AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS

Transcription

1 AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN Abstract. Associated to any simplicial graph Γ is a right-angled Artin group A Γ. This class of groups includes free groups and free abelian groups and may be said to interpolate between these two extremes. We study the outer automorphism group of a right-angled Artin group in the case where the defining graph is connected and triangle-free. We give an algebraic description of Out(A Γ) in terms of maximal join subgraphs in Γ and we construct an analogue of outer space for these groups. We prove that this outer space is finite dimensional, contractible, and has a proper action of Out(A Γ), from which it follows that Out(A Γ) has finite virtual cohomological dimension. We give upper and lower bounds on this dimension. 1. Introduction A right-angled Artin group is a finitely presented group whose relations are simply commutators of generators. These groups have nice algorithmic properties and act naturally on CAT(0) cube complexes. Also known as graph groups, they occur in many different mathematical contexts; for some particularly interesting examples we refer to the work of Bestvina-Brady [BB97] on finiteness properties of groups, Croke-Kleiner [CK00] on boundaries of CAT(0) spaces, and Abrams [Abr02] and Ghrist [Ghr01],[GP06] on configuration spaces in robotics. A right-angled Artin group can be efficiently described by giving a finite simplicial graph Γ. If U is the vertex set of Γ, then the group A Γ is defined by the presentation A Γ = U vw = wv if v and w are connected by an edge in Γ. At the two extremes of this construction are the case of a graph with n vertices and no edges, in which case A Γ is a free group of rank n, and that of a complete graph on n vertices, in which case A Γ is a free abelian group of rank n. In general, right-angled Artin groups can be thought of as interpolating between these two extremes. Thus it seems reasonable to consider automorphism groups of right-angled Artin groups as interpolating between Aut(F n ), the automorphism group of a free group, and GL n (Z), the automorphism group of a free abelian group. The automorphism groups of free groups and of free abelian groups have been extensively studied but, beyond work of Servatius [Ser89] and Laurence [Lau95] on generating sets, there seems to be little known about the automorphism groups of general right-angled Artin groups. In this paper we begin a systematic study of automorphism groups of right-angled Artin groups. We restrict our our attention to the case that the defining graph Γ is connected 1

2 2 RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN and triangle-free or, equivalently, in which A Γ is freely indecomposable and contains no abelian subgroup of rank greater than two. A key example is when Γ is a complete bipartite graph, which we call a join since Γ is the simplicial join of two disjoint sets of vertices. The associated Artin group A Γ is a product of two free groups. If neither of the free groups is cyclic, then the automorphism group of A Γ is just the product of the automorphism groups of the two factors (or possibly index two in this product). If one of the free groups is cyclic, however, the automorphism group is much larger, containing in addition an infinite group generated by transvections from the cyclic factor to the other factor. We will show that for any connected, triangle-free graph Γ, the maximal join subgraphs of Γ play a key structural role in the automorphism group. We do this by proving that the subgroups A J generated by maximal joins J are preserved up to conjugacy (and up to diagram symmetry) by automorphisms of A Γ. This gives rise to a homomorphism Out 0 (A Γ ) Out(A J ), where Out 0 (A Γ ) is a finite index normal subgroup of Out(A Γ ) which avoids certain diagram symmetries. We use algebraic arguments to prove that the kernel of this homomorphism is a finitely generated free abelian group. Elements of this kernel commute with certain transvections, called leaf transvections, and adding them forms an even larger free abelian subgroup of Out(A Γ ). The second part of the paper takes a geometric turn. The group Out(F n ) can be usefully represented as symmetries of a topological space known as outer space. This space, introduced by Culler and Vogtmann in [CV86], may be described as a space of actions of F n on trees. Outer space has played a key role in the study of the groups Out(F n ) (see for example, the survey article [Vog02]). For GL(n, Z) the analogous outer space of actions of Z n on R n is the classical homogeneous space SL(n, R)/SO(n, R). In Section 4 of this paper we construct an outer space O(A Γ ) for any right-angled Artin group A Γ. If Γ is a single join, O(A Γ ) consists of actions of A Γ on products of trees T T. For a general Γ, a point in outer space is a graph of such actions, parameterized by a collection of maximal joins in Γ. We prove that for any connected, triangle-free graph Γ, the space O(A Γ ) is finite dimensional, contractible, and has a proper action of Out 0 (A Γ ). Moreover, it retracts equivariantly onto a lower dimensional spine. This gives an upper bound on the virtual cohomological dimension of Out(A Γ ) complementing the lower bound given by the free abelian subgroup found in the first part of the paper. We isolate a certain subset U 0 of the vertex set U and prove, Theorem 1.1. The virtual cohomological dimension of Out(A Γ ) is finite and satisfies ( ˆm v 1) vcd(out(a Γ )) (3l v 4), v U 0 v U 0 where l v denotes the valence of v and ˆm v the number of connected components of Γ {v}. If Γ is a tree, for example, then U 0 consists of all interior vertices of Γ and the above formula reduces to (e 1) vcd(out(a Γ )) (2e + l 4),

