Arxiv:1905.00915V3 [Math.GT] 6 Mar 2021 1

LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION

YUSHENG LUO

Abstract. In this paper, we show that one can naturally associate a limiting dynamical system F : T −→ T on an R-tree to any degenerating sequence of rational maps fn : Cˆ −→ Cˆ of ﬁxed degree. The construction 3 of F is in 2 steps: ﬁrst, we use barycentric extension to get E fn : H −→ 3 H ; second, we take appropriate limit on rescalings of hyperbolic space. An important ingredient we prove here is that the Lipschitz constant depends only on the degree of the rational map. We show that the dynamics of F records the limiting length spectra of the sequence fn.

Contents 1. Introduction 1 2. The barycentric extension 6 3. Extremal length and width 8 4. The uniform bound for barycentric extensions 10 3 5. Limits of barycentric extensions on H 13 6. The Carath´eodory convergence of annuli 16 7. R-trees and branched coverings 20 3 8. Limits of barycentric extensions on rescalings of H 21 9. Length spectra and translation lengths 32 Appendix A. The barycentric extension of z2 33 Appendix B. The uniform bound for barycentric extensions in higher dimensions 36 References 39 arXiv:1905.00915v3 [math.GT] 6 Mar 2021 1. Introduction A rich theory is developed in the study of the dynamics of rational maps f : Cˆ −→ Cˆ with d = deg(f) ≥ 2. The space Ratd(C) of all rational maps of degree d is not compact. Hence it is useful to construct limiting dynamical systems as fn → ∞ in Ratd(C). In this paper, we show that to any sequence fn → ∞, one can naturally associate a degree d branched covering on an R-tree F : T −→ T . The map F is constructed in two steps.

Date: March 9, 2021. 1 2 YUSHENG LUO

(1) By applying the barycentric extension, we extend each rational map 3 3 fn : Cˆ −→ Cˆ to a map E fn : H −→ H (see §2, or [DE86, Pet11]). 3 (2) Using the standard idea that H takes on the appearance of an R-tree under rescaling (cf. [Gro81]), we construct the branched covering by taking limits of barycentric extensions E fn with appropriate rescalings. The following equicontinuity theorem plays a crucial role.

Theorem 1.1. For any rational map f : Cˆ −→ Cˆ of degree d, the norm of 3 3 the derivative of its barycentric extension E f : H −→ H satisﬁes sup kDE f(x)k ≤ C deg(f). x∈H3 Here the norm is computed with respect to the hyperbolic metric, and C is a universal constant.

Remark. The theorem in particular says that the Lipschitz constant of E f 3 is C deg(f), i.e., ∀x, y ∈ H ,

dH3 (E f(x), E f(y)) ≤ C deg(f) dH3 (x, y). In dimension 1, if f : S1 −→ S1 is the restriction of a Blaschke product, 2 then the barycentric extension E f : ∆ −→ H satisﬁes E f = f. Hence, the Schwarz lemma implies that E f is a distance non-increasing map with respect to hyperbolic metric, i.e., uniformly 1-Lipschitz. Theorem 1.1 is a replacement of this geometric bound in higher dimensions. 27 Our proof shows that one can take the constant C = 2 log 3 ≈ 12.3. On the other hand, the barycentric extension of zd has Lipschitz constant d on the geodesic connecting 0, ∞ ∈ S2. Hence C ≥ 1. We conjecture that C = 1 gives the sharp bound for the inequality. There is a more general statement for quasiregular maps and in higher dimensions. See Theorem B.1 in Appendix B for details.

Geometric limits as branched coverings on an R-tree. Back to the study of fn → ∞, Theorem 1.1 shows E fn is equicontinuous. However, 3 (1) E fn is not uniformly bounded in general: the image of 0 ∈ H may escape to infinity (see Proposition 5.5); (2) even if E fn is uniformly bounded, the degree will definitely drop: some preimages of 0 will escape to infinity (see Proposition 8.1). 3 It is thus natural to consider limits of rescalings of H . Given a sequence of rescalings rn → ∞, we will construct the asymptotic r 3 0 3 cone ( H , x , d) for the sequence of pointed metric spaces (H , 0, dH3 /rn). This construction uses a non-principal ultrafilter ω which will be fixed once for all. The ultrafilter ω allows any sequences in compact spaces to converge automatically without the need to pass to a subsequence. We will denote the limit with respect to the ultrafilter by limω(see §8.2). LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 3

r 3 0 It is well-known that the asymptotic cone ( H , x , d) is an R-tree [Roe03]. r 3 This means that any two points x, y ∈ H can be joined by a unique arc r 3 [x, y] ⊆ H which is isometric to an interval of R. The appropriate scale we consider here is

rn := max d 3 (0, y), −1 H y∈E fn (0) as it brings all the preimages of 0 in view. We show that

Theorem 1.2. Let fn → ∞ in Ratd(C). Let

rn := max d 3 (0, y), −1 H y∈E fn (0) r 3 3 and H be the asymptotic cone of H with rescaling rn. Then we have a limiting map r 3 r 3 F = lim E fn : −→ , ω H H r 3 and it is a degree d branched covering of the R-tree H .

Remark. Let sn → ∞ be a rescaling with dH3 (0, E fn(0)) ≤ Ksn for some 0 s 3 s 3 K > 0. Then our construction gives a limiting map F : H −→ H (see Proposition 8.3).

• If limω sn/rn = M ∈ (0, ∞), the two limiting maps are conjugate to each other. More precisely, there exists a homeomorphism Φ : r 3 s 3 0 H → H , which is a dilation by M, conjugating F and F . • If limω sn/rn = 0, the degree of the limiting map will drop. 0 r 3 • If limω sn/rn = ∞, then x ∈ H is totally invariant, resulting in a less interesting dynamical system.

We have a version for sequences [fn] → ∞ in the moduli space of rational maps Md = Ratd(C)/ PSL2(C) (see Theorem 8.11). We also show how to extract standard geometric limits of E fn from the r 3 r 3 ultralimit F : H −→ H (see Theorem 8.12). Although the barycentric extension does not behave well under composi- tion, we show that the limiting map is dynamically natural.

Theorem 1.3 (Dynamical naturality). Let fn → ∞ in Ratd(C), with lim- r 3 r 3 iting map F : H −→ H . Then for any N ∈ N, F N = lim E (f N ). ω n Length spectra and translation lengths. A ray α in the R-tree T is a subtree isometric to [0, ∞) ⊆ R. Two rays α1, α2 are equivalent, denoted by α1 ∼ α2, if α1 ∩α2 is still a ray. The collection (T ) of all equivalence classes of rays forms the set of ends of T . We will use α to denote both a ray and the end it represents. We say a sequence of points xi converges to an end α, denoted by xi → α, if for any ray β ∼ α, xi ∈ β for all suﬃciently large i. 4 YUSHENG LUO

We deﬁne the translation length of an end α by 0 0 L(α, F ) = lim d(xi, x ) − d(F (xi), x ) xi→α for any xi ∈ T with xi → α. We show that the translation length is well- r 3 r 3 deﬁned for the limiting map F : H −→ H (see §9). If C = {α1, ..., αq} is a cycle of periodic ends, then the translation length of the periodic cycle C is q X L(C,F ) = L(αi,F ). i=1

Let f ∈ Ratd(C), and C = {z1, ..., zq} ⊆ Cˆ be a periodic cycle of f of period q. We deﬁne the length of the periodic cycle C by q q 0 X 0 L(C, f) := log |(f ) (z1)| = log |f (zi)|. i=1 r 3 For the asymptotic cone H , any sequence (zn) ∈ Cˆ determines an end r 3 α ∈ ( H ) (see §8.2). We denote it by zn ω α. r 3 More generally, if Cn = {z1,n, ..., zq,n} ⊆ Cˆ and C = {α1, ..., αq} ⊆ ( H ) with zi,n ω αi, then we denote it by Cn ω C. We show that we can recover the limiting length spectra from the translation lengths for the limiting map F . r 3 r 3 Theorem 1.4. Let fn → ∞ in Ratd(C), with limiting map F : H −→ H . Let Cn ⊆ Cˆ be a sequence of periodic cycles of fn of period q, with Cn ω C. Then C is periodic with period q, and

L(Cn, fn) L(C,F ) = lim . ω rn Supplement: the barycentric extension of z2. We conclude our discussion by studying the barycentric extension of the map f(z) = z2. The exact formula for the extension is hard to compute, even in this simple case. 3 Let (r, θ, h) be the cylindrical coordinates for the hyperbolic 3-space H with respect to the hyperbolic metric. In Appendix A, we show that: Theorem A.1. Let f(z) = z2. In the cylindrical coordinate, there exists a real analytic function δ : [0, ∞) −→ R with • δ(0) = 0, • δ(r) > 0 when r > 0, • limr→∞ δ(r) = 0, such that the barycentric extension E f is given by E f(r, θ, h) = (log(cosh(r)) − δ(r), 2θ, 2h). Since in (r, θ) coordinate, z2 has the form (r, θ) 7→ (log(cosh(r)), 2θ), LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 5

Theorem A.1 implies the restriction of E f on the equatorial hyperbolic plane 2 2 H0 deﬁned by h = 0 is not z . This gives a counterexample to a conjecture of Petersen [Pet11]. 3 3 Since Isom(H ) × Isom(H ) acts transitively on Rat2(C), by naturality, E f| 2 =6 f for any degree 2 Blaschke product f. It seems unlikely that the H0 restriction E f| 2 of a Blaschke product f would ever agree with f. H0

Notes and discussion. An important ingredient of Thurston’s theory on surface automorphisms is his construction of an equivariant compactification of the Teichmüllerspace [Thu88, FLP79]. Isometric actions on R-trees provide such compactifications for various spaces of geometric structures. In [MS84], Morgan and Shalen showed how to assign an isometric action on a R- tree to a degenerating sequence of representations. This shed new light and generalized parts of Thurston’s Geometrization Theorem for 3-manifolds. Bestvina and Paulin gave a new geometric perspective on this theory in [Bes88, Pau88]. R-trees were also used by Otal to give a proof of Thurston’s Double Limit Theorem and the Hyperbolization Theorem for 3-manifolds that fiber over the circle [Ota96]. The use of asymptotic cone and its connection with Gromov-Hausdorff limit are explained in [Ota96, KL95, Chi91]. Various constructions of R-trees also appear in complex dynamics. In [McM09], McMullen constructed a branched covering on a Ribbon R-tree as geometric limits of a divergent sequence of Blaschke products. Using Berkovich spaces, R-trees also appear naturally in the study of degenerating families of rational maps [Kiw15, BR10]. Other application of trees in complex dynamics can be found in [Shi89, DM08]. The barycentric extension method has been used in various geometric settings. In [BCG95, BCG96], Besson, Coutois and Gallot used the explicit bound on the Jacobian of the barycentric extension of the Patterson-Sullivan measure to prove rigidity results on negatively curved Riemannian manifold. Other applications can be found in [EMS02, Sto07]. In comparison with the previous work, the geometric limits constructed in Theorem 1.2 and Theorem 8.11 generalize the Ribbon R-tree construction in [McM09] and give analogous compactifications as in [MS84, Bes88, Pau88]. The limiting ratios of lengths of marked geodesics for a degenerating sequence of Riemann surfaces or Kleinian groups are naturally recorded in the limiting action on the R-tree [MS84, Ota96]. Theorem 1.4 gives the analogous result in complex dynamics.

Upcoming sequel. In the sequel [Luo19], we will compare our construction of branched coverings of the R-tree using the barycentric extension to the Berkovich dynamics of the complexiﬁed Robinson’s ﬁeld. We will show that the two approaches are equivalent. We will also apply the relation in Theorem 1.4 to classify those hyperbolic components admitting a degenerating sequence where all the multipliers stay bounded. The connections with the theory of Kleinian groups will be explored further there. 6 YUSHENG LUO

The structure of the paper. In §2 and §3, we review the construction of the barycentric extension and the theory of extremal lengths. Using the classical length-area principle, the proof the Theorem 1.1 is given in §4. Theorem 1.1 allows us to construct various limits for barycentric extensions of a sequence of rational maps. We study the limits of barycentric 3 3 extensions on H in §5 and the limits on rescalings of H in §8. Some back- grounds on convergence of annuli and R-trees are provided in §6 and §7. We study the translation lengths and length spectra in §9. Finally, the barycentric extension of z2 is studied in Appendix A, and a higher dimensional analogue of Theorem 1.1 is proved in Appendix B.

Acknowledgments. The author thanks C. T. McMullen for his advice and helpful discussion on this problem. The author gratefully thanks the anonymous referee for valuable comments and suggestions.

