The geometry of the disk complex

We give a distance estimate for the metric on the disk complex and show that it is Gromov hyperbolic. As another application of our techniques, we find an algorithm which computes the Hempel distance of a Heegaard splitting, up to an error depending only on the genus.

1. Introduction 2 2. Background on complexes 4 3. Background on coarse geometry 7 4. Natural maps 10 5. Holes in general and the lower bound on distance 13 6. Holes for the non-orientable surface 16 7. Holes for the arc complex 17 8. Background on three-manifolds 18 9. Holes for the disk complex 20 10. Holes for the disk complex -annuli 22 11. Holes for the disk complex -compressible 24

Introduction
In this paper we initiate the study of the geometry of the disk complex of a handlebody V . The disk complex D(V ) has a natural simplicial inclusion into the curve complex C(S) of the boundary of the handlebody. Surprisingly, this inclusion is not a quasi-isometric embedding; there are disks which are close in the curve complex yet very far apart in the disk complex. As we will show, any obstruction to joining such disks via a short path is a topologically meaningful subsurface of S = ∂V . We call such subsurfaces holes. A path in the disk complex must travel into and then out of these holes; paths in the curve complex may skip over a hole by using the vertex representing the boundary of the subsurface. We classify the holes: Theorem 1.1. Suppose V is a handlebody. If X ⊂ ∂V is a hole for the disk complex D(V ) of diameter at least 61 then: • X is not an annulus.
• If X is compressible then there are disks D, E with boundary contained in X so that the boundaries fill X. • If X is incompressible then there is an I-bundle ρ F : T → F so that T is a component of V ∂ v T and X is a component of ∂ h T .
See Theorems 10.1, 11.6 and 12.1 for more precise statements. The Ibundles appearing in the classification lead us to study the arc complex A(F ) of the base surface F . Since the I-bundle T may be twisted the surface F may be non-orientable.
Thus, as a necessary warm-up to the difficult case of the disk complex, we also analyze the holes for the curve complex of an non-orientable surface, as well as the holes for the arc complex.
Topological application. It is a long-standing open problem to decide, given a Heegaard diagram, whether the underlying splitting surface is reducible. This question has deep connections to the geometry, topology, and algebra of the ambient three-manifold. For example, a resolution of this problem would give new solutions to both the threesphere recognition problem and the triviality problem for three-manifold groups. The difficulty of deciding reducibility is underlined by its connection to the Poincaré conjecture: several approaches to the Poincaré Conjecture fell at essentially this point. See [10] for a entrance into the literature.
One generalization of deciding reducibility is to find an algorithm that, given a Heegaard diagram, computes the distance of the Heegaard splitting as defined by Hempel [20]. (For example, see [5,Section 2].) The classification of holes for the disk complex leads to a coarse answer to this question.
Theorem 21.1. In every genus g there is a constant K = K(g) and an algorithm that, given a Heegaard diagram, computes the distance of the Heegaard splitting with error at most K.
In addition to the classification of holes, the algorithm relies on the Gromov hyperbolicity of the curve complex [24] and the quasi-convexity of the disk set inside of the curve complex [26]. However the algorithm does not depend on our geometric applications of Theorem 1.1.
Geometric application. The hyperbolicity of the curve complex and the classification of holes allows us to prove: Theorem 20.3. The disk complex is Gromov hyperbolic.
Again, as a warm-up to the proof of Theorem 20.3 we prove that C(F ) and A(S) are hyperbolic in Corollary 6.4 and Theorem 20.2. Note that Bestvina and Fujiwara [4] have previously dealt with the curve complex of a non-orientable surface, following Bowditch [6].
These results cannot be deduced from the fact that D(V ), C(F ), and A(S) can be realized as quasi-convex subsets of C(S). This is because the curve complex is locally infinite. As simple example consider the Cayley graph of Z 2 with the standard generating set. Then the cone C(Z 2 ) of height one-half is a Gromov hyperbolic space and Z 2 is a quasi-convex subset. Another instructive example, very much in-line with our work, is the usual embedding of the three-valent tree T 3 into the Farey tessellation.
The proof of Theorem 20.3 requires the distance estimate Theorem 19.1: the distance in C(F ), A(S), and D(V ) is coarsely equal to the sum of subsurface projection distances in holes. However, we do not use the hierarchy machine introduced in [25]. This is because hierarchies are too flexible to respect a symmetry, such as the involution giving a non-orientable surface, and at the same time too rigid for the disk complex. For C(F ) we use the highly rigid Teichmüller geodesic machine, due to Rafi [33]. For D(V ) we use the extremely flexible train track machine, developed by ourselves and Mosher [27].
Theorems 19.1 and 20.3 are part of a more general framework. Namely, given a combinatorial complex G we understand its geometry by classifying the holes: the geometric obstructions lying between G and the curve complex. In Sections 13 and 14 we show that any complex G satisfying certain axioms necessarily satisfies a distance estimate. That hyperbolicity follows from the axioms is proven in Section 20.
Our axioms are stated in terms of a path of markings, a path in the the combinatorial complex, and their relationship. For the disk complex the combinatorial paths are surgery sequences of essential disks while the marking paths are provided by train track splitting sequences; both constructions are due to the first author and Minsky [26] (Section 18). The verification of the axioms (Section 19) relies on our work with Mosher, analyzing train track splitting sequences in terms of subsurface projections [27].
We do not study non-orientable surfaces directly; instead we focus on symmetric multicurves in the double cover. This time marking paths are provided by Teichmüller geodesics, using the fact that the symmetric Riemann surfaces form a totally geodesic subset of Teichmüller space. The combinatorial path is given by the systole map. We use results of Rafi [33] to verify the axioms for the complex of symmetric curves. (See Section 16.) Section 17 verifies the axioms for the arc complex again using Teichmüller geodesics and the systole map. It is interesting to note that the axioms for the arc complex can also be verified using hierarchies or, indeed, train track splitting sequences.
The distance estimates for the marking graph and the pants graph, as given by the first author and Minsky [25], inspired the work here, but do not fit our framework. Indeed, neither the marking graph nor the pants graph are Gromov hyperbolic. It is crucial here that all holes interfere; this leads to hyperbolicity. When there are non-interfering holes, it is unclear how to partition the marking path to obtain the distance estimate.

Background on complexes
We use S g,b,c to denote the compact connected surface of genus g with b boundary components and c cross-caps. If the surface is orientable we omit the subscript c and write S g,b . The complexity of S = S g,b is ξ(S) = 3g − 3 + b. If the surface is closed and orientable we simply write S g .
2.1. Arcs and curves. A simple closed curve α ⊂ S is essential if α does not bound a disk in S. The curve α is non-peripheral if α is not isotopic to a component of ∂S. A simple arc β ⊂ S is proper if β ∩ ∂S = ∂β. An isotopy of S is proper if it preserves the boundary setwise. A proper arc β ⊂ S is essential if β is not properly isotopic into a regular neighborhood of ∂S.
Define C(S) to be the set of isotopy classes of essential, non-peripheral curves in S. Define A(S) to be the set of proper isotopy classes of essential arcs. When S = S 0,2 is an annulus define A(S) to be the set of essential arcs, up to isotopies fixing the boundary pointwise. For any surface define AC(S) = A(S) ∪ C(S).
For α, β ∈ AC(S) the geometric intersection number ι(α, β) is the minimum intersection possible between α and any β equivalent to β. When S = S 0,2 we do not count intersection points occurring on the boundary. If α and β realize their geometric intersection number then α is tight with respect to β. If they do not realize their geometric intersection then we may tighten β until they do.
Define ∆ ⊂ AC(S) to be a multicurve if for all α, β ∈ ∆ we have ι(α, β) = 0. Following Harvey [18] we may impose the structure of a simplical complex on AC(S): the simplices are exactly the multicurves. Also, C(S) and A(S) naturally span sub-complexes.
Note that the curve complexes C(S 1,1 ) and C(S 0,4 ) have no edges. It is useful to alter the definition in these cases. Place edges between all vertices with geometric intersection exactly one if S = S 1,1 or two if S = S 0,4 . In both cases the result is the Farey graph. Also, with the current definition C(S) is empty if S = S 0,2 . Thus for the annulus only we set AC(S) = C(S) = A(S).
Definition 2.1. For vertices α, β ∈ C(S) define the distance d S (α, β) to be the minimum possible number of edges of a path in the one-skeleton C 1 (S) which starts at α and ends at β.
Note that if d S (α, β) ≥ 3 then α and β fill the surface S. We denote distance in the one-skeleton of A(S) and of AC(S) by d A and d AC respectively. Recall that the geometric intersection of a pair of curves gives an upper bound for their distance.

Subsurfaces.
Suppose that X ⊂ S is a connected compact subsurface. We say X is essential exactly when all boundary components of X are essential in S. We say that α ∈ AC(S) cuts X if all representatives of α intersect X. If some representative is disjoint then we say α misses X.
Definition 2.4. An essential subsurface X ⊂ S is cleanly embedded if for all components δ ⊂ ∂X we have: δ is isotopic into ∂S if and only if δ is equal to a component of ∂S.
Definition 2.5. Suppose X, Y ⊂ S are essential subsurfaces. If X is cleanly embedded in Y then we say that X is nested in Y . If ∂X cuts Y and also ∂Y cuts X then we say that X and Y overlap.
A compact connected surface S is simple if AC(S) has finite diameter.
Lemma 2.6. Suppose S is a connected compact surface. The following are equivalent: • S is not simple.
• The diameter of AC(S) is at least five. • S admits an ending lamination or S = S 1 or S 0,2 . • S admits a pseudo-Anosov map or S = S 1 or S 0,2 .
Lemma 4.6 of [24] shows that pseudo-Anosov maps have quasi-geodesic orbits, when acting on the associated curve complex. A Dehn twist acting on C(S 0,2 ) has geodesic orbits.
Note that Lemma 2.6 is only used in this paper when ∂S is non-empty. The closed case is included for completeness.
Proof sketch of Lemma 2.6. If S admits a pseudo-Anosov map then the stable lamination is an ending lamination. If S admits a filling lamination then, by an argument of Kobayashi [21], AC(S) has infinite diameter. (This argument is also sketched in [24], page 124, after the statement of Proposition 4.6.) If the diameter of AC is infinite then the diameter is at least five. To finish, one may check directly that all surfaces with χ(S) > −2, other than S 1,1 , S 1 and the annulus have AC(S) with diameter at most four. (The difficult cases, S 012 and S 003 , are discussed by Scharlemann [35].) Alternatively, all surfaces with χ(S) < −1, and also S 1,1 , admit pseudo-Anosov maps. The orientable cases follow from Thurston's construction [38]. Penner's generalization [32] covers the non-orientable cases.

2.3.
Handlebodies and disks. Let V g denote the handlebody of genus g: the three-manifold obtained by taking a closed regular neighborhood of a polygonal, finite, connected graph in R 3 . The genus of the boundary is the genus of the handlebody. A properly embedded disk D ⊂ V is essential if ∂D ⊂ ∂V is essential.
Let D(V ) be the set of essential disks D ⊂ V , up to proper isotopy. A subset ∆ ⊂ D(V ) is a multidisk if for every D, E ∈ ∆ we have ι(∂D, ∂E) = 0. Following McCullough [28] we place a simplical structure on D(V ) by taking multidisks to be simplices. As with the curve complex, define d D to be the distance in the one-skeleton of D(V ).

Markings.
A finite subset µ ⊂ AC(S) fills S if for all β ∈ C(S) there is some α ∈ µ so that ι(α, β) > 0. For any pair of finite subsets µ, ν ⊂ AC(S) we extend the intersection number: We say that µ, ν are L-close if ι(µ, ν) ≤ L. We say that µ is a Kmarking if ι(µ, µ) ≤ K. For any K, L we may define M K,L (S) to be the graph where vertices are filling K-markings and edges are given by L-closeness.
As defined in [25] we have: • A collection of base curves base(µ) = {α i }: a maximal simplex in C(S). • A collection of transversal curves {β i }: for each i define X i = S j =i α j and take β i ∈ C(X i ) to be a Farey neighbor of α i . If µ is a complete clean marking then ι(µ, µ) ≤ 2ξ(S) + 6χ(S). As discussed in [25] there are two kinds of elementary moves which connected markings. There is a twist about a pants curve α, replacing its transversal β by a new transversal β which is a Farey neighbor of both α and β. We can flip by swapping the roles of α i and β i . (In the case of the flip move, some of the other transversals must be cleaned.) It follows that for any surface S there are choices of K, L so that M(S) is non-empty and connected. We use d M (µ, ν) to denote distance in the marking graph.

Background on coarse geometry
Here we review a few ideas from coarse geometry. See [8], [12], or [15] for a fuller discussion.
3.1. Quasi-isometry. Suppose r, s, A are non-negative real numbers, with A ≥ 1. If s ≤ A · r + A then we write s ≤ A r. If s ≤ A r and r ≤ A s then we write s = A r and call r and s quasi-equal with constant A. We also define the cut-off function . The relation f is a quasi-isometry, and X is quasi-isometric to Y, if f is an A-quasi-isometric embedding and the image of f is A-dense: the A-neighborhood of the image equals all of Y.
Often the exact choice of f is unimportant and all that matters are the endpoints x and y. We then denote the image of f by (Compare to the definition of (K, δ, s)-quasi-geodesics found in [24].) A subset Y ⊂ X is Q-quasi-convex if every X -geodesic connecting a pair of points of Y lies within a Q-neighborhood of Y.
3.3. Hyperbolicity. We now assume that X is a connected graph with metric induced by giving all edges length one.
Definition 3.1. The space X is δ-hyperbolic if, for any three points x, y, z in X and for any geodesics k = [x, y], g = [y, z], h = [z, x], the triangle ghk is δ-slim: the δ-neighborhood of any two sides contains the third.
An important tool for this paper is the following theorem of the first author and Minsky [24]: Theorem 3.2. The curve complex of an orientable surface is Gromov hyperbolic.
For the remainder of this section we assume that X is δ-hyperbolic graph, x, y, z ∈ X are points, and k = [x, y], g = [y, z], h = [z, x] are geodesics. Definition 3.3. We take ρ k : X → k to be the closest points relation: We now list several lemmas useful in the sequel.
Lemma 3.4. There is a point on g within distance 2δ of ρ k (z). The same holds for h.
Lemma 3.5. The closest points ρ k (z) have diameter at most 4δ.
Lemma 3.8. Suppose that k is another geodesic in X so that the endpoints of k are within distance R of the points x and y. Then We now turn to a useful consequence of the Morse stability of quasigeodesics in hyperbolic spaces. Lemma 3.9. For every δ and A there is a constant C with the following property: If X is δ-hyperbolic and g : [0, N ] → X is an Aunparameterized quasi-geodesic then for any m < n < p in [0, N ] we have: d X (x, y) + d X (y, z) < d X (x, z) + C where x, y, z = g(m), g(n), g(p).