3 AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS 3 where e is the number of edges in Γ and l is the number of leaves. 2. Algebraic Preliminaries 2.1. Special subgroups. A graph Γ with vertices V is called simplicial if it is a simplicial complex, i.e., if it has no loops or multiple edges. Vertices of valence one are called leaves, and all other vertices are interior vertices. To each finite simplicial graph Γ we associate the right-angled Artin group A Γ as in the introduction. For any full subgraph Θ Γ, the Artin group A Θ is naturally isomorphic to the subgroup of A Γ generated by the vertices in Θ. Such a subgroup is called a special subgroup of A Γ. In the discussion which follows, we will need to know the normalizer N(A Θ ), the centralizer C(A Θ ), and the center Z(A Θ ) of a special subgroup A Θ A Γ. These are easily derived from the work of Servatius [Ser89] and Godelle [God03]. For any subgraph Θ of Γ, let Θ denote the subgraph of Γ spanned by the set of vertices in Γ which commute with every vertex of Θ. For example, if Θ consists of a single vertex v, then Θ = st(v), the (closed) star of v. (This notation differs slightly from that of Godelle whose definition of Θ excludes vertices in Θ.) Proposition 2.1. For any graph Γ, and any special subgroups A Θ and A Λ of A Γ, (1) the normalizer, centralizer, and center of A Θ are given by N(A Θ ) = A Θ Θ C(A Θ ) = A Θ Z(A Θ ) = A Θ Θ. (2) If ga Θ g 1 A Λ, then Θ Λ and g = g 1 g 2 for some g 1 N(A Λ ), g 2 N(A Θ ). Proof. In [Ser89], Servatius proves that the centralizer of a single vertex v is the special subgroup generated by st(v) and hence for a set of vertices Θ, the centralizer, C(A Θ ), is generated by the intersection of these stars, which is exactly Θ. It follows that the center of A Θ is A Θ C(A Θ ) = A Θ Θ. In [God03], Godelle considers normalizers and centralizers of special subgroups in a larger class of Artin groups, Artin groups of FC type. He defines the quasi-centralizer, QZ(A Θ ), of a special subgroup to be the group of elements g which conjugate the set Θ to itself, and he proves that N(A Θ ) = A Θ QZ(A Θ ). In the right-angled case, no two generators are conjugate, hence QZ(A Θ ) = C(A Θ ) = A Θ and N(A Θ ) = A Θ Θ. Godelle also describes the set of elements which conjugate one special subgroup A Θ into another A Λ in terms of a category Ribb(V ) whose objects are subsets of the generating set V and whose morphisms conjugate one subset of V into another. In the case of a right-angled Artin group, since no two generators are conjugate, there are no morphisms between distinct objects of Ribb(V ) and the group of morphisms from an object Θ to itself is precisely the centralizer C(A Θ ) = A Θ. Proposition 3.2 of [God03] asserts that ga Θ g 1 A Λ if and only if g = g 1 g 2 where g 1 A Λ and g 2 is a morphism in a certain subcategory of Ribb(V ). In the right-angled case, it is straightforward to verify that such a morphism exists if and only if Θ Λ and g 2 A Θ. (In Godelle s notation, he decomposes Θ into disjoint subsets Θ = Θ s Θ as where Θ s is the set of generators lying in the center of A Θ. His theorem states that g 2 must commute with Θ s and conjugate Θ as to a set R

4 4 RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN w 1 v w 2 Figure 1. Graph with a partial conjugation w i v 1 w i v with R Θ s Λ. In the right-angled case, this is possible only if R = Θ as and g 2 also commutes with Θ as.) Thus g A Λ A Θ N(A Λ )N(A Θ ). Associated to each right-angled Artin group A Γ, there is a CAT(0) cube complex C Γ on which A Γ acts, constructed as follows. The 1-skeleton of C Γ is the Cayley graph of A Γ with generators U. There is a cube of dimension k > 1 glued in wherever possible, i.e. wherever the 1-skeleton of a cube exists in the Cayley graph. In the quotient by the action there is a k-dimensional torus for each complete subgraph of Γ with k vertices. The cube complex associated to a free group is simply the Cayley graph of the free group, i.e. a tree on which the free group acts freely. The cube complex associated to the complete graph on n vertices is the standard cubulation of R n. The cube complex associated to a join V W is a product T V T W, where T V (resp T W ) is a tree on which the free group F V (resp F W ) acts freely with quotient a rose Generators for the automorphism group of a right-angled Artin group. A set of generators for Aut(A Γ ) was found by M. Laurence [Lau95], extending work of H. Servatius [Ser89]. There are five classes of generators: (1) Inner automorphisms (2) Inversions (3) Partial conjugations (4) Transvections (5) Symmetries Inversions send a standard generator of A Γ to its inverse. A partial conjugation exists when removal of the (closed) star of some vertex v disconnects the graph Γ. In this case one obtains an automorphism by conjugating all of the generators in one of the components by v. (See example in Figure 1.) Transvections occur whenever there are vertices v and w such that st(v) lk(w); in this case the transvection sends w wv. There are two essentially different types of transvections, depending on whether or not v and w commute: (1) Type I transvections. v and w are not connected by an edge. (2) Type II transvections. v and w are connected by an edge. (See examples in Figure 2.) Finally, symmetries are induced by symmetries of the graph, and permute the generators.

5 AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS 5 v w v w Type I Type II Figure 2. Graphs with transvections w wv We will be especially interested in the subgroup we obtain by leaving out the graph symmetries: Definition 2.2. The subgroup of Aut(A Γ ) generated by inner automorphisms, inversions, partial conjugations and transvections is called the pure automorphism group and is denoted Aut 0 (A Γ ). The image of Aut 0 (A Γ ) in Out(A Γ ) is the group of pure outer automorphisms and is denoted Out 0 (A Γ ). The subgroups Aut 0 (A Γ ) and Out 0 (A Γ ) are easily seen to be normal and of finite index in Aut(A Γ ) and Out(A Γ ) respectively. We remark that if A Γ is a free group or free abelian group, then Aut 0 (A Γ ) = Aut(A Γ ). 3. Maximal Joins 3.1. Restriction to connected, two-dimensional right-angled Artin groups. If Γ is disconnected, then A Γ is a free product of the groups associated to the components of Γ. Guirardel and Levitt [GLb] have constructed a type of outer space for a free product with at least one non-cyclic factor, which can be used to reduce the problem of understanding the outer automorphism group to understanding the outer automorphism groups of the free factors. Therefore, in this paper we will consider only connected graphs Γ. We will further restrict ourselves to the case that Γ has no triangles. In this case, the associated cube complex C Γ is 2-dimensional so we call these two-dimensional right-angled Artin groups. To avoid technicalities, we also assume that Γ has at least 2 edges. Type II transvections are severely limited in the two-dimensional case. Since v and w are connected by an edge, the vertex w must actually be a terminal vertex i.e. a leaf: if there were another vertex u v connected to w, then the conditon st(v) lk(w) would imply that u,v and w form a triangle in the graph (see Figure 2). For this reason, we call Type II transvections leaf transvections. If Γ is triangle-free and Θ is a subgraph with at least one edge, then Θ Θ, so by Proposition 2.1, N(A Θ ) = A Θ and C(A Θ ) = Z(A Θ ) = A Θ. If Θ contains two non-adjacent edges, then the latter groups are trivial. Key example. A join Γ = V W, has no triangles. As we noted above, the associated right-angled Artin group A Γ is F(V ) F(W). It is easy to deduce the structure of the