2. The barycentric extension In this section, we will brieﬂy review the theory of barycentric extensions. The theory is extensively studied for circle homeomorphisms in [DE86]. The construction can be easily generalized to any continuous maps on the unit sphere Sn (see [McM96, Pet11]). We will then use the implicit diﬀerentiation to compute its derivatives.

Deﬁnition of barycentric extensions. Throughout the paper, we iden- n+1 ∼ n+1 n+1 tify H = B(0, 1) ⊆ R as the unit ball centered at the origin 0 ∈ R . n n+1 Let µ be a probability measure on the unit sphere S ⊆ R . We say µ is balanced if Z n+1 1 ζdµ(ζ) = ~0 ∈ R . Sn n+1 We say µ is balanced at a point y ∈ H if the push forward measure M∗µ n+1 is balanced, where M ∈ Isom H is an isometry sending y to 0. Since the balanced condition is invariant under rotation, it follows that this deﬁnition does not depend on the choice of M. If µ is a probability measure on Sn with no atoms, then there is a unique n+1 point β(µ) ∈ H , called the barycenter of µ, for which the measure is balanced at β(µ) (see [DE86, Hub06] or [Pet11]). Let µSn be the probability measure coming from the spherical metric on n S . We say a map f is admissible if f∗µSn has no atoms. n n n+1 Let f : S −→ S be an admissible continuous map and x ∈ H . Let n+1 M ∈ Isom H be an isometry sending 0 to x. The barycentric extension is the real analytic map n+1 n+1 E f : H −→ H

1 n+1 This integral has a natural interpretation of a vector in the tangent space T0H (see [DE86] for details), thus we use the usual vector notation ~0 instead of 0. LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 7 that sends x to the barycenter of the measure (f ◦ M)∗µSn . Since the measure µSn is invariant under rotation, it follows that this deﬁnition does not depend on the choice of M. One of the desired properties of the barycentric extension is that it is n+1 conformally natural. This means that if M1,M2 ∈ Isom H , then

M1 ◦ E f ◦ M2 = E (M1 ◦ f ◦ M2), n+1 n where we have identified Isom H with the group Conf S of conformal n n+1 maps on S = ∂H . n+1 ∼ Implicit formula for barycentric extensions. Given a point x ∈ H = n+1 n+1 n+1 B(0, 1) ⊆ R , we define Mx : H −→ H by z(1 − kxk2) + x(1 + kzk2 + 2hx, zi) Mx(z) = . 1 + kxk2kzk2 + 2hx, zi n+1 −1 Then Mx is an isometry on H sending 0 to x, with Mx = M−x. n n Note that Mx extends to a map Mx : S −→ S . An easy computation shows that the Jacobian at ζ ∈ Sn is 2 1 − kxk n JacMx(ζ) = ( ) . kζ + xk2 n+1 n+1 n+1 Let F : H × H −→ R be the function Z −1 F (x, y) = My (f(Mx(ζ)))dµSn (ζ) Sn Z = M−y(f(ζ))d(Mx)∗µSn (ζ) Sn Z 2 1 − kxk n n = M−y(f(ζ))( 2 ) dµS (ζ). Sn kζ − xk With the above notations, the barycentric extension of f is the unique solution of the implicit formula: n+1 F (x, E f(x)) = ~0 ∈ R . Therefore, implicit differentiation gives: −1 DE f(x) = −Fy (x, E f(x))Fx(x, E f(x)). n+1 n+1 n+1 Since Isom H × Isom H acts transitively on pairs of points in H , by naturality of the barycentric extension, it suffices to compute the derivative at the origin 0 under the assumption E f(0) = 0, i.e. Z f(ζ)dµSn (ζ) = ~0. Sn For this, we have n+1 Proposition 2.1. Assume that E f(0) = 0. Then for all ~v ∈ R , Z Fy(0, 0)(~v) = −2~v + 2 h~v, f(ζ)if(ζ)dµSn (ζ), Sn 8 YUSHENG LUO and Z Fx(0, 0)(~v) = 2n h~v, ζif(ζ)dµSn (ζ). Sn Proof. For the first equality, we have 1 Fy(0, 0)(~v) = lim · F (0, t~v) t→0 t Z 1 = lim · M−t~v(f(ζ))dµSn (ζ) t→0 Sn t Z 1 f(ζ)(1 − t2k~vk2) − t~v(1 + kf(ζ)k2 − 2th~v, f(ζ)i) n = lim · 2 2 2 dµS (ζ) t→0 Sn t 1 − 2th~v, f(ζ)i + t k~vk kf(ζ)k Z 1 2 2 = lim · (f(ζ) − 2t~v + O(t ))(1 + 2th~v, f(ζ)i + O(t ))dµSn (ζ) t→0 Sn t Z = −2~v + 2 h~v, f(ζ)if(ζ)dµSn (ζ). Sn Similarly, for the second equality, we have 1 Fx(0, 0)(~v) = lim · F (t~v, 0) t→0 t Z 2 1 1 − kt~vk n n = lim · f(ζ)( 2 ) dµS (ζ) t→0 Sn t kζ − t~vk Z 1 2 2 n 2 2 −n = lim · f(ζ)(1 − t k~vk ) (1 − 2th~v, ζi + t k~vk ) dµSn (ζ) t→0 Sn t Z 1 2 2 = lim · f(ζ)(1 + O(t ))(1 + 2nth~v, ζi + O(t ))dµSn (ζ) t→0 Sn t Z = 2n h~v, ζif(ζ)dµSn (ζ). Sn

3. Extremal length and width There is a wealth of sources containing background material on extremal length (see [Ahl73] or [KL09, Appendix 4]). We will brieﬂy summarize the necessary minimum.

Extremal length and extremal width. Let Γ be a family of (multi)curves2 on U ⊆ C. Given a (measurable) conformal metric ρ = ρ(z)|dz| on U, let L(Γ, ρ) := inf L(γ, ρ), γ∈Γ where L(γ, ρ) stands for the ρ-length of γ. The extremal length of Γ is deﬁned as L(Γ, ρ)2 L(Γ) := sup , ρ A(U, ρ)

2An element γ ∈ Γ is a ﬁnite union of arcs or curves. LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 9 where A(U, ρ) is the area of U with respect to the measure ρ2, and the supremum is taken over all ρ subject to the condition 0 < A(U, ρ) < ∞. L(Γ,ρ)2 We call a conformal metric ρ an extremal metric if L(Γ) = A(U,ρ) . The extremal width of Γ is deﬁned as the inverse of extremal length: 1 W(Γ) = . L(Γ)

Extremal distance and separating curves. To avoid delicate discussions of the boundary behavior, we deﬁne an admissible triple (U, E1,E2) by following conditions3: (1) U ⊆ C is a bounded open set whose boundary consists of a ﬁnite number of Jordan curves; (2) Each of E1 and E2 is a union of boundary components of U with ∂U = E1 ∪ E2 and E1 ∩ E2 = ∅.

Let Γ be the set of connected arcs in U which join E1 and E2. In other words, each γ ∈ Γ shall have one end point in E1 and one in E2, and all other points shall be in U. The extremal distance modU (E1,E2) is deﬁned by

modU (E1,E2) = L(Γ). The extremal distance and the corresponding extremal metric can be computed explicitly using the solution for the Dirichlet problem as follows (see [Ahl73, §4.9]). Let u be the unique harmonic function in U equal to 1 on E2 and vanishing on E1. Then 2 2 1/2 • ρ0 = |∇u| = (ux + uy) is an extremal metric; 1 • modU (E1,E2) = D(u) , where ZZ D(u) = |∇u|2dxdy U is the Dirichlet energy of the function u. Let σ be a ﬁnite union of closed curves in U. We say σ is separating in (U, E1,E2) if there are two (relative) open sets V1,V2 of U with E1 ⊆ V1, E2 ⊆ V2 and U − σ = V1 ∪ V2. Let Σsep be the set of all separating (multi)curves. It follows from the discussion in [Ahl73, §4.9] that L(Σsep) = D(u). Therefore, we have:

Theorem 3.1. Let (U, E1,E2) be an admissible triple. Then

sep 1 modU (E1,E2) = W(Σ ) = . L(Σsep)

3See Appendix B for more general settings. 10 YUSHENG LUO

Transformation rules. Both extremal length and extremal width are conformal invariants. They also behave nicely under holomorphic branched coverings by the following theorem (see [KL09, Lemma 4.5]). 0 0 0 Theorem 3.2. Let (U, E1,E2) and (U ,E1,E2) be two admissible triples. Let f : U −→ U 0 be a holomorphic branched covering of degree d with 0 0 f(E1) = E1 and f(E2) = E2. Then 0 0 modU 0 (E1,E2) ≤ d · modU (E1,E2).

4. The uniform bound for barycentric extensions In this section, we will prove Theorem 1.1. Let f : Cˆ −→ Cˆ be a rational map of degree d. Recall that Z −1 F (x, y) = My (f(Mx(ζ)))dµS2 (ζ). S2 The key is to establish the following estimate.

Proposition 4.1. Let f : Cˆ −→ Cˆ be a rational map of degree d with E f(0) = 0. Then −1 kFy(0, 0) k ≤ Md, for some universal constant M. Once the above estimate is established, Theorem 1.1 follows immediately. Proof of Theorem 1.1 assuming Proposition 4.1. If E f(0) = 0, by Proposi- tion 2.1, Z Fx(0, 0)(~v) = 4 h~v, ζif(ζ)dµS2 (ζ). S2 2 Since ζ, f(ζ) ∈ S , kf(ζ)k = 1 and h~v, ζi ≤ k~vk. Hence kFx(0, 0)k ≤ 4. −1 By Proposition 4.1, kFy(0, 0) k ≤ Md. Thus −1 kDE f(0)k ≤ kFy(0, 0) kkFx(0, 0)k ≤ 4Md. 3 3 3 Since Isom H × Isom H acts transitively on pairs of points in H , by 4 naturality of the barycentric extension, the theorem is proved. To prove Proposition 4.1, we perform two further reductions. First, by Proposition 2.1, Z Fy(0, 0)(~v) = −2~v + 2 h~v, f(ζ)if(ζ)dµS2 (ζ). S2 Let Z T : ~v 7→ h~v, f(ζ)if(ζ)dµS2 (ζ). S2

4 27 From the proof of Proposition 4.1, we can take M = 8 log 3 . Thus, we can take the 27 constant C for Theorem 1.1 as 2 log 3 . LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 11

Since T is a self-adjoint operator with spectral radius ≤ 1, to prove Propo- sition 4.1, it suﬃces to bound the largest eigenvalue from 1. After a change 0 3 of variable, we assume the largest eigenvalue is associated to ~e3 = 0 ∈ R . 1 Second, we will relate the eigenvalue h~e3,T (~e3)i to the spherical area of certain sets. Let √ 0 1 U = B(0, 3) − B(0, √ ) ⊆ C 3 ∼ 2 be an annulus. Under the identiﬁcation Cˆ = S by the stereographic projection, U 0 corresponds to the ‘belt’ x 2 1 1 2 {y ∈ S : − < z < } ⊆ S . z 2 2 √ B0 B , √1 B0 ˆ − B , U Let 1 = (0 3 ) and 2 = C (0 3) be the two ‘caps’. Let = −1 0 −1 0 −1 0 f (U ), B1 = f (B1) and B2 = f (B2) be the corresponding preimages. Let V := µS2 (U), V1 := µS2 (B1) and V2 := µS2 (B2) be the normalized spherical area of the corresponding sets. Note that

( 1 2 4 if ζ ∈ U h~e3, h~e3, f(ζ)if(ζ)i = (h~e3, f(ζ)i) ≤ . 1 if ζ ∈ B1 ∪ B2

Therefore, Z h~e3,T (~e3)i = h~e3, h~e3, f(ζ)if(ζ)dµS2 (ζ)i S2 Z Z = h~e3, h~e3, f(ζ)if(ζ)idµS2 (ζ) + h~e3, h~e3, f(ζ)if(ζ)idµS2 (ζ) U B1∪B2 1 3 ≤ · V + 1 · (1 − V ) = 1 − V. 4 4 We show we can bound V from below. 16 log 3 Lemma 4.2. With the above notations, V ≥ 81d . The proof uses the classical length-area principle developed by Gr¨otzsch and Ahlfors. If V is very small, then V1 and V2 are both big by the condition R ~ S2 f(ζ)dµS2 (ζ) = 0. Thus, any separating curve must bound a deﬁnite amount of spherical area. So the isoperimetric inequality gives a lower bound on its spherical length. This gives a lower bound of V in terms of the extremal width of separating curves. Using Theorem 3.1 and Theorem 3.2, we can bound the extremal width of separating curves from below. This gives the desired lower bound for V . 12 YUSHENG LUO

1 Proof of Lemma 4.2. If V ≥ 3 , then the lemma holds. Thus, we may assume 1 R ~ V < 3 . Since E f(0) = 0, we have S2 f(ζ)dµS2 (ζ) = 0. Therefore, Z 0 = h~e3, f(ζ)dµS2 (ζ)i S2 Z Z Z = h~e3, f(ζ)idµS2 (ζ) + h~e3, f(ζ)idµS2 (ζ) + h~e3, f(ζ)idµS2 (ζ) B1 U B2 1 1 1 1 ≤ (− ) · V1 + · V + 1 · V2 = (− )(1 − V2 − V ) + V + V2 2 2 2 2 3 1 3 1 = V2 + V − ≤ V2 − . 2 2 2 6 V ≥ 1 V ≥ 1 Hence 2 9 . Similarly, 1 9 . √ E0 ∂B , √1 E0 ∂B , E f −1 E0 E Let 1 = (0 3 ) and 2 = (0 3). Let 1 = ( 1) and 2 = −1 0 0 0 0 f (E2). Then (U, E1,E2) and (U ,E1,E2) are admissible triples.

