On train track splitting sequences

Let \tau be a complete track extending \sigma [PennerHarer92, Corollary 1.4.2]. Section 3.4 of [PennerHarer92] and a dimension count gives a lamination \lambda \pitchfork \tau so that i(\lambda, \alpha) \neq 0 for all \alpha \in \mathcal{C}(S). Now an argument of Kobayashi [Kobayashi88b], refined by Luo [MasurMinsky99, page 124], implies that \mathcal{C}^*(\tau) \subset \mathcal{C}^*(\sigma) has infinite diameter.

Some multiple of \alpha \prec \tau is a sum of vertices: m \cdot w_\alpha = \sum n_i w_i where w_i is the integral transverse measure for the vertex \beta_i \in V(\tau). Via a sequence of tie-preserving isotopies of N we may arrange for all of the \beta_i to realize their geometric intersection with each other. Note that there is an isotopy representative of \alpha contained inside of a small neighborhood of the union B = \cup \beta_i.

To prove the contrapositive, suppose that none of the \beta_i cut X. It follows that X may be isotoped in S to be disjoint from B. Thus \alpha misses X, as desired. A similar discussion applies when \alpha \pitchfork \tau.

Note train-routes in \tau^X that are mapped properly to S^X are uniform quasi-geodesics in S^X [Mosher03, Proposition 3.3.3].

Thus there is a compact core X' \subset S^X, homeomorphic to X, so that any route meeting S^X {\smallsetminus} X' has one endpoint on the Gromov boundary of S^X. It follows that \tau|X \subset X'.

The first claim follows from the definitions. An index argument proves the second claim.

The proof is modelled on that of Lemma 2.8. If \alpha \prec \tau^X is wide we are done. If not, as \alpha is a quasi-geodesic [Mosher03, Proposition 3.3.3],

orient \alpha so that \alpha is wide outside of a compact core for S^X. Now we induct on the total number of arcs of intersection between \alpha and rectangles R_b \subset N(\tau^X) meeting the compact core.

Let t be a tie of R_b. Orient t. Suppose that x, y are consecutive (along t) points of \alpha \cap t. Suppose that the sign of intersection at x equals the sign at y. Let [x,y] be the subarc of t bounded by x and y. As in Lemma 2.8 surger \alpha along [x,y] to form an arc \beta' and a curve \gamma. See Figure ? with \beta' substituted for \beta.

Note that \gamma is essential in S^X, by an index argument. If \gamma is non-peripheral then the second conclusion holds. So suppose that \gamma is peripheral. Then \alpha is obtained by Dehn twisting \beta' about \gamma. So \beta' is properly isotopic to \alpha and has smaller intersection with R_b; thus we are done by induction.

Suppose instead that x, y, z are consecutive (along t) points of \alpha \cap t, with alternating sign. Surger \alpha along [x,z] to form an arc \beta' and a curve \gamma. See Figure ?, with \beta' substituted for \beta, for one of the possible arrangements of \alpha, \beta', and \gamma. Again \gamma is essential. If \gamma is non-peripheral then the second conclusion holds and we are done. If \gamma is peripheral then, as \alpha and \beta' differ by a half-twist about \gamma, we find that \beta' is properly isotopic to \alpha. Since \beta' has smaller intersection with R_b we are done by induction.

All that remains is the case that \alpha meets every rectangle R_b in most a pair of arcs of opposite orientation. For every branch b where \alpha meets R_b twice, choose a subarc t_b of a tie in R_b so that \alpha \cap t_b = \partial t_b. We call t_b a chord for \alpha. For every t_b there is a subarc \alpha_b \subset \alpha so that \partial t_b = \partial \alpha_b. A chord t_b is innermost if there is no chord t_c with \alpha_c strictly contained in \alpha_b. Let t_b be the first innermost chord. Let \alpha' be the component of \alpha {\smallsetminus} \alpha_b before \alpha_b. Build a route \beta by taking two copies of \alpha' and joining them to \alpha_b. Note that \beta \cap R_c is a single arc or a pair of arcs exactly as \alpha_b or \alpha' meets R_c. Thus \beta is wide. Also, \beta is essential: otherwise t_b \cup \alpha_b bounds a disk with index one-half, a contradiction. By construction i(\alpha, \beta) = 0 and Lemma 3.3 is proved.