6 6 RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN automorphism group from Laurence s generators. If V and W each contain at least two elements then every automorphism preserves the two factors (or possibly switches them if V = W ). Thus Out(A Γ ) contains Out(F(V )) Out(F(W)) as a subgroup of index at most 2. If V = {v} and W = l 2, then A Γ = Z F(W) with the center Z generated by v. The elements of W are all leaves. Any automorphism of A Γ must preserve the center and hence induces an automorphism of F(W), as well as an automorphism of the center Z. The map Out(A Γ ) Out(Z) Out(F(W)) splits and its kernel is the group generated by leaf transvections. The leaf transvections commute, so Out(A Γ ) = Z l (Z/2 Out(F(W))). We assume for the rest of this paper that that Γ is connected and triangle-free. In addition, we assume that Γ contains at least two edges Restricting automorphisms to joins. A connected, triangle-free graph Γ can be covered by subgraphs which are joins. For example, to each interior vertex v we can associate the join J v = L v L v, where L v = lk(v). Note that L v always contains v. If L v = {v}, then J v = st(v); in this case we say that v is a cyclic vertex since F(L v ) = Z is cyclic. Lemma 3.1. If v and w are interior vertices joined by an edge of Γ, then L v L w, so we have In particular, J w J v = L w L v. J v = L v L v J w = L w L w The join J v is not properly contained in any other join subgraph of Γ, i.e. J v is a maximal join in Γ. The following proposition shows that the special subgroups A J associated to maximal join subgraphs J of Γ are preserved up to conjugacy by pure automorphisms: Proposition 3.2. Let φ Aut 0 (A Γ ) be a pure automorphism of A Γ and let J = V W be a maximal join in Γ. Then φ maps A J = F(V ) F(W) to a conjugate of itself. Moreover, if V contains no leaves, then φ preserves the factor F(V ) up to conjugacy. Proof. It suffices to verify the proposition for the generators of Aut 0 (A Γ ). Inner automorphisms. These obviously send each A J to a conjugate of itself. Inversions. An inversion sends each A J to itself and preserves the factors. Partial conjugations. If φ is a partial conjugation by a vertex u, we claim that φ either fixes all of A J or conjugates it by u. Suppose first that u is not in V W. The link of u cannot contain vertices of both V and W, since there would then be a triangle in Γ. Furthermore, lk(u) cannot contain all of V, since then adding it to W would make a larger join, contradicting maximality. Therefore the subgraph of V W spanned by vertices not in lk(u) is still connected, so φ has the same effect (either trivial or conjugation by u) on generators corresponding to all vertices in V W. Next, suppose that u is actually in V W, say u V. The resulting partial conjugation restricted to A V W is an internal

7 AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS 7 automorphism of A V W which may conjugate some generators of F(V ) by u, but has no effect on generators of F(W). Transvections. We claim that a transvection either fixes A J or acts as an internal automorphism of A J. If φ is a transvection sending s su, then φ is the identity on A V W unless s V W, say s V. If s is not a leaf, then the condition st(u) lk(s) and maximality imply that u is also in V, so the restriction of φ is an internal automorphism of A V W preserving the factor F(V ) and fixing F(W). If s is a leaf, then W = {u} and A V W = F(V ) Z. In this case, φ fixes the (central) Z factor and multiplies x F(V ) by the generator of Z. This proposition has two easy corollaries. Let Sym(Γ) denote the group of diagram symmetries of Γ and Sym 0 (Γ) = Sym(Γ) Aut 0 (Γ). Clearly Aut(A Γ )/Aut 0 (A Γ ) = Out(A Γ )/Out 0 (A Γ ) = Sym(Γ)/Sym 0 (Γ). Denote this quotient group by Q(Γ). Corollary 3.3. The quotient maps from Aut(A Γ ), Out(A Γ ), and Sym(Γ) to Q(Γ) split. Proof. It suffices to define a splitting of the projection Sym(Γ) Q(Γ). Composing with the inclusion of Sym(Γ) into Aut(A Γ ) or Out(A Γ ), gives a splitting in the other two cases. We first characterize elements of Sym 0 (Γ) as those graph symmetries which simply permute sets of vertices with the same link. Define an equivalence relation on the vertices of Γ by v w if lk(v) = lk(w). The elements of an equivalence class [v] generate a free subgroup, and any automorphism of this subgroup extends (via the identity) to an automorphism of the whole Artin group A Γ. Since Aut 0 = Aut for a free group, any permutation of [v] can be realized by an element of Sym 0 (Γ) which is the identity outside [v]. Composing these gives an automorphism in Sym 0 (Γ) realizing any permutation of the elements of each equivalence class. Conversely, if a graph symmetry is in Sym 0 (Γ) we claim that it acts by permuting the elements of each equivalence class [v]. If v is a leaf, it is connected to a vertex w which is not a leaf, so by the last statement of Proposition 3.2 the symmetry must fix w and permute [v], which consists of all leaves connected to w. If v is not a leaf, then L v contains no leaves, so by the last statement of Proposition 3.2 applied to the maximal join L v L v, the symmetry must permute the vertices of L v = [v]. Now choose an ordering on the vertices in each equivalence class. Then we can define a splitting of Sym(Γ) Q(Γ) by mapping a coset to the unique element of the coset which is order preserving on every equivalence class. Corollary 3.4. For every maximal join J Γ, there is a restriction homomorphism R J : Out 0 (A Γ ) Out(A J ). Proof. Fix J. Then for any element of Out 0 (A Γ ), there is a representative φ Aut 0 (A Γ ) which maps A J to itself. Any two such representatives differ by conjugation by an element of the normalizer of A J. But the normalizer N(A J ) is equal to A J since J is a maximal join, so the restriction of φ to A J is a well-defined element of Out(A J ).