B1 U

B1 B2 U

B2 B2 B2

Figure 4.1. A schematic conﬁguration of U, B1,B2. An example of a separating (multi)curve is colored in red.

sep Let σ ∈ Σ be a separating curve for (U, E1,E2), and let Ω be the set bounded by σ with B1 ⊆ Ω and B2 ∩ Ω = ∅. Let ρS2 be the standard 2 1 spherical metric on S . Since V1 = µS2 (B1),V2 = µS2 (B2) ≥ 9 , the spherical area of Ω is bounded between 1 8 · 4π ≤ A(Ω, ρ 2 ) ≤ · 4π. 9 S 9 √ 4π 8 2π Since a spherical disk with area 9 has perimeter 9 , by the isoperimetric inequality for the sphere (see [L´ev22,Troisi`emePartie, Chapitre I] or [Sch39]), √ 8 2π L(σ, ρ 2 ) ≥ . S 9 LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 13

Since this holds for any separating curve, we have √ sep 2 L(Σ , ρ 2 ) 8 2π 32π L(Σsep) ≥ S ≥ ( )2/(4π · V ) = . A(U, ρS2 ) 9 81V By Theorem 3.1, we have 1 81V modU (E1,E2) = ≤ . (4.1) L(Σsep) 32π 0 0 0 log 3 Since U is an annulus, the extremal distance modU 0 (E1,E2) = 2π equals to the modulus of U 0. Since f : U −→ U 0 is a degree d branched covering 0 0 with f(E1) = E1 and f(E2) = E2, by Theorem 3.2,

1 0 0 log 3 modU (E1,E2) ≥ · modU 0 (E ,E ) = . (4.2) d 1 2 2πd Combining the inequalities 4.1 and 4.2, we get 16 log 3 V ≥ . 81d We are now ready to prove Proposition 4.1. Proof of Proposition 4.1. Since T is self-adjoint, all of its eigenvalues are 3 4 log 3 real. By Lemma 4.2, the largest eigenvalue of T is at most 1− 4 V ≤ 1− 27d . Since Fy(0, 0) = 2T − 2I, the smallest absolute value of its eigenvalues is 8 log 3 −1 27 at least 27d . Thus, kFy(0, 0) k ≤ 8 log 3 · d.

3 5. Limits of barycentric extensions on H

In this section, we will study the limit of barycentric extensions E fn on 3 H for a sequence of rational maps fn ∈ Ratd(C). We will give a condition 3 3 for E fn to converge compactly to a map F : H −→ H (see Theorem 5.6).

The space of rational maps. The space Ratd(C) of rational maps of d ≥ 2d+1 degree 1 is an open variety of PC . More concretely, fixing a coordinate ˆ ∼ 1 system on C = PC, then a rational map can be expressed as a ratio of homogeneous polynomials f(z : w) = (P (z, w): Q(z, w)), where P and Q have degree d with no common divisors. Using the coefficients of P and Q as parameters, then 2d+1 − V , Ratd(C) = PC (Res) where V (Res) is the vanishing locus for the resultant of P and Q. 2d+1 One natural compactification of Ratd(C) is Ratd(C) = PC . Every map f ∈ Ratd(C) determines the coefficients of a pair of homogeneous polynomials. We write f = (P : Q) = (Hp : Hq) = Hϕf , 14 YUSHENG LUO where H = gcd(P,Q) and ϕf = (p : q) is a rational map of degree at most d. A zero of H is called a hole of f and the set of zeros of H is denoted by H(f). We define the degree of f ∈ Ratd(C) as the deg of ϕf . The following lemma is well-known (see [DeM05, Lemma 4.2]):

Lemma 5.1. Let fn ∈ Ratd(C) converge to f ∈ Ratd(C). Then fn converges compactly to ϕf on Cˆ − H(f). The next lemma follows from the proof of [DeM05, Lemma 4.5 and 4.6]:

Lemma 5.2. Let fn ∈ Ratd(C) converge to f ∈ Ratd(C) and a ∈ H(f). Let U be a neighborhood of a. • If deg(ϕf ) ≥ 1, then there exists K > 1 so that fn(U) = Cˆ for all n ≥ K. Moreover, there exists a sequence of critical points cn of fn converging to a. • If ϕf ≡ C is a constant map, then for any compact set M ⊆ Cˆ −{C}, there exists K > 1 so that M ⊆ fn(U) for all n ≥ K. We also need the following lemma (see [DeM07, Lemma 2.6]).

Lemma 5.3. Let fn ∈ Ratd1 (C) and gn ∈ Ratd2 (C) with fn → f ∈ Ratd1 (C) and gn → g ∈ Ratd2 (C). If ϕf is not a constant map taking value in H(g), ˆ −1 then gn ◦ fn converges compactly to ϕg ◦ ϕf on C − H(f) − ϕf (H(g)). −1 Since H(f) and ϕf (H(g)) are both ﬁnite sets if ϕf is not a constant function taking value in H(g), the above lemma immediately implies that

Lemma 5.4. Let fn ∈ Ratd1 (C) and gn ∈ Ratd2 (C) with fn → f ∈ Ratd1 (C) and gn → g ∈ Ratd2 (C). If ϕf is not a constant map taking value in H(g), then (gn ◦ fn)∗µS2 converges weakly to (ϕg ◦ ϕf )∗µS2 . 3 The limit of E fn on H . Using Lemma 5.4, we have: Proposition 5.5. Let fn ∈ Ratd(C). Then E fn(0) is uniformly bounded if and only if degree 0 maps are not in the limit set of {fn} in Ratd(C). Proof. If a degree 0 map is in the limit set, then after passing to a subsequence, we assume fn → f ∈ Ratd(C) of degree 0. Suppose for contradiction that E fn(0) stays bounded. Let Mn ∈ PSL2(C) with E fn(0) = Mn(0). Then after passing to a further subsequence, Mn → M ∈ PSL2(C). By naturality of the barycentric extension, −1 −1 E (Mn ◦ fn)(0) = Mn ◦ E fn(0) = 0. −1 Hence, the measure (Mn ◦ fn)∗µS2 is balanced for all n. −1 By Lemma 5.4, (Mn ◦ fn)∗µS2 converges weakly to the delta measure −1 (M ◦ ϕf )∗µS2 , giving a contradiction. Conversely, assume that E fn(0) is unbounded. Let Mn ∈ PSL2(C) with E fn(0) = Mn(0). Then after passing to a subsequence, Mn → M ∈ Rat1(C) of degree 0. By naturality of the barycentric extension, −1 −1 E (Mn ◦ fn)(0) = Mn ◦ E fn(0) = 0. LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 15

−1 Hence, the measure (Mn ◦ fn)∗µS2 is balanced for all n. Suppose for contradiction that degree 0 maps are not in the limit set of {fn}. Then after passing to a further subsequence, fn → f ∈ Ratd(C) of −1 degree ≥ 1. By Lemma 5.4, (Mn ◦ fn)∗µS2 converges to the delta measure (ϕM −1 ◦ ϕf )∗µS2 , giving a contradiction. Using Theorem 1.1, we show:

Theorem 5.6. Let fn ∈ Ratd(C) converge to f ∈ Ratd(C) with deg(ϕf ) ≥ 3 3 1. Then E fn converges compactly to E ϕf : H −→ H .

Proof. We ﬁrst claim that E fn converges to E ϕf pointwise. By naturality of the barycentric extension, it suﬃces to show that E fn(0) converges to 0 under the assumption that E ϕf (0) = 0. Let Mn ∈ PSL2(C) with Mn(0) = E fn(0). By Proposition 5.5, Mn is bounded in PSL2(C). Suppose for contradiction that Mn(0) does not converge to 0. Then after passing to a subsequence, we assume that Mn → M ∈ PSL2(C) with M(0) =6 0. −1 −1 By Lemma 5.4, (Mn ◦ fn)∗µS2 converges weakly to (M ◦ ϕf )∗µS2 . −1 −1 Since the measure (Mn ◦ fn)∗µS2 is balanced for any n,(M ◦ ϕf )∗µS2 is also balanced. Hence by naturality,

E ϕf (0) = M(0) =6 0, which is a contradiction. Hence E fn converges to E ϕf pointwise. By Theorem 1.1, the sequence E fn is Cd-Lipschitz. Hence, the pointwise convergence can be promoted to uniform convergence on any compact set by Arzel`a-Ascolitheorem. Therefore, E fn converges compactly to E ϕf . As an application of Theorem 5.6, we record the following proposition for future references.

Proposition 5.7. Let fn ∈ Ratd(C) converge to f ∈ Ratd(C) with deg(ϕf ) ≥ −1 1. Then there exists R > 0 and xn ∈ E fn (0) so that for all suﬃciently large n,

dH3 (0, xn) ≤ R. 3 Proof. Fix a Euclidean ball B(0, r) ⊆ R with r < 1 such that −1 E ϕf (0) ∩ (B(0, 1) − B(0, r)) = ∅.

Such r exists as E ϕf is proper. −1 It suffices to show that for all sufficiently large n, E fn (0) ∩ B(0, r) =6 ∅. Suppose not. Then after passing to a subsequence, we can define 2 Fn : B(0, r) −→ S

x 7→ E fn(x)/kE fn(x)k, and F : ∂B(0, r) −→ S2

x 7→ E ϕf (x)/kE ϕf (x)k. 16 YUSHENG LUO

Since E fn converges uniformly to E ϕf on ∂B(0, r) by Theorem 5.6, Fn|∂B(0,r) is homotopic to F . Since there are no preimages of 0 of E ϕf in B(0, 1) − B(0, r), F is homotopic to ϕf on Cˆ. So Fn|∂B(0,r) is homotopic to ϕf . Since ϕf has homological degree ≥ 1, Fn|∂B(0,r) has homological degree ≥ 1 as well. Therefore, Fn cannot be extended to a continuous map from 2 B(0, r) to S , giving a contradiction. This proves the proposition. 6. The Caratheodory´ convergence of annuli In this section, we will first define the Carathéodory topology for annuli (cf. Carathéodory topology for pointed disks in [McM94, §5]). We will then use it to study the barycentric extensions on large scale. Carathéodory topology on the spaces of annuli. Let E ⊆ C be a continuum, i.e., a connected compact set. Let V ⊆ C − E be the unbounded component. We define the hull of E by Hull(E) = C − V . Note that the Hull(E) is a full continuum, i.e., a connected compact set whose complement is also connected.

Definition 6.1. Let An = Un −Kn be an annulus where Un is a topological disk of C, and Kn ⊆ Un is a full continuum. We say An converges to an annulus A = U − K (in the Carathéodory topology) if (1) Kn converges in the Hausdorff topology on compact subset of Cˆ to E, and K = Hull(E); and (2) Cˆ − Un converges in the Hausdorff topology on compact subset of Cˆ to D, and U is the component of Cˆ − D containing K.

Equivalently, An = Un−Kn converges to A = U −K if there exist un ∈ Kn and u ∈ K so that • (Un, un) converges to (U, u) in the Carath´eodory topology; and • Kn converges to E in the Hausdorﬀ topology, and K = Hull(E).

Figure 6.1. An illustration of the convergence of annuli. LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 17

Let mod(A) be the modulus of an annulus A. The following theorem follows from the proof of [McM94, Theorem 5.8]. We include the proof here for completeness.