A hyperbolicity criterion.
Here we give a hyperbolicity criterion tailored to our setting. We thank Brian Bowditch for both finding an error in our first proof of Theorem 3.11 and for informing us of Gilman's work [13,14].
Suppose that X is a graph with all edge-lengths equal to one. Suppose that γ : [0, N ] → X is a loop in X with unit speed. Any pair of points a, b ∈ [0, N ] gives a chord of γ. If a < b, N/4 ≤ b − a and N/4 ≤ a + (N − b) then the chord is 1/4-separated. The length of the chord is d X (γ(a), γ(b)).
Following Gilman's proof goes via the subquadratic isoperimetric inequality. We now give our criterion, noting that it is closely related to another paper of Gilman [14]. Theorem 3.11. Suppose that X is a graph with all edge-lengths equal to one. Then X is Gromov hyperbolic if and only if there is a constant M ≥ 0 and, for all unordered pairs x, y ∈ X 0 , there is a connected subgraph g x,y containing x and y with the following properties: • (Local) If d X (x, y) ≤ 1 then g x,y has diameter at most M .
• (Slim triangles) For all x, y, z ∈ X 0 the subgraph g x,y is contained in an M -neighborhood of g y,z ∪ g z,x .
Proof. Suppose that γ : [0, N ] → X is a loop. If is the empty string let I = [0, N ]. For any binary string ω let I ω0 and I ω1 be the first and second half of I ω . Note that if |ω| ≥ log 2 N then |I ω | ≤ 1. Fix a string ω and let [a, b] = I ω . Let g ω be the subgraph connecting γ(a) to γ(b). Note that g 0 = g 1 because γ(0) = γ(N ). Also, for any binary string ω the subgraphs g ω , g ω0 , g ω1 form an M -slim triangle. If |ω| ≤ log 2 N then every x ∈ g ω has some point b ∈ I ω so that Since g 0 is connected there is a point x ∈ g 0 that lies within the M -neighborhoods both of g 00 and of g 01 . Pick some b ∈ I 1 so that d X (x, γ(b)) is bounded as in the previous paragraph. It follows that there is a point a ∈ I 0 so that a, b are 1/4-separated and so that Thus there is an additive error K large enough so that X satisfies the criterion of Theorem 3.10 and we are done.

Natural maps
There are several natural maps between the complexes and graphs defined in Section 2. Here we review what is known about their geometric properties, and give examples relevant to the rest of the paper. 4.1. Lifting, surgery, and subsurface projection. Suppose that S is not simple. Choose a hyperbolic metric on the interior of S so that all ends have infinite areas. Fix a compact essential subsurface X ⊂ S which is not a peripheral annulus. Let S X be the cover of S so that X lifts homeomorphically and so that S X ∼ = interior(X). For any α ∈ AC(S) let α X be the full preimage.
Since there is a homeomorphism between X and the Gromov compactification of S X in a small abuse of notation we identify AC(X) with the arc and curve complex of S X .
Definition 4.1. We define the cutting relation κ X : AC(S) → AC(X) as follows: α ∈ κ X (α) if and only if α is an essential non-peripheral component of α X .
Note that α cuts X if and only if κ X (α) is non-empty. Now suppose that S is not an annulus.
Definition 4.2. We define the surgery relation σ X : AC(S) → C(S) as follows: α ∈ σ S (α) if and only if α ∈ C(S) is a boundary component of a regular neighborhood of α ∪ ∂S.
With S and X as above: The subsurface projection relation π X : AC(S) → C(X) is defined as follows: If X is not an annulus then define π X = σ X • κ X . When X is an annulus π X = κ X .
See Lemma 2.3 of [25] and the remarks in the section Projection Bounds in [29].
It is crucial to note that if some vertex of {β i } misses X then the projection distance d X (β 0 , β n ) may be arbitrarily large compared to n. Corollary 4.5 can be greatly strengthened when the path is a geodesic [25]: Here is a converse for Lemma 4.4.
Lemma 4.7. For every a ∈ N there is a number b ∈ N with the following property: for any α, Corollary D of [11] gives a more precise relation between projection distance and intersection number.
Proof of Lemma 4.7. We only sketch the contrapositive: Suppose we are given a sequence of curves α n , β n so that ι(α n , β n ) tends to infinity. Passing to subsequences and applying elements of the mapping class group we may assume that α n = α 0 for all n. Setting c n = ι(α 0 , β n ) and passing to subsequences again we may assume that β n /c n converges to λ ∈ PML(S), the projectivization of Thurston's space of measured laminations. Let Y be any connected component of the subsurface filled by λ, chosen so that α 0 cuts Y . Note that π Y (β n ) converges to λ| Y . Again applying Kobayashi's argument [21], the distance d Y (α 0 , β n ) tends to infinity.
Brian Bowditch raised the question, at the Newton Institute in August 2003, of the geometric properties of the inclusion A(S) → AC(S). The natural assumption, that this inclusion is again a quasi-isometric embedding, is false. In this paper we will exactly characterize how the inclusion distorts distance.
We now move up a dimension. Suppose that V is a handlebody and S = ∂V . We may take any disk D ∈ D(V ) to its boundary ∂D ∈ C(S), giving an inclusion ν : D(V ) → C(S). It is important to distinguish the disk complex from its image ν(D(V )); thus we will call the image the disk set.
The first author and Minsky [26] have shown: Theorem 4.9. The disk set is a quasi-convex subset of the curve complex.
It is natural to ask if this map is a quasi-isometric embedding. If so, the hyperbolicity of C(V ) immediately follows. In fact, the inclusion again badly distorts distance and we investigate exactly how, below.

4.3.
Markings and the mapping class group. Once the connectedness of M(S) is in hand, it is possible to use local finiteness to show that M(S) is quasi-isometric to the Cayley graph of the mapping class group [25].
Using subsurface projections the first author and Minsky [25] obtained a distance estimate for the marking complex and thus for the mapping class group. Theorem 4.10. There is a constant C 0 = C 0 (S) so that, for any c ≥ C 0 there is a constant A with independent of the choice of µ and µ . Here the sum ranges over all essential, non-peripheral subsurfaces X ⊂ S.
This, and their similar estimate for the pants graph, is a model for the distance estimates given below. Notice that a filling marking µ ∈ M(S) cuts all essential, non-peripheral subsurfaces of S. It is not an accident that the sum ranges over the same set.

Holes in general and the lower bound on distance
Suppose that S is a compact connected surface. In this paper a combinatorial complex G(S) will have vertices being isotopy classes of certain multicurves in S. We will assume throughout that vertices of G(S) are connected by edges only if there are representatives which are disjoint. This assumption is made only to simplify the proofs -all arguments work in the case where adjacent vertices are allowed to have uniformly bounded intersection. In all cases G will be connected. There is a natural map ν : G → AC(S) taking a vertex of G to the isotopy classes of the components. Examples in the literature include the marking complex [25], the pants complex [9] [2], the Hatcher-Thurston complex [19], the complex of separating curves [7], the arc complex and the curve complexes themselves.
For any combinatorial complex G defined in this paper other than the curve complex we will denote distance in the one-skeleton of G by d G (·, ·). Distance in C(S) will always be denoted by d S (·, ·). 5.1. Holes, defined. Suppose that S is non-simple. Suppose that G(S) is a combinatorial complex. Suppose that X ⊂ S is an cleanly embedded subsurface. A vertex α ∈ G cuts X if some component of α cuts X.
Almost equivalently, if X is a hole then the subsurface projection π X : G(S) → C(X) never takes the empty set as a value. Note that the entire surface S is always a hole, regardless of our choice of G. A boundary parallel annulus cannot be cleanly embedded (unless S is also an annulus), so generally cannot be a hole. A hole X ⊂ S is strict if X is not homeomorphic to S.
We now classify the holes for A(S).
Example 5.2. Suppose that S = S g,b with b > 0 and consider the arc complex A(S). The holes, up to isotopy, are exactly the cleanly embedded surfaces which contain ∂S. So, for example, if S is planar then only S is a hole for A(S). The same holds for S = S 1,1 . In these cases it is an exercise to show that C(S) and A(S) are quasi-isometric. In all other cases the arc complex admits infinitely many holes.
Definition 5.3. If X is a hole and if π X (G) ⊂ C(X) has diameter at least R we say that the hole X has diameter at least R.
Example 5.4. Continuing the example above: Since the mapping class group acts on the arc complex, all non-simple holes for A(S) have infinite diameter.
Suppose now that X, X ⊂ S are disjoint holes for G. In the presence of symmetry there can be a relationship between π X |G and π X |G as follows: Definition 5.5. Suppose that X, X are holes for G, both of infinite diameter. Then X and X are paired if there is a homeomorphism τ : X → X and a constant L 4 so that d X (π X (γ), τ (π X (γ))) ≤ L 4 for every γ ∈ G. Furthermore, if Y ⊂ X is a hole then τ pairs Y with Y = τ (Y ). Lastly, pairing is required to be symmetric; if τ pairs X with X then τ −1 pairs X with X.
Definition 5.6. Two holes X and Y interfere if either X ∩ Y = ∅ or X is paired with X and X ∩ Y = ∅.
Examples arise in the symmetric arc complex and in the discussion of twisted I-bundles inside of a handlebody.

5.2.
Projection to holes is coarsely Lipschitz. The following lemma is used repeatedly throughout the paper: Lemma 5.7. Suppose that G(S) is a combinatorial complex. Suppose that X is a hole for G. Then for any α, β ∈ G we have The additive error is required only when α = β.
Proof. This follows directly from Corollary 4.5 and our assumption that vertices of G connected by an edge represent disjoint multicurves.

Infinite diameter holes.
We may now state a first answer to Bowditch's question.
Lemma 5.8. Suppose that G(S) is a combinatorial complex. Suppose that there is a strict hole X ⊂ S having infinite diameter. Then ν : G → AC(S) is not a quasi-isometric embedding.
This lemma and Example 5.2 completely determines when the inclusion of A(S) into AC(S) is a quasi-isometric embedding. It quickly becomes clear that the set of holes tightly constrains the intrinsic geometry of a combinatorial complex. Lemma 5.9. Suppose that G(S) is a combinatorial complex invariant under the natural action of MCG(S). Then every non-simple hole for G has infinite diameter. Furthermore, if X, Y ⊂ S are disjoint non-simple holes for G then there is a quasi-isometric embedding of Z 2 into G.
We will not use Lemmas 5.8 or 5.9 and so omit the proofs. Instead our interest lies in proving the far more powerful distance estimate (Theorems 5.10 and 13.1) for G(S).

A lower bound on distance.
Here we see that the sum of projection distances in holes gives a lower bound for distance.
Theorem 5.10. Fix S, a compact connected non-simple surface. Suppose that G(S) is a combinatorial complex. Then there is a constant C 0 so that for all c ≥ C 0 there is a constant A satisfying Here α, β ∈ G and the sum is taken over all holes X for the complex G.
The proof follows the proof of Theorems 6.10 and 6.12 of [25], practically word for word. The only changes necessary are to • replace the sum over all subsurfaces by the sum over all holes, • replace Lemma 2.5 of [25], which records how markings differing by an elementary move project to an essential subsurface, by Lemma 5.7 of this paper, which records how G projects to a hole.
One major goal of this paper is to give criteria sufficient obtain the reverse inequality; Theorem 13.1.

Holes for the non-orientable surface
Fix F a compact, connected, and non-orientable surface. Let S be the orientation double cover with covering map ρ F : S → F . Let τ : S → S be the associated involution; so for all x ∈ S, ρ F (x) = ρ F (τ (x)).
The same definitions holds for subsurfaces X ⊂ S. Definition 6.2. The invariant complex C τ (S) is the simplicial complex with vertex set being isotopy classes of invariant multicurves. There is a k-simplex for every collection of k + 1 distinct isotopy classes having pairwise disjoint representatives.
Notice that C τ (S) is simplicially isomorphic to C(F ). There is also a natural map ν : C τ (S) → C(S). We will prove: It thus follows from the hyperbolicity of C(S) that: We begin the proof of Lemma 6.3: since ν sends adjacent vertices to adjacent edges we have as long as α and β are distinct in C τ (S). In fact, since the surface S itself is a hole for C τ (S) we may deduce a slightly weaker lower bound from Lemma 5.7 or indeed from Theorem 5.10. The other half of the proof of Lemma 6.3 consists of showing that S is the only hole for C τ (S) with large diameter. After a discussion of Teichmüller geodesics we will prove: Lemma 16.4. There is a constant K with the following property: Suppose that α, β are invariant multicurves in S. Suppose that X ⊂ S is an essential subsurface where d X (α, β) > K. Then X is symmetric.
From this it follows that: Corollary 6.6. With K as in Lemma 16.4: If X ⊂ S is a hole for C τ (S) with diameter greater than K then X = S.
Proof. Suppose that X ⊂ S is a strict subsurface, cleanly embedded. Suppose that diam X (C τ (S)) > K. Thus X is symmetric. It follows that ∂X ∂S is also symmetric. Since ∂X does not cut X deduce that X is not a hole for C τ (S).
This corollary, together with the upper bound (Theorem 13.1), proves Lemma 6.3.