8 8 RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN v w Figure 3. Example with Γ 0 = [v,w] We can put all of the homomorphisms R J together to obtain a homomorphism R = R J : Out 0 (A Γ ) Out(A J ). J M To understand Out 0 (A Γ ), then, we would like to understanding the image and kernel of this homomorphism. However, we note that there is a lot of redundant information in the set of all maximal joins used to define R; for example, just the maximal joins J v already cover Γ. Even the covering of Γ by the maximal joins J v is inefficient. If v and w have the same link, then J v = J w so we don t need them both. With this in mind, we now specify a subgraph Γ 0 of Γ which will turn out to contain all of the information we need. The key idea is that of vertex equivalence, which we already encountered in the proof of Corollary 3.3. Definition 3.5. Vertices v and w of Γ are called equivalent if they have the same link, i.e. L v = L w. Equivalence classes of vertices are partially ordered by the relation [v] [w] if L v L w. To define Γ 0, we choose a vertex in each maximal equivalence class and let Γ 0 be the full subgraph of Γ spanned by these vertices. In the special case that Γ is a star {v} W, set Γ 0 = {v}. Up to isomorphism, Γ 0 is independent of the choice of representatives. We denote by U 0 the set of vertices in Γ 0. Examples 3.6. (i) If Γ is a tree, then Γ 0 is the subtree spanned by the vertices which are not leaves. (ii) If Γ is the graph in Figure 3.2, then Γ 0 is the single edge spanned by v and w. Whether we are working with vertices of Γ or Γ 0, the notation L v will always refer to the link of v in the original graph Γ. Likewise, a vertex is considered a leaf if it is a leaf of Γ, not just of Γ 0. Recall that a vertex v of Γ is called a cyclic vertex if the associated maximal join J v is equal to st(v). The following lemma specifies the properties of Γ 0 which we will need. Lemma 3.7. Γ 0 is a connected subgraph of Γ. The vertex set U 0 of Γ 0 contains every cyclic vertex and no leaf vertices. If Γ is not a star, then every vertex of Γ lies in L w for at least one w U 0 and lies in L w for at most one w U 0. Proof. Let v,w be two vertices in Γ 0 and let J M v = v 0,v 1,... v k = w

9 AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS 9 be an edgepath in Γ connecting v to w. If v 1 does not lie in Γ 0, then there is a vertex v 1 Γ 0 with L v1 L v 1. Replacing v 1 by v 1 gives another edgepath in Γ from v to w whose first edge lies in Γ 0. The first statement of the lemma now follows by induction on k. For the second statement note that [v] L v for any v, and [v] = L v if and only if [v] is maximal. This is always the case when v is cyclic, since then L v = {v}. If v is a leaf then v is connected to some interior vertex w by an edge (recall that we have assumed that the diameter of Γ is at least two). If Γ = st(w) we have defined Γ 0 = {w} so v / U 0. If Γ is not a star, w must be connected to some other interior vertex u. It follows that L v = {w} L u, hence v / U 0. For the third statement, note that since Γ is connected, every vertex v lies in the link of some other vertex and hence lies in the link of some maximal vertex. Since L w = [w] for every vertex w in Γ 0, v lies in at most one such L w The kernel of the restriction and projection homomorphisms. We are interested in determining the kernel of the map R constructed from the restriction homomorphisms R J : Out 0 (A Γ ) Out(A J ). We first consider the kernel of the analogous map R 0 defined by looking only at the R J for maximal joins J = J v, for v U 0 : R 0 = R Jv : Out 0 (A Γ ) Out(A Jv ) v U 0 v U 0 This kernel consists of outer automorphisms such that any representative in Aut 0 (A Γ ) acts by conjugation on each A Jv. Thus for any φ Out 0 (A Γ ) and any interior vertex v, we can choose a representative automorphism φ v such that φ v (A Jv ) = A Jv. Since L v contains no leaves, φ v must also preserve A L v by Propositon 3.2. If v and w are interior vertices connected by an edge, then the representatives φ v and φ w are related as follows. Lemma 3.8. Suppose v,w are interior vertices connected by an edge in Γ, and φ Out 0 (A Γ ) is represented by automorphisms φ v and φ w with φ v (A Jv ) = A Jv and φ w (A Jw ) = A Jw. Then there exists g v A Jv and g w A Jw such that c(g v ) φ v = c(g w ) φ w, where c(g) denotes conjugation by g. Proof. Since φ v and φ w represent the same element in Out 0 (A Γ ), φ w φ 1 v = c(g) for some g A Γ. Since φ v preserves A L v, and A L v A Jw, we have ga L v g 1 A Jw. By Proposition 2.1 (2), we must then have g = g 1 g 2 with g 1 N(A Jw ) = A Jw and g 2 N(A L v ) = A Jv. Taking g w = g1 1 and g v = g 2 gives the desired formula. A vertex v is a separating vertex if Γ {v} is disconnected. It is easy to see that if v is separating, then v is cyclic. It is also easy to see that conjugating one component of Γ {v} gives an element of the kernel of R. We remark, however, that one component of Γ {v} may contain more than one component of Γ st(v), so not every partial conjugation by v lies in this kernel. In the graph in Figure 4, for example, Γ {v} has two components while Γ st(v) has four. The partial conjugation of u by v restricts non-trivially in Out(A Jw ). Proposition 3.9. The kernel K of the homomorphism R 0 is a finitely generated, free abelian group. It is generated by conjugations by separating vertices v on components of

10 10 RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN u w v Figure 4. Γ st(v) has more components than Γ {v} Γ {v} and has rank v U 0 (m v 1) where m v is the number of connected components of Γ {v} containing a non-leaf vertex. Proof. If Γ 0 consists of a single vertex v, then Γ is the star of v so J v = Γ and R 0 is injective. Therefore we may assume that Γ 0 has at least two vertices. Suppose φ is an element of the kernel of R 0. Then for each v Γ 0 we may choose a representative φ v Aut 0 (A Γ ) which acts as the identity on A Jv. If v and w in Γ 0 are connected by an edge, then by Lemma 3.8 there are elements g v A Jv and g w A Jw such that c(g v ) φ v = c(g w ) φ w, i.e. φ v = c(gv 1 g w ) φ w. Since φ w acts as the identity on J w, the action of φ v on J w is conjugation by g = gv 1 g w. Since both φ v and φ w act as the identity on A Jv A Jw, g is in the centralizer of A Jv A Jw = F(L v ) F(L w). This centralizer is easily computed using Proposition 2.1; it is contained in the free abelian group v,w generated by v and w, and is contained in v if w is not cyclic. If w is cyclic, then w is in the center of A Jw ; thus in all cases φ v is equivalent on A Jw to conjugation by v nw for some n w Z. Fix v G 0. We claim that if u and w in Γ 0 are both in the link of v and are both in the same component of Γ {v}, then n u = n w. To prove this, let C be the component of Γ {v} containing u and w, and choose a path u,v 1,...,v k,w in C from u to w. Now choose u i in Γ 0 with lk(u i ) lk(v i ). Then u,u 2,...,u k 1,w is a path in Γ 0, with all vertices in C {v}. If u i is never equal to v, we have a loop γ Γ 0 with vertices v = u 0,u = u 1,...,u k = w,u k+1 = v. Replace this by a simple loop, if necessary, with all vertices except v in Γ 0 {v}. By our above analysis of edges in Γ 0, φ ui = c(g i ) φ ui+1 for some g i u i,u i+1, and φ ui+1 acts on vertices of J ui+1 as conjugation by some power of u i. Thus φ v acts on vertices of J ui+1 as conjugation by g 0 g 1...g i 1 u n i i = v nu u m u m i i. Taking i = k, we get that φ v acts on J v as conjugation by v nu u m u m k k. Thus vnu u m u m k k lies in the center of A Jv, i.e. it is a power of v. But this is impossible unless n i = 0 for i = 1,...,k. It follows that in this case φ v acts on J ui as conjugation by the same power of v for all i, and in particular n u = n w. If u i = v for some i, then u i 1 and u i+1 are both in lk(v), so φ v conjugates u i 1 by some power v n i 1 and u i+1 by some power v n i+1. Since v i lies in both J ui 1 and J ui+1, the element v n i 1 n i+1 centralizes v i. Now v is not in st(v i ) since Γ has no triangles, so v does not commute with v i and we must have n i 1 = n i+1. Thus by decomposing the loop γ into