Theorem 6.2. Given m > 0, the space of annuli A ⊆ C with mod(A) ≥ m is compact up to aﬃne transformation. More precisely, let An = Un −Kn with mod(An) ≥ m. Then after passing to a subsequence, either

(1) Kn is degenerate for all n, and An normalized so that Kn = {0} and B(0, 1) ⊆ Un, has a convergent subsequence; or (2) Kn is nondegenerate for all n, and An normalized so that 0 ∈ Kn and the Euclidean diameter of Kn is equal to 1, has a convergent subsequence.

Proof. For the first case, after passing to a subsequence, (Un, 0) converges to some pointed disk (U, 0) in the Carathéodory topology by [McM94, Theorem 5.2]. Thus An converges to A = U − {0}, and mod(A) = ∞ ≥ m. For the second case, since the Euclidean diameter of 0 ∈ Kn is equal to 1 and mod(An) ≥ m, the Euclidean distance from 0 to ∂Un is greater than C > 0, where C is a constant depending only on m by [McM94, Theorem 2.5]. Thus, after passing to a subsequence, (Un, 0) converges to some pointed disk (U, 0) in the Carathéodory topology by [McM94, Theorem 5.2]. The lower bound mod(An) ≥ m provides an upper bound on the diameter of Kn in the hyperbolic metric on Un, by [McM94, Theorem 2.4]. Thus, after passing to a subsequence, the Hausdorff limit E of Kn is a compact set of U by [McM94, Theorem 5.3], so K = Hull(E) is also a compact set of U. Therefore An converges to A = U − K. Let hn : A(m) −→ An be a univalent map from the standard round annulus A(m) of modulus m into An. Then one can extract a limiting injection into A by [McM94, Corollary 2.8], so mod(A) ≥ m. Carathéodory topology on functions on annuli. The following notion is the analogue of the Carathéodory topology for functions on pointed disks [McM94, §5].

Deﬁnition 6.3. Let fn : An −→ C be a sequence of holomorphic functions on annuli An ⊆ C. Then fn is said to converge to f : A −→ C if • An converges to A; and • for all suﬃciently large n, fn converges to f uniformly on compact subsets of A.

Note that any compact set M ⊆ A is eventually contained in An, so fn is deﬁned on M for all suﬃciently large n. 0 Lemma 6.4. Let fn : An −→ A ⊆ C be a sequence of degree e holomorphic 0 0 coverings between annuli An and A . If fn converges to f : A −→ A , then f : A −→ A0 is a degree e covering. 18 YUSHENG LUO

Proof. We will first show that f : A −→ A0 is proper. Post-composing with a uniformization map, we may assume A0 = B(0,R) − B(0, 1). 0 Let > 0 and A := B(0,R − ) − B(0, 1 + ). Since fn is a degree e −1 0 ± ± covering map, An − fn (A) consists of two annuli An,, with mod(An,) −1 0 uniformly bounded from below. Therefore, fn (A) converges to a compact subset of A by the same argument in Theorem 6.2. Since fn converges −1 0 0 compactly to f, f (A) is compact in A. Since any compact set of A is 0 contained in A for some , f is proper. By the argument principle, the degree is e. Thus f : A −→ A0 is a degree e covering. The next lemma is useful to bound the degrees for limits of rational maps. 0 Lemma 6.5. Let fn ∈ Ratd(C) with fn → f ∈ Ratd(C). Let An, A, A ⊆ C 0 be annuli with An converging to A. If fn : An −→ A is a degree e covering, 0 then ϕf : A −→ A is a degree e covering. In particular, deg(ϕf ) ≥ e. Proof. We claim A ∩ H(f) = ∅. Suppose not, and let U ⊆ A be a neighborhood of a hole a ∈ H(f). Let M = Cˆ if ϕf is non-constant, and let M 0 be a compact subset of Cˆ − {C} with M ∩ (Cˆ − A ) =6 ∅ if ϕf ≡ C. By Lemma 5.2, M ⊆ fn(U) for sufficiently large n. This is a contradiction as 0 fn(U) ⊆ A . 0 0 Therefore, by Lemma 5.1, fn : An −→ A converges to ϕf : A −→ A . By Lemma 6.4, ϕf is a degree e covering. Maps on annuli with large modulus. In the following, we will discuss how the Carathéodory convergence allows us to control barycentric extensions on large scale. 3 Given two points x, y ∈ H , we can associate an annulus, denoted by A(x, y), as follows. Let Hx and Hy be the two geodesic planes perpendicular to the geodesic segment [x, y] that pass through x and y respectively. The ∼ 3 boundaries of Hx and Hy in Cˆ = ∂H are two circles Cx and Cy. The annulus A(x, y) is defined as the region bounded by Cx and Cy. Note that mod(A(x, y)) = dH3 (x, y)/2π. 3 We say a pair of points (x, y) in H approximates an annulus A if • A(x, y) is essentially contained in A; and 5 log 2 • mod(A) ≤ mod(A(x, y)) + 2π . Here essential means π1(A(x, y)) injects into π1(A). By [McM94, Theorem 2.1], any annulus A with sufficiently large modulus 3 can be approximated by a pair of points in H (see also [Ahl73, §4.11-12]). The following proposition shows how the maps on large annuli control the corresponding barycentric extensions. We will use it to study the limiting 3 map on rescalings of H in §8. 3 Proposition 6.6. Let fn ∈ Ratd(C). Let xn, yn ∈ H with dH3 (xn, yn) → 0 ∞. Assume fn : An −→ An := A(xn, yn) is a degree e covering, and (an, bn) approximates An. Then dH3 (xn, E fn(an)) and dH3 (yn, E fn(bn)) are bounded. LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 19

Proof. We shall prove dH3 (xn, E fn(an)) is bounded, the second statement is proved similarly. 0 0 By naturality, we assume xn = an = 0 with An = B(0,Rn) − B(0, 1) and A(an, bn) = B(0,Rn) − B(0, 1) ⊆ An ⊆ C. Let Dn be the bounded component of C − An. We assume that ∂Dn is mapped to ∂B(0, 1). Under the above normalization, by Proposition 5.5, it suﬃces to show that any limit point of fn in Ratd(C) has degree ≥ 1. Suppose for contradiction that this is not the case. Then after passing to a subsequence, we assume fn → f ∈ Ratd(C) of degree 0.

0 Cyn = ∂B(0,Rn) −1 ∂An ∩ fn (Cyn )

∂B(0,K)

Cb = ∂B(0,Rn) 0 Aen n Ae

Dn −1 ∂An ∩ fn (Cxn ) C ∂B , Can = ∂B(0, 1) xn = (0 1)

Figure 6.2

0 −1 0 0 0 Let Ae := B(0,K) − B(0, 1) and Aen := fn (Ae ) ∩ An. Then Ae ⊆ An for 0 all large n, and fn : Aen −→ Ae is a degree e covering. Since mod(B(0,Rn) − Dn) ≤ mod(B(0,Rn) − B(0, 1)) + O(1), the Eu- clidean diameter of Dn is bounded from below. Therefore, after passing to a subsequence, Aen converges to an annulus Ae by Theorem 6.2. Hence, by Lemma 6.5, ϕf has degree ≥ e ≥ 1, which is a contradiction. The following proposition gives us control on the preimages of barycentric extensions on large scale. We will use it in §8 to show the degree of the limiting map is d. 3 Proposition 6.7. Let fn ∈ Ratd(C). Let xn, yn ∈ H with dH3 (xn, yn) → −1 ∞. Let an ∈ E fn (xn). Then after passing to a subsequence, there exists −1 3 bn ∈ E fn (yn) ∈ H with

lim d 3 (an, bn)/ d 3 (xn, yn) ≤ 1. n→∞ H H 0 Proof. By naturality, we assume xn = an = 0 with An := A(xn, yn) = 0 B(0,Rn) − B(0, 1). Since E fn(0) = 0, after passing to a subsequence, we assume fn → f ∈ Ratd(C) of degree ≥ 1 by Proposition 5.5. 0 −1 0 Assume ﬁrst that An contains no critical values of fn. Then fn (An) consists of d annuli counted multiplicities. Since ϕf has degree ≥ 1, there −1 0 exists a sequence of components An ⊆ fn (An) converging to an annulus A. ˆ Let (ˆan, bn) approximate An. Then it is easy to see that dH3 (an, aˆn) is bounded. Let Mn(0) = ˆbn, and Ln(0) = yn. By Proposition 6.6, 20 YUSHENG LUO

ˆ dH3 (yn, E fn(bn)) is bounded. Thus after passing to a subsequence, we as- −1 sume Ln ◦ fn ◦ Mn → g ∈ Ratd(C) of degree ≥ 1 by Proposition 5.5. −1 Applying Proposition 5.7 to Ln ◦ fn ◦ Mn, there exists R > 0 and bn ∈ −1 ˆ E fn (yn) with dH3 (bn, bn) ≤ R. ˆ Hence, dH3 (an, bn) = dH3 (ˆan, bn) + O(1) = 2π mod(An) + O(1). Since 0 0 fn : An −→ An is a covering, mod(An) ≤ mod(An). 0 Since dH3 (xn, yn) = 2π mod(An) → ∞, 0 lim d 3 (an, bn)/ d 3 (xn, yn) = lim mod(An)/ mod(A ) ≤ 1. n→∞ H H n→∞ n More generally, after passing to a subsequence, we can decompose

[xn, yn] = [x0,n = xn, x1,n] ∪ [x1,n, x2,n] ∪ ... ∪ [xk−1,n, xk,n = yn] so that each annulus A(xi,n, xi+1,n) contains no critical values of fn. Then the statement follows by induction.

7. R-trees and branched coverings In this section, we will deﬁne branched coverings between R-trees.

R-trees. An R-tree is a nonempty metric space (T, d) such that any two points x, y ∈ T are connected by a unique topological arc [x, y] ⊆ T , and every arc of T is isometric to an interval in R. We say x is an endpoint of T if T − {x} is connected; otherwise x is an interior point. If T −{x} has three or more components, we say x is a branch point. The set of branch points will be denoted B(T ). We say T is a finite tree if B(T ) is finite. Note that we allow a finite tree to have an infinite end, so a finite tree may not be compact. We will write [x, y) and (x, y) for [x, y] with one or both of its endpoints removed.

Ends of an R-tree. A ray α in the R-tree T is a subtree isometric to [0, ∞) ⊆ R. Two rays α1, α2 are equivalent, denoted by α1 ∼ α2, if α1 ∩ α2 is still a ray. The collection (T ) of all equivalence classes of rays forms the set of ends of T . We will use α to denote both a ray and the end it represents. We say a sequence of points xi converges to an end α, denoted by xi → α, if for all β ∼ α, xi ∈ β for all suﬃciently large i.

Convexity and subtrees. A subset S of T is called convex if x, y ∈ S =⇒ [x, y] ⊆ S. The smallest convex set containing E ⊆ T is called the convex hull of E, and is denoted by hull(E). More generally, we can easily extend the above definitions to subsets of T ∪ (T ). Note that a subset S ⊆ T is convex ⇐⇒ S is connected ⇐⇒ S is a subtree. Moreover, S is a finite subtree of T ⇐⇒ S is the convex hull of a finite set E ⊆ T ∪ (T ). LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 21

Branched coverings between R-trees. We now give the deﬁnition of a branched covering between R-trees:

Deﬁnition 7.1. Let f : T1 −→ T2 be a continuous map between two R- trees. We say f is a degree d branched covering if there is a ﬁnite subtree S ⊆ T1 such that

(1) S is nowhere dense in T1, and f(S) is nowhere dense in T2. (2) For every y ∈ T2 − f(S), there are exactly d preimages in T1. (3) For every x ∈ T1 − S, f is a local isometric homeomorphism at x. (4) For every x ∈ S, and any suﬃciently small neighborhood U of f(x), f : V − f −1(f(V ∩ S)) −→ U − f(V ∩ S) is an isometric covering5, where V is the component of f −1(U) containing x.

Local degree and critical locus. Let f : T1 −→ T2 be a degree d branched covering, and x ∈ T1, we define the local degree at x, denoted as degx(f) as the degree of the isometric covering of f : V −f −1(f(V ∩S)) −→ U −f(V ∩S) for sufficiently small neighborhood U of f(x). We define

C(f) = {x ∈ T1 : degx(f) ≥ 2} as the critical locus of f. Note that C(f) ⊆ S. One shall think of the finite subtree S in the definition as an intermediate step recording all ‘potential’ critical points for the branched covering, while C(f) is the set of critical points. For our purposes, it is more convenient to introduce this intermediate subtree S. Here are some properties which are ready to check using definitions.

