Gamma_0} g W Γ 0 and g W Γ 1 gW_{\Gamma_1} g W Γ 1 . W Γ W_{\Gamma} W Γ acts on T T T by automorphisms.
Now we want to show X Δ X_{\Delta} X Δ coarsely separates X X X . Let X ′ X' X ′ be the barycentric subdivision of X X X and T ′ T' T ′ be the barycentric subdivision of T T T . Let e e e be the edge of T T T stabilized by W Δ W_{\Delta} W Δ . The midpoint a a a of e e e is a vertex of T ′ T' T ′ . Since every vertex of X X X is a group element of W Γ W_{\Gamma} W Γ , we define a map p : X ( 0 ) → T ′ p : X^{(0)} \to T' p : X ( 0 ) → T ′ by p ( g ) = g ⋅ a p(g) = g \cdot a p ( g ) = g ⋅ a . For every generator v ∈ V ( Γ ) v \in V(\Gamma) v ∈ V ( Γ ) , v ⋅ a = a v \cdot a = a v ⋅ a = a if v ∈ V ( Δ ) v \in V(\Delta) v ∈ V ( Δ ) , v ⋅ a v \cdot a v ⋅ a is the midpoint of the edge connecting W Γ 0 W_{\Gamma_0} W Γ 0 and v W Γ 1 vW_{\Gamma_1} v W Γ 1 if v ∈ V ( Γ 0 ∖ Δ ) v \in V(\Gamma_0 \setminus\Delta) v ∈ V ( Γ 0 ∖ Δ ) , and v ⋅ a v \cdot a v ⋅ a is the midpoint of the edge connecting W Γ 1 W_{\Gamma_1} W Γ 1 and v W Γ 0 vW_{\Gamma_0} v W Γ 0 if v ∈ V ( Γ 1 ∖ Δ ) v \in V(\Gamma_1 \setminus\Delta) v ∈ V ( Γ 1 ∖ Δ ) . Therefore, d T ′ ( g ⋅ a , g v ⋅ a ) = 0 d_{T'}(g \cdot a,gv \cdot a) = 0 d T ′ ( g ⋅ a , g v ⋅ a ) = 0 or 2 2 2 for each generator v ∈ V ( Γ ) v \in V(\Gamma) v ∈ V ( Γ ) and every group element g ∈ W Γ g \in W_{\Gamma} g ∈ W Γ . Then every edge ( g , g v ) (g,gv) ( g , g v ) in X X X can be mapped by p p p to either a point or a path with length 2 2 2 in T ′ T' T ′ . Hence, we can extend p p p to a simplicial map from X ′ X' X ′ to T ′ T' T ′ . By construction, p − 1 ( a ) = X Δ p^{-1}(a) = X_{\Delta} p − 1 ( a ) = X Δ . Furthermore, for vertices g , h ∈ X g,h \in X g , h ∈ X , d T ′ ( p ( g ) , p ( h ) ) ≤ d X ′ ( g , h ) d_{T'}(p(g),p(h)) \le d_{X'}(g,h) d T ′ ( p ( g ) , p ( h )) ≤ d X ′ ( g , h ) .
Let s ∈ Γ 0 ∖ Δ s \in\Gamma_0 \setminus\Delta s ∈ Γ 0 ∖ Δ and t ∈ Γ 1 ∖ Δ t \in\Gamma_1 \setminus\Delta t ∈ Γ 1 ∖ Δ . In particular, s s s and t t t are non-adjacent. Let Π \Pi Π be the maximal subgraph of Γ \Gamma Γ consisting of cone vertices of Γ \Gamma Γ . Let Λ = Γ ∖ Π \Lambda= \Gamma\setminus\Pi Λ = Γ ∖ Π . Assume V ( Λ ) = { s , t , v 1 , … , v n } V(\Lambda) = \{s,t,v_1,\ldots,v_n\} V ( Λ ) = { s , t , v 1 , … , v n } . Let g = s v 1 ⋯ v n t g = sv_1 \cdots v_nt g = s v 1 ⋯ v n t and h = t v 1 ⋯ v n s h = tv_1 \cdots v_ns h = t v 1 ⋯ v n s . g m g^m g m and h m h^m h m are geodesic for all m ∈ N m \in\mathbb{N} m ∈ N . Let m ∈ N m \in\mathbb{N} m ∈ N . Since the first letter of g g g is s ∈ Γ 0 ∖ Δ s \in\Gamma_0 \setminus\Delta s ∈ Γ 0 ∖ Δ , the geodesic in T ′ T' T ′ from a a a to p ( g m ) p(g^m) p ( g m ) passes through the vertex W Γ 0 W_{\Gamma_0} W Γ 0 . Similarly, the geodesic from a a a to p ( h m ) p(h^m) p ( h m ) passes through W Γ 1 W_{\Gamma_1} W Γ 1 . Hence, p ( g m ) p(g^m) p ( g m ) and p ( h m ) p(h^m) p ( h m ) are in different components of T ′ ∖ { a } T' \setminus\{a\} T ′ ∖ { a } . Let β \beta β be a path from g m g^m g m to h m h^m h m in X X X . β \beta β has a unique image in X ′ X' X ′ . Then p ( β ) p(\beta) p ( β ) passes through a a a . Hence, β \beta β intersects X Δ X_{\Delta} X Δ . By the construction of g g g and h h h , d T ′ ( a , g ⋅ a ) ≥ 2 d_{T'}(a,g \cdot a) \ge2 d T ′ ( a , g ⋅ a ) ≥ 2 , d T ′ ( a , h ⋅ a ) ≥ 2 d_{T'}(a,h \cdot a) \ge2 d T ′ ( a , h ⋅ a ) ≥ 2 and they have cyclically reduced amalgam normal forms of positive syllable length. Hence, g g g and h h h act on T ′ T' T ′ loxodromically with translation length ≥ 2 \ge2 ≥ 2 and d T ′ ( a , g m ⋅ a ) ≥ m d_{T'}(a,g^m \cdot a) \ge m d T ′ ( a , g m ⋅ a ) ≥ m and d T ′ ( a , h m ⋅ a ) ≥ m d_{T'}(a,h^m \cdot a) \ge m d T ′ ( a , h m ⋅ a ) ≥ m . Since p p p is distance non-increasing, d X ( g m , X Δ ) ≥ m / 2 d_X(g^m,X_{\Delta}) \ge m/2 d X ( g m , X Δ ) ≥ m /2 and d X ( h m , X Δ ) ≥ m / 2 d_X(h^m,X_{\Delta}) \ge m/2 d X ( h m , X Δ ) ≥ m /2 for all m ∈ N m \in\mathbb{N} m ∈ N . Hence, X Δ X_{\Delta} X Δ coarsely separates X X X .
Let γ 0 \gamma_0 γ 0 and γ 1 \gamma_1 γ 1 be the geodesic rays obtained as limits of ( g m ) (g^m) ( g m ) and ( h m ) (h^m) ( h m ) , respectively. The image in T ′ T' T ′ of every path from γ 0 \gamma_0 γ 0 to γ 1 \gamma_1 γ 1 passes through a a a , so every such path meets X Δ X_\Delta X Δ . Thus the rays lie in different deep components of X ∖ X Δ X \setminus X_\Delta X ∖ X Δ . Because Label ( g ) = Label ( h ) = V ( Λ ) \operatorname{Label}(g) = \operatorname{Label}(h) = V(\Lambda) Label ( g ) = Label ( h ) = V ( Λ ) is not contained in a nontrivial join, both rays are contracting, as in Lemma 3.4 .
Let π : X → X ^ \pi: X \to\widehat{X} π : X → X be the inclusion map. Clearly, π ( X Δ ) \pi(X_\Delta) π ( X Δ ) has uniformly bounded diameter. Indeed, if Δ \Delta Δ is a clique, then diam ( π ( X Δ ) ) ≤ M \operatorname{diam}(\pi(X_\Delta)) \le M diam ( π ( X Δ )) ≤ M , where M M M is the maximum cardinality of a clique in Γ \Gamma Γ . And if Δ \Delta Δ is a subgraph of a nontrivial join, then diam ( π ( X Δ ) ) ≤ 1 \operatorname{diam}(\pi(X_\Delta)) \le1 diam ( π ( X Δ )) ≤ 1 . Then we let K = M + 2 K = M + 2 K = M + 2 . We want to show that X ^ ∖ B K ( o ) \widehat{X} \setminus B_K(o) X ∖ B K ( o ) has at least two unbounded components.
Let x , y x, y x , y be vertices in distinct deep components of X ∖ X Δ X \setminus X_\Delta X ∖ X Δ such that the images π ( x ) \pi(x) π ( x ) and π ( y ) \pi(y) π ( y ) in X ^ \widehat{X} X are not contained in the ball B K ( o ) B_K(o) B K ( o ) . We identify the vertex set of X X X with the vertex set of X ^ \widehat{X} X . Suppose there is a path x = x 0 , ⋯ , x n = y x = x_0, \cdots, x_n = y x = x 0 , ⋯ , x n = y from x x x to y y y in X ^ \widehat{X} X avoiding B K ( o ) B_K(o) B K ( o ) . For each i = 1 , ⋯ , n i = 1, \cdots, n i = 1 , ⋯ , n , the vertices x i − 1 x_{i-1} x i − 1 and x i x_i x i in X X X are either connected by a 1-cube or contained in the same standard product region. Let α i \alpha_i α i be a combinatorial geodesic in X X X connecting x i − 1 x_{i-1} x i − 1 and x i x_i x i for each i i i . Then, the path α = α 1 ⋯ α n \alpha= \alpha_1 \cdots\alpha_n α = α 1 ⋯ α n obtained by concatenation is a path in X X X from x x x to y y y . Since x i x_i x i ’s are not in X Δ X_\Delta X Δ , there is a pair x i − 1 x_{i-1} x i − 1 and x i x_i x i that are contained in different components of X ∖ X Δ X \setminus X_\Delta X ∖ X Δ for some i i i and the intersection α i ∩ X Δ \alpha_i \cap X_\Delta α i ∩ X Δ is nonempty. In particular, α i \alpha_i α i is not a 1-cube. Then x i − 1 x_{i-1} x i − 1 and x i x_i x i are contained in the same standard product region P P P .
Let z ∈ α i ∩ X Δ z \in\alpha_i \cap X_\Delta z ∈ α i ∩ X Δ . Since α i \alpha_i α i is geodesic and P P P is convex, α i ⊂ P \alpha_i \subset P α i ⊂ P . Hence, d X ^ ( z , x i ) = 1 d_{\widehat{X}}(z, x_i) = 1 d X ( z , x i ) = 1 . Since diam X ^ ( π ( X Δ ) ) ≤ M \operatorname{diam}_{\widehat{X}}(\pi(X_\Delta)) \le M diam X ( π ( X Δ )) ≤ M , d X ^ ( o , x i ) ≤ M + 1 ≤ K d_{\widehat{X}}(o, x_i) \le M + 1 \le K d X ( o , x i ) ≤ M + 1 ≤ K . Contradiction. Therefore, the images under π \pi π of distinct deep components of X ∖ X Δ X \setminus X_\Delta X ∖ X Δ are separated by B K ( o ) B_K(o) B K ( o ) in X ^ \widehat{X} X . By [1 Theorem D] , π ( γ 0 ) \pi(\gamma_0) π ( γ 0 ) and π ( γ 1 ) \pi(\gamma_1) π ( γ 1 ) are quasi-geodesic rays in distinct components of X ^ ∖ B K ( o ) \widehat{X} \setminus B_K(o) X ∖ B K ( o ) . □
We make a remark on right-angled Artin groups.
Remark 3.7. By an argument similar to the one above, we could show that the maximized hyperbolic space of a right-angled Artin group A Γ A_\Gamma A Γ is either empty, or has a bounded separating set, hence a disconnected Gromov boundary as follows. If the defining graph Γ \Gamma Γ is disconnected, then A Γ A_\Gamma A Γ has infinitely many ends. In this case, a bounded separating set in A Γ A_\Gamma A Γ has an image in the maximized hyperbolic space which is also separating. If the defining graph Γ \Gamma Γ is a join, then A Γ A_\Gamma A Γ is a direct product of two infinite subgroups. Hence the maximized hyperbolic space has finite diameter and the Gromov boundary is empty. Now suppose Γ \Gamma Γ is connected, not a join and has more than one vertex. Then the star of each vertex generates an unbounded product subgroup, i.e. for a vertex v ∈ V ( Γ ) v \in V(\Gamma) v ∈ V ( Γ ) , A s t ( v ) ≅ A v × A l k ( v ) A_{\mathrm{st}(v)} \cong A_v \times A_{\mathrm{lk}(v)} A st ( v ) ≅ A v × A lk ( v ) where both A v A_v A v and A l k ( v ) A_{\mathrm{lk}(v)} A lk ( v ) are infinite. The stabilizer of each hyperplane of type v v v is a conjugate of A l k ( v ) A_{\mathrm{lk}(v)} A lk ( v ) . Hence each hyperplane is contained in a standard product region, a left coset of a special subgroup containing A s t ( v ) A_{\mathrm{st}(v)} A st ( v ) . Then the image of A l k ( v ) A_{\mathrm{lk}(v)} A lk ( v ) has a bounded diameter in the maximized hyperbolic space. On the other hand, let Γ 0 = s t ( v ) \Gamma_0 = \mathrm{st}(v) Γ 0 = st ( v ) and Γ 1 = Γ ∖ { v } \Gamma_1 = \Gamma\setminus\{v\} Γ 1 = Γ ∖ { v } . This provides a group splitting A Γ ≅ A Γ 0 ∗ A l k ( v ) A Γ 1 A_\Gamma\cong A_{\Gamma_0} *_{A_{\mathrm{lk}(v)}} A_{\Gamma_1} A Γ ≅ A Γ 0 ∗ A lk ( v ) A Γ 1 . Hence A Γ A_\Gamma A Γ is coarsely separated by A l k ( v ) A_{\mathrm{lk}(v)} A lk ( v ) . Therefore the image of A l k ( v ) A_{\mathrm{lk}(v)} A lk ( v ) in the maximized hyperbolic space is a bounded separating subset.
We remark that the maximized hyperbolic space of an RAAG is quasi-isometric to the contact graph and also to the extension graph ([1 Section 5] ), which is a quasi-tree.
We conclude the section with the forward implication of Theorem 1.1 .
Proposition 3.8. Let W Γ W_\Gamma W Γ be a right-angled Coxeter group. Suppose that at least one of the following holds:
Γ \Gamma Γ is a clique, is disconnected, or has a separating clique;
Γ \Gamma Γ has a separating nontrivial join;
lk ( v ) \operatorname{lk}(v) lk ( v ) is contained in a nontrivial join for some v ∈ V ( Γ ) v \in V(\Gamma) v ∈ V ( Γ ) .
Then the Gromov boundary ∂ X ^ \partial\widehat{X} ∂ X of the maximized hyperbolic space is either empty or disconnected. Moreover, ∂ X ^ \partial\widehat{X} ∂ X is empty if and only if Γ \Gamma Γ is a clique or a nontrivial join.
Proof. The final assertion follows from Lemma 3.4. Assume that Γ \Gamma Γ is neither a clique nor a nontrivial join, and fix o ∈ X ^ o \in\widehat{X} o ∈ X . By Proposition 3.6, some complement X ^ ∖ B K ( o ) \widehat{X} \setminus B_K(o) X ∖ B K ( o ) has more than one unbounded component. Let ( A i ) i ∈ N (A_i)_{i \in\mathbb{N}} ( A i ) i ∈ N be its unbounded components, and let ∂ A i \partial A_i ∂ A i consist of boundary classes represented by sequences eventually contained in A i A_i A i . We may assume that A 1 A_1 A 1 and A 2 A_2 A 2 contain quasi-geodesic rays, so ∂ A 1 \partial A_1 ∂ A 1 and ∂ A 2 \partial A_2 ∂ A 2 are nonempty. We claim that ∂ X ^ = ⋃ i ∈ N ∂ A i \partial\widehat{X} = \bigcup_{i \in\mathbb{N}} \partial A_i ∂ X = ⋃ i ∈ N ∂ A i .
Let ( x n ) (x_n) ( x n ) represent a point of ∂ X ^ \partial\widehat{X} ∂ X . It suffices to show that the sequence eventually lies in a single component A i A_i A i . Otherwise, some subsequence visits more than one component. Whenever two terms of this subsequence lie in different components, every geodesic between them intersects B K ( o ) B_K(o) B K ( o ) . Hyperbolicity then gives a constant K ′ = K ′ ( K , δ ) K' = K'(K,\delta) K ′ = K ′ ( K , δ ) such that their Gromov product based at o o o is at most K ′ K' K ′ . This contradicts the fact that ( x n ) (x_n) ( x n ) converges at infinity.
If ( y n ) (y_n) ( y n ) represents a point of ∂ A 1 \partial A_1 ∂ A 1 and ( z n ) (z_n) ( z n ) represents a point of ∂ A i \partial A_i ∂ A i with i ≠ 1 i \ne1 i = 1 , then ⟨ y n , z m ⟩ o ≤ K ′ \langle y_n,z_m\rangle_o \le K' ⟨ y n , z m ⟩ o ≤ K ′ for all sufficiently large n , m n,m n , m . Hence any visual metric separates ∂ A 1 \partial A_1 ∂ A 1 uniformly from the other ∂ A i \partial A_i ∂ A i . Thus ∂ A 1 \partial A_1 ∂ A 1 is a nonempty proper open and closed subset of ∂ X ^ \partial\widehat{X} ∂ X . Therefore, the boundary is disconnected. □ \square □
Morse boundaries and Connectedness of Gromov boundaries
In this section, we deduce the connectivity of certain Gromov boundaries from the connectivity of Morse boundaries. We record a general formulation for cobounded Morse-recognizing actions on hyperbolic spaces, using known results on boundary maps and standard facts about loxodromic dynamics.
We apply this formulation to maximized hyperbolic spaces of HHGs and to the hyperbolic models for Coxeter groups constructed by Cashen–Chu–Tao [20 ] . For right-angled Coxeter groups, which are contained in both categories, this gives an alternative proof of the connectedness conclusion of Theorem 1.11 . The direct geometric argument in Section 9 establishes the stronger conclusion of linear connectedness.
Density for Morse-recognizing actions
Let G G G be a finitely generated group and Cay ( G , A ) \operatorname{Cay}(G,A) Cay ( G , A ) be the Cayley graph of G G G with respect to a finite generating set A A A . We denote by ∂ M G \partial_M G ∂ M G the Morse boundary of G G G with the direct limit topology. We refer readers to [26 ] for details of the Morse boundary.
Suppose that G G G acts on a geodesic hyperbolic space Y Y Y by isometries. Denote by ∂ Y \partial Y ∂ Y the Gromov boundary of Y Y Y . Fix y 0 ∈ Y y_0 \in Y y 0 ∈ Y and consider the orbit map
π : G ⟶ Y , π ( g ) = g ⋅ y 0 . \pi\colon G \longrightarrow Y,\qquad\pi(g)=g\cdot y_0. π : G ⟶ Y , π ( g ) = g ⋅ y 0 .
Note that π \pi π is Lipschitz since G G G is finitely generated.
Definition 4.1. We say that the orbit map π \pi π is Morse recognizing if, for every ( k , c ) (k,c) ( k , c ) -quasi-geodesic γ : I → Cay ( G , A ) \gamma\colon I \to\operatorname{Cay}(G,A) γ : I → Cay ( G , A ) , the following hold:
For each Morse gauge N N N , if γ \gamma γ is N N N -Morse, then π ∘ γ \pi\circ\gamma π ∘ γ is a quasi-geodesic with constants depending only on N N N , k k k , and c c c .
If π ∘ γ \pi\circ\gamma π ∘ γ is a ( K , C ) (K,C) ( K , C ) -quasi-geodesic, then γ \gamma γ is Morse with a gauge depending only on k k k , c c c , K K K , and C C C .
The interval I I I may be bounded, a ray, or the whole real line.
This is the Morse-detectability property from [68 Definition A.4] , applied to an orbit map.
Proposition 4.2 (Density for Morse-recognizing actions). Let G G G be a finitely generated group acting coboundedly on a geodesic hyperbolic space Y Y Y by isometries. Suppose that the orbit map π : G → Y \pi:G \to Y π : G → Y is Morse recognizing. Then π \pi π induces a natural G G G -equivariant continuous injection
ι : ∂ M G ⟶ ∂ Y \iota:\partial_M G \longrightarrow\partial Y ι : ∂ M G ⟶ ∂ Y with a dense image.
Proof. By [68 Lemma A.6] , Morse recognition gives a continuous injection
ι : ∂ M G ⟶ ∂ Y . \iota:\partial_M G \longrightarrow\partial Y. ι : ∂ M G ⟶ ∂ Y .
For a Morse geodesic ray γ \gamma γ in G G G , the point ι ( [ γ ] ) \iota([\gamma]) ι ([ γ ]) is the endpoint of the quasi-geodesic ray π ∘ γ \pi\circ\gamma π ∘ γ . By the definition of the orbit map π \pi π , the induced boundary map is G G G -equivariant.
Now we want to show the image ι ( ∂ M G ) \iota(\partial_M G) ι ( ∂ M G ) is dense in ∂ Y \partial Y ∂ Y by proving that it contains all fixed points of loxodromic elements and that these points are dense in ∂ Y \partial Y ∂ Y .
If Y Y Y has finite diameter, then ∂ Y \partial Y ∂ Y is empty. Moreover, ∂ M G \partial_M G ∂ M G is empty, since a Morse geodesic ray in Cay ( G , A ) \operatorname{Cay}(G,A) Cay ( G , A ) would map to an unbounded quasi-geodesic ray in Y Y Y . Thus the conclusion holds in this case. Henceforth assume that Y Y Y has infinite diameter.
Let g ∈ G g \in G g ∈ G act loxodromically on Y Y Y with fixed points g + , g − ∈ ∂ Y g^+,g^- \in\partial Y g + , g − ∈ ∂ Y . There exist constants a > 0 a > 0 a > 0 and b ≥ 0 b \ge0 b ≥ 0 such that
a ∣ m − n ∣ − b ≤ d Y ( g m y 0 , g n y 0 ) for all m , n ∈ Z . a\lvert m-n\rvert-b \le d_Y(g^m y_0,g^n y_0) \qquad\text{for all }m,n \in\mathbb{Z}. a ∣ m − n ∣ − b ≤ d Y ( g m y 0 , g n y 0 ) for all m , n ∈ Z .
Set
L = max ( { 1 } ∪ { d Y ( y 0 , s y 0 ) : s ∈ A } ) . L=\max\left(\{1\}\cup\{d_Y(y_0,sy_0):s\in A\}\right). L = max ( { 1 } ∪ { d Y ( y 0 , s y 0 ) : s ∈ A } ) .
Since the orbit map π \pi π is L L L -Lipschitz, we have
a ∣ m − n ∣ − b L ≤ d G ( g m , g n ) ≤ ∣ g ∣ A ∣ m − n ∣ . \frac{a\lvert m-n\rvert-b}{L}\le d_G(g^m,g^n)\le\lvert g\rvert_A\lvert m-n\rvert. L a ∣ m − n ∣ − b ≤ d G ( g m , g n ) ≤ ∣ g ∣ A ∣ m − n ∣ .
Hence, n ↦ g n n \mapsto g^n n ↦ g n is a quasi-isometric embedding of Z \mathbb{Z} Z into G G G . We extend it to a quasi-geodesic γ g : R → Cay ( G , A ) \gamma_g:\mathbb{R}\to\operatorname{Cay}(G,A) γ g : R → Cay ( G , A ) such that γ g ( n ) = g n \gamma_g(n)=g^n γ g ( n ) = g n by concatenating consecutive pairs g n , g n + 1 g^n,g^{n+1} g n , g n + 1 by a fixed geodesic path representing the word g ∈ G g\in G g ∈ G .
Clearly, π ∘ γ g \pi\circ\gamma_g π ∘ γ g is a quasi-geodesic in Y Y Y . By the Morse-recognizing property, γ g ⊂ Cay ( G , A ) \gamma_g \subset\operatorname{Cay}(G,A) γ g ⊂ Cay ( G , A ) is a Morse quasi-geodesic. Its two half-rays determine points γ g + , γ g − ∈ ∂ M G \gamma_g^+,\gamma_g^- \in\partial_M G γ g + , γ g − ∈ ∂ M G , and
ι ( γ g + ) = g + , ι ( γ g − ) = g − . \iota(\gamma_g^+)=g^+,\qquad\iota(\gamma_g^-)=g^-. ι ( γ g + ) = g + , ι ( γ g − ) = g − .
Thus every loxodromic fixed point lies in ι ( ∂ M G ) \iota(\partial_M G) ι ( ∂ M G ) .
Let Λ Y ( G ) \Lambda_Y(G) Λ Y ( G ) denote the limit set of the action G ↷ Y G \curvearrowright Y G ↷ Y . Coboundedness implies
Λ Y ( G ) = ∂ Y . \Lambda_Y(G)=\partial Y. Λ Y ( G ) = ∂ Y .
Indeed, any sequence representing a boundary point can be approximated within a uniform distance by points of the orbit $Gy_0. By [2 Theorem 4.2] , an unbounded cobounded action has either a two-point or an infinite limit set. In the two-point case, these points are the fixed points of a loxodromic element, so both belong to ι ( ∂ M G ) \iota(\partial_M G) ι ( ∂ M G ) . In the infinite case, the attracting fixed points of loxodromic elements are dense in Λ Y ( G ) \Lambda_Y(G) Λ Y ( G ) [33 Proposition 7.4.6] . Since these points belong to ι ( ∂ M G ) \iota(\partial_M G) ι ( ∂ M G ) and Λ Y ( G ) = ∂ Y \Lambda_Y(G)=\partial Y Λ Y ( G ) = ∂ Y , the image of ι \iota ι is dense in ∂ Y \partial Y ∂ Y .
Corollary 4.3. Under the hypotheses of Proposition 4.2, if ∂ M G \partial_M G ∂ M G is connected, then ∂ Y \partial Y ∂ Y is connected. If ∂ M G \partial_M G ∂ M G is also nonempty, then ∂ Y \partial Y ∂ Y is nonempty.
Proof. The continuous image ι ( ∂ M G ) \iota(\partial_M G) ι ( ∂ M G ) is connected. By Proposition 4.2 ,
∂ Y = ι ( ∂ M G ) ‾ . \partial Y=\overline{\iota(\partial_M G)}. ∂ Y = ι ( ∂ M G ) .
The closure of a connected subset is connected, so ∂ Y \partial Y ∂ Y is connected. Nonemptiness follows from the definition of ι \iota ι .