Let \Gamma = \{ \gamma_i \}_{i=1}^{k}. By induction, the curves of \Gamma' = \Gamma {\smallsetminus} \{ \gamma_k \} realize their pairwise geometric intersection numbers. If \gamma_k meets some \gamma_i \in \Gamma' non-minimally then by the Bigon Criterion [FLP91, page 46] there is an innermost bigon between \gamma_k and some \gamma_j \in \Gamma'. We now may reduce the intersection number following the proof of uniqueness of efficient position.

Note that there are infinitely many half-branches on both sides of \rho: if not then \partial_h N(\tau) would have a component without corners, contrary to assumption.

Suppose that there are only finitely many rising branches along \rho. Then there is a curve \gamma \prec \tau so that w_\gamma(b) \leq 1 for every branch b and so that the two sides of \gamma are combed in opposite directions.

Thus \tau is not recurrent, a contradiction. The same contradiction is obtained if there are only finitely many falling branches along \rho.

Let x = \rho(\infty) \in \partial_\infty \mathbb{H}. Consider the subsequence \{ \beta_n \} where the first branch of each \beta_n is falling. Let P_n = \rho|(-\infty, n] \cup \beta_n, oriented away from \rho(-\infty). Note that P_n(\infty) = \beta_n(\infty). Recall that \rho and P_n are both uniformly close to geodesics [PennerHarer92, pages 61–62]. Thus P_n(\infty) \to x as n \to \infty.

Now consider the subsequence \{ \beta_n \} where the first branch of each \beta_n is rising. Let P_n = \beta_n \cup \rho|[n, \infty) oriented towards x; so P_n(\infty) = x for all n. Note that P(-\infty) = \beta_n(\infty). Since all complementary regions of \tau^\mathbb{H} have negative index none of the P_n may cross each other. It follows that either the P_n exit compact subsets of \mathbb{H}, and we are done, or the P_n converge [PennerHarer92, Theorem 1.5.4] to P, a train-route with P(\infty) = x. Since P does not cross any P_n deduce that P and \rho are disjoint. But this contradicts [Mosher03, Corollary 3.3.4]: train-routes that share an endpoint must share a half-route.

The endpoints of train-routes are dense in S^1 = \partial_\infty \mathbb{H}.

Fix x \in (z, y) so that x is the endpoint of a train-route \gamma. Since there are infinitely many rising branches along \gamma (Claim 4.4) the claim follows from the rising case of Claim 4.5.

This is proved in detail on pages 122-123 of [MasurMinsky99].

Pick a sequence of subintervals (x_n, y_n) \subset (x,y) so that x_n, y_n \to x as n \to \infty. By Claim 4.6 for every n there is a route \rho_n and a half-plane H_n = H(\rho_n) so that \partial_\infty H_n \subset (x_n, y_n). It follows that H_n \subset \mathcal{H}_{x,y}. Let r_n be any bi-infinite geodesic perpendicular to \partial H_{x,y} and meeting H_n. Thus r_n \to x as n \to \infty. By Claim 4.7 the intersection r_n \cap \partial (\mathbb{H} {\smallsetminus} \mathcal{H}_{x,y}) is nonempty, and we are done.

Let y, z be arbitrary points of S^1 so that x, y, z are ordered counterclockwise. It suffices to construct a train-route separating x from (y,z).

First assume that x is the endpoint of a route \rho. Claim 4.4 implies that there are infinitely many rising branches \{ a_m \} on one side of \rho and infinitely many falling branches \{ c_n \} on the other side. Run half-routes \alpha_m and \gamma_n through a_m and c_n; so each half-route meets \rho in a single switch. By Claim 4.5 the endpoints converge \alpha_m(\infty), \gamma_n(\infty) \to x. Thus sufficiently large m, n give a train-route

that separates x from (y,z), as desired.

For the general case consider

Note that x is an accumulation point of \mathcal{K} (by Claim 4.9 and because \mathcal{H}_{z,x} cannot contain points of \partial (\mathbb{H} {\smallsetminus} \mathcal{H}_{x,y})).

Fix any basepoint w \in \mathcal{K}. By Claim 4.8 \mathcal{K} is \epsilon'–convex. Thus there is a path r \subset \mathcal{K} from w to x which has geodesic curvature less than \epsilon' at every point. Since x is not a parabolic point the projection of r to S recurs to the thick part of S; thus r meets infinitely many branches \{ b_n \} of \tau^\mathbb{H}.