Holes for the arc complex
Here we generalize the definition of the arc complex and classify its holes.
Definition 7.1. Suppose that S is a non-simple surface with boundary. Let ∆ be a non-empty collection of components of ∂S. The arc complex A(S, ∆) is the subcomplex of A(S) spanned by essential arcs α ⊂ S with ∂α ⊂ ∆.
Note that A(S, ∂S) and A(S) are identical. This follows directly from the definition of a hole. We now have an straight-forward observation: The proof follows immediately from Lemma 7.2. Lemma 5.9 indicates that Lemma 7.3 is essential to proving that A(S, ∆) is Gromov hyperbolic.
In order to prove the upper bound theorem for A we will use pants decompositions of the surface S. In an attempt to avoid complications in the non-orientable case we must carefully lift to the orientation cover.
Suppose that F is non-simple, non-orientable, and has non-empty boundary. Let ρ F : S → F be the orientation double cover and let τ : S → S be the induced involution. Fix ∆ ⊂ ∂F and let ∆ = ρ −1 F (∆ ). Definition 7.4. We define A τ (S, ∆) to be the invariant arc complex: vertices are invariant multi-arcs and simplices arise from disjointness.
Again, A τ (S, ∆) is simplicially isomorphic to A(F, ∆ ). If X ∩τ (X) = ∅ and ∆ ⊂ X ∪ τ (X) then the subsurfaces X and τ (X) are paired holes, as in Definition 5.5. Notice as well that all non-simple symmetric holes X ⊂ S for A τ (S, ∆) have infinite diameter.
Unlike A(F, ∆ ) the complex A τ (S, ∆) may have disjoint holes. Nonetheless, we have: Lemma 7.5. Any two non-simple holes for A τ (S, ∆) interfere.
Proof. Suppose that X, Y are holes for the τ -invariant arc complex, A τ (S, ∆). It follows from Lemma 16.4 that X is symmetric with ∆ ⊂ X ∪ τ (X). The same holds for Y . Thus Y must cut either X or τ (X).

Background on three-manifolds
Before discussing the holes in the disk complex, we record a few facts about handlebodies and I-bundles.
Fix M a compact connected irreducible three-manifold. Recall that M is irreducible if every embedded two-sphere in M bounds a three-ball. Recall that if N is a closed submanifold of M then fr(N ), the frontier of N in M , is the closure of ∂N ∂M .
In this situation form a new surface F as follows: Let N be a closed regular neighborhood of B. First remove from F the annulus N ∩ F . Now form F by gluing on both disk components of ∂N F . We say that F is obtained by compressing F along B. If no such disk exists we say F is incompressible.
• ∂B is a union of connected arcs α and β, A disk, like B, with boundary partitioned into two arcs is called a bigon. Note that this definition of boundary compression is slightly weaker than some found in the literature; the arc α is often required to be essential in F . We do not require this additional property because, for us, F will usually be a properly embedded disk in a handlebody.
Just as for compressing disks we may boundary compress F along B to obtain a new surface F : Let N be a closed regular neighborhood of B. First remove from F the rectangle N ∩ F . Now form F by gluing on both bigon components of fr(N ) F . Again, F is obtained by boundary compressing F along B. Note that the relevant boundary components of F and F cobound a pair of pants embedded in ∂M . If no boundary compression exists then F is boundary incompressible.
Remark 8.2. Recall that any surface F properly embedded in a handlebody V g , g ≥ 2, is either compressible or boundary compressible.
Suppose now that F is properly embedded in M and Γ is a multicurve in ∂M .
It is often useful to restrict our attention to boundary compressions meeting a single subsurface of ∂M . So suppose that X ⊂ ∂M is an essential subsurface. Suppose that ∂F is tight with respect to ∂X. Suppose B is a boundary compression of F . If B ∩ ∂M ⊂ X we say that F is boundary compressible into X.
Suppose that ∂X and ∂F are tight and that X compresses in M . Then either: Proof. Suppose that X is compressible via a disk E. Isotope E to make ∂E tight with respect to ∂F . This can be done while maintaining ∂E ⊂ X because ∂F and ∂X are tight. Since M is irreducible and F is incompressible we may isotope E, rel ∂, to remove all simple closed curves of F ∩ E. If F ∩ E is non-empty then an outermost bigon of E gives the desired boundary compression lying in X. Suppose instead that F ∩ E = ∅ but F does cut X. Let δ ⊂ X be a simple arc meeting each of F and E in exactly one endpoint. Let N be a closed regular neighborhood of δ ∪ E. Note that fr(N ) F has three components. One is a properly embedded disk parallel to E and the other two B, B are bigons attached to F . At least one of these, say B is trivial in the sense that B ∩ ∂M is a trivial arc embedded in ∂M ∂F . If B is non-trivial then B provides the desired boundary compression.
Suppose that B is also trivial. It follows that ∂E and one component

Band sums.
A band sum is the inverse operation to boundary compression: Fix a pair of disjoint properly embedded surfaces F 1 , F 2 ⊂ M . Let F = F 1 ∪ F 2 . Fix a simple arc δ ⊂ ∂M so that δ meets each of F 1 and F 2 in exactly one point of ∂δ. Let N ⊂ M be a closed regular neighborhood of δ. Form a new surface by adding to F N the rectangle component of fr(N ) F . The surface F obtained is the result of band summing F 1 to F 2 along δ. Note that F has a boundary compression dual to δ yielding F : that is, there is a boundary compression B for F so that δ ∩ B is a single point and compressing F along B gives F .
Suppose that F is a compact connected surface with at least one boundary component. Let T be the orientation I-bundle over F . If F is orientable then T ∼ = F × I. If F is not orientable then T is the unique I-bundle over F with orientable total space. We call T the I-bundle and F the base space. Let ρ F : T → F be the associated bundle map. Note that T is homeomorphic to a handlebody.
If A ⊂ T is a union of fibers of the map ρ F then A is vertical with respect to T . In particular take ∂ v T = ρ −1 F (∂F ) to be the vertical boundary of T . Take ∂ h T to be the union of the boundaries of all of the fibers: this is the horizontal boundary Note that, as |∂ v T | ≥ 1, any vertical surface in T can be boundary compressed. However no vertical surface in T may be boundary compressed into ∂ h T .
We end this section with: • ∂D and ∂X are tight, and • D cannot be boundary compressed into X then D may be properly isotoped to be vertical with respect to T .

Holes for the disk complex
Here we begin to classify the holes for the disk complex, a more difficult analysis than that of the arc complex. To fix notation let V be a handlebody. Let S = S g = ∂V . Recall that there is a natural inclusion ν : D(V ) → C(S).
Remark 9.1. The notion of a hole X ⊂ ∂V for D(V ) may be phrased in several different ways: • every essential disk D ⊂ V cuts the surface X, The classification of holes X ⊂ S for D(V ) breaks roughly into three cases: either X is an annulus, is compressible in V , or is incompressible in V . In each case we obtain a result: Theorem 10.1. Suppose X is a hole for D(V ) and X is an annulus. Then the diameter of X is at most 5.
Theorem 11.6. Suppose X is a compressible hole for D(V ) with diameter at least 15. Then there are a pair of essential disks D, E ⊂ V so that • ∂D, ∂E ⊂ X and • ∂D and ∂E fill X.
• F supports a pseudo-Anosov map.
As a corollary of these theorems we have: Proof. If X is a hole with diameter at least 61 then either Theorem 11.6 or Theorem 12.1 applies. If X is compressible then Dehn twists, in opposite directions, about the given disks D and E yields an automorphism f : V → V so that f |X is pseudo-Anosov. This follows from Thurston's construction [38]. By Lemma 2.6 the hole X has infinite diameter.
If X is incompressible then X ⊂ ∂ h T where ρ F : T → F is the given I-bundle. Let f : F → F be the given pseudo-Anosov map. So g, the suspension of f , gives a automorphism of V . Again it follows that the hole X has infinite diameter.
Applying Lemma 5.8 we find another corollary: Theorem 9.3. If S = ∂V contains a strict hole with diameter at least 61 then the inclusion ν : D(V ) → C(S) is not a quasi-isometric embedding.

Holes for the disk complex -annuli
The proof of Theorem 10.1 occupies the rest of this section. This proof shares many features with the proofs of Theorems 11.6 and 12.1. However, the exceptional definition of C(S 0,2 ) prevents a unified approach. Fix V , a handlebody.
Theorem 10.1. Suppose X is a hole for D(V ) and X is an annulus. Then the diameter of X is at most 5.
We begin with: Proof. Since X is a hole, every disk cuts X. Since X is an annulus, let α be a core curve for X. If |D ∩ X| = 1, then we may band sum parallel copies of D along an subarc of α. The resulting disk misses α, a contradiction.
Assume, to obtain a contradiction, that X has diameter at least 6.
Isotope D and E to make the boundaries tight and also tight with respect to ∂X. Tightening triples of curves is not canonical; nonetheless there is a tightening so that S (∂D ∪ ∂E ∪ X) contains no triangles. See Figure 1. After this tightening we have: Claim. Every arc of ∂D ∩ X meets every arc of ∂E ∩ X at least once.
Proof. Fix components arcs α ⊂ D ∩ X and β ⊂ E ∩ X. Let α , β be the corresponding arcs in S X the annular cover of S corresponding to X. After the tightening we find that Claim. There is an outermost bigon B ⊂ E D with the following properties: See the lower right of Figure 2 for a picture.
Proof. Consider the intersection of D and E, thought of as a collection of arcs and curves in E. Any simple closed curve component of D ∩ E can be removed by an isotopy of E, fixed on the boundary. (This follows from the irreducibility of V and an innermost disk argument.) Since we have assumed that |D ∩ E| is minimal it follows that there are no simple closed curves in D ∩ E.
So consider any outermost bigon B ⊂ E D. Let α = B ∩ D. Let β = ∂B α = B ∩ ∂V . Note that β cannot completely contain a component of E ∩ X as this would contradict either the fact that B is outermost or the claim that every arc of E ∩ X meets some arc of D ∩ X. Using this observation, Figure 2 lists the possible ways for B to lie inside of E.  Figure 2. The arc α cuts a bigon B off of E. The darker part of ∂E are the arcs of E ∩X. Either β is disjoint from X, β is contained in X, β meets X in a single subarc, or β meets X in two subarcs.
Let D and D be the two essential disks obtained by boundary compressing D along the bigon B. Suppose α is as shown in one of the first three pictures of Figure 2. It follows that either D or D has, after tightening, smaller intersection with X than D does, a contradiction. We deduce that α is as pictured in lower right of Figure 2.
Boundary compressing D along B still gives disks D , D ∈ D(V ). As these cannot have smaller intersection with X we deduce that |D ∩ X| ≤ 2 and the claim holds.
Using the same notation as in the proof above, let B be an outermost bigon of E D. We now study how α ⊂ ∂B lies inside of D.
Claim. The arc α ⊂ D connects distinct components of D ∩ X.
Proof. Suppose not. Then there is a bigon C ⊂ D α with ∂C = α ∪ γ and γ ⊂ ∂D ∩ X. The disk C ∪ B is essential and intersects X at most once after tightening, contradicting our first claim.
We finish the proof of Theorem 10.1 by noting that D ∪ B is homeomorphic to Υ×I where Υ is the simplicial tree with three edges and three leaves. We may choose the homeomorphism so that (D∪B)∩X = Υ×∂I. It follows that we may properly isotope D ∪ B until (D ∪ B) ∩ X is a pair of arcs. Recall that D and D are the disks obtained by boundary compressing D along B. It follows that one of D or D (or both) meets X in at most a single arc, contradicting our first claim.