Applications to hierarchically hyperbolic groups
We apply the preceding results to the maximized structure on a hierarchically hyperbolic group. We refer readers to [3 Section 3] for related definitions.
Corollary 4.4. Let G G G be a hierarchically hyperbolic group equipped with its maximized structure, and let C S CS C S be its maximized hyperbolic space. Then the natural map
∂ M G ⟶ ∂ C S \partial_M G\longrightarrow\partial CS ∂ M G ⟶ ∂ C S is a continuous injection with a dense image. In particular, if ∂ M G \partial_M G ∂ M G is nonempty and connected, then ∂ C S \partial CS ∂ C S is nonempty and connected.
Proof. For the maximized structure, the action on C S CS C S is cobounded, and the top-level projection is Morse-recognizing by [1 Corollary 6.2] . The projection is coarsely equivariant, so it is at a uniformly bounded distance from an orbit map. This orbit map is therefore also Morse recognizing. The conclusions follow from Proposition 4.2 and Corollary 4.3 .
Remark 4.5. For an HHG containing a Morse element, the density assertion has been made in [31 ] and [3 Remark 5.11] . To see this, let T \mathcal{T} T be the maximized structure with nested-maximal hyperbolic space C S CS C S and let ∂ ( G , T ) \partial(G,\mathcal{T}) ∂ ( G , T ) be its HHS boundary. Both maps in
∂ M G ↪ ∂ C S ↪ ∂ ( G , T ) \partial_M G\hookrightarrow\partial CS\hookrightarrow\partial(G,\mathcal{T}) ∂ M G ↪ ∂ C S ↪ ∂ ( G , T )
are continuous injections with dense images.
Hyperbolic models for Coxeter groups
Proposition 4.2 also applies to hyperbolic models for Coxeter groups.
A recent result by Cashen–Chu–Tao shows that every Coxeter group admits a hyperbolic space which is Morse recognizing in [20 ] . Let ( W , S ) (W,S) ( W , S ) be a finite-rank Coxeter system with Davis complex Σ \Sigma Σ . The hyperbolic model Σ ^ \widehat{\Sigma} Σ is obtained by coning off the wide parabolic subcomplexes of Σ \Sigma Σ [20 Definition 4.18] . Note that if the Coxeter group is right-angled, this hyperbolic model Σ ^ \widehat{\Sigma} Σ is quasi-isometric to the maximized hyperbolic space X ^ \widehat{X} X (Definition 2.22).
Corollary 4.6. Let ( W , S ) (W,S) ( W , S ) be a finite-rank Coxeter system, and let Σ ^ \widehat{\Sigma} Σ be the hyperbolic model of [20 ] . Then the natural map
∂ M W ⟶ ∂ Σ ^ \partial_M W\longrightarrow\partial\widehat{\Sigma} ∂ M W ⟶ ∂ Σ is a continuous injection with dense image. Consequently, if ∂ M W \partial_M W ∂ M W is nonempty and connected, then ∂ Σ ^ \partial\widehat{\Sigma} ∂ Σ is nonempty and connected.
Proof. Equip Σ ( 1 ) \Sigma^{(1)} Σ ( 1 ) with its combinatorial metric and identify it with Cay ( W , S ) \operatorname{Cay}(W,S) Cay ( W , S ) . By [20 Corollary 4.20] , the space Σ ^ \widehat{\Sigma} Σ is geodesic and hyperbolic, and the inclusion
Σ ( 1 ) ⟶ Σ ^ \Sigma^{(1)} \longrightarrow\widehat{\Sigma} Σ ( 1 ) ⟶ Σ
is Morse recognizing. Clearly, W W W acts on Σ ^ \widehat{\Sigma} Σ coboundedly. The conclusions follow from Proposition 4.2 and Corollary 4.3 . □
Combining this result with [24 Corollary 4.10] gives the following corollary.
Corollary 4.7. Let W W W be a one-ended, wide-spherical-avoidant Coxeter group. Then the Gromov boundary ∂ Σ ^ \partial\widehat{\Sigma} ∂ Σ of its Morse-recognizing space is connected.
For the definition of wide-spherical-avoidant , we refer readers to [25 Definition 3.5] .
In the case of right-angled Coxeter groups, combining the preceding corollaries with the characterization of Cordes–Levcovitz recovers the connectedness implication of our main theorem.
Corollary 4.8. Let W Γ W_{\Gamma} W Γ be a right-angled Coxeter group whose defining graph satisfies conditions (a)–(c) of Theorem 1.1 . Then ∂ X ^ Γ \partial\widehat{X}_{\Gamma} ∂ X Γ is nonempty and connected.
Proof. Condition (a) of Theorem 1.1 and Theorem 2.7 imply that W Γ W_{\Gamma} W Γ is one-ended. Conditions (b) and (c) then imply that Γ \Gamma Γ is wide-avoidant, by Lemma 2.14 .
By [25 Theorem B] , the Morse boundary ∂ M W Γ \partial_M W_{\Gamma} ∂ M W Γ is nonempty and connected. Corollary 4.4 or Corollary 4.6 shows that ∂ X ^ Γ \partial\widehat{X}_{\Gamma} ∂ X Γ is nonempty and connected. □
The arguments in this section establish the connectedness of the Gromov boundary. Continuity and density alone do not imply path connectedness, local connectedness (for instance, consider the topologist’s sine curve), or linear connectedness.
Relation between X and the maximized hyperbolic space
The purpose of this section is to establish tools relating the cubical geometry of X X X to the fine geometry of the maximized hyperbolic space X ^ \widehat{X} X . These tools will be used to establish visibility in § 7 and to control the fan-and-filter construction in § 8 . The main difficulty in studying X ^ \widehat{X} X is that coning off standard product regions in X X X simplifies the large-scale geometry but generally destroys properness and can produce infinitely many geodesics between two vertices in X ^ \widehat{X} X ; see Example 6.2 . We therefore need controlled representatives of geodesics in X ^ \widehat{X} X , together with estimates that give lower bounds on distance in X ^ \widehat{X} X in terms of combinatorial information in X X X .
Recall that we identify the vertex set of the Davis complex X X X with the vertex set of X ^ \widehat{X} X via the natural inclusion X ↪ X ^ X \hookrightarrow\widehat{X} X ↪ X .
This section proceeds in three parts. First, we introduce taut geodesics in X ^ \widehat{X} X (Definition 5.1 ) and prove their existence in Proposition 5.2 . The vertices of a taut geodesic lie in the cubical convex hull in X X X of its endpoints. Since the convex hull of two vertices is finite, only finitely many taut geodesics join any fixed pair of vertices. Thus, taut geodesics allow us to study geodesics in X ^ \widehat{X} X by using the hyperplanes of a finite cubical convex hull.
Second, we study cubical convex hulls and gate maps with respect to the metric d X ^ d_{\widehat{X}} d X . Lemma 5.5 shows that gate maps to convex subcomplexes are 1 1 1 –Lipschitz with respect to d X ^ d_{\widehat{X}} d X . Lemma 5.6 shows that the distance from a fixed basepoint is nondecreasing along combinatorial geodesics starting at that basepoint. We use these estimates to find a point far from the basepoint in the intersection of two convex hulls whose terminal vertices are close in X ^ \widehat{X} X ((Lemma 5.7)). We also prove that every point of a cubical convex hull lies uniformly close in X ^ \widehat{X} X to a taut geodesic joining its endpoints ((Lemma 5.8)).
Finally, we introduce wide subpaths , which record the portions of a combinatorial geodesic that lie inside standard product regions and are therefore collapsed by the coning-off procedure. Lemma 5.10 shows that a uniform bound on the lengths of the wide subpaths of one combinatorial geodesic gives a uniform bound for every combinatorial geodesic with the same endpoints. Using taut geodesics, Lemma 5.11 then converts this combinatorial information into progress in X ^ \widehat{X} X . More precisely, if the wide subpaths of every combinatorial geodesic from x x x to y y y have length at most C ′ C' C ′ , then
d X ( x , y ) ≥ n ( C ′ + 1 ) ⟹ d X ^ ( x , y ) ≥ n . d_X(x,y) \ge n(C' + 1) \quad\Longrightarrow\quad d_{\widehat{X}}(x,y) \ge n. d X ( x , y ) ≥ n ( C ′ + 1 ) ⟹ d X ( x , y ) ≥ n .
These results will control the geometry of the filters that we will construct in § 8 and ensure that sufficiently deep layers of a filter remain outside of prescribed balls in X ^ \widehat{X} X .
Fix o ∈ X ^ o \in\widehat{X} o ∈ X and write
S r ( o ) = { x ∣ d X ^ ( o , x ) = r } , B r ( o ) = { x ∣ d X ^ ( o , x ) ≤ r } . S_r(o) = \{x \mid d_{\widehat{X}}(o,x) = r\}, \qquad B_r(o) = \{x \mid d_{\widehat{X}}(o,x) \le r\}. S r ( o ) = { x ∣ d X ( o , x ) = r } , B r ( o ) = { x ∣ d X ( o , x ) ≤ r } .
Taut geodesics in X ^ \widehat{X} X
Definition 5.1 (Taut geodesic). Let x , y ∈ X ^ x,y \in\widehat{X} x , y ∈ X . A geodesic x = x 0 , … , x n = y x=x_0,\ldots,x_n=y x = x 0 , … , x n = y in X ^ \widehat{X} X is taut if every vertex x i x_i x i lies in the cubical convex hull [ x , y ] ⊂ X [x,y] \subset X [ x , y ] ⊂ X .
It is an immediate consequence of the above definition that, since the convex hull of two points in X X X is finite, for any pair of vertices x , y ∈ X ^ x,y \in\widehat{X} x , y ∈ X there are only finitely many taut geodesics from x x x to y y y in X ^ \widehat{X} X .
Now we show the existence of taut geodesics. The proof is inspired by [40 Lemma 4.11] .
Proposition 5.2 (Existence of taut geodesics). Let x , y ∈ X ^ x,y \in\widehat{X} x , y ∈ X . There is a combinatorial geodesic α \alpha α in X X X from x x x to y y y and a geodesic x = x 0 , x 1 , ⋯ , x n = y x=x_0,x_1,\cdots,x_n=y x = x 0 , x 1 , ⋯ , x n = y in X ^ \widehat{X} X such that each x i x_i x i is a vertex of α \alpha α . In particular, this geodesic in X ^ \widehat{X} X is taut.
Proof. Let x = x 0 , x 1 , ⋯ , x n = y x=x_0,x_1,\cdots,x_n=y x = x 0 , x 1 , ⋯ , x n = y be a geodesic in X ^ \widehat{X} X from x x x to y y y , where each x i x_i x i is a vertex and a consecutive pair is joined by an edge. Therefore, for each 1 ≤ i ≤ n 1 \le i \le n 1 ≤ i ≤ n , x i − 1 x_{i-1} x i − 1 and x i x_i x i are either adjacent by a 1-cube in X X X or contained in the same standard product region P P P . Let α i \alpha_i α i be a combinatorial geodesic from x i − 1 x_{i-1} x i − 1 to x i x_i x i in X X X . Note that α i \alpha_i α i is either a 1-cube containing x i − 1 x_{i-1} x i − 1 and x i x_i x i or a combinatorial geodesic contained in a product region P P P from x i − 1 x_{i-1} x i − 1 to x i x_i x i . Then α = α 1 ⋯ α n \alpha=\alpha_1\cdots\alpha_n α = α 1 ⋯ α n obtained by concatenation is a combinatorial path from x x x to y y y in X X X .
Define the complexity of the pair ( ( x 0 , … , x n ) , ( α 1 , … , α n ) ) ((x_0,\ldots,x_n),(\alpha_1,\ldots,\alpha_n)) (( x 0 , … , x n ) , ( α 1 , … , α n )) to be the n n n -tuple
( ∣ α 1 ∣ , ∣ α 2 ∣ , … , ∣ α n ∣ ) , (|\alpha_1|,|\alpha_2|,\ldots,|\alpha_n|), ( ∣ α 1 ∣ , ∣ α 2 ∣ , … , ∣ α n ∣ ) ,
ordered lexicographically. Suppose that the geodesic x 0 , x 1 , … , x n x_0,x_1,\ldots,x_n x 0 , x 1 , … , x n in X ^ \widehat{X} X and combinatorial geodesics α 1 , … , α n \alpha_1,\ldots,\alpha_n α 1 , … , α n in X X X have been chosen to minimize the complexity. This means that we prioritize α 1 \alpha_1 α 1 being as short as possible, and among the choices such that α 1 \alpha_1 α 1 is as short as possible, we minimize the length of α 2 \alpha_2 α 2 and continue. We claim that α \alpha α is a combinatorial geodesic in X X X from x x x to y y y . It is enough to show that every hyperplane W W W crosses α \alpha α at most once. Suppose a hyperplane W W W crosses α \alpha α more than once at an edge e e e and then at an edge f f f . Suppose e ⊂ α i e \subset\alpha_i e ⊂ α i and f ⊂ α j f \subset\alpha_j f ⊂ α j with i < j i<j i < j . Note that i ≠ j i \ne j i = j by construction since e ≠ f e \ne f e = f and each α i \alpha_i α i is geodesic.
Case 1: j = i + 1 j=i+1 j = i + 1 . If both α i \alpha_i α i and α i + 1 \alpha_{i+1} α i + 1 are single edges of X X X , then the two edges incident to x i x_i x i have the same label and hence coincide. Thus x i − 1 = x i + 1 x_{i-1}=x_{i+1} x i − 1 = x i + 1 , contradicting that x 0 , … , x n x_0,\ldots,x_n x 0 , … , x n is geodesic in X ^ \widehat{X} X .
Suppose, without loss of generality, that α i \alpha_i α i is a combinatorial geodesic lying in a standard product region P P P and α i + 1 \alpha_{i+1} α i + 1 is a 1-cube. By the convexity of the product region, the latter 1-cube has the same label as e e e , hence lies in P P P . Thus x i − 1 x_{i-1} x i − 1 and x i + 1 x_{i+1} x i + 1 lie in the same standard product region and are adjacent in X ^ \widehat{X} X , again contradicting geodesicity.
Now suppose both α i \alpha_i α i and α i + 1 \alpha_{i+1} α i + 1 are not 1-cubes. The hyperplane W W W separates the CAT(0) cube complex X X X into two halfspaces W ← \overleftarrow{W} W and W → \overrightarrow{W} W . Suppose x i − 1 , x i + 1 ∈ W → x_{i-1},x_{i+1}\in\overrightarrow{W} x i − 1 , x i + 1 ∈ W and x i ∈ W ← x_i\in\overleftarrow{W} x i ∈ W . Let A A A and C C C be the endpoints of e e e and f f f , respectively, that lie in W → \overrightarrow{W} W , and let B = x i B=x_i B = x i . Let m = m ( A , B , C ) m=m(A,B,C) m = m ( A , B , C ) be the median of A , B , C A,B,C A , B , C . Note that, by construction, the geodesics α i , α i + 1 \alpha_i,\alpha_{i+1} α i , α i + 1 are contained in standard product regions P i , P i + 1 P_i,P_{i+1} P i , P i + 1 , respectively. P i P_i P i and P i + 1 P_{i+1} P i + 1 are convex. Since A , B ∈ P i A,B\in P_i A , B ∈ P i , m ∈ P i m\in P_i m ∈ P i . Similarly, m ∈ P i + 1 m\in P_{i+1} m ∈ P i + 1 . Also, since A , C ∈ W → A,C\in\overrightarrow{W} A , C ∈ W and the halfspace W → \overrightarrow{W} W is convex, m ∈ W → m\in\overrightarrow{W} m ∈ W . In particular, m ≠ B m\ne B m = B since B ∈ W ← B\in\overleftarrow{W} B ∈ W . By the definition of median, there is a geodesic β i \beta_i β i from A A A to B B B passing through m m m . Thus, d ( A , m ) < d ( A , B ) d(A,m)<d(A,B) d ( A , m ) < d ( A , B ) .
Since m ∈ P i ∩ P i + 1 m\in P_i\cap P_{i+1} m ∈ P i ∩ P i + 1 , the path x 0 , … , x i − 1 , m , x i + 1 , … , x n x_0,\ldots,x_{i-1},m,x_{i+1},\ldots,x_n x 0 , … , x i − 1 , m , x i + 1 , … , x n is a geodesic in X ^ \widehat{X} X . Let α i ′ \alpha_i' α i ′ be the combinatorial geodesic from x i − 1 x_{i-1} x i − 1 to m m m inside P i P_i P i . Clearly, ∣ α i ′ ∣ < ∣ α i ∣ |\alpha_i'|<|\alpha_i| ∣ α i ′ ∣ < ∣ α i ∣ , hence contradicting the lexicographical minimality of α \alpha α .
Case 2: j > i + 1 j>i+1 j > i + 1 . Choose e e e and f f f innermost, so the subpath β \beta β between them crosses no hyperplane twice and is therefore geodesic. Since each combinatorial hyperplane is convex, β \beta β is contained in a combinatorial hyperplane of W W W , say W × { 1 2 } W\times\{\frac{1}{2}\} W × { 2 1 } . Sliding β \beta β across the carrier of W W W gives a parallel geodesic β ′ \beta' β ′ in the opposite combinatorial hyperplane W × { − 1 2 } W\times\{-\frac{1}{2}\} W × { − 2 1 } . For i + 1 ≤ k ≤ j − 1 i+1\le k\le j-1 i + 1 ≤ k ≤ j − 1 , let x k ′ x_k' x k ′ be the vertex opposite x k x_k x k across W W W . Convexity of the carrier shows that x k ′ x_k' x k ′ and x k + 1 ′ x_{k+1}' x k + 1 ′ are adjacent in X ^ \widehat{X} X exactly when x k x_k x k and x k + 1 x_{k+1} x k + 1 are. Replacing the portion of the path between e e e and f f f by β ′ \beta' β ′ preserves the length of the X ^ \widehat{X} X -geodesic but shortens the first affected subpath α i \alpha_i α i , contradicting lexicographic minimality.
Therefore, α \alpha α is a geodesic from x x x to y y y in X X X . □
Remark 5.3. By Proposition 5.2 , we may choose a taut geodesic x = x 0 , … , x n = y x=x_0,\ldots,x_n=y x = x 0 , … , x n = y whose vertices occur, in this order, on a combinatorial geodesic in X X X . Consequently, each hyperplane of X X X separates at most one consecutive pair x i , x i + 1 x_i,x_{i+1} x i , x i + 1 .
Remark 5.4. Suppose d X ^ ( x , y ) = n d_{\widehat{X}}(x,y)=n d X ( x , y ) = n . Then by the existence of taut geodesics, there is x ′ ∈ [ x , y ] x'\in[x,y] x ′ ∈ [ x , y ] with d X ^ ( x , x ′ ) = n − 1 d_{\widehat{X}}(x,x')=n-1 d X ( x , x ′ ) = n − 1 and d X ^ ( x ′ , y ) = 1 d_{\widehat{X}}(x',y)=1 d X ( x ′ , y ) = 1 . Equivalently, if every x ′ ∈ [ x , y ] ∩ S 1 ( y ) x'\in[x,y]\cap S_1(y) x ′ ∈ [ x , y ] ∩ S 1 ( y ) is at distance at least n − 1 n-1 n − 1 from x x x , then d X ^ ( x , y ) ≥ n d_{\widehat{X}}(x,y)\ge n d X ( x , y ) ≥ n . This will be an essential observation used in Lemma 5.11.
Image of cubical convex hulls in X ^ \widehat{X} X
Now we show that the gate map g \mathfrak{g} g from X X X to a convex subcomplex is 1-Lipschitz with respect to the d X ^ d_{\widehat{X}} d X metric.
Lemma 5.5 (Gate maps are 1-Lipschitz). Let x , y ∈ X x,y\in X x , y ∈ X with d X ^ ( x , y ) = n d_{\widehat{X}}(x,y)=n d X ( x , y ) = n . Let C ⊂ X C\subset X C ⊂ X be a convex subcomplex of X X X . Let x ′ = g C ( x ) x'=\mathfrak{g}_C(x) x ′ = g C ( x ) and y ′ = g C ( y ) y'=\mathfrak{g}_C(y) y ′ = g C ( y ) be the closest point projections. Then d X ^ ( x ′ , y ′ ) ≤ n d_{\widehat{X}}(x',y')\le n d X ( x ′ , y ′ ) ≤ n .
Proof. We argue by induction on the distance n = d X ^ ( x , y ) n=d_{\widehat{X}}(x,y) n = d X ( x , y ) . Suppose n = 1 n=1 n = 1 . By the property of closest point projection, every hyperplane separating x ′ x' x ′ and y ′ y' y ′ also separates x x x and y y y . Thus, Label ( [ x ′ , y ′ ] ) ⊂ Label ( [ x , y ] ) \operatorname{Label}([x',y']) \subset\operatorname{Label}([x,y]) Label ([ x ′ , y ′ ]) ⊂ Label ([ x , y ]) . Then d X ^ ( x ′ , y ′ ) ≤ 1 d_{\widehat{X}}(x',y') \le1 d X ( x ′ , y ′ ) ≤ 1 .
Assume d X ^ ( x ′ , y ′ ) ≤ n − 1 d_{\widehat{X}}(x',y') \le n-1 d X ( x ′ , y ′ ) ≤ n − 1 whenever d X ^ ( x , y ) ≤ n − 1 d_{\widehat{X}}(x,y) \le n-1 d X ( x , y ) ≤ n − 1 . Now suppose d X ^ ( x , y ) = n d_{\widehat{X}}(x,y)=n d X ( x , y ) = n . Then there is z ∈ [ x , y ] z \in[x,y] z ∈ [ x , y ] such that d X ^ ( x , z ) = n − 1 d_{\widehat{X}}(x,z)=n-1 d X ( x , z ) = n − 1 and d X ^ ( z , y ) = 1 d_{\widehat{X}}(z,y)=1 d X ( z , y ) = 1 by Proposition 5.2 . Let z ′ = g C ( z ) z'=g_C(z) z ′ = g C ( z ) . By the induction hypothesis, d X ^ ( x ′ , z ′ ) ≤ n − 1 d_{\widehat{X}}(x',z') \le n-1 d X ( x ′ , z ′ ) ≤ n − 1 . For the same reason as above, d X ^ ( z ′ , y ′ ) ≤ 1 d_{\widehat{X}}(z',y') \le1 d X ( z ′ , y ′ ) ≤ 1 . Hence d X ^ ( x ′ , y ′ ) ≤ d X ^ ( x ′ , z ′ ) + d X ^ ( z ′ , y ′ ) ≤ n d_{\widehat{X}}(x',y') \le d_{\widehat{X}}(x',z')+d_{\widehat{X}}(z',y') \le n d X ( x ′ , y ′ ) ≤ d X ( x ′ , z ′ ) + d X ( z ′ , y ′ ) ≤ n .
The existence of taut geodesics allows us to study d X ^ ( x , y ) d_{\widehat{X}}(x,y) d X ( x , y ) using only the vertices of the cubical convex hull [ x , y ] ⊂ X [x,y]\subset X [ x , y ] ⊂ X .
Lemma 5.6 (Monotonicity in a convex hull). Let o , x o,x o , x be vertices in X ^ \widehat{X} X . For every y ∈ [ o , x ] y\in[o,x] y ∈ [ o , x ] , d X ^ ( o , y ) ≤ d X ^ ( o , x ) d_{\widehat{X}}(o,y)\le d_{\widehat{X}}(o,x) d X ( o , y ) ≤ d X ( o , x ) .
Proof. Since y ∈ [ o , x ] y\in[o,x] y ∈ [ o , x ] , we have [ o , y ] ⊂ [ o , x ] [o,y]\subset[o,x] [ o , y ] ⊂ [ o , x ] . Furthermore, g [ o , y ] ( x ) = y g_{[o,y]}(x)=y g [ o , y ] ( x ) = y . By applying Lemma 5.5 , we have
d X ^ ( o , y ) = d X ^ ( g [ o , y ] ( o ) , g [ o , y ] ( x ) ) ≤ d X ^ ( o , x ) . d_{\widehat{X}}(o,y)=d_{\widehat{X}}(g_{[o,y]}(o),g_{[o,y]}(x))\le d_{\widehat{X}}(o,x). d X ( o , y ) = d X ( g [ o , y ] ( o ) , g [ o , y ] ( x )) ≤ d X ( o , x ) .
The next lemma shows that if two points are close in X ^ \widehat{X} X then the convex hulls from a fixed point o o o to each of them have a large intersection. It will be used in the filter construction to control the diameter of the top boundary path in § 8 .
Lemma 5.7 (Large intersection of hulls). Suppose x , y ∈ S R ( o ) x,y\in S_R(o) x , y ∈ S R ( o ) and d X ^ ( x , y ) = E d_{\widehat{X}}(x,y)=E d X ( x , y ) = E where E < R E<R E < R . Then there is z ∈ [ o , x ] ∩ [ o , y ] z\in[o,x]\cap[o,y] z ∈ [ o , x ] ∩ [ o , y ] with d X ^ ( o , z ) ≥ R − E d_{\widehat{X}}(o,z)\ge R-E d X ( o , z ) ≥ R − E .
Proof. Let z : = g [ o , y ] ( x ) z:=g_{[o,y]}(x) z := g [ o , y ] ( x ) , the closest point projection. Since [ o , y ] ∩ [ o , x ] ≠ ∅ [o,y]\cap[o,x]\ne\varnothing [ o , y ] ∩ [ o , x ] = ∅ , by Lemma 2.16 , z ∈ [ o , x ] ∩ [ o , y ] z\in[o,x]\cap[o,y] z ∈ [ o , x ] ∩ [ o , y ] . Furthermore, by Lemma 5.5 ,
d X ^ ( z , y ) = d X ^ ( g [ o , y ] ( x ) , g [ o , y ] ( y ) ) ≤ d X ^ ( x , y ) = E . d_{\widehat{X}}(z,y)=d_{\widehat{X}}(g_{[o,y]}(x),g_{[o,y]}(y))\le d_{\widehat{X}}(x,y)=E. d X ( z , y ) = d X ( g [ o , y ] ( x ) , g [ o , y ] ( y )) ≤ d X ( x , y ) = E .
Hence, d X ^ ( o , z ) ≥ d X ^ ( o , y ) − d X ^ ( z , y ) ≥ R − E d_{\widehat{X}}(o,z)\ge d_{\widehat{X}}(o,y)-d_{\widehat{X}}(z,y)\ge R-E d X ( o , z ) ≥ d X ( o , y ) − d X ( z , y ) ≥ R − E .
The following lemma shows that the image of a convex hull of two vertices in X X X is uniformly close to a taut geodesic in X ^ \widehat{X} X .