11 AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS 11 a product of loops at v, we see that n u = n i 1 = n i+1 =... = n w. This finishes the proof of the claim. Let C be a connected component of Γ {v} which contains a non-leaf vertex w. Then lk(w) contains at least one vertex x v. Since lk(w) lk(u) for some u Γ 0, the path from w to x to u is a path in in Γ {v} connecting w to Γ 0. Thus C contains at least one vertex of Γ 0. Since Γ 0 is connected and v Γ 0, we can in fact find a vertex of C Γ 0 which is in the link of v. Choose such a vertex for each component of Γ {v}. The associated conjugating powers give us an m v -tuple of integers (n w ). Since the choice of φ v is unique up to conjugation by a power of v, this m v -tuple is well-defined up to the diagonal action of Z on Z mv. Set µ w (φ) = (n v ) Z mw / (Z) = Z mw 1. It is straightforward to verify that µ w defines a homomorphism from K to Z mw 1. Let U cyc denote the set of all cyclic vertices, i.e. U cyc = {v V L v = {v} } U 0 We claim that the product homomorphism µ = µ w : K w U cyc w U cyc Z mw 1 is an isomorphism. Note that taking the product over w U 0 instead of w U cyc changes nothing since for w U 0 U cyc, Γ {w} is connected so m w 1 = 0. Surjectivity of µ is clear since the partial conjugations described in the proposition map onto a generating set for Z mw 1. Suppose φ lies in the kernel of µ w. Then φ w acts as conjugation by the same power of w, say w n, on A Jv for every v L w U 0. Replacing φ w by c(w n ) φ w we see that it is possible to choose φ w so that it acts as the identity on all vertices in J w and in J v for v L w U 0. The reader can easily check that the union of these joins includes all vertices in the original graph Γ of distance at most 2 from w. We will call this the 2-neighborhood of w and denote it by N 2 (w). Now suppose φ K lies in the kernel of µ w for every w U cyc. Then for each w U 0 (including those w not in U cyc ), we can choose a representative φ w which acts as the identity on the 2-neighborhood of w. If v and w are connected by an edge in Γ 0, then the chosen representatives, φ v and φ w, must differ by an element of the centralizer of N 2 (v) N 2 (w). But this intersection contains an edge-path of length three, so it s centralizer is trivial. In other words, φ v = φ w. Since Γ 0 is connected and the neighborhoods N 2 (w) cover Γ, we conclude that φ w is the identity on all of A Γ. This proves injectivity and completes the proof of the theorem. Remark It follows from the theorem that the homomorphism R, which was defined over all maximal joins J instead of just the joins J v with v U 0, has the same kernel as R 0. That is, if the restriction of φ to the subgroup A Jv is trivial for each v in U 0, then the same is true for the restriction to every A J.

12 12 RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN One advantage of restricting attention to the joins J v for v U 0 is that we can further define a projection homomorphism, as follows. For v U 0, L v contains no leaves, hence by Proposition 3.2, every φ Out 0 (A Γ ) has a representative φ v which preserves both A Jv and A L v. Thus φ v descends to an automorphism φ v of A Lv = A Jv /A L v. This gives rise to a homomorphism P v : Out 0 (A Γ ) Out(F(L v )). Let P be the product homomorphism P = P v : Out 0 (A Γ ) Out(F(Lv )). v U 0 Proposition Let ˆm v denote the number of connected components in Γ {v} The kernel ˆK of P is a free abelian group of rank v U 0 ( ˆm v 1). It is generated by K and the set of leaf transvections. Proof. Let l be the number of leaves in Γ. It is clear that leaf transvections are contained in ˆK and that they generate a free abelian group of rank l. It is also easy to see that leaf transvections commute with the generators of K since if u uv is a leaf transvection, then u and w are connected by an edge and hence belong to the same component of Γ {w} for any w Γ 0. Together with K, the leaf transvections thus generate a free abelian subgroup of rank l + v U 0 (m v 1) = v U 0 ( ˆm v 1). It remains only to show that this subgroup is all of ˆK. Consider an element φ in ˆK. For v U 0, let φ v denote a representative automorphism which preserves A Jv and hence also A L v. Then for w L v, we have φ v (w) = wg for some g A L v. Since w, and hence φ v (w), commutes with A L v, g must lie in the center of A L v. If v is not cyclic, the center is trivial, so g = 1. If v is cyclic, then then g = v k. If k 0, then φ v (C(w)) = C(wv k ) = w,v, so C(w) is free abelian of rank 2. This implies that w is a leaf. We conclude that φ v acts as the identity on non-leaf elements of L v and as leaf transvections on leaf elements. It follows that we can compose φ with an appropriate product of leaf transvections to obtain an outer automorphism θ and representative automorphisms θ v which restrict (as opposed to project) to the identity on A Lv for every v U 0. Taking an adjacent vertex w U 0, θ w restricts to the identity on A L v A Lw. Since θ v and θ w differ by an inner automorphism (which necessarily normalizes A L v ), θ v acts on A L v as conjugation by some h in the normalizer N(A L v ) = A Jv = A Lv A L v. Writing h as a product h 1 h 2 of an element in A Lv and an element in A L v, conjugation by h agrees with conjugation by h 2 on A L v, while on A Lv conjugation by h 2 is the identity. We conclude that θ v restricts to conjugation by h 2 on all of A Jv. In other words, θ lies in K. 4. Outer Space In this section we introduce outer space for a right-angled Artin group A Γ. We continue to assume that Γ is a connected, triangle-free graph and has diameter 2. Let O(F) denote the unreduced, unprojectivized version of Culler and Vogtmann s outer space for a free group F [CV86]. This space can be described as the space of minimal, free, isometric actions of F on simplicial trees. (The terms unreduced and unprojectivized specify that quotient graphs may have separating edges, and that we are not considering

