# Full text of "Convex cocompact subgroups of mapping class groups"

## See other formats

ISSN 1364-0380 (on line) 1465-3060 (printed) 91 Qeometry & Topology Volume 6 (2002) 91-152 Published: 14 March 2002 Convex cocompact subgroups of mapping class groups Benson Fare Lee Mosher Department of Mathematics, University of Chicago 5734 University Ave, Chicago, U 60637, USA and Department of Mathematics and Computer Science Rutgers University, Newark, NJ 07102, USA Email: fcirb@math.uchicago.edu and mosher@eindromeda.rutgers.edu Abstract We develop a theory of convex cocompact subgroups of the mapping class group MCG of a closed, oriented surface S of genus at least 2, in terms of the action on Teichmiiller space. Given a subgroup G of MCG defining an extension 1 — TTi (S) -^Vc^G^l^we prove that if Vq is a word hyperbolic group then G is a convex cocompact subgroup of MCG. When G is free and convex cocompact, called a Schottky subgroup of MCG, the converse is true as well; a semidirect product of 7ri(S') by a free group G is therefore word hyperbolic if and only if G is a Schottky subgroup of MCG. The special case when G = Z follows from Thurston's hyperbolization theorem. Schottky subgroups exist in abundance: sufficiently high powers of any independent set of pseudo-Anosov mapping classes freely generate a Schottky subgroup. AMS Classification numbers Primary: 20F67, 20F65 Secondary: 57M07, 57S25 Keywords: Mapping class group, Schottky subgroup, cocompact subgroup, convexity, pseudo-Anosov Proposed: Walter Neumann Received: 20 October 2001 Seconded: Shigeyuki Morita, Robion Kirby Accepted: 20 February 2002 © Qeometry & Topology Vublications 92 Benson Farb and Lee Mosher 1 Introduction 1.1 Convex cocompact groups A convex cocompact subgroup of Isom(H") , the isometry group of hyperbolic n-space, is a discrete subgroup G < Isom(H"), with hmit set Ac C 9H"^, such that G acts cocompactly on the convex hull Hullc C H" of its limit set Aq. It follows that G is a word hyperbolic group with model geometry HuIIg and Gromov boundary Ac- Given any finitely generated, discrete subgroup G < Isom(H") , G is convex cocompact if and only if any orbit of G is a quasiconvex subset of H". Convex cocompact subgroups satisfy several useful properties: every infinite order element of G is loxodromic; Ac is the smallest nontrivial G-invariant closed subset of H = H'* U ; the action of G on \ Ac is properly discontinuous; assuming Ac 7^ , the stabilizer subgroup of Ac is a finite index supergroup of G, and it is the relative commensurator of G in Isom(H"). A Schottky group is a convex cocompact subgroup of Isom(H") which is free. Schottky subgroups of Isom(H") exist in abundance and can be constructed using the classical ping-pong argument, attributed to Klein: if </>i, . . . ,(pn are loxodromic elements whose axes have pairwisc disjoint cndpoints at infinity, then sufficiently high powers of (pi, . . . ,(t>n freely generate a Schottky group. ^ We shall investigate the notions of convex cocompact groups and Schottky groups in the context of Teichmiiller space. Given a closed, oriented surface S of genus > 2, the mapping class group MGG acts as the full isometry group of the Teichmiiller space T [45].^ This action extends to the Thurston compacti- fication T = TuPMJ^ [16]. Teichmiiller space is not Gromov hyperbolic [34], no matter what finite covolume, equivariant metric one picks [10], and yet it exhibits many aspects of a hyperbolic metric space [38] [32]. A general theory of limit sets of finitely generated subgroups of MGG is developed in [36]. In this paper we develop a theory of convex cocompact subgroups and Schottky subgroups of MGG acting on T, and we show that Schottky subgroups exist in abundance. We apply this theory to relate convex cocompactness of sub- groups of MGG with the large scale geometry of extensions of surface groups by subgroups of MGG. ^Thc term "Schottky group" sometimes refers explicitly to a subgroup of Isom(H") produced by the ping-pong argument, but the broader reference to free, convex cocom- pact subgroups has become common. ^In this paper, MCG includes orientation reversing mapping classes, and so repre- sents what is sometimes called the "extended" mapping class group. Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 93 Our first result establishes the concept of convex cocompactness for subgroups of MCG, by proving the equivalence of several properties: Theorem 1.1 (Characterizing convex cocompactness) Given a Gnitely gen- erated subgroup G < MCG, the following statements are equivalent: • Some orbit of G is quasiconvex in T. • Every orbit of G is quasiconvex in T . • G is word hyperbolic, and there is a G~equivariant embedding df : dG PMJ^ with image Aq such that the following properties hold: — Any two distinct points ^,rj E Aq are the ideal endpoints of a unique i > geodesic {^,ri) in T. — Let WHc be the "weak hull" of G , namely the union of the geodesies i > iCiV) J ^ V ^ ^G- Then the action of G on WHq is cocompact, and if f : G ^ WHg is any G-equivariant map then f is a quasi- isometry and the following map is continuous: f = fUdf: GUdG^T = TuPMJ^ Any such subgroup G is said to be convex cocompact. This theorem is proved in Section 3.3. A convex cocompact subgroup G < MCG shares many properties with convex cocompact subgroups of Isom(H") . Every infinite order element of G is pseudo- Anosov (Proposition 3.1). The limit set Aq is the smallest nontrivial closed subset of T invariant under the action of G, and the action of G on PM^F — Aq is properly discontinuous (Proposition 3.2); this depends on work of McCarthy and Papadoupolos [36]. The stabilizer of Aq is a finite index supergroup of G in MCG, and it is the relative commensurator of G in MCG (Corollary 3.3). A Schottky subgroup of MCG = Isom(T) is defined to be a convex cocompact subgroup which is free of finite rank. In Theorem 1.4 we prove that if 0i, . . . , are pseudo-Anosov elements of MCG whose axes have pairwise disjoint end- points in PAiJ^, then for all sufficiently large positive integers ai, . . . ,a„ the mapping classes (?!)"\ . . . freely generate a Schottky subgroup of MCG. Warning Our formulation of convex cocompactness in T is not as strong as in H"' . Although there is a general theory of limit sets of finitely generated subgroups of MCG [36], wc have no general theory of their convex hulls. Such a theory would be tricky, and unnecessary for our purposes. In particular, when Qeometry & Topology, Volume 6 (2002) 94 Benson Farb and Lee Mosher G is convex cocompact, we do not know whether there is a closed, convex, G-equivariant subset of T on which G acts cocompactly. One could attempt to construct such a subset by adding to WHg any geodesies with endpoints in WHg, then adding geodesies with endpoints in that set, etc, continuing transfinitely by adding geodesies and taking closures until the result stabilizes; however, there is no guarantee that G acts cocompactly on the result. 1.2 Surface group extensions There is a natural isomorphism of short exact sequences 1 ^ MS,P) -^-^ MCG{S,p) — - — ^ MCG{S) ^ 1 1 ^MS,P) > Aut(7ri(S',p)) ^Out(7ri(S',p)) > 1 where MCG{S,p) is the mapping class group of S punctured at the base point p. In the bottom sequence, the inclusion TTi{S,p) is obtained by identifying T^iiSjp) with its group of inner automorphisms, an injection since 7ri(S',p) is centerless. For each based loop £ in S, i{£) is the punctured mapping class which "pushes" the base point p around the loop £ (see Section 2.2 for the exact definition). The homomorphism q is the map which "forgets" the puncture p. Exactness of the top sequence is proved in [7]. The isomorphism MCG{S) « Out(7ri(S',p)) follows from work of Dehn-Nielsen [43], Baer [3], and Epstein [13]. As a consequence, either of the above sequences is natural for extensions of TTi{S) , in the following sense. For any group homomorphism G MGG{S) , by applying the fiber product construction to the homomorphisms MCG{S,p) G MCG{S) we obtain a group Fq and a commutative diagram of short exact sequences 1 ^7ri(5) ^Fg >G ^1 1 ^ TTl (5) ^ MCGiS, p) ^ MCG{S) 1 Note that we are suppressing the homomorphism G MCG{S) in the notation Fg- If G is free then the top sequence splits and we can write Fg = t^i{S) x Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 95 G, where again our notation suppresses a lift G — Aut(7ri(S')) of the given homomorphism G MCG{S) ^ Out(7ri(S')). Every group extension 1 7ri(S') ^ E ^ G ^ 1 arises from the above construction, because the given extension determines a homomorphism G — > Out(7ri(S')) « MCG{S) which in turn determines an extension 1 — > 7ri(S') — Vg ^ G ^ \ isomorphic to the given extension. When P is a cycUc subgroup of MGG, Thurston's hyperbohzation theorem for mapping tori (see, eg, [44]) shows that '7ri(S') x P is the fundamental group of a closed, hyperbolic 3-manifold if and only if P is a pseudo-Anosov subgroup. In particular, 7ri(S') xi P is a word hyperbolic group if and only if P is a convex cocompact subgroup of MGG. Our results about the extension groups Fg are aimed towards generalizing this statement as much as possible. The theme of these results is that the geometry of Fg is encoded in the geometry of the action of G on r. From [39] it follows that if Fg is word hyperbolic then G is word hyperbolic. Our next result gives much more precise information: Theorem 1.2 (Hyperbolic extension has convex cocompact quotient) If Fg is word hyperbolic then the homomorphism G — > MGG has hnite kernel and convex cocompact image. This theorem is proved in Section 5. Finiteness of the kernel of G ^ MGG is easy to prove, using the fact that 7ri(5) x K is a, subgroup of Fg- UK is infinite, then either it is a torsion group, or it has an infinite order element and so Fg has a Z © Z subgroup; in either case, Fg cannot be word hyperbolic. Because one can mod out by a finite kernel without affecting word hyperbolicity of the extension group, this brings into focus the extensions defined by inclusion of subgroups of MGG. We are particularly interested in free subgroups of MGG. A finite rank, free, convex cocompact subgroup is called a Schottky subgroup. For Schottky sub- groups we have a converse to Theorem 1.2, giving a complete characterization of word hyperbolic groups Fp when F < MGG is free: Theorem 1.3 (Surface- by-Schottky group has hyperbolic extension) If F is a finite rank, free subgroup of MGG then the extension group Fp = T^iiS) xi F is word hyperbolic if and only if F is a Schottky group. This is proved in Section 6. Some special cases of this theorem are immediate. It is not hard to see that tti (S) xi F has a Z © Z subgroup if and only if there Geometry & Topology, Volume 6 (2002) 96 Benson Farb and Lee Mosher exists a nontrivial element f & F which is not pseudo-Anosov. Such an element /, being infinite order, must be reducible. Assuming f £ F is nontrivial and reducible, the group 7ri(S') x F contains the subgroup 7ri{S) x (/) which is the fundamental group of a closed 3-manifold that contains an incompressible torus. Conversely, when 7ri(5) x F has a Z Z subgroup then that subgroup must map onto an infinite cyclic subgroup (/) C F whose action on 7ri(S') preserves a nontrivial conjugacy class, and so / is not pseudo-Anosov. Theorem 1.3 is therefore mainly about free, pseudo-Anosov subgroups of MCG (see Question 1.5 below). The abundance of word hyperbolic extensions of the form tti {S) x F was proved in [40]. It was shown by McCarthy [35] and Ivanov [23] that if 0i, . . . , are pseudo-Anosov elements of MCG which are pairwise independent, meaning that their axes have distinct endpoints in the Thurston boundary VAiJ^ , then sufficiently high powers of these elements freely generate a pseudo-Anosov sub- group F . The main result of [40] shows in addition that, after possibly making the powers higher, the group 7ri(S) x F is word hyperbolic. The nature of the free subgroups F < MCG produced in [40] was somewhat mysterious, but Theorems 1.2 and 1.3 clear up this mystery by characterizing the subgroups F using an intrinsic property, namely convex cocompactness. By combining [40] and Theorem 1.3, we immediately have the following result: Theorem 1.4 (Abundance of Schottky subgroups) If (pi,. . . ,(pn G MCG are pairwise independent pseudo-Anosov elements, then for all sufficiently large positive integers ai, . . . ,an the mapping classes , ■ ■ ■ , 0n" freely generate a Schottky subgroup F of MCG. Finally, we shall show in Section 7 that all of the above results generalize to the setting of closed hyperbolic 2-orbifolds. These generalized results find ap- plication in the results of [15] , as we now recall. 1.3 An application In the paper [15] we apply our theory of Schottky subgroups of MCG to inves- tigate the large-scale geometry of word hyperbolic surface-by-free groups: Theorem [15] Let F c MCG{S) be Schottky Then the group Tp = 711(5) x F is quasi-isometrically rigid in the strongest sense: • Tp embeds with finite index in its quasi-isometry group QI(ri?) . Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 97 It follows that: • Let H be any finitely generated group. If H is quasi-isometric to Tp, then there exists a Gnite normal subgroup N <] H such that H/N and Tp are abstractly commensurable. • The abstract commensurator group Comm(rF) is isomorphic to Ql{rp) , and can be computed explicitly. The computation of Comm(rF) ~ Ql(rir) goes as follows. Among all orb- ifold subcovers S ^ O there exists a unique minimal such subcovcr such that the subgroup F < MCG{S) descends isomorphically to a subgroup F' < MCG{0). The whole theory of Schottky groups extends to general closed hyperbolic orbifolds, as wc show in Section 7 of this paper. In particular, F' is a Schottky subgroup of MCG{0) . By Corollary 3.3 it follows that F' has finite index in its relative commensurator N < MCG{0) , which can be regarded as a virtual Schottky group. The inclusion N < MCG{0) determines a canonical extension 1 7ri(C) Tn ^ N ~> 1, and we show in [15] that the extension group r^v is isomorphic to Q^Fp). 1.4 Some questions Our results on convex cocompact and Schottky subgroups of MCG motivate several questions. Proposition 3.1 implies that if F is a Schottky subgroup of MCG then every nontrivial element of F is pseudo-Anosov. Question 1.5 Suppose F < MCG is a finite rank, free subgroup all of whose nontrivial elements are pseudo-Anosov. Is F convex cocompact? In other words, is F a Schottky group? A non-Schottky example F would be very interesting for the following reasons. There exist examples of infinite, finitely presented groups which are not word hyperbolic and whose solvable subgroups are all virtually cyclic, but all known examples fail to be of finite type; see for example [9]. If there were a non- Schottky subgroup F < MCG as in Question 1.5, then the group 7ri(S') x F would be of finite type (being the fundamental group of a compact aspherical 3-complex), it would not be word hyperbolic (since F is not Schottky), and every nontrivial solvable subgroup H < 7ri(S') x F would be infinite cyclic. To see why the latter holds, since vri(5) xi F is a torsion free subgroup of Qeometry & Topology, Volume 6 (2002) 98 Benson Farb and Lee Mosher MCG{S,p) it follows by [8] that the subgroup H is finite rank free abelian. Under the homomorphism H ^ F, the groups image (i7 F) < F and kernel(i? — > F) < Tri{S) each are free abelian of rank at most 1, and so it suffices to rule out the case where the image and kernel both have rank 1 . But in that case we would have a pseudo-Anosov element of MCG{S) which fixes the conjugacy class of some infinite order element of ttiS , a contradiction. Note that Question 1.5 has an analogue in the theory of Kleinian groups: if G is a discrete, cocompact subgroup of Isom(H^), is every free subgroup of G a Schottky subgroup? More generally, if G is a discrete, cofinite volume subgroup of Isom(H^), is every free loxodromic subgroup of G a Schottky group? The first question, at least, would follow from Simon's tame ends conjecture [11]. For a source of free, pseudo-Anosov subgroups on which to test question 1.5, consider Whittlesey's group [47], an infinite rank, free, normal, pseudo-Anosov subgroup of the mapping class group of a closed, oriented surface of genus 2. Question 1.6 Is every finitely generated subgroup of Wliittlesey's group a Scliottky group? Concerning non-free subgroups of MCG, note first that Question 1.5 can also be formulated for any finitely generated subgroup of MCG , though we have no examples of non-free pseudo-Anosov subgroups. This invites comparison with the situation in Isom(H") where it is known for any n > 2 that there exist convex cocompact subgroups which are not Schottky, indeed are not virtually Schottky. Question 1.7 Docs there exist a convex cocompact subgroup G < MCG wliich is not Schottky, nor is virtually Schottky? The converse to Theorem 1.2, while proved for free subgroups in Theorem 1.3, remains open in general. This issue becomes particularly interesting if Ques- tion 1.7 is answered affirmatively: Question 1.8 If G < MCG is convex cocompact, is the extension group Tq word hyperbolic? Surface subgroups of mapping class groups are interesting. Gonzalez- Diez and Harvey showed that MCG can contain the fundamental group of a closed, oriented surface of genus > 2 [19], but their construction always produces subgroups containing mapping classes that are not pseudo-Anosov. If questions 1.7 and 1.8 were true, it would raise the stakes on the fascinating question of whether there exist surface-by-surface word hyperbolic groups: Qeometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 99 Question 1.9 Does there exist a convex cocompact subgroup G < MCG isomorphic to the fundamental group of a closed, oriented surface Sg of genus g > 2? If so, is the surface-by-surface extension group Tq word hyperbohc? Misha Kapovich shows in [25] that when G is a surface group, the extension group Tq cannot be a lattice in Isom(CH^). 