Lemma 5.8 (Hulls are uniformly close to taut geodesics). Let M = max { ∣ K ∣ ∣ K ⊂ Γ is a clique } M=\max\{|K|\mid K\subset\Gamma\text{ is a clique}\} M = max { ∣ K ∣ ∣ K ⊂ Γ is a clique } . Let o , z o,z o , z be two 0 0 0 -cubes in X X X . Then for every x ∈ [ o , z ] x\in[o,z] x ∈ [ o , z ] , there is y y y on a taut geodesic from o o o to z z z with d X ^ ( x , y ) ≤ M + 1 d_{\widehat{X}}(x,y)\le M+1 d X ( x , y ) ≤ M + 1 .
Proof. Let M ′ = dim ( [ o , z ] ) M'=\dim([o,z]) M ′ = dim ([ o , z ]) , the dimension of the convex hull of { o , z } \{o,z\} { o , z } . Note that M ′ ≤ M M'\le M M ′ ≤ M . Let o = y 0 , ⋯ , y n = z o=y_0,\cdots,y_n=z o = y 0 , ⋯ , y n = z be a taut geodesic in X ^ \widehat{X} X from o o o to z z z such that every hyperplane in X X X separates at most one consecutive pair. For two vertices a , b a,b a , b , denote by W ( a , b ) W(a,b) W ( a , b ) the set of hyperplanes separating a a a and b b b . Then for each 0 ≤ i ≤ n 0\le i\le n 0 ≤ i ≤ n , define
A i = W ( o , x ) ∖ W ( o , y i ) , B i = W ( o , y i ) ∖ W ( o , x ) . A_i=W(o,x)\setminus W(o,y_i),\qquad B_i=W(o,y_i)\setminus W(o,x). A i = W ( o , x ) ∖ W ( o , y i ) , B i = W ( o , y i ) ∖ W ( o , x ) .
Claim. For each 0 ≤ i ≤ n 0\le i\le n 0 ≤ i ≤ n , W ( x , y i ) = A i ⊔ B i W(x,y_i)=A_i\sqcup B_i W ( x , y i ) = A i ⊔ B i . Moreover, for every hyperplane V ∈ A i V\in A_i V ∈ A i and H ∈ B i H\in B_i H ∈ B i , V V V crosses H H H .
Proof. Let W W W be a hyperplane separating x x x and y i y_i y i . Since x , y i ∈ [ o , z ] x,y_i\in[o,z] x , y i ∈ [ o , z ] , W W W also separates o o o and z z z . If W ∉ A i W\notin A_i W ∈ / A i , then it must separate { o , x } \{o,x\} { o , x } from { z , y i } \{z,y_i\} { z , y i } . Hence, W ∈ B i W\in B_i W ∈ B i . Also, clearly, A i ∩ B i = ∅ A_i\cap B_i=\varnothing A i ∩ B i = ∅ . Thus, W ( x , y i ) = A i ⊔ B i W(x,y_i)=A_i\sqcup B_i W ( x , y i ) = A i ⊔ B i .
Let V ∈ A i V \in\mathcal{A}_i V ∈ A i and H ∈ B i H \in\mathcal{B}_i H ∈ B i . Suppose H H H has two halfspaces H ← \overleftarrow{H} H and H → \overrightarrow{H} H such that { o , x } ⊂ H ← \{o,x\} \subset\overleftarrow{H} { o , x } ⊂ H and { z , y i } ⊂ H → \{z,y_i\} \subset\overrightarrow{H} { z , y i } ⊂ H . Suppose V V V does not cross H H H . Then V ⊂ H → V \subset\overrightarrow{H} V ⊂ H . Hence, V V V would separate x x x from { o , z } \{o,z\} { o , z } . Contradiction. □
Then for each i i i , Label ( A i ) \operatorname{Label}(\mathcal{A}_i) Label ( A i ) and Label ( B i ) \operatorname{Label}(\mathcal{B}_i) Label ( B i ) span a join in the defining graph Γ \Gamma Γ . Furthermore, by the choice of taut geodesic, A i ⊂ A i − 1 \mathcal{A}_i \subset\mathcal{A}_{i-1} A i ⊂ A i − 1 and B i − 1 ⊂ B i \mathcal{B}_{i-1} \subset\mathcal{B}_i B i − 1 ⊂ B i for i = 1 , … , n i=1,\ldots,n i = 1 , … , n .
Let k ∈ { 0 , 1 , … , n } k \in\{0,1,\ldots,n\} k ∈ { 0 , 1 , … , n } be the minimal index such that Label ( A k ) \operatorname{Label}(\mathcal{A}_k) Label ( A k ) is a clique. Such an index k k k exists. Indeed, since x ∈ [ o , z ] x \in[o,z] x ∈ [ o , z ] , we have W ( o , x ) ⊆ W ( o , z ) \mathcal{W}(o,x) \subseteq\mathcal{W}(o,z) W ( o , x ) ⊆ W ( o , z ) , and A n = ∅ \mathcal{A}_n=\varnothing A n = ∅ . We consider the empty set as a clique.
Suppose k = 0 k=0 k = 0 . Then y 0 = o y_0=o y 0 = o and x x x are contained in a cube. Then, d X ^ ( x , o ) ≤ d ( x , o ) ≤ dim ( [ o , z ] ) = M ′ d_{\widehat{X}}(x,o) \le d(x,o) \le\dim([o,z])=M' d X ( x , o ) ≤ d ( x , o ) ≤ dim ([ o , z ]) = M ′ .
Suppose k ≥ 1 k \ge1 k ≥ 1 . Then Label ( A k ) \operatorname{Label}(\mathcal{A}_k) Label ( A k ) is a clique and Label ( A k − 1 ) \operatorname{Label}(\mathcal{A}_{k-1}) Label ( A k − 1 ) is not a clique.
Case 1: Suppose Label ( B k − 1 ) \operatorname{Label}(\mathcal{B}_{k-1}) Label ( B k − 1 ) is not a clique. Then Label ( A k − 1 ) ∗ Label ( B k − 1 ) \operatorname{Label}(\mathcal{A}_{k-1}) * \operatorname{Label}(\mathcal{B}_{k-1}) Label ( A k − 1 ) ∗ Label ( B k − 1 ) is a nontrivial join. Hence, d X ^ ( x , y k − 1 ) ≤ 1 d_{\widehat{X}}(x,y_{k-1}) \le1 d X ( x , y k − 1 ) ≤ 1 .
Case 2: Suppose Label ( B k − 1 ) \operatorname{Label}(\mathcal{B}_{k-1}) Label ( B k − 1 ) is a clique. Let x ′ = g [ y k − 1 , y k ] ( x ) x' = \mathfrak{g}_{[y_{k-1},y_k]}(x) x ′ = g [ y k − 1 , y k ] ( x ) . See Figure 4 .
Figure 4. Vertices x x x , y k y_k y k , y k − 1 y_{k-1} y k − 1 in the convex hull of o o o and z z z .
Claim. W ( x , x ′ ) = A k ⊔ B k − 1 \mathcal{W}(x,x')=\mathcal{A}_k \sqcup\mathcal{B}_{k-1} W ( x , x ′ ) = A k ⊔ B k − 1 . Furthermore, every hyperplane in A k \mathcal{A}_k A k crosses every hyperplane in B k − 1 \mathcal{B}_{k-1} B k − 1 .
Proof. By the gate-map property, a hyperplane is in W ( x , x ′ ) \mathcal{W}(x,x') W ( x , x ′ ) if and only if it separates x x x from { y k − 1 , y k } \{y_{k-1},y_k\} { y k − 1 , y k } .
A k ∪ B k − 1 ⊂ W ( x , x ′ ) \mathcal{A}_k \cup\mathcal{B}_{k-1} \subset\mathcal{W}(x,x') A k ∪ B k − 1 ⊂ W ( x , x ′ ) : Let V ∈ A k V \in\mathcal{A}_k V ∈ A k . Since A k ⊆ A k − 1 \mathcal{A}_k \subseteq\mathcal{A}_{k-1} A k ⊆ A k − 1 , V V V separates o o o from x x x , but separates neither y k − 1 y_{k-1} y k − 1 nor y k y_k y k from o o o . Thus V V V separates x x x from { y k − 1 , y k } \{y_{k-1},y_k\} { y k − 1 , y k } , and hence V ∈ W ( x , x ′ ) V \in\mathcal{W}(x,x') V ∈ W ( x , x ′ ) .
W ( x , x ′ ) ⊂ A k ∪ B k − 1 \mathcal{W}(x,x') \subset\mathcal{A}_k \cup\mathcal{B}_{k-1} W ( x , x ′ ) ⊂ A k ∪ B k − 1 : Let W ∈ W ( x , x ′ ) W \in\mathcal{W}(x,x') W ∈ W ( x , x ′ ) . Suppose W W W separates x x x and o o o . Then it does not separate o o o and y k y_k y k ; otherwise W ∉ W ( x , x ′ ) W \notin\mathcal{W}(x,x') W ∈ / W ( x , x ′ ) . Hence, W ∈ A k W \in\mathcal{A}_k W ∈ A k . Now suppose W W W does not separate x x x and o o o . Then it must separate { o , x } \{o,x\} { o , x } from { y k , y k − 1 } \{y_k,y_{k-1}\} { y k , y k − 1 } . Hence, W ∈ B k − 1 W \in\mathcal{B}_{k-1} W ∈ B k − 1 .
Therefore, W ( x , x ′ ) = A k ∪ B k − 1 \mathcal{W}(x,x')=\mathcal{A}_k \cup\mathcal{B}_{k-1} W ( x , x ′ ) = A k ∪ B k − 1 . Since B k − 1 ⊂ B k \mathcal{B}_{k-1} \subset\mathcal{B}_k B k − 1 ⊂ B k and B k ∩ A k = ∅ \mathcal{B}_k \cap\mathcal{A}_k=\varnothing B k ∩ A k = ∅ , B k − 1 ∩ A k = ∅ \mathcal{B}_{k-1} \cap\mathcal{A}_k=\varnothing B k − 1 ∩ A k = ∅ . Hence, W ( x , x ′ ) = A k ⊔ B k − 1 \mathcal{W}(x,x')=\mathcal{A}_k \sqcup\mathcal{B}_{k-1} W ( x , x ′ ) = A k ⊔ B k − 1 . By the previous claim, every hyperplane in B k \mathcal{B}_k B k crosses every hyperplane in A k \mathcal{A}_k A k . Since B k − 1 ⊂ B k \mathcal{B}_{k-1} \subset\mathcal{B}_k B k − 1 ⊂ B k , the furthermore part follows. □
Therefore, Label ( [ x , x ′ ] ) = Label ( A k ) ∗ Label ( B k − 1 ) \operatorname{Label}([x,x'])=\operatorname{Label}(\mathcal{A}_k)*\operatorname{Label}(\mathcal{B}_{k-1}) Label ([ x , x ′ ]) = Label ( A k ) ∗ Label ( B k − 1 ) is a clique. d X ^ ( x , x ′ ) ≤ d ( x , x ′ ) ≤ M ′ d_{\widehat{X}}(x,x') \le d(x,x') \le M' d X ( x , x ′ ) ≤ d ( x , x ′ ) ≤ M ′ .
Since y k − 1 y_{k-1} y k − 1 and y k y_k y k are adjacent in X ^ \widehat{X} X , they are either connected by a 1 1 1 -cube or contained in the same product region in X X X . Since x ′ ∈ [ y k − 1 , y k ] x' \in[y_{k-1},y_k] x ′ ∈ [ y k − 1 , y k ] , in either case d X ^ ( x ′ , y k ) ≤ 1 d_{\widehat X}(x',y_k) \le1 d X ( x ′ , y k ) ≤ 1 . Therefore,
d X ^ ( x , y k ) ≤ d X ^ ( x , x ′ ) + d X ^ ( x ′ , y k ) ≤ M ′ + 1. d_{\widehat X}(x,y_k) \le d_{\widehat X}(x,x') + d_{\widehat X}(x',y_k) \le M' + 1. d X ( x , y k ) ≤ d X ( x , x ′ ) + d X ( x ′ , y k ) ≤ M ′ + 1.
Wide combinatorial paths in X X X .
Definition 5.9 (Wide path). We say a combinatorial path α \alpha α in X Γ X_\Gamma X Γ is wide if α \alpha α is contained in a standard product region. Equivalently, α \alpha α is wide if and only if Label ( α ) \operatorname{Label}(\alpha) Label ( α ) is contained in a nontrivial join.
The next two lemmas relate wide subpaths in X X X to distance in X ^ \widehat X X . Lemma 5.10 transfers a bound on wide subpath lengths from one combinatorial geodesic to every combinatorial geodesic with the same endpoints. Lemma 5.11 then converts this bound into a lower bound on distance in X ^ \widehat X X . Together, these estimates will control progress in the filter construction in § 8 .
Lemma 5.10 (Longest wide subpath). Let α \alpha α be a combinatorial geodesic from x x x to y y y such that every wide subpath has length at most C C C for some constant C ≥ 1 C \ge1 C ≥ 1 . Then there is a constant C ′ = C ′ ( C , Γ ) C' = C'(C,\Gamma) C ′ = C ′ ( C , Γ ) such that every wide subpath of any combinatorial geodesic from x x x to y y y has length at most C ′ C' C ′ .
Proof. Let M M M be the maximal cardinality of a clique in Γ \Gamma Γ . We will prove that one may take
C ′ = ( 2 ∣ V ( Γ ) ∣ + 1 ) max { C , M } . C'=(2|V(\Gamma)|+1)\max\{C,M\}. C ′ = ( 2∣ V ( Γ ) ∣ + 1 ) max { C , M } .
Let γ \gamma γ be a combinatorial geodesic from x x x to y y y , and let γ ′ \gamma' γ ′ be a wide subpath of γ \gamma γ . We assume that γ ′ \gamma' γ ′ has positive length. There is a nontrivial join Λ ⊆ Γ \Lambda\subseteq\Gamma Λ ⊆ Γ containing Label ( γ ′ ) \operatorname{Label}(\gamma') Label ( γ ′ ) .
Both α \alpha α and γ \gamma γ cross precisely the hyperplanes separating x x x from y y y , each exactly once. Thus the hyperplanes crossed by γ ′ \gamma' γ ′ provide a collection of 1-cubes of α \alpha α . Let α 1 , … , α n \alpha_1,\ldots,\alpha_n α 1 , … , α n be the maximal nonempty subpaths of α \alpha α formed by concatenating these 1-cubes. Without loss of generality, we assume α i \alpha_i α i appears before α i + 1 \alpha_{i+1} α i + 1 when traveling from x x x to y y y along α \alpha α . For 1 ≤ t < n 1 \le t < n 1 ≤ t < n , let β t \beta_t β t be the nonempty subpath between α t \alpha_t α t and α t + 1 \alpha_{t+1} α t + 1 . Denote
A i = Label ( α i ) , B t = Label ( β t ) . A_i=\operatorname{Label}(\alpha_i), \qquad B_t=\operatorname{Label}(\beta_t). A i = Label ( α i ) , B t = Label ( β t ) .
Clearly, for each i = 1 , … , n i=1,\ldots,n i = 1 , … , n , A i ⊂ Λ A_i \subset\Lambda A i ⊂ Λ and α i \alpha_i α i is a wide subpath. Moreover, by the construction, we have
∣ γ ′ ∣ = ∑ i = 1 n ∣ α i ∣ . |\gamma'|=\sum_{i=1}^{n}|\alpha_i|. ∣ γ ′ ∣ = i = 1 ∑ n ∣ α i ∣.
Let B \mathcal{B} B be the set of hyperplanes crossed by γ \gamma γ but not crossed by γ ′ \gamma' γ ′ . Since γ ′ \gamma' γ ′ is a subpath of γ \gamma γ , each hyperplane in B \mathcal{B} B separates γ ′ \gamma' γ ′ from either x x x or y y y . We define
B x = { W ∈ B ∣ W separates x from γ ′ } , B y = { W ∈ B ∣ W separates y from γ ′ } . \mathcal{B}^{x}=\{W\in\mathcal{B}\mid W\text{ separates }x\text{ from }\gamma'\},\qquad
\mathcal{B}^{y}=\{W\in\mathcal{B}\mid W\text{ separates }y\text{ from }\gamma'\}. B x = { W ∈ B ∣ W separates x from γ ′ } , B y = { W ∈ B ∣ W separates y from γ ′ } .
Hence, B = B x ⊔ B y \mathcal{B}=\mathcal{B}^{x}\sqcup\mathcal{B}^{y} B = B x ⊔ B y . For 1 ≤ t < n 1\le t<n 1 ≤ t < n , let
B t x = Label { W ∈ B x : W crosses β t } , B t y = Label { W ∈ B y : W crosses β t } . B_t^{x}=\operatorname{Label}\{W\in\mathcal{B}^{x}:W\text{ crosses }\beta_t\},\qquad
B_t^{y}=\operatorname{Label}\{W\in\mathcal{B}^{y}:W\text{ crosses }\beta_t\}. B t x = Label { W ∈ B x : W crosses β t } , B t y = Label { W ∈ B y : W crosses β t } .
Thus B t = B t x ∪ B t y B_t=B_t^{x}\cup B_t^{y} B t = B t x ∪ B t y . See Figure 5 .
Figure 5. Label(W W W ) ∈ B t x \in B^{x}_{t} ∈ B t x and Label(V V V ) ∈ B t y \in B^{y}_{t} ∈ B t y .
Suppose that W ∈ B t x W\in B_t^{x} W ∈ B t x crosses β t \beta_t β t . For every i ≤ t i\le t i ≤ t , the subpath α i \alpha_i α i and γ ′ \gamma' γ ′ are separated by W W W . Since by construction every hyperplane crossing α i \alpha_i α i also crosses γ ′ \gamma' γ ′ , it must cross W W W as well. Hence
A i ⊆ lk ( B t x ) (for i ≤ t ) . A_i\subseteq\operatorname{lk}(B_t^{x})\qquad\text{(for }i\le t\text{)}. A i ⊆ lk ( B t x ) (for i ≤ t ) .
Similarly,
A i ⊆ lk ( B t y ) ( i > t ) . A_i \subseteq\operatorname{lk}(B^{y}_{t}) \qquad(i > t). A i ⊆ lk ( B t y ) ( i > t ) .
For 1 ≤ i ≤ n 1 \leq i \leq n 1 ≤ i ≤ n , let
X i = ⋃ t = i n − 1 B t x , Y i = ⋃ t = 1 i − 1 B t y . X_i=\bigcup_{t=i}^{n-1}B^{x}_{t}, \qquad Y_i=\bigcup_{t=1}^{i-1}B^{y}_{t}. X i = t = i ⋃ n − 1 B t x , Y i = t = 1 ⋃ i − 1 B t y .
The preceding observation gives
A i ⊆ lk ( X i ) ∩ lk ( Y i ) . A_i \subseteq\operatorname{lk}(X_i)\cap\operatorname{lk}(Y_i). A i ⊆ lk ( X i ) ∩ lk ( Y i ) .
Furthermore, by construction we have
X 1 ⊇ X 2 ⊇ ⋯ ⊇ X n , Y 1 ⊆ Y 2 ⊆ ⋯ ⊆ Y n . X_1 \supseteq X_2 \supseteq\cdots\supseteq X_n,\qquad Y_1 \subseteq Y_2 \subseteq\cdots\subseteq Y_n. X 1 ⊇ X 2 ⊇ ⋯ ⊇ X n , Y 1 ⊆ Y 2 ⊆ ⋯ ⊆ Y n .
Partition { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } into maximal consecutive intervals I 1 , … , I m I_1,\ldots,I_m I 1 , … , I m on which the pair ( X i , Y i ) (X_i,Y_i) ( X i , Y i ) is constant. Each change of this pair removes at least one vertex from X i X_i X i or adds at least one vertex to Y i Y_i Y i . Since each vertex can be removed once and added once, there are at most 2 ∣ V ( Γ ) ∣ 2|V(\Gamma)| 2∣ V ( Γ ) ∣ changes. Thus
m ≤ 2 ∣ V ( Γ ) ∣ + 1. m \leq2|V(\Gamma)|+1. m ≤ 2∣ V ( Γ ) ∣ + 1.
Fix an interval I ℓ = { p , … , q } I_\ell=\{p,\ldots,q\} I ℓ = { p , … , q } , and let
K = ⋃ j = p q A j , T = ⋃ t = p q − 1 B t . K=\bigcup_{j=p}^{q}A_j,\qquad T=\bigcup_{t=p}^{q-1}B_t. K = j = p ⋃ q A j , T = t = p ⋃ q − 1 B t .
Let τ ℓ \tau_\ell τ ℓ be the subpath of α \alpha α beginning with α p \alpha_p α p and ending with α q \alpha_q α q , i.e.
τ ℓ = α p β p α p + 1 ⋯ β q − 1 α q . \tau_\ell=\alpha_p\beta_p\alpha_{p+1}\cdots\beta_{q-1}\alpha_q. τ ℓ = α p β p α p + 1 ⋯ β q − 1 α q .
When p = q p=q p = q , we just have τ ℓ = α p \tau_\ell=\alpha_p τ ℓ = α p . In particular,
Label ( τ ℓ ) = K ∪ T . \operatorname{Label}(\tau_\ell)=K\cup T. Label ( τ ℓ ) = K ∪ T .
We claim that every element of T T T commutes with every element of K K K . Indeed, for any b ∈ T b\in T b ∈ T , there is p ≤ t ≤ q − 1 p\leq t\leq q-1 p ≤ t ≤ q − 1 such that either b ∈ B t x b\in B^{x}_{t} b ∈ B t x or b ∈ B t y b\in B^{y}_{t} b ∈ B t y . If b ∈ B t x b\in B^{x}_{t} b ∈ B t x , since X t = X q X_t=X_q X t = X q by constancy on I ℓ I_\ell I ℓ , we have K ⊆ lk ( b ) K\subseteq\operatorname{lk}(b) K ⊆ lk ( b ) . Similarly, if b ∈ B t y b\in B^{y}_{t} b ∈ B t y , since Y t + 1 = Y p Y_{t+1}=Y_p Y t + 1 = Y p by constancy on I ℓ I_\ell I ℓ , we have K ⊆ lk ( b ) K\subseteq\operatorname{lk}(b) K ⊆ lk ( b ) . Thus, K ∪ T K\cup T K ∪ T spans a join in Γ \Gamma Γ .
Now we claim that
∑ j = p q ∣ α j ∣ ≤ max { C , M } . \sum_{j=p}^{q} |\alpha_j| \le\max\{C,M\}. j = p ∑ q ∣ α j ∣ ≤ max { C , M } .
Case 1: K K K is not a clique. Let
Δ = lk ( T ) ∩ Λ . \Delta= \operatorname{lk}(T) \cap\Lambda. Δ = lk ( T ) ∩ Λ.
Since K ⊆ Δ K \subseteq\Delta K ⊆ Δ , the subgraph Δ \Delta Δ is not a clique. Lemma 2.8 implies that T ∪ Δ T \cup\Delta T ∪ Δ is contained in a nontrivial join. Hence,
Label ( τ ℓ ) = K ∪ T ⊆ Δ ∪ T \operatorname{Label}(\tau_\ell) = K \cup T \subseteq\Delta\cup T Label ( τ ℓ ) = K ∪ T ⊆ Δ ∪ T
is contained in a nontrivial join. Thus τ ℓ \tau_\ell τ ℓ is a wide subpath of α \alpha α , and
∑ j = p q ∣ α j ∣ ≤ ∣ τ ℓ ∣ ≤ C . \sum_{j=p}^{q} |\alpha_j| \le|\tau_\ell| \le C. j = p ∑ q ∣ α j ∣ ≤ ∣ τ ℓ ∣ ≤ C .
Case 2: K K K is a clique. Since every element of T T T commutes with every element of K K K , there is a cube spanned by (possibly translates of) α p , … , α q \alpha_p,\ldots,\alpha_q α p , … , α q in the convex hull [ x , y ] [x,y] [ x , y ] . Hence,
∑ j = p q ∣ α j ∣ ≤ ∣ K ∣ ≤ M . \sum_{j=p}^{q} |\alpha_j| \le|K| \le M. j = p ∑ q ∣ α j ∣ ≤ ∣ K ∣ ≤ M .
Therefore,
∣ γ ′ ∣ = ∑ ℓ = 1 m ∑ j ∈ I ℓ ∣ α j ∣ ≤ m max { C , M } ≤ ( 2 ∣ V ( Γ ) ∣ + 1 ) max { C , M } . |\gamma'| = \sum_{\ell=1}^{m}\sum_{j\in I_\ell} |\alpha_j| \le m\max\{C,M\} \le(2|V(\Gamma)|+1)\max\{C,M\}. ∣ γ ′ ∣ = ℓ = 1 ∑ m j ∈ I ℓ ∑ ∣ α j ∣ ≤ m max { C , M } ≤ ( 2∣ V ( Γ ) ∣ + 1 ) max { C , M } .
Lemma 5.11 (Bounded wide subpaths in X X X force progress in X ^ \widehat{X} X ). Let x , y x,y x , y be two 0 0 0 -cubes in X Γ X_\Gamma X Γ with the property that every wide subpath of any combinatorial geodesic from x x x to y y y has length at most C ′ C' C ′ for some constant C ′ ≥ 1 C' \ge1 C ′ ≥ 1 . For each n ∈ N n \in\mathbb{N} n ∈ N , if d ( x , y ) ≥ n ( C ′ + 1 ) d(x,y) \ge n(C'+1) d ( x , y ) ≥ n ( C ′ + 1 ) , then d X ^ ( x , y ) ≥ n d_{\widehat{X}}(x,y) \ge n d X ( x , y ) ≥ n .
Proof. We argue by induction on n n n . Suppose d ( x , y ) ≥ C ′ + 1 d(x,y) \ge C'+1 d ( x , y ) ≥ C ′ + 1 . Then x ≠ y x \ne y x = y . Hence, d X ^ ( x , y ) ≥ 1 d_{\widehat{X}}(x,y) \ge1 d X ( x , y ) ≥ 1 .