Holes for the disk complex -compressible
The proof of Theorem 11.6 occupies the second half of this section. 11.1. Compression sequences of essential disks. Fix a multicurve Γ ⊂ S = ∂V . Fix also an essential disk D ⊂ V . Properly isotope D to make ∂D tight with respect to Γ.
If D ∩ Γ = ∅ we may define: starting at D has ∆ 1 = {D} and ∆ k+1 is obtained from ∆ k via a boundary compression, disjoint from Γ, and tightening. Note that ∆ k is a collection of exactly k pairwise disjoint disks properly embedded in V . We further require, for k ≤ n, that every disk of ∆ k meets some component of Γ. We call a compression sequence maximal if either • no disk of ∆ n can be boundary compressed into S Γ or • there is a component Z ⊂ S Γ and a boundary compression of ∆ n into S Γ yielding an essential disk E with ∂E ⊂ Z. We say that such maximal sequences end essentially or end in Z, respectively.
All compression sequences must end, by Remark 8.3. Given a maximal sequence we may relate the various disks in the sequence as follows: If α = α then the two disjointness pairs (α, β) and (α , β) are distinct, even if α is properly isotopic to α . A similar remark holds for the second coordinate.
The following lemma controls how subsurface projection distance changes in maximal sequences.
where each of the arcs {α i } appears as the first coordinate of at most two pairs.
Proof. We induct on n − k. If D k is contained in ∆ n there is nothing to prove. If D k is contained in ∆ k+1 we are done by induction. Thus we may assume that D k is the disk of ∆ k which is boundary compressed at stage k. Let D k+1 , D k+1 ∈ ∆ k+1 be the two disks obtained after boundary compressing D k along the bigon B. See Figure 3 for a picture of the pair of pants cobounded by ∂D k and ∂D k+1 ∪ ∂D k+1 . Figure 3. All arcs connecting D k to itself or to D k+1 ∪ D k+1 are arcs of Γ ∩ P . The boundary compressing arc B ∪ S meets D k twice and is parallel to the vertical arcs of Γ ∩ P .
Let δ be a band sum arc dual to B (the dotted arc in Figure 3). We may assume that |Γ∩δ| is minimal over all arcs dual to B. It follows that the band sum of D k+1 with D k+1 along δ is tight, without any isotopy.
(This is where we use the fact that B is a boundary compression in the complement of Γ, as opposed to being a general boundary compression of D k in V .) There are now three possibilities: neither, one, or both points of ∂δ are contained in X.
First suppose that X ∩ ∂δ = ∅. Then every arc of D k+1 ∩ X is parallel to an arc of D k ∩ X, and similarly for D k+1 . If D k+1 and D k+1 are both components of ∆ n then choose any arcs β, β of D k+1 ∩ X and of D k+1 ∩ X. Let α, α be the parallel components of D k ∩ X. The four disjointness pairs are then (α, β), (α, β ), (α , β), (α , β ). Suppose instead that D k+1 is not a component of ∆ n . Then D k inherits four disjointness pairs from D k+1 .
Second suppose that exactly one endpoint x ∈ ∂δ meets X. Let γ ⊂ D k+1 be the component of D k+1 ∩ X containing x. Let X be the component of X ∩P that contains x and let α, α be the two components of D k ∩ X . Let β be any arc of D k+1 ∩ X.
If D k+1 / ∈ ∆ n and γ is not the first coordinate of one of D k+1 's four pairs then D k inherits disjointness pairs from D k+1 . If D k+1 / ∈ ∆ n then D k inherits disjointness pairs from D k+1 .
Thus we may assume that both D k+1 and D k+1 are in ∆ n or that only D k+1 ∈ ∆ n while γ appears as the first arc of disjointness pair for D k+1 . In case of the former the required disjointness pairs are (α, β), (α , β), (α, γ), and (α , γ). In case of the latter we do not know if γ is allowed to appear as the second coordinate of a pair. However we are given four disjointness pairs for D k+1 and are told that γ appears as the first coordinate of at most two of these pairs. Hence the other two pairs are inherited by D k . The pairs (α, β) and (α , β) give the desired conclusion.
Third suppose that the endpoints of δ meet γ ⊂ D k+1 and γ ⊂ D k+1 . Let X be a component of X ∩ P containing γ. Let α and α be the two arcs of D k ∩ X . Suppose both D k+1 and D k+1 lie in ∆ n . Then the desired pairs are (α, γ), (α , γ), (α, γ ), and (α , γ ). If D k+1 ∈ ∆ n while D k+1 is not then D k inherits two pairs from D k+1 . We add to these the pairs (α, γ ), and (α , γ ). If neither disk lies in ∆ n then D k inherits two pairs from each disk and the proof is complete.
Given a disk D ∈ D(V ) and a hole X ⊂ S our Lemma 11.3 allows us to adapt D to X.
there is a disk D with the following properties: • ∂X and ∂D are tight.
• If X is incompressible then D is not boundary compressible into Here A = A(X) and AC = AC(X).
Proof. If ∂D ⊂ X then the lemma is trivial. So assume, by Remark 9.1, that D cuts ∂X. Choose a maximal sequence with respect to ∂X starting at D.
Suppose that the sequence is non-trivial (n > 1). By Lemma 11.3 there is a disk E ∈ ∆ n so that D ∩ X and E ∩ X contain disjoint arcs.
If the sequence ends essentially then choose D = E and the lemma is proved. If the sequence ends in X then there is a boundary compression of ∆ n , disjoint from ∂X, yielding the desired disk D with ∂D ⊂ X. Since E ∩ D = ∅ we again obtain the desired bound.
Assume now that the sequence is trivial (n = 1). Then take E = D ∈ ∆ n and the proof is identical to that of the previous paragraph.
Remark 11.5. Lemma 11.4 is unexpected: after all, any pair of curves in C(X) can be connected by a sequence of band sums. Thus arbitrary band sums can change the subsurface projection to X. However, the sequences of band sums arising in Lemma 11.4 are very special. Firstly they do not cross ∂X and secondly they are "tree-like" due to the fact every arc in D is separating.
When D is replaced by a surface with genus then Lemma 11.4 does not hold in general; this is a fundamental observation due to Kobayashi [21] (see also [17]). Namazi points out that even if D is only replaced by a planar surface Lemma 11.4 does not hold in general.
11.2. Proving the theorem. We now prove: Theorem 11.6. Suppose X is a compressible hole for D(V ) with diameter at least 15. Then there are a pair of essential disks D, E ⊂ V so that • ∂D, ∂E ⊂ X and • ∂D and ∂E fill X.
Proof. Choose disks D and E in D(V ) so that d X (D , E ) ≥ 15. By Lemma 11.4 there are disks D and E so that ∂D, ∂E ⊂ X, d X (D , D) ≤ 6, and d X (E , E) ≤ 6. It follows from the triangle inequality that d X (D, E) ≥ 3.

Holes for the disk complex -incompressible
This section classifies incompressible holes for the disk complex.
Here is a short plan of the proof: We are given X, an incompressible hole for D(V ). Following Lemma 11.4 we may assume that D, E are essential disks, without boundary compressions into X or S X, with d X (D, E) > 43. Examine the intersection pattern of D and E to find two families of rectangles R and Q. The intersection pattern of these rectangles in V will determine the desired I-bundle T . The third conclusion of the theorem follows from standard facts about primitive annuli. The fourth requires another application of Lemma 11.4 as well as Lemma 2.6.

Diagonals of polygons.
To understand the intersection pattern of D and E we discuss diagonals of polygons. Let D be a 2n sided regular polygon. Label the sides of D with the letters X and Y in alternating fashion. Any side labeled X (or Y ) will be called an X side (or Y side). Proof. A counting argument shows that |Γ| ≤ 4n − 3. If every X side meets at least nine non-parallel diagonals then |Γ| ≥ 9 2 n > 4n − 3, a contradiction.
12.2. Improving disks. Suppose now that X is an incompressible hole for D(V ) with diameter at least 61. Note that, by Theorem 10.1, X is not an annulus. Let Y = S X.
Choose disks D and E in V so that d X (D , E ) ≥ 61. By Lemma 11.4 there are a pair of disks D and E so that both are essential in V , cannot be boundary compressed into X or into Y , and so that d A(X) (D , D) ≤ 3 and d A(X) (E , E) ≤ 3. Thus d X (D , D) ≤ 9 and d X (E , E) ≤ 9 (Lemma 5.7). By the triangle inequality d X (D, E) ≥ 61 − 18 = 43.
Recall, as well, that ∂D and ∂E are tight with respect to ∂X. We may further assume that ∂D and ∂E are tight with respect to each other. Also, minimize the quantities |X ∩(∂D ∩∂E)| and |D ∩E| while keeping everything tight. In particular, there are no triangle components of ∂V (D ∪ E ∪ ∂X). Now consider D and E to be even-sided polygons, with vertices being the points ∂D ∩ ∂X and ∂E ∩ ∂X respectively. Let Γ = D ∩ E. See Figure 4 for one a priori possible collection Γ ⊂ D. From our assumptions and the irreducibility of V it follows that Γ contains no simple closed curves. Suppose now that there is a γ ⊂ Γ so that, in D, both endpoints of γ lie in the same side of D. Then there is an outermost such arc, say γ ⊂ Γ, cutting a bigon B out of D. It follows that B is a boundary compression of E which is disjoint from ∂X. But this contradicts the construction of E. We deduce that all arcs of Γ are diagonals for D and, via a similar argument, for E.
Let α ⊂ D ∩ X be an X side of D meeting at most eight distinct types of diagonal of Γ. Choose β ⊂ E ∩ X similarly. As d X (D, E) ≥ 43 we have that d X (α, β) ≥ 43 − 6 = 37. Now break each of α and β into at most eight subarcs {α i } and {β j } so that each subarc meets all of the diagonals of fixed type and only of that type. Let R i ⊂ D be the rectangle with upper boundary α i and containing all of the diagonals meeting α i . Let α i be the lower boundary of R i . Define Q j and β j similarly. See Figure 5 for a picture of R i .
Call an arc α i large if there is an arc β j so that |α i ∩ β j | ≥ 3. We use the same notation for β j . Let Θ be the union of all of the large α i and β j . Thus Θ is a four-valent graph in X. Let Θ be the union of the corresponding large α i and β i .  Proof. If Θ = ∅, then all α i are small. It follows that |α ∩ β| ≤ 128 and thus d X (α, β) ≤ 16, by Lemma 2.2. As d X (α, β) ≥ 37 this is a contradiction.
Let Z ⊂ ∂V be a small regular neighborhood of Θ and define Z similarly. Proof. For a contradiction suppose that W is a component of Z contained in a disk. Then there is some pair α i , β j having a bigon in ∂V . This contradicts the tightness of ∂D and ∂E. The same holds for Z .
Suppose now that some component W is contained in an annulus A, peripheral in X. Thus W fills A. Suppose that α i and β j are large and contained in W . By the classification of arcs in A we deduce that either α i and β j form a bigon in A or ∂X, α i and β j form a triangle. Either conclusion gives a contradiction.
Proof. Suppose not. Fix attention on any component W ⊂ Z. Since Θ does not fill, the previous claim implies that there is a component γ ⊂ ∂W that is essential and non-peripheral in X. Note that any large α i meets ∂W in at most two points, while any small α i meets ∂W in at most 32 points. Thus |α ∩ ∂W | ≤ 256 and the same holds for β. Thus d X (α, β) ≤ 36 by the triangle inequality. As d X (α, β) ≥ 37 this is a contradiction.
The previous two claims imply: Claim 12.7. The graph Θ is connected.
There are now two possibilities: either Θ ∩ Θ is empty or not. In the first case set Σ = Θ and in the second set Σ = Θ ∪ Θ . By the claims above, Σ is connected and fills X. Let R = {R i } and Q = {Q j } be the collections of large rectangles.
12.3. Building the I-bundle. We are given Σ, R and Q as above. Note that R∪Q is an I-bundle and Σ is the component of its horizontal boundary meeting X. See Figure 6 for a simple case. Let T 0 be a regular neighborhood of R ∪ Q, taken in V . Again T 0 has the structure of an I-bundle. Note that ∂ h T 0 ⊂ ∂V , ∂ h T 0 ∩ X is a component of ∂ h T 0 , and this component fills X due to Claim 12.6. We will enlarge T 0 to obtain the correct I-bundle in V .
Begin by enumerating all annuli {A i } ⊂ ∂ v T 0 with the property that some component of ∂A i is inessential in ∂V . Suppose that we have built the I-bundle T i and are now considering the annulus A = A i . Let γ ∪ γ = ∂A ⊂ ∂V with γ inessential in ∂V . Let B ⊂ ∂V be the disk which γ bounds. By induction we assume that no component of ∂ h T i is contained in a disk embedded in ∂V (the base case holds by Claim 12.5). It follows that B ∩ T i = ∂B = γ. Thus B ∪ A is isotopic, rel γ , to be a properly embedded disk B ⊂ V . As γ lies in X or Y , both incompressible, γ must bound a disk C ⊂ ∂V . Note that C ∩ T i = ∂C = γ , again using the induction hypothesis.
It follows that B ∪ A ∪ C is an embedded two-sphere in V . As V is a handlebody V is irreducible.
, and A is identified with ∂B × I. We form T i+1 = T i ∪ U i and note that T i+1 still has the structure of an I-bundle.
After dealing with all of the annuli {A i } in this fashion we are left with an I-bundle T . Now all components of ∂∂ v T [sic] are essential in ∂V . All of these lying in X are peripheral in X. This is because they are disjoint from Σ ⊂ ∂ h T , which fills X, by induction. It follows that the component of ∂ h T containing Σ is isotopic to X.
This finishes the construction of the promised I-bundle T and demonstrates the first two conclusions of Theorem 12.1. For future use we record:

A vertical annulus parallel into the boundary.
Here we obtain the third conclusion of Theorem 12.1: at least one component of ∂ v T is boundary parallel in ∂V .
Fix T an I-bundle with the incompressible hole X a component of Proof. Suppose that A ⊂ ∂ v T was compressible. By Remark 12.8 we may compress A to obtain a pair of essential disks B and C. Note that ∂B is isotopic into the complement of ∂ h T . So S X is compressible, contradicting Remark 9.1. Proof. Since ∂ v T is incompressible (Claim 12.9) by Remark 8.2, we find that ∂ v T is boundary compressible in V . Let B be a boundary compression for The arc α is either essential or inessential in A. Suppose α is inessential in A. Then α cuts a bigon, C, out of A. Since B was a boundary compression the disk D = B ∪ C is essential in V . Since B meets ∂ v T in a single arc, either D ⊂ T or D ⊂ V T . The former implies that ∂ h T is compressible and the latter that X is not a hole. Either gives a contradiction.
It follows that α is essential in A. Now carefully boundary compress A: Let N be the closure of a regular neighborhood of B, taken in V A. It follows that D is inessential in V . Thus D cuts a closed three-ball U out of V . There are two final cases: either N ⊂ U or N ∩U = B ∪B . If U contains N then U contains A. Thus ∂A is contained in the disk U ∩∂V . This contradicts Remark 12.8. Deduce instead that W = U ∪N is a solid torus with meridional disk B. Thus W gives a parallelism between A and the annulus ∂V ∩ ∂W , as desired.
Remark 12.11. Similar considerations prove that the multicurve 12.5. Finding a pseudo-Anosov map. Here we prove that the base surface F of the I-bundle T admits a pseudo-Anosov map.
As in Section 12.2, pick essential disks D and E in V so that d X (D , E ) ≥ 61. Lemma 11.4 provides disks D and E which cannot be boundary compressed into X or into S X -thus D and E cannot be boundary compressed into ∂ h T . Also, as above, d X (D, E) ≥ 61 − 18 = 43.
After isotoping D to minimize intersection with ∂ v T it must be the case that all components of D ∩ ∂ v T are essential arcs in ∂ v T . By Lemma 8.5 we conclude that D may be isotoped in V so that D ∩ T is vertical in T . The same holds of E. Choose A and B, components of D∩T and E ∩T . Each are vertical rectangles. Since diam X (π X (D)) ≤ 3 (Lemma 4.4) we now have d X (A, B) ≥ 43 − 6 = 37.
We now begin to work in the base surface F . Recall that ρ F : T → F is an I-bundle. Take α = ρ F (A) and β = ρ F (B). Note that the natural map C(F ) → C(X), defined by taking a curve to its lift, is distance non-increasing (see Equation 6.5). Thus d F (α, β) ≥ 37. By Theorem 10.1 the surface F cannot be an annulus. Thus, by Lemma 2.6 the subsurface F supports a pseudo-Anosov map and we are done.
12.6. Corollaries. We now deal with the possibility of disjoint holes for the disk complex.
Lemma 12.12. Suppose that X is a large incompressible hole for D(V ) supported by the I-bundle ρ F : T → F . Let Y = ∂ h T X. Let τ : ∂ h T → ∂ h T be the involution switching the ends of the I-fibres. Suppose that D ∈ D(V ) is an essential disk. • Proof. By Lemma 11.4 there is a disk D ⊂ V which is tight with respect to ∂ h T and which cannot be boundary compressed into ∂ h T (or into the complement). Also, for any component Proof. Suppose that X ∩ Z = ∅. It follows from Remark 9.1 that both X and Z are incompressible. Let ρ F : T → F be the I-bundle in V with X ⊂ ∂ h T , as provided by Theorem 12.1. We also have a component A ⊂ ∂ v T so that A is boundary parallel. Let U be the solid torus component of V A. Note that Z cannot be contained in ∂U A because Z is not an annulus (Theorem 10.1).
Let α = ρ F (A). Choose any essential arc δ ⊂ F with both endpoints in α ⊂ ∂F . It follows that ρ −1 F (δ), together with two meridional disks of U , forms an essential disk D in V . Let W = ∂ h T ∪ (U A) and note that ∂D ⊂ W .
If F is non-orientable then Z ∩ W = ∅ and we have a contradiction. Deduce that F is orientable. Now, if Z misses Y then Z misses W and we again have a contradiction. It follows that Z cuts Y and we are done.