Suppose that b is a branch of \tau^\mathbb{H} lying in \mathcal{K}. If the two sides of b meet \mathcal{H}_{z,x} and \mathcal{H}_{x,y} then b is a bridge of \mathcal{K}. If the sides of b meet neither \mathcal{H}_{z,x} nor \mathcal{H}_{x,y} then b is an interior branch of \mathcal{K}. If exactly one side of b lies in \mathcal{K} then b is a boundary branch. If both sides lie in \mathcal{H}_{z,x} (or both sides lie in \mathcal{H}_{x,y}) then b is an exterior branch. See Figure ? below.

By convexity the path r is disjoint from the exterior branches. After a small isotopy the path r is also disjoint from the boundary branches, meets the interior branches transversely, and still has geodesic curvature less than \epsilon' at every point.

Now, if r travels along a bridge then there are routes \rho^\pm cutting off half-planes H^\pm lying in \mathcal{H}_{z,x} and \mathcal{H}_{x,y} respectively. Then either x is the endpoint of a train-route or a cut and paste of \rho^\pm gives the desired route \Gamma separating x from (y,z). In either case we are done.

So suppose that r only meets interior branches \{ b_n \} of \mathcal{K}. Let \gamma_n be any train-route travelling along b_n. If any of the \gamma_n land at x we are done, as above. Supposing not: Fixing orientations and passing to a subsequence we may assume that \gamma_n(\infty) \to x as n \to \infty. There are now two cases: suppose that for infinitely many n we find that \gamma_n(-\infty) \in [y,z]. Then passing to a further subsequence we have that \gamma_n \to \Gamma where \Gamma(\infty) = x [PennerHarer92, Theorem 1.5.4]; thus x is the endpoint of a train-route and we are done as above. The other possibility is that for some sufficiently large n both endpoints \gamma_n(\pm\infty) lie in (z, y). Since b_n is an interior branch \gamma_n separates x from (y,z) and Lemma 4.10 is proved.

In the proof we use V, V^* to represent V(\tau|X) and V^*(\tau|X). Since \tau|X \neq \emptyset it follows that \mathcal{A}(\tau|X) is nonempty. The first conclusion of Lemma 3.4 now implies that V is nonempty.

Since V, V^* are nonempty it follows that \operatorname{diam}(V \cup V^*) \leq 8.

In the proof we use W, W^*, \mathcal{AC}, \mathcal{AC}^* to represent W(\tau^X) and so on. Since \pi_X(\tau) \neq \emptyset there is some vertex cycle \alpha \in V(\tau) so that \alpha cuts X. Since \alpha is wide (Lemma 2.8) there is a lift \alpha' \prec \tau^X which is also wide; deduce that W is nonempty.

Now isotope \partial X into efficient position. Let X' be the compact component of the preimage of X under the covering map S^X \to S. Note that \partial X' is in efficient position with respect to \tau^X. Note that the covering map S^X \to S induces a homeomorphism between X' and X. Let N^X = N(\tau^X) \subset S^X be the preimage of N = N(\tau). Let N' = X' \cap N^X. Again, the covering map induces a homeomorphism between N' and N \cap X.

Suppose that \partial X, with its induced orientation, is not wide. If \partial X fails to be wide in S {\smallsetminus} N(\tau) then there is a properly embedded, essential arc \gamma \subset X disjoint from N(\tau). Lift \gamma to \gamma' \subset X'. Adjoin to \gamma' geodesic rays in S^X {\smallsetminus} X' to obtain an essential, properly embedded arc \gamma'' \subset S^X. Note that i(\gamma'', \alpha) = 0 for every \alpha \in \mathcal{AC}; only intersections in X' contribute to geometric intersection number as computed in S^X. This implies that \operatorname{diam}_X(\mathcal{C}(\tau|X)) \leq 2. Furthermore, i(\gamma'', \beta) \leq 2 for every \beta \in W^*. This gives the desired diameter bound for W \cup W^*.

If, instead, \partial X fails to be wide in N(\tau) then there is a properly embedded, essential arc \gamma \subset X that is a subarc of a tie. Again, lift and extend to an essential arc \gamma'' \subset S^X so that i(\gamma'', \beta) = 0 for any \beta \in \mathcal{AC}^*. This implies that \operatorname{diam}_X(\mathcal{C}^*(\tau|X)) \leq 2. We also have i(\gamma'', \alpha) \leq 2 for any \alpha \in W. Again the diameter is bounded.