Assume the property holds for n − 1 n-1 n − 1 . Now we assume that d ( x , y ) ≥ n ( C ′ + 1 ) d(x,y) \ge n(C'+1) d ( x , y ) ≥ n ( C ′ + 1 ) . Let z ∈ [ x , y ] z \in[x,y] z ∈ [ x , y ] with d X ^ ( y , z ) ≤ 1 d_{\widehat{X}}(y,z) \le1 d X ( y , z ) ≤ 1 . Then d ( y , z ) ≤ C ′ d(y,z) \le C' d ( y , z ) ≤ C ′ . Hence d ( x , z ) ≥ ( n − 1 ) ( C ′ + 1 ) + 1 d(x,z) \ge(n-1)(C'+1)+1 d ( x , z ) ≥ ( n − 1 ) ( C ′ + 1 ) + 1 . Since z ∈ [ x , y ] z \in[x,y] z ∈ [ x , y ] , every combinatorial geodesic from x x x to z z z extends to a combinatorial geodesic from x x x to y y y . Consequently, each of its wide subpaths has length at most C ′ C' C ′ . Then by the induction hypothesis, d X ^ ( x , z ) ≥ n − 1 d_{\widehat{X}}(x,z) \ge n-1 d X ( x , z ) ≥ n − 1 . By the arbitrariness of z z z and Remark 5.4, d X ^ ( x , y ) ≥ n d_{\widehat{X}}(x,y) \ge n d X ( x , y ) ≥ n . □
Dead ends in the maximized hyperbolic space
Birman and Menasco showed that the curve graph of a closed orientable surface of genus at least two contains dead ends [16 ] . In particular, they constructed dead ends at arbitrarily large distance from a fixed vertex. Since the curve graph is a basic model for hyperbolic spaces arising in hierarchical structures, it is natural to ask whether maximized hyperbolic spaces of right-angled Coxeter groups exhibit a similar phenomenon.
A dead end results in the failure of exact geodesic extension. A geodesic segment ending at a dead end cannot be extended further as a geodesic from a basepoint. Thus, exact geodesic extension cannot be assumed in X ^ \widehat{X} X . The examples below illustrate this obstruction and motivate the coarse extension results proved in the following section.
We first recall the definition of dead ends and then give two examples. The first produces dead ends on the sphere with radius one and pairs of vertices joined by infinitely many geodesics. The second produces dead ends arbitrarily far from a fixed basepoint. Although this resembles the phenomenon in the curve graph, the underlying technique is different: in the curve graph, dead ends arise from separating curves, whereas in our examples they arise from the arrangement of maximal nontrivial joins in the defining graph.
Definition 6.1 (Dead ends). Let G \mathcal{G} G be a simplicial graph equipped with its path metric d d d , and fix a vertex o ∈ G o \in\mathcal{G} o ∈ G . A vertex x ∈ G x \in\mathcal{G} x ∈ G is a dead end with respect to o o o if
d ( o , z ) ≤ d ( o , x ) d(o,z) \le d(o,x) d ( o , z ) ≤ d ( o , x ) for every vertex z z z adjacent to x x x .
The following example produces dead ends on the sphere with radius one about any chosen basepoint. It also gives pairs of vertices joined by infinitely many geodesics in X ^ \widehat{X} X .
Example 6.2. Let X = X Γ 1 X = X_{\Gamma_1} X = X Γ 1 , where Γ 1 \Gamma_1 Γ 1 is the left graph in Figure 6 . It has two maximal nontrivial joins:
Λ 1 = { a , c } ∗ { b , d , v } , Λ 2 = { b , d } ∗ { a , c , w } . \Lambda_1 = \{a,c\} * \{b,d,v\}, \qquad\Lambda_2 = \{b,d\} * \{a,c,w\}. Λ 1 = { a , c } ∗ { b , d , v } , Λ 2 = { b , d } ∗ { a , c , w } . Their intersection Δ = { a , c } ∗ { b , d } \Delta= \{a,c\} * \{b,d\} Δ = { a , c } ∗ { b , d } is itself a nontrivial join. Because v v v and w w w do not lie in a common nontrivial join, X ^ \widehat{X} X has infinite diameter.
Dead ends on the sphere with radius 1. Fix o ∈ X ^ o \in\widehat{X} o ∈ X and take x ≠ o ∈ X ^ x \ne o \in\widehat{X} x = o ∈ X such that Label ( [ o , x ] ) ⊂ Δ \operatorname{Label}([o,x]) \subset\Delta Label ([ o , x ]) ⊂ Δ . Then d X ^ ( o , x ) = 1 d_{\widehat{X}}(o,x) = 1 d X ( o , x ) = 1 . Let z ∈ X ^ z \in\widehat{X} z ∈ X be adjacent to x x x . Then z z z and x x x are either connected by a 1-cube or contained in the same standard product region in X X X .
Case 1: Suppose z z z and x x x are adjacent by a 1-cube labeled by u ∈ V ( Γ 1 ) u \in V(\Gamma_1) u ∈ V ( Γ 1 ) . Since for any u ∈ V ( Γ 1 ) u \in V(\Gamma_1) u ∈ V ( Γ 1 ) , u u u and Δ \Delta Δ are contained in a nontrivial join, Label ( [ o , z ] ) ⊂ Δ ∪ { u } \operatorname{Label}([o,z]) \subset\Delta\cup\{u\} Label ([ o , z ]) ⊂ Δ ∪ { u } is contained in a nontrivial join. Hence d X ^ ( o , z ) ≤ 1 d_{\widehat{X}}(o,z) \le1 d X ( o , z ) ≤ 1 .
Case 2: Otherwise, suppose d ( x , z ) ≥ 2 d(x,z) \ge2 d ( x , z ) ≥ 2 and Label ( [ x , z ] ) \operatorname{Label}([x,z]) Label ([ x , z ]) is contained in a nontrivial join. Then Label ( [ x , z ] ) ⊆ Λ i \operatorname{Label}([x,z]) \subseteq\Lambda_i Label ([ x , z ]) ⊆ Λ i for i = 1 i = 1 i = 1 or $2$. Since Δ = Label ( [ o , x ] ) ⊆ Λ i \Delta= \operatorname{Label}([o,x]) \subseteq\Lambda_i Δ = Label ([ o , x ]) ⊆ Λ i for $i = 1, 2 as well, Label ( [ o , z ] ) ⊆ Λ i \operatorname{Label}([o,z]) \subseteq\Lambda_i Label ([ o , z ]) ⊆ Λ i . Then d X ^ ( o , z ) ≤ 1 d_{\widehat{X}}(o,z) \le1 d X ( o , z ) ≤ 1 .
Figure 6. Defining graphs Γ 1 \Gamma_1 Γ 1 (left) and Γ 2 \Gamma_2 Γ 2 (right) of RACGs whose maximized hyperbolic spaces contain dead ends.
Therefore, x x x is a dead end with respect to o o o .
Infinitely many geodesics between a pair of points in X ^ Γ 1 \widehat{X}_{\Gamma_1} X Γ 1 . We use this example to show the existence of infinitely many geodesics between a pair of points in X ^ Γ 1 \widehat{X}_{\Gamma_1} X Γ 1 as well. Let p , q p,q p , q be two 0-cubes in X Γ 1 X_{\Gamma_1} X Γ 1 which are separated exactly by two hyperplanes labeled by v v v and w w w . Since v , w v,w v , w are not in a nontrivial join in Γ 1 \Gamma_1 Γ 1 , d X ^ ( p , q ) = 2 d_{\widehat{X}}(p,q)=2 d X ( p , q ) = 2 . Let z z z be the 0-cube which is adjacent to both p p p and q q q in X Γ 1 X_{\Gamma_1} X Γ 1 . Since the convex hull [ p , q ] ⊂ X Γ 1 [p,q]\subset X_{\Gamma_1} [ p , q ] ⊂ X Γ 1 is the combinatorial geodesic from p p p to q q q , the path p , z , q p,z,q p , z , q in X ^ Γ 1 \widehat{X}_{\Gamma_1} X Γ 1 is the unique taut geodesic from p p p to q q q . For every vertex z ′ ≠ z z'\ne z z ′ = z contained in the copy of the flat X Δ X_{\Delta} X Δ passing through z z z , the path p , z ′ , q p,z',q p , z ′ , q in X ^ Γ 1 \widehat{X}_{\Gamma_1} X Γ 1 is also a geodesic but not taut.
The next example contains dead ends arbitrarily far from the basepoint.
Example 6.3. Let X Γ 2 X_{\Gamma_2} X Γ 2 be the Davis complex of the RACG W Γ 2 W_{\Gamma_2} W Γ 2 whose defining graph appears in Figure 6 . The graph Γ 2 \Gamma_2 Γ 2 has four maximal nontrivial joins:
Λ 1 = { v 1 , v 2 } ∗ { b , d , v } , Λ 2 = { w 1 , w 2 } ∗ { a , c , w } , \Lambda_1=\{v_1,v_2\}*\{b,d,v\},\qquad\Lambda_2=\{w_1,w_2\}*\{a,c,w\}, Λ 1 = { v 1 , v 2 } ∗ { b , d , v } , Λ 2 = { w 1 , w 2 } ∗ { a , c , w } ,
Δ 1 = { a , c } ∗ { b , d , w 1 , w 2 } , Δ 2 = { b , d } ∗ { a , c , v 1 , v 2 } . \Delta_1=\{a,c\}*\{b,d,w_1,w_2\},\qquad\Delta_2=\{b,d\}*\{a,c,v_1,v_2\}. Δ 1 = { a , c } ∗ { b , d , w 1 , w 2 } , Δ 2 = { b , d } ∗ { a , c , v 1 , v 2 } .
Fix o ∈ X ^ Γ 2 o\in\widehat{X}_{\Gamma_2} o ∈ X Γ 2 . Let γ \gamma γ be the combinatorial geodesic in X Γ 2 X_{\Gamma_2} X Γ 2 starting at o o o labeled by the word
γ = v w a c b d . \gamma=vwacbd. γ = v w a c b d .
Let x x x be the terminal point of γ \gamma γ . Applying Proposition 5.2 , we know that d X ^ ( o , x ) = 3 d_{\widehat{X}}(o,x)=3 d X ( o , x ) = 3 by enumerating paths in X ^ Γ 2 \widehat{X}_{\Gamma_2} X Γ 2 with vertices in [ o , x ] [o,x] [ o , x ] . There are two points y 1 , y 2 ∈ [ o , x ] y_1,y_2\in[o,x] y 1 , y 2 ∈ [ o , x ] such that y 1 y_1 y 1 and x x x are separated by two hyperplanes of types b b b and d d d , and y 2 y_2 y 2 and x x x are separated by two hyperplanes of types a a a and c c c . Note that y 1 , y 2 ∈ S 2 ( o ) y_1,y_2\in S_2(o) y 1 , y 2 ∈ S 2 ( o ) .
We claim that x x x is a dead end with respect to o o o . Let z z z be a vertex in X ^ Γ 2 \widehat{X}_{\Gamma_2} X Γ 2 such that d X ^ ( x , z ) = 1 d_{\widehat{X}}(x,z)=1 d X ( x , z ) = 1 and z ∉ [ o , x ] z\notin[o,x] z ∈ / [ o , x ] . Let Π = Label ( [ x , z ] ) \Pi=\operatorname{Label}([x,z]) Π = Label ([ x , z ]) . The set Π \Pi Π is either a singleton or is contained in a nontrivial join; hence it lies in one of Λ 1 , Λ 2 , Δ 1 , Δ 2 \Lambda_1,\Lambda_2,\Delta_1,\Delta_2 Λ 1 , Λ 2 , Δ 1 , Δ 2 . Consequently, Π \Pi Π lies in a common nontrivial join with either { b , d } \{b,d\} { b , d } or { a , c } \{a,c\} { a , c } . If Π \Pi Π lies in a common nontrivial join with { b , d } \{b,d\} { b , d } , then Label ( [ z , y 1 ] ) ⊆ Π ∪ { b , d } \operatorname{Label}([z,y_1])\subseteq\Pi\cup\{b,d\} Label ([ z , y 1 ]) ⊆ Π ∪ { b , d } is contained in a nontrivial join. Hence d X ^ ( z , y 1 ) ≤ 1 d_{\widehat{X}}(z,y_1)\le1 d X ( z , y 1 ) ≤ 1 . Similarly, if Π \Pi Π lies in a common nontrivial join with { a , c } \{a,c\} { a , c } , then d X ^ ( z , y 2 ) ≤ 1 d_{\widehat{X}}(z,y_2)\le1 d X ( z , y 2 ) ≤ 1 . Therefore,
d X ^ ( o , z ) ≤ 3 = d X ^ ( o , x ) . d_{\widehat{X}}(o,z)\le3=d_{\widehat{X}}(o,x). d X ( o , z ) ≤ 3 = d X ( o , x ) .
Neighbors of x x x in [ o , x ] [o,x] [ o , x ] are not farther from o o o , so x x x is a dead end with respect to o o o .
Furthermore, for each n ∈ N n\in\mathbb{N} n ∈ N , let x n x_n x n be the terminal point of the geodesic γ n \gamma_n γ n labeled by the word ( v w a c b d ) n (vwacbd)^n ( v w a c b d ) n with initial point o o o . By this construction, d X ^ ( o , x n ) = 2 n + 1 d_{\widehat{X}}(o,x_n)=2n+1 d X ( o , x n ) = 2 n + 1 . (The existence of taut geodesics enables us to focus on combinatorial geodesics in the convex hull when computing the distance in X ^ Γ 2 \widehat{X}_{\Gamma_2} X Γ 2 .) By the same analysis, x n x_n x n is a dead end with respect to o o o for every n ∈ N n\in\mathbb{N} n ∈ N .
These examples highlight a distinction from the curve graph C S \mathcal{C}S C S of a finite-type surface. Following [Wri24, Definition 4.2], a simple closed curve is essentially non-separating if it is non-separating, or bounds a pair of pants. The bounded geodesic image theorem [57 Theorem 3.1] implies that such a curve is never a dead end, regardless of the base vertex (see also [Wri24, Corollary 4.3]). We do not have an analogous characterization for X ^ \widehat{X} X . Proposition 7.1 only forces a geodesic in X ^ \widehat{X} X to pass uniformly close to a fixed point, not through that point.
Bounded geodesic image and applications
The examples in the preceding section show that exact geodesic extension may fail in X ^ \widehat{X} X . Moreover, since X ^ \widehat{X} X need not be proper, the usual Arzelà–Ascoli argument cannot be used to extract limiting geodesic rays. In this section, the bounded geodesic image theorem and the HHS distance formula provide a substitute for the missing compactness.
We first give a proof of the bounded geodesic image statement for X ^ \widehat{X} X (Proposition 7.1 ). Together with the distance formula and the local finiteness of the Davis complex X X X , it allows us to find several properties of ∂ X ^ \partial\widehat{X} ∂ X which are familiar from boundaries of hyperbolic groups. We use them to prove that every boundary point in ∂ X ^ \partial\widehat{X} ∂ X can be represented by a geodesic ray in X ^ \widehat{X} X (Lemma 7.4 ) and that every pair of distinct boundary points can be joined by a bi-infinite geodesic (Proposition 7.5 ). These visibility results imply that geodesic ideal triangles in X ^ \widehat{X} X are uniformly slim and that X ^ \widehat{X} X is geodesic almost extendable (Proposition 7.7). Finally, when ∂ X ^ Γ \partial\widehat{X}_{\Gamma} ∂ X Γ is nonempty, we use bounded geodesic image once more to show that ∂ X ^ Γ \partial\widehat{X}_{\Gamma} ∂ X Γ is noncompact exactly when Γ \Gamma Γ contains an induced 4-cycle (Theorem 7.8 ).
Bounded geodesic image
The bounded geodesic image property for maximized hyperbolic spaces of HHSs with unbounded product regions follows from [1 Theorem 3.7] . For completeness, we give a direct proof in the case of right-angled Coxeter groups.
Recall that g K : X → K \mathfrak{g}_K:X\to K g K : X → K denotes the gate map to a convex subcomplex K ⊂ X K\subset X K ⊂ X . Let U ∈ S U\in\mathcal{S} U ∈ S be a domain represented by a convex subcomplex K ⊂ X K\subset X K ⊂ X . If
d U ( x , y ) > dim X , d_U(x,y)>\dim X, d U ( x , y ) > dim X ,
then g K ( x ) \mathfrak{g}_K(x) g K ( x ) and g K ( y ) \mathfrak{g}_K(y) g K ( y ) cannot lie in a common cube (see Lemma 2.16 ). Thus, sufficiently large projection distance under the gate map gives the hypothesis in the following proposition.
Proposition 7.1 (Bounded geodesic image). There is a constant C = C ( Γ ) C=C(\Gamma) C = C ( Γ ) satisfying the following property. Let x , y x,y x , y be any two 0-cubes in X X X . Suppose there is a domain U ∈ S U\in\mathcal{S} U ∈ S represented by a convex subcomplex K = g X Λ K=gX_{\Lambda} K = g X Λ for some g ∈ W Γ g\in W_{\Gamma} g ∈ W Γ and Λ ⊂ Γ \Lambda\subset\Gamma Λ ⊂ Γ such that g K ( x ) , g K ( y ) \mathfrak{g}_K(x),\mathfrak{g}_K(y) g K ( x ) , g K ( y ) are not contained in a cube. Let ρ X ^ U ∈ g X Λ ⊂ X ^ \rho_{\widehat{X}}^U\in gX_{\Lambda}\subset\widehat{X} ρ X U ∈ g X Λ ⊂ X be the fixed point in X ^ \widehat{X} X for the domain U U U . Then for any geodesic α ⊂ X ^ \alpha\subset\widehat{X} α ⊂ X from x x x to y y y , we have
α ∩ N C ( ρ X ^ U ) ≠ ∅ . \alpha\cap N_C(\rho_{\widehat{X}}^U)\ne\varnothing. α ∩ N C ( ρ X U ) = ∅ .
The proof has two steps. First, we construct a standard product region P P P containing [ x K , y K ] [x_K,y_K] [ x K , y K ] and show that the gate g P ( x ) \mathfrak{g}_P(x) g P ( x ) belongs to the cubical interval [ x , y ] [x,y] [ x , y ] . Second, we show that this gate is uniformly close to ρ X ^ U \rho_{\widehat{X}}^U ρ X U and transfer this bound from the cubical interval to an arbitrary geodesic in X ^ \widehat{X} X .
Proof. By the construction of the maximized HHS structure, neither Λ \Lambda Λ nor lk ( Λ ) \operatorname{lk}(\Lambda) lk ( Λ ) is a clique. Put x K = g K ( x ) x_K=\mathfrak{g}_K(x) x K = g K ( x ) and y K = g K ( y ) y_K=\mathfrak{g}_K(y) y K = g K ( y ) . Every hyperplane separating x K x_K x K from y K y_K y K also separates x x x from y y y . Moreover,
Λ ′ = Label ( [ x K , y K ] ) ⊂ Λ \Lambda'=\operatorname{Label}([x_K,y_K])\subset\Lambda Λ ′ = Label ([ x K , y K ]) ⊂ Λ
is not a clique, because x K x_K x K and y K y_K y K do not lie in a common cube by assumption. Let Δ = lk ( Λ ′ ) \Delta=\operatorname{lk}(\Lambda') Δ = lk ( Λ ′ ) . The graph Δ \Delta Δ is not a clique, since it contains lk ( Λ ) \operatorname{lk}(\Lambda) lk ( Λ ) , and Λ ′ ⊂ lk ( Δ ) \Lambda'\subset\operatorname{lk}(\Delta) Λ ′ ⊂ lk ( Δ ) . Thus Λ ′ ∪ Δ \Lambda'\cup\Delta Λ ′ ∪ Δ lies in a nontrivial join. Let P P P be a standard product region containing [ x K , y K ] [x_K,y_K] [ x K , y K ] and associated to a maximal join containing Λ ′ ⋆ Δ \Lambda'\star\Delta Λ ′ ⋆ Δ . See Figure 7 .
Figure 7. The standard product region P P P and the subcomplex K K K are in gray. x K x_K x K (resp. y K y_K y K ) is the image of x x x (resp. y y y ) under the gate map g K \mathfrak{g}_K g K . Every geodesic from x x x to y y y in X ^ \widehat{X} X passes within a uniform distance of ρ X ^ U \rho_{\widehat{X}}^U ρ X U .
Let x P = g P ( x ) x_P=\mathfrak{g}_P(x) x P = g P ( x ) and y P = g P ( y ) y_P=\mathfrak{g}_P(y) y P = g P ( y ) . Let γ 1 \gamma_1 γ 1 be a combinatorial geodesic from x x x to x P x_P x P , γ 2 \gamma_2 γ 2 be a combinatorial geodesic from x P x_P x P to y P y_P y P and γ 3 \gamma_3 γ 3 be a combinatorial geodesic from y P y_P y P to y y y . Since the product region P P P is convex, γ 2 ⊂ P \gamma_2\subset P γ 2 ⊂ P . We claim that the concatenated combinatorial path γ = γ 1 γ 2 γ 3 \gamma=\gamma_1\gamma_2\gamma_3 γ = γ 1 γ 2 γ 3 is a geodesic from x x x to y y y . Since each γ i \gamma_i γ i is geodesic for i = 1 , 2 , 3 i = 1,2,3 i = 1 , 2 , 3 , it is enough to show that no hyperplane crosses both γ i \gamma_i γ i and γ j \gamma_j γ j for i ≠ j i \ne j i = j . By the bridge lemma (see for example Lemma 2.16 (4)), any hyperplane crossing γ 1 \gamma_1 γ 1 (resp. γ 3 \gamma_3 γ 3 ) separates P P P from x x x (resp. y y y ). Hence it does not cross γ 2 \gamma_2 γ 2 . Now suppose there is a hyperplane crossing γ 1 \gamma_1 γ 1 and γ 3 \gamma_3 γ 3 . Equivalently, there is a hyperplane separating x x x and y y y from P P P . Let W W W be such a hyperplane which contacts P P P . By the bridge lemma again,
Label ( W ) ∈ lk ( Label ( [ x P , y P ] ) ) ⊂ lk ( Λ ′ ) = Δ . \operatorname{Label}(W) \in\operatorname{lk}(\operatorname{Label}([x_P,y_P])) \subset\operatorname{lk}(\Lambda') = \Delta. Label ( W ) ∈ lk ( Label ([ x P , y P ])) ⊂ lk ( Λ ′ ) = Δ.
Hence W ∩ P ≠ ∅ W \cap P \ne\varnothing W ∩ P = ∅ . Contradiction. Hence γ \gamma γ is a combinatorial geodesic from x x x to y y y that has nonempty intersection with P P P .
Since d X ^ ( x K , ρ X ^ U ) ≤ 1 d_{\widehat{X}}(x_K,\rho_{\widehat{X}}^U) \le1 d X ( x K , ρ X U ) ≤ 1 and x K x_K x K and x P x_P x P lie in the same standard product region, we have
d X ^ ( x P , ρ X ^ U ) ≤ 2. d_{\widehat{X}}(x_P,\rho_{\widehat{X}}^U) \le2. d X ( x P , ρ X U ) ≤ 2.
Lemma 5.8 provides a point of a taut geodesic from x x x to y y y in X ^ \widehat{X} X within M + 1 M+1 M + 1 of x P x_P x P , where M M M is the maximal size of a clique in Γ \Gamma Γ . Any geodesic in X ^ \widehat{X} X from x x x to y y y fellow travels that taut geodesic within a constant depending only on the hyperbolicity constant δ \delta δ . Combining these bounds gives the required constant C = C ( Γ ) C=C(\Gamma) C = C ( Γ ) .
Visibility of geodesic rays in X ^ \widehat{X} X
In a proper hyperbolic space, any two distinct boundary points are joined by a bi-infinite geodesic. This need not hold in a general nonproper hyperbolic space because the Arzelà–Ascoli theorem is not applicable. Nevertheless, in this subsection, we will show that the points in ∂ X ^ \partial\widehat{X} ∂ X are still visible , that is, they can be joined by geodesics. Our argument is inspired by Minsky’s proof in curve graphs [56 Lemma 5.14] .
First, we prove a technical lemma showing that the combinatorial geodesic in X X X constructed from a taut geodesic in X ^ \widehat{X} X in Proposition 5.2 is a hierarchy path . Let ( X , S ) (X,\mathcal{S}) ( X , S ) be a maximized HHS structure on W Γ W_{\Gamma} W Γ in the sense of [1 ] . For each domain U ∈ S U \in\mathcal{S} U ∈ S , the associated hyperbolic space is either a point corresponding to a dummy domain, the nested-maximal space X ^ \widehat{X} X , or a coned-off contact graph, denoted by C U ^ \widehat{C U} C U , with a coarse Lipschitz projection π U : X → C U ^ \pi_U:X\to\widehat{C U} π U : X → C U ; see Subsection 2.6. We call π U \pi_U π U a coarse projection because the image π U ( x ) \pi_U(x) π U ( x ) is a subset of C U ^ \widehat{C U} C U . But diam C U ^ π U ( x ) ≤ 1 \operatorname{diam}_{\widehat{C U}}\pi_U(x)\leq1 diam C U π U ( x ) ≤ 1 , so π U \pi_U π U is coarsely well-defined. We recall the definition of the hierarchy path in HHSs.
Definition 7.2 (Hierarchy path). For each D ≥ 1 D\geq1 D ≥ 1 , a path γ \gamma γ in X X X is a D D D -hierarchy path if
(1) γ \gamma γ is a ( D , D ) (D,D) ( D , D ) -quasi-geodesic;
(2) for each domain U ∈ S U\in\mathcal{S} U ∈ S , π U ∘ γ \pi_U\circ\gamma π U ∘ γ is an unparametrized ( D , D ) (D,D) ( D , D ) -quasi-geodesic.
Lemma 7.3 (Hierarchy paths for taut geodesics). There exists a constant 1 ≤ D = D ( X , S ) 1\leq D=D(X,\mathcal{S}) 1 ≤ D = D ( X , S ) with the following property. Let ( x 0 , … , x n ) (x_0,\ldots,x_n) ( x 0 , … , x n ) be a geodesic in X ^ \widehat{X} X and suppose that γ \gamma γ is a combinatorial geodesic in X X X from x 0 x_0 x 0 to x n x_n x n such that vertices x 0 , … , x n x_0,\ldots,x_n x 0 , … , x n lie on γ \gamma γ in the same order. Then γ \gamma γ is a D D D -hierarchy path. In particular, combinatorial geodesics constructed in Proposition 5.2 for taut geodesics in X ^ \widehat{X} X are hierarchy paths.
Proof. It is enough to show that for each U ∈ S U\in\mathcal{S} U ∈ S , the image π U ( γ ) \pi_U(\gamma) π U ( γ ) is a uniform unparametrized quasi-geodesic.
On the ⊑ \sqsubseteq ⊑ -maximal domain X ^ \widehat{X} X . Let γ i \gamma_i γ i be a subpath of γ \gamma γ from x i − 1 x_{i-1} x i − 1 to x i x_i x i . Then γ = γ 1 ⋯ γ n \gamma=\gamma_1\cdots\gamma_n γ = γ 1 ⋯ γ n . Since γ \gamma γ is a combinatorial geodesic in X X X , so is γ i \gamma_i γ i for each i i i . Each γ i \gamma_i γ i is either a 1-cube or a geodesic contained in a standard product region (by convexity). Hence,
diam X ^ ( γ i ) ≤ 1. \operatorname{diam}_{\widehat{X}}(\gamma_i)\leq1. diam X ( γ i ) ≤ 1.