Axioms for combinatorial complexes
The goal of this section and the next is to prove, inductively, an upper bound on distance in a combinatorial complex G(S) = G. This section presents our axioms on G: sufficient hypotheses for Theorem 13.1. The axioms, apart from Axiom 13.2, are quite general. Axiom 13.2 is necessary to prove hyperbolicity and greatly simplifies the recursive construction in Section 14.
Theorem 13.1. Fix S a compact connected non-simple surface. Suppose that G = G(S) is a combinatorial complex satisfying the axioms of Section 13. Let X be a hole for G and suppose that α X , β X ∈ G are contained in X. For any constant c > 0 there is a constant A satisfying: where the sum is taken over all holes Y ⊆ X for G.
The proof of the upper bound is more difficult than that of the lower bound, Theorem 5.10. This is because naturally occurring paths in G between α X and β X may waste time in non-holes. The first example of this is the path in C(S) obtained by taking the short curves along a Teichmüller geodesic. The Teichmüller geodesic may spend time rearranging the geometry of a subsurface. Then the systole path in the curve complex must be much longer than the curve complex distance between the endpoints.
In Sections 16, 17, 19 we will verify these axioms for the curve complex of a non-orientable surface, the arc complex, and the disk complex.
13.1. The axioms. Suppose that G = G(S) is a combinatorial complex. We begin with the axiom required for hyperbolicity.
Axiom 13.2 (Holes interfere). All large holes for G interfere, as given in Definition 5.6.
Fix vertices α X , β X ∈ G, both contained in a hole X. We are given Λ = {µ n } N n=0 , a path of markings in X. Axiom 13.3 (Marking path). We require: (1) The support of µ n+1 is contained inside the support of µ n .
(2) For any subsurface Y ⊆ X, if π Y (µ k ) = ∅ then for all n ≤ k the map n → π Y (µ n ) is an unparameterized quasi-geodesic with constants depending only on G.
The second condition is crucial and often technically difficult to obtain. We are given, for every essential subsurface Y ⊂ X, a perhaps empty interval J Y ⊂ [0, N ] with the following properties.
Axiom 13.4 (Accessibility). The interval for X is J X = [0, N ]. There is a constant B 3 so that There is a combinatorial path Γ = {γ i } K i=0 ⊂ G starting with α X ending with β X and each γ i is contained in X. There is a strictly increasing reindexing function r : [0, K] → [0, N ] with r(0) = 0 and r(K) = N .
Axiom 13.5 (Combinatorial). There is a constant C 2 so that: • d Y (γ i , µ r(i) ) < C 2 , for every i ∈ [0, K] and every hole Y ⊂ X, Axiom 13.6 (Replacement). There is a constant C 4 so that: (1) If Y ⊂ X is a hole and r(i) ∈ J Y then there is a vertex γ ∈ G so that γ is contained in Y and d G (γ i , γ ) < C 4 . (2) If Z ⊂ X is a non-hole and r(i) ∈ J Z then there is a vertex γ ∈ G so that d G (γ i , γ ) < C 4 and so that γ is contained in Z or in X Z.
There is one axiom left: the axiom for straight intervals. This is given in the next subsection.
13.2. Inductive, electric, shortcut and straight intervals. We describe subintervals that arise in the partitioning of [0, K]. As discussed carefully in Section 13.3, we will choose a lower threshold L 1 (Y ) for every essential Y ⊂ X and a general upper threshold, L 2 .
When X is the only relevant hole we have a simpler definition: Electric intervals will be further partitioned into shortcut and straight intervals. , µ r(q) ) < L 2 for all non-holes Y .
Our final axiom is: Axiom 13.11 (Straight). There is a constant A depending only on X and G so that for every straight interval [p, q]: 13.3. Deductions from the axioms. Axiom 13.3 and Lemma 3.9 imply that the reverse triangle inequality holds for projections of marking paths.
Lemma 13.12. There is a constant C 1 so that for every essential Y ⊂ X and for every m < n < p in [0, N ].
We record three simple consequences of Axiom 13.4.
Lemma 13.13. There is a constant C 3 , depending only on B 3 , with the follow properties: Proof. We first prove conclusion (i): Since Y is strictly nested in Z and since Y is contained in the support of µ m (part (1) of Axiom 13.4), both ∂Y and µ m cut Z. By Axiom 13.4, part (2), we have that ι(∂Y, µ m ) ≤ B 3 . It follows that ι(∂Y, π Z (µ m )) ≤ 2B 3 . By Lemma 2.2 we deduce that d Z (∂Y, µ m ) ≤ 2 log 2 B 3 + 3. We take C 3 larger than this right hand side. Conclusion (ii) follows from a pair of applications of conclusion (i) and the triangle inequality.
For conclusion (iii): As in (ii), to bound d Z (µ m , µ n ) it suffices to note that ∂Y cuts Z and that ∂Y has bounded intersection with both of µ m , µ n .
We now have all of the constants C 1 , C 3 , C 2 , C 4 in hand. Recall that L 4 is the pairing constant of Definition 5.5 and that M 0 is the constant of 4.6. We must choose a lower threshold L 1 (Y ) for every essential Y ⊂ X. We must also choose the general upper threshold, L 2 and general lower threshold L 0 . We require, for all essential Z, Y in X, with ξ(Z) < ξ(Y ) ≤ ξ(X): L 0 > C 3 + 2C 2 + 2L 4 (13.14)

Partition and the upper bound on distance
In this section we prove Theorem 13.1 by induction on ξ(X). The first stage of the proof is to describe the inductive partition: we partition the given interval [0, K] into inductive and electric intervals. The inductive partition is closely linked with the hierarchy machine [25] and with the notion of antichains introduced in [34].
We next give the electric partition: each electric interval is divided into straight and shortcut intervals. Note that the electric partition also gives the base case of the induction. We finally bound d G (α X , β X ) from above by combining the contributions from the various intervals.
14.1. Inductive partition. We begin by identifying the relevant surfaces for the construction of the partition. We are given a hole X for G and vertices α X , β X ∈ G contained in X. Define We now partition [0, K] into inductive and electric intervals. Begin with the partition of one part P X = {[0, K]}. Recursively P X is a partition of [0, K] consisting of intervals which are either inductive, electric, or undetermined. Suppose that [i, j] ∈ P X is undetermined.
In the former case, as L 1 (X) < L 2 , we are done. So suppose the latter holds. Now, by the reverse triangle inequality (Lemma 13.12), Since r(0) = 0 and r(K) = N we find: Deduce that This completes the proof.
Thus if B X (i, j) is empty then [i, j] ∈ P X is determined to be electric. Proceed on to the next undetermined element. Suppose instead that B X (i, j) is non-empty. Pick a hole Y ∈ B X (i, j) so that Y has maximal ξ(Y ) amongst the elements of B X (i, j) Let p, q ∈ [i, j] be the first and last indices, respectively, so that r(p), r(q) ∈ J Y . (If Y is paired with Y then we take the first and last indices that, after reindexing, lie inside of J Y ∩ J Y .) Claim. The indices p, q are well-defined.
Proof. By assumption d Y (γ i , γ j ) ≥ L 1 (X). By Equation 13.14, We deduce from Axiom 13.4 and Axiom 13.5 that J Y ∩ r([i, j]) is non-empty. Thus, if Y is not paired, the indices p, q are well-defined.
Suppose instead that Y is paired with Y . Recall that measurements made in Y and Y differ by at most the pairing constant L 4 given in Definition 5.5. By (13.16), We deduce again from Axiom 13.4 that J Y ∩ r([i, j]) is non-empty.
Suppose now, for a contradiction, that Without loss of generality we may assume that h < k. It follows that the last inequality by (13.16). This is a contradiction to the assumption.

Claim. The interval [p, q] is inductive for Y .
Proof. We must check that d Y (γ p , γ q ) ≥ L 1 (Y ). Suppose first that Y is not paired. Then by the definition of p, q, (2) of Axiom 13.4, and the triangle inequality we have d Y (µ r(i) , µ r(j) ) ≤ d Y (µ r(p) , µ r(q) ) + 2C 3 .
Thus, when B X (i, j) is non-empty we may find a hole Y and indices p, q as above. In this situation, we subdivide the element [i, j] ∈ P X into the elements [i, p − 1], [p, q], and [q + 1, j]. (The first or third intervals, or both, may be empty.) The interval [p, q] ∈ P X is determined to be inductive and associated to Y . Proceed on to the next undetermined element. This completes the construction of P X . As a bit of notation, if [i, j] ∈ P X is associated to Y ⊂ X we will sometimes write I Y = [i, j].

Properties of the inductive partition.
Lemma 14.1. Suppose that Y, Z are holes and I Z is an inductive element of P X associated to Z. Suppose that r( • Z is nested in Y or • Z and Z are paired and Z is nested in Y . Proof. Let I Z = [i, j]. Suppose first that Y is strictly nested in Z. Then by (ii) of Lemma 13.13, d Z (µ r(i) , µ r(j) ) < C 3 . Then by Axiom 13.5 a contradiction. We reach the same contradiction if Y and Z overlap using (iii) of Lemma 13.13. Now, if Z and Y are disjoint then there are two cases: Suppose first that Y is paired with Y . Since all holes interfere, Y and Z must meet. In this case we are done, just as in the previous paragraph.
Suppose now that Z is paired with Z . Since all holes interfere, Z and Y must meet. If Z is nested in Y then we are done. If Y is strictly nested in Z then, as r([i, j]) ⊂ J Y , we find that as above by Axioms 13.5 and (ii) of Lemma 13.13 that d Z (γ i , γ j ) < C 3 + 2C 2 and so d Z (γ i , γ j ) < C 3 + 2C 2 + 2L 4 < L 1 (Z), a contradiction. We reach the same contradiction if Y and Z overlap.
X is a hole for G.
(1) If Y is associated to an inductive interval I Y ∈ P X and Y is paired with Y then Y is not associated to any inductive interval in P X . (2) There is at most one inductive interval I Y ∈ P X associated to Y .
The second statement is essentially similar. Finally suppose that Z and W are the first and last holes, if any, satisfying the hypotheses of the third claim. Since d Y (γ h , γ i ) ≥ L 0 we find by Axiom 13.5 that By (13.14), L 0 − 2C 2 > C 3 so that If Y is paired then, again by (13.14) we have L 0 > C 3 + 2C 2 + 2L 4 , we also find that J Y ∩ r(I Z ) = ∅. Symmetrically, J Y ∩ r(I W ) (and J Y ∩ r(I W )) are also non-empty.
It follows that the interval between I Z and I W , after reindexing, is contained in J Y (and J Y , if Y is paired). Thus for any inductive interval Thus there are no inductive intervals between I Z and I W satisfying the hypotheses of the third claim.
The following lemma and proposition bound the number of inductive intervals. The discussion here is very similar to (and in fact inspired) the antichains defined in [34,Section 5]. Our situation is complicated by the presence of non-holes and interfering holes. Lemma 14.4. Suppose that X, α X , β X are given, as above. For any Proof. Let g X be a geodesic in C(X) joining α X , β X . By the Bounded Geodesic Image Theorem (Theorem 4.6), since Thus assume there do not exist at least L 2 distinct ω i . Then there is some fixed ω among these ω i such that at least Thus one component, call it W , of X ω contains at least (3 · L 2 ) ξ(X)−1 of the Y i . Let g W be a geodesic in C(W ) joining α W = π W (α X ) and β W = π W (β X ). Notice that because we are projecting to nested subsurfaces. This follows for example from Lemma 4.4. Hence d Y i (α W , β W ) ≥ L 1 (W ). Again apply Theorem 4.6. Since L 1 (W ) > M 0 , for every remaining If there are at least L 2 distinct η i then we take Z = W . Otherwise we repeat the argument. Since the complexity of each successive subsurface is decreasing by at least 1, we must eventually find the desired Z containing at least L 2 of the Y i , each disjoint from distinct vertices of g Z .
So suppose that there are at least L 2 distinct η i associated to distinct Y i and we have taken Z = W . Now we must find at least L 2 − 4(C 1 + To this end we focus attention on a small subset {Y j } 5 j=1 ⊂ {Y i }. Let η j be the vertex of g Z = g W associated to Y j . We choose these Y j so that • the η j are arranged along g Z in order of index and • d Z (η j , η j+1 ) > C 1 + C 3 + 2C 3 + 2, for j = 1, 2, 3, 4. This is possible by (13.15) because Set J j = J Y j and pick any indices m j ∈ J j . (If Z is paired then Y j is as well and we pick m j ∈ J Y j ∩ J (Y j ) .) We use µ(m j ) to denote µ m j . Since ∂Y j is disjoint from η j , Axiom 13.4 and Lemma 2.2 imply Since the sequence π Z (µ n ) satisfies the reverse triangle inequality (Lemma 13.12), it follows that the m j appear in [0, N ] in order agreeing with their index. The triangle inequality implies that possible Y i 's which satisfy the hypothesis on Y 3 . This completes the proof.