1.5 Sketches of proofs Although Teichmiiller space T is not hyperbolic in any reasonable sense [34], [10] , nevertheless it possesses interesting and useful hyperbolicity properties. To formulate these, recall that the action of MCG by isometries on T is smooth and properly discontinuous, with quotient orbifold M. = T /MCG called the moduli space of S. The action is not cocompact, and we define a subset A C T to be cobounded if its image under the universal covering map T — > Al has compact closure in M , equivalently there is a compact subset of T whose translates under Isom(T) cover A. In [38], Minsky proves (see Theorem 3.6 below) that if ^ is a cobounded geodesic in T then any projection T ^ £ that takes each point of T to a closest point on £ satisfies properties similar to a closest point projection from a (5-hyperbolic metric space onto a bi-infinite geodesic. This projection property is a key step in the proof of the Masur-Minsky theorem [32] that Harvey's curve complex is a 5-hyperbolic metric space. These results say intuitively that T exhibits hyperbolicity as long as one focusses only on cobounded aspects. Keeping this in mind, the tools of [38] and [32] can be used to prove Theorem 1.1 along the classical lines of the proof for subgroups of Isom(H"). The proof of Theorem 1.3, that 7ri(S') x: _F is word hyperbolic if F is Schottky, uses the Bestvina-Feighn combination theorem [6] . Consider a tree t on which F acts freely and cocompactly, and choose an F-equivariant mapping ^ : t — > T. Let H — s- T be the canonical hyperbolic plane bundle over Teichmiiller space. Pulling back via cf) we obtain a hyperbolic plane bundle vr: TCi t, and 7ri(S') X F acts properly discontinuously and cocompactly on Hi. This shows that Ht is a model geometry for the group 7ri(S') xi F, and in particular Ht is a 5-hyperbolic metric space if and only if 7ri(S') xi F is word hyperbolic. By the Bestvina-Feighn combination theorem [6] and its converse due to Ger- sten [18], hyperbolicity of Tit is equivalent to (5 -hyperbolicity of each "hyper- plane" He = 7r~^{£), where £ ranges over all the bi-infinite lines in t and 6 is independent of £. Geometry & Topology, Volume 6 (2002) 100 Benson Farb and Lee Mosher Recall that for each Teichmiiller geodesic g, the canonical marked Riemann surface bundle Sg over g carries a natural singular SOLV metric; the bundle Sg equipped with this metric is denoted . Lifting the metric to the universal cover Hg we obtain a singular SOLV space denoted Hg°^^ . When F is a Schottky group, convex cocompactness tells us that for each bi- infinite geodesic ^ in t, the map ^ — ^ T is a quasigeodesic and there is a unique Teichmiiller geodesic g within finite Hausdorff distance from (f)(i) . This feeds into Proposition 4.2, a basic construction principle for quasi-isometries which will be used several times in the paper. The conclusion is: Fact 1.10 The hyperplane Tii is uniformly quasi-isometric to the singular SOLV-space Hg^^"^, by a quasi-isometry which is a lift of a closest point map i^g. Uniform hyperbolicity of singular SOLV-spaces 7^^°^^, where y is a uniformly cobounded geodesic in T, is then easily checked by another application of the Bestvina-Feighn combination theorem, and Theorem 1.3 follows. For Theorem 1.2, we first outline the proof in the special case of a free subgroup of MCG. As noted above, using Gersten's converse to the Bestvina-Feighn combination theorem, word hyperbolicity of '7ri(5') x F implies uniform hyper- bolicity of the hyperplanes He- Now we use a result of Mosher [41], which shows that from uniform hyperbolicity of the hyperplanes it follows that the lines £ are uniform quasigeodesics in T, and each I has uniformly finite Hausdorff distance from some Teichmiiller geodesic g. Piecing together the geodesies g in T, one for each geodesic ^ in t, we obtain the data we need to prove that F is Schottky. The general proof of Theorem 1.2 follows the same outline, except that we cannot apply Gersten's converse to the Bestvina-Feighn combination theorem. That result applies only to the setting of groups acting on trees, not to the setting of Theorem 1.2 where Vq acts on the Cayley graph of G. To handle this problem we need a new idea: a generalization of Gersten's converse to the Bestvina-Feighn combination theorem, which holds in a much broader setting. This generalization is contained in Lemma 5.2. The basis of this result is an analogy between the "flaring property" of Bestvina-Feighn and the divergence of geodesies in a word hyperbolic group [12]. Acknowledgements We are grateful to the referee for a thorough reading of the paper, and for making numerous useful comments. Both authors are supported in part by the National Science Foundation. Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 101 2 Background 2.1 Coeirse language Quasi-isometries and uniformly proper maps Given a metric space X and two subsets A,BgX, the Hausdorjf distance dHaus(^) B) is the infimum of all real numbers r such that each point of A is within distance r of a point of B, and vice versa. A quasi-isometric embedding between two metric spaces X,Y is a map / : X — > Y such that for some K > 1, C > , we have ^d{x, y)-C< d{fx, fy) < Kd{x, y)+C for each x,y & X . To refer to the constants we say that / is a if, C-quasi- isometric embedding. For example, a quasigeodesic embedding R ^ X is called a quasigeodesic line in X. We also speak of quasigeodesic rays or segments with the domain is a half-line or a finite segment, respectively. Since every map of a segment is a quasi-isomctry, it usually behooves one to include the constants and speak about a (K, C) -quasi- isometric segment. A quasi- is ometry between two metric spaces X,Y is a map f : X Y which, for some K > 1, C > is a K,C quasi-isometry and has the property that image(/) has Hausdorff distance < C from Y . Every quasi-isometry f : X ^ Y has a coarse inverse, which is a quasi-isometry f:Y^X such that fof: X — ^ X is a bounded distance in the sup norm from Idx , and similarly for fof : y — > Y; the sup norm bounds and the quasi-isometry constants of / depend only on the quasi-isometry constants of /. More general than a quasi-isometric embedding is a uniformly proper embedding f:X Y, which means that there exists K > 1, C > 0, and a function r: [0, go) — [0, go) satisfying r{t) — oo as t — oo, such that r{d{x, y)) < d{fx, fy) < Kd{x, y) + C for each x,y e X . Geodesic and quasigeodesic metric spaces A metric space is proper if closed balls are compact. A metric d on a space X is called a path metric if for any x,y £ X the distance d{x, y) is the infimum of the path lengths of rectifiable paths between x and ?/, and d is called a geodesic metric if d{x,y) Geometry & Topology, Volume 6 (2002) 102 Benson Farb and Lee Mosher equals the length of some rectifiable path between x and y . The following fact is an immediate consequence of the Ascoli-Arzela theorem: Fact 2.1 A compact path metric space is a geodesic metric space. More gen- erally, a proper path metric is a geodesic metric. □ The Ascoli-Arzela theorem also shows that for any proper geodesic metric space X , every path homotopy class contains a shortest path. This implies that the metric on X lifts to a geodesic metric on any covering space of X . A metric space X is called a quasigeodesic metric space if there exists constants A, e such that for any x,y € X there exists an interval [a, 6] C R and a A, e quasigeodesic embedding a: [a,b] ^ X such that cr(a) = x and a{b) = y. For example, if F is a geodesic metric space and X is a subset of Y such that dB.aus{X,Y) < oo then X is a quasigeodesic metric space. The fundamental theorem of geometric group theory, first known to Efremovich, to Schwarzc, and to Milnor, can be given a general formulation as follows. Let X be a proper, quasigeodesic metric space, and let the group G act on X properly discontinuously and cocompactly, by an action denoted {g,x) i-^ g ■ x. Then G is finitely generated, and for any base point xq E X the map G ^ X defined hy g i-^ g ■ xq is a quasi-isometry between the word metric on G and the metric space X. Uniform families of quasi-isometries The next lemma says a family of geodesic metrics which is "compact" in a suitable sense has the property that any two metrics in the family are uniformly quasi-isometric, with respect to the identity map. Given a compact space X , let M{X) denote the space of metrics generating the topology of X , regarded as a subspace of [0, oo)"'^^"'^ with the topology of uniform convergence. Lemma 2.2 Let X be a compact, path connected space with universal cover X . Let D C AI{X) be a compact family of geodesic metrics. Let D be the set of lifted metrics on X . Then there exist K > 1, C > such that for any d, d' e D the identity map on X is a K,C quasi-isometry between {X, d) and {X,d'). Proof By compactness of D, the metric spaces X^ have a uniform injectivity radius — that is, there exists e > such that for each d & D every homotopically Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 103 nontrivial closed curve in has length > 4e, and it follows that every closed e ball in lifts isometrically to X^- Let P C X x X be the set of pairs {x,y) E X X X such that for some d & D we have d{x, y) < e. Evidently tti (X) acts cocompactly on P , and so we have a finite supremum A = sup{d{x, y) \ d £ D and {x, y) G P} Given d G D and x,y G X, choose a d-geodesic 7 from x to y and let X = xo,xi,... ,Xn-i,Xn = y he a monotonic sequence along 7 such that d{xi-i,Xi) = e for i = 1, . . . ,n — 1 and d{xn-i,Xn) < e. For any d' E D it follows that: c^'l^;, y) < An = A Setting K = ^ and C = A the lemma follows. □ Hyperbolic metric spaces A geodesic metric space X is hyperbolic if there exists (5 > such that for any x,y,z E X and any geodesies xy, yz, 'zx, any point on xy has distance < 5 from some point on yzUzx. A finitely generated group is word hyperbolic if the Cayley graph of some (any) finite generating set, equipped with the geodesic metric making each edge of length 1, is a hyperbolic metric space. If X is (5-hyperbolic, then for any A > 1 , e > there exists A , depending only on S,X,e, such that the following hold: for any x,y £ X , any A, e quasigeodesic segment between x and y has Hausdorff distance < A from any geodesic seg- ment between x and y; for any x & X , any A, e quasigeodesic ray starting at x has Hausdorff distance < A from some geodesic ray starting at x; and any A, e quasigeodesic line in X has Hausdorff distance < A from some geodesic line in X. The boundary of X , denoted dX , is the set of coarse equivalence classes of geodesic rays in X, where two rays are coarsely equivalent if they have finite Hausdorff distance. For any ^ G dX and xq & X , there is a ray based at xq representing ^; we denote such a ray [xq,^). For any ^ 7^ G dX there is a geodesic line £ in X such that any point on £ divides it into two rays, one representing ^ and the other representing rj. Assuming X is proper, there is a compact topology on X U dX in which X is dense, which is characterized by the following property: a sequence G XUdX > converges to ^ £ dX if and only if, for any base point p G X, if [p, ^i) denotes Geometry & Topology, Volume 6 (2002) dix, y) A~, . <-d{x,y) + A 104 Benson Farb and Lee Mosher either a segment from p to G X , or a ray from p with ideal endpoint G dX , then any subsequential hmit of the sequence \p, is a ray with ideal endpoint ^ . It follows that any quasi-isometric embedding between (5-hyperbolic geodesic metric spaces extends to a continuous embedding of boundaries. In particular, if X is hyperbolic then the action of Isom(X) on X extends continuously to an action on X U dX . The following fundamental fact is easily proved by considering what happens to geodesies in a 5-hyperbolic metric space under a quasi-isometry. Lemma 2.3 For allS>0,K>l,C>0 there exists A> such that the following holds. If X, Y are two S -hyperbolic metric spaces and if f,g: X ^ Y are two K, C quasi-isometries such that df = dg: dX dY , then: dsupif, g) = sup d{f{x),g{x)) < A □ xex 2.2 Teichmiiller space and the Thurston boundary Fix once and for all a closed, oriented surface S of genus g > 2. Let C be the set of isotopy classes of nontrivial simple closed curves on S . The fundamental notation for the paper is as follows. Let T be the Teichmiiller space of S . Let be the space of measured foliations on S , and let FMJ^ be the space of projective measured foliations on S , with projectivization map P : AiJ^ — > FAdJ-^. The Thurston compactification of Teichmiiller space is 7" = TUPA^.;^. Let MCG be the mapping class group of S, and let M = T /MCG be the moduli space of S. Definitions of these objects are all recalled below. The Teichmiiller space T is the set of hyperbolic structures on S modulo iso- topy, with the structure of a smooth manifold diffeomorphic to R^^^® given by Fenchel-Nielsen coordinates. The Riemann mapping theorem associates to each conformal structure on S a unique hyperbolic structure in that confor- mal class, and hence we may naturally identify T with the set of conformal structures on S modulo isotopy. Given a conformal structure or a hyperbolic structure a , we will often confuse a with its isotopy class by writing a . There is a length pairing T x C ^ R-|- which associates to each a ^ T , C ^ C the length of the unique simple closed geodesic on the hyperbolic surface a in the isotopy class C. We obtain a map T — [0, oo)*^ which is an embedding with image homeomorphic to an open ball of dimension 6(/ — 6. Moreover, under projectivization [0, oo)^ — > P[0, oo)^, T embeds in P[0, oo)^ with precompact image. Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 105 Thurston's boundary A measured foliation ^ on S is a foliation with finitely many singularities equipped with a positive transverse Borel measure, with the property that for each singularity s there exists n > 3 such that in a neighborhood of s the foliation is modelled on the horizontal measured foliation of the quadratic differential z"'~^dz^ in the complex plane. A saddle connection of .7^ is a leaf segment connecting two distinct singularities; col- lapsing a saddle connection to a point yields another measured foliation on S . The set of measured foliations on S modulo the equivalence relation generated by isotopy and saddle collapse is denoted MT. Given a measured foliation J^, its equivalence class is denoted G A4J-; elements of AiJ- will often be represented by the letters X,Y, Z . For each measured foliation JT, there is a function Ij^: C [0, oo) defined as follows. Given a simple closed curve c, we may pull back the transverse measure on to obtain a measure on c, and then integrate over c to obtain a number j^J' ■ Define i-j^{c) = i{J-,c) to be the infimum of J^, J" as c' ranges over the isotopy class of c. The function ijr is well-defined up to equivalence, thereby defining an embedding AiJ^ — > [0, oo)'' whose image is homeomorphic to R69-6_|o}. Given a measured foliation J^, multiplying the transverse measure by a positive scalar r defines a measured foliation denoted r ■ J^, yielding a positive scalar multiplication operation R x AiJ^ AiJ^. With respect to the equivalence relation .F ~ r ■ J^, r > 0, the set of equivalence classes is denoted PMJ^ and the projection is denoted P: — > PMT. We obtain an embedding PMT P[0,oo)<^ whose image is homeomorphic to a sphere of dimension 6g — 7. We often use the letters ^,rf,( to represent points of PMJ^. Thurston's compactification theorem [16] says, by embedding into P[0, 00)*^, that there is a homeomorphism of triples: We will also need the standard embedding C AiJ-, defined on [c] as follows. Take an embedded annulus A C S foliated by circles in the isotopy class [c] , and assign total transverse measure 1 to the annulus. Choose a deformation retraction of each component of the closure of S — A onto a finite 1-complex, and extend to a map f : S ^ S homotopic to the identity and which is an embedding on int(^). The measured foliation on A pushes forward under / to the desired measured foliation on S, giving a well-defined point in MJ^ depending only on [c] . The intersection number AiJ- x C ^ ' [0, 00) extends continuously to AiJ- x Geometry & Topology, Volume 6 (2002) 106 Benson Farb and Lee Mosher MJ^ — '-^ [0, co) . This intersection number is most efficaciously defined in terms of measured geodesic laminations. Marked surfaces Having fixed once and for all the surface S, a marked surface is a pair (F, (p) where i*" is a surface and (p: S ^ F is a homeomorphism. Thus we may speak about a marked hyperbolic surface, a marked Riemann surface, a marked measured foliation on a surface, etc. Given a marked hyperbolic surface (j): S ^ F, pulling back via ^ determines a hyperbolic structure on S and a point of t. Two marked hyperbolic surfaces (f>: S ^ F and <p' : S ^ F' give the same element of T if and only if they are equivalent in the