Then π X ^ ( γ ) \pi_{\widehat{X}}(\gamma) π X ( γ ) is an unparametrized quasi-geodesic.
On a proper domain U ∈ S U\in\mathcal{S} U ∈ S . Since the hyperbolic space associated with a dummy domain is a point, we may assume U U U is not a dummy domain. Then the associated hyperbolic space C U ^ \widehat{C U} C U is the coned-off contact graph. Suppose U U U is represented by a convex subcomplex K ⊂ X K\subset X K ⊂ X . The coarse projection π U : X → C U ^ \pi_U:X\to\widehat{C U} π U : X → C U is the following composition:
X → g K K → c U C U → i C U ^ . X \xrightarrow{g_K} K \xrightarrow{c_U} C U \xrightarrow{i} \widehat{C U}. X g K K c U C U i C U .
where g K g_K g K is the gate map, c U c_U c U sends each vertex to the set of hyperplanes whose carriers contain the vertex, and i i i is the inclusion map.
We first establish uniform control on the projection of every subpath of γ \gamma γ . Let a , z , b a,z,b a , z , b be vertices on γ \gamma γ such that z ∈ [ a , b ] z\in[a,b] z ∈ [ a , b ] , where [ a , b ] [a,b] [ a , b ] is the combinatorial convex hull in X X X containing a a a and b b b . Denote
a K = g K ( a ) , z K = g K ( z ) , b K = g K ( b ) . a_K=g_K(a),\qquad z_K=g_K(z),\qquad b_K=g_K(b). a K = g K ( a ) , z K = g K ( z ) , b K = g K ( b ) .
Since z ∈ [ a , b ] z\in[a,b] z ∈ [ a , b ] , we have
z K ∈ [ a K , b K ] . z_K\in[a_K,b_K]. z K ∈ [ a K , b K ] .
Choose H a ∈ π U ( a ) H_a\in\pi_U(a) H a ∈ π U ( a ) and H b ∈ π U ( b ) H_b\in\pi_U(b) H b ∈ π U ( b ) , and let
σ = ( H a = H 0 , … , H m = H b ) \sigma=(H_a=H_0,\ldots,H_m=H_b) σ = ( H a = H 0 , … , H m = H b )
be a geodesic in C U ^ \widehat{C U} C U from H a H_a H a to H b H_b H b .
Claim. π U ( z ) ⊆ N 2 ( σ ) \pi_U(z) \subseteq N_2(\sigma) π U ( z ) ⊆ N 2 ( σ ) .
Proof. If z K = a K z_K=a_K z K = a K , then it is clear that π U ( z ) ⊂ N 2 ( H a ) ⊂ N 2 ( σ ) \pi_U(z) \subset N_2(H_a) \subset N_2(\sigma) π U ( z ) ⊂ N 2 ( H a ) ⊂ N 2 ( σ ) . Now let W W W be the hyperplane dual to the last edge of a combinatorial geodesic from a K a_K a K to z K z_K z K . Since z K ∈ [ a K , b K ] z_K \in[a_K,b_K] z K ∈ [ a K , b K ] , the hyperplane W W W separates a K a_K a K from b K b_K b K . Now we construct a path α \alpha α from a K a_K a K to b K b_K b K as follows. For each pair of consecutive hyperplanes H i − 1 H_{i-1} H i − 1 and H i H_i H i in σ \sigma σ , we choose a vertex y i ∈ N ( H i − 1 ) ∩ N ( H i ) ∩ K y_i \in N(H_{i-1}) \cap N(H_i) \cap K y i ∈ N ( H i − 1 ) ∩ N ( H i ) ∩ K when they contact, or a vertex y i ∈ N ( H i ) y_i \in N(H_i) y i ∈ N ( H i ) when H i − 1 H_{i-1} H i − 1 and H i H_i H i do not contact and they both cross a proper convex subcomplex F i ⊂ K F_i \subset K F i ⊂ K whose contact graph is coned off in C ^ U \widehat{\mathcal C}_U C U . Also let y 0 = a K y_0=a_K y 0 = a K . Then we choose paths α i \alpha_i α i from y i − 1 y_{i-1} y i − 1 to y i y_i y i contained in N ( H i − 1 ) N(H_{i-1}) N ( H i − 1 ) when H i − 1 H_{i-1} H i − 1 and H i H_i H i contact or in N ( H i − 1 ) ∪ F i N(H_{i-1}) \cup F_i N ( H i − 1 ) ∪ F i for a proper subcomplex F i F_i F i when H i − 1 H_{i-1} H i − 1 and H i H_i H i do not contact. Such paths exist since N ( H i − 1 ) N(H_{i-1}) N ( H i − 1 ) and F i F_i F i are convex. Let α = α 1 ⋯ α m \alpha=\alpha_1\cdots\alpha_m α = α 1 ⋯ α m be the concatenation. Then α \alpha α is a combinatorial path from a K a_K a K to b K b_K b K and contained in K K K by construction. Since W W W is a hyperplane separating a K a_K a K and b K b_K b K , α \alpha α crosses W W W . Therefore, W W W either contacts H i H_i H i or crosses F i F_i F i for some i ∈ { 0 , ⋯ , m } i \in\{0,\cdots,m\} i ∈ { 0 , ⋯ , m } . In either case, d C ^ U ( W , H i ) ≤ 1 d_{\widehat{\mathcal C}_U}(W,H_i) \le1 d C U ( W , H i ) ≤ 1 . Since every hyperplane incident to z K z_K z K contacts W W W at least at z K z_K z K , this proves the claim. □
Now for an integer i i i , choose
p i ∈ π U ( γ ( i ) ) . p_i \in\pi_U(\gamma(i)). p i ∈ π U ( γ ( i )) .
Applying the preceding claim to every subpath gives, whenever r ≤ s ≤ t r \le s \le t r ≤ s ≤ t ,
d C ^ U ( p r , p s ) + d C ^ U ( p s , p t ) ≤ d C ^ U ( p r , p t ) + 4. (1) d_{\widehat{\mathcal C}_U}(p_r,p_s)+d_{\widehat{\mathcal C}_U}(p_s,p_t) \le d_{\widehat{\mathcal C}_U}(p_r,p_t)+4. \tag*{(1)} d C U ( p r , p s ) + d C U ( p s , p t ) ≤ d C U ( p r , p t ) + 4. (1)
Moreover, consecutive points p t , p t + 1 p_t,p_{t+1} p t , p t + 1 are at distance at most 2.
Therefore, ( p t ) (p_t) ( p t ) is a uniform unparametrized quasi-geodesic. Indeed, fix R > 4 R>4 R > 4 and select points successively, starting with p 0 p_0 p 0 , taking the next point at distance at least R R R from the previously selected point. Include the terminal point if it has not been selected, and denote the resulting sequence by q 0 , … , q ℓ q_0,\ldots,q_\ell q 0 , … , q ℓ . Then consecutive selected points are at distance between R R R and R + 2 R+2 R + 2 , except the last pair which are at distance at most R + 2 R+2 R + 2 . Applying (1) successively gives, for 0 ≤ i , j ≤ ℓ 0 \le i,j \le\ell 0 ≤ i , j ≤ ℓ ,
( R − 4 ) ∣ i − j ∣ − R ≤ d C ^ U ( q i , q j ) ≤ ( R + 2 ) ∣ i − j ∣ . (R-4)|i-j|-R \le d_{\widehat{\mathcal C}_U}(q_i,q_j) \le(R+2)|i-j|. ( R − 4 ) ∣ i − j ∣ − R ≤ d C U ( q i , q j ) ≤ ( R + 2 ) ∣ i − j ∣.
Thus the selected sequence forms a uniform quasi-geodesic. Since each p t p_t p t which is not contained in the sequence q 0 , … , q ℓ q_0,\ldots,q_\ell q 0 , … , q ℓ lies in R R R -neighborhood of some q t ′ q_{t'} q t ′ , the projection ( p t ) (p_t) ( p t ) is an unparametrized quasi-geodesic with constants independent of U U U or γ \gamma γ .
Now, by using taut geodesics and hierarchy paths, we construct a geodesic ray in X ^ \widehat X X from a basepoint to a boundary point.
Lemma 7.4. For each o ∈ X ^ o \in\widehat X o ∈ X and ν ∈ ∂ X ^ \nu\in\partial\widehat X ν ∈ ∂ X , there is a geodesic ray μ : [ 0 , ∞ ) → X ^ \mu:[0,\infty) \to\widehat X μ : [ 0 , ∞ ) → X with μ ( 0 ) = o \mu(0)=o μ ( 0 ) = o and μ ( ∞ ) = ν \mu(\infty)=\nu μ ( ∞ ) = ν .
Proof. Let ( x 0 , x 1 , … , x n , … ) (x_0,x_1,\ldots,x_n,\ldots) ( x 0 , x 1 , … , x n , … ) be a sequence of points in X ^ \widehat X X representing the boundary point ν ∈ ∂ X ^ \nu\in\partial\widehat X ν ∈ ∂ X . For each i i i , take a taut geodesic c i c_i c i from o o o to x i x_i x i , together with the ordered combinatorial geodesic lift α i \alpha_i α i provided in Proposition 5.2. By Lemma 7.3, the combinatorial geodesics α i \alpha_i α i are uniform hierarchy paths.
Our goal is to extract a limiting ray from c i c_i c i 's. By the δ \delta δ -hyperbolicity of X ^ \widehat X X and the definition of convergence at infinity, we have the following:
( ∗ ) (*) ( ∗ ) For each R > 0 R>0 R > 0 , there exists n n n so that for i , j ≥ n i,j \ge n i , j ≥ n the initial segment of length R R R in c i c_i c i is within δ \delta δ of c j c_j c j , and vice versa.
We plan to improve this coarse stabilization to exact stabilization in the locally infinite graph X ^ \widehat{X} X .
Let C C C be the constant from Proposition 7.1 (BGI). Pick R > 20 ( δ + C + 1 ) R > 20(\delta+ C + 1) R > 20 ( δ + C + 1 ) and let n n n be as in (*). For j ≥ n j \ge n j ≥ n , let v j ′ v'_j v j ′ be a vertex of c j c_j c j within distance R / 2 R/2 R /2 of o o o in X ^ \widehat{X} X .
Claim. The set ⋃ j ≥ n ( c j ( 0 ) ∩ B R / 2 ( o ) ) \bigcup_{j \ge n}\left(c_j^{(0)} \cap B_{R/2}(o)\right) ⋃ j ≥ n ( c j ( 0 ) ∩ B R /2 ( o ) ) is finite.
Proof. Let v v v be a vertex of c n c_n c n such that d X ^ ( v , v j ′ ) ≤ δ d_{\widehat{X}}(v,v'_j) \le\delta d X ( v , v j ′ ) ≤ δ . Such v v v exists by (*). Let W ∈ S W \in\mathcal{S} W ∈ S be a domain (in the HHS structure of W Γ W_\Gamma W Γ ) relevant to the pair ( v , v j ′ ) (v,v'_j) ( v , v j ′ ) , that is, d W ( v , v j ′ ) > M d_W(v,v'_j) > M d W ( v , v j ′ ) > M , where M = dim ( X ) M = \dim(X) M = dim ( X ) .
Recall that there is a fixed point ρ X ^ W ∈ X ^ \rho^W_{\widehat{X}} \in\widehat{X} ρ X W ∈ X . Since d W ( v , v j ′ ) > M d_W(v,v'_j) > M d W ( v , v j ′ ) > M , every geodesic from v v v to v j ′ v'_j v j ′ in X ^ \widehat{X} X passes through a C C C -neighborhood of ρ X ^ W \rho^W_{\widehat{X}} ρ X W by BGI. In particular, d X ^ ( ρ X ^ W , v j ′ ) ≤ δ + C d_{\widehat{X}}(\rho^W_{\widehat{X}},v'_j) \le\delta+ C d X ( ρ X W , v j ′ ) ≤ δ + C and d X ^ ( o , ρ X ^ W ) ≤ R 2 + ( δ + C ) d_{\widehat{X}}(o,\rho^W_{\widehat{X}}) \le\frac{R}{2} + (\delta+ C) d X ( o , ρ X W ) ≤ 2 R + ( δ + C ) .
We claim that for every j ≥ n j \ge n j ≥ n , d W ( x n , x j ) d_W(x_n,x_j) d W ( x n , x j ) is uniformly bounded from above. Let w w w and w j ′ w'_j w j ′ be points on c n c_n c n and c j c_j c j that are at least 5 C + 5 δ + 1 5C + 5\delta+ 1 5 C + 5 δ + 1 further from o o o than v v v and v j ′ v'_j v j ′ , respectively, and such that d X ^ ( w , w j ′ ) ≤ δ d_{\widehat{X}}(w,w'_j) \le\delta d X ( w , w j ′ ) ≤ δ . This is possible by (*) since R R R is large enough. Then we can connect w w w and w j ′ w'_j w j ′ by a geodesic in X ^ \widehat{X} X all of whose vertices are at distance at least C C C from ρ X ^ W \rho^W_{\widehat{X}} ρ X W . Then by applying BGI, d W ( w , w j ′ ) ≤ M d_W(w,w'_j) \le M d W ( w , w j ′ ) ≤ M . Since the geodesic segment from w w w to x n x_n x n and the geodesic segment from w j ′ w'_j w j ′ to x j x_j x j are at least C C C -apart from ρ X ^ W \rho^W_{\widehat{X}} ρ X W , by BGI again, d W ( w , x n ) d_W(w,x_n) d W ( w , x n ) and d W ( w j ′ , x j ) d_W(w'_j,x_j) d W ( w j ′ , x j ) are uniformly bounded from above. Hence, d W ( x n , x j ) ≤ 3 M d_W(x_n,x_j) \le3M d W ( x n , x j ) ≤ 3 M .
Recall that the image of a hierarchy path in each hyperbolic space is a uniform unparametrized quasi-geodesic. In particular, since v ∈ α n v \in\alpha_n v ∈ α n , v j ′ ∈ α j v'_j \in\alpha_j v j ′ ∈ α j , where α n , α j \alpha_n,\alpha_j α n , α j are hierarchy paths, there is a constant E = E ( X , S ) E = E(X,\mathcal{S}) E = E ( X , S ) such that
d W ( v , x n ) ≤ d W ( o , x n ) + E , d W ( x j , v j ′ ) ≤ d W ( x j , o ) + E . d_W(v,x_n) \le d_W(o,x_n) + E,\qquad d_W(x_j,v'_j) \le d_W(x_j,o) + E. d W ( v , x n ) ≤ d W ( o , x n ) + E , d W ( x j , v j ′ ) ≤ d W ( x j , o ) + E .
Then for all j ≥ n j \ge n j ≥ n , we have
d W ( v , v j ′ ) ≤ d W ( v , x n ) + d W ( x n , x j ) + d W ( x j , v j ′ ) ≤ d W ( o , x n ) + E + d W ( x n , x j ) + d W ( o , x j ) + E ≤ d W ( o , x n ) + d W ( x n , x j ) + d W ( o , x n ) + d W ( x n , x j ) + 2 E ≤ 2 d W ( o , x n ) + 6 M + 2 E . \begin{aligned}
d_W(v,v'_j) &\le d_W(v,x_n) + d_W(x_n,x_j) + d_W(x_j,v'_j) \\
&\le d_W(o,x_n) + E + d_W(x_n,x_j) + d_W(o,x_j) + E \\
&\le d_W(o,x_n) + d_W(x_n,x_j) + d_W(o,x_n) + d_W(x_n,x_j) + 2E \\
&\le2d_W(o,x_n) + 6M + 2E.
\end{aligned} d W ( v , v j ′ ) ≤ d W ( v , x n ) + d W ( x n , x j ) + d W ( x j , v j ′ ) ≤ d W ( o , x n ) + E + d W ( x n , x j ) + d W ( o , x j ) + E ≤ d W ( o , x n ) + d W ( x n , x j ) + d W ( o , x n ) + d W ( x n , x j ) + 2 E ≤ 2 d W ( o , x n ) + 6 M + 2 E .
For domains not satisfying the relevance inequality, the projection distance is at most M M M . Let s 0 s_0 s 0 be the constant from the Distance Formula (Theorem 2.24 ). Take a threshold s ≥ 2 s 0 + ( 6 M + 2 E ) + δ s \ge2s_0 + (6M + 2E) + \delta s ≥ 2 s 0 + ( 6 M + 2 E ) + δ . By the preceding computation and Remark 2.25, every domain contributing a nonzero term to the thresholded sum in the distance formula is contained in a fixed finite set of domains
S ~ x n = { W ∈ S ∣ d W ( o , x n ) ≥ s 0 } . \widetilde{\mathcal{S}}_{x_n} = \{W \in\mathcal{S} \mid d_W(o,x_n) \ge s_0\}. S x n = { W ∈ S ∣ d W ( o , x n ) ≥ s 0 } .
Thus,
∑ W ∈ S { d W ( v , v j ′ ) } s ≤ ∑ W ∈ S ~ x n d W ( v , v j ′ ) ≤ ∑ W ∈ S ~ x n 2 d W ( o , x n ) + ( 6 M + 2 E ) . \sum_{W\in\mathcal{S}}\{d_W(v,v'_j)\}_s \le\sum_{W\in\widetilde{\mathcal{S}}_{x_n}}d_W(v,v'_j) \le\sum_{W\in\widetilde{\mathcal{S}}_{x_n}}2d_W(o,x_n) + (6M + 2E). W ∈ S ∑ { d W ( v , v j ′ ) } s ≤ W ∈ S x n ∑ d W ( v , v j ′ ) ≤ W ∈ S x n ∑ 2 d W ( o , x n ) + ( 6 M + 2 E ) .
Since the right-hand side does not depend on j j j , v j ′ v'_j v j ′ or v v v , the distance formula provides a uniform upper bound for d ( v , v j ′ ) d(v,v'_j) d ( v , v j ′ ) . Since v v v is a vertex of a fixed path c n c_n c n and X X X is proper (locally finite), we have that the set
⋃ j ≥ n ( c j ( 0 ) ∩ B R / 2 ( o ) ) \bigcup_{j\geq n}\left(c_j^{(0)}\cap B_{R/2}(o)\right) j ≥ n ⋃ ( c j ( 0 ) ∩ B R /2 ( o ) )
is finite.
Thus, there are only finitely many possibilities for the initial segment of length R / 2 R/2 R /2 of c j c_j c j for all j ≥ n j\geq n j ≥ n . Then we can take a subsequence c j ( R ) c_j(R) c j ( R ) for which these initial segments are all the same. Then increase R R R and make the usual diagonalization argument. There is a convergent subsequence of { c i } \{c_i\} { c i } ; call the limit μ \mu μ . It accumulates on ν \nu ν since its initial segments are equal to the initial segments of c j c_j c j for all large j j j in the subsequence. Thus, μ \mu μ is the desired geodesic ray.
We can now join any two distinct boundary points in ∂ X ^ \partial\widehat{X} ∂ X by a bi-infinite geodesic.
Proposition 7.5 (Visibility of ∂ X ^ \partial\widehat{X} ∂ X ). For every pair of distinct points a , b ∈ ∂ X ^ a,b\in\partial\widehat{X} a , b ∈ ∂ X , there is a geodesic γ : ( − ∞ , ∞ ) → X ^ \gamma:(-\infty,\infty)\to\widehat{X} γ : ( − ∞ , ∞ ) → X such that γ ( − ∞ ) = a \gamma(-\infty)=a γ ( − ∞ ) = a and γ ( ∞ ) = b \gamma(\infty)=b γ ( ∞ ) = b .
Proof. Fix o ∈ X ^ o\in\widehat{X} o ∈ X . Let ν , μ \nu,\mu ν , μ be geodesic rays based at o o o converging to a , b a,b a , b , respectively. Let ν i = ν ( i ) \nu_i=\nu(i) ν i = ν ( i ) , μ i = μ ( i ) \mu_i=\mu(i) μ i = μ ( i ) be points on ν , μ \nu,\mu ν , μ , respectively. Let γ i \gamma_i γ i be a taut geodesic from μ i \mu_i μ i to ν i \nu_i ν i . Consider the geodesic triangle with sides ( o , μ i ) (o,\mu_i) ( o , μ i ) , ( o , ν i ) (o,\nu_i) ( o , ν i ) and γ i \gamma_i γ i . There is T ∈ N T\in\mathbb{N} T ∈ N such that for each i ≥ T i\geq T i ≥ T , μ i ∉ N δ ( ν ) \mu_i\notin N_\delta(\nu) μ i ∈ / N δ ( ν ) . Since each geodesic triangle is δ \delta δ -slim, γ i ∩ B δ ( μ T ) ≠ ∅ \gamma_i\cap B_\delta(\mu_T)\ne\varnothing γ i ∩ B δ ( μ T ) = ∅ . Suppose p i ∈ B δ ( μ T ) ∩ γ i p_i\in B_\delta(\mu_T)\cap\gamma_i p i ∈ B δ ( μ T ) ∩ γ i for each i ≥ T i\geq T i ≥ T . In particular, d X ^ ( p i , p j ) ≤ 2 δ d_{\widehat{X}}(p_i,p_j)\leq2\delta d X ( p i , p j ) ≤ 2 δ for any i , j ≥ T i,j\geq T i , j ≥ T . We claim that there is a constant δ ′ \delta' δ ′ satisfying the following.
(**) For any R > 0 R>0 R > 0 there exists N N N such that for all i ≥ N i\geq N i ≥ N the subsegments of γ i \gamma_i γ i of radius R R R centered on p i p_i p i are within δ ′ \delta' δ ′ of each other.
Indeed, we fix p i p_i p i with i ≥ T i\geq T i ≥ T . By δ \delta δ -hyperbolicity and the definition of the geodesic ray μ \mu μ , there is N 1 ∈ N N_1\in\mathbb{N} N 1 ∈ N such that for any j , k ≥ N 1 j,k\geq N_1 j , k ≥ N 1 the initial segment of length R R R in c j c_j c j is within δ \delta δ of c k c_k c k , and vice versa, where c j , c k c_j,c_k c j , c k are taut geodesics from p i p_i p i to μ j , μ k \mu_j,\mu_k μ j , μ k , respectively. Similarly, there is N 2 ∈ N N_2\in\mathbb{N} N 2 ∈ N such that for any j , k ≥ N 2 j,k\geq N_2 j , k ≥ N 2 the initial segment of length R + 2 δ R+2\delta R + 2 δ in a taut geodesic from p i p_i p i to ν j \nu_j ν j is within δ \delta δ of a taut geodesic from p i p_i p i to ν k \nu_k ν k , and vice versa. Take N = max { N 1 , N 2 , T } N=\max\{N_1,N_2,T\} N = max { N 1 , N 2 , T } .
Let γ i + \gamma_i^+ γ i + be a subpath of γ i \gamma_i γ i from p i p_i p i to μ i \mu_i μ i of length R R R . We want to show that γ i + \gamma_i^+ γ i + and γ j + \gamma_j^+ γ j + are within distance δ ′ = 3 δ \delta'=3\delta δ ′ = 3 δ for any i , j ≥ N i,j\geq N i , j ≥ N . Since i ≥ N ≥ N 1 i\geq N\geq N_1 i ≥ N ≥ N 1 , the initial segment with length R R R of c j c_j c j is within δ \delta δ of γ i + \gamma_i^+ γ i + , and vice versa. Furthermore, since the distance between p i p_i p i and p j p_j p j is bounded from above by 2 δ 2\delta 2 δ , the initial segment with length R R R of c j c_j c j is within 2 δ 2\delta 2 δ of γ j + \gamma_j^+ γ j + , and vice versa.
The same argument works for subpaths of γ i , γ j \gamma_i,\gamma_j γ i , γ j from p i , p j p_i,p_j p i , p j , respectively, to points on ν \nu ν . This proves the claim.
Then an argument similar to the proof of Lemma 7.4 then produces a subsequence converging to a bi-infinite geodesic with endpoints a a a and $b.
Geodesic almost extendability of X ^ \widehat{X} X
In a geodesic hyperbolic space, ideal triangles are uniformly slim; see [48 Lemma 1.84] . A geodesic triangle is ideal if at least one of its vertices lies in the Gromov boundary.
We use the uniform slimness for X ^ \widehat{X} X and the W Γ W_\Gamma W Γ -action to show that X ^ \widehat{X} X is geodesic almost extendable . Mihalik observed that infinite hyperbolic groups have this property [15 Lemma 3.1] .
Definition 7.6. (Geodesic almost extendability) Let Y Y Y be a geodesic metric space. We say that Y Y Y is geodesic almost extendable if there is a constant E E E such that for all points p p p and q q q in Y Y Y there is a geodesic ray emanating from p p p that passes within a distance E E E of q q q . The constant E E E is called a geodesic almost extendability constant of Y Y Y .
Note that in the literature, geodesic almost extendable spaces are sometimes called visual spaces (for example, in [18 ] ).
Proposition 7.7 (Geodesic almost extendability). Let X ^ \widehat{X} X be the maximized hyperbolic space associated with the right-angled Coxeter group W Γ W_{\Gamma} W Γ . Suppose that X ^ \widehat{X} X has nonempty Gromov boundary. Then there is a constant C = C ( Γ ) C=C(\Gamma) C = C ( Γ ) such that X ^ \widehat{X} X is geodesic almost extendable with the constant C C C .
Proof. Let p , q ∈ X ^ p,q\in\widehat{X} p , q ∈ X . Lemma 3.4 and Proposition 7.5 give a bi-infinite geodesic α \alpha α . Choose a vertex a ∈ α a\in\alpha a ∈ α . Since W Γ W_{\Gamma} W Γ acts transitively on the vertex set, a translate β = g α \beta=g\alpha β = g α passes through q q q .