Now suppose that \partial X is wide. Thus, for every b \in \mathcal{B}(\tau), the rectangle R_b meets \partial X in at most a pair of arcs. It follows that N \cap X, and thus N', is a union of at most 2|\mathcal{B}(\tau)| subrectangles of the form R', R'' \subset R_b. Suppose that \alpha \in W and \beta \in W^*. Then \alpha and \beta each meet a subrectangle R' in at most two arcs. Thus \alpha and \beta intersect in at most four points inside of R'. Thus i(\alpha, \beta) \leq 8 |\mathcal{B}(\tau)| and Lemma 5.2 is proved.

Fix an interval [a,b] \subset [0,N]. Note that \tau_b \prec \tau_a and so \tau_b^X \prec \tau_a^X. Thus \mathcal{AC}(\tau_b^X) \subset \mathcal{AC}(\tau_a^X) while \mathcal{AC}^*(\tau_a^X) \subset \mathcal{AC}^*(\tau_b^X).

Now fix i \in I_X.

Thus Theorem 5.3 is proved.

By the first conclusion of Theorem 5.3 we may restrict attention to the subinterval [p,q] = I_X \subset [0,N].

Fix any vertex \alpha \in V(\tau_q|X). Note \alpha is carried by \tau_i|X for all i \leq q. So define \sigma_i \subset \tau_i|X to be the minimal pretrack carrying \alpha. Since \sigma_i does not carry any peripheral curves \sigma_i is a train track. Note that \sigma_i is recurrent by definition and transversely recurrent by Lemma 3.2. Applying Theorem 5.3, for all i \in [p, q-1] the track \sigma_{i+1} is a slide, a split, or identical to the track \sigma_i.

Theorem 5.4 implies the sequence \{ V(\sigma_i) \} is a Q–unparameterized quasi-geodesic in \mathcal{C}(X). Note that d_X(\sigma_i, \tau_i|X) is uniformly bounded because \sigma_i is a subtrack.

Since \alpha \prec \tau_i|X the curve \alpha is also carried by \tau_i. By Lemma 2.9 there is a vertex cycle \beta_i \prec \tau_i that cuts X. Since \beta_i is wide (Lemma 2.8) any element \beta_i' \in \pi_X(\beta) is carried by and wide in \tau_i^X. It follows that \beta_i' and the vertex cycles of \tau_i|X have bounded intersection. Thus d_X(\tau_i, \tau_i|X) is uniformly bounded and we are done.

Let \{ \tau_i \} be the given sliding and splitting sequence in \mathcal{T}(S). Let I = [p,q] \subset [0,N] be a subinterval. Note that d_{\mathcal{T}(S)}(I) \leq |I| because \{ \tau_i \} is an edge-path in \mathcal{T}(S).

Define \mathcal{S}(I) to be the set of indices r \in I where r + 1 \in I and \tau_{r+1} is a splitting of \tau_r. Thus |I| \leq_A |\mathcal{S}(I)|, where A is the constant of Lemma 6.3. Now, Theorem 6.1 implies |\mathcal{S}(I)| \leq_Q d_{\mathcal{M}(S)}(I). Finally, since \tau \mapsto V(\tau) is distance non-increasing we have d_{\mathcal{M}(S)}(I) \leq d_{\mathcal{T}(S)}(I). Deduce that |I| \leq_Q d_{\mathcal{M}(S)}(I), for a somewhat larger value of Q.

Fix an essential Y \subset X with \xi(Y) < \xi(X). It suffices to show, for every subinterval J \subset I, that d_Y(J) \leq {\sf T_1}(X).

If J \cap I_Y = \emptyset then Theorem 5.3 implies d_Y(J) \leq {\sf K_0}.

Suppose that J meets I_Y; thus J_Y = \emptyset by hypothesis and so Y is not inductive. It follows that d_Y(I) < {\sf T_0}(X). By Lemma 6.6 we have d_Y(J) < {\sf T_0}(X) + 2{\sf R_0}.

If X is an annulus then, by Theorem 5.3, for every r \in I the core curve \alpha \subset X is carried by and wide in \tau_r. It follows that the number of switches in \alpha \subset \tau_r|X is bounded by some constant K = K(S). Let q = \max I and pick any \beta \in V(\tau_q|X). As in the proof of Theorem 5.5 let \sigma_r \subset \tau_r|X be the minimal subtrack carrying \beta. Thus \sigma_r has either exactly four branches and two switches, or is an embedded arc. It follows that every K^2/4 consecutive splittings in S_X(I) induces at least one splitting in the sequence of tracks \{ \sigma_r \}. Therefore the singleton sets V(\sigma_r) form a quasi-geodesic in \mathcal{A}(X). Since V(\sigma_r) \subset V(\tau_r|X) the proof is complete when X is an annulus.