following sense: there exists an isometry h: F ^ F' such that o h o (f>: S ^ S is isotopic to the identity. In this manner, we can identify the collection of marked hyperbolic surfaces up to equivalence with the Teichmiillcr space T of 5*. This allows us the freedom of representing a point of T by a hyperbolic structure on some other surface F , assuming implicitly that we have a marking cp: S ^ F. The same discussion holds for marked Riemann surfaces, marked measured foliations on surfaces, etc. Given two marked surfaces (f): S ^ F, cf)' : S ^ F' , a marked map is a homeo- morphism ip: F ^ F' such that o (f) is isotopic to (p' . Mapping class groups and moduli space Let Homeo(S') be the group of homeomorphisms of S , let Homeoo(5') be the normal subgroup consisting of homeomorphisms isotopic to the identity, and let MCG = MCG{S) = Homeo(iS')/Homeoo(S') be the mapping class group of S. Pushing a hyperbolic structure on S forward via an clement of Homeo(S') gives a well-defined action of MCG on T. This action is smooth and properly discontinuous but not cocompact. It follows that the moduli space M. = T/MCG is a smooth, noncompact orbifold with fundamental group MCG and universal covering space T . Let Homeo(S',p) be the group of homeomorphisms of S preserving a base point p, let Homeoo (S*, p) be the normal subgroup consisting of those home- omorphisms which are isotopic to the identity leaving p stationary, and let MCG{S,p) = Homeo(5,p)/IIomeoo(5',p). Recall the short exact sequence: 1 ^ TTiiS,p) ^MCG{S,p) ^MCG{S) -> 1 The map q is the map which "forgets" the puncture p . To define the map t , for each closed loop i: [0, 1] ^ 5 based at p, choose numbers = xq < xi < . . . < Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 107 Xn = I and embedded open balls Bi, Bn C S so that £[xi-i,Xi] C -Bj for i = I, . . . ,n, and let VTj : S" ^ S" be a homeomorphism which is the identity on S — Bi and such that Tri{£{xi-i)) = i{xi) . Then l{£) is defined to be the isotopy class rel p of the homeomorphism 7r„ o Tin-i o • • • o tti : {S,p) iS,p) , which we say is obtained by "pushing" the point p around the loop £. The mapping class L{i) is well-defined independent of the choices made, and independent of the choice of £ in its path homotopy class. When £ is simple, l{£) may also be described as the composition of opposite Dehn twists on the two boundary components of a regular neighborhood of £. For details see [7]. As noted in the introduction, by the Dehn-Nielsen-Baer-Epstein theorem, the above sequence is naturally isomorphic to the sequence 1 ^ 7ri{S,p) ^ Aut(7ri(5,p)) ^ Out(7ri(S,p)) ^ 1 Canonical bundles Over the Tcichmiillcr space T of S there is a canonical marked hyperbolic surface bundle S ^ T, defined as follows. Topologically S = S X T , with the obvious marking S ^ S x a = ioi each a G 7". As a varies over T, one can assign a hyperbolic structure on S in the class of a, varying continuously in the C°° topology on Riemannian metrics; this follows from the description of Fcnchel-Nielsen coordinates. It follows that on each fiber Sa of S there is a hyperbolic structure which varies continuously in a. Note that by the Riemann mapping theorem we can also think of S as the canonical marked Riemann surface bundle over T. The action of MCG on T lifts uniquely to an action on S, such that for each fiber Sa and each [h] G MCG the map [h] is an isometry, and the map is in the mapping class [h] . The universal cover of S is called the canonical -bundle over T, denoted 7i ^ T . There is a fibration preserving, isometric action of the once-punctured mapping class group MCG{S,p) on the total space 7i such that the quotient action of MCG{S,p) on S has kernel Tri{S,p), and corresponds to the given action of MCG{S) = MCG{S,p)/Tri{S,p) on S. Also, the action of 7ri{S,p) on any fiber of TC is conjugate to the action on the universal cover S by deck transformations. Bers proved in [4] that W is a Teichmiiller space in its own Geometry & Topology, Volume 6 (2002) 108 Benson Farb and Lee Mosher right: there is an MCG{S,p) equivariant homeomorphism between Ti and the Teichmiiller space of the once-punctured surface S — p. The tangent bundle TS has a smooth 2-dimcnsional vertical sub-bundle T^S consisting of the tangent planes to fibers of the fibration S ^ T . A connection on the bundle 5 — > T is a smooth codimension-2 sub-bundle of TS complemen- tary to TyS. The existence of an MCG-equivariant connection on S can be derived following standard methods, as follows. Choose a locally finite, equiv- ariant open cover of T, and an equivariant partition of unity dominated by this cover. For each MCG-orbit of this cover, choose a representative U G T of this orbit, and choose a linear retraction TSu — TySu- Pushing these retrac- tions around by the action of MCG and taking a linear combination using the partition of unity, we obtain an equivariant linear retraction TS — > TyS , whose kernel is the desired connection. By lifting to H we obtain a connection on the bundle 7i —> T , equivariant with respect to the action of the group MCG(S,p) . Notation Given any subset ^ C T, or more generally any continuous map A T, hy pulling back the bundle S ^ T we obtain a bundle Sa ^ A, as shown in the following diagram: Sa >5 A — Similarly, the pullback of the bundle H ^ T is denoted Ha A. Quadratic differentials Given a conformal structure a on S , & quadratic differenti,al q on S^ assigns to each conformal coordinate z an expression of the form q[z)dz'^ where q{z) is a complex valued function on the domain of the coordinate system, and q(z) I -— I = q(w), for overlapping coordinates z,w. \dw J We shall always assume that the functions q{z) are holomorphic, in other words, our quadratic differentials will always be "holomorphic" quadratic differentials. A quadratic differential q is trivial if q{z) is always the zero function. Given a nontrivial quadratic differential q on S^ , & point p & S^ is a zero of q in one coordinate if and only if it is a zero in any coordinate; also, the order of Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 109 the zero is well-defined. If p is not a zero then there is a coordinate z near p, unique up to multiplication by ±1, such that p corresponds to the origin and such that q{z) = 1; this is called a regular canonical coordinate. If p is a zero of order n > 1 then up to multiplication by the (n + 2)^^^ roots of unity there exists a unique coordinate z in which p corresponds to the origin and such that q(z) = ; this is called a singular canonical coordinate. There is a well-defined singular Euclidean metric \q{z) \ \dz\^ on S , which in any regular canonical coordinate z = x + iy takes the form dx^ + dy^ . In any singular canonical coordinate this metric has finite area, and so the total area of S in this singular Euclidean metric is finite, denoted ||g|| . We say that q is normalized if ||g|| = 1. By the Riemann-Roch theorem, the quadratic differentials on S,j form a com- plex vector space QD^ of complex dimension 3(7 — 3, and these vector spaces fit together, one for each a G T, to form a complex vector bundle over T de- noted QD T . Teichmxiller space has a complex structure whose cotangent bundle is canonically isomorphic to the bundle QD. The Teichmiiller metric on T induces a Finsler metric on the (real) tangent bundle of T, and the norm \\q\\ is dual to this metric. The normalized quadratic differentials form a sphere bundle QD^ — > T of real dimension 6g — 7 embedded in QD. Corresponding to each quadratic differential q on there is a pair of measured foliations, the horizontal foliation J^x{q) and the vertical foliation J^yiq). In a regular canonical coordinate z = x + iy, the leaves of J^x (?) Sire parallel to the X-axis and have transverse measure \dy\, and the leaves of J^y{q) arc parallel to the y-axis and have transverse measure \dx\. The foliations JF^.(g), J-y{q) have the zero set of q as their common singularity set, and at each zero of order n both have an [n + 2) -pronged singularity, locally modelled on the singularity at the origin of the horizontal and vertical measured foliations of z'^dz^ . Conversely, consider a transverse pair of measured foliations {!Fx,Ty) on S which means that Tx^J'y have the same singular set, are transverse at all regular points, and at each singularity s there is a number n > 3 such that Tx and J-y are locally modelled on the horizontal and vertical measured foliations of z'^'''^dz'^ . Associated to the pair Tx-,^y there are a conformal structure and a quadratic differential defined as follows. Near each regular point, there is an oriented coordinate z = x + iy in which J^x is the horizontal foliation with transverse measure \dy\, and J^y is the vertical foliation with transverse measure \dx\. These regular coordinates have conformal overlap. Near any singularity s, at which J^xi ^y are locally modelled on the the horizontal and vertical foliations of z'^dz'^ , the coordinate z has conformal overlap with any regular coordinate. We therefore obtain a conformal structure a{J^xi^y) on S, on Geometry & Topology, Volume 6 (2002) 110 Benson Farb and Lee Mosher which we have a quadratic differential q{J^x,^y) defined in regular coordinates by dz'^ . A pair of measured foliations (X, Y) G A4T{F) x M.J^{F) is said to jointly fill the surface F if, for every Z G MT{F), either i{X,Z) / or i{Y,Z) / 0. This condition is invariant under positive scalar multiplication on AiJ^(F) , and so joint filling is well-defined for a pair of points in PM.J^{F) . A basic fact is that a pair X,Y e M.T{F) jointly fills F if and only if there exist a transverse pair of measured foliations Tx^J'y representing X^Y\ moreover, such a pair Txi is unique up to joint isotopy, meaning that for any other transverse pair J^xj^y representing X,Y respectively, there exists h G Homeoo(<S') such that J^'^ = h{Fx)i = h{Ty). These facts may be proved by passing back and forth between measured geodesic laminations and measured foliations. By uniqueness up to joint isotopy as just described, it follows that for each jointly filling pair (X, Y) G MJ-{F) x A4J-{F) there is a conformal structure a{J^x,^y) and quadratic differential q{J-'x,^y) on a{X,Y), well-defined up to isotopy independent of the choice of a transverse pair TxiJ^y representing X, Y . We thus have a well-defined point a{X, Y) G T{F) and a well-defined element q{X,Y)eqD,^x,Y)nF)- Geodesies and a metric on T We shall describe geodesic lines in T follow- ing [17] and [21]; of course everything depends on Teichmiiller's theorem (see eg, [1] or [22]). Let TV C M!F x MJ^ denote the set of jointly filling pairs, and let PJ^V be the image of J^V under the product of projection maps P x P : MJ^ x MJ^ — PMJ" X PMJ^. Associated to each jointly filling pair {^,rj) G fJ-V we associate a Teichmiiller i > line (^,T/), following [17]. Choosing a transverse pair of measured foliations Tx^Ty representing ^,r/ respectively, we obtain a parameterized Teichmiiller geodesic given by the map t (T(e~*.?^^;, e*jFy) ; it follows from Teichmiiller's theorem that this map is an embedding R — > T. Uniqueness of J^xi^y up to joint isotopy and positive scalar multiplication imply that the map t ^ a{e'^^J-x, e^Ty) is well-defined up to translation of the t-paramctcr, as is easily checked. Thus, the image of this map is well defined and is denoted (^, r/) ; in ad- dition, parameter difference between points on the line is well-defined, and there is a well-defined orientation. The positive direction of the geodesic is defined to be the point r] = FJ^y G PAiJ-', the projective class of the vertical measured foliation; the negative direction is the point ^ = PJ^x £ PMJ^- Note that as Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 111 t +00 the vertical measured foliation becomes "exponentially thicker" and so dominates over the horizontal foliation which becomes "exponentially thinner" , a useful mnemonic for remembering which direction is which. Teichmiiller's theorem says that any two distinct points of T lie on a unique Teichmiiller line: for any a ^ t £ T there exists a unique pair r/) G PJ-V < > such that cr, T G 77) . Moreover, if d{a, r) is the parameter difference between a and r along this geodesic, then d is a metric on T, called the Teichmiiller metric. In particular, each line 77) is, indeed, a geodesic for the Teichmiiller i > metric. It is also true that the segment [a, r] C rj) is the unique geodesic seg- ment connecting a to r, and hence geodesic segments are uniquely extensible. Thus we obtain a 1-1 correspondence between oriented geodesic segments and the set T X T. Also, every bi-infinite geodesic line in T is uniquely expressible in the form (,^, 77) , and so we obtain a 1-1 correspondence between oriented geodesic lines and the set PjFP C PMJ^ x PA^JF. There is a also 1-1 correspondence between geodesic rays in T and the set T X for any a G T and rj £ PAiJ^ there is a unique geodesic ray, denoted [a, rj) , whose endpoint is u and whose direction is 77 G PM.T , and every geodesic ray has this form. This is an immediate consequence of the Hubbard- Masur theorem [21], which says that for each o" G T the map QD^ — >■ M.J- taking g 7^ G QD^. to [J'y{q)] is a homcomorphism. Throughout the paper, the term "geodesic" will refer to any geodesic segment, ray, or line in T . Geodesies in T are uniquely extendable: any geodesic segment or ray is contained in a unique geodesic line. Since T is a complete metric space, an argument using the Ascoli-Arzela theorem shows that any sequence of geodesies, each element of which intersects a given bounded subset of T, has a subsequence converging pointwise to a geodesic. By unique extendability of geodesies it follows that T is a proper, geodesic metric space. Prom the definitions it follows that the action of MCG on T is isometric, and so the metric on T descends to a proper, geodesic metric on M=T/MCG. > The reader is cautioned that a geodesic ray [a, 77) is not known to converge in T to its direction 77 G PMJ^. However, consider the case where 77 is uniquely ergodic, which means that for any measured foliation T representing 77, every transverse measure on the underlying singular foliation of is a scalar multiple of the given measure on .7^. In this case the ray [a, 77) does converge to r/, as is proved by Masur [30] , and so in this situation the direction 77 is also called the end or endpoint of the ray. Geometry & Topology, Volume 6 (2002) 112 Benson Farb and Lee Mosher Cobounded geodesies in T A subset A c T is cobounded if the image of A under the projection T ^ M is a bounded subset of Ai; equivalently, there is a bounded subset of T whose translates by the action of MCG cover A. If the bounded set B C M contains the projected image of A then we also say that A is B -cobounded. Since At is a proper metric space it follows that A is cobounded in T if and only if A is "co-precompact" , meaning that the projection of ^ to has compact closure. One common gauge for coboundedness, as noted by Mumford [42], is the injec- tivity radius of a hyperbolic structure, or to put it another way, the length £{a) of the shortest closed geodesic in a hyperbolic structure a For each e > the "e-thick subset" of T, namely the set % = {a e T \ i{a) > e}, is an MCG equivariant subset of T projecting to a compact subset of ^A , and as e ^ this gives an exhaustion of M. by compact sets. A subset of T is therefore cobounded if and only if it is contained in the e-thick subset of T for some e > 0. Extremal length, rather than hyperbolic length, is used to obtain another com- mon gauge of coboundedness, and is comparable to the length of the shortest geodesic by Maskit's work [27]. We rarely use any particular gauge for coboundedness. Instead, the primary way in which we use coboundedness is in carrying out compactness arguments over closed, bounded subsets. For this reason we rarely refer to any gauge, instead sticking with coboundedness as the more primitive mathematical concept. > One important fact we need is that if p = [a, rj) is a cobounded geodesic ray in Teichmiiller space then p converges to r] in Thurston's compactification T = TU PMJ^. This follows from two theorems of Masur. First, since p is cobounded, the direction rj G FMJ- is uniquely ergodic; this result, proved in [29], was later sharpened in [31] to show that if r] is not uniquely ergodic then the > projection of [a, rj) to moduli space leaves every compact subset. Second, when rj is uniquely ergodic, any ray with direction 77 converges to rj in Thurston's compactification. This is a small part of a Masur's Two Boundaries Theorem [30] , concerning relations between the Teichmiiller boundary and the Thurston boundary of T (we will use the full power of this theorem in the proof of Theorem 1.1). The following result is essentially a consequence of [38] : ^Also called the "systole" in the differential geometry literature. Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 113 Lemma 2.4 (End Uniqueness) If [cr, [r, ry) are two cobounded rays in T i > < > which have finite Hausdorff distance in T then ^ = t]. If ^'), {"Uif]') are two cobounded lines in T which have hnite Hausdorff distance then, up to relabelling the ends of one of the lines, we have ^ = rj and ^' = rj' , and so Proof For the proof wc review briefly notions of extremal length, in the clas- sical setting of simple closed curves, as well as Kerckhoff's extension to the setting of measured foliations [26]. Recall that for any conformal structure on an open annulus A there is a unique Euclidean annulus of the form x (0, M) conformally equivalent to A , with M G R+ U {oo}; the modulus of A, denoted M{A), is defined to be the number M . For any Riemann surface S^j and any isotopy class of simple closed curves [c] G C, the extremal length ^cxt(o', [c]) is the infimum of 1/M{A) taken over all annuli A C F whose core is in the isotopy class [c] . Kcrckoff proved [26] that the function £ext • ^ ' C) — > (0, oo) defined by £cxt{(^- r[c]) ^ r£ext{(^, [c]) extends continuously to a function 4xt '■ ^ x [0, oo). Moreover, for any transverse pair of measured foliations Tx^Ty with associated conformal structure a = a{J-x,J'y) and quadratic differential q = Q{J^x,J^y), we have Given X £ MJ- , the extremal length horohall based at X is defined to be H{X) = {a \ lcxt{cr,X) < 1}. Note for example that, setting ^ = PX , for i > every rj G PAiT the extremal length of X at points of (rj,^) decreases strictly monotonically to zero as the point moves towards ^, and so every Teichmiiller geodesic with positive direction PX eventually enters H{X) in the positive direction and, once in, never leaves. Given ^ G PAiJ^, there is a one parameter family of extremal length horoballs based at ^, namely H(X) for all X G A4J-^ such that PX = ^. > > For the first sentence of the theorem, consider two geodesic rays [a, 4) , [t, rj) such that ^ ^ rf e PMJ^. Pick any extremal length horoball H based at rj. The proof of Theorem 4.3 of [38] shows that H n [a, ^) is bounded. However, > > > H n [r, 77) is an infinite subray of [r, 77) , and moreover as a point p G [r, 7;) travels to infinity in [r, rf) the horoball H contains a larger and larger ball in T > > centered on p. It follows that [o", ^) and [r, 77) have infinite Hausdorff distance in r. Geometry & Topology, Volume 6 (2002) 114 Benson Farb and Lee Mosher The second sentence follows from the first, by dividing each line into two rays. □ Remark Combining results of Masur mentioned above, one can show that even more is true: two cobounded geodesic rays which have finite Hausdorff distance are asymptotic, meaning that as they go to oo , the distance between the rays approaches zero. To see why, as mentioned earlier Masur proves that if [a, rj) is cobounded then 77 is uniquely ergodic. Furthermore, two rays [a, ry) , > [r, T]) with uniquely ergodic endpoint 77 are asymptotic, according to [28] . 2.3 Singular SOLV spaces Consider a geodesic g = {^,r]) in T, and let Sg ^ g he the canonical marked Riemann surface bundle over g , obtained by pulling back the canonical marked Ricmann surface bundle S ^ T . Topologically we identify Sg = S x g . Choos- ing a transverse pair of measured foliations J-x,^y representing ^,rj respec- tively, we have g{t) = a{e~^J^xi&^^y)- Let \dy\ be the transverse measure on the horizontal measured foliation J^x and let \dx\ be the transverse measure on the vertical measured foliation !Fy. We may assume that the pair J^xi^y is normalized, meaning that the Euclidean area equals 1 : \\q{Tx,:Fy)\\= ! \dx\x\dy\ = l Js and hence for all t G R the pair e~^Tx-,^J^y is normalized: \q{e-^TxjTy)\ = I \e*dx\x \e~*dy\ = 1 s Note that the singular Euclidean metric on each fiber , may be expressed as dsl = e^* \dxf + Define the singular SOLV metric on Sg to be the singular Riemannian metric given by the formula: dsl = I'^^l^ + \dy\'^ + '^^^ We use the notation <S|°^^ to denote Sg equipped with this metric. The univer- sal cover of 5^ is the canonical Poincare disc bundle TCg over g, and lifting the singular SOLV metric from to Tig we obtain a singular SOLV space denoted ^soLv_ rpj^g singular locus of = S x g is the union of the singular lines Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 115 s X g, one for each singularity s of the pair Tx^^y Away from the singular lines, and 'W'^^^ are locally modelled on 3-dimensional SOLV-geometry. On each singular line the metric is locally modelled by gluing together several copies of the half-plane y > in SOLV-geometry. 2.4 Comparing hyperbolic and singular Euclidean structures Given cr G T, the Riemann surface Sa has several important metrics in its conformal class: a unique hyperbolic metric; and one singular Euclidean metric for each q G QD(t- These lift to the universal cover l-i^. Given a, r G T, if each Riemann surface »So-, St is given either its unique hyperbolic metric or one of its singular Euclidean metrics, then for any marked map ^: Sr ^ each lift : Ti^ ~* is a quasi-isometry. We are interested in how the quasi- isometry constants of ^ compare to the Teichmiiller distance d((T, r) , although we need only the crudest estimates. Proposition 2.5 shows how to bound the quasi-isometry constants in terms of d((T, r) . Part 1 of this proposition was first proved by Minsky in [37], Lemma 3.3; we give a quicker proof using Lemma 2.2. Proposition 2.5 For each bounded subset B C M and each r > there exists K>1,C>0,^>0 such that the following hold: (1) Suppose that a,T G T are each B-cobounded and d{a,T) < r. Let f^^ : Sa ^ Sr be the canonical marked map Sa = Sxa^SxT = Sr- If we impose on S„ and S-j- either the hyperbolic metric or the singular Euclidean metric associated to some normalized quadratic differential, then any lift far '■ 'Ha — of f^T is a K,C quasi-isometry. (2) Let (Tj G T, i = 1, 2, 3, be B cobounded and have pairwisc distances < r, let metrics be imposed on S^^ as above, and let fij : — > S^-^ , etc. be the marked maps as above, with K, C -quasi-isometric lifts fij: Ti-aj ■ If /i3 is the unique lift of /13 such that 9/23 o dfi2 = dfi3, then if 23 ° /l2,/l3) < A. Proof Part (1) is an easy consequence of Lemma 2.2, as follows. Choose a com- pact subset A C T whose image in covers B and such that over any point of B there exists a point a G A such that B't{(T, r) d A. It follows that the points (J, r in (1) may be translated to lie in A. Identifying Sj\_ diffeomorphically with Geometry & Topology, Volume 6 (2002) 116 Benson Farb and Lee Mosher S X A, compactness of A produces a compact family of hyperbolic metrics on S , and compactness of the restriction of QD^ to A produces a compact family of singular Euclidean metrics. Now apply Lemma 2.2. For part (2), note that by compactness of A and of the compactness of the restriction of QD^ to A, there exists a uniform 6 such that any hyperbolic metric and any normalized singular Euclidean structure determined by an el- ement a e A has a (5-hyperbolic universal cover. Part (2) is now a direct consequence of Lemma 2.3. □ 3 Convex cocompact subgroups of Isom(T) 3.1 Variations of convex cocompactness Given a proper, geodesic metric space X, a subset L C X is quasiconvex if there exists ^ > such that every geodesic segment in X with endpoints in L is contained in the ^-neighborhood of L . When G is a finitely generated, discrete subgroup of the isometry group of H" , it is well known that the following properties of G are all equivalent to each other: Orbit Quasiconvexity Any orbit of G is a quasiconvex subset of . Single orbit quasiconvexity There exists an orbit of G which is quasicon- vex in H". Convex cocompact G acts cocompactly on the convex hull of its limit set A. Moreover, these properties imply that G is word hyperbolic, and there is a continuous G-equivariant embedding of the Gromov boundary dG into OH" whose image is the limit set A. Similar facts hold for finitely generated groups acting discretely on any Gromov hyperbolic space, for example finitely gener- ated subgroups of Gromov hyperbolic groups. In this section we prove Theorem 1.1, which is a list of similar equivalences for finitely generated subgroups of the isometry group of the Teichmiiller space T oi S. In this case the entire isometry group Isom(T) acts discretely on T, and in fact by Royden's Theorem [45], [24] the canonical homomorphism MCG Isom(T) is an isomorphism, except in genus 2 where the kernel is cyclic of order 2. Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 117 Although T fails to be negatively curved in any reasonable sense, nevertheless one can say that it behaves in a negatively curved manner as long as one focusses only on cobounded aspects. This, at least, is one way to interpret the projection properties introduced by Minsky in [38] and further developed by Masur and Minsky in [32]. Given a B-cobounded geodesic g in T, Minsky's projection property says that a closest point projection map of T onto g behaves in a negatively curved manner, such that the quality of the negative curvature depends only on B. See Theorem 3.6 for the precise statement. For a finitely generated subgroup G C Isom(T) we can obtain equivalences as above, as long as we tack on an appropriate uniform coboundedness property; in some cases the desired property comes for free by uniform coboundedness of the action of G on any of its orbits. First we have some properties of G which are variations on orbit quasiconvexity: Orbit quasiconvexity Any orbit of G is quasiconvex in T. Single orbit quasiconvexity There exists an orbit of G that is quasiconvex in T . Weak orbit quasiconvexity There exists a constant A and an orbit O of G, and for each x,y & O there exists a geodesic segment [x',y'] in T, such that d{x,x') < A, d{y,y') < A, and [x',y'] is in the ^-neighborhood of O. The latter is a more technical version of orbit quasiconvexity which is quite useful in several settings. Another property of G is a version of convex cocompactness, into which we incorporate the hyperbolicity properties mentioned above: Convex cocompact The group G is word hyperbolic, and there exists a con- tinuous G-equivariant embedding : dG PAiJ^ with image Aq , such that A(3 x Aq — Ac P!FV, and the following holds. Letting WHg = u{(cO I C / C G Ag} be the weak hull of A^ , if / : G ^ WHg is any G-equivariant map, then / is a quasi-isometry and the map / = / U /oo : GU dG — > WHg UAg is continuous. In this definition, WH^ is metrized by restricting the Teichmiiller metric on T, which a posteriori has the effect of making WHq into a quasigeodesic metric space. The definition implies that G acts cocompactly on WH^: since A^ x Aq — a is a closed subset of PJ^V it follows that WHc is a closed subset of T ; Geometry & Topology, Volume 6 (2002) 118 Benson Farb and Lee Mosher and since G acts coboundedly on itself it follows that G acts coboundedly on WHc ; thus, the image of WHfj in moduli space is closed and bounded, hence compact. 3.2 Properties of convex cocompact subgroups In this section we prove several properties of convex cocompact subgroups of Isom(T) which are analogues of well known properties in Isom(H") . Proposition 3.1 Every infinite order element g of a convex cocompact sub- group G < Isom(7") fs MCG is a pseudo-Anosov mapping class. Proof Any infinite order element of a word hyperbolic group has source-sink dynamics on its Gromov boundary, and so g has source-sink dynamics on dG Ag- It follows that g has an axis in WHg. But the elements of Isom(T) MCG having an axis in T are precisely the pseudo-Anosovs [5] . □ The following is a consequence of work of McCarthy and Papadoupolos [36] . Proposition 3.2 If G is a convex cocompact subgroup of Isom(T) then: (1) Ag is the smallest nontrivial closed subset of T = T U PAiJ^ invariant under G. (2) The action of G on PMJ^ \ Aq is properly discontinuous. Proof The Gromov boundary of a word hyperbolic group is the closure of the fixed points of infinite order elements in the group, and so by Proposition 3.1 the set Ag is the closure of the fixed points of the pseudo-Anosov elements of G. Item (1) now follows from Theorem 4.1 of [36]. To prove (2), let Z{A) = {C e PMJ^ I there exists C' e A such that i{C, C) = 0} Theorem 6.16 of [36] says that G acts properly discontinuously on PAiT — Z{A) , and so it suffices to prove that A = Z(A) . Each point G A is the ideal endpoint of a cobounded geodesic ray, which implies that (' is uniquely ergodic and fills the surface [29], and so if C') = then C = C'- □ Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 119 Remark One theme of [36] is that for a general finitely generated subgroup G < MCG, there are several different types of "limit sets" for the action of G on PMT. Assuming that G contains a pseudo-Anosov element, the two sets mentioned in the proof above play key roles in [36]: A(G) which is the closure of the fixed points of pseudo-Anosov elements of the subgroup, and is also the smallest nontrivial closed G-invariant subset; and the set Z{A{G)). What we have proved is that for a convex cocompact subgroup G, these two sets are identical. Henceforth we refer to A^ as the limit set for the action of G on The analogue of the following result is true for convex cocompact discrete sub- groups of H" , as well as for word hyperbolic groups [2] ; the proof here is similar. Proposition 3.3 Let G be a convex cocompact subgroup of Isom(T), and let Nq and Comm^ be the normalizer and the relative commensurator of G in Isom(T) . Then each of the inclusions G < Nq < Commc is of hnite index, and we have Commc = Stab(AG') = Stab(WHG) . Proof Let Aq be the limit set of G , with weak hull WHo , and note that we trivially have Stab(WHG) = Stab(AG) . Note that Stab(WHG) acts properly on WH^. Indeed, Isom(T) acts properly on T, and so any subgroup of Isom(T) acts properly on any subset of T which is invariant under that subgroup. Since G C Stab(WHG'), and since G acts cocompactly on WHg , it follows that G is contained with finite index in Stab(WHG). This implies that Stab(WH(3) C Commc. To complete the proof we only have to prove the reverse inclusion Commo C Stab(WH(3). Given g G Isom(T), suppose that g G Commc, and choose finite index sub- groups H,K < G such that g~^Hg = K. By the definition of convex cocom- pactness it follows that WH^ = WHg = WHi^. Since ^(WHk) = WH^ it follows that g G Stab(WHG) . □ Remark Another natural property for subgroups G < MCG is quasiconvex- ity with respect to the word metric on MCG. It seems possible to us that this is not equivalent to orbit quasiconvexity of G in Isom(T) . Masur and Minsky [33] give an example of an infinite cyclic subgroup of Isom(T) which is not orbit quasiconvex, and yet this subgroup is quasi- isometrically embedded in MCG [14]; it may also be quasiconvex in MCG, but we have not investigated this. Qeometry & Topology, Volume 6 (2002) 120 Benson Farb and Lee Mosher 3.3 Equivalence of definitions: Proof of Theorem 1.1 Here is our main result equating the various quasiconvexity properties with convex cocompactness: Theorem 1.1 If G is a Rnitely generated subgroup of Isom(T) , the following are equivalent: (1) Orbit quasiconvexity (2) Single orbit quasiconvexity (3) Weak orbit quasiconvexity (4) Convex cocompactness Because of this theorem we are free to refer to "quasiconvexity" or "convex cocompactness" of G without any ambiguity. Proof of Theorem 1.1 The key ingredients in the proof are results of Minsky from [38] concerning projections from balls and horoballs in T to geodesies in T, and results of Masur-Minsky [32] characterizing J-hyperbolicity of proper geodesic metric spaces in terms of projections properties to paths. To begin with, note that the implications (1) ^ (2) ^ (3) are obvious. We now prove that (3) (1). Suppose we have an orbit O oi G and a constant A, and for each x,y e O we have two points x',y' G T, endpoints of a unique geodesic segment [x',y'] in r, such that d{x,x') < A, d{y,y') < A, and [x',y'] C NaIO). The set O maps to a single point in Ai and so the projection of Na{0) to is a bounded set B. It follows that each [x',y'] is B-cobounded. Now consider an arbitrary orbit Oi of G; we must prove that Oi is quasiconvex in T. The orbits 0,0i have finite Hausdorff distance C in T. Given xi,yi € Oi, choose x,y ^ O within distance C of respectively, and consider the geodesic segment [x', y'] and the piecewise geodesic path -/ =[x',x]* {x,xi] * [xi,yi] * {yi,y] * [y,y'] Of the five subsegments of 7, all but the middle subsegment have length < Max{A, C}, and it follows that 7 is a (1, D)-quasigeodesic in T, with D depending only on A,C. Since the geodesic [x',y'] is ^B-cobounded we can apply the following result of Minsky [38] to obtain 6 , depending only on B and D, such that 7 C Ns[x',y']. Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 121 Theorem 3.4 (Stability of cobounded geodesies) For any bounded subset B of Ai and any K > 1,C > there exists 5 > such that if 7 is a K,C quasigeodesic in T with endpoints x,y, and if [x,y] is B-cobounded, then C Ns[x,y]. □ It follows that C 7 C Ns+aO C Ns+a+c^^i, proving quasiconvexity of Oi in r. Weak orbit quasiconvexity implies convex cocompactness Fix an orbit O of G, and so O is quasiconvex in T. Let Q be the set of all geodesic segments, rays, and lines that are obtained as pointwisc limits of sequences of geodesies with endpoints in O. Let UQ C T he the union of the elements of Q. The left action of G on O is evidently cobounded. By quasiconvexity of O it follows that the action of G on the union of geodesic segments with endpoints in O is cobounded, which implies in turn that the action of G on UQ is cobounded. Since UQ is closed and T is locally compact, it follows that the G action on UQ is cocompact. The set UQ therefore projects to a compact subset of M which we denote B. All geodesies in Q are therefore H-cobounded. Let UQ be equipped with the restriction of the Teichmiiller metric. Note that while UQ is not a geodesic metric space, it is a quasigeodesic metric space: there exists yl > such that any x,y £ L)Q are within distance A of points x',y' £ O C UG , and the geodesic is contained in L)Q . To prepare for the proof that G is word hyperbolic, fix a finite generating set for G with Cayley graph F , and fix a G-equivariant map f-.F^ UQ taking the vertices of F to O and taking each edge of F to an element of Q . Since G acts properly and coboundedly on both F and UQ , and since both are quasigeodesic metric spaces, it follows that the equivariant map / is a quasi-isometry between F and UQ; pick a coarse inverse F: UQ ^F. By definition the group G is word hyperbolic if and only if the Cayley graph F is (5-hyperbolic for some S > 0. Our proof that G is word hyperbolic will use a result of Masur and Minsky, Theorem 2.3 of [32]: Theorem 3.5 Let X be a geodesic metric space and suppose that there is a set of paths P in X with the following properties: Coarse transitivity There exists C > such that for any x,y E X with d{x, y) >C there is a path in P joining x and y . Contracting projections: There exist a,b,c> 0, and for each path 7: I — X in P there exists a map n: X ^ I such that: Qeometry & Topology, Volume 6 (2002) 122 Benson Farb and Lee Mosher Coarse projection For each t e I we have diam (7[t, 7r(7t)]) < c. Coarse lipschitz If d{x,y) < 1 then diam (7[7rx, Try]) < c. Contraction If d{x,^{Trx)) > a and d{x,y) < b ■ d{x,^{Trx)) then diam (7[7ra;, Try]) < c Then X is S-hyperboUc for some S >(). To prove that G is (5-hyperbolic we take P to be the set of geodesic segments in G, and we look at the set of paths / oP = {/o7|7£ P} in UQ . Using some results of Minsky [38] , we will show that / o P satisfies the hypotheses of Theorem 3.5. Then we shall pull the hypotheses back to P and apply Theorem 3.5. The first result of Minsky that we need is the main theorem of [38] : Theorem 3.6 (Contraction Theorem) For every bounded subset B of M. there exists c > such that if 7 is any B-cobounded geodesic in T then the closest point projection T — > 7 satisfies the (a, 6, c) contracting projection property with (a, b) = (0, 1) . In our context, where we have a uniform B such that each geodesic in Q is ;B-cobounded, it follows that there is a uniform c such that each geodesic in Q satisfies the (0, l,c) contracting projection property. Now consider 7 = [xo,xi,. . . ,rc„] a geodesic in the Cay ley graph F, mapping via / to a piecewise geodesic /7 = [fxQ, fxi] U • • • U [fxn-i, fxn] in UQ , with each subsegment [/x,, /xj+i] an element of ^. It follows that fj is a K,C quasigeodesic in T , for K > 1,C > independent of the given geodesic in r. The T-geodesic [fxQ,fxn] is B-cobounded. Applying Theorem 3.4 it fol- lows that /7 C Nj:i[fxo,fxn], where D depends only on B, K,G . As noted above, closest point projection from T onto [fxo,fxn] satisfies the (0,1, c) contracting projection property. From this it follows that closest point projec- tion tt: T ^ /7 satisfies the {a',b',c') contraction property where {a',b',c') depend only on B,K,C. Now define the projection F ^ 7 to be the composi- tion r ^ UQ ^ fj —>■ T —>■ J where the last map is closest point projection in F. This composition clearly satisfies the {a",b",cf') projection property where {a",b",c") depend only on {a',b',c.) and the quasi-isometry constants and coarse inverse constants for f,F. Geodesies in F are clearly coarsely transitive, and applying Theorem 3.5 it follows that G is word hyperbolic. This means that geodesic triangles in F are Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 123 uniformly thin, and it implies that for each K, C there is a 5 such that K, C quasigeodesic triangles in T are (5-thin. Applying the quasi-isometry between r and UQ , it follows that there is a uniform 5 such that for each x,y,z G O the geodesic triangle A[x, y, z] in UQ is (5-thin; we fix this 6 for the arguments below. Now we turn to a description of the "limit set" A C PMJ^ of G, with the ultimate goal of identifying it with the Gromov boundary dG. ^ Each geodesic ray in Q has the form [a;, 77), for some x G O, rj G PA4J-; define A C PMT be the set of all such points rj, over all geodesic rays in Q . The set A is evidently G-equivariant. Fact 1 For any x 77 G A, the ray [x,r]) in T is an element of Q. ^ To prove this, by definition of A there exists a ray [y,r)) in Q for some y E O. > Choose a sequence yi,y2,... G O staying uniformly close to [y,r]) and going to infinity. Pass to a subsequence so that the sequence of segments [x, y„] > converges to some ray [x,rj') G Q; it suffices to show that r]' = r]. Since x is > fixed and the points y„ stay uniformly close to [y,r]), it follows by Theorem > 3.4 that the segments stay uniformly close to [y,r]), and so [x,rf ) is in a finite neighborhood of [y, rf) . The reverse inclusion, that [y, rj) is in a finite neighborhood of [x, 77'), is a standard argument: as points move to infinity in [x, T]') taking bounded steps, uniformly nearby points move to infinity in [y, rj) also taking bounded steps, and thus must come uniformly close to an arbitrary > > > point of [y, r/). This shows that the rays [x,r]'), [y,r]) have finite Hausdorff distance, and applying Lemma 2.4 (End Uniqueness) shows that r) = rj' . Note that in the proof of Fact 1 we have established a little more, namely that for any x,y e O and G A the rays [x,i]) and [y,r)) have finite Hausdorff distance. This will be useful below. Fact 2 For any 77 7^ ^ G A there exists a line (77, Q contained in Q. From Fact 2 it immediately follows that A x A — A C PTV , that the weak hull WHg of A is defined, and that G acts coboundedly on WHo, since G acts coboundedly on . To prove Fact 2, pick a point x G O, and note that by Fact 1 wc have two > > rays [x,r]), [x,C,) in Q- Pick a sequence yn ^ O staying uniformly close to Qeometry & Topology, Volume 6 (2002) 124 Benson Farb and Lee Mosher > [x, rf) and going to infinity, and a sequence Zn & O staying uniformly close to [x,C) and going to infinity. We have a sequence of triangles [x,y„,2;„] in all (5-thin. Applying Theorem 3.4 there is a D such that the sides > are contained in the D-neighborhood of [x, rj) , and the sides [x, Zn] are con- > tained in the D-neighborhood of [x, () . Each side [y„, Zn] , being contained in the (5-neighborhood of [x, U [x, Zn] , is therefore contained in the D + S- > > neighborhood of [x,rf) U [x,Q. We claim that the point x is uniformly close to the segments Zn] ■ If not, then from uniform thinness of the triangles [x,yn,Zn] it follows that there are points y'j € [x, Un] and z'^ G [x, Zn] such that the segments [x, y'^] and [x, z'J get arbitrarily long while the Hausdorff distance between them stays uniformly bounded. This implies that there are sequences y'^ G [x, rj) going to infinity and G [a;, C,) going to infinity such that the Hausdorff distance between the segments and [a;,2;^] stays uniformly bounded, which implies in turn that the rays [x, rj) and [x, Q have finite Hausdorff distance. Applying End Uniqueness 2.4, it follows that rj = contradicting the hypothesis of Fact 2, and the claim follows. Passing to a subsequence and applying Ascoli-Arzcla it follows that [yn,Zn] > converges to a line in Q . One ray of this line is Hausdorff close to [x,r]) and so > has endpoint r], and the other ray is Hausdorff close to [x, () and so has endpoint < > C, by End Uniqueness. We therefore have lim[?/„,Zn] = (t/,C)) completing the proof of Fact 2. Now we define a map foo '■ dG A . Recall that the relation of finite Hausdorff distance is an equivalence relation on geodesic rays in the Cayley graph F of G, and dG is the set of equivalence classes. Consider then a point ^ G dG represented by two geodesic rays [xq, xi, . . . ) and [yo,yi,---) with finite Hausdorff distance in F. These map to piecewise geodesic, quasigeodcsic rays P = [fxQ, fxi] U [/xi, fx2] U • • • and a = [fyo, fyi] U [fyi, fy2] U • • • with finite Hausdoff distance in U^. The sequence of geodesic segments [fxo,fxn] in Q has a subsequence converging to some ray [fxo,Q in Q, and [fyo,fyn\ has a subsequence converging to some ray [fyo,C') in ^. To obtain a well defined map dG — A it suffices to prove that C = d ^nd then we can set fooiO = C- To prove that ^ = ^' it suffices, by End Uniqueness 2.4, to prove that the rays > > [fxoX) &nd [fyo,C') have finite Hausdorff distance in T. Since the piecewise geodesic rays p, a have finite Hausdorff distance in T , it suffices to prove that p Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 125 has finite Hausdorff distance from [/xq, C) , ^'^d similarly a has finite Hausdorff > distance from [fyo,C')- Consider a point p £ p. For sufficiently large n we have p £ pn = [fxo,fxi] U ••• U [fxn-i,fxn]- Applying Theorem 3.4 there is a uniform constant D such that pn C Nd ( [fxQ , fxn] ) , and so p is within > distance D of some point in [fxo,fxn\- Since [/xo,C) is the pointwise limit of [fxo, fXn] as n — GO it follows that p is within a uniformly bounded distance > > of [fxo, () . This shows that p is within a finite neighborhood of [/xq, C) ■ The reverse inclusion is a standard argument: as points move along p towards the end taking bounded steps, uniformly nearby points move along [/xq, C) towards the end also taking bounded steps, and thus must come uniformly close to some point of [/xo,C)- Hence foo '■ dG — A is well defined. Observe that a similar argument proves a little more: if Xj G G converges to ^ G dG then the segments [/xq, /xj] > converge in the compact-open topology to the ray [/xq, fCl'i details are left to the reader. We now turn to verifying required properties of /oo • > To see that /oo is surjective, consider a point G A and pick a ray [x, rj) 'm.Q . It follows that p = F ^[x, r/)^ is a quasigcodesic ray in F. Since F is (5-hyperbolic it follows that p has finite Hausdorff distance from some geodesic ray /?' in F, with endpoint C,' G dG. As shown above, f{p') has finite Hausdorff distance > from some geodesic ray [x', /ooC') • Since /, F are coarse inverses it follows that > > [x,?7) has finite Hausdorff distance from [x',/ooC')! ^-^d so by End Uniqueness it follows that rj = fooC ■ To see that /oo is injective, consider two points r/, C G dG and suppose that /oo(^) = fooiO'i let ^ G A be this point. Pick rays p,a in T representing r],C, respectively. As wc have just seen, the images /(/)), /(o") have finite Hausdorff > > distance in T to rays [y, ^) , [z, ^) in Q , respectively. As noted at the end of the > > proof of Fact 1, the rays [y,^) and [z,^) have finite Hausdorff distance in T; applying the coarse inverse F it follows that p, a have finite Hausdorff distance in F and therefore rj = (. We have shown that /oo is a bijection between dG and A. We want to prove that foo is a homeomorphism, and that the extension f = f U foo '■ G U dG — ^ T = T U PAiJ^ is continuous. For this purpose first we establish: Fact 3 A is a closed subset of PA^JT, and therefore compact. Geometry & Topology, Volume 6 (2002) 126 Benson Farb and Lee Mosher To prove this, choose a sequence G A so that hm (n = Coo in PMJ^; we must prove that Coo ^ Choose a point x E O, and apply Fact 1 to obtain rays > > [x, Cn) • Passing to a subsequence these converge to a limiting ray lim [x, Cm) = > [a;,C^) in Q, and so C^jo ^ ^- Looking in the unit tangent bundle of T at the point X it follows that lim Cn = C^ ! *nd so Coo = C^ ^ ^ • Fact 4 /oo : dG — >^ A is a homeomorphism. Since both the domain and range are compact Hausdorff spaces it suffices to prove continuity in one direction. Continuity of follows by simply noting that for fixed x e O and for a convergent sequence ^„ — >^ ^ in A C PMJ^, the sequence of rays [x, converges in the compact open topology to the ray Fact 5 The map J = f U foo - G Li dG = T \J PMJ^ is continuous. To be precise, this map is continuous using the Thurston compactification T of T . We prove this by showing first that the map is continuous using the Te- ichmiiller compactification, and then we apply Masur's Two Boundaries Theo- rem [30] which says that the map from the Teichmiiller compactification to the Thurston compactification is continuous at uniquely ergodic points of PM!F. First we recall the Teichmiiller compactification in a form convenient for our cur- rent purposes. There are actually many different Teichmiiller compactifications, one for each choice of a base point in T; we shall fix a base point z = f{x) € O for some x £ G. As we have seen, there is a unique geodesic segment [z,z'] > for each z' G T , and a unique geodesic ray [z,Q for each C £ PAiJ^. The Teichmiiller topology on T = T U PMT restricts to the standard topologies on T and on PAiJ^, it has T as a dense open subset, and a sequence Zi E T converges to C £ PMJ- if and only if the sequence of segments [z, Zi] converges > to the ray [z, () in the compact open topology; equivalently, letting B denote the unit ball in T centered on z, the distance d(z,Zi) goes to infinity and the > set [z, Zi] n B converges to the set [z, C) n 5 in the Hausdorff topology. We already proved in Fact 4 that /oo is continuous; for this we implicitly used the fact that the Thurston topology on PMJ^ is identical to the Teichmiiller topology, defined by identifying PMJ^ with the unit tangent bundle at x. We also observed earlier, after the proof that /oo is well-defined, that if Xj G G converges to ^ G dG, then /(xj) G T converges to /oo(0 ^ PMJ-^ in Qeometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 127 the Teichmiiller topology on T. Putting these together it follows that / is continuous using the Teichmiiller topology on T. Since A = foo{dG) consists entirely of uniquely ergodic points in PAiJ^, Masur's Two Boundaries Theorem [30] implies that the identity map on T is continuous from the Teichmiiller topology to the Thurston topology at each point of A , and so / is continuous using the Thurston topology on T. We now put the pieces together to complete the proof of convex cocompact- ness. Let f : G ^ WHq be an arbitrary G-equivariant map, and define : do PAiJ^ to be equal to foo ■ We must prove that /' is a quasi-isometry and that the extension f = f 'O f'^: GLidG WHg UAg is continuous. Prom Pacts 1-5 above, it follows that the quasi-isometry f : G ^ UQ has continuous extension / : GUdG — ^ U^U A, and so UQ is a Gromov hyperbolic metric space with Gromov compactification UQ U A . Since WHc C UQ is a G— invariant sub- set, it follows that WHc is Gromov hyperbolic with Gromov compactification WHgUA. The map /' is a G-equivariant map between quasigeodesic metric spaces on which G acts properly and coboundedly by isometries, and hence /' is a quasi-isometry. Since d{f'{x),f{x)) is uniformly bounded for x EG, then from the fact that = foo it follows that /' is continuous. This completes the proof that weak orbit quasiconvexity implies convex cocom- pactness. Convex cocompact implies weak orbit quasiconvexity Assuming G is convex cocompact, pick a finite generating set for G with Cayley graph P and G-equivariant, coarsely inverse quasi-isometries /: P ^ WHc, /: WHg — > P. Let O be an orbit of G in T. Since G acts coboundedly on WHg it follows that O has finite Hausdorff distance from WHg in T. It suffices to show that for any two points x,y E O there is a geodesic line whose infinite ends are in A such that x, y come within a uniformly finite distance of that line. Pick a G-equivariant map g:T ^ T taking the vertices of P bijectively to O and each edge of P to a geodesic segment, so / and g differ by a bounded amount. Since P is (5-hyperbolic it follows that there is a constant A such that any two vertices of P lie within distance A of some bi-infinite geodesic. Pick x,y G O, and pick a bi-infinitc geodesic 7 in P such that g~^{x), (y) are within distance ^ of 7. Let S,,rj E dG be the two ends of 7. By the statement of convex cocompactness, there is & K,G quasigeodesic line in P of the form / ( (/ooC) fooV) ) whose two infinite ends are ^,77, where K, G are independent Qeometry & Topology, Volume 6 (2002) 128 Benson Farb and Lee Mosher of ^,77. It follows that 7 and / (j^foo^jfoovij uniformly close, and so 7(7) and (foo^, foov) uniformly close, and so the points x, y are uniformly close < > to (/oo^,/oo??)- □ 4 Hyperbolic surface bundles over graphs In this section our goal is to give an explicit construction of model geometries for surface group extensions, and to study regularity properties of these geometries. Here is a brief outline; detailed constructions follow. Consider a finitely generated group G and a homomorphism f : G ^ Isom(T) ~ MCG. Let X be a Cayley graph for G. Choose a map X ^ T which is equivariant with respect to the homomorphism /, that is, ^{g-x) = f{g) • , X e X,g e G, where we use the • notation to denote an action. By pulling back the canonical marked hyperbolic surface bundle S ^ T via the map $ we obtain a marked hyperbolic surface bundle 5^ — > X. By pulling back the canonical hyperbolic plane bundle H ^ T we obtain a hyperbolic plane bundle Hx X , and a covering map Hx Sx with deck transformation group 7ri(S'). There is an action of the extension group Tq on TCx such that the covering map Hx Sx is equivariant with respect to the homomorphism By imposing a G -equivariant, proper, geodesic metric on Sx and lifting to Hx , we can then use Hx as a model geometry for the extension group Fq ■ We may summarize all this in the following commutative diagrams: H ^S MGG{S,p) ^MCG{S) Each group in the right hand diagram acts on the corresponding space in the left hand diagram, and each map in the left hand diagram is equivariant with respect to the corresponding group homomorphism in the right hand diagram. We will impose several -equivariant structures on the space Hx , by finding appropriate G-equivariant structures on Sx and lifting. Qeometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 129 For example, we put an equivariant, proper, geodesic metric on TCx by lifting an equivariant, proper, geodesic metric on Sx ■ These metrics will have the property that the topological fibrations Sx X , Hx X are also "metric fibrations" in the following sense. In a metric space given subsets A^B d Z , denote the min distance by dmin(^, -B) = inf{d(a, 6) | a G A, 6 G -B}, and the Hausdorff distance by (iHaus(^,-B) = inf{r | A C Nr{B),B C Nr{A)}. Metric fibration property A map of metric spaces f-.Z^Y satisfies the metric fibration property if Y is covered by neighborhoods U such that a y,z E U then dnnn{r\y),r\z)) = ^^Haus (?