Let β + \beta^{+} β + and β − \beta^{-} β − be the endpoints of β \beta β . By Lemma 7.4 , choose geodesic rays γ + \gamma^{+} γ + and γ − \gamma^{-} γ − based at p p p and ending at β + \beta^{+} β + and β − \beta^{-} β − , respectively. These rays together with β \beta β form an ideal triangle. Since the ideal triangles are uniformly slim for some constant δ ′ \delta' δ ′ only depending on δ = δ ( Γ ) \delta=\delta(\Gamma) δ = δ ( Γ ) , q ∈ β q\in\beta q ∈ β lies within δ ′ \delta' δ ′ of γ + ∪ γ − \gamma^{+}\cup\gamma^{-} γ + ∪ γ − . One of these two rays therefore has nonempty intersection with B δ ′ ( q ) B_{\delta'}(q) B δ ′ ( q ) , so δ ′ \delta' δ ′ is an almost-extendability constant. □
Noncompactness of ∂ X ^ \partial\widehat{X} ∂ X
In this subsection, we will show that when ∂ X ^ \partial\widehat{X} ∂ X is nonempty and W Γ W_{\Gamma} W Γ is not hyperbolic, the boundary ∂ X ^ \partial\widehat{X} ∂ X is non-compact, as are the boundaries of curve graphs and fine curve graphs. This contrasts with the boundary of a proper hyperbolic space, which is always compact. The proof is inspired by [53 Theorem 1.4] .
The strategy is similar to showing that the Gromov boundary of an infinite-valence regular tree is non-compact, i.e. constructing a sequence of boundary points with no convergent subsequence. In the case of X ^ \widehat{X} X , the key is to find an element g ∈ W Γ g\in W_{\Gamma} g ∈ W Γ whose orbit is bounded in X ^ \widehat{X} X but whose action on the hyperbolic space C U ^ \widehat{\mathcal{C}U} C U associated with a nested domain is loxodromic. Bounded geodesic image then shows that a suitable subsequence of the orbit { g n p } \{g^{n}p\} { g n p } is uniformly separated in a visual metric on ∂ X ^ \partial\widehat{X} ∂ X .
Theorem 7.8 (Non-compactness of ∂ X ^ \partial\widehat{X} ∂ X ). Let W Γ W_{\Gamma} W Γ be a right-angled Coxeter group and X ^ = X ^ Γ \widehat{X}=\widehat{X}_{\Gamma} X = X Γ be the maximized hyperbolic space. Then the Gromov boundary ∂ X ^ \partial\widehat{X} ∂ X is nonempty and non-compact if and only if Γ \Gamma Γ is neither a clique nor a nontrivial join and contains an induced 4 4 4 -cycle.
Proof. Necessity. By Lemma 3.4 , the boundary is nonempty exactly when Γ \Gamma Γ is neither a clique nor a nontrivial join. If Γ \Gamma Γ has no induced square, then W Γ W_{\Gamma} W Γ is hyperbolic and no standard product regions are coned off. Thus X ^ = X ( 1 ) \widehat{X}=X^{(1)} X = X ( 1 ) is proper and its Gromov boundary is compact.
Sufficiency. Conversely, suppose that Γ \Gamma Γ is neither a clique nor a nontrivial join and that it contains an induced 4 4 4 -cycle labeled by { a , b , c , d } \{a,b,c,d\} { a , b , c , d } in order. Let Λ = { a , c } \Lambda=\{a,c\} Λ = { a , c } . Then Λ \Lambda Λ is not a clique and lk ( Λ ) \operatorname{lk}(\Lambda) lk ( Λ ) contains the non-adjacent vertices b , d b,d b , d . Hence Λ \Lambda Λ determines a non-maximal domain U U U of the maximized HHS structure, and its associated coned-off contact graph C U ^ \widehat{\mathcal{C}U} C U has infinite diameter. The element g = a c ∈ W Γ g=ac\in W_{\Gamma} g = a c ∈ W Γ acts loxodromically on C U ^ \widehat{\mathcal{C}U} C U . On the other hand, g g g preserves X Λ X_{\Lambda} X Λ , which lies in the standard product region X Λ ∗ lk ( Λ ) X_{\Lambda*\operatorname{lk}(\Lambda)} X Λ ∗ lk ( Λ ) and therefore has diameter at most one in X ^ \widehat{X} X .
Fix a basepoint o ∈ X ^ o \in\widehat{X} o ∈ X , a boundary point p ∈ ∂ X ^ p \in\partial\widehat{X} p ∈ ∂ X and a visual metric d ∞ d_{\infty} d ∞ on ∂ X ^ \partial\widehat{X} ∂ X . Choose a representing sequence ( x i ) (x_i) ( x i ) for p p p . For n ∈ Z n \in\mathbb{Z} n ∈ Z set p n = g n p p_n = g^n p p n = g n p . We want to show that the sequence ( p n ) ⊂ ∂ X ^ (p_n) \subset\partial\widehat{X} ( p n ) ⊂ ∂ X has no convergent subsequence.
Let τ > 0 \tau> 0 τ > 0 be the stable translation length of g g g on C U ^ \widehat{C U} C U . Choose L L L so that L τ > M L\tau> M Lτ > M , where M = dim X M = \dim X M = dim X . For any s , t ∈ L Z s,t \in L\mathbb{Z} s , t ∈ L Z with s ≠ t s \ne t s = t and every i i i ,
d U ( g s x i , g t x i ) ≥ ∣ s − t ∣ τ > M . d_U(g^s x_i,g^t x_i) \ge|s-t|\tau> M. d U ( g s x i , g t x i ) ≥ ∣ s − t ∣ τ > M .
Proposition 7.1 therefore implies that every geodesic in X ^ \widehat{X} X from g s x i g^s x_i g s x i to g t x i g^t x_i g t x i meets a fixed C C C -neighborhood of ρ X ^ U \rho^{U}_{\widehat{X}} ρ X U . In a hyperbolic space, the Gromov product based at o o o coarsely equals the distance from o o o to a geodesic between the two points. Consequently there is a constant
D = D ( δ , C , d X ^ ( o , ρ X ^ U ) ) D = D\bigl(\delta,C,d_{\widehat{X}}(o,\rho^{U}_{\widehat{X}})\bigr) D = D ( δ , C , d X ( o , ρ X U ) )
such that
⟨ g s x i , g t x i ⟩ o ≤ D \langle g^s x_i,g^t x_i\rangle_o \le D ⟨ g s x i , g t x i ⟩ o ≤ D
for all i i i and all distinct s , t ∈ L Z s,t \in L\mathbb{Z} s , t ∈ L Z . Since ( g s x i ) (g^s x_i) ( g s x i ) and ( g t x i ) (g^t x_i) ( g t x i ) represent p s p_s p s and p t p_t p t , respectively, the definition of the boundary Gromov product gives ⟨ p s , p t ⟩ o ≤ D \langle p_s,p_t\rangle_o \le D ⟨ p s , p t ⟩ o ≤ D . Proposition 2.20 now provides a constant ϵ > 0 \epsilon> 0 ϵ > 0 , independent of s s s and t t t , such that
d ∞ ( p s , p t ) ≥ ϵ ( s , t ∈ L Z , s ≠ t ) . d_{\infty}(p_s,p_t) \ge\epsilon\qquad(s,t \in L\mathbb{Z},\ s \ne t). d ∞ ( p s , p t ) ≥ ϵ ( s , t ∈ L Z , s = t ) .
Thus ( p n L ) n ≥ 0 (p_{nL})_{n\ge0} ( p n L ) n ≥ 0 has no convergent subsequence. Since the boundary is metrizable, it is non-compact. □ \square □
Remark 7.9. By arguing as in the proof of Theorem 7.8 , one can show the Gromov boundary ∂ X ^ Γ \partial\widehat{X}_{\Gamma} ∂ X Γ is nowhere locally compact when Γ \Gamma Γ is neither a clique nor a nontrivial join and contains an induced 4-cycle. We leave it as an exercise for readers.
The fan-and-filter construction
The fan-and-filter construction is the main combinatorial ingredient in the proof of Theorem 1.1 . The aim is, by using fans and filters, to show that two points in a sphere with radius R R R in X ^ \widehat{X} X can be connected by a path such that (1) vertices of this path are at distance at least R R R from the center of the sphere; and (2) the diameter of this path has a linear upper bound with respect to the distance between endpoints with the linearity constant only depending on Γ \Gamma Γ (Proposition 8.8 ).
A fan is a van Kampen diagram, which locally pushes a path in X X X away from one standard product region. Iterating this operation produces a filter, which is also a van Kampen diagram of W Γ W_{\Gamma} W Γ whose directed paths are combinatorial geodesics in X Γ X_{\Gamma} X Γ and do not have arbitrarily long subpaths in standard product regions. By using filters, one can push a path in X X X away from certain unions of standard product regions whose image in X ^ \widehat{X} X forms a closed ball.
We begin with a warm-up to show when a standard product region does not separate X X X .
Throughout this section, we denote by X = X Γ X = X_{\Gamma} X = X Γ the Davis complex associated with the graph Γ \Gamma Γ and we assume that the defining graph Γ \Gamma Γ satisfies the following conditions:
(a) Γ \Gamma Γ is connected, is not a clique, and has no separating clique;
(b) Γ \Gamma Γ has no separating nontrivial join;
(c) for every vertex v ∈ V ( Γ ) v \in V(\Gamma) v ∈ V ( Γ ) , the link of v v v is not contained in a nontrivial join.
Avoiding one standard product region
In Proposition 8.1, we will show that an edge path in Γ \Gamma Γ avoiding a nontrivial join could produce a combinatorial path in X = X Γ X = X_\Gamma X = X Γ avoiding a standard product region. This idea will motivate the construction of fans in Subsection 8.2.
Recall that a combinatorial path α \alpha α in X X X determines a word α ˉ \bar{\alpha} α ˉ by reading edge labels, and Label ( α ) ⊂ V ( Γ ) \operatorname{Label}(\alpha) \subset V(\Gamma) Label ( α ) ⊂ V ( Γ ) is the set of letters occurring in that word. If α \alpha α and β \beta β are combinatorial geodesics with the same endpoints, they cross the same set of hyperplanes and hence Label ( α ) = Label ( β ) \operatorname{Label}(\alpha) = \operatorname{Label}(\beta) Label ( α ) = Label ( β ) . We therefore write Label ( [ x , y ] ) \operatorname{Label}([x,y]) Label ([ x , y ]) for this vertex set.
Proposition 8.1. Let X = X Γ X = X_\Gamma X = X Γ , where Γ \Gamma Γ satisfies conditions (a), (b) and (c) . If P ⊂ X P \subset X P ⊂ X is a standard product region, then X ∖ P X \setminus P X ∖ P is path-connected.
Proof. Let x , y ∉ P x,y \notin P x , y ∈ / P and choose a combinatorial geodesic α \alpha α from x x x to y y y . There is nothing to prove if α ∩ P = ∅ \alpha\cap P = \varnothing α ∩ P = ∅ . Otherwise let P = g X Λ P = gX_\Lambda P = g X Λ , where Λ ⊂ Γ \Lambda\subset\Gamma Λ ⊂ Γ is a maximal nontrivial join and g ∈ W Γ g \in W_\Gamma g ∈ W Γ . Convexity of P P P implies that α ∩ P \alpha\cap P α ∩ P is a (possibly trivial) subpath α ′ ⊂ α \alpha' \subset\alpha α ′ ⊂ α . Write
α ′ ‾ = a 1 ⋯ a n , \overline{\alpha'} = a_1 \cdots a_n, α ′ = a 1 ⋯ a n ,
where a i ∈ V ( Γ ) a_i \in V(\Gamma) a i ∈ V ( Γ ) for i = 1 , … , n i = 1,\ldots,n i = 1 , … , n . Denote the successive vertices of α ′ \alpha' α ′ by p 0 , … , p n p_0,\ldots,p_n p 0 , … , p n . Let e v e_v e v and e w e_w e w be the edges of α ∖ α ′ \alpha\setminus\alpha' α ∖ α ′ incident to p 0 p_0 p 0 and p n p_n p n , with labels v v v and w w w in V ( Γ ) V(\Gamma) V ( Γ ) , respectively. Note that every 1-cube in X X X containing a 0-cube in g X Λ ( 0 ) gX_\Lambda^{(0)} g X Λ ( 0 ) lies in g X Λ gX_\Lambda g X Λ exactly when its label lies in Λ \Lambda Λ . Since e v e_v e v and e w e_w e w are not in P P P , while p 0 , p n ∈ P p_0,p_n \in P p 0 , p n ∈ P , we have v , w ∉ Λ v,w \notin\Lambda v , w ∈ / Λ . Since α ′ ⊂ P \alpha' \subset P α ′ ⊂ P , a i ∈ Λ a_i \in\Lambda a i ∈ Λ for each i = 1 , … , n i = 1,\ldots,n i = 1 , … , n . See Figure 8 .
Figure 8. A subpath α ′ \alpha' α ′ of α \alpha α contained in a product region P P P is pushed to a subpath of β \beta β across two hyperplane carriers and three fans.
For each 1 ≤ i ≤ n 1 \leq i \leq n 1 ≤ i ≤ n , condition (c) gives a vertex a i ′ ∈ lk ( a i ) ∖ Λ a_i' \in\operatorname{lk}(a_i) \setminus\Lambda a i ′ ∈ lk ( a i ) ∖ Λ , because, otherwise, lk ( a i ) \operatorname{lk}(a_i) lk ( a i ) would be contained in the nontrivial join Λ \Lambda Λ . condition (c) also guarantees the defining graph Γ \Gamma Γ is not a nontrivial join. In particular, Γ ≠ Λ \Gamma\neq\Lambda Γ = Λ . Then condition (b) implies that Γ ∖ Λ \Gamma\setminus\Lambda Γ ∖ Λ is connected. Set a 0 ′ = v a_0' = v a 0 ′ = v and a n + 1 ′ = w a_{n+1}' = w a n + 1 ′ = w . Hence, for 0 ≤ i ≤ n 0 \leq i \leq n 0 ≤ i ≤ n , choose an edge path
τ i = ( a i ′ = t i 0 , t i 1 , … , t i r i = a i + 1 ′ ) \tau_i = (a_i' = t_i^0,t_i^1,\ldots,t_i^{r_i} = a_{i+1}') τ i = ( a i ′ = t i 0 , t i 1 , … , t i r i = a i + 1 ′ )
in Γ ∖ Λ \Gamma\setminus\Lambda Γ ∖ Λ .
At p i p_i p i , the edge path τ i ⊂ Γ \tau_i \subset\Gamma τ i ⊂ Γ determines a chain of squares in X X X . Remove from the boundary of this chain of squares its first and last 1-cubes, both of which are incident to p i p_i p i in X X X , and denote the remaining path by β i \beta_i β i . Since every 0-cube of β i \beta_i β i is separated from p i p_i p i by a hyperplane with label not in Λ \Lambda Λ , β i ∩ P = ∅ \beta_i \cap P = \varnothing β i ∩ P = ∅ .
Since a i ′ ∈ lk ( a i ) a_i' \in\operatorname{lk}(a_i) a i ′ ∈ lk ( a i ) , the corresponding square labeled by a i a i ′ a i a i ′ a_i a_i' a_i a_i' a i a i ′ a i a i ′ translates the a i a_i a i -edge of α ′ \alpha' α ′ to an edge e i e_i e i connecting the terminal vertex of β i − 1 \beta_{i-1} β i − 1 with the initial vertex of β i \beta_i β i . Both endpoints of e i e_i e i lie outside of P P P (since a i ′ ∉ Λ a'_i \notin\Lambda a i ′ ∈ / Λ ), and so does e i e_i e i . The concatenation
β = β 0 e 1 β 1 e 2 ⋯ e n β n \beta=\beta_0e_1\beta_1e_2\cdots e_n\beta_n β = β 0 e 1 β 1 e 2 ⋯ e n β n
has the same endpoints as e v α ′ e w e_v\alpha'e_w e v α ′ e w and is disjoint from P P P . Replacing e v α ′ e w e_v\alpha'e_w e v α ′ e w by β \beta β gives a path from x x x to y y y in X ∖ P X\setminus P X ∖ P .
Fans
We now define a fan , a particular connected van Kampen diagram of W Γ W_\Gamma W Γ . A van Kampen diagram for W Γ W_\Gamma W Γ is a planar square complex D D D equipped with a cellular map D → X D\to X D → X whose restriction
D ( 1 ) → X ( 1 ) D^{(1)}\to X^{(1)} D ( 1 ) → X ( 1 )
preserves edge labels. Hence every square records a commuting relation in W Γ W_\Gamma W Γ . We note that there may be distinct vertices and edges of D D D with the same image in X X X .
Let γ \gamma γ be a path with label γ ‾ = s 0 ⋯ s n \overline{\gamma}=s_0\cdots s_n γ = s 0 ⋯ s n . Its wide tail γ ′ \gamma' γ ′ is its longest terminal subpath whose labels lie in a nontrivial join. Equivalently, γ ′ \gamma' γ ′ has the image contained in a standard product region of X Γ X_\Gamma X Γ . The wide tail may be empty. Let f 0 f_0 f 0 and f r f_r f r be edges incident to the terminal vertex of γ \gamma γ , with labels v 0 , v r v_0,v_r v 0 , v r , such that γ f 0 \gamma f_0 γ f 0 and γ f r \gamma f_r γ f r are geodesic. The following definition is from [25 Definition 6.1] for right-angled Coxeter groups.
Definition 8.2 (Fan). A fan F F F is a van Kampen diagram consisting of a path γ \gamma γ , a sequence of at least three edges f 0 , … , f r f_0,\ldots,f_r f 0 , … , f r incident to the terminal vertex of γ \gamma γ , and, for each 0 ≤ i < r 0\le i<r 0 ≤ i < r , a square c i c_i c i spanned by f i f_i f i and f i + 1 f_{i+1} f i + 1 with the following properties:
(1) γ f i ‾ \overline{\gamma f_i} γ f i is geodesic in W Γ W_\Gamma W Γ for each 0 ≤ i ≤ r 0\le i\le r 0 ≤ i ≤ r .
(2) If γ ′ \gamma' γ ′ is the wide tail of γ ‾ \overline{\gamma} γ , then
(a) either Label ( γ ′ ) \operatorname{Label}(\gamma') Label ( γ ′ ) is a clique;
(b) or there is a nontrivial join subgraph Δ ⊂ Γ \Delta\subset\Gamma Δ ⊂ Γ such that Label ( γ ′ ) ⊂ Δ \operatorname{Label}(\gamma')\subset\Delta Label ( γ ′ ) ⊂ Δ and v i ∉ Δ v_i\notin\Delta v i ∈ / Δ , where v i = Label ( f i ) v_i=\operatorname{Label}(f_i) v i = Label ( f i ) , for each 0 < i < r 0<i<r 0 < i < r .
We call γ \gamma γ the base path , f i f_i f i the fan edges . In particular, f 0 , f r f_0,f_r f 0 , f r are called the left and right fan edges, and the other f i f_i f i ’s are interior fan edges . Orient γ \gamma γ toward the common initial vertex of the fan edges and order the fan edges from left to right. A path is directed if it follows the resulting orientation. In each square c i c_i c i , the vertex opposite the terminal vertex of γ \gamma γ is called the top vertex . Moreover, the edge opposite f i + 1 f_{i+1} f i + 1 is the top-left edge and the edge opposite f i f_i f i is the top-right edge .
The next lemma gives the required fans. For proofs, see [61 Lemmas 5.5 and 5.6] in the right-angled case and [25 Lemma 6.4] for wide-spherical-avoidant Coxeter groups. By Lemma 2.14 , the graph hypotheses (conditions (a)–(c)) in our main theorem imply the assumptions needed for the fan-and-filter construction in the literature.
Lemma 8.3 (Existence of fans). Suppose Γ \Gamma Γ satisfies conditions (a)–(c). Let γ \gamma γ be an abstract labeled path with γ ‾ ∈ W Γ \overline{\gamma}\in W_\Gamma γ ∈ W Γ being geodesic, and let e l , e r e_l,e_r e l , e r be distinct abstract edges incident to the terminal vertex of γ \gamma γ . Suppose that γ e l \gamma e_l γ e l and γ e r \gamma e_r γ e r are geodesic. Then there is a fan
F ( γ , e l , e r ) F(\gamma,e_l,e_r) F ( γ , e l , e r ) whose base path is γ \gamma γ with left and right fan edges e l e_l e l and e r e_r e r , respectively.
The following lemma is from [25 Lemma 6.5] and also outlined in [61 Lemma 3.9] . We give a proof in the case of right-angled Coxeter groups for completeness.
Lemma 8.4. Let F F F be a fan with base path γ \gamma γ . Let ρ \rho ρ be a directed path in F F F starting at the endpoint of γ \gamma γ . Then γ ‾ ρ \overline{\gamma}\rho γ ρ is geodesic.
Proof. A directed path beginning at the terminal vertex of γ \gamma γ consists of one fan edge, possibly followed by one top fan edge contained in a square. The one-edge case is guaranteed to be geodesic by the definition of the fan. In the two-edge case, write γ = s 0 ⋯ s n \gamma= s_0\cdots s_n γ = s 0 ⋯ s n and suppose the two edge labels are s , t s,t s , t . By the definition of a fan, both s 0 ⋯ s n s s_0\cdots s_ns s 0 ⋯ s n s and s 0 ⋯ s n t s_0\cdots s_nt s 0 ⋯ s n t are geodesic. If s 0 ⋯ s n s t s_0\cdots s_nst s 0 ⋯ s n s t were not geodesic, the deletion condition would produce an earlier occurrence of t t t commuting with every letter in between. The same cancellation would shorten s 0 ⋯ s n t s_0\cdots s_nt s 0 ⋯ s n t , a contradiction. These two cases exhaust the directed paths in a single fan. □
Filters
Given two geodesic rays based at the same point in X X X , we will build a one-ended van Kampen diagram, called a filter , by attaching fans inductively. This filter construction was used to establish local connectivity of the visual boundary of RACGs in [61 ] and connectivity and local connectivity of the Morse boundary of more general Coxeter groups in [25 ] . We study a truncated filter whose image in X ^ \widehat{X} X produces a path in spheres with the desired radii.
Let α \alpha α and β \beta β be combinatorial geodesic rays with a common initial vertex o ∈ X o \in X o ∈ X . Consider α \alpha α and β \beta β as abstract labeled paths mapping to X ( 1 ) X^{(1)} X ( 1 ) . Let η \eta η be their maximal common initial subpath and let z 0 z_0 z 0 be its terminal vertex. Assume that the rays have distinct first edges after z 0 z_0 z 0 , and write
α = η a 0 a 1 ⋯ , β = η b 0 b 1 ⋯ . \alpha= \eta a_0a_1\cdots,\qquad\beta= \eta b_0b_1\cdots. α = η a 0 a 1 ⋯ , β = η b 0 b 1 ⋯ .
The path η \eta η is called the common stem .
Begin with the abstract union F 0 = α ∪ η β \mathcal{F}_0 = \alpha\cup_{\eta} \beta F 0 = α ∪ η β and attach the fan
F ( η , a 0 , b 0 ) . F(\eta,a_0,b_0). F ( η , a 0 , b 0 ) .
This is the level-one fan, and the resulting diagram is F 1 \mathcal{F}_1 F 1 . We orient α \alpha α and β \beta β from o o o and F 1 \mathcal{F}_1 F 1 has the induced orientation. After deleting the top-left edge of every square in the level-one fan, the remaining edges, together with F 0 ( 1 ) \mathcal{F}_0^{(1)} F 0 ( 1 ) , form a spanning tree T 1 T_1 T 1 .
Suppose that F N \mathcal{F}_N F N and its spanning tree T N T_N T N have been constructed. If v v v is the terminal vertex of a fan edge in a level-N N N fan, there are two outgoing edges immediately above v v v , ordered from left to right based on the orientation of F N \mathcal{F}_N F N . Let γ v \gamma_v γ v be the unique path in T N T_N T N from o o o to v v v , and denote the two outgoing edges by e v l , e v r e_v^l,e_v^r e v l , e v r . The simultaneous induction in Lemma 8.5 says that both γ v e v l \gamma_ve_v^l γ v e v l and γ v e v r \gamma_ve_v^r γ v e v r are geodesic. By applying Lemma 8.3, we attach a fan
F ( γ v , e v l , e v r ) F(\gamma_v,e_v^l,e_v^r) F ( γ v , e v l , e v r )
at every such vertex. These are the level-( N + 1 ) (N+1) ( N + 1 ) fans. The resulting diagram is F N + 1 \mathcal{F}_{N+1} F N + 1 , and T N + 1 T_{N+1} T N + 1 is obtained by deleting the top-left edge of every square in every fan. Define
F = lim ⟶ F N , T = lim ⟶ T N \mathcal{F}=\lim_{\longrightarrow}\mathcal{F}_N,\qquad T=\lim_{\longrightarrow}T_N F = ⟶ lim F N , T = ⟶ lim T N
by taking the direct limit. We call F \mathcal{F} F the full filter spanning α \alpha α and β \beta β , and T T T its spanning tree .
We record some properties of full filters.
Lemma 8.5 (Basic properties of filters). For every N ≥ 1 N \ge1 N ≥ 1 ,
(1) T N T_N T N is a spanning tree of F N ( 1 ) \mathcal{F}_N^{(1)} F N ( 1 ) , and T N ⊂ T N + 1 T_N \subset T_{N+1} T N ⊂ T N + 1 ;
(2) every directed path in F N \mathcal{F}_N F N maps to a combinatorial geodesic in X X X ;
every vertex beyond the common stem has one or two incoming edges in F N \mathcal{F}_N F N , depending on whether it is the top vertex of a square in a fan. And exactly one of these incoming edges belongs to T N T_N T N . Consequently, the unique path in T N T_N T N from o o o to any vertex is directed;
the path η \eta η is contained in the tree path from o o o to every vertex in the part of the filter beyond z 0 z_0 z 0 ;
every terminal vertex of a level-N N N fan edge has exactly two outgoing edges, and these are the left and right fan edges of the unique fan attached there at level N + 1 N+1 N + 1 .
We refer readers to the discussion before [61 Lemma 5.8] or [25 Lemmas 6.6 and 6.7] for a proof.
We may also use finite paths instead of rays for a van Kampen diagram. Recall that every finite combinatorial geodesic in X X X extends to a combinatorial geodesic ray. At each endpoint the set of labels that shorten the current word spans a clique (Remark 2.18), and this clique is a proper subgraph of Γ \Gamma Γ because Γ \Gamma Γ is not a clique. Thus finite side paths of a van Kampen diagram may be extended to rays, and the corresponding full filter may be fixed once and for all.
Recall that we identify the vertex set of the Davis complex X ( 0 ) X^{(0)} X ( 0 ) and the vertex set of the maximized hyperbolic space X ^ ( 0 ) \widehat{X}^{(0)} X ( 0 ) by the natural inclusion map X → X ^ X \to\widehat{X} X → X . We denote the sphere in ( X ^ , d X ^ ) (\widehat{X},d_{\widehat{X}}) ( X , d X ) with radius R R R based at a vertex o o o by
S R ( o ) = { x ∈ X ^ ∣ d X ^ ( o , x ) = R } . S_R(o) = \{x \in\widehat{X} \mid d_{\widehat{X}}(o,x) = R\}. S R ( o ) = { x ∈ X ∣ d X ( o , x ) = R } .