13 AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS 13 homothetic actions to be equal.) Our initial approach to constructing outer space for A Γ was to consider minimal, free, isometric actions of A Γ on CAT(0) 2-complexes. In the case of a single join, these turn out to be products of trees. More generally, (under mild hypotheses) such a 2-complex is a union of geodesic subspaces which are products of trees. However, the interaction between these subspaces proved difficult to control and we ultimately found that it was easier to work directly with the tree-products Outer space for a join. Let us examine more closely the case when Γ = V W is a single join. Suppose that A Γ = F(V ) F(W) acts freely and cocompactly by isometries on a piecewise Euclidean CAT(0) 2-complex X with no proper invariant subspace. If neither V nor W is a singleton, then A Γ = F(V ) F(W) has trivial center, so the splitting theorem for CAT(0) spaces ([BH99], Theorem 6.21) says that X splits as a product of two one-dimensional CAT(0) complexes (i.e. trees) T V T W, and the action of A Γ is orthogonal, i.e. it is the product of the actions of F(V ) on T V and F(W) on T W. Twisting an action by an element of Out(A Γ ), which contains Out(F(V )) Out(F(W)) as a subgroup of index at most two, preserves the product structure. Thus it makes sense to take as our outer space the product of the Culler-Vogtmann outer spaces O(F(V )) O(F(V )). If V = {v} is a single vertex then A Γ = Z F(W) where Z is generated by v. In this case, X is equal to the min set for v, so that X splits as a product α v T W, where α v is an axis for v ([BH99], Theorem 6.8) and T W is a tree. We identify the axis α v with a real affine line, with v acting by translation in the positive direction by an amount t v > 0. The tree T W has a free F(W)-action induced by the projection A Γ F(W), but the action of A Γ on X need not be a product (orthogonal) action. Recall that Out(A Γ ) = Z l (Z/2 Out(F(W)) where l = W and Z l is generated by leaf transvections. Twisting an orthogonal action by a leaf transvection w wv results in an action which is no longer orthogonal. Instead, w now acts as translation in a diagonal direction on the plane in α v T W spanned by an axis for v and an axis for w, i.e. the translation vector has a non-trivial α v -component. More generally, for any action of A Γ on α v T W, v acts only in the α v -direction, v (r,x) = (r + t v,x), while an element w i W has a skewing constant λ(w i ), i.e. w i acts by w i (r,x) = (r + λ(w i ),w i x). A free action of A Γ on X is thus determined by an F(W)-tree T W, the translation length t v and an l-tuple of real numbers (λ(w 1 ),...,λ(w l )). So it is reasonable in the case, to take for outer space the product R >0 R l O(F(W) Tree spaces. Our understanding of the join case motivates the following definition. Let J = V W be a join in Γ. An admissible tree-space X J for J is a product of two simplicial, metric trees, T V and T W with free, minimal, isometric actions of F(V ) and F(W) respectively, and an action of A J = F(V ) F(W) on X J = T V T W of the following type. (1) If J contains no leaves, then the action is the product of the given actions, (g 1,g 2 ) (x 1,x 2 ) = (g 1 x 1,g 2 x 2 ).

14 14 RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN (2) Suppose J contains leaf vertices, say in W. This forces V = {v}, so T V = R and v acts on T V as translation by some positive real number t v. Then there exists a homomorphism λ : F(W) R which is zero on non-leaf vertices of W, such that (v n,g) (r,x) = (r + nt v + λ(g), g x). We remark that the definition of admissible depends not only on the join J, but on the graph Γ as well since Γ determines which vertices are considered as leaves. Since Γ is fixed throughout, this should not cause any confusion. A point in outer space for A Γ will be a collection of admissible tree-spaces satisfying certain compatibility conditions. Recall that to each interior vertex of Γ we have associated a maximal join J v = L v L v. If e is an edge from v to w, set J e = J v J w = L v L w. If e lies in Γ 0, then J e contains no leaves. Definition 4.1. A graph of tree-spaces X = {X v,x e,i e,v } for A Γ consists of the following data. (1) For each vertex v Γ 0, an admissible tree-space X v for J v, (2) For each edge e Γ 0 with vertices v and w, an admissible tree space X e for J e and a pair of A Je -equivariant isometric embeddings. We define outer space for A Γ to be the set i e,v i e,w X v X e X w. O(A Γ ) = {X X is a graph of tree-spaces for A Γ }/ where is the equivalence relation induced by replacing any X v (respectively X e ) with an equivariantly isometric space X v (respectively X e ), and composing the associated connecting maps i e,v, by the equivariant isometry. A natural topology for O(A Γ ) will be described in Section 4.2. Since J e contains no leaves, the group A Je = F(L v ) F(L w ) acts orthogonally on X e and the maps i e,v and i e,w split as products. Write X v = T v Tv, where T v is an F(L v )- tree, and Tv is an F(L v )-tree. Then up to equivariant isometry we may assume that X e = Tv T w, and that i e,v = (i 1,i 2 ) is the identity on the first factor while i e,w = (j 1,j 2 ) is the identity on the second. The image of i 2 : Tw T v is uniquely determined, namely it is the minimal F(L w )-invariant subtree of T v. However, the map i 2 is not necessarily unique. If w is not cyclic, then there are no non-trivial equivariant isometries of Tw, so i 2 is unique. On the other hand, if w is cyclic then Tw is a real line and the possible equivariant inclusions i 2 are parameterized by R. To keep track of these inclusions, it will be convenient to introduce basepoints. These will also play a crucial role in the proof of the contractibility of O(A Γ ) Basepoints for free actions of free groups. Let F(S) be a finitely generated free group with a specified basis S of cardinality at least 2, and let T be a metric tree with a minimal, free, isometric action of F(S). Each generator s S preserves a unique line