We assume for the rest of the proof that X is not an annulus. The map i \mapsto V(\tau_i|X), taking tracks to their vertex cycles, is generally not injective. (For example, see [PennerHarer92, Proposition 2.2.2].) However:

Notice that if V(\tau_{i+1}|X) \neq V(\tau_i|X) then V(\tau_{i+1}|X) \neq V(\tau_j|X) for j \leq i. This is because P(\tau_{k+1}|X) \subset P(\tau_k|X) for all k. Using C = 1 + \max \{ {\sf C}(X), {\sf T_1}(X) \} as the cut-off in Theorem 6.7 gives some constant of quasi-equality, say {\sf E}. Define {\sf R_1} = {\sf E} + 1.

Suppose that [p,q] = I, the straight subinterval of J_X given by Lemma 6.13. We define a function \rho \colon [0,M] \to I as follows: let \rho(0) = p and let \rho(n+1) be the smallest element in [\rho(n), q] with d_{\mathcal{M}(X)}(\tau_{\rho(n)}, \tau_{\rho(n+1)}) = {\sf R_1}. (If \rho(n+1) is undefined then take M = n+1 and \rho(M) = q.) Let B(\mu) be the ball of radius {\sf R_1} about the marking \mu \in \mathcal{M}(X). Define

Deduce from Claim ? and the remark immediately following that, for all n \in [0, M - 1],

So to prove Lemma 6.13 it suffices to bound M from above in terms of d_X(I).

Thus we have

where the first and second lines follow from Lemma 6.6 and Claim ? respectively. This completes the proof of Lemma 6.13.

Let [p,q] = I. Applying Theorem 5.3, as p \in J_Y \subset I_Y, the multicurve \partial Y is wide with respect to \tau_p. It follows that \partial Y is also wide with respect to \tau_p^X. Note that the curves of V(\tau_p|X) are also wide with respect to \tau_p^X. Repeating this discussion for q, and then applying Lemma 5.2 and the triangle inequality gives a uniform bound for d_X(\tau_p, \tau_q). The lemma now follows from Lemma 6.13.

From Lemma 6.13 we deduce

where the first term on the right arises from addition of additive errors.

By the definition of a long straight interval and from Lemma 6.6 deduce

Thus 2{\sf R_0}|{\sf Long}| \leq d_X(J_X). These inequalities combine to prove the lemma, for a somewhat larger value of A = A(X).

By Lemma 6.13 the number of splittings in any short straight interval is a priori bounded (depending only on X). Since |{\sf Short}| \leq |{\sf Ind}'| the lemma follows.

Fix Y \in {\sf Ind}. Consider an index r \in J_Y that is assigned to X. Let I \subset J_Y be the maximal interval containing r so that all indices in \mathcal{S}_X(I) are assigned to X. We now show that I is straight. Let Z be any essential subsurface of X with \xi(Z) < \xi(X) and let [r,s] = J \subset I be any subinterval. If J \cap I_Z = \emptyset then Theorem 5.3 implies that d_Z(J) \leq {\sf K_0}. If J meets I_Z then, as no splittings of J are assigned to Z we deduce that \tau_s|Z is obtained from \tau_r|Z by sliding and taking subtracks only. Thus d_Z(J) \leq {\sf B_0} {\sf N_2} \leq {\sf T_1}(X), as desired.

By Lemma 6.16 we find that |\mathcal{S}_X(I)| is bounded. It follows that the number of splittings in the inductive intervals is quasi-bounded by \sum_{Y \in {\sf Ind}} |\mathcal{S}_Y(J_Y)|.

By induction, Proposition 6.9 gives

Taking a cutoff of C = 1 + \max \{ {\sf C}(Y), {\sf T_0}(X) + 2{\sf R_0} \} and applying the distance estimate Theorem 6.7 we have a quasi-inequality

Since d_Z(J_Y) \leq d_Z(J_X) + 2{\sf R_0} for all Z \subset Y, it follows that non-zero terms in the sum only arise for subsurfaces in {\sf Ind}'(Y) = \{ Z \in {\sf Ind}' \mathbin{\mid} Z \subset Y \}. Since 2{\sf R_0} \leq {\sf T_0}(X)

we have [d_Z(J_Y)]_C \leq 2\cdot d_Z(J_X). Making A = A(X) larger if necessary we have

where the final quasi-inequality follows from Lemma 6.20, taking A larger as necessary. Note that the term |{\sf Ind}| on the middle line arises by adding additive errors. This proves Lemma 6.22.