Define
P ind = {I ∈ P X | I is inductive}.
Proposition 14.6. The number of inductive intervals is a lower bound for the projection distance in X: Proof. Suppose, for a contradiction, that the conclusion fails. Let g X be a geodesic in C(X) connecting α X to β X . Then, as in the proof of Lemma 14.4, there is a vertex ω of g X and a component W ⊂ X ω where at least (3 · L 2 ) ξ(X)−1 of the inductive intervals in I X have associated surfaces, Y i , contained in W .
Since ξ(X) − 1 ≥ ξ(W ) we may apply Lemma 14.4 inside of W . So we find a surface Z ⊆ W X so that Let Y 1 ∈ Y and η 1 ∈ g Z satisfy ∂Y 1 ∩ η 1 = ∅ and η 1 is the first such. Choose Y 2 and η 2 similarly, so that η 2 is the last such. By Lemma 14.4 (1) of Axiom 13.4, and Lemma 2.2, d Z (µ r(p) , ∂Y 1 ) < C 3 . It follows that d Z (µ r(p) , η 1 ) ≤ C 3 + 1 and the same bound applies to d Z (µ r(q) , η 2 ). Combined with (14.7) we find that By the reverse triangle inequality (Lemma 13.12), for any p ≤ p, q ≤ q , Finally by Axiom 13.5 and the above inequality we have By (13.15) the right-hand side is greater than L 1 (X) + 2L 4 so we deduce that Z ∈ B X (p , q ), for any such p , q . (When Z is paired deduce also that Z ∈ B X (p , q ).) Let I V be the first inductive interval chosen by the procedure with the property that I V ∩ [p, q] = ∅. Note that, since I Y 1 and I Y 2 will also be chosen, I V ⊂ [p, q]. Let p , q be the indices so that V is chosen from B X (p , q ). Thus p ≤ p and q ≤ q . However, since I V ⊂ [p, q] ⊂ J Z , Lemma 14.1 implies that V is strictly nested in Z. (When pairing occurs we may find instead that V ⊂ Z or V ⊂ Z.) Thus ξ(Z) > ξ(V ) and we find that Z would be chosen from B X (p , q ), instead of V . This is a contradiction.
14.3. Electric partition. The goal of this subsection is to prove: We begin by building a partition of [i, j] into straight and shortcut intervals. Define We also define, for all Our recursion starts with the partition of one part, Proof. We show the contrapositive. Suppose that Y is a non-hole with d Y (µ r(p) , µ r(q) ) ≥ L 2 . Since L 2 > C 3 , Axiom 13.4 implies that J Y ∩ [r(p), r(q)] is non-empty. Also, the reverse triangle inequality (Lemma 13.12) gives: d Y (µ r(p) , µ r(q) ) < d Y (µ r(i) , µ r(j) ) + 2C 1 .
Since L 2 > L 1 (X) + 2C 1 , we find that Y ∈ C X . It follows that Y ∈ C X (p, q).
So when C X (p, q) is empty the interval [p, q] is determined to be straight. Proceed onto the next undetermined element of P(i, j). Now suppose that C X (p, q) is non-empty. Then we choose any Y ∈ C X (p, q) so that Y has maximal ξ(Y ) amongst the elements of C X (p, q). Notice that by the accessibility requirement that J Y ∩ [r(p), r(q)] is non-empty.
There are two cases. If J Y ∩ r([p, q]) is empty then let p ∈ [p, q] be the largest integer so that r(p ) < min J Y . Note that p is well-defined. Now divide the interval [p, q] into the two undetermined intervals [p, p ], [p + 1, q]. In this situation we say Y is associated to a shortcut of length one and we add the element [p + 1 2 ] to P(i, j). Next suppose that J Y ∩ r([p, q]) is non-empty. Let p , q ∈ [p, q] be the first and last indices, respectively, so that r(p ), r(q ) ∈ J Y . (Note that it is possible to have p = q .) Partition [p, q] = [p, p −1]∪[p , q ]∪[q +1, q].
The first and third parts are undetermined; either may be empty. This completes the recursive construction of the partition.
Define P short = {I ∈ P(i, j) | I is a shortcut} and P str = {I ∈ P(i, j) | I is straight}.
Proposition 14.9. With P(i, j) as defined above, Proof. The proof is identical to that of Proposition 14.6 with the caveat that in Lemma 14.4 we must use the markings µ r(i) and µ r(j) instead of the endpoints γ i and γ j . Now we "electrify" every shortcut interval using Theorem 13.1 recursively.
Proof. As [p, q] is a shortcut we are given a non-hole Z ⊂ X so that r([p, q]) ⊂ J Z . Let Y = X Z. Thus Axiom 13.6 gives vertices γ p , γ q of G lying in Y or in Z, so that d G (γ p , γ p ), d G (γ q , γ q ) ≤ C 4 .
If one of γ p , γ q lies in Y while the other lies in Z then d G (γ p , γ q ) < 2C 4 + 1.
If both lie in Z then, as Z is a non-hole, there is a vertex δ ∈ G(S) disjoint from both of γ p and γ q and we have If both lie in Y then there are two cases. If Y is not a hole for G(S) then we are done as in the previous case. If Y is a hole then by the definition of shortcut interval, Lemma 5.7, and the triangle inequality we have d W (γ p , γ q ) < 6 + 6C 4 + L 2 for all holes W ⊂ Y . Notice that Y is strictly contained in X. Thus we may inductively apply Theorem 13.1 with c = 6 + 6C 4 + L 2 . We deduce that all terms on the right-hand side of the distance estimate vanish and thus d G (γ p , γ q ) is bounded by a constant depending only on X and G. The same then holds for d G (γ p , γ q ) and we are done.
We are now equipped to give: Proof of Proposition 14.8. Suppose that P(i, j) is the given partition of the electric interval [i, j] into straight and shortcut subintervals. As a bit of notation, if [p, q] = I ∈ P(i, j), we take d G (I) = d G (γ p , γ q ) and d X (I) = d X (γ p , γ q ). Applying Axiom 13.5 we have The last term arises from connecting left endpoints of intervals with right endpoints. We must bound the three terms on the right.
We begin with the third; recall that |P(i, j)| = |P short | + |P str |, that |P str | ≤ |P short |+1, and that |P short |≤ A d X (γ i , γ j ). The second inequality follows from the construction of the partition while the last is implied by Proposition 14.9. Thus the third term of Equation 14.11 is quasibounded above by d X (γ i , γ j ).
By Lemma 14.10, the second term of Equation 14.11 at most L 3 |P short |. Finally, by Axiom 13.11, for all I ∈ P str we have Also, it follows from the reverse triangle inequality (Lemma 13.12) that We deduce that I∈Pstr d G (I) is also quasi-bounded above by d X (γ i , γ j ). Thus for a somewhat larger value of A we find This completes the proof.
14.4. The upper bound. We will need: Proposition 14.12. For any c > 0 there is a constant A with the following property. Suppose that [i, j] = I Y is an inductive interval in P X . Then we have: where Z ranges over all holes for G contained in X.
Proof. Axiom 13.6 gives vertices γ i , γ j ∈ G, contained in Y , so that d G (γ i , γ i ) ≤ C 4 and the same holds for j. Since projection to holes is coarsely Lipschitz (Lemma 5.7) for any hole Z we have d Z (γ i , γ i ) ≤ 3 + 3C 3 .
Fix any c > 0. Now, since to find the required constant A it suffices to bound d G (γ i , γ j ). Let c = c + 6C 3 + 6. Since Y X, induction gives us a constant A so that where N is the number of non-zero terms in the final sum. Also, the sum ranges over sub-holes of Y . We may take A somewhat larger to deal with the term (6C 3 + 6)N and include all holes Z ⊂ X to find where the sum is over all holes Z ⊂ X.
14.5. Finishing the proof. Now we may finish the proof of Theorem 13.1. Fix any constant c ≥ 0. Suppose that X, α X , β X are given as above. Suppose that Γ = {γ i } K i=0 is the given combinatorial path and P X is the partition of [0, K] into inductive and electric intervals. So we have: Again, the last term arises from adjacent right and left endpoints of different intervals.
We must bound the terms on the right-hand side; begin by noticing that |P X | = |P ind | + |P ele |, |P ele | ≤ |P ind | + 1 and |P ind | ≤ A d X (α X , β X ). The second inequality follows from the way the partition is constructed and the last follows from Proposition 14.6. Thus the third term of Equation 14.13 is quasi-bounded above by d X (α X , β X ).
Next consider the second term of Equation 14.13: with the first inequality following from Proposition 14.8 and the second from the reverse triangle inequality (Lemma 13.12).
Finally we bound the first term of Equation 14.13. Let c = c + L 0 . Thus, Here A Y and the first inequality are given by Proposition 14.12. Also A = max{A Y | Y X}. In the last line, each sum of the form I∈P ind [d Z (I)] c has at most three terms, by Remark 14.3 and the fact that c > L 0 . For the moment, fix a hole Z and any three elements I, I , I ∈ P ind .
By the reverse triangle inequality (Lemma 13.12) we find that Thus, where N is the number of non-zero terms in the final sum. Also, the sum ranges over all holes Z X.
Combining the above inequalities, and increasing A once again, implies that where the sum ranges over all holes Z ⊆ X. This completes the proof of Theorem 13.1.

Background on Teichmüller space
Our goal in Sections 16, 17 and 19 will be to verify the axioms stated in Section 13 for the complex of curves of a non-orientable surface, for the arc complex, and for the disk complex. Here we give the necessary background on Teichmüller space.
Fix now a surface S = S g,n of genus g with n punctures. Two conformal structures on S are equivalent, written Σ ∼ Σ , if there is a conformal map f : Σ → Σ which is isotopic to the identity. Let T = T (S) be the Teichmüller space of S; the set of equivalence classes of conformal structures Σ on S.
Define the Teichmüller metric by, where the infimum ranges over all quasiconformal maps f : Σ → Σ isotopic to the identity and where K(f ) is the maximal dilatation of f . Recall that the infimum is realized by a Teichmüller map that, in turn, may be defined in terms of a quadratic differential.

Quadratic differentials.
Definition 15.1. A quadratic differential q(z) dz 2 on Σ is an assignment of a holomorphic function to each coordinate chart that is a disk and of a meromorphic function to each chart that is a punctured disk. If z and ζ are overlapping charts then we require in the intersection of the charts. The meromorphic function q z (z) has at most a simple pole at the puncture z = 0.
At any point away from the zeroes and poles of q there is a natural coordinate z = x + iy with the property that q z ≡ 1. In this natural coordinate the foliation by lines y = c is called the horizontal foliation. The foliation by lines x = c is called the vertical foliation. Now fix a quadratic differential q on Σ = Σ 0 . Let x, y be natural coordinates for q. For every t ∈ R we obtain a new quadratic differential q t with coordinates Also, q t determines a conformal structure Σ t on S. The map t → Σ t is the Teichmüller geodesic determined by Σ and q.

15.2.
Marking coming from a Teichmüller geodesic. Suppose that Σ is a Riemann surface structure on S and σ is the uniformizing hyperbolic metric in the conformal class of Σ. In a slight abuse of terminology, we call the collection of shortest simple non-peripheral closed geodesics the systoles of σ. Fix a constant smaller than the Margulis constant. The -thick part of Teichmüller space consists of those Riemann surfaces such that the hyperbolic systole has length at least . We define P = P (σ), a Bers pants decomposition of S, as follows: pick α 1 , any systole for σ. Define α i to be any systole of σ restricted to S (α 1 ∪ . . . ∪ α i−1 ). Continue in this fashion until P is a pants decomposition. Note that any curve with length less than the Margulis constant will necessarily be an element of P .
Suppose that Σ, Σ ∈ T (S). Suppose that P, P are Bers pants decompositions with respect to Σ and Σ . Suppose also that d T (Σ, Σ ) ≤ 1. Then the curves in P have uniformly bounded lengths in Σ and conversely. By the Collar Lemma, the intersection ι(P, P ) is bounded, solely in terms of ξ(S).
Suppose now that {Σ t | t ∈ [−M, M ]} is the Teichmüller geodesic defined by the quadratic differentials q t . Let σ t be the hyperbolic metric uniformizing Σ t . Let P t = P (σ t ) be a Bers pants decomposition.
We now find transversals in order to complete P t to a Bers marking ν t . Suppose that P t = {α i }. For each i, let A i be the annular cover of S corresponding to α i . Note that q t lifts to a singular Euclidean metric q i t on A i . Let α i be a geodesic representative of the core curve of A i with respect to the metric q i t . Choose γ i ∈ C(A i ) to be any geodesic arc, also with respect to q i t , that is perpendicular to α i . Let β i be any curve in S ({α j } j =i ) which meets α i minimally and so that d Suppose that Σ t and Σ s are surfaces in the -thick part of T (S). We take B 0 sufficiently large so that if ι(ν t , ν s ) ≥ B 0 then d T (Σ t , Σ s ) ≥ 1.
15.3. The marking axiom. We construct a sequence of markings µ n , for n ∈ [0, N ] ⊂ N, as follows. Take µ 0 = ν −M . Now suppose that µ n = ν t is defined. Let s > t be the first time that there is a marking with ι(ν t , ν s ) ≥ B 0 , if such a time exists. If so, let µ n+1 = ν s . If no such time exists take N = n and we are done.
We now show that µ n = ν t and µ n+1 = ν s have bounded intersection. By the above lemma there is a marking ν r with t ≤ r < s and ι(ν r , ν s ) ≤ B 0 .
Since intersection number bounds distance in the marking complex we find that by the triangle inequality, ν t and ν s are bounded distance in the marking complex. Conversely, since distance bounds intersection in the marking complex we find that ι(µ n , µ n+1 ) is bounded. It follows that d Y (µ n , µ n+1 ) is uniformly bounded, independent of Y ⊂ S and of n ∈ [0, N ]. It now follows from Theorem 6.1 of [33] that, for any subsurface Y ⊂ S, the sequence {π Y (µ n )} ⊂ C(Y ) is an unparameterized quasigeodesic. Thus the marking path {µ n } satisfies the second requirement of Axiom 13.3. The first requirement is trivial as every µ n fills S.
15.4. The accessibility axiom. We now turn to Axiom 13.4. Since µ n fills S for every n, the first requirement is a triviality.
In Section 5 of [33] Rafi defines, for every subsurface Y ⊂ S, an interval of isolation I Y inside of the parameterizing interval of the Teichmüller geodesic. Note that I Y is defined purely in terms of the geometry of the given quadratic differentials. Further, for all t ∈ I Y and for all components α ⊂ ∂Y the hyperbolic length of α in Σ t is less than the Margulis constant. Furthermore, by Theorem 5.3 [33], there is So define J Y ⊂ [0, N ] to be the subinterval of the marking path where the time corresponding to µ n lies in I Y . The third requirement follows. Finally, if m ∈ J Y then ∂Y is contained in base(µ m ) and thus ι(∂Y, µ m ) ≤ 2 · |∂Y |.
where µ and µ are Bers markings on Σ and Σ , Y ⊂ S ranges over non-annular surfaces and α ranges over vertices of C(S).