Proposition 7.2. Let f : T1 −→ T2 be a degree d branched covering. Then (1) C(f) is a closed set. P (2) For y ∈ T2, f(x)=y degx(f) = d. (3) If degx(f) = 1, then f is a local isometric homeomorphism at x. (4) For every x ∈ T1, and any suﬃciently small neighborhood U of f(x), f : V − f −1(f(V ∩ C(f))) −→ U − f(V ∩ C(f)) is an isometric covering where V is the component of f −1(U) containing x. (5) If U is a subtree disjoint from C(f), then f maps U isometrically to f(U).

3 8. Limits of barycentric extensions on rescalings of H In this section, we will construct an R-tree, called the asymptotic cone, 3 by taking the limit of rescalings of H . We will construct a limiting map F on the asymptotic cone, and will prove Theorem 1.2 and Theorem 1.3 by analyzing the basic properties of the map F .

5Note that the covering is isometric if f is isometric restricting on each connected component of V − f −1(f(V ∩ C(f))). 22 YUSHENG LUO

8.1. An equivalent condition on degeneration. Let fn ∈ Ratd(C). We say fn is degenerating, denoted by fn → ∞, if fn escapes every compact set of Ratd(C). We ﬁrst show that fn is degenerating if and only if there is ‘loss of mass’ for the barycentric extension.

Proposition 8.1. Let fn ∈ Rat ( ), and rn := max −1 d 3 (0, y). d C y∈E fn (0) H Then fn → ∞ if and only if rn → ∞.

Proof. We first prove the ‘if’ direction. Assume rn → ∞, and suppose for contradiction that fn 6→ ∞. Let Mn ∈ PSL2(C) such that E fn(Mn(0)) = 0 and dH3 (0,Mn(0)) → ∞. After passing to a subsequence, we may assume that fn → f ∈ Ratd(C) and Mn → M ∈ Rat1(C) with deg(ϕM ) = 0. By Lemma 5.4, (fn ◦ Mn)∗µS2 converges to the delta measure (f ◦ ϕM )∗µS2 which is a contradiction to (fn ◦ Mn)∗µS2 is balanced. Conversely, assume fn → ∞. Suppose for contradiction that rn 6→ ∞. After passing to a subsequence, we may assume rn → r, fn → f ∈ Ratd(C) with deg(ϕf ) < d. We may also assume the set of critical values of fn converges to a finite set V . 0 Let A be an annulus of modulus K > 0 compactly contained in Cˆ − V . 0 0 We assume C/∈ A if ϕf ≡ C. Then for all sufficiently large n, A contains −1 0 no critical values of fn, so fn (A ) consists of d annuli counted multiplicities. Since fn → ∞, the set of holes H(f) =6 ∅. Let a ∈ H(f). By Lemma 5.2, after passing to a subsequence, there exists a sequence of components −1 0 An ⊆ fn (A ) converging to a. −1 Let Mn ∈ PSL2(C) so that Mn (An) ⊆ C separates 0 and ∞ with the −1 diameter of the bounded component of C − Mn (An) equal to 1. Since −1 −1 mod(Mn (An)) ≥ K/d, after passing to a subsequence, Mn (An) converges to an annulus A by Theorem 6.2. After passing to a subsequence, we may assume −1 0 fn ◦ Mn : Mn (An) −→ A is a degree e covering map, and fn ◦ Mn → g ∈ Ratd(C). By Lemma 6.5, deg(ϕg) ≥ e ≥ 1. Applying Proposition 5.7 to fn ◦ Mn, we conclude that there exists R > 0 −1 and a sequence xn ∈ (E fn ◦ Mn) (0), with dH3 (0, xn) ≤ R. Note that −1 Mn(xn) ∈ E fn (0). Since An converges to a, Mn → ∞. Therefore, dH3 (0,Mn(xn)) → ∞, so rn → ∞, which is a contradiction. A similar proof yields the following, which gives a criteria for a point a ∈ Cˆ to be a hole. Lemma 8.2. Let fn ∈ Ratd(C) converge to f ∈ Ratd(C) of degree ≥ 1. Let 3 a ∈ H(f) ⊆ Cˆ and yn ∈ H be a sequence. After passing to a subsequence, −1 there exists a sequence xn ∈ E fn (yn) converging to a. Proof. Let Ln ∈ PSL2(C) so that Ln(0) = yn. Let Vn be the set of critical values of fn. After passing to a subsequence, we may assume Ln → L ∈ LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 23

−1 0 Rat1(C), and Ln (Vn) converges to a ﬁnite set V . Let A be an annulus with mod(A) = K > 0 compactly contained in Cˆ − V − H(L). 0 0 0 Since A ∩ H(L) = ∅, Ln(A ) converges to L(A ), which is a point if deg(L) = 0 or an annulus if deg(L) = 1. 0 −1 −1 0 −1 0 Since A ∩ V = ∅, for suﬃciently large n,(Ln ◦ fn) (A ) = fn (Ln(A )) consists of d annuli counted multiplicities. By Lemma 5.2, after passing to −1 0 a subsequence, there exists a sequence of components An ⊆ fn (Ln(A )) converging to the hole a. −1 Let Mn ∈ PSL2(C) so that Mn (An) ⊆ C separates 0 and ∞ with the −1 diameter of the bounded component of C − Mn (An) equal to 1. By the −1 same argument as in Proposition 8.1, after passing to a subsequence, Ln ◦ fn ◦ Mn → g ∈ Ratd(C) of degree ≥ 1. −1 Applying Proposition 5.7 to Ln ◦ fn ◦ Mn, there exists R > 0 and a 0 −1 −1 0 sequence xn ∈ (Ln ◦ E fn ◦ Mn) (0), with dH3 (0, xn) ≤ R. Let xn := 0 −1 Mn(xn) ∈ E fn (yn). Since An → a, Mn(0) → a. Hence, xn → a.

8.2. Ultralimits and asymptotic cones of metric spaces. Proposition 8.1 suggests the natural objects to consider are the rescalings of metric 3 spaces. In the following, we will construct asymptotic cones of H . We refer the readers to [Roe03, KL95] for details.

Ultrafilters. We begin by briefly reviewing the theory of ultrafilters on N. A subset ω ⊆ ℘(N) of the power set of N is called an ultrafilter if it satisfies the following 4 properties: (1) If A, B ∈ ω, then A ∩ B ∈ ω; (2) If A ∈ ω and A ⊆ B, then B ∈ ω; (3) ∅ ∈/ ω; (4) If A ⊆ N, then either A ∈ ω or N − A ∈ ω. By virtue of the 4 properties, one can think of an ultrafilter ω as defining a finitely additive {0, 1}-valued probability measure on N. The sets of measure 1 are precisely those belonging to the ultrafilter ω. We will call a set in ω as ω-big or simply big. Its complement is called ω-small or simply small. If a specific property is satisfied by a ω-big set, we will say this property holds ω-almost surely.

Example. Let a ∈ N. We deﬁne

ωa := {A ⊆ N : a ∈ A}.

Then it can be easily verified that ωa is an ultrafilter on N. An ultrafilter of the above type will be called a principal ultrafilter. It can be shown that an ultrafilter is principal if and only if it contains a finite set. An ultrafilter that is not principal is called a non-principal ultrafilter. The existence of a non-principal ultrafilter is guaranteed by Zorn’s lemma. 24 YUSHENG LUO

Let ω be a non-principal ultraﬁlter on N. Let xn be a sequence in a metric space (X, d) and x ∈ X. We say x is the ω-limit of xn, denoted by

lim xn = x, ω if for every > 0, the set {n : d(xn, x) < } is big. It can be easily veriﬁed (see [KL95]) that (1) If the ω-limit of xn exists, then it is unique. (2) If xn is contained in a compact set, then the ω-limit exists. (3) If x = limn→∞ xn in the standard sense, then x = limω xn. (4) If x = limω xn, then there exists a subsequence nk such that x =

limk→∞ xnk in the standard sense. From these properties, one should intuitively think (as one of the ben- efits) of the non-principal ultrafilter ω as performing all the subsequence- selection in advance. The non-principal ultrafilter ω allows all sequences in compact spaces to converge automatically, without the need to pass to a subsequences. Convention. From now on and throughout the rest of the paper, we will fix a non-principal ultrafilter ω on N. The statements shall be interpreted as ω-almost surely without explicit mention.

Ultralimits of pointed metric spaces. Let (Xn, pn, dn) be a sequence of pointed metric spaces with basepoint pn. Let X denote the set of sequences (xn) ⊆ Xn such that dn(xn, pn) is a bounded function of n. We deﬁne an equivalence relation ∼ on X by

(xn) ∼ (yn) ⇔ lim dn(xn, yn) = 0. ω

We use [(xn)] to denote the equivalence class of the sequence (xn). Let Xω = X / ∼. We deﬁne

dω([(xn)], [(yn)]) := lim dn(xn, yn). ω The function dω makes Xω a metric space, and is called the ultralimit of (Xn, pn, dn) with respect to the ultrafilter ω. It is written as limω(Xn, pn, dn) or simply limω(Xn, pn) for short. The ultralimit of Xn has many of the desired properties (see [Roe03, §7.5] and [KL95] for associated definitions and proofs): (1) The ultralimit is always a complete metric space. (2) The ultralimit of a sequence of length spaces is a length space. (3) The ultralimit of a sequence of geodesic spaces is a geodesic space. (4) If Xn is proper and (Xn, pn) → (Y, y) in the sense of Gromov- Hausdorff, then (Y, y) is canonically isometric to limω(Xn, pn). 3 Asymptotic cones of H . Given a positive sequence rn → ∞, we define r 3 0 the asymptotic cone ( H , x , d) with rescaling rn as the ultralimit r 3 0 3 1 ( H , x , d) = lim(H , 0, dH3 ). ω rn LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 25

r 3 3 From the definition, a point x ∈ H is represented by a sequence xn ∈ H d 3 (0,xn) d 3 (xn,yn) H < ∞ x y ⇐⇒ H with limω rn , and [( n)] = [( n)] limω rn = 0. r 3 It is well known that H is an R-tree that has uncountably many branches at every point (see [Roe03, KL95]). In the sequel [Luo19], we will show r 3 that H is canonically isometric to the Berkovich hyperbolic space of the complexified Robinson’s field. ∼ 2 3 Let z ∈ Cˆ = S . We denote γ(t, z) ∈ H as the point of distance t from 0 in the direction of z. Given any sequence zn ∈ Cˆ, the ray

s(t) = [(γ(t · rn, zn))] r 3 is a geodesic ray parameterized by arc length in H . Thus, (zn) corresponds r 3 to an end α ∈ ( H ). We denote this by zn ω α. r 3 8.3. The construction of the limiting map F on H . We ﬁrst show that Theorem 1.1 allows us to construct a natural limiting map on the asymptotic cone.

Proposition 8.3. Let fn ∈ Ratd(C), and rn → ∞. If there exists K > 0 such that dH3 (0, E fn(0)) ≤ Krn, then the natural limiting map r 3 r 3 F = lim E fn : −→ ω H H [(xn)] 7→ [(E fn(xn))] is well-deﬁned and uniformly Lipschitz. r 3 Proof. We ﬁrst show (E fn(xn)) represents a point in H . Since [(xn)] ∈ r 3 H , there exists M > 0 so that

dH3 (xn, 0) ≤ Mrn. Hence Theorem 1.1 implies that

dH3 (E fn(0), E fn(xn)) ≤ Cd · Mrn.

By our assumption, dH3 (0, E fn(0)) ≤ Krn. Thus,

dH3 (0, E fn(xn)) ≤ dH3 (0, E fn(0)) + dH3 (E fn(0), E fn(xn)) ≤ (K + Cd · M)rn. r 3 Therefore, (E fn(xn)) represents a point in H . We now show the deﬁnition F does not depend on the choice of represen- r 3 tatives. If [(xn)] = [(yn)] ∈ H , then

lim d 3 (xn, yn)/rn = 0. ω H By Theorem 1.1,

lim d 3 (E fn(xn), E fn(yn))/rn ≤ lim Cd · d 3 (xn, yn)/rn = 0, ω H ω H so [(E fn(xn))] = [(E fn(yn))]. Thus, F is well-deﬁned. By the same computation, F is Cd-Lipschitz. 26 YUSHENG LUO

If fn → ∞ in Rat , then rn := max −1 d 3 (0, y) → ∞ by Proposi- d y∈E fn (0) H −1 tion 8.1. For yn ∈ E fn (0), Theorem 1.1 gives

dH3 (0, E fn(0)) ≤ Cd · dH3 (0, yn) ≤ Cd · rn. Hence, we have the following immediate corollary.