/), (^)) = dY{y,z) 4.1 Metrics and connections on surface bundles over paths The marked hyperbolic surface bundle over a path in T Consider first a smooth path a: I ^ T, defined on a closed connected subset / C R, that is, a closed interval, a closed ray, or the whole line. Pulling back the canonical marked hyperbolic surface bundle S ^ T via the map a we obtain a marked hyperbolic surface bundle Sq, ^ I ■ We impose a Riemannian metric on Sq, as follows. Recall that we have chosen a connection on the bundle iS — ^ T. By pulling back the connection on the bundle S ^ T we obtain a connection on the bundle Sa ^ 1 1 that is, a 1-dimensional sub-bundle of TSa which is complementary to the vertical sub-bundle TySa ■ There is a unique vector field V on Sa parallel to the connection such that the projection map Sa ^ I takes each vector of V to a positive unit vector in the tangent bundle of / C R. There is now a unique Riemannian metric on S whose restriction to T^Sa is the given hyperbolic metric along leaves of Sa, and such that F is a unit vector field orthogonal to TySa- Since / is closed subset of R, the path metric on S^ induced from this Riemannian metric is proper, and so by Fact 2.1 we may regard Sa as a geodesic metric space. Here is another description of the Riemannian metric on . Integrcition of the connection sub-bundle defines a 1-dimensional foliation on Sa transverse to the surface fibration, whose leaves are called connection paths. Choosing a base leaf of the fibration Sa ~^ I , and identifying this base leaf with S , we may project along connection paths to define a fibration Sa ^ S . Combining this with the fibration Sa ^ I we obtain a diffeomorphism Sa ^ S x I . Letting gt Geometry & Topology, Volume 6 (2002) 130 Benson Farb and Lee Mosher be the given Riemannian metric of curvature —1 on the leaf St ^ S x t, t E I, we obtain the Riemannian metric on Sq, via the formula Remark The metric on Sa depends on the choice of a connection on the bundle 5 — > T. However, when a is cobounded, two different connections on S will induce metrics on Sa which axe bilipschitz equivalent, with bilips- chitz constant depending only on the pair of connections and on the cobound- edness of a, not on a itself. For each s,t E I we have a connection map hst' Sg ^ St, defined by moving each point of Ss along a connection path until it hits St- Clearly wc have hst ° = hrt, {r, s,t G /). Notice that the map hgt takes each point of Sg to the unique closest point on St, and that point is at distance |s — i|. In fact, starting from an arbitrary point on Ss, all paths to St have length > |s — f|, and the connection path is the unique one with length = |s — t| . It follows that the map Sa ^ I satisfies the metric fibration property. Consider more generally a piecewise smooth path a: I T . On each subinter- val I' G I over which a is smooth, there is a Riemannian metric as constructed above. At a point t e I where two such subintervals meet, the Riemannian metrics on the two sides agree when restricted to St- We therefore have a piecewise Riemannian metric on Sa , inducing a proper geodesic metric. The connection paths which are defined over each smooth subinterval I' d I piece together to give connection paths on all of Sa , and we obtain connection maps hst '■ Ss ^ St for all s,t E I. Note that since the connection on S ^ T is equivariant with respect to the ac- tion of MCG, the piecewise Riemannian metric on each Sa is natural, meaning that for any h G MCG , the induced map Sa Shoa is an isometry. Similarly, the connection paths and connection maps are also natural. Each connection map hst '■ Ss ^ St is clearly a diffeomorphism, and since its domain is compact it follows that hst is bilipschitz. The next proposition ex- hibits some regularity, bounding the bilipschitz constant of hst by a function of |s — t| that depends only on the coboundedness of the path a: I ^ T , and a lipschitz constant for a. For technical reasons we state the lemma only for paths a: I ^ T which are piecewise affine, meaning that / is a concatenation of subintervals /' such that a | /' is an affine path, a constant speed reparam- eterization of a Teichmiiller geodesic. Piecewise affine paths are sufficient for all of what follows. Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 131 Lemma 4.1 For each bounded subset B C M. and each p > 1 there exists K > 1 such that the following happens. If a: I T is a B-cobounded, p-lipschitz, piecewise afhne path, then for each s,t & I the connection map hgt : Sg ^ St is Xl*"*! -bilipschitz. In what follows we shall describe the conclusion of this proposition by saying that K is a bilipschitz constant for the connection maps on 5q, . Proof A standard lemma found in most O.D.E. textbooks shows that if $ is a smooth flow on a compact manifold then there is a constant K >1 such that ||^t(^)|| < -f^'*' ll^ll- We can plug into this argument as follows. The conclusion of the lemma is local, and so it suffices to prove it under the assumption that / = [0, 1] and that a is affine. There exists a compact subset A (Z T such that any ;B-cobounded, p-lipschitz path a: [0,1] — > T, can be translated by the action of MCG to lie in the set A. Let C{Ajp) be the set of all /9-lipschitz affine paths [0, 1] ^, a compact space in the compact open topology. By naturality of the metric on 5« , it suffices to prove the lemma for a G C{A,p). For each a G C{A,p) and each vector w tangent to a fiber Sg, s G [0,1], define: l{w) = lim - log t^o t \ \\w\\ J dt log =0 \Dhs^s+t{w)\\ \w\ Since l{cw) = l{w) for c ^ 0, we may regard l{w) as a function defined on the projective tangent bundle of S crossed with /, a compact space. As w varies, and as a varies over the compact space C{A, p) , the function l{w) varies continuously, and so by compactness l{w) has a finite upper bound I. Setting = e', it now follows by standard methods that ||/is^s_|_t(it;)|| < \\w\\ when w is tangent to Sg, and so hs,s+t is -f^'*' bilipschitz. □ The hyperbolic plane bundle over a path in T Letting a : I — > T be a piecewise affine path as above, by pulling back the canonical hyperbolic plane bundle 7i — ^ T we obtain a bundle Tia I ■ Note that there is a universal covering map Tia — > Sa with deck transformation group '7ri(S') such that the composition Tia Sa ^ S equals the composition TLa H ^ S , and also the composition Ha Sa ^ I equals the fibration map Ha I ■ By lifting the piecewise Riemannian metric from Sa we obtain a piecewise Riemannian metric on Ha , inducing a proper, geodesic metric. The map Ha I satisfies the metric fibration property. The connection paths on Sa lift to connection Geometry & Topology, Volume 6 (2002) 132 Benson Farb and Lee Mosher paths on Ha , and we obtain connection maps hgt '■ Hs Ht ■ By applying Lemma 4.1 it follows that if a is ;S-cobounded and /9-lipschitz then the same constant K = K{B,p) is a bilipschitz constant for the connection maps on Tia- 4.2 Metrics and connections on surface bundles over graphs Let f:G^ MCG be a homomorphism defined on a finitely generated group G. We have a canonical extension 1 — tti (S) Tq ^ G ^ 1. Fix once and for all a Cayley graph X for G , on which G acts cocompactly with quotient a rose. Fix a geodesic metric on X with each edge having length 1. Choose a G-equivariant map $ : X ^ T taking each edge of X to an affine path in T. Letting ||$|| be the maximum speed of the map ie, the maximal length of the image of an edge of X under it follows that $ is a ||$||-lipschitz map. Evidently the image of <1> is a cobounded subset of T , because the vertices of X map to a single orbit and each edge of X maps to a geodesic of length < ||$||. Choose a compact set B C M so that image($) is B-cobounded. Using the method of Section 4.1, for each edge e of X we have a bundle <Se — >■ e equipped with a Riemannian metric. Given any vertex v oi X , for any two edges e, e' incident to v the Riemannian metrics on Se and S^' fit together isometrically at ■ We may therefore paste together the Riemannian metrics on Se for all edges e to obtain a marked hyperbolic surface bundle Sx X equipped with a piecewise Riemannian metric. The induced path metric on Sx is a proper, geodesic metric. By naturality of the metrics on the bundles Sg, the action of G on X lifts to an isometric action on Sx ■ By lifting the metric from Sx to its universal cover Ttx we obtain a hyper- bolic plane bundle Hx X on which the extension group acts cocom- pactly, equipped with a Fq equivariant, piecewise Riemannian metric, inducing a proper, geodesic metric on Hx ■ Note in particular that Tq is thus quasi- isometric to Hx ■ Note that this construction produces bundles Sx X and Hx X isomor- phic to the pullback bundles described at the beginning of Section 4. Since each map Se ^ e, He ^ e satisfies the metric fibration property, it follows that the maps Sx — X , Hx — X also satisfy that property. The connections on the spaces Se , for edges e of X , piece together to define a G— equivariant connection on Sx- To make sense out of this, we consider only the connection map defined for a piecewise path 7 : [a, ^ X , as follows. The bundle Sx X pulls back to give a bundle S^ — > [a, 5] , and the connection Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 133 paths over each edge of X piece together to give connection paths on , with an induced connection map : 5^(a) '^7(6) • follows immediately from Lemma 4.1 that is i^i'^^W-bilipschitz, where K = K{B, ||$||). By lifting to Tix , for each picccwise geodesic path 7 : [a,b] X we similarly obtain a K^^^^'^'> bilipschitz connection map /i-y : 'H-y(a) — >■ ^7(5) ■ 4.3 Lcirge scale geometry of surface bundles over paths Our goal now is to compare metrics on Ti.^ and Tip for paths 7, /3 in T which are closely related. Given a metric space Z, two paths 7,/?: / Z, and a constant ^ > 0, we say that 7, /3 are A-fellow travellers if d{'-f{t), P{t)) < A for all t & I. More generally, given paths I^Z,^: J^Z,a constant A> 0, and constants A > l,e > 0, we say that 7, (3 are asynchronous A-fellow travellers with respect to a A, e quasi-isometry (j): I ^ J if the paths 7 and (3 o (j) are fellow travellers. It is a well known and simple fact that given a quasigeodesic j: I ^ Z and another path p-. J ^ Z , the following are equivalent: (1) P is a quasigeodesic and /3,7 have finite Hausdorff distance; (2) P is an asynchronous fellow traveller of 7. Moreover, the constants are uniformly related: in 1 =^ 2, there exist asyn- chronous fellow traveller constants A, A, e depending only on the quasigeodesic constants for /3 and the Hausdorff distance of /9,7; in 2 => 1, there exist quasigeodesic constants for (3 and a bound on the Hausdorff distance between (3 and 7 depending only on the asynchronous fellow traveller constants. The following proposition says that if j: I ^ T, (3: J ^ T are asynchronous fellow travellers in T, then there is a fiber preserving quasi-isometry Hj -^Hp. Moreover, if 7 is a geodesic, and if instead of wc use the singular SOLV space 7^^°^^, then there is a fiber preserving quasi-isometry H^°^^ Hp. Proposition 4.2 For each bounded subset B C M, and each p > 1, \ > I, e>0,A>0,K>l, there exists K' >1, C >0 such that the following hold. Suppose that ^f. I ^ T , j3: J ^ T are B-cobounded p-Lipschitz, piecewise afhne paths in T . Suppose also that 7, f3 are asynchronous A-fellow travellers, with respect to a X,e quasi-isometry cp: / — > J. Then: Qeometry & Topology, Volume 6 (2002) 134 Benson Farb and Lee Mosher (1) There exists a commutative diagram such that the top row preserves markings, and such that any Ufted map Hj — Hp is a K',C' quasi-isometry. (2) If 7 is a geodesic, then there exists a commutative diagram such that the top row preserves markings, and such that any Ufted map H^°^'^ H/3 is a K',C' quasi-isometry One way to interpret item (1) of this proposition is that a cobounded, hpschitz path in Teichmiiller space has a well-defined geometry associated to it: ap- proximate the given path by a piecewise afiine path and take the associated hyperbolic plane bundle; the metric on that bundle is well-defined up to quasi- isometry, independent of the approximation. A further argument shows that the geometry is independent of the choice of an equivariant connection on the bundle S ^ T: any two equivariant connections are related in a uniformly bilipschitz manner over any cobounded subset of 7". Proof Both (1) and (2) are proved in the same manner using Proposition 2.5; we prove only (1). To smooth the notation in the proof we denote t' = (p{t) , we let St denote the fiber of S^, we let 5^, denote the corresponding fiber 5/3(,^(f')) of Sg, etc. To prove (1), by applying Proposition 2.5(1) we choose for each t G R a marked map $t : »St — S'^, for which any lift ^t'-Ht ^ H[, is a if i , Ci quasi-isometry, where the constants Ki,Ci depend only on B,A. Since each preserves markings we may choose the lifts $t so that for any s, t we have a commutative diagram of induced boundary maps: I dhst Qeometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 135 Applying Proposition 2.5(2) it follows that if we strip off the d symbols from the above diagram, and if we choose s, t so that |s — t| < 1, then we obtain the following diagram, a coarsely commutative diagram in the sense that the two paths around the diagram differ in the sup norm by a constant C2 depending only on B,p,X,e,A,K: is n, hs,t H Define Hj -^Hp so that $ | T^s = To prove that $ is a quasi-isometry we need only show that if x, y G satisfy d{x,y) < 1 then r/($(x), $(y)) is bounded by a constant depending only on B, p, X,e, A, K , and then carry out the similar argument with inverses. Given x,y & H.^ with d{x,y) < 1, choose s,t so that x G Tig, y ^TLt- By the metric fibration property we have |s — t| < 1. Changing notation if necessary we may assume that s <t. Let a be the geodesic in H^y connecting x and y, and by the metric fibration property note that a C Ti.