Fix a full filter F \mathcal{F} F spanning α \alpha α and β \beta β based at o o o . Let γ x \gamma_x γ x and γ y \gamma_y γ y be finite initial subpaths of α \alpha α and β \beta β , respectively, ending at x , y ∈ S R ( o ) x,y \in S_R(o) x , y ∈ S R ( o ) . We define the radius-R R R truncation F R \mathcal{F}^R F R as a subdiagram of F \mathcal{F} F as follows.
Initially, z 0 z_0 z 0 is eligible for extension and its prescribed fan from F \mathcal{F} F is F ( η , a 0 , b 0 ) F(\eta,a_0,b_0) F ( η , a 0 , b 0 ) . Thereafter, every endpoint of a fan edge is eligible for extension . Lemma 8.5 provides its unique next-level fan from the construction of F \mathcal{F} F . We say that an eligible vertex z z z is pending if
z ∉ { x , y } and d X ^ ( o , z ) < R . z \notin\{x,y\} \qquad\text{and} \qquad d_{\widehat{X}}(o,z) < R. z ∈ / { x , y } and d X ( o , z ) < R .
We attach the prescribed fan at every pending eligible vertex at every stage, and do not attach a fan at an eligible vertex whose distance from o o o is at least R R R with respect to the X ^ \widehat{X} X -metric. We also stop at x x x and y y y . Denote the finite-stage diagrams and trees by F N R F_N^R F N R and T N R T_N^R T N R . Proposition 8.8 will show that the construction terminates after finitely many stages.
At each finite stage F N R F_N^R F N R , a part of its boundary opposite the common stem is a nonbacktracking edge path τ N \tau_N τ N from x x x to y y y , called the top boundary . The basic filter properties continue to hold for the truncated diagrams, since they are subdiagrams of the fixed full filter.
Lemma 8.6 (Distance control in a truncated filter). If e e e is a directed edge of T N R T_N^R T N R from vertex u u u to vertex v v v , then
d X ^ ( o , u ) ≤ d X ^ ( o , v ) ≤ d X ^ ( o , u ) + 1. (2) d_{\widehat{X}}(o,u) \le d_{\widehat{X}}(o,v) \le d_{\widehat{X}}(o,u) + 1. \tag*{(2)} d X ( o , u ) ≤ d X ( o , v ) ≤ d X ( o , u ) + 1. (2) If the truncation process has no pending eligible boundary vertices at stage N N N , then
τ N ( 0 ) ⊂ S R ( o ) ∪ S R + 1 ( o ) . \tau_N^{(0)} \subset S_R(o) \cup S_{R+1}(o). τ N ( 0 ) ⊂ S R ( o ) ∪ S R + 1 ( o ) . Proof. The tree path from o o o to v v v is a combinatorial geodesic containing u u u , so u ∈ [ o , v ] u \in[o,v] u ∈ [ o , v ] . Lemma 5.6 gives the first inequality, and the second follows because u u u and v v v are adjacent in X X X , hence also adjacent in X ^ \widehat{X} X .
Suppose that no pending eligible boundary vertex remains. Every eligible vertex p p p of τ N \tau_N τ N is the endpoint of a fan edge based at a pending eligible vertex p ′ p' p ′ . Since d X ^ ( o , p ′ ) ≤ R − 1 d_{\widehat{X}}(o,p') \le R-1 d X ( o , p ′ ) ≤ R − 1 , we have d X ^ ( o , p ) ≤ R d_{\widehat{X}}(o,p) \le R d X ( o , p ) ≤ R . Furthermore, since p p p is not pending, d X ^ ( o , p ) = R d_{\widehat{X}}(o,p)=R d X ( o , p ) = R , so p ∈ S R ( o ) p \in S_R(o) p ∈ S R ( o ) .
Each non-eligible vertex q ∉ { x , y } q \notin\{x,y\} q ∈ / { x , y } of τ N \tau_N τ N is adjacent to an eligible vertex. Hence by the above inequality 2, R ≤ d X ^ ( o , q ) ≤ R + 1 R \le d_{\widehat{X}}(o,q) \le R+1 R ≤ d X ( o , q ) ≤ R + 1 . Therefore, every vertex of τ N \tau_N τ N lies in S R ( o ) ∪ S R + 1 ( o ) S_R(o) \cup S_{R+1}(o) S R ( o ) ∪ S R + 1 ( o ) . □
Another property needed below is that a directed path away from the side rays α \alpha α and β \beta β cannot remain for too long in one standard product region.
Lemma 8.7 (Bounded wide paths). There is a constant C 0 = C 0 ( Γ ) C_0=C_0(\Gamma) C 0 = C 0 ( Γ ) with the following property. Let ζ \zeta ζ be a directed path in the spanning tree T T T . Suppose that, as a path in the abstract diagram, ζ \zeta ζ meets α ∪ β \alpha\cup\beta α ∪ β in at most its initial vertex. If Label ( ζ ) \operatorname{Label}(\zeta) Label ( ζ ) is contained in a nontrivial join, then
∣ ζ ∣ ≤ C 0 . |\zeta| \le C_0. ∣ ζ ∣ ≤ C 0 .
We refer readers to [61 Lemmas 5.8–5.9] for a proof. Cordes and Levcovitz prove the corresponding qualitative boundedness statement for the filters for general Coxeter groups in [25 Proposition 6.19] .
Coarse connectedness of spheres in X ^ \widehat{X} X
We now apply the filter to show the spheres in X ^ \widehat{X} X are coarsely connected. There are two estimates in the proof. First, the bounded-wide-path lemma forces every branch to reach radius R R R after finitely many stages in the construction of truncated filters. Second, a long common stem keeps the entire top boundary in a uniformly bounded subset of X ^ \widehat{X} X .
Recall that the truncated filter F n R F_n^R F n R is constructed as a van Kampen diagram whose image is contained in X X X and we measure the distance under the metric d X ^ d_{\widehat{X}} d X .
Proposition 8.8 (Coarse connectedness of spheres). Fix o ∈ X ( 0 ) = X ^ ( 0 ) o \in X^{(0)}=\widehat{X}^{(0)} o ∈ X ( 0 ) = X ( 0 ) . For every E ≥ 0 E \ge0 E ≥ 0 there is a constant D = D ( Γ , E ) D=D(\Gamma,E) D = D ( Γ , E ) with the following property. If R ∈ N R \in\mathbb{N} R ∈ N and x , y ∈ S R ( o ) x,y \in S_R(o) x , y ∈ S R ( o )
satisfy
d X ^ ( x , y ) ≤ E , d_{\widehat{X}}(x,y) \le E, d X ( x , y ) ≤ E ,
then there is a combinatorial path
τ = ( x = x 0 , x 1 , … , x m = y ) \tau=(x=x_0,x_1,\ldots,x_m=y) τ = ( x = x 0 , x 1 , … , x m = y )
in X ( 1 ) X^{(1)} X ( 1 ) such that
τ ( 0 ) ⊂ S R ( o ) ∪ S R + 1 ( o ) \tau^{(0)} \subset S_R(o)\cup S_{R+1}(o) τ ( 0 ) ⊂ S R ( o ) ∪ S R + 1 ( o )
and
diam X ^ { x 0 , … , x m } ≤ D . \operatorname{diam}_{\widehat{X}}\{x_0,\ldots,x_m\}\le D. diam X { x 0 , … , x m } ≤ D .
One can take D = 4 E + 4 M + 6 D=4E+4M+6 D = 4 E + 4 M + 6 , where M = max { ∣ K ∣ : K ⊂ Γ is a clique } M=\max\{|K|:K\subset\Gamma\text{ is a clique}\} M = max { ∣ K ∣ : K ⊂ Γ is a clique } .
Proof. If x = y x=y x = y , take the constant path. Now we can assume E ∈ N E\in\mathbb{N} E ∈ N and E ≥ 1 E\ge1 E ≥ 1 . If E < R E<R E < R , then d X ^ ( x , y ) < R d_{\widehat{X}}(x,y)<R d X ( x , y ) < R , and Lemma 5.7 gives a vertex z ∈ [ o , x ] ∩ [ o , y ] z\in[o,x]\cap[o,y] z ∈ [ o , x ] ∩ [ o , y ] with d X ^ ( o , z ) ≥ R − d X ^ ( x , y ) ≥ R − E d_{\widehat{X}}(o,z)\ge R-d_{\widehat{X}}(x,y)\ge R-E d X ( o , z ) ≥ R − d X ( x , y ) ≥ R − E . If E ≥ R E\ge R E ≥ R , take z = o z=o z = o . In either case,
d X ^ ( o , z ) ≥ R − E . (3) d_{\widehat{X}}(o,z)\ge R-E. \tag*{(3)} d X ( o , z ) ≥ R − E . (3)
Choose a combinatorial geodesic from o o o to z z z and extend it to combinatorial geodesics γ x \gamma_x γ x and γ y \gamma_y γ y from o o o to x x x and y y y , respectively. Let η \eta η be their maximal common initial subpath, let z 0 z_0 z 0 be its terminal vertex, and let
s = d X ^ ( o , z 0 ) . s=d_{\widehat{X}}(o,z_0). s = d X ( o , z 0 ) .
The common stem contains the chosen path from o o o to z z z . Hence Lemma 5.6 and the inequality (3) give
R − E ≤ d X ^ ( o , z ) ≤ s . (4) R-E\le d_{\widehat{X}}(o,z)\le s. \tag*{(4)} R − E ≤ d X ( o , z ) ≤ s . (4)
Since z 0 ∈ [ o , x ] ∩ [ o , y ] z_0\in[o,x]\cap[o,y] z 0 ∈ [ o , x ] ∩ [ o , y ] , by Lemma 5.6 again, we have s ≤ R s\le R s ≤ R .
If s = R s=R s = R , no filter is needed. Every vertex u u u on either terminal subpath from z 0 z_0 z 0 to x x x or y y y satisfies
R = d X ^ ( o , z 0 ) ≤ d X ^ ( o , u ) ≤ R , R=d_{\widehat{X}}(o,z_0)\le d_{\widehat{X}}(o,u)\le R, R = d X ( o , z 0 ) ≤ d X ( o , u ) ≤ R ,
so the concatenation of these two subpaths lies in S R ( o ) S_R(o) S R ( o ) . Moreover, z 0 ∈ [ o , u ] z_0\in[o,u] z 0 ∈ [ o , u ] . Lemma 5.8 provides a point z u z_u z u on a taut geodesic from o o o to u u u with d X ^ ( z 0 , z u ) ≤ M + 1 d_{\widehat{X}}(z_0,z_u)\le M+1 d X ( z 0 , z u ) ≤ M + 1 . Since both z 0 z_0 z 0 and u u u have distance R R R from o o o , it follows that d X ^ ( u , z 0 ) ≤ 2 ( M + 1 ) d_{\widehat{X}}(u,z_0)\le2(M+1) d X ( u , z 0 ) ≤ 2 ( M + 1 ) . The resulting path has diameter at most 4 ( M + 1 ) 4(M+1) 4 ( M + 1 ) , as required.
Assume from now on that s < R s<R s < R . In particular, both γ x \gamma_x γ x and γ y \gamma_y γ y continue after z 0 z_0 z 0 . Extend them to combinatorial geodesic rays α \alpha α and β \beta β , build the full filter with common stem η \eta η , and construct its radius-R R R truncation, stopping at x x x and $y.
We will show that this truncation terminates after finitely many stages. Let C 0 = C 0 ( Γ ) C_0=C_0(\Gamma) C 0 = C 0 ( Γ ) be the constant from Lemma 8.7 and let C 1 = C 1 ( C 0 , Γ ) C_1=C_1(C_0,\Gamma) C 1 = C 1 ( C 0 , Γ ) be the constant given by Lemma 5.10 . Set
L = ( E + 2 M + 2 ) ( C 1 + 1 ) . L=(E+2M+2)(C_1+1). L = ( E + 2 M + 2 ) ( C 1 + 1 ) .
Let p p p be a pending eligible vertex in a finite stage of the truncated tree. If p p p does not lie on γ x \gamma_x γ x or γ y \gamma_y γ y , let q q q be the last vertex at which the unique tree path from o o o to p p p meets the union of paths γ x ∪ γ y \gamma_x\cup\gamma_y γ x ∪ γ y , and let ζ \zeta ζ be the subpath from q q q to p p p . Every wide subpath of ζ \zeta ζ therefore has length at most C 0 C_0 C 0 , by Lemma 8.7 . The path ζ \zeta ζ maps to a combinatorial geodesic by Lemma 8.5 , so d X ( q , p ) = ∣ ζ ∣ d_X(q,p)=|\zeta| d X ( q , p ) = ∣ ζ ∣ . Hence, by Lemma 5.10 , every wide subpath of every combinatorial geodesic from q q q to p p p has length at most C 1 C_1 C 1 . Suppose that ∣ ζ ∣ ≥ L |\zeta| \ge L ∣ ζ ∣ ≥ L . Lemma 5.11 shows that
d X ^ ( q , p ) ≥ E + 2 ( M + 1 ) . (5) d_{\widehat{X}}(q,p) \ge E + 2(M+1). \tag*{(5)} d X ( q , p ) ≥ E + 2 ( M + 1 ) . (5)
Let t = d X ^ ( o , q ) t=d_{\widehat{X}}(o,q) t = d X ( o , q ) . By the choice of q q q , q q q lies in [ z 0 , x ] [z_0,x] [ z 0 , x ] or [ z 0 , y ] [z_0,y] [ z 0 , y ] . Thus,
R − E ≤ s ≤ t . R-E \le s \le t. R − E ≤ s ≤ t .
It follows from (5) that
d X ^ ( q , p ) ≥ R − t + 2 ( M + 1 ) . d_{\widehat{X}}(q,p) \ge R-t+2(M+1). d X ( q , p ) ≥ R − t + 2 ( M + 1 ) .
The directed tree path from o o o to p p p is a combinatorial geodesic passing through q q q , so q ∈ [ o , p ] q\in[o,p] q ∈ [ o , p ] . Lemma 5.8 gives a point q ′ q' q ′ on a taut geodesic from o o o to p p p such that
d X ^ ( q , q ′ ) ≤ M + 1. d_{\widehat{X}}(q,q') \le M+1. d X ( q , q ′ ) ≤ M + 1.
Consequently,
d X ^ ( q ′ , p ) ≥ R − t + M + 1 and d X ^ ( o , q ′ ) ≥ t − ( M + 1 ) . d_{\widehat{X}}(q',p) \ge R-t+M+1 \qquad\text{and}\qquad d_{\widehat{X}}(o,q') \ge t-(M+1). d X ( q ′ , p ) ≥ R − t + M + 1 and d X ( o , q ′ ) ≥ t − ( M + 1 ) .
Since q ′ q' q ′ lies on a geodesic from o o o to p p p in X ^ \widehat{X} X , these inequalities imply
d X ^ ( o , p ) = d X ^ ( o , q ′ ) + d X ^ ( q ′ , p ) ≥ R . d_{\widehat{X}}(o,p)=d_{\widehat{X}}(o,q')+d_{\widehat{X}}(q',p)\ge R. d X ( o , p ) = d X ( o , q ′ ) + d X ( q ′ , p ) ≥ R .
This contradicts the assumption that p p p is pending. Thus every directed subpath to a pending eligible vertex has length less than L L L after its last intersection with γ x ∪ γ y \gamma_x\cup\gamma_y γ x ∪ γ y . The two paths γ x \gamma_x γ x and γ y \gamma_y γ y are finite, and the abstract spanning tree is locally finite by Lemma 8.5 . The radius-R R R truncated filter F R F^R F R can be obtained within an L L L -neighborhood of γ x ∪ γ y \gamma_x\cup\gamma_y γ x ∪ γ y . Hence only finitely many fans are attached in the process, and the truncation terminates after finitely many steps. Let
τ = ( x = x 0 , x 1 , … , x m = y ) \tau=(x=x_0,x_1,\ldots,x_m=y) τ = ( x = x 0 , x 1 , … , x m = y )
be the terminal top boundary of F R F^R F R . Consecutive vertices are adjacent in X X X (hence adjacent in X ^ \widehat{X} X as well), and Lemma 8.6 gives, for each 0 ≤ i ≤ m 0\le i\le m 0 ≤ i ≤ m ,
x i ∈ S R ( o ) ∪ S R + 1 ( o ) . x_i\in S_R(o)\cup S_{R+1}(o). x i ∈ S R ( o ) ∪ S R + 1 ( o ) .
It remains to control the diameter. The unique tree path from o o o to each x i x_i x i contains the common stem η \eta η , and hence z 0 ∈ [ o , x i ] z_0\in[o,x_i] z 0 ∈ [ o , x i ] for each i = 0 , … , m i=0,\ldots,m i = 0 , … , m . Lemma 5.8 gives a point z i z_i z i on a taut geodesic from o o o to x i x_i x i such that
d X ^ ( z 0 , z i ) ≤ M + 1. d_{\widehat{X}}(z_0,z_i)\le M+1. d X ( z 0 , z i ) ≤ M + 1.
By (4) ,
d X ^ ( o , z i ) ≥ s − ( M + 1 ) ≥ R − E − ( M + 1 ) . d_{\widehat{X}}(o,z_i)\ge s-(M+1)\ge R-E-(M+1). d X ( o , z i ) ≥ s − ( M + 1 ) ≥ R − E − ( M + 1 ) .
Because d X ^ ( o , x i ) ≤ R + 1 d_{\widehat{X}}(o,x_i)\le R+1 d X ( o , x i ) ≤ R + 1 and z i z_i z i lies on a geodesic in X ^ \widehat{X} X from o o o to x i x_i x i , we obtain
d X ^ ( x i , z i ) ≤ E + M + 2. d_{\widehat{X}}(x_i,z_i)\le E+M+2. d X ( x i , z i ) ≤ E + M + 2.
Therefore
d X ^ ( x i , z 0 ) ≤ E + 2 M + 3 d_{\widehat{X}}(x_i,z_0)\le E+2M+3 d X ( x i , z 0 ) ≤ E + 2 M + 3
and, for all i , j i,j i , j ,
d X ^ ( x i , x j ) ≤ 4 E + 4 M + 6. d_{\widehat{X}}(x_i,x_j)\le4E+4M+6. d X ( x i , x j ) ≤ 4 E + 4 M + 6.
We conclude this section by a corollary about the number of ends of X ^ \widehat{X} X . We say a connected, unbounded graph is one-ended if the complement of every bounded subset has exactly one unbounded connected component.
Corollary 8.9 (One-endedness of X ^ \widehat X X ). Suppose Γ \Gamma Γ satisfies conditions (a), (b) and (c). Then the maximized hyperbolic space X ^ Γ \widehat X_\Gamma X Γ is one-ended.
Proof. By conditions (a) and (c), we know that Γ \Gamma Γ is neither a clique nor a nontrivial join. Thus X ^ Γ \widehat X_\Gamma X Γ has infinite diameter by Lemma 3.4 .
Fix o ∈ X ^ Γ o \in\widehat X_\Gamma o ∈ X Γ and N ∈ N N \in\mathbb{N} N ∈ N . Let u , v ∉ B N ( o ) u,v \notin B_N(o) u , v ∈ / B N ( o ) . Choose geodesics in X ^ Γ \widehat X_\Gamma X Γ from o o o to u u u and v v v , and let u ′ , v ′ u',v' u ′ , v ′ be the vertices in the intersection of chosen geodesics with S N + 1 ( o ) S_{N+1}(o) S N + 1 ( o ) , respectively. The terminal geodesic subpaths from u ′ u' u ′ to u u u and from v ′ v' v ′ to v v v avoid B N ( o ) B_N(o) B N ( o ) . Applying Proposition 8.8 with R = N + 1 R=N+1 R = N + 1 and E = d X ^ ( u ′ , v ′ ) E=d_{\widehat X}(u',v') E = d X ( u ′ , v ′ ) gives an edge path in
S N + 1 ( o ) ∪ S N + 2 ( o ) S_{N+1}(o) \cup S_{N+2}(o) S N + 1 ( o ) ∪ S N + 2 ( o )
joining u ′ u' u ′ to v ′ v' v ′ . Concatenating the above three paths connects u u u to v v v in X ^ ∖ B N ( o ) \widehat X \setminus B_N(o) X ∖ B N ( o ) . Hence the complement of every bounded ball based at o o o is connected. Since W Γ W_\Gamma W Γ acts on X ^ Γ \widehat X_\Gamma X Γ transitively, the complement of every bounded ball in X ^ Γ \widehat X_\Gamma X Γ is connected.
Every bounded subset is contained in such a ball. Therefore its complement has exactly one unbounded component, so X ^ Γ \widehat X_\Gamma X Γ is one-ended. □
Linear Connectedness of the Gromov boundary
We now combine the results of the preceding sections. The obstruction established in Proposition 3.8 proves one direction of Theorem 1.1 . For the other direction, we use the controlled connectivity of spheres from Proposition 8.8 to construct a path with controlled diameter joining a pair of points in the boundary. The construction proceeds by repeatedly inserting new boundary points between two previously chosen ones. At each stage the new points are represented by geodesic rays that fellow travel farther from the basepoint, so the diameters of the inserted boundary chains decrease geometrically.
The proof is inspired by [Wri24, Proposition 2.2], [15 Proposition 3.2] and [62 Proposition 5.2] .
Proof of Theorem 1.1. Suppose first that one of conditions (a)–(c) fails. Proposition 3.8 shows that ∂ X ^ \partial\widehat X ∂ X is either empty or disconnected.
Conversely, assume that Γ \Gamma Γ satisfies conditions (a)–(c). We will prove that ∂ X ^ \partial\widehat X ∂ X , equipped with a visual metric d ∞ d_\infty d ∞ , is linearly connected. It is sufficient to show the linear connectivity since linear connectivity implies path connectivity, hence connectivity. The Gromov boundary may admit many visual metrics, but they are quasisymmetric to each other and linear connectivity is preserved under a quasisymmetry. Hence, it is enough to show the linear connectivity for the chosen visual metric d ∞ d_\infty d ∞ .
Condition (a) says that Γ \Gamma Γ is not a clique. Moreover, Γ \Gamma Γ cannot itself be a nontrivial join, since otherwise lk ( v ) \operatorname{lk}(v) lk ( v ) would be contained in the nontrivial join Γ \Gamma Γ for every vertex v v v , contrary to condition (c). Lemma 3.4 therefore shows that ∂ X ^ \partial\widehat X ∂ X is nonempty and contains at least two points.
Let p , q ∈ ∂ X ^ p,q \in\partial\widehat X p , q ∈ ∂ X . Fix a vertex o ∈ X ^ o \in\widehat X o ∈ X . We will construct a continuous map γ : [ 0 , 1 ] → ∂ X ^ \gamma\colon[0,1] \to\partial\widehat X γ : [ 0 , 1 ] → ∂ X such that γ ( 0 ) = p \gamma(0)=p γ ( 0 ) = p , γ ( 1 ) = q \gamma(1)=q γ ( 1 ) = q and diam ( γ ( [ 0 , 1 ] ) ) \operatorname{diam}(\gamma([0,1])) diam ( γ ([ 0 , 1 ])) is bounded above by a constant multiple of d ∞ ( p , q ) d_\infty(p,q) d ∞ ( p , q ) , where the constant is independent of p p p and q q q . See Figure 10 for an illustration. The key idea is to construct a chain of sequences of points in the boundary.
Figure 10. At each step, two points on the previously chosen rays are connected by an edge path on the spheres in X ^ \hat{X} X ^ , provided by Proposition 8.8 . The path in red is the edge path in the base step and the path in blue is an edge path chosen in the next step.
By Lemma 7.4 , there are geodesic rays α \alpha α and β \beta β based at o o o and converging to p p p and q q q , respectively. Fix b > 1 b>1 b > 1 and assume the visual metric d ∞ d_\infty d ∞ is as in Proposition 2.20 . The usual fellow-traveling estimate for geodesic rays in a hyperbolic space gives [Vä05, DSU17] gives constants E 0 = E 0 ( δ ) ≥ 2 E_0=E_0(\delta)\ge2 E 0 = E 0 ( δ ) ≥ 2 and A 0 ≥ 1 A_0\ge1 A 0 ≥ 1 , depending only on δ \delta δ and the chosen visual metric, with the following property. Let R R R be the largest integer such that
d X ^ ( α ( R ) , β ( R ) ) ≤ E 0 . d_{\widehat X}\bigl(\alpha(R),\beta(R)\bigr)\le E_0. d X ( α ( R ) , β ( R ) ) ≤ E 0 .
Then R R R differs from the Gromov product ⟨ p , q ⟩ o \langle p,q\rangle_o ⟨ p , q ⟩ o by a uniformly bounded constant, and hence
A 0 − 1 b − R ≤ d ∞ ( p , q ) ≤ A 0 b − R . (6) A_0^{-1}b^{-R}\le d_\infty(p,q)\le A_0b^{-R}. \tag*{(6)} A 0 − 1 b − R ≤ d ∞ ( p , q ) ≤ A 0 b − R . (6)
The integer R R R exists because the two geodesic rays represent distinct points in the boundary.
We next fix the constants that will be used in the construction. Let C C C be the geodesic almost extendability constant from Proposition 7.7. We may assume that C C C is an integer. Let
K = 6 C + 3 , E 1 = K + 2 C + 2. K=6C+3,\qquad E_1=K+2C+2. K = 6 C + 3 , E 1 = K + 2 C + 2.
For i = 0 , 1 i=0,1 i = 0 , 1 , let D i = D ( Γ , E i ) D_i=D(\Gamma,E_i) D i = D ( Γ , E i ) be the constant supplied by Proposition 8.8 .
We will use the following standard consequence of hyperbolicity and the property of a visual metric. Suppose that ρ \rho ρ and ρ ′ \rho' ρ ′ are geodesic rays based at o o o , and that they contain points, each at distance at least T T T from o o o , whose distance is at most L L L . Then
d ∞ ( ρ ( ∞ ) , ρ ′ ( ∞ ) ) ≤ A ( L ) b − T . (7) d_\infty\bigl(\rho(\infty),\rho'(\infty)\bigr)\le A(L)b^{-T}. \tag*{(7)} d ∞ ( ρ ( ∞ ) , ρ ′ ( ∞ ) ) ≤ A ( L ) b − T . (7)
where A ( L ) A(L) A ( L ) is a constant depending only on L L L , δ \delta δ , and the chosen visual metric d ∞ d_\infty d ∞ . Indeed, the Gromov product of the two endpoints is at least T − O ( L + δ ) T-O(L+\delta) T − O ( L + δ ) , and (7) follows from Proposition 2.20 .