Paths for the non-orientable surface
Fix F a compact, connected, and non-orientable surface. Let S be the orientation double cover with covering map ρ F : S → F . Let τ : S → S be the associated involution. Note that C(F ) = C τ (S). Let C τ (S) → C(S) be the relation sending a symmetric multicurve to its components.
Our goal for this section is to prove Lemma 16.4, the classification of holes for C(F ). As remarked above, Lemma 6.3 and Corollary 6.4 follow, proving the hyperbolicity of C(F ).
16.1. The marking path. We will use the extreme rigidity of Teichmüller geodesics to find τ -invariant marking paths. We first show that τ -invariant Bers pants decompositions exist.
Proof. Let P 0 = ∅. Suppose that 0 ≤ k < ξ(S) curves have been chosen to form P k . By induction we may assume that P k is τ -invariant. Let Y be a component of S P k with ξ(Y ) ≥ 1. Note that since τ is orientation reversing, τ does not fix any boundary component of Y .
Pick any systole α for Y .
If β is null-homotopic then α ∪ τ (α) cuts a monogon or a bigon out of S, contradicting our assumption that α was a geodesic. Suppose, by way of contradiction, that β is homotopic to some boundary curve b ⊂ ∂Y . Since τ (β ) = β , it follows that τ (b) and β are also homotopic. Thus b and τ (b) cobound an annulus, implying that Y is an annulus, a contradiction. The same holds for γ .
Let β and γ be the geodesic representatives of β and γ . Since β and τ (β) meet transversely, β has length in σ strictly smaller than 2 σ (β). Similarly the length of γ is strictly smaller than 2 σ (γ). Suppose that β is shorter then γ . It follows that β strictly shorter than α. If β is embedded then this contradicts the assumption that α was shortest.
If β is not embedded then there is an embedded curve β inside of a regular neighborhood of β which is again essential, non-peripheral, and has geodesic representative shorter than β . This is our final contradiction and the claim is proved.
Transversals are chosen with respect to a quadratic differential metric. Suppose that α, β ∈ C τ (S). If α and β do not fill S then we may replace S by the support of their union. Following Thurston [38] there exists a square-tiled quadratic differential q with squares associated to the points of α ∩ β. (See [6] for analysis of how the square-tiled surface relates to paths in the curve complex.) Let q t be image of q under the Teichmüller geodesic flow. We have: Proof. Note that τ preserves α and also β. Since τ permutes the points of α ∩ β it permutes the rectangles of the singular Euclidean metric q t while preserving their vertical and horizontal foliations. Thus τ is an isometry of the metric and the conclusion follows.
We now choose the Teichmüller geodesic {Σ t | t ∈ [−M, M ]} so that the hyperbolic length of α is less than the Margulis constant in σ −M and the same holds for β in σ M . Also, α is the shortest curve in σ −M and similarly for β in σ M Lemma 16.3. Fix t. There are transversals for P t which are close to being quadratic perpendicular in q t and which are τ -invariant.
Proof. Let P = P t and fix α ∈ P . Let X = S (P α). There are two cases: either τ (X) ∩ X = ∅ or τ (X) = X. Suppose the former. So we choose any transversal β ⊂ X close to being q t -perpendicular and take τ (β) to be the transversal to τ (α).
Suppose now that τ (X) = X. It follows that X is a four-holed sphere. The quotient X/τ is homeomorphic to a twice-holed RP 2 . Therefore there are only four essential non-peripheral curves in X/τ . Two of these are cores of Möbius bands and the other two are their doubles. The cores meet in a single point. Perforce α is the double cover of one core and we take β the double cover of the other.
It remains only to show that β is close to being q t -perpendicular. Let S α be the annular cover of S and lift q t to S α . Let ⊥ be the set of q α t -perpendiculars. This is a τ -invariant diameter one subset of C(S α ).
Thus τ -invariant markings exist; these have bounded intersection with the markings constructed in Section 15. It follows that the resulting marking path satisfies the marking path and accessibility requirements, Axioms 13.3 and 13.4. To choose the combinatorial path, pick γ n ∈ base(µ n ) so that γ n is a τ -invariant curve or pair of curves and so that γ n is shortest in base(µ n ).
16.3. The classification of holes. We now finish the classification of large holes for C τ (S). Fix L 0 > 3C 3 + 2C 2 + 2C 1 . Note that these constants are available because we have verified the axioms that give them.
Proof. Let (Σ t , q t ) be the Teichmüller geodesic defined above and let σ t be the uniformizing hyperbolic metric. Since L 0 > C 3 + 2C 2 it follows from the accessibility requirement that J X = [m, n] is non-empty. Now for all t in the interval of isolation I X σt (δ) < , where δ is any component of ∂X and is the Margulis constant. Let Y = τ (X). Since τ is an isometry (Lemma 16.2) and since the interval of isolation is metrically defined we have I Y = I X and thus J Y = J X .
Deduce that ∂Y is also short in σ t . This implies that ∂X ∩ ∂Y = ∅. If X and Y overlap then by (iii) of Lemma 13.13 we have d X (µ m , µ n ) < C 3 and so by the triangle inequality, two applications of (2) of Axiom 13.4, we have d X (µ 0 , µ N ) < 3C 3 . By the combinatorial axiom it follows that d X (α, β) < 3C 3 + 2C 2 a contradiction. Deduce that either X = Y or X ∩Y = ∅ as desired.
As noted in Section 6 this shows that the only hole for C τ (S) is S itself. Thus all holes trivially interfere, verifying Axiom 13.2.
16.4. The replacement axiom. We now verify Axiom 13.6 for subsurfaces Y ⊂ S with d Y (α, β) ≥ L 0 . (We may ignore all subsurfaces with smaller projection by taking L 1 (Y ) > L 0 .) By Lemma 16.4 the subsurface Y is symmetric. If Y is a hole then Y = S and the first requirement is vacuous. Suppose that Y is not a hole. Suppose that γ n is such that n ∈ J Y . Thus γ n ∈ base(µ n ). All components of ∂Y are also pants curves in µ n . It follows that we may take any symmetric curve in ∂Y to be γ and we are done. 16.5. On straight intervals. Lastly we verify Axiom 13.11. Suppose that [p, q] is a straight interval. We must show that d C τ (γ p , γ q ) ≤ d S (γ p , γ q ). Suppose that µ p = ν s and µ q = ν t ; that is, s and t are the times when µ p , µ q are short markings. Thus d X (µ p , µ q ) ≤ L 2 for every X S. This implies that the Teichmüller geodesic, along the straight interval, lies in the thick part of Teichmüller space.
Notice that d C τ (γ p , γ q ) ≤ C 2 |p − q|, since for all i ∈ [p, q − 1], d C τ (γ i , γ i+1 ) ≤ C 2 . So it suffices to bound |p − q|. By our choice of B 0 and because the Teichmüller geodesic lies in the thick part we find that |p − q| ≤ d T (Σ s , Σ t ). Rafi's distance estimate (Theorem 15.3) gives: Since ν s = µ p , ν t = µ q , and since γ p ∈ base(µ p ), γ q ∈ base(µ q ) deduce that d S (µ p , µ q ) ≤ d S (γ p , γ q ) + 4. This verifies Axiom 13.11. Thus the distance estimate holds for C τ (S) = C(F ). Since there is only one hole for C(F ) we deduce that the map C(F ) → C(S) is a quasi-isometric embedding. As a corollary we have: Theorem 16.5. The curve complex C(F ) is Gromov hyperbolic.

Paths for the arc complex
Here we verify that our axioms hold for the arc complex A(S, ∆). It is worth pointing out that the axioms may be verified using Teichmüller geodesics, train track splitting sequences, or resolutions of hierarchies. Here we use the former because it also generalizes to the non-orientable case; this is discussed at the end of this section.
First note that Axiom 13.2 follows from Lemma 7.3.
Recall that σ S : A(X) → C(X) is the surgery map, defined in Definition 4.2. Let α = σ S (α) and define β similarly. Note that α cuts a pants off of S. As usual, we may assume that α and β fill X. If not we pass to the subsurface they do fill. As in the previous sections let q be the quadratic differential determined by α and β . Exactly as above, fix a marking path {µ n } N n=0 . This path satisfies the marking and accessibility axioms (13.3, 13.4).
17.2. The combinatorial path. Let Y n ⊂ X be any component of X base(µ n ) meeting ∆. So Y n is a pair of pants. Let γ n be any essential arc in Y n with both endpoints in ∆. Since α ⊂ base(µ 0 ) and β ⊂ base(µ N ) we may choose γ 0 = α and γ N = β.
As in the previous section the reindexing map is the identity. It follows immediately that ι(γ n , µ n ) ≤ 4. This bound, the bound on ι(µ n , µ n+1 ), and Lemma 4.7 imply that ι(γ n , γ n+1 ) is likewise bounded. The usual surgery argument shows that if two arcs have bounded intersection then they have bounded distance. This verifies Axiom 13.5.
17.3. The replacement and the straight axioms. Suppose that Y ⊂ X is a subsurface and γ n has n ∈ J Y . Let µ n = ν t ; that is t is the time when µ n is a short marking. Thus ∂Y ⊂ base(µ n ) and so γ n ∩ ∂Y = ∅. So regardless of the hole-nature of Y we may take γ = γ n and the axiom is verified.
Axiom 13.11 is verified exactly as in Section 16.
17.4. Non-orientable surfaces. Suppose that F is non-orientable and ∆ F is a collection of boundary components. Let S be the orientation double cover and τ : S → S the involution so that S/τ = F . Let ∆ be the preimage of ∆ F . Then A τ (S, ∆) is the invariant arc complex. Suppose that α F , β F are vertices in A(F, ∆ ). Let α, β be their preimages. As above, without loss of generality, we may assume that σ F (α F ) and σ F (β F ) fill F . Note that σ F (α F ) cuts a surface X off of F . The surface X is either a pants or a twice-holed RP 2 . When X is a pants we define α ⊂ S to be the preimage of σ F (α F ). When X is a twice-holed RP 2 we take γ F to be a core of one of the two Möbius bands contained in X and we define α to be the preimage of γ F ∪ σ F (α F ). We define β similarly. Notice that α and α meet in at most four points.
We now use α and β to build a τ -invariant Teichmüller geodesic. The construction of the marking and combinatorial paths for A τ (S, ∆) is unchanged. Notice that we may choose combinatorial vertices because base(µ n ) is τ -invariant. There is a small annoyance: when X is a twice-holed RP 2 the first vertex, γ 0 , is disjoint from but not equal to α. Strictly speaking, the first and last vertices are γ 0 and γ N ; our constants are stated in terms of their subsurface projection distances. However, since α∩γ 0 = ∅, and the same holds for β, γ N , their subsurface projection distances are all bounded.

Background on train tracks
Here we give the necessary definitions and theorems regarding train tracks. The standard reference is [31]. See also [30]. We follow closely the discussion found in [27].
18.1. On tracks. A generic train track τ ⊂ S is a smooth, embedded trivalent graph. As usual we call the vertices switches and the edges branches. At every switch the tangents of the three branches agree. Also, there are exactly two incoming branches and one outgoing branch at each switch. See Figure 7 for the local model of a switch. Let B(τ ) be the set of branches. A transverse measure on τ is function w : B → R ≥0 satisfying the switch conditions: at every switch the sum of the incoming measures equals the outgoing measure. Let P (τ ) be the projectivization of the cone of transverse measures. Let V (τ ) be the vertices of P (τ ). As discussed in the references, each vertex measure gives a simple closed curve carried by τ .
For every track τ we refer to V (τ ) as the marking corresponding to τ (see Section 2.4). Note that there are only finitely many tracks up to the action of the mapping class group. It follows that ι(V (τ )) is uniformly bounded, depending only on the topological type of S.
If τ and σ are train tracks, and Y ⊂ S is an essential surface, then define We also adopt the notation π Y (τ ) = π Y (V (τ )). A train track σ is obtained from τ by sliding if σ and τ are related as in Figure 8. We say that a train track σ is obtained from τ by splitting if σ and τ are related as in Figure 9.  Again, since the number of tracks is bounded (up to the action of the mapping class group) if σ is obtained from τ by either a slide or a split we find that ι(V (τ ), V (σ)) is uniformly bounded.
18.2. The marking path. We will use sequences of train tracks to define our marking path. of train tracks so that τ n+1 is obtained from τ n by a slide or a split.
The sequence {τ n } gives a sequence of markings via the map τ n → V n = V (τ n ). Note that the support of V n+1 is contained within the support of V n because every vertex of τ n+1 is carried by τ n . Theorem 5.5 of [27] verifies the remaining half of Axiom 13.3. Theorem 18.2. Fix a surface S. There is a constant A with the following property. Suppose that {τ n } N n=0 is a sliding and splitting sequence in S of birecurrent tracks. Suppose that Y ⊂ S is an essential surface. Then the map n → π Y (τ n ), as parameterized by splittings, is an A-unparameterized quasi-geodesic.
Note that, when Y = S, Theorem 18.2 is essentially due to the first author and Minsky; see Theorem 1.3 of [26].
In Section 5. 18.3. Quasi-geodesics in the marking graph. We will also need Theorem 6.1 from [27]. (See [16] for closely related work.) Theorem 18.3. Fix a surface S. There is a constant A with the following property. Suppose that {τ n } N n=0 is a sliding and splitting sequence of birecurrent tracks, injective on slide subsequences, where V N fills S. Then {V (τ n )} is an A-quasi-geodesic in the marking graph.