Corollary 8.4. Let fn → ∞ in Ratd(C), and

rn := max d 3 (0, y). −1 H y∈E fn (0) Then the natural limiting map r 3 r 3 F = lim E fn : −→ ω H H [(xn)] 7→ [(E fn(xn))] is well-deﬁned and uniformly Lipschitz.

8.4. The limiting map F is dynamically natural. Let fn → ∞ in r 3 r 3 Rat ( ) with rn = max −1 d 3 (0, y). Let F : −→ be the d C y∈E fn (0) H H H limiting map. In the following, we will verify F is dynamically natural. We record the following lemma, which follows immediately from the naturality of the barycentric extension and Proposition 5.5.

Lemma 8.5. Let fn ∈ Ratd(C). Then dH3 (Ln(0), E fn(Mn(0))) is bounded −1 ω-almost surely if and only if deg(limω Ln ◦ fn ◦ Mn) ≥ 1. For barycentric extensions, E f ◦ E g is usually diﬀerent from E (f ◦ g). We now show this discrepancy disappears when we pass to the limit. r 3 N N Proof of Theorem 1.3. Given x ∈ H , we will show F (x) = limω E (fn )(x). Let M0,n ∈ PSL2(C) with x = [(M0,n(0))]. Let Mi,n ∈ PSL2(C) with N Mi,n(0) = E fn(Mi−1,n(0)). Note that F (x) = [(MN,n(0))]. −1 Since E fn ◦ Mi−1,n(0) = Mi,n(0), by Lemma 8.5, limω Mi,n ◦ fn ◦ Mi−1,n −1 N has degree ≥ 1, for i = 1, ..., N. Hence limω MN,n ◦ fn ◦ M0,n has degree ≥ 1 by Lemma 5.3. Thus, N dH3 (MN,n(0), E (fn )(M0,n(0))) is bounded ω-almost surely by Lemma 8.5. Therefore, N N N lim E (f )(x) = [(E (f )(M0,n(0)))] = [(MN,n(0))] = F (x). ω n n 8.5. The limiting map F is a branched covering. In the following, we will verify that F is a branched covering between R-trees. Let c1,n, ..., c2d−2,n be the 2d − 2 critical points of fn. Let κ1, ..., κ2d−2 be r 3 the corresponding ends of H , i.e., ci,n ω κi for all i = 1, ..., 2d − 2. We deﬁne the critical subtree by r 3 S = hull(κ1, ..., κ2d−2) ⊆ H . LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 27

Let fn(c1,n), ..., fn(c2d−2,n) be the 2d − 2 critical values of fn. The corresponding ends are called the ends of critical values. The following lemma follows from the analysis of the dynamics fn on annuli with large modulus. r 3 Lemma 8.6. Let x, y ∈ H . Assume the projections of the ends of critical values onto [x, y] are not interior points. Then there are segments r 3 [a1, b1], ..., [ak, bk] ⊆ H with disjoint interiors6 such that (1) F maps [ai, bi] linearly onto [x, y] with derivative ei ∈ Z≥1; and Pk (2) i=1 ei = d; and −1 Sk (3) F ([x, y]) = i=1[ai, bi]. Proof. Since the projections of the ends of critical values onto [x, y] are not interior points, we can choose representatives (xn), (yn) of x, y so that the 7 annulus An := A(xn, yn) contains no critical values of fn. −1 8 Let fn (An) = {A1,n, ..., Ak,n}, and ei be the degree of the covering fn : Ai,n −→ An. 7 Let (ai,n, bi,n) approximate Ai,n . By Proposition 6.6, dH3 (xn, E fn(ai,n)) −1 is bounded. By Proposition 6.7, there exists un ∈ E fn (0) with

lim d 3 (un, ai,n)/ d 3 (0, E fn(ai,n)) ≤ 1. ω H H 0 Thus, limω dH3 (un, ai,n)/rn ≤ limω dH3 (0, E fn(ai,n))/rn = d(x , x). By our choice of rn, limω dH3 (0, un)/rn ≤ 1. Thus, (ai,n) represents a r 3 point in H . Let ai := [(ai,n)]. Similarly, let bi = [(bi,n)]. By construction, F (ai) = x and F (bi) = y. Moreover, −1 d(ai, bi) = lim r · (2π mod Ai,n + O(1)) ω n −1 = lim r · (2π mod An/ei + O(1)) ω n −1 = lim r · (d 3 (xn, yn)/ei + O(1)) = d(x, y)/ei. ω n H Since the same arguement works for any pair of points in [x, y], we conclude that [ai, bi] maps linearly onto [x, y] with derivative ei ∈ Z≥1. Since the components A1,n, ..., Ak,n are disjoint, [a1, b1], ..., [ak, bk] have Pk disjoint interior. Since the degree of fn is d, i=1 ei = d. −1 To prove the last property, it suﬃces to show F (x) = {a1, ..., ak}. By 0 naturality, it suﬃces to show x ∈ {a1, ..., ak} assuming E fn(0) = 0. By Lemma 8.5, deg(limω fn) ≥ 1. Therefore, Ai,n converges to an annulus Ai 0 for some i. So x ∈ {a1, ..., ak}. Similar to Lemma 8.6, we have

6 Note that ai (or bi) may equal to aj (or bj ) for distinct i, j. 7See page 18 for the deﬁnition. 8 Note that the number of components k and the degree ei are deﬁned ω-almost surely. 28 YUSHENG LUO

r 3 Lemma 8.7. Let x, y ∈ H . If the projections of the critical ends onto [x, y] are not interior points, then F maps [x, y] linearly onto [F (x),F (y)] with derivative e ∈ Z≥1. r 3 Moreover, if [x, y] ⊆ H − S, then F maps [x, y] onto [F (x),F (y)] isometrically.

Proof. The first part follows from a similar proof as in Lemma 8.6. r 3 If [x, y] ⊆ H − S, then the projections of the critical ends onto [x, y] is a single point a ∈ [x, y]. To prove the moreover part, we consider two cases. If a = x (or a = y), then we can choose representatives xn, yn so that all the critical points of fn are contained in the component of Cˆ − A(xn, yn) corresponding to x (or y respectively). Then the degree of the associated covering map is 1. Therefore, F is an isometry. If a ∈ (x, y), then by the above arguement, F is an isometry on [x, a] and [a, y]. Thus, it suffices to show the image of [x, a) and (a, y] are disjoint. Let xn, yn represent x, y respectively. Let an be in the geodesic segment connecting xn, yn such that a = [(an)]. By naturality, we may assume that an = 0 with E fn(0) = 0, and xn, yn are contained in the geodesic with ends 0, ∞ ∈ Cˆ. Let f = limω fn. By Lemma 8.5, deg(ϕf ) ≥ 1. If deg(ϕf ) > 1, then there are at least two distinct limits of critical points of fn in Cˆ, contradicting r 3 [x, y] ⊆ H − S. Thus deg(ϕf ) = 1 Modifying an if necessary, we may assume that all the critical points converge to 1 ∈ Cˆ. So by Lemma 5.2, H(f) = {1}. Thus, fn converges to ϕf uniformly near 0, ∞ by Lemma 5.1. Since deg(ϕf ) = 1, ϕf (0) =6 ϕf (∞), so the image of [x, a) and (a, y] are disjoint. r 3 Lemma 8.8. Let y ∈ H − F (S). Then it has exactly d preimages. r 3 r 3 Proof. Since y ∈ H − F (S), we can find two points u, v ∈ H so that r 3 y ∈ (u, v) ⊆ [u, v] ⊆ H − F (S). By Lemma 8.6, there are segments [a1, b1], ..., [ak, bk] such that F maps Pk r 3 [ai, bi] to [x, y] with derivative ei and i=1 ei = d. Since [ai, bi] ∈ H − S, by Lemma 8.7, ei = 1, so k = d. Since [ai, bi] have disjoint interiors, y has exactly d preimages.

r 3 The tangent space Tx H at x is deﬁned as the set of components of r 3 r 3 H − {x}. Let a ∈ Tx H be a tangent direction with the corresponding r 3 component Sa. For all y ∈ Sa ⊆ H − {x} suﬃciently close to x, the projections of the critical ends onto [x, y] are not interior points. Thus by Lemma 8.7, F induces a map on the tangent space

r 3 r 3 DFx : Tx H −→ TF (x) H .

r 3 We define the local degree of F at the direction a ∈ Tx H by degx,a F = e, where e is the local derivative in the direction of a in Lemma 8.7. We define LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 29 the local degree of F at x as X degx F := degx,a F, {a:DFx(a)=b} r 3 where b ∈ TF (x) H . Using a similar argument as Lemma 8.6, it is not hard to check that the definition is independent of the choice of b and every point has exactly d preimages counted multiplicities. We remark that it follows from Lemma 8.9 below that this definition of local degrees agrees with the definition in §7. Once the connection with the Berkovich dynamics is established in the sequel [Luo19], the tangent spaces r 3 1 of H are naturally isomorphic to PK for some field K. The tangent map 1 DFx is a rational map on PK , and degx F is the degree of the rational map. r 3 Lemma 8.9. Let x ∈ H . Then for sufficiently small neighborhood U of F (x), F : V − F −1(F (V ∩ S)) −→ U − F (V ∩ S) is an isometric covering of degree degx F where V is the component of F −1(U) containing x. Proof. We choose a neighborhood U of F (x) so that the component V of F −1(U) containing x contains no other preimages of F (x). Let y ∈ U −F (S), we will show that y has exactly degx F preimages in V . Suppose that the projections of the ends of critical values onto [F (x), y] are not interior points. Since V contains no other preimages of F (x), k −1 [ F ([F (x), y]) ∩ V = [x, yi] i=1 by Lemma 8.6. By definition of the local degree, y has degx F preimages in V counted multiplicities. By Lemma 8.8, each yi has multiplicity 1. So k = degx F and y has exactly degx F preimages in V . More generally, we can decompose

[F (x), y] = [x0 = F (x), x1] ∪ ... ∪ [xn−1, xn = y] into ﬁnitely many segments where xi are the projections of the ends of critical values on [F (x), y]. Then by induction, each xi has degx F preimages in V counted multiplicities. By Lemma 8.8, each preimage of y has multiplicity 1. Therefore, y has exactly degx F preimages in V . −1 r 3 Since each component of V − F (F (V ∩ S)) is contained in H − S, by Lemma 8.7, the covering map is isometric. We are ready to prove Theorem 1.2:

r 3 Proof of Theorem 1.2. By Corollary 8.4, we have a limiting map F : H −→ r 3 r 3 H . It is easy to check that S and F (S) are both nowhere dense in H . So by Lemma 8.8 and Lemma 8.9, F is a branched covering of degree d. 30 YUSHENG LUO

8.6. Degenerating sequences in Md. In the following, we will give a version of Theorem 1.2 in the moduli space of rational maps

Md = Ratd(C)/ PSL2(C).

A sequence [fn] ∈ Md is degenerating, denoted by [fn] → ∞, if [fn] escapes every compact set of Md. Equivalently, [fn] → ∞ if and only if every sequence of representatives fn ∈ Ratd is degenerating. For [f] ∈ Md, we deﬁne 9 r([f]) := inf max dH3 (x, y). x∈H3 y∈E f −1(x) It follows immediately from the naturality and Proposition 8.1 that:

Proposition 8.10. [fn] → ∞ in Md if and only if r([fn]) → ∞.

Thus, the natural scale to consider is r([fn]), and we have

Theorem 8.11. Let [fn] → ∞ in Md and rn := r([fn]). Let fn be a representative of [fn] with

max d 3 (0, y) ≤ rn + 1. −1 H y∈E fn (0) Then the corresponding limiting map r 3 r 3 F = lim E fn : −→ ω H H is a branched covering of degree d with no totally invariant point. Proof. By Theorem 1.2, it suﬃces to show there are no totally invariant r 3 points. Since rn := r([fn]), for any point x ∈ H , max d(x, y) ≥ 1. y∈F −1(x) Therefore, there are no totally invariant points. 8.7. Geometric Limits. In the following, we will illustrate how to extract a geometric limit of E fn in the standard sense. 3 3 Let F : T −→ T be a map on an R-tree T and Fn : H −→ H . Following the deﬁnitions in [McM09, §11], we say Fn converges geometrically to F if there exists a sequence of maps 3 hn : T −→ H and a sequence of rescalings rn → ∞ such that (1) Rescaling: We have −1 d(x, y) = lim r d 3 (hn(x), hn(y)) n→∞ n H for all x, y ∈ T ;

9We do not know if the inﬁmum is achieved. Consider the function R(x) := 3 −1 3 y, x R x → ∞ x → ∞ maxy∈E f (x) dH ( ). It is not hard to show that ( ) as in H . We do not know if R(x) is continuous. LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 31

(2) Conjugacy: For all x ∈ T , we have −1 lim r d 3 (hn(F (x)),Fn(hn(x))) = 0. n→∞ n H r 3 r 3 Theorem 8.12. Let fn → ∞ in Ratd(C), with limiting map F : H → H . r 3 Let T = hull(A) ⊆ H be an invariant subtree where A is countable. After passing to a subsequence, E fn converges geometrically to F : T −→ T .