[s-i,t+i] ■ Consider the map p: TC[s-i,t+i] ~^ T~^t whose restriction to TCj. is the connection map hrt', it follows that p is bilipschitz with constant X*-s+2 ^ The distance in Ht between the point p{x) = hst{x) and the point y is therefore at most . Mapping over to 7^^ we have d{^x)My)) < d{^x),hs't'mx)))+d{hs't'{Hx))Mhst{x))) + d{^hstix))My)) < \s' -f'\+C2 + {KiK^ + Ci) and since \s' — t'\ < A |s — i| + e < A + e, the proof is done. □ 5 Hyperbolic extension implies convex cocompact quotient In this section we prove Theorem 1.2. Fix a homomorphism f : G ^ MCG defined on a finitely generated group G, and suppose that the extension group Tq is word hyperbolic. We must prove that / has finite kernel and that f{G) is a convex cocompact subgroup of MCG. Qeometry & Topology, Volume 6 (2002) 136 Benson Farb and Lee Mosher Fix a Cayley graph X for G and an /-equivariant map X — > G which is affine on edges of X . Choose a bounded subset B C M and a number p > 1 such that is ;S-cobounded and p-hpschitz. We have a hyperbohc plane bundle Hx X, and an action of on Hx, such that the fibration Hx — > X is equivariant with respect to the homomorphism To — G. We also have a piecewise Riemannian metric for which TCx X satisfies the metric fibration property. We also have a connection on TCx , in the form of a connection map h^: 'H^(a) ^7(6) for geodesic path 7: [a,6] — > X. The connection and metric are each equivariant with respect to Tq- Since Hx is a proper geodesic metric space, it follows that TCx is a model geometry for Tq- Since Tq is word hyperbolic, it follows that TCx is (^-hyperbolic for some 5>0. Fact 5.1 For each point x £ X , the inclusion map TCx ^ T^x is uniformly proper, with uniform properness data independent of x . Proof This follows because the subgroup of Vq stabilizing TCx is the normal subgroup 7ri(S'), and the inclusion map 7ri(S') Tq is uniformly proper with respect to word metrics, a fact that holds for any finitely generated subgroup of a finitely generated group. □ For each geodesic path 7: / ^ X, / a closed, connected subset of R, we obtain a piecewise affine path <1> o 7: / T and a hyperbolic plane bundle Ti^ — > /, which can be regarded either as the pullback of the bundle TC ^ T via $ o 7, or as the restriction of the bundle TCx X to ^. In either case, we obtain a piecewise Riemannian metric and connection on TC-f, natural with respect to the action of 7ri(S'). The connection on TC-f has bilipschitz constant K depending only on B and p, meaning that for any s,i G R, the connection map hst: Hs TCt is kI^"*! -bilipschitz. Here is an outline of the proof of Theorem 1.2. Our main task will be to prove that for each geodesic path 7 : I X , the space TC~^ is a 5' -hyperbolic metric space, for some constant S' depending only on B, p, and S. Of course, when / is a finite segment the space TC-y is quasi-isometric to the hyperbolic plane and so TCy is a hyperbolic metric space, but uniformity of the hyperbolicity constant 6' is crucial. This is obtained using the concept of flaring, introduced by Bestvina and Feighn for their combination theorem [6], and further developed by Gersten in [18]. The combination theorem says, in an appropriate context, that flaring implies hyperbolicity. Gersten's converse, proved in the same context, says that hyperbolicity implies flaring. We shall Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 137 give a new technique for proving the converse, which apphes in a much broader, "higher-dimensional" context, and using this technique we show that since Tix is 5-hyperbohc it follows that each satisfies flaring, with uniformity of constants. Then we shall apply the Bestvina-Feighn combination theorem in its original context to conclude that is (5' -hyperbolic. Next we will apply a result of Mosher [41] which says that since is hyper- bolic, the path <^o^:7— >Tisa quasigeodesic which is Hausdorff close to a Teichmiillcr geodesic, again with uniformity of constants. This will quickly im- ply finiteness of the kernel of /. The collection of these Teichmiiller geodesies, one for each geodesic 7 in X, will be used to verify the orbit quasiconvexity property for the group f{G). In what follows, a path I ^ X will often be confused with the composed path I — ^ X — ^ T; the context should make the meaning clear. Remark The context of the Bestvina-Fcighn combination theorem, and Ger- sten's converse, is the following. Consider a finite graph of groups F, with word hyperbolic vertex and edge groups, such that each edge-to-vertex group injec- tion is a quasi-isometric embedding. Associated to this is the Bass-Serre tree T, and a graph of spaces X T on which ttiF acts properly discontinuously and cocompactly. For each path in the tree T , Bestvina-Feighn define a flaring condition on the portion of X lying over that path. The combination theorem combined with Gersten's converse says that flaring is satisfied uniformly over all paths in the Bass-Serre tree if and only if ttiF is word hyperbolic. When G is a free group mapped to MCG then the extension 1 ttiS ^ Tq ^ G ^ 1 fits into this context, because Tq is the fundamental group of a graph of groups with edge and vertex groups isomorphic to ttiS , and with isomorphic edge-to- vertex injections, where the underlying graph is a rose with fundamental group G. This was the technique used in [40] to construct examples where F^ is word hyperbolic. When G is not free then this doesn't work, motivating our "higher-dimensional" version of Gersten's result. 5.1 Flciring Motivated by the statement of the Bestvina-Feighn combination theorem, we make the following definitions. Consider a sequence of positive real numbers (rj)j^j, indexed by a subinterval J of Z. Geometry & Topology, Volume 6 (2002) 138 Benson Farb and Lee Mosher The L-lipschitz condition says that r^/r^ < lI' ■'I for ah i,j, or equivalently Given k > 1, an integer n > 1, and ^4 > 0, we say that (rj) satisfies the (k, n, A) -flaring property if, whenever the three integers j — n, j , j + n are ah in J, we have: The number A is called the flaring threshold. Having a positive flaring threshold A allows the sequence to stay bounded by A on arbitrarily long intervals. However, at any place where the sequence has a value larger than A , exponential growth kicks in inexorably, in either the positive or the negative direction. Consider a piecewise affine, cobounded, lipschitz path 7 : / ^ T and the corre- sponding hyperbolic plane bundle ^ I. A X-quasivertical path in 7i.y is a A-lipschitz path a: I' —>■ H^y, defined on a subinterval I' C I , which is a section of the projection map Hj I. For example, a A-quasivertical path is a con- nection path if and only if it is 1-quasivertical. Note that each A-quasivertical path is a (A, 0)-quasigeodesic. The vertical flaring property for the fibration TCy 7 says that there exists K > 1, an integer n > 1, and a function A{X): [l,oo) (0, 00), such that if a,/3: / — Hj are two A-quasivertical paths with the same domain I' , then setting J = 7' n Z the sequence satisfies the K,n,A{X) flaring property, where dj is the distance function on TCj , j £ J . One can check that if the vertical flaring property holds for some function A(X) then it holds for a function which grows linearly. Lemma 5.2 (Hyperbolicity of Hx implies vertical flaring of Hj) With notation as above, for every 5 there exists k, n, A{X) such that if Ttx is 5~hypcrboMc then for each bi-infinite geodesic 7 in X the Hbration Hj — >■ / satisfies k, n, A{X) vertical flaring. The intuition behind the proof is that the flaring property is exactly analogous to the geodesic divergence property in hyperbolic groups, described by Cannon in [12]. The geodesic divergence property says that in a (5 -hyperbolic metric space, if p is a base point and if a, (3 are a pair of geodesic rays based at p, and if di is the shortest length of a path between a{i) and (5{i) that stays outside of the ball of radius i centered on p, then the sequence di satisfies a flaring property with constants independent of a,/?. In our context, a and (5 will no fi/fj < L whenever \i — j\ = 1. Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 139 longer have one endpoint in common. But the quasivertical property together with the metric fibration property give us just what we need to adapt Cannon's proof of geodesic divergence given in [12], substituting the geodesic triangles in Cannon's proof with geodesic rectangles. Proof We use d for the metric on Hx ■ First observe that any A -quasivertical path a in H^y is a (A, 0)-quasigeodesic in Hx ) in feet |s - i| < d{a{s),a{t)) < A |s - i| The upper bound is just the fact that a is A-lipschitz, and the lower bound follows from the metric fibration property for Hx X , together with the fact that 7 is a geodesic in X . Consider then a pair of A-quasivertical paths a,/3: /' Hj defined on a subinterval I' C I , and let J = I' Ci Z = We assume that j+ — j- is even and let jo = € J. For each j G J we have a fiber Hj isometric to H^, with metric denoted dj. We must prove that the sequence ^0 ~ satisfies K,n,A flaring, with K,n independent of A and with K,n,A independent of a, /?, and 7. For j,k E J let hjk ■ Hj —>■ Hk be the connection map, a K\^~''\ bilipschitz map. For each j e J we have an Hj geodesic pj : [0, Dj] — Hj with endpoints a{j) , Claim 5.3 There is a family of quasivertical paths v described as follows: • For each j £ J and each t G [0, Dj] the family contains a unique qua- sivertical path Vjt'. — > Hry that passes through the point Pj{t). If we fix j E J, we thus obtain a parameterization of the family Vjt by points t e [0,Dj]. • The ordering of the family Vjt induced by the order on t e [0, Dj] is independent of j . The first path Vjo in the family is identified with a , and the last path vjOj is identified with [3. • Each Vjt is X' -quasivertical, where X' depends only on X and K . When j is assumed fixed, we write vt for the path Vjt ■ Qeometry & Topology, Volume 6 (2002) 140 Benson Farb and Lee Mosher Proof of claim Given G J, consider the following {K, 0)-quasigeodesic in Hj'. Pj = hj^ij o pj_i : [0, Dj^i] -> Tij Since connection paths are geodesies, and since a,P are A-quasivertical, it fol- lows that the endpoint p'j{0) = hj-ij{a(j — 1)) and the corresponding endpoint Pj{0) = a{j) have distance in Hx at most A-|- 1, and similarly for the opposite endpoints p'-{Dj^i) = hj^ij{(3{j — 1)) and Pj{Dj) = (3{j). Each endpoint of p'^ and the corresponding endpoint of pj therefore have distance in Tij bounded by a constant depending only on A; this follows from Fact 5.1. Since the spaces Hj are all isometric to , it follows that the Hausdorff distance between pj and p'j in Tij is bounded by a constant depending only on K , A, which implies in turn that there is a quasi-isometric reparameterization rj : [0, -Dj-i] [0, Dj] such that dj{p'j{t),Pj{rj{t)))<D where the constant D and the quasi-isometry constants for Vj depend only on K, A. By possibly increasing the quasi-isometry constants we may assume fur- thermore that rj is an orientation preserving homeomorphism. It follows that we may connect the point pj-i{t) to the point Pj{rj{t)) by a A'-quasivertical path defined over the interval [j — C H, where A' depends only on K, A; when t = we may choose the path to be a | [j — 1, j] , and when t = -Dj-i we may choose the path /3 | [j — By piecing together these paths as j varies over J, we obtain the required family of paths v. □ We use (5-hyperbolicity of TCx in the following manner. First, for any geodesic rectangle a * b * c * d in TCx it follows that any point on a is within distance 26 of 6 U c U d. Second, for any (A',0) quasigeodesic in Hx, the Hausdorff distance to any geodesic with the same endpoints is bounded by a constant Si depending only on 5, X' . For any rectangle of the form v*a*w*a' where a, a' are geodesies and v,w are (A',0) quasigeodesics, it follows that any point on v is within distance 82 = 25 + 25i of aU wU a' . By Fact 5.1 there exists a constant ^3 such that: for all j & J,x,y & Hj, if d{x, y) < (1 + X')S2 then dj{x, y) < 63 We are now ready to define the flaring parameters K,n,A. Let 3 n=[S2 + 353\ + 1 A = 63 Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 141 where [x\ is the greatest integer < x . Assuming as we may that j± = jo ±n (and so the Hausdorff distance between Hj^ and Hj^ in Hx equals n), we must prove: • if Djg > A then max{Dj_ , > kDj^. Case 1 m8iX.{Dj_,Dj_^} < 663 It follows that there is a rectangle in Hx of the form a*a- * fd * a+ where a± is a geodesic in Tlx with the same endpoints as pj_j_, and where a± has length < 653. Consider now the point a{jo), whose distance from some point z e a- U (3 U a-^- is at most 62- li z e a- then it follows that 6(^3 d{a{jo),'Hj^) <S2 + — <n, a contradiction. We reach a similar contradiction if 2; G (J+. Therefore z e (3. It follows that z = P{s) G Tis for some s such that \s — jo\ < S2, and so by following along /3 a length at most X'62 we reach the point /3(jo) • This shows that d{a{jo), P{jo)) < (1 + X')S2, and so Dj^^ < 63, that is, Dj^ < A. Case 2 max{Dj , Dj^} > 35^ In the family v, we claim that there is a discrete subfamily a = vtQ,vtj^, . . . , vtj^ = /?, with to < ti < ■ ■ ■ < tx , such that the following property is satisfied: for each k = 1,... ,K, letting then we have max{Afc_,Afc+} G [3(53,653]. By assumption of Case 2, the subfamily {a = vto, P = vt^^} has the property max{Ajt_, Ajt+} = msLX.{Dj_, Dj_^} > 3Ss (for k = I). Suppose by induction that we have a subfamily a = vtQ,vt-i^, . . . , vtj^ = P , with to < ti < ■ ■ ■ < tx , such that max{A^._, A^^} > 36^ for all k, but suppose that max{A^._, A^.+} > 6(^3 for some k. If, say, Ay^^ > 6^3, then we subdivide the geodesic segment Pj+b*fe-i(i+)'^*fe(i+)] ^^^f ^ point t G pj_^_, yielding two subsegments of length > 3Ss , and we add the path Vj_^_t to our subfamily; similarly, if Aj._ > 6^3 then we subdivide the interval pj_ [vti^_i (j_), (i-)] i^i ^sili. This process must eventually stop, because thereby proving the claim. Geometry & Topology, Volume 6 (2002) 142 Benson Farb and Lee Mosher Prom the exact same argument as in Case 1, using the fact that max{Afe_, Afe+} < 6^3, it now follows that Afeo = (%_iOo),%Oo)) < S3 for all A; = 1,... We therefore have: K k=l K K Dj_ + Dj^ = ^ Afe_ + Afe+ > max{Afc_, Ak+} k=l k=l >K-3S3 3 max{i:)j_,Dj_^} > ^K63 - 2 JO This completes the proof of Lemma 5.2. Remark The argument given in Lemma 5.2, while stated explicitly only for groups of the form Tq, generalizes to a much broader context. Graphs of groups, the context for the Bestvina-Feighn combination theorem [6] and Ger- sten's converse [18], have been generalized to triangles of groups by Gersten and Stallings [46], and to general complexes of groups by Haefliger [20]. The arguments of Lemma 5.2 will also apply to show that a developable complex of groups with word hyperbolic fundamental group satisfies a flaring property over any geodesic in the universal covering complex. A converse would also be nice, giving a higher dimensional generalization of the Bestvina-Feighn combination theorem, but we do not know how to prove such a converse, nor do we have any examples to which it might apply (see Question 1.7 in the introduction). Next we have: Lemma 5.4 (Flaring implies hyperbolic) For each bounded subset B <Z M, each p> I, and each set of flaring data k > 1, n > 1, A{X), there exists S >0 such that the following holds. If 'j: I ^ T is a B -coboundcd, p-lipschitz, piecewise afEne path defined on a subinterval / C R, and if the metric Hbration H'y — > / satisfies K,n,A{X) vertical flaring, then H.^ is 5 -hyperbolic. Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 143 Proof This is basically an immediate application of the Bestvina-Feighn com- bination theorem [6]. To be formally correct, some remarks are needed to trans- late from our present geometric setting, of a hyperbolic plane bundle I , to the combinatorial setting of [6], and to justify that our vertical flaring prop- erty for H-y corresponds to the "hallways flare condition" of [6] . We may assume that the endpoints of the interval /, if any, are integers. The first observation is that there is a 7ri(S')-equivariant triangulation r of H-y with the following properties: Graph of spaces • For each n G J = I HZ there is a 2-dimensional subcomplex r„ which is a triangulation of the hyperbolic plane Hn ■ • Each 1-cell of r is either horizontal (a 1-cell of some r„ ) , or vertical (connecting a vertex of some r„ to a vertex of some r„+i); • each 2-cell of r is either horizontal (a 2-cell of some Tn), or vertical (meaning that the boundary contains exactly two vertical 1-cells) . Bounded combinatorics There is an upper bound depending only on B, p for the valence of each 0-cell and the number of sides of each 2-cell. Quasi-isometry The inclusion of the 1-skeleton of r into is a quasi- isometry with constants depending only on B and p . To see why r exists as described, consider the marked hyperbolic surface bundle ^ I . For each hyperbolic surface Sn, n & J , there is a geodesic triangulation Tn of Sn with One vertex, whose edges have length bounded only in terms of B . It follows that there are constants K' , C' depending only on B , such that if r„ is the lifted triangulation in , then the inclusion of the 1-skeleton of Tn into Tin is a {K',C') quasi-isometry. Then, regarding {JneJ^^>- '^^ triangulation of U neJ'^n, we can extend to a cell-decomposition r of which is a graph of spaces of bounded combinatorics. The existence of r uses the fact that each connection map /in,n+i • ^n+i is iiT-bilipschitz, so by moving each vertex of r„ along a connection path into 5^+1 and them moving a finite distance to a vertex of r„+i we obtain a {K" , C") -quasi-isometry h'^^n+i '■ —>■ T^n+i , with {K", C") depending only on K , and from this we easily construct r so that its lift r has the desired properties. The second observation is that vertical flaring in is equivalent to the "hall- way flare condition" of [6] for r , and this equivalence is uniform with respect to the parameters in each property. To see why, note that quasivertical paths in Ti^ correspond to thin paths in r as defined implicitly in [6] Section 2: an edge Geometry & Topology, Volume 6 (2002) 144 Benson Farb and Lee Mosher path a: I' = [m, n] — ^ r is p-thin if the restriction of a to each subinterval [i, i + 1] Hcs in T[j and is a concatenation of at most p edges. Under the quasi-isometry r H.^ and its coarse inverse H.^ — > r, A-quasivertical paths in correspond to p-thin paths with a uniform relation between A and p. In order to complete the translation from the geometric setting to the combi- natorial setting, while the results of [6] are stated only when r is the universal cover of a finite graph of spaces, nevertheless, the proofs hold as stated for any graph of spaces with uniformly bounded combinatorics: all the steps in the proof extend to such graphs of spaces, regardless of the presence of a deck transformation group with compact quotient. The conclusion of the combina- tion theorem is the 6' hyperbolicity of the 1-skeleton of r , with 5' depending only on the flaring constants for r, which depend in tTirn only on ;B, p, and the flaring constants for Ti^. It follows that Ti^ is 6 hyperbolic with the correct dependency for the constant 5. □ 5.2 Proof of Theorem 1.2 We adopt the notation from the beginning of Section 5: a homomorphism /: G — ^ MCG determining the group Fc, a Cayley graph X for G, and a piecewise affine /-equivariant map $ : X ^ T which is 6-cobounded and p- lipschitz. We have already proved, in Section 1.2, that word hyperbolicity of Tg implies finiteness of the kernel of / . Letting be the 0-skeleton, on which G acts transitiveily, it follows that $(X°) is an orbit of f{G) in T . We prove that /(G) is convex cocompact by proving that satisfies orbit quasiconvexity. Choose two points x, y G X^ . Let 7 : 7 X be a geodesic segment connecting X to y . Consider the composed path I — > X — > T , which by abuse of notation we shall also denote 7. There is a corresponding hyperbolic plane bundle I . Recall that 7 is ;B-cobounded and p-lipschitz in T, with B, p independent of 7. Now apply Lemmas 5.2 and 5.4, to conclude that Hj is 5-hyperbolic, with S independent of 7. Now we quote the following result to obtain a Teichmiiller geodesic: Theorem 5.5 [41] For every bounded set B C Ai , p > 1, and 6 > 0, there exists A > 1, e > 0, and A such that the following hold. If j: I ^ T is B-cobounded and p-lipschitz, and if Ti.