Paths for the disk complex
Suppose that V = V g is a genus g handlebody. The goal of this section is to verify the axioms of Section 13 for the disk complex D(V ) and so complete the proof of the distance estimate.
Remark 19.3. For the details of the proof we refer to [26]. Note that the double-wave curve replacements of that paper are not needed here; as X is a hole, no curve of ∂X compresses in V . It follows that consecutive disks in the surgery sequence are disjoint (as opposed to meeting at most four times). Also, in the terminology of [27], the disk D i is a wide dual for the track τ r(i) . Finally, recurrence of τ n follows because E is fully carried by τ N . Transverse recurrence follows because D is fully dual to τ 0 .
Thus V n will be our marking path and D i will be our combinatorial path. The requirements of Axiom 13.5 are now verified by Theorem 19.2.
19.3. The replacement axiom. We turn to Axiom 13.6. Suppose that Y ⊂ X is an essential subsurface and D i has r(i) ∈ J Y . Let n = r(i). From Theorem 19.2 we have that ι(∂D i , V n ) is uniformly bounded. By Axiom 13.4 we have Y ⊂ supp(V n ) and ι(∂Y, µ n ) is bounded. It follows that there is a constant K depending only on ξ(S) so that ι(∂D i , ∂Y ) < K. Isotope D i to have minimal intersection with ∂Y . As in Section 11.1 boundary compress D i as much as possible into the components of X ∂Y to obtain a disk D so that either • D cannot be boundary compressed any more into X ∂Y or • D is disjoint from ∂Y . We may arrange matters so that every boundary compression reduces the intersection with ∂Y by at least a factor of two. Thus: Suppose now that Y is a compressible hole. By Lemma 8.4 we find that ∂D ⊂ Y and we are done.
Suppose now that Y is an incompressible hole. Since Y is large there is an I-bundle T → F , contained in the handlebody V , so that Y is a component of ∂ h T . Isotope D to minimize intersection with ∂ v T . Let ∆ be the union of components of ∂ v T which are contained in ∂V . Let Γ = ∂ v T ∆. Notice that all intersections D ∩ Γ are essential arcs in Γ: simple closed curves are ruled out by minimal intersection and inessential arcs are ruled out by the fact that D cannot be boundary compressed in the complement of ∂Y . Let D be a outermost component of D Γ. Then Lemma 8.5 implies that D is isotopic in T to a vertical disk.
If D = D then we may replace D i by the arc ρ F (D ). The inductive argument now occurs inside of the arc complex A(F, ρ F (∆)).
Suppose that D = D . Let A ∈ Γ be the vertical annulus meeting D . Let N be a regular neighborhood of D ∪ A, taken in T . Then the frontier of N in T is again a vertical disk, call it D . Note that ι(D , D ) < K − 1. Finally, replace D i by the arc ρ F (D ).
Suppose now that Y is not a hole. Then some component S Y is compressible. Applying Lemma 8.4 again, we find that either D lies in Z = X Y or in Y . This completes the verification of Axiom 13.6.

Straight intervals.
We end by checking Axiom 13.11. Suppose that [p, q] ⊂ [0, K] is a straight interval. Recall that d Y (µ r(p) , µ r(q) ) < L 2 for all strict subsurfaces Y ⊂ X. We must check that Note that |p−q| ≤ |r(p)−r(q)| because the reindexing map is increasing. Now, |r(p) − r(q)| ≤ A d M(X) (µ r(p) , µ r(q) ) because the sequence {µ n } is a quasi-geodesic in M(X) (Theorem 18.3). Increasing A as needed and applying Theorem 4.10 we have d M (µ r(p) , µ r(q) ) ≤ A Y [d Y (µ r(p) , µ r(q) )] L 2 and the right hand side is thus less than d X (µ r(p) , µ r(q) ) which in turn is less than d X (D p , D q ) + 2C 2 . This completes our discussion of Axiom 13.11 and finishes the proof of Theorem 19.1.

Hyperbolicity
The ideas in this section are related to the notion of "time-ordered domains" and to the hierarchy machine of [25] (see also Chapters 4 and 5 of Behrstock's thesis [1]). As remarked above, we cannot use those tools directly as the hierarchy machine is too rigid to deal with the disk complex. 20.1. Hyperbolicity. We prove: Theorem 20.1. Fix G = G(S), a combinatorial complex. Suppose that G satisfies the axioms of Section 13. Then G is Gromov hyperbolic. In fact, Theorem 20.1 follows quickly from: Theorem 20.4. Fix G, a combinatorial complex. Suppose that G satisfies the axioms of Section 13. Then for all A ≥ 1 there exists δ ≥ 0 with the following property: Suppose that T ⊂ G is a triangle of paths where the projection of any side of T into into any hole is an A-unparameterized quasi-geodesic. Then T is δ-slim.

As corollaries we have
Proof of Theorem 20.1. As laid out in Section 14 there is a uniform constant A so that for any pair α, β ∈ G there is a recursively constructed path P = {γ i } ⊂ G so that • for any hole X for G, the projection π X (P) is an A-unparameterized quasi-geodesic and For any hole X and for any geodesic k ∈ C(X) connecting a point of π X (α) to a point of π X (β) we also define ρ k : G → k to be the relation π X |G : G → C(X) followed by taking closest points in k. Since the diameter of ρ k (γ) is uniformly bounded, we may simplify our formulas by treating ρ k as a function. Define index X : G → N to be the index in X: index X (σ) = d X (α, ρ k (σ)).
Remark 20.5. Suppose that k is a different geodesic connecting π X (α) to π X (β) and index X is defined with respect to k . Then | index X (σ) − index X (σ)| ≤ 17δ + 4 by Lemma 3.7 and Lemma 3.8. After permitting a small additive error, the index depends only on α, β, σ and not on the choice of geodesic k.
We say σ precedes τ by at most K if the inequality is reversed. If σ precedes τ then we say τ succeeds σ. Now take P = {σ i } to be a path in G connecting α to β. Recall that we have made the simplifying assumption that σ i and σ i+1 are disjoint.
We formalize a pair of properties enjoyed by unparameterized quasigeodesics. The path P backtracks at most K if for every hole X and all indices i < j we find that σ j precedes σ i by at most K. The path P sidetracks at most K if for every hole X and every index i we find that d X (σ i , ρ k (σ i )) ≤ K, for some geodesic k connecting a point of π X (α) to a point of π X (β). Now choose K 1 ≥ max{M 0 , 4M 1 , M 2 , 8} + 8δ sufficiently large so that any A-unparameterized quasi-geodesic in any hole back and side tracks at most K 1 .
Claim 20.9. If σ i precedes γ in X and σ j succeeds γ in Y , both by at least 2K 1 , then i < j.
Proof. To begin, as X and Y are holes and all holes interfere, we need not consider the possibility that X ∩ Y = ∅. If X = Y we immediately deduce that index X (σ i ) + 2K 1 ≤ index X (γ) ≤ index X (σ j ) − 2K 1 .
Thus index X (σ i ) + 4K 1 ≤ index X (σ j ). Since P backtracks at most K 1 we have i < j, as desired.
Suppose instead that X ⊂ Y . Since σ i precedes γ in X we immediately find d X (α, β) ≥ 2K 1 ≥ M 0 and d X (α, γ) ≥ 2K 1 −2δ ≥ M 2 . Apply Lemma 20.8 to deduce index Y (∂X) − 4 ≤ index Y (γ). Since σ j succeeds γ in Y it follows that index Y (∂X) − 4 + 2K 1 ≤ index Y (σ j ). Again using the fact that σ i precedes γ in X we have that d X (σ i , β) ≥ M 2 . We deduce from Lemma 20.8 that index Y (σ i ) ≤ index Y (∂X) + 4. Thus Since P backtracks at most K 1 in Y we again deduce that i < j. The case where Y ⊂ X is similar.
Suppose now that X and Y overlap. Applying Lemma 20.7 and breaking symmetry, we may assume that d X (γ, ∂Y ) < M 1 . Since σ i precedes γ we have index X (γ) ≥ 2K 1 . Lemma 3.7 now implies that index X (∂Y ) ≥ 2K 1 − M 1 − 6δ. Thus, where the first inequality follows from Lemma 3.4.
Thus d X (γ, σ j ) ≤ 2M 1 . Finally, Lemma 3.7 implies that the difference in index (in X) between σ i and σ j is at least 2K 1 − 2M 1 − 6δ. Since this is greater than the backtracking constant, K 1 , it follows that i < j.
We define σ β to be the first σ i to succeed γ by at least 2K 1 -if no such vertex of P exists take σ β = β. If α = β then σ α = σ β . Otherwise, from Claim 20.9, we immediately deduce that σ α comes before σ β in P. A symmetric version of Claim 20.10 applies to σ β : for every hole X d X (ρ h (γ), σ β ) ≤ 3K 1 + 6δ + 3. 20.6. Another side of the triangle. Recall now that we are also given a path R = {τ i } connecting α to γ in G. As before, R has bounded back and sidetracking. Thus we again find vertices τ α and τ γ the last/first to precede/succeed β by at least 2K 1 . Again, this is defined in terms of the closest points projection of β to a geodesic of the form h = [π X (α), π X (γ)]. By Claim 20.10, for every hole X, τ α and τ γ are close to ρ h (β).
This claim and Claim 20.10 imply that the body of the triangle PQR is bounded in size. We now show that the legs are narrow.
Claim 20.12. There is a constant N 2 = N 2 (S) with the following property. For every σ i ≤ σ α in P there is a τ j ≤ τ α in R so that d X (σ i , τ j ) ≤ N 2 for every hole X.
Proof. We only sketch the proof, as the details are similar to our previous discussion. Fix σ i ≤ σ α .
Suppose first that no vertex of R precedes σ i by more than 2K 1 in any hole. So fix a hole X and geodesics k = [π X (α), π X (β)] and h = [π X (α), π X (γ)]. Then ρ h (σ i ) is within distance 2K 1 of π X (α). Appealing to Claim 20.11, bounded sidetracking, and hyperbolicity of C(X) we find that the initial segments [π X (α), ρ k (σ α )], [π X (α), ρ h (τ α )] of k and h respectively must fellow travel. Because of bounded backtracking along P, ρ k (σ i ) lies on, or at least near, this initial segment of k. Thus by Lemma 3.8 ρ h (σ i ) is close to ρ k (σ i ) which in turn is close to π X (σ i ), because P has bounded sidetracking. In short, d X (α, σ i ) is bounded for all holes X. Thus we may take τ j = τ 0 = α and we are done. Now suppose that some vertex of R precedes σ i by at least 2K 1 in some hole X. Take τ j to be the last such vertex in R. Following the proof of Claim 20.9 shows that τ j comes before τ α in R. The argument now required to bound d X (σ i , τ j ) is essentially identical to the proof of Claim 20.10.
By the distance estimate, we find that there is a uniform neighborhood of [σ 0 , σ α ] ⊂ P, taken in G, which contains [τ 0 , τ α ] ⊂ P. The slimness of PQR follows directly. This completes the proof of Theorem 20.4.

Coarsely computing Hempel distance
We now turn to our topological application. Recall that a Heegaard splitting is a triple (S, V, W ) consisting of a surface and two handlebodies where V ∩ W = ∂V = ∂W = S. Hempel [20] defines the quantity  Theorem 21.2. There is a constant R 0 = R 0 (V ) and an algorithm that, given an essential curve α ⊂ S and a cut system D ⊂ D(V ), finds a disk C ∈ D(V ) so that d S (C, ρ V (α)) ≤ R 0 .
Proof of Theorem 21.1. Suppose that (S, D, E) is a Heegaard diagram. Using Theorem 21.2 we find a disk D within distance R 0 of ρ V (E). Again using Theorem 21.2 we find a disk E within distance R 0 of ρ W (D). Notice that E is defined using D and not the cut system D.
Since computing distance between fixed vertices in the curve complex is algorithmic [22,37] we may compute d S (D, E). By the hyperbolicity of C(S) (Theorem 3.2) and by the quasi-convexity of the disk set (Theorem 4.9) this is the desired estimate.
Very briefly, the algorithm asked for in Theorem 21.2 searches an R 2 -neighborhood in M(S) about a splitting sequence from D to α.
Here are the details. Algorithm 21.3. We are given α ∈ C(S) and a cut system D ⊂ D(V ). Build a train track τ in S = ∂V as follows: make D and α tight. Place one switch on every disk D ∈ D. Homotope all intersections of α with D to run through the switch. Collapse bigons of α inside of S D to create the branches. Now make τ a generic track by combing away from D [31, Proposition 1.4.1]. Note that α is carried by τ and so gives a transverse measure w.
Build a splitting sequence of measured tracks {τ n } N n=0 where τ 0 = τ , τ N = α, and τ n+1 is obtained by splitting the largest switch of τ n (as determined by the measure imposed by α).
Let µ n = V (τ n ) be the vertices of τ n . For each filling marking µ n list all markings in the ball B(µ n , R 2 ) ⊂ M(S), where R 2 is given by Lemma 21.5 below. (If µ 0 does not fill S then output D and halt.) For every marking ν so produced we use Whitehead's algorithm (see Lemma 21.4) to try and find a disk meeting some curve γ ∈ ν at most twice. For every disk C found compute d S (α, C) [22,37]. Finally, output any disk which minimizes this distance, among all disks considered, and halt.
We use the following form of Whitehead's algorithm [3]: Lemma 21.4. There is an algorithm that, given a cut system D ⊂ V and a curve γ ⊂ S, outputs a disk C ⊂ V so that ι(γ, ∂C) = min{ι(γ, ∂E) | E ∈ D(V )}. constant larger than max{R 3 , C 0 , M 3 + δ + Q} ensures that all terms on the right-hand side vanish. In this case the additive error in Theorem 4.10 is the desired constant R 2 and the lemma is proved.
If there are such surfaces then choose one, say Z, that minimizes = min J Z . Thus d Y (µ n , ν ) < C 3 for all strict non-holes and all strict large holes. Since d S (µ n , E n ) ≤ M 3 + δ + Q and {ν m } is an unparameterized quasi-geodesic [33, Theorem 6.1] we find that d S (µ n , ν l ) is uniformly bounded. The claim above bounds distances in small holes. As before we find a sufficiently large cutoff so that all terms on the right-hand side of Theorem 4.10 vanish. Again the additive error of Theorem 4.10 provides the constant R 2 . Since ∂Z ⊂ base(ν ) the lemma is proved.
To prove the correctness of Algorithm 21.3 it suffices to show that the disk produced is close to ρ V (α). Let m be the largest index so that for all n ≤ m we have d S (µ n , E n ) ≤ M 3 + δ + Q.
If no such n exists then take n = 0. Now, Lemma 21.5 implies that there is a disk C with d S (C, µ n ) ≤ 4R 2 and this disk is found during the running of Algorithm 21.3. It follows from the above inequalities that d S (C, α) ≤ 4R 2 + 5M 3 + 5δ + 3Q + 5 + 2C 1 + d S (α, ρ V (α)).
This completes the proof of Theorem 21.2.