Proof. Enumerate A = {x0, x1, x2, ...} and let Tk = hull({x0, x1, ..., xk}) ⊆ r 3 H . Then T1 ⊆ T2 ⊆ ... is an increasing sequence of ﬁnite trees. Note ∞ [ T = Tk, k=1 S∞ as k=1 Tk is the smallest connected subset containing A. Since each Tk is a ﬁnite tree, it is separable, i.e., it contains a countable S∞ dense subset. Since T = k=1 Tk, T is also separable. Thus by adding countably many points, we may assume A is dense in T . By induction on k, we will construct hn on Tk with:

(1) x = [(hn(x))] for all x ∈ Tk; and (2)d H3 (hn(x), hn(y)) ≤ 2rn · d(x, y) for all x, y ∈ Tk. As the base case, we choose representatives x0 = [(x0,n)], x1 = [(x1,n)] so that

dH3 (x0,n, x1,n) ≤ 2rn · d(x0, x1). Deﬁne 3 hn : T1 = [x0, x1] −→ H by hn(x0) = x0,n, hn(x1) = x1,n, and extending linearly between [x0, x1] and the geodesic segment connecting hn(x0) and hn(x1). By construction, the two properties are satisﬁed. 3 Suppose hn : Tk −→ H is constructed. If Tk+1 = Tk, then we do nothing. Otherwise, Tk+1 = Tk ∪ [a, xk+1], where a is the projection of xk+1 onto Tk. Choose a representative xk+1 = [(xk+1,n)] with

dH3 (xk+1,n, hn(a)) ≤ 2rn · d(xk+1, a).

Define the map hn on [a, xk+1] by hn(xk+1) = xk+1,n, and extending linearly between [a, xk+1] and the geodesic segment connecting hn(a) and hn(xk+1). 3 By induction hypothesis and triangle inequality, hn : Tk+1 −→ H satisfies the two properties. S∞ 3 Let hn : T = k=1 Tk −→ H . By property (2), hn is 2rn-Lipschitz, so it is continuous. By property (1) and the definition of F , for x, y ∈ T , −1 d(x, y) = lim r d 3 (hn(x), hn(y)), ω n H and −1 lim r d 3 (hn(F (x)), E fn(hn(x))) = 0. ω n H 32 YUSHENG LUO

Since the ultralimit of a sequence of real numbers is in the limit set, we can ﬁnd a subsequence so that for all x, y ∈ {x0, ..., xk}, −1 d(x, y) = lim r d 3 (hn(x), hn(y)), n→∞ n H and −1 lim r d 3 (hn(F (x)), E fn(hn(x))) = 0. n→∞ n H Using a standard diagonal argument, we can pass to a further subsequence so that the rescaling and conjugacy conditions hold for all x, y ∈ A. −1 Since hn is 2-Lipschitz with respect to the metric rn dH3 in the target, and F , E fn are Cd-Lipschitz with a universal constant C, the rescaling and conjugacy conditions hold for all x, y ∈ T = A.

Remark. Let fn be a degenerating sequence of Blaschke product, with r 3 r 3 0 limiting map F : H → H . Let A be the grand orbit of the base point x , r 3 and T = hull(A) ⊆ H . Then it can be checked that Theorem 8.12 recovers the geometric limit constructed in [McM09, Theorem 1.3].

9. Length spectra and translation lengths In this section, we will show that we can recover the limiting length spectra from the translation lengths for the limiting map, and prove Theorem 1.4. Let fn → ∞ in Ratd(C) with

rn := max d 3 (0, y). −1 H y∈E fn (0) r 3 r 3 Let F : H −→ H be the limiting map. r 3 Let α : [0, ∞) −→ H represent an end. By Lemma 8.7, for all large t, F is either an isometry or expanding with constant derivative e ∈ Z≥2 on α([t, ∞)). If F is eventually an isometry, then d(α(t), x0) − d(F (α(t)), x0) is eventually constant. Otherwise, d(α(t), x0) − d(F (α(t)), x0) → −∞. Therefore, the translation length 0 0 L(α, F ) = lim d(xi, x ) − d(F (xi), x ) xi→α is well defined, which can possibly be −∞. We are ready to prove Theorem 1.4. Proof of Theorem 1.4. We consider the case for q = 1. By Theorem 1.3, the general case can be proved by taking iterations.10 Let zn be a fixed point for fn, with zn ω α. By naturality, we assume zn = 0. −t·rn First, assume α is not critical. Let Mt,n(z) = e z. Then α(t) = [(Mt,n(0))] represents the end. Let Lt,n(0) = E fn ◦Mt,n(0) with Lt,n(0) = 0. To simplify the notations, we define −1 ft,n := Lt,n ◦ fn ◦ Mt,n

10It can be veriﬁed from the argument below that the period of C is q. LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 33 and −1 ft := lim L ◦ fn ◦ Mt,n ∈ Ratd( ). ω t,n C

Note deg(ϕft ) ≥ 1 by Lemma 8.5. Since α is not critical, deg(ϕft ) = 1 for all large t. Since F −1(x0) is ﬁnite, for all large t, α(t) separates the end of α from F −1(x0). Therefore, for all large t −1 −1 −1 ˆ E ft,n (Lt,n(0)) = Mt,n (E fn (0)) → ∞ ∈ C.

Thus, by applying Lemma 8.2 to ft,n and Lt,n(0), 0 is not a hole. Hence, ft,n converges uniformly to ft near 0 by Lemma 5.1. This gives fn(0) = 0 ω F (α), so α is a ﬁxed end. Moreover, for all suﬃciently large t, −1 0 0 lim(L ◦ fn ◦ Mt,n) (0) = ϕ (0) =6 0. ω t,n ft By the chain rule, we get lim r−1(− log |L0 (0)| + log |f 0 (0)| + log |M 0 (0)|) = lim r−1 log |ϕ0 (0)| = 0. ω n t,n n t,n ω n ft Therefore, for all large t, we have L(α, F ) = d(α(t), x0) − d(F (α(t)), x0) = − lim r−1 log |M 0 (0)| + lim r−1 log |L0 (0)| ω n t,n ω n t,n = lim r−1 log |f 0 (0)|. ω n n

0 α log |(fn) (0)| −∞ If is critical, then a similar argument shows limω rn = .

Appendix A. The barycentric extension of z2 In this section, we will compute the extension of a degree 2 rational map 2 3 3 on S . Since Isom H × Isom H acts transitively on Rat2(C), by naturality, there is only one map z2 to study. Let (r, θ, h) be the cylindrical coordinate system for the hyperbolic 3- 3 space H . More precisely, (r, θ) is the polar coordinates for the projection 3 3 of p ∈ H onto the hyperbolic plane P ⊆ H and h is the signed hyperbolic distance of p to the plane. We will prove Theorem A.1. Let f(z) = z2. In the cylindrical coordinate, there exists a real analytic function δ : [0, ∞) −→ R with • δ(0) = 0, • δ(r) > 0 when r > 0, • limr→∞ δ(r) = 0, such that the barycentric extension E f is given by E f(r, θ, h) = (log(cosh(r)) − δ(r), 2θ, 2h). 34 YUSHENG LUO

Proof. Since the barycentric extension is natural and e2tf(z/et) = f(z), 2 2 E f sends the hyperbolic plane Ht of h = t to H2t. Hence, E f preserves the 2 2 2 hyperbolic plane H0 of h = 0. By naturality, the restriction E f : Ht −→ H2t is the same as E f| 2 . Therefore, it suﬃces to show H0 2 2 E f| 2 : −→ H0 H0 H0 (r, θ) 7→ (log(cosh(r)) − δ(r), 2θ). Since f(z) is invariant under complex conjugation, by naturality,

E f| 2 (r, 0) = (g(r), 0). H0 A priori, g(r) can be negative. We ﬁrst show g(r) > 0 when r > 0. t 3 3 Let u = 0 ∈ H ⊆ R . For 0 < t < 1, the x-coordinate of the integral 0 Z Z 2 1 − kuk 2 3 ζd(f ◦ Mu)∗µS2 (ζ) = f(ζ)( 2 ) dµS2 (ζ) ∈ R S2 S2 kζ − uk is (1 − t2)2 Z π Z 2π h(ϕ) cos(2θ) dθ sin ϕdϕ π t2 2 2t 2 4 (1 + ) 0 0 (1 − 1+t2 sin ϕ cos θ) 2 2 ∞ (1 − t ) Z π Z 2π X 2t = h(ϕ) cos(2θ) n( sin ϕ cos θ)n−1dθ sin ϕdϕ > 0. π t2 2 t2 4 (1 + ) 0 0 n=1 1 + Here h(ϕ) ≥ 0 is some non-zero function of ϕ. The last inequality holds as the integral ( Z 2π = 0 if n − 1 is odd cos(2θ)(cos θ)n−1dθ . 0 > 0 if n − 1 is even

Thus, the barycenter of (f ◦ Mu)∗µS2 is on the positive x-axis. Hence g(r) > 0 when r > 0. Since f(z) = ei2θf(z/eiθ), by naturality,

E f| 2 (r, θ) = R2θ(E f| 2 (r, 0)) = (g(r), 2θ), H0 H0 where R2θ is rotation by 2θ. Therefore, it suﬃces to compute g(r). We need the following lemma: z−t Lemma A.2. Let ft(z) = z 1−tz for t ∈ (0, 1). Then there exists a positive function κ : (0, 1) −→ R+ such that E ft(0) = (κ(t), π, 0) for all t ∈ (0, 1).

Proof of Lemma A.2. Let ∼ 3 P : C −→ C × R = R 2z |z|2 − 1 z 7→ ( , ) 1 + |z|2 |z|2 + 1 LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 35

Figure A.1. The graph of the function It. be the stereographic projection. Note that E ft(0) = 0 if and only if Z −1 ~ 3 ∼ P (ft(P (ζ)))dµS2 (ζ) = 0 ∈ R = C × R. S2 By symmetry, the second component in C × R of the integral is always 0. Changing the variables to z, the ﬁrst component of the integral equals to z−t Z 2z 4 i I 1−tz |dz|2. t := z−t 2 2 2 C 1 + |z 1−tz | (1 + |z| ) 2 In polar coordinate, we may express the integral as z−t Z ∞ Z 2π 2z 4 I 1−tz rdθdr t = z−t 2 2 2 0 0 1 + |z 1−tz | (1 + |z| ) z−t Z ∞ Z 2z 4 |z|dz 1−tz dr = ( z−t 2 2 2 ) 0 ∂Br 1 + |z 1−tz | (1 + |z| ) iz where Br is the disk centered at 0 of radius r. Let Jt(r) be the inner integral. We will show that Jt(r) is negative for all 2 r =6 1. Note that on ∂Br,z ¯ = r /z, so we have z−t Z 2z 4 |z|dz J r 1−tz t( ) = z−t 2 2 2 ∂Br 1 + |z 1−tz | (1 + |z| ) iz Z z−t 2z 1−tz 4 rdz = (z−t)(¯z−t) 2 2 ∂B 2 (1 + r ) iz r 1 + r (1−tz)(1−tz¯) z−t 8r Z 1−tz dz = 2 2 (z−t)(r2/z−t) i(1 + r ) ∂B 2 r 1 + r (1−tz)(1−tr2/z) 8r Z (z − t)(z − tr2) = 2 2 2 2 2 dz i(1 + r ) ∂Br (1 − tz)(z − tr ) + r (z − t)(r − tz) 16rπ (z − t)(z − tr2) = Resz∈B . (1 + r2)2 r (1 − tz)(z − tr2) + r2(z − t)(r2 − tz) Let F (z) = (z −t)(z −tr2) and G(z) = (1−tz)(z −tr2)+r2(z −t)(r2 −tz). Note G(r) = r(1 − tr)2 + r3(r − t)2 > 0 as t ∈ (0, 1) and r > 0. Since the 36 YUSHENG LUO