We now recursively construct increasing finite sets
I 0 ⊂ I 1 ⊂ ⋯ ⊂ [ 0 , 1 ] I_0\subset I_1\subset\cdots\subset[0,1] I 0 ⊂ I 1 ⊂ ⋯ ⊂ [ 0 , 1 ]
and associated maps
γ r : I r ⟶ ∂ X ^ . \gamma_r:I_r\longrightarrow\partial\widehat X. γ r : I r ⟶ ∂ X .
For each a ∈ I r a\in I_r a ∈ I r , we choose a particular geodesic ray ρ a \rho_a ρ a based at o o o and satisfying γ r ( a ) = ρ a ( ∞ ) \gamma_r(a)=\rho_a(\infty) γ r ( a ) = ρ a ( ∞ ) . Figure 10 illustrates one step of the construction.
The base step. Apply Proposition 8.8 to α ( R ) \alpha(R) α ( R ) and β ( R ) \beta(R) β ( R ) . We obtain an edge path
α ( R ) = x 0 , x 1 , … , x m = β ( R ) \alpha(R)=x_0,x_1,\ldots,x_m=\beta(R) α ( R ) = x 0 , x 1 , … , x m = β ( R )
in X ( 1 ) X^{(1)} X ( 1 ) such that
x i ∈ S R ( o ) ∪ S R + 1 ( o ) for every i , x_i\in S_R(o)\cup S_{R+1}(o)\quad\text{for every }i, x i ∈ S R ( o ) ∪ S R + 1 ( o ) for every i ,
and d X ^ ( x i , x j ) ≤ D 0 d_{\widehat X}(x_i,x_j)\le D_0 d X ( x i , x j ) ≤ D 0 for each 0 ≤ i , j ≤ m 0\le i,j\le m 0 ≤ i , j ≤ m . If m ≥ 2 m\ge2 m ≥ 2 , let ρ 0 = α \rho_0=\alpha ρ 0 = α and ρ m = β \rho_m=\beta ρ m = β . For each 0 < i < m 0<i<m 0 < i < m , Proposition 7.7 provides a geodesic ray ρ i \rho_i ρ i based at o o o that passes within C C C of x i x_i x i . Define
I 0 = { i / m : 0 ≤ i ≤ m } , γ 0 ( i / m ) = ρ i ( ∞ ) . I_0=\{i/m:0\le i\le m\},\qquad\gamma_0(i/m)=\rho_i(\infty). I 0 = { i / m : 0 ≤ i ≤ m } , γ 0 ( i / m ) = ρ i ( ∞ ) .
If m ≤ 1 m\le1 m ≤ 1 , then we just let I 0 = { 0 , 1 / 2 , 1 } I_0=\{0,1/2,1\} I 0 = { 0 , 1/2 , 1 } , and
ρ 0 = ρ 1 / 2 = α , ρ 1 = β , γ 0 ( a ) = ρ a ( ∞ ) for each a ∈ I 0 . \rho_0=\rho_{1/2}=\alpha,\qquad\rho_1=\beta,\qquad\gamma_0(a)=\rho_a(\infty)\quad\text{for each }a\in I_0. ρ 0 = ρ 1/2 = α , ρ 1 = β , γ 0 ( a ) = ρ a ( ∞ ) for each a ∈ I 0 .
Let
T 0 = max { 0 , R − C } . T_0=\max\{0,R-C\}. T 0 = max { 0 , R − C } .
We claim that, whenever a , a ′ ∈ I 0 a,a'\in I_0 a , a ′ ∈ I 0 are consecutive, their assigned rays satisfy
d X ^ ( ρ a ( T 0 ) , ρ a ′ ( T 0 ) ) ≤ K . (8) d_{\widehat X}\bigl(\rho_a(T_0),\rho_{a'}(T_0)\bigr)\le K. \tag*{(8)} d X ( ρ a ( T 0 ) , ρ a ′ ( T 0 ) ) ≤ K . (8)
To see this when R ≥ C R \ge C R ≥ C , choose on each ray a point, say y i y_i y i , within C C C of its corresponding x i x_i x i . Then
R − C ≤ d X ^ ( o , y i ) ≤ R + C + 1. R-C \le d_{\hat{X}}(o,y_i) \le R+C+1. R − C ≤ d X ^ ( o , y i ) ≤ R + C + 1.
Since x i x_i x i and x i + 1 x_{i+1} x i + 1 are adjacent,
d X ^ ( ρ a ( T 0 ) , ρ a ′ ( T 0 ) ) ≤ 2 ( ( 2 C + 1 ) + C ) + 1 = K , d_{\hat{X}}\bigl(\rho_a(T_0),\rho_{a'}(T_0)\bigr) \le2\bigl((2C+1)+C\bigr)+1=K, d X ^ ( ρ a ( T 0 ) , ρ a ′ ( T 0 ) ) ≤ 2 ( ( 2 C + 1 ) + C ) + 1 = K ,
for each consecutive pair a , a ′ ∈ I 0 a,a' \in I_0 a , a ′ ∈ I 0 . When R < C R<C R < C , both rays begin at o o o and the estimate at time T 0 = 0 T_0=0 T 0 = 0 is immediate.
The points x i x_i x i have diameter at most D 0 D_0 D 0 . Hence, for 0 ≤ i , j ≤ m 0\le i,j\le m 0 ≤ i , j ≤ m , the distance between y i y_i y i and y j y_j y j is at most D 0 + 2 C D_0+2C D 0 + 2 C , and both points lie at distance at least R − C R-C R − C from o o o . It follows from (7) that, for some constant A 1 A_1 A 1 (depending only on D 0 D_0 D 0 and C C C ),
diam ( γ 0 ( I 0 ) ) ≤ A 1 b − R . (9) \operatorname{diam}\bigl(\gamma_0(I_0)\bigr)\le A_1b^{-R}. \tag*{(9)} diam ( γ 0 ( I 0 ) ) ≤ A 1 b − R . (9)
The inductive step. Suppose that I r I_r I r , γ r \gamma_r γ r , and the rays ρ a \rho_a ρ a , for a ∈ I r a\in I_r a ∈ I r , are defined so that (8) holds for every pair of consecutive points of I r I_r I r at time
T r = T 0 + r . T_r=T_0+r. T r = T 0 + r .
We will construct the finite set I r + 1 I_{r+1} I r + 1 and the corresponding map γ r + 1 \gamma_{r+1} γ r + 1 . Let ( s , t ) (s,t) ( s , t ) be a component of [ 0 , 1 ] ∖ I r [0,1]\setminus I_r [ 0 , 1 ] ∖ I r . Its endpoints belong to I r I_r I r . Let
L r = T r + C + 1 , x = ρ s ( L r ) , y = ρ t ( L r ) . L_r=T_r+C+1,\qquad x=\rho_s(L_r),\qquad y=\rho_t(L_r). L r = T r + C + 1 , x = ρ s ( L r ) , y = ρ t ( L r ) .
Hence,
d X ^ ( x , y ) ≤ K + 2 C + 2 = E 1 . d_{\widehat{X}}(x,y)\le K+2C+2=E_1. d X ( x , y ) ≤ K + 2 C + 2 = E 1 .
We apply Proposition 8.8 again to obtain an edge path
x = z 0 , z 1 , … , z n = y x=z_0,z_1,\ldots,z_n=y x = z 0 , z 1 , … , z n = y
in S L r ( o ) ∪ S L r + 1 ( o ) S_{L_r}(o)\cup S_{L_{r+1}}(o) S L r ( o ) ∪ S L r + 1 ( o ) whose diameter in X ^ \widehat{X} X is at most D 1 D_1 D 1 . For each 0 < i < n 0<i<n 0 < i < n , choose a geodesic ray based at o o o that passes within C C C of z i z_i z i . If n ≥ 2 n\ge2 n ≥ 2 , subdivide [ s , t ] [s,t] [ s , t ] into n n n subintervals equally and assign these rays to the new subdivision points in the order determined by the path. If n ≤ 1 n\le1 n ≤ 1 , insert the midpoint of [ s , t ] [s,t] [ s , t ] and assign to it the ray ρ s \rho_s ρ s .
We verify that the induction invariant persists. A point on a newly chosen ray lying within C C C of some z i z_i z i has distance from o o o between
T r + 1 and T r + 2 C + 2. T_r+1\quad\text{and}\quad T_r+2C+2. T r + 1 and T r + 2 C + 2.
For consecutive vertices z i , z i + 1 z_i,z_{i+1} z i , z i + 1 , the corresponding nearby ray points are at distance at most 2 C + 1 2C+1 2 C + 1 . Note that T r + 1 = T r + 1 T_{r+1}=T_r+1 T r + 1 = T r + 1 . Then we have
d X ^ ( ρ a ( T r + 1 ) , ρ a ′ ( T r + 1 ) ) ≤ 2 ( 2 C + 1 ) + ( 2 C + 1 ) = K , d_{\widehat{X}}\bigl(\rho_a(T_{r+1}),\rho_{a'}(T_{r+1})\bigr)\le2(2C+1)+(2C+1)=K, d X ( ρ a ( T r + 1 ) , ρ a ′ ( T r + 1 ) ) ≤ 2 ( 2 C + 1 ) + ( 2 C + 1 ) = K ,
where a , a ′ a,a' a , a ′ are consecutive points of a component of [ 0 , 1 ] ∖ I r [0,1]\setminus I_r [ 0 , 1 ] ∖ I r . When n ≤ 1 n\le1 n ≤ 1 , the inserted ray is ρ s \rho_s ρ s , and the estimate follows directly from d X ^ ( x , y ) ≤ 1 d_{\widehat{X}}(x,y)\le1 d X ( x , y ) ≤ 1 at time L r L_r L r . This proves (8) at step r + 1 r+1 r + 1 .
The points z i z_i z i have diameter at most D 1 D_1 D 1 , and their nearby ray points lie at distance at least T r + 1 T_{r+1} T r + 1 from o o o . Consequently, (7) gives a constant A 2 A_2 A 2 , independent of r , s , t r,s,t r , s , t , such that
diam { γ r + 1 ( u ) ∣ u ∈ I r + 1 ∩ [ s , t ] } ≤ A 2 b − T r + 1 . (10) \operatorname{diam}\{\gamma_{r+1}(u)\mid u\in I_{r+1}\cap[s,t]\}\le A_2b^{-T_{r+1}}. \tag*{(10)} diam { γ r + 1 ( u ) ∣ u ∈ I r + 1 ∩ [ s , t ]} ≤ A 2 b − T r + 1 . (10)
Extension of the map. Let
I = ⋃ r ≥ 0 I r . I=\bigcup_{r\ge0} I_r. I = r ≥ 0 ⋃ I r .
Every component of [ 0 , 1 ] ∖ I r [0,1]\setminus I_r [ 0 , 1 ] ∖ I r is divided into at least two subintervals at the next step. Thus the largest length of a complementary interval converges to 0 0 0 ; hence I I I is dense in [ 0 , 1 ] [0,1] [ 0 , 1 ] . We define the map
γ : I ⟶ ∂ X ^ \gamma:I\longrightarrow\partial\widehat{X} γ : I ⟶ ∂ X
by γ ( a ) = γ r ( a ) \gamma(a)=\gamma_r(a) γ ( a ) = γ r ( a ) if a ∈ I r a\in I_r a ∈ I r .
For a component ( s , t ) (s,t) ( s , t ) of [ 0 , 1 ] ∖ I r [0,1]\setminus I_r [ 0 , 1 ] ∖ I r , by applying (10) iteratively, we have
diam { γ ( u ) ∣ u ∈ I ∩ [ s , t ] } ≤ 2 ∑ k ≥ r + 1 A 2 b − T k . \operatorname{diam}\{\gamma(u)\mid u\in I\cap[s,t]\}\le2\sum_{k\ge r+1}A_2b^{-T_k}. diam { γ ( u ) ∣ u ∈ I ∩ [ s , t ]} ≤ 2 k ≥ r + 1 ∑ A 2 b − T k .
Denote the series on the right-hand side by Q r Q_r Q r . Then Q r → 0 Q_r\to0 Q r → 0 as r → ∞ r\to\infty r → ∞ since it is a geometric series with common ratio less than 1 1 1 .
We claim that γ \gamma γ is uniformly continuous.
Given ε > 0 \varepsilon>0 ε > 0 , first choose r r r so that Q r < ε / 2 Q_r<\varepsilon/2 Q r < ε /2 , and then choose δ > 0 \delta>0 δ > 0 so that every interval of length δ \delta δ contains at most one point of I r I_r I r . This is possible since I r I_r I r is a finite subset of [ 0 , 1 ] [0,1] [ 0 , 1 ] . If v , w ∈ I v,w\in I v , w ∈ I and ∣ v − w ∣ < δ |v-w|<\delta ∣ v − w ∣ < δ , then they lie either in the closure of one component of [ 0 , 1 ] ∖ I r [0,1]\setminus I_r [ 0 , 1 ] ∖ I r , or in the closures of two components having a common endpoint in I r I_r I r . In the first case their images in ∂ X ^ \partial\widehat{X} ∂ X are at distance at most Q r Q_r Q r . In the second case, their images are at distance at most 2 Q r 2Q_r 2 Q r . Thus d ∞ ( γ ( v ) , γ ( w ) ) < ε d_\infty(\gamma(v),\gamma(w)) < \varepsilon d ∞ ( γ ( v ) , γ ( w )) < ε .
Moreover,
diam ( γ ( I ) ) ≤ A 1 b − R + 2 ∑ r ≥ 1 A 2 b − ( T 0 + r ) ≤ A 3 b − R \begin{aligned}
\operatorname{diam}(\gamma(I)) \le A_1 b^{-R} + 2 \sum_{r \ge1} A_2 b^{-(T_0+r)} \\
&\le A_3 b^{-R}
\end{aligned} diam ( γ ( I )) ≤ A 1 b − R + 2 r ≥ 1 ∑ A 2 b − ( T 0 + r ) ≤ A 3 b − R
for a uniform constant A 3 A_3 A 3 (recall that T 0 = max { 0 , R − C } T_0 = \max\{0,R-C\} T 0 = max { 0 , R − C } ).
Note that the Gromov boundary ( ∂ X ^ , d ∞ ) (\partial\widehat{X},d_\infty) ( ∂ X , d ∞ ) is a complete metric space. Since I I I is dense in [ 0 , 1 ] [0,1] [ 0 , 1 ] , the uniformly continuous map γ \gamma γ extends to a unique uniformly continuous map
γ ‾ : [ 0 , 1 ] ⟶ ∂ X ^ . \overline{\gamma}\colon[0,1] \longrightarrow\partial\widehat{X}. γ : [ 0 , 1 ] ⟶ ∂ X .
Its image is a compact connected set containing p p p and q q q . Finally, (6) gives
diam ( γ ‾ ( [ 0 , 1 ] ) ) ≤ A 3 b − R ≤ A 0 A 3 d ∞ ( p , q ) , \operatorname{diam}(\overline{\gamma}([0,1])) \le A_3b^{-R} \le A_0A_3d_\infty(p,q), diam ( γ ([ 0 , 1 ])) ≤ A 3 b − R ≤ A 0 A 3 d ∞ ( p , q ) ,
where the constant A 0 A 3 A_0A_3 A 0 A 3 is independent of p p p and q q q . Therefore, ∂ X ^ \partial\widehat{X} ∂ X is linearly connected. □
Application to group splittings
As an application of Theorem 1.1 , in this section, we relate the connectivity of ∂ X ^ Γ \partial\widehat{X}_\Gamma ∂ X Γ to a group splitting of W Γ W_\Gamma W Γ . The strategy is the following. The equivalence between the connectivity of ∂ X ^ \partial\widehat{X} ∂ X and a graph-theoretic condition on Γ \Gamma Γ is established in Theorem 1.1 . The equivalence between the algebraic splitting of W Γ W_\Gamma W Γ and the same graph-theoretic condition will be obtained by using the visual decomposition theorem from [63 ] .
We say a group G G G splits over a subgroup H ≤ G H \le G H ≤ G if G G G splits over H H H as a free product with amalgamation or an HNN extension, that is, there are proper subgroups A A A and B B B of G G G with A ≠ G ≠ B A \ne G \ne B A = G = B such that
G ≅ A ∗ H B or G ≅ A ∗ H . G \cong A *_H B \qquad\text{or} \qquad G \cong A *_H . G ≅ A ∗ H B or G ≅ A ∗ H .
We note that every group which splits as an HNN extension has a surjective homomorphism to Z \mathbb{Z} Z by sending the stable element to 1 1 1 . But there is no surjective homomorphism from a Coxeter group to Z \mathbb{Z} Z , since the abelianization of a Coxeter group is the finite abelian group ( Z / 2 Z ) ∣ V ( Γ ) ∣ (\mathbb{Z}/2\mathbb{Z})^{|V(\Gamma)|} ( Z /2 Z ) ∣ V ( Γ ) ∣ . Therefore, no Coxeter group splits as an HNN extension.
The main ingredient of this section is the following proposition where we show that a group splitting over a subgroup of an unbounded product subgroup forces one of the two graph-theoretic obstructions in Theorem 1.1 . The visual decomposition theorem of Mihalik–Tschantz provides the bridge from the arbitrary splitting to a visual splitting on the defining graph.
Proposition 10.1. If a one-ended right-angled Coxeter group W Γ W_\Gamma W Γ splits over a subgroup of an unbounded product subgroup, then either Γ \Gamma Γ has a separating nontrivial join or lk ( v ) \operatorname{lk}(v) lk ( v ) is contained in a nontrivial join for some v ∈ V ( Γ ) v \in V(\Gamma) v ∈ V ( Γ ) .
Proof. Suppose C C C is a subgroup of an unbounded product subgroup H = H 0 × H 1 H=H_0 \times H_1 H = H 0 × H 1 of W Γ W_\Gamma W Γ such that both factors H 0 H_0 H 0 and H 1 H_1 H 1 are infinite and W Γ W_\Gamma W Γ splits over C C C as an amalgamation W Γ ≅ A ∗ C B W_\Gamma\cong A *_C B W Γ ≅ A ∗ C B , where A A A and B B B are subgroups of W Γ W_\Gamma W Γ . Since W Γ W_\Gamma W Γ is one-ended, Stallings’ theorem implies that C C C is infinite [71 ] .
Convert to a visual decomposition. The visual decomposition theorem [63 ] , Theorem 1 and Remark 1 gives a finite reduced visual graph-of-groups decomposition Ψ \Psi Ψ of W Γ W_\Gamma W Γ with the following properties. Every vertex group lies in a conjugate of A A A or B B B , and every edge group lies in a conjugate of C C C . The underlying graph of Ψ \Psi Ψ is a tree, since otherwise W Γ W_{\Gamma} W Γ would be an HNN extension. Moreover, there is an edge e e e such that neither of the two components Ψ 1 \Psi_1 Ψ 1 and Ψ 2 \Psi_2 Ψ 2 of Ψ ∖ e \Psi\setminus e Ψ ∖ e contains all the standard generators among its vertex groups.
For i ∈ { 1 , 2 } i \in\{1,2\} i ∈ { 1 , 2 } , let
S i = { s ∈ V ( Γ ) : s belongs to a vertex group over Ψ i } , S_i=\{s\in V(\Gamma):s\text{ belongs to a vertex group over }\Psi_i\}, S i = { s ∈ V ( Γ ) : s belongs to a vertex group over Ψ i } ,
and let Γ i \Gamma_i Γ i be the induced subgraph of Γ \Gamma Γ spanned by S i S_i S i . For each standard generator s ∈ V ( Γ ) s \in V(\Gamma) s ∈ V ( Γ ) , the vertices and edges of Ψ \Psi Ψ whose groups contain s s s form a nonempty subtree by [63 Lemma 4] . It follows that
S 1 ∪ S 2 = V ( Γ ) . S_1\cup S_2=V(\Gamma). S 1 ∪ S 2 = V ( Γ ) .
A generator belongs to both S 1 S_1 S 1 and S 2 S_2 S 2 precisely when it belongs to the special edge group Ψ ( e ) \Psi(e) Ψ ( e ) . Let Λ \Lambda Λ be the induced subgraph spanned by S 1 ∩ S 2 S_1\cap S_2 S 1 ∩ S 2 . Then
Ψ ( e ) = W Λ , \Psi(e)=W_{\Lambda}, Ψ ( e ) = W Λ ,
and neither S 1 ∖ V ( Λ ) S_1\setminus V(\Lambda) S 1 ∖ V ( Λ ) nor S 2 ∖ V ( Λ ) S_2\setminus V(\Lambda) S 2 ∖ V ( Λ ) is empty.
Mihalik–Tschantz’s separator property [63 Corollary 5] implies that Λ \Lambda Λ separates Γ \Gamma Γ . In particular, there is no edge between Γ 1 ∖ Λ \Gamma_1\setminus\Lambda Γ 1 ∖ Λ and Γ 2 ∖ Λ \Gamma_2\setminus\Lambda Γ 2 ∖ Λ , and Γ = Γ 1 ∪ Γ 2 \Gamma=\Gamma_1\cup\Gamma_2 Γ = Γ 1 ∪ Γ 2 .
Collapsing every edge of Ψ \Psi Ψ other than e e e gives a nontrivial splitting of W Γ W_{\Gamma} W Γ over W Λ W_{\Lambda} W Λ . Since W Γ W_{\Gamma} W Γ is one-ended, W Λ W_{\Lambda} W Λ cannot be finite. Thus Λ \Lambda Λ is not a clique.
The visual edge group is contained in a standard product region. We claim that Λ \Lambda Λ is contained in a nontrivial join. Suppose not. Lemma 3.2 provides a contracting element g ∈ W Λ g\in W_{\Lambda} g ∈ W Λ . Then a conjugate
h = a g a − 1 ∈ C ≤ H h=aga^{-1}\in C\le H h = a g a − 1 ∈ C ≤ H
would be a contracting element in the unbounded product subgroup H H H . Then the centralizer of a finite power of h h h would contain a Z 2 \mathbb{Z}^2 Z 2 subgroup. By [8 Theorem 5.4] , an element h ∈ W Γ h\in W_{\Gamma} h ∈ W Γ is contracting if and only if it is rank-one. A finite power of a rank-one element is still rank-one. And by [CF10, Corollary 4.6(ii)], the centralizer of a rank-one element is virtually cyclic. Contradiction.
Hence Λ ⊂ J \Lambda\subset J Λ ⊂ J for some nontrivial join subgraph J J J of Γ \Gamma Γ .
Graph-theoretic conditions. If J J J separates Γ \Gamma Γ , then we are done. Suppose that J J J does not separate Γ \Gamma Γ . Then Γ ∖ J \Gamma\setminus J Γ ∖ J is empty or connected. If Γ ∖ J \Gamma\setminus J Γ ∖ J is empty, then Γ = J \Gamma=J Γ = J is a nontrivial join and for each vertex v ∈ V ( Γ ) v\in V(\Gamma) v ∈ V ( Γ ) , the link lk ( v ) \operatorname{lk}(v) lk ( v ) is contained in the nontrivial join Γ = J \Gamma=J Γ = J .
Now suppose Γ ∖ J \Gamma\setminus J Γ ∖ J is nonempty and connected. The subgraph Λ \Lambda Λ separates Γ \Gamma Γ , so there is a component U U U of Γ ∖ Λ \Gamma\setminus\Lambda Γ ∖ Λ contained in J J J . In particular,
U ⊂ J ∖ Λ . U\subset J\setminus\Lambda. U ⊂ J ∖ Λ.
Choose a vertex u ∈ V ( U ) u\in V(U) u ∈ V ( U ) . Every neighbor of u u u belongs either to U U U or to Λ \Lambda Λ , because U U U is a component of Γ ∖ Λ \Gamma\setminus\Lambda Γ ∖ Λ . Consequently,
lk ( u ) ⊂ U ∪ Λ ⊂ J . \operatorname{lk}(u)\subset U\cup\Lambda\subset J. lk ( u ) ⊂ U ∪ Λ ⊂ J .
We can now characterize these splittings through the connectivity of ∂ X ^ \partial\widehat{X} ∂ X .
Theorem 10.2. Let W Γ W_{\Gamma} W Γ be a one-ended right-angled Coxeter group. The following are equivalent:
W Γ W_{\Gamma} W Γ splits over a subgroup of an unbounded product subgroup;
(2) Γ \Gamma Γ has a separating nontrivial join, or lk ( v ) \operatorname{lk}(v) lk ( v ) is contained in a nontrivial join for some v ∈ V ( Γ ) v \in V(\Gamma) v ∈ V ( Γ ) ;
(3) ∂ X ^ Γ \partial\widehat{X}_{\Gamma} ∂ X Γ is empty or disconnected.
Proof. For one-ended RACGs, condition (a) of Theorem 1.1 holds. Hence, the equivalence of (2) and (3) is given by Theorem 1.1 . (1) implies (2) by Proposition 10.1 . It remains to prove that (2) implies (1). A separating nontrivial join gives the required visual splitting immediately.
Now suppose lk ( v ) \operatorname{lk}(v) lk ( v ) is contained in a nontrivial join. If Γ ∖ st ( v ) \Gamma\setminus\operatorname{st}(v) Γ ∖ st ( v ) is nonempty, then lk ( v ) \operatorname{lk}(v) lk ( v ) separates v v v from that set, and W Γ W_{\Gamma} W Γ splits over W lk ( v ) W_{\operatorname{lk}(v)} W lk ( v ) . Otherwise, V ( Γ ) = { v } ∪ lk ( v ) V(\Gamma)=\{v\}\cup\operatorname{lk}(v) V ( Γ ) = { v } ∪ lk ( v ) and Γ \Gamma Γ is a nontrivial join. Because W Γ W_{\Gamma} W Γ is one-ended, Γ \Gamma Γ is not a clique, so lk ( v ) \operatorname{lk}(v) lk ( v ) contains non-adjacent vertices x , y x,y x , y . Let Δ = Γ ∖ { x , y } \Delta=\Gamma\setminus\{x,y\} Δ = Γ ∖ { x , y } . Then Δ \Delta Δ lies in a nontrivial join and separates Γ \Gamma Γ , hence
W Γ ≅ W Δ ∪ { x } ∗ W Δ W Δ ∪ { y } . W_{\Gamma}\cong W_{\Delta\cup\{x\}}*_{W_{\Delta}}W_{\Delta\cup\{y\}}. W Γ ≅ W Δ ∪ { x } ∗ W Δ W Δ ∪ { y } .
References
Department of Mathematics, The Graduate Center of City University of New York, 365 Fifth Ave, New York, NY, USA, 10016
Email address: zmu@gc.cuny.edu