15 AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS 15 α(t) p(t, s) = b(s) p(t, s) p(t,s) α(s) α(t ) α(t ) Figure 5. Projections and basepoint on the axis α(s) α(s) in T called the axis for s. Orient the axis so that s acts as translation in the positive direction. We choose a base point b(s) on α(s) as follows. For each generator t s, the set of points on α(s) of minimal distance from α(t) is a closed connected interval (possibly a single point). Define the projection p(t, s) of α(t) on α(s) to be the initial point of this interval, and the basepoint b(s) to be the minimum of these projections, with respect to the ordering given by the orientation of α(s). (See Figure 5). The basepoints b(w) will be called the unrestricted basepoints of the action Basepoints in X. Suppose now that we have a graph X of tree-spaces for A Γ. Basepoints in Tw. Assume Γ is not a star, so Γ 0 contains at least two vertices. If L w contains at least two vertices, then for each u L w we let b(u) be the unrestricted basepoint on the axis α(u) for the action of F(L w) on Tw. If w is cyclic the axis α(w) is the entire tree Tw, in which case we cannot use this method to choose a basepoint. If v is adjacent to w in Γ 0, we have an equivariant isometry i(w,v): Tw T v. In T v, the element w has an axis α v (w) with an unrestricted basepoint b v (w). We can use i(w,v) to pull this back to a point c v (w) on α(w). In this way, we get a point c v (w) on α(w) for each v adjacent to w in Γ 0. We take the minimum b(w) of these points as a basepoint for α(w). (See Figure 6.) Basepoints in T v. Suppose v is connected to w by an edge in Γ 0. In Tw, we have a basepoint b(w) on α(w); we also have an equivariant isometry i(w,v): Tw T v sending α(w) to α v (w). We take the image s v (w) of b(w) as a basepoint for α v (w). Note that if w is cyclic, the basepoint s v (w) may be different from the unrestricted basepoint on α v (w); it depends not just on information associated to the link L v but also to vertices in the 2-neighborhood of v Topology on O(A Γ ). We will define the topology on O(A Γ ) by embedding O(A Γ ) in a product of topological spaces and giving it the resulting subspace topology. In the case that Γ is a star {v} W, we have seen that O(A Γ ) can be identified with O(F(W)) R l R >0, where l = W. We take this to be a homeomorphism. For any graph Γ, let U cyc denote the set of cyclic vertices.

16 16 RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN v Γ w u α u (w) u α v (w) α(w) i(w,u) T Lu b u (w) b v (w) T Lv i(w,v) c v (w) c u (w) i(w,u ) α u (w) c u (w) = b(w) T Lu b u (w) T Lw Figure 6. Basepoint on the axis for w in T w Proposition 4.2. Let Γ be a graph which is not a star, which has l 0 leaves. Let k = v U cyc k v, where k v = L v Γ 0. Then there is an injective map O(A Γ ) O(F(L v )) R l R k. v Γ 0 The image of this map is of the form Y R l Q, where Y is a subspace of v U 0 O(F(L v )) and Q is a piecewise linear subspace of R k of dimension v U cyc (k v 1). Proof. Let X be an element of O(A Γ ). The vertex spaces X v are determined by free group actions on two trees, Tv and T v, together with a real number λ(w) for each leaf vertex w L v. Since Γ is not a star, there is at least one vertex w of L v which is also in Γ 0. The tree Tv is equivariantly isometric to the minimal F(L v ) subtree of T w for this w. Thus the adjacency relations of Γ, the trees T u for u U 0 and the real numbers λ(w) for w a leaf completely determine the vertex spaces X v. As noted above, the edge spaces are of the form X e = Tu T v with orthogonal action of F(L u ) F(L v ), so that they too are determined by the actions on the trees T u.

17 AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS 17 It remains to account for the connecting maps between edge spaces and vertex spaces. If v is not cyclic, then for any vertex w L v Γ 0, there is a unique equivariant map from Tv into T w, so the connecting map X e X w is uniquely determined. If v is cyclic, then Tv is a copy of the real line, and an equivariant embedding of Tv T w is determined by the position of the image of b(v) on α w (v) i.e. by the difference v (w) = b w (v) s w (v) = c w (v) b(v), where the points b w (v),s w (v),c w (v) and b(v) are as defined at the beginning of Section 4.4. Thus for each v U 0 we have a point T v O(F(L v )), real numbers λ(w) for w a leaf in L v, and if v is cyclic, real numbers v (w) for all w L v U 0. This data completely determines X. Note that if w is a leaf in Γ, it is connected to a unique vertex v which is necessarily cyclic, hence lies in U 0. Thus there is exactly one λ(w) for each leaf vertex, and we have an injective map f : O Γ O(F(L v )) R l R k v U 0 where k and l are as defined in the statement of the proposition. Now consider the image of f. There are no restrictions at all on the numbers λ(w); given any X in O(A Γ ), arbitrarily changing λ(w) for any leaf w gives rise to another valid point in O(A Γ ). As for the numbers v (w) = c v (w) b(w), since b(w) was defined to be the infimum of the c v (w), we must have v (w) 0 for all w L v U 0, and at least one of these must equal 0. There are no other restrictions. Recall that k v = L v U 0 and let Q v R kv be defined by Q v = {(r 1,... r kv ) r i 0 for all i and r j = 0 for some j}. In other words, Q v is the boundary of the positive orthant of R kv. Then the image of f is of the form Y R l v U cyc Q v where Y is the projection of the image of f onto v U 0 O(F(L v )). Remark 4.3. If Γ = V W is a join with V, W 1, then Γ 0 consists of a single edge e joining a pair of vertices v V and w W. In this case, J e = J v = J w = Γ and a point in O(A Γ ) is determined by a single tree space for Γ (since there is a unique equivariant isometry of such a tree-space). In this case, l = k v 1 = k w 1 = 0 and Y is all of O(F(V )) O(F(W)). Thus, the definition of outer space for a join agrees with that proposed in Section 4.1. From now on, O(A Γ ) will be viewed as a topological space with the subspace topology induced by the embedding f described above. Each space O(F) is endowed with the equivariant Gromov-Hausdorff topology. In this topology, a neighborhood basis of an action of F on a tree T is given by the sets N(X,H,ǫ), where X = {x 1,...,x N } is a finite set of points in T and H = {g 1,...,g M } is a finite set of elements of F. An action of F on T is in this neighborhood if there is a subset X = {x 1,...,x k } of T such that for all i,j,k. d T (x i,g j x k ) d T (x i,g j x k ) < ǫ