2 coeﬃcient of z in G(z) is negative, G(z) has two real roots x1, x2 with x1 < r < x2. Note that G(t) = t(1 − t2)(1 − r2) and G(tr2) = tr4(r2 − 1)(1 − t2) have 2 diﬀerent signs when r =6 1, and G(max{t, tr }) > 0, so x1 lies between t and 2 tr . Hence, F (x1) < 0. Therefore, when r =6 1,

2 (z − t)(z − tr ) F (x1) Resz∈Br 2 2 2 = 2 < 0. (1 − tz)(z − tr ) + r (z − t)(r − tz) −t(1 + r )(x1 − x2)

Hence, Jt(r) < 0 for all r =6 1, so It < 0. So   Z −v(t) −1 3 P (ft(P (ζ)))dµS2 (ζ) =  0  ∈ R 2 S 0 for some positive function v(t) > 0. Thus, the barycenter of (ft)∗µS2 is on the negative x-axis, so

E ft(0) = (κ(t), π, 0) for some positive function κ(t) > 0.

z−t −1 Let Mt(z) = 1−tz . By comparing zeros, Mt2 ◦ f ◦ M−t (z) = f 2t (z). 1+t2 1+t Note that dH3 (0,Mt(0)) = log 1−t . Hence, by Lemma A.2 and naturality, 1 + t 1 + t 1 + t2 E f| 2 (log , 0) = E f| 2 (log , π) = (log − δ˜(t), 0), H0 1 − t H0 1 − t 1 − t2 where δ˜(t) > 0 for t ∈ (0, 1). 1+t Thus, by substituting r = log 1−t , there exists δ(r) with δ(r) > 0 for r > 0 so that

E f| 2 (r, 0) = (log(cosh(r)) − δ(r), 0). H0

Since both E f| 2 and log(cosh(r)) are real analytic, δ(r) is real analytic. H0 ∼ 2 Since the spherical derivative of f at 1 ∈ Cˆ = S is 2,

log 2 = lim d 3 ((r, 0, 0), 0) − d 3 (E f(r, 0, 0), 0) r→∞ H H = lim (r − (log(cosh(r)) − δ(r))). r→∞

Since limr→∞(r − log(cosh(r))) = log 2, limr→∞ δ(r) = 0.

Appendix B. The uniform bound for barycentric extensions in higher dimensions In this section, we explain how our method can be generalized to give uniform bounds of barycentric extensions of proper quasiregular maps in higher dimensions. LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 37

Theorem B.1. Let f : Sn −→ Sn be a proper K-quasiregular map of degree d. Then the norm of the derivative of its barycentric extension E f : n+1 n+1 H −→ H satisﬁes 1 sup kDE f(x)k ≤ C · (Kd) n−1 . x∈Hn+1 Here the norm is computed with respect to the hyperbolic metric, and C = C(n) depends only on the dimension n. The proof of the above theorem is almost identical to the proof of Theorem 1.1 once we have all the ingredients. We refer the readers to [Ric93, Geh61, Geh62, Zie67] for detailed discussions on quasiconformal and quasiregular maps in higher dimensions.

Conformal capacity and separating set. The theory of extremal length and extremal width extends naturally to higher dimensions. For our pur- pose, we consider an equivalent formulation of capacity of a condenser. n In the literature, a condenser in R is a pair (A, C) where A is open in n R and C =6 ∅ is a compact subset of A. Let U = A − C,E1 = ∂C and E2 = ∂A − E1. To make our discussion parallel with §3, we shall also call the triple (U, E1,E2) arising in this way n a condenser in R .

Definition B.2. The (conformal) capacity of (U, E1,E2) is defined by Z n capU (E1,E2) := inf |∇u| dm u U n where dm is the Lebesgue measure on R , and the infimum is taken over all 0 1 11 nonnegative functions u ∈ C ∩Wn,loc with compact support and u|E1 ≥ 1. The extremal distance is defined by 1 modU (E1,E2) = . capU (E1,E2)

Let (U, E1,E2) be a condenser with U bounded. We say a closed set σ is separating if σ ⊆ U and if there are two (relative) open sets V1 and V2 of U with E1 ⊆ V1, E2 ⊆ V2 and U − σ = V1 ∪ V2. The following is a higher dimensional generalization of extremal width of family of separating curves. Deﬁnition B.3. Let Σsep be the set of all separating sets of a bounded con- n 12 denser (U, E1,E2) in R . The extremal width of separating sets is deﬁned by Z sep n W f n−1 dm. (Σ ) := infsep f∧Σ Rn

11 1 n Wn consists of all real valued functions u ∈ L with weak ﬁrst order partial deriva- n 1 1 tives which are themselves in L . Wn,loc consists those functions that are locally in Wn . 12It is usually referred to the modulus of separating sets in the literature. We use the terminology extremal width to make the discussion parallel with §3. 38 YUSHENG LUO

sep n Here, f ∧ Σ means f is a nonnegative Borel function on R such that Z fdHn−1 ≥ 1 for all σ ∈ Σsep, σ where dHn−1 is the n − 1-dimensional Hausdorﬀ measure. The following theorem is proved in [Zie67, Geh62], and is a direct analogue of Theorem 3.1.

n sep Theorem B.4. Let (U, E1,E2) be a bounded condenser in R and Σ be the set of all separating sets. Then

sep n−1 modU (E1,E2) = W(Σ ) .

n n Quasiregular maps. A mapping f :Ω −→ R of a domain Ω ⊆ R is K-quasiregular if 0 1 (1) f ∈ C ∩ Wn,loc; and (2) there exists K, 1 ≤ K < ∞ such that |f 0(x)|n ≤ K Jac f(x) a.e.. Quasiregular maps share many nice properties with holomorphic functions. For example, a non-constant quasiregular map is discrete and open (see [Ric93, Chapter I §4]). The following theorem, proved in [Ric93, Chapter II §1 and §10], is an analogue of Theorem 3.2.

0 0 0 Theorem B.5. Let (U, E1,E2) and (U ,E1,E2) be two condensers. Let 0 0 f : U −→ U be a proper quasiregular map of degree d with f(E1) = E1 and 0 f(E2) = E2. Then 0 0 modU 0 (E1,E2) ≤ Kd · modU (E1,E2). Remark. The theorem in [Ric93] is stated for connected U. It is ready to verify that the proof of the theorem does not require this assumption. Proof of Theorem B.1. Now that we have all the ingredients, the proof is almost identical to the proof of Theorem 1.1. Using Theorem B.4 and Theorem B.5 as substitutes for Theorem 3.1 and Theorem 3.2, we can show the following analogue of Proposition 4.1. Proposition B.6. Let f : Sn −→ Sn be a proper K-quasiregular map of degree d with E f(0) = 0. Then

−1 1 kFy(0, 0) k ≤ M(Kd) n−1 for some constant M = M(n) depending only on the dimension n. Theorem B.1 now follows immediately from Proposition B.6. LIMITS OF RATIONAL MAPS, R-TREES AND BARYCENTRIC EXTENSION 39

References [Ahl73] Lars Ahlfors. Conformal invariants: topics in geometric function theory. Mc- Graw Hill, 1973. [BCG95] GérardBesson, Gilles Courtois, and Sylvestre Gallot. Entropies et rigiditédes espaces localement symétriquesde courbure strictement. Geom. Funct. Anal., 5(5):731–799, 1995. [BCG96] GérardBesson, Gilles Courtois, and Sylvestre Gallot. Minimal entropy and Mostow’s rigidity theorems. Ergodic Theory and Dynam. Systems, 16(4):623– 649, 1996. [Bes88] Mladen Bestvina. Degenerations of the hyperbolic space. Duke Math. J., 56(1):143–161, 1988. [BR10] Matthew Baker and Robert Rumely. Potential theory and dynamics on the Berkovich projective line, volume 159. American Mathematical Society, Prov- idence, 2010. [Chi91] Ian Chiswell. Non-standard analysis and the Morgan-Shalen compactification. Q. J. Math., 42(1):257–270, 1991. [DE86] Adrien Douady and Clifford Earle. Conformally natural extension of homeomorphisms of the circle. Acta Math., 157:23–48, 1986. [DeM05] Laura DeMarco. Iteration at the boundary of the space of rational maps. Duke Math. J., 130(1):169–197, 10 2005. [DeM07] Laura DeMarco. The moduli space of quadratic rational maps. J. Amer. Math. Soc., 20:321–355, 2007. [DM08] Laura DeMarco and Curtis T. McMullen. Trees and the dynamics of polynomials. Ann. Sci. Ec.´ Norm. Supér., Ser. 4, 41(3):337–383, 2008. [EMS02] Clifford Earle, Vladimir Markovic, and Dragomir Saric. Barycentric extension and the bers embedding for asymptotic teichmüllerspace. Contemp. Math., 311:87–105, 2002. [FLP79] Albert Fathi, Fran¸coisLaudenbach, and Valentin Poénaru.Travaux de thurston sur les surfaces. Asterisque, (66), 1979. [Geh61] Frederick Gehring. Symmetrization of rings in space. Trans. Amer. Math. Soc., 101(3):499–519, 1961. [Geh62] Frederick Gehring. Extremal length definitions for the conformal capacity of rings in space. Michigan Math. J., 9(2):137–150, 1962. [Gro81] Michael Gromov. Groups of polynomial growth and expanding maps (with an appendix by Jacques Tits). Publ. Math. Inst. Hautes Etudes´ Sci., 53:53–78, 1981. [Hub06] John Hubbard. Teichmüller theory and applications to geometry, topology, and dynamics: Teichmüller theory. TeichmüllerTheory and Applications to Geom- etry, Topology, and Dynamics. Matrix Editions, 2006. [KL09] Jeremy Kahn and Mikhail Lyubich. The quasi-additivity law in conformal geometry. Ann. of Math., 169(2):561–593, 2009. [KL95] Michael Kapovich and Bernhard Leeb. On asymptotic cones and quasi-isometry classes of fundamental groups of 3-manifolds. Geom. Funct. Anal., 5(3):582–603, 1995. [Kiw15] Jan Kiwi. Rescaling limits of complex rational maps. Duke Math. J., 164(7):1437–1470, 2015. [Lév22] Paul Lévy. Lecons d’analyse fonctionnelle,. Collection de monographies sur la théoriedes fonctions. Gauthier-Villars, Paris, 1922. [Luo19] Yusheng Luo. Trees, length spectra for rational maps via barycentric extensions and Berkovich spaces. arXiv:1907.06220, 2019. [Mag12] Francesco Maggi. Sets of finite perimeter and geometric variational problems: an introduction to geometric measure theory. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 2012. 40 YUSHENG LUO

[McM94] Curtis McMullen. Complex dynamics and renormalization (AM-135). Princeton University Press, 1994. [McM96] Curtis McMullen. Renormalization and 3-Manifolds which fiber over the circle (AM-142). Princeton University Press, 1996. [McM09] Curtis McMullen. Ribbon R-trees and holomorphic dynamics on the unit disk. J. Topol., 2(1):23–76, 2009. [MS84] John Morgan and Peter Shalen. Valuations, trees, and degenerations of hyperbolic structures, I. Ann. of Math., 120(3):401–476, 1984. [Ota96] Jean-Pierre Otal. Le théorèmed’hyperbolisation pour les variétésfibréesde dimension 3. Asterisque, (235), 1996. [Pau88] FrédéricPaulin. Topologie de gromov équivariante, structures hyperboliques et arbres réels. Invent. Math., 94(1):53–80, 1988. [Pet11] Carsten Petersen. Conformally natural extensions revisited, arXiv:1102.1470, 2011. [Ric93] Seppo Rickman. Quasiregular mappings. Springer, 1993. [Roe03] John Roe. Lectures on coarse geometry. American Mathematical Society, Prov- idence, 2003. [Sch39] Erhard Schmidt. Uber das isoperimetrische problem im raum von n dimensionen. Math. Z., 44:689–788, 1939. [Shi89] Mitshuhiro Shishikura. Trees associated with the configuration of Herman rings. Ergodic Theory and Dynam. Systems, 9(3):543–560, 1989. [Sto07] Peter Storm. The barycenter method on singular spaces. Comment. Math. Helv., 82:133–173, 2007. [Thu88] William Thurston. On the geometry and dynamics of diffeomorphisms of surfaces. Bull. Amer. Math. Soc. (N.S.), 19(2):417–431, 1988. [Zie67] William Ziemer. Extremal length and conformal capacity. Trans. Amer. Math. Soc., 126(3):460–473, 1967.

Deptment of Mathematics, University of Michigan, East Hall, 530 Church Street, Ann Arbor, Michigan 48109 USA Email address: [email protected]