^ is 6 -hyperbolic, then 7 is a (A, e)- quasigeodesic, and there exists a Teichmiiller geodesic g, sharing any endpoints of 7 , such that 7 and g have Hausdorff distance at most A . □ Qeometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 145 Letting g be the Teichmiiller geodesic connecting x to y provided by the theo- rem, it fohows that g is contained in the A + p neighborhood of ^{X^). Since x,y G are arbitrary, this proves orbit quasiconvexity, and so f{G) is convex cocompact. 6 Schottky groups Definition A Schottky subgroup of MCG is a free, convex cocompact sub- group. The hmit set A C ^M.T of a Schottky subgroup is therefore a Cantor set, and every nontrivial element is pseudo-Anosov. In this section we prove Theorem 1.3, that a surfacc-by-frcc group is word hyperbolic if and only if the free group is Schottky. One direction is already proved by Theorem 1.2, and so we need only prove that when F C MCG is a Schottky subgroup then Vp tti{S) xi F is word hyperbolic. Continuing with earlier notation, let A C PMJ^ be the limit set of F with weak hull WH^. Let t be a Cayley graph for the group F, a tree on which F acts properly discontinuously with quotient a rose. Let t ^ T be an F-equivariant map, affine on each edge, and p-lipschitz for some p > 1. There is a bounded subset B C M so that both WHa and $(t) are H-cobounded. We have a hyperbolic plane bundle Ht — > t, on which 7ri(5) xi F acts properly discontinuously and cocompactly, and we have a piecewise Riemannian metric on Til on which 7ri(S') xi F acts by isometrics. We must prove that Tit is 5 -hyperbolic. By the Bestvina-Feighn combina- tion theorem [6], it is enough to show that for each bi-infinite geodesic 7 in t, the bundle Hj — R satisfies vertical flaring, with flaring data K,n,A{X) independent of the choice of 7 (see the proof of Lemma 5.4 for translating the combinatorial setting of [6] to our present geometric setting). Since F is convex cocompact, there is a geodesic line g in WHa which has finite Hausdorff distance from ^>(7). Let 7ig"^^ be the singular SOLV-space thereby obtained. By Proposition 4.2, the closest point map ^ ^ g lifts to a quasi- isometry T-i-./ 7i^°^^ , with quasi-isometry constants independent of 7, depending only on B and p. It therefore suffices to check the flaring condition in 7Y^°^^, with flaring data independent of anything. 2 Take any k with 1 < k < i^^-, say k = 2.6. Let n = 2. We show that for any A there is an A such that any two A quasivertical lines in "W^^^ satisfy Geometry & Topology, Volume 6 (2002) 146 Benson Farb and Lee Mosher the (k, 2, ^) -flaring condition. For this argument we do not need that g is cobounded (although in that case may not have bounded geometry). Let a, a' : [—2, 2] Ti.g°^^ be two A quasivertical Hnes, lying over a length 4 subsegment [r — 2, r + 2] of (7 ~ R. Let Xi,yi be the points where a, a' respectively intersect 7ir+i ■ Let ^0 = xq and let be obtained by flowing xq vertically into Hr+i] define rjo = yo and rji similarly. Note that for i G [—2,2] the points and Xi are connected in Hg°^^ by a path which goes along a from Ci to ^0 travelling a distance at most 2A, and then vertically from ,^0 = to Xi; the vertical projection of this path into Hi has length at most 2e^A, and so di{xi,^i) < 2e^A. Similarly, di{yi,r]i) <2e'^X. We turn for the moment to showing that the sequence dr+i{Ci,Vi), ^ = -2,-1,0,1,2 2 satisfies the (^^, 2, 0) -flaring condition. In the singular Euclidean surface Hr+i, let £i be the geodesic from ^i to rji, so the above sequence becomes: len(£i), i = -2,-1,0,1,2 The singular Euclidean geodesic £0 is a concatenation of subsegments of con- stant slope, two consecutive subsegments meeting at a singularity. If at least half of io has slope of absolute value > 1 then: ^len(4)-^-e2<len(£2) If at least half of £0 has slope of absolute value < 1 , we get a similar inequality but with len(£_2) on the right hand side. We have therefore shown: e2 max{dr+2{^2,m),dr-2{^-2,r]-2)} > -^-j=doi^Q,r]o) It follows that e2 max{dr+2ix2,y2),dr-2ix-2,y-2)} > ■^-y=doixo,yo) -2e A > Kdo(xo,yo) where the last inequality holds as long as: "o(a;o, yo) > ^ = 2V2 "'^ This ends the proof that 7ri(5) xi F is word hyperbolic when F is Schottky. Qeometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 147 7 Extending the theory to orbifolds In this section we sketch how the theory can be extended to 2-dimensional orbifolds. We shall consider only those compact orbifolds whose underlying 2-mamfold is closed, and whose orbifold locus therefore consists only of cone points, what we shall call a cone orbifold. The reason for this restriction is that if the underlying 2-manifold has nonempty boundary then the orbifold does not support any pseudo-Anosov homeomorphisms, since the isotopy classes of the boundary curves must be permuted.^ As it turns out, the mapping class group and Teichmiiller space of a cone orbifold depend not on the actual orders of the different cone points, but only on the partition of the set of cone points into subsets of constant order. For example, a spherical orbifold with one Z/2 cone point and three Z/4 cone points has the same mapping class group and Teichmiiller space as a spherical orbifold with three Z/42 cone points and one Z/1000 cone point. The relevant structures can therefore be described more directly and economically in the following manner. Let 5 be a closed surface, not necessarily orientable. Let P = {Pj}jg7 be a finite, pairwise disjoint collection of finite, nonempty subsets of S. Let Homeo(S', P) be the group of homeomorphisms of S which leave invariant each of the sets Pi, i e I. Let Homeoo(S', P) be the component of the iden- tity of Homeo(<S', P) with respect to the compact open topology; equivalently, Homeoo(S', P) consists of all elements of Homeo(S', P) which are isotopic to the identity through elements of Homeo(S', P) . The mapping class group is MCG{S,P) = Homeo(5,P)/Homeoo(5,P). To define the Teichmiiller space, first we must widen the concept of a conformal structure so that it applies to non-orientable surfaces, and we do this by allowing overlap maps which arc anticonformal as well as conformal. The Teichmiiller space T{S, P) is then defined to be the set of conformal structures on S modulo the action of Homeoo(>S', P) . Quadratic differentials and measured foliations on (S*, P) are defined using the usual local models at points oi S — UP, but at a point of P a quadratic differential can have the local model z'^^'^dz^ for any n > 1; the horizontal measured foliation of z'^^'^dz'^ is the local model for an n-pronged singularity of a measured foliation. Thus, at a point of UP a measured foliation can have any number of prongs > 1 , whereas a singularity in S — UP must have > 3 prongs as usual. With these definitions, Teichmiiller maps are defined as usual, making T{S, P) into a proper geodesic metric space ''While the monograph [16] develops a kind of pseudo-Anosov theory on a bounded surface, it is not appropriate for our present purposes. Geometry & Topology, Volume 6 (2002) 148 Benson Farb and Lee Mosher on which MCG{S, P) acts properly discontinuously, but not cocompactly; also, pseudo-Anosov homeomorphisms of {S, P) are defined as usual. We shall assume that {S, P) actually supports a pseudo-Anosov homeomor- phism which has an n-pronged singularity with n =^ 2. This rules out a small number of special cases, as follows. When 5 is a sphere, UP must have at least four points. When S is a projective plane, UP must have at least two points. When 5 is a torus or Klein bottle, UP must have at least one point. When S is the surface of Euler characteristic — 1 , namely the connected sum of a torus and a projective plane, the curve along which the torus and the projective plane are glued is actually a characteristic curve for S, meaning that it is preserved up to isotopy by any mapping class; therefore, in order for (S, P) to support a pseudo-Anosov homeomorphism, UP must have at least one point. Now we apply these concepts to 2-dimensional cone orbifolds. Suppose O is a cone orbifold with underlying surface S . Let P„ be the set of Z/n cone points, and let P = {Pn}n>2- Then we may define the mapping class group MCG{0) to be MCG(S',P), and the Teichmiiller space T{0) to be 7(5", P). Note that with the restrictions above on the type of (S", P) , the orbifold O has negative Euler characteristic. It follows that if O — > O is the orbifold universal covering map, then for any conformal structure on O the lifted conformal structure is isomorphic to the Riemann disc. It follows that any conformal structure on O can be TmiqTicly uniformized to produce a hyperbolic structure, with a cone angle of 27r/n at each Z/n cone point. At this stage we must confront the fact that the universal extension for surface groups, as formulated in Section 1.2, must be reformulated before it can be applied to orbifolds. The Dehn-Nielsen-Baer-Epstein theorem is still true, as long as one uses orbifold fundamental groups: if p is a generic point of the cone orbifold O, and if ■Ki[0,p) is the orbifold fundamental group, then we have MCG{0) k, 0\xt{'Ki{0,p)). However, the "once-punctured" mapping class group MCG{0,p) is not isomorphic to Aut(7ri(0,p)) . For example, take a based simple loop i which bounds a disc whose interior contains a single Z/n cone point. In the group 7ri{0,p) , the loop i represents an element of order n, and under the usual injection 7ri{0,p) ^ Aut(7ri(0,p)) we obtain an element of order n. However, the element of MCG{0,p) obtained by pushing p around £ has infinite order in MCG{0,p). To repair this we need another group to take over the role of MCG{0,p). Let Homeo(C') denote the group of homeomorphisms of O which are lifts of homeomorphisms of O, that is, a homeomorphism f : O ^ O is in the group Homeo(O) if and only if there exists a homeomorphism f : O ^ O such that Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 149 the following diagram commutes: O f O With respect to the compact open topology, Homeo(O) becomes a topological group. Let Homeoo(C) be the component of the identity Homeo(O) . Equiv- alently, Homeoo(O) is the subgroup of elements of Homeo(O) isotopic to the identity through elements of Homeo(C) ; alternatively it is the subgroup of Homeo(O) acting trivially on the circle at infinity of O ^ . Define MCG{0) = Hotmo(C>)/Ho^oo(C'). Note that universal covering map O ^ O induces a surjective homomorphism MCG{0) MCG{0), and the kernel is the group of deck transformations, isomorphic to 7Ti{0). We now have a natural isomorphism of short exact se- quences ■7ri(0) ^MCG{0) MCG{0) -^1 ■7ri(0) Aut(7ri(0)) > Out(7ri(0)) -^1 where we have suppressed the generic base point needed to define -KiiO). We are now in a position to state that our main results, Theorem 1.1, 1.2, 1.3, and 1.4, are true with the orbifold O in place of the surface 5, and the proofs are unchanged. Although the references that we quote are stated solely in terms of surfaces, namely [38] and [32] for Theorem 1.1, [39] for Theorem 1.2, and [40] for Theorem 1.4, nevertheless all the proofs in those references work just as well for orbifolds instead of surfaces. References [1] W Abikoff, The real analytic theory of Teichmuller space, volume 820 of Lecture Notes in Mathematics, Springer (1980) [2] G N Arzhantseva, On quasiconvex subgroups of word hyperbolic groups, Ge- ometriae Dedicata 98 (2001) 191-208 Qeometry & Topology, Volume 6 (2002) 150 Benson Farb and Lee Mosher [3] R Baer, Isotopien von Kurven auf orientierbaren, geschlossenen Fldchen und ihr Zusammenhang mit der topologischen Deformation der Fldchen, J. Reine Angew. Math. 159 (1928) 101-116 L Bers, Fiber spaces over Teichmuller spaces, Acta Math. 130 (1973) 89-126 L Bers, An extremal problem for quasiconformal mappings and a theorem by Thurston, Acta Math. 141 (1978) 73 98 M Bestvina, M Feighn, A combination theorem for negatively curved groups, J. Diff. Geom. 35 (1992) 85-101 J Birman, Braids, links, and mapping class groups, volume 82 of Annals of Math. Studies, Princeton University Press (1974) J Birman, A Lubotzky, J McCarthy, Abelian and solvable subgroups of the mapping class groups, Duke Math. J. 50 (1983) 1107-1120 N Brady, Branched coverings of cubical complexes and subgroups of hyperbolic groups, J. London Math. Soc. (2) 60 (1999) 461-480 J Brock, B Farb, Curvature and rank of Teichmuller space (2001), preprint, arXiv : math . GT/0109045 R D Canary, Covering theorem,s for hyperbolic 3 manifolds, preprint J Cannon, The theory of negatively curved spaces and groups, from: "Er- godic theory, symboHc dynamics, and hyperbolic spaces", (C Series T Bedford, MKeane, editor), Oxford Univ. Press (1991) DBA Epstein, Curves on 2-manifolds and isotopies. Acta Math. 115 (1966) 83-107 B Farb, A Lubotzky, Y Minsky, Rank one phenomena for mapping class groups, Duke Math. J. 106 (2001) 581-597 B Farb, L Mosher, The geometry of surface-by-free groups, Geom. Punct. Anal. (2002), to appear, preprint, arXiv:math.GR/0008215 A Fathi, F Laudenbach, V Poenaru, et al., Travaux de Thurston sur les surfaces, volume 66-67 of Asterisque, Societe Mathematique de France (1979) F Gardiner, H Masur, Extremal length geometry of Teichmuller space. Com- plex Variables Theory Appl. 16 (1991) 209-237 S Gersten, Cohomological lower bounds for isoperimetric functions on groups. Topology 37 (1998) 1031-1072 G Gonzalez-Diez, WJ Harvey, Surface subgroups inside mapping class groups. Topology 38 (1999) 57-69 A Haefliger, Complexes of groups and orbihedra, from: "Group theory from a geometrical viewpoint (Trieste, 1990)", World Sci. Publishing, River Edge, NJ (1991) 504-540 J Hubbard, H Maisur, Quadratic differentials and foliations. Acta Math. 142 (1979) 221-274 Geometry & Topology, Volume 6 (2002) Convex cocompact subgroups of mapping class groups 151 [22] Y Imayoshi, M Taniguchi, An introduction to Teichmiiller spaces, Springer (1992) [23] N V Ivanov, Subgroups of TeichmiiUer modular groups, volume 115 of Trans- lations of Mathematical Monographs, Amer. Math. Soc. (1992) [24] N V Ivanov, Automorphisms of complexes of curves and Teichmiiller spaces, from: "Progress in knot theory and related topics", Travaux en Cours 56, Her- mann, Paris (1997) 113 120 [25] M Kapovich, On normal subgroups in the fundamental groups of complex sur- faces (1998), preprint, arXiv : math . GT/9808085 [26] S Kerckhoff, The asymptotic geometry of Teichmiiller space, Topology 19 (1980) 23-41 [27] B Maskit, Comparison of hyperbolic and extremal lengths, Ann. Acad. Sci. Fenn. Scries A I Math. 10 (1985) 381-386 [28] H Masur, Uniquly ergodic quadratic differentials. Comment. Math. Helv. 55 (1980) 255-266 [29] H Masur, Interval exchange transformations and measured foliations, Ann. of Math. 115 (1982) 169 200 [30] H Masur, Two boundaries of Teichmiiller space, Duke Math. J. 49 (1982) 183- 190 [31] H Masur, Hausdorff dimension of the set of nonergodic foliations of a quadratic differential, Duke Math. J. 66 (1992) 387-442 [32] H Masur, Y Minsky, Geometry of the complex of curves, I. Hyperbolicity, Invent. Math. 138 (1999) 103 149 [33] H Masur, Y Minsky, Unstable quasigeodesics in Teichmiiller space, from: "In the tradition of Ahlfors and Bers (Stony Brook, NY, 1998)" , Contemp. Math. 256, Amer. Math. Soc. (2000) 239-241 [34] H Masur, M Wolf, TeichmAiller space is not Gromov hyperbolic, Ann. Acad. Sci. Fenn. Ser. A I Math. 20 (1995) 259 267 [35] J McCarthy, A "Tits-alternative" for subgroups of surface mapping class groups. Trans. AMS 291 (1985) 582-612 [36] J McCarthy. A Papadopoulos, Dynamics on Thurston's sphere of projective measured, foliations. Comment. Math. Helv. 64 (1989) 133-166 [37] Y Minsky, On rigidity, limit sets, and end invariants of hyperbolic 3-manifolds, Jour. AMS 7 (1994) 539—588 [38] Y Minsky, Quasi-projections in Teichmiiller space, J. Reine Angew. Math. 473 (1996) 121-136 [39] L Mosher, Hyperbolic extensions of groups, J. Pure and Appl. Alg. 110 (1996) 305-314 Geometry & Topology, Volume 6 (2002) 152 Benson Farb and Lee Mosher [40] L Mosher, A hyperbolic-by-hyperbolic hyperbolic group, Proc. AMS 125 (1997) 3447-3455 [41] L Mosher, Stable quasigeodesics in Teichmiiller space and ending laminations (2001), preprint, arXiv:math.GR/0107035 [42] D Mumford, A remark on Mahler's compactness theorem, Proc. Amer. Math. Sec. 28 (1971) 289-294 [43] J Nielsen, Untersuchungen zur Topologie der geschlossenen zweiseitigen Flachen, Acta Math. 50 (1927) 189-358 [44] J-P Otal, Le theoreme d'hyperbolisation pour les varietes fibrees de dimension 3, Asterisque 235, Societe Mathematique de France (1996) [45] H Royden, Report on the Teichmiiller metric, Proc. Nat. Acad. Sci. U.S.A. 65 (1970) 497-499 [46] J StalHngs, Non-positively curved triangles of groups, from: "Group theory from a geometrical viewpoint (Trieste, 1990)", World Sci. Publishing (1991) 491 503 [47] K Whittlesey, Normal all pseudo-Anosov subgroups of mapping class groups. Geometry and Topology 4 (2000) 293-307 Geometry & Topology, Volume 6 (2002)