18 18 RUTH CHARNEY, JOHN CRISP, AND KAREN VOGTMANN 4.6. Contractibility. In this section we prove the following theorem: Theorem 4.4. For any connected, triangle-free graph Γ, the space O(A Γ ) is contractible. The proof follows ideas of Skora [Sko89] and Guirardel-Levitt [GLb, GLa] on unfolding trees. In their work, however, the unfolding of a tree was defined with respect to a single basepoint. In our case we will need to preserve several basepoints. We first consider a single action. Let F(S) be a free group with a preferred generating set S and let T O(F(S)) be an F(S)-tree. Proposition 4.5. Let P = {p(t,s)} be any set of projections, with s,t S. The convex hull B(P,T) of P in T depends continuously on T. Proof. If T is any other F(S)-tree, let P = {p (t,s)} be the corresponding projections in T. We must show: given ǫ > 0 and T O(F(S)), there is a neighborhood N of T such that B(P,T ) is ǫ-close to B(P,T) for T in N. To show B(P,T ) is ǫ-close to B(P,T) in the Gromov-Hausdorff topology, we need to take an arbitrary finite set of points X in B(P,T) and find corresponding points X in B(P,T ) such that the distances between points in X are within ǫ of the distances between the corresponding points in X. For T close to T in the equivariant Gromov-Hausdorff topology, we can find points X in T with the required properties, but that is not good enough; we need the points X to be in B(P,T ). We can fix this by projecting each point x X onto B(P,T ), but we need to be sure that this projection is sufficiently close to x. Each projection p = p(t, s) is uniquely determined by the following set of equations (see [GLb]): (1) d(p,sp) = l(s) d(p,tp) = l(t) + 2D d(s 1 p,t 1 p) = l(s) + l(t) + 2D d(s 1 p,tp) = l(s) + l(t) + 2D where D is the distance from α(s) to α(t). If T is in the N({p}, {s 1,t,t 1 },ǫ)-neighborhood of T, then we can find q in T satisfying equations (1) up to ǫ. By ([Pau89]), the corresponding lengths l (s),l (t) and D are within ǫ of l(s),l(t) and D, so that q satisfies the analogous equations (1) up to 4ǫ. The projection p = p (s,t) in T satisfies the equations (1) exactly. It is an easy exercise to verify that this implies that d(p,q ) < 3ǫ/2. Now let Y = X P, and let H = S S 1. If T is in the N(Y,H,ǫ)-neighborhood of T, then we can find q (s,t) and X in T with d(q (s,t),p (s,t)) < 3ǫ/2 and d(x 1,x 2 ) d(x 1,x 2 ) < ǫ for all x 1,x 2 X. We claim that the projection of x onto B(P,T ) is within 9ǫ of x for each x X. Each point x of B(P,T) is determined by its distances to the points p(s,t) (x is on some straight arc between p 1 = p(s 1,t 1 ) and p 2 = p(s 2,t 2 ); both the fact that it lies on this arc and its position on the arc are determined by its distances to p 1 and p 2 ).

19 and AUTOMORPHISMS OF 2-DIMENSIONAL RIGHT-ANGLED ARTIN GROUPS 19 If x is on [p 1,p 2 ] then d(p 1,x ) + d(x,p 2 ) d(p 1,q 1 ) + d(q 1,x ) + d(x,q 2 ) + d(q 2,p 2 ) d(p 1,x) + d(x,p 2 ) + 3ǫ/2 + 3ǫ/2 + 2ǫ = d(p 1,p 2 ) + 5ǫ d(p 1,p 2 ) d(q 1,q 2 ) d(p 1,q 1 ) d(p 2,q 2 ) d(p 1,p 2 ) ǫ 3ǫ d(p 1,p 2 ) 4ǫ For any three points in a tree, we have d(a,[b,c]) = 1 2 (d(b,a) + d(a,c) d(b,c)), so d(x,[p 1,p 2]) = d(p 1,x ) + d(x,p 2 ) d(p 1,p 2 ) 9ǫ. 2 Since [p 1,p 2 ] B(P,T ), the projection of x onto B(P,T ) is within 9ǫ of x, and we may replace x by this projection. Now let P be a set of projections with at least one projection on each axis α(s), for s S. We use the convex hull B(P,T) to define a new F(S)-tree T 0 (P,T) as follows. For each s S, let b(s) be the minimum of the projections p(t,s) in P (where α(s) is oriented in the direction of translation by s.) Form a labeled graph R(P, T) by attaching an oriented circle to b(s) labeled s, whose length is the translation length of s acting on T. We call R(P,T) a stemmed rose. The subgraph B(P,T) is the stem and the circles are the petals. Note that for any point x B(P,T), there is a canonical identification of π 1 (R,x) with F(S). Lifting to the universal cover of R(P,T) defines an action T 0 (P,T) of F(S) on a tree. Lemma 4.6. The action T 0 (P,T) depends continuously on T. Proof. By Proposition 4.5 the subtree B(P, T) and basepoints b(s) depend continuously on T, and by [Pau89] the lengths l(s) for s S depend continuously on T. Together, this data completely determines T 0 (P,T). Now fix a copy of B(P,T) in T 0 (P,T). Then there is a unique equivariant map f T : T 0 (P,T) T which is the identity on B(P,T). This map is a folding in the sense of Skora: for any segment [x,y] on T 0 (P,T), there exist a non-trivial initial segment [x,z] [x,y], x z, such that f restricted to [x,z] is an isometry. Following [GLb], we can use this morphism to define a path from T to T 0 (P,T) in O(F(S)) as follows. For r [0, ), let T r (P,T) be the quotient of T 0 (P,T) by the equivalence relation x r y if f T (x) = f T (y) and f T [x,y] is contained in the ball of radius r around f T (x). Since f T is a morphism, this gives T 0 (P,T) for r = 0. By [Sko89], T r (P,T) is a tree for all r, the action of F(S) on T 0 (P,T) descends to an action on T r (P,T) and for r sufficiently large, T r (P,T) = T. Furthermore, the path p T : [0,1] O(F(S)) by p T (0) = T and for p T (r) = T (1+r)/2r (P,T) for r > 0 is continuous.