arXiv reader

Stratification of opers with apparent singularities and isomonodromic bubbling

Duong Dinh, Troy Figiel and Joerg Teschner

arXiv
2610.22326v1 · PDF · 15 October 2026
Categories
math.DG, hep-th, math-ph, math.AG, math.MP
Licence
CC BY 4.0 The site shows the text.

Abstract

On a compact Riemann surface of genus g \geq 2, we study Simpson’s stratification of the de Rham moduli space of rank-2 flat connections. Upon representing flat \mathrm{SL}_2(\mathbb{C})-connections by branched projective structures, or equivalently by opers with apparent singularities, the stratum containing a flat connection is determined by the minimal number of apparent singularities needed to represent it. We construct explicit holomorphic families in which two simple apparent singularities collide and disappear, so that the family crosses into a smaller stratum; this realizes Simpson’s closure relation between strata near generic points. The analogous families of Higgs bundles are described by Baker–Akhiezer divisors on the spectral curve in which two points converge to a pair of points exchanged by the spectral involution. We show that, before the collision, the two apparent singularities are joined by twin paths bounding a bubble, so that these families arise from the isomonodromic bubbling surgery discussed by Calsamiglia, Deroin and Francaviglia.

Items 79

Proposition

Proposition A (Reduction of apparent singularities)

ZR4F

Let \mathcal{D} be an oper on C with d-2 simple apparent singularities x_3, \dots, x_d, where d \leq 3g-3, and let x_0 \in C be generic. Then there exists a holomorphic family \{\mathcal{D}_u\}_{u\in z(U)} with \mathcal{D}_0 = \mathcal{D} such that, for u \neq 0, \mathcal{D}_u has simple apparent singularities at x_+(u), x_-(u), x_3, \dots, x_d, and the accessory parameters of x_\pm(u) with respect to the projective chart z are of the form

\nu_\pm(u) = \pm \frac{3}{4u} + \mathcal{O}(u).

(See Proposition 5.14; the precise form of \nu_\pm(u) is (?).)

Theorem

Theorem B (Crossing to smaller strata)

LQBD

Suppose moreover that d is even and that the projective monodromy of \mathcal{D} is irreducible. Then the holomorphic family \{\mathcal{D}_u\} of Proposition A is defined by a holomorphic family \{[E_u, \nabla_u]\}_u in \mathcal{M}_{dR} together with line subbundles L_u \hookrightarrow E_u of degree

\deg L_u = g-1-\frac{d}{2} \quad \text{for } u \neq 0, \qquad \qquad \deg L_0 = g - \frac{d}{2},

and D_{L_u}(\nabla_u) = x_+(u) + x_-(u) + x_3 + \dots + x_d for u \neq 0, D_{L_0}(\nabla_0) = x_3 + \dots + x_d. For d < 2g-2, [E_u, \nabla_u] lies in the partial oper stratum labelled by d for u \neq 0, and in the partial oper stratum labelled by d-2 at u = 0. (See Proposition 5.15, Corollary 5.16.)

Theorem

Theorem C (Bubbling)

EGIM

For sufficiently small u \neq 0, the branched projective structure corresponding to \mathcal{D}_u of Proposition A is the output of a bubbling surgery that creates branch points that correspond to the apparent singularities x_+(u) and x_-(u). Specifically, this bubbling surgery takes as an input the branched projective structure \sigma_u, which is obtained by removing the bubble bounded by the twin paths and identifying these twin paths, and has simple apparent singularities at x_3, \dots, x_d. (See Theorem 6.3.)

Theorem

Theorem 2.1

2VT2

Let \alpha be a stable Hodge bundle, contained in X \in \pi_0(\mathcal{M}_H^{\mathbb{C}^\ast}). Then \alpha^+_0 \subset \mathcal{M}_H and \alpha^+_1 \subset \mathcal{M}_{dR} are holomorphic Lagrangian submanifolds, isomorphic as varieties with \mathbb{C}^\ast-action to the positive weighted subspace of T_\alpha \mathcal{M}_H. In particular, for G = \mathrm{SL}(2, \mathbb{C}), \alpha^+_0 and \alpha^+_1 are isomorphic to \mathbb C^{3g-3}.

Proposition

Proposition 2.2

5ANQ

Let E be strictly unstable with destabilizing line subbundle L_E \hookrightarrow E.

  1. Let [E, \phi] \in \mathcal{M}_H(\Lambda) and \alpha = [E_M, \phi_D] \in N_d for some d < 2g-2. Then [E, \phi] \in \alpha^+_0 if and only if M is the unique destabilizing line subbundle of E and

    D_{M}(\phi) = D.
  2. Let [E, \nabla] \in \mathcal{M}_{dR} and \alpha = [E_M, \phi_D] \in N_d for some d < 2g-2. Then [E, \nabla] \in \alpha^+_1 if and only if M is the unique destabilizing line subbundle of E and

    D_{M}(\nabla) = D.
Theorem

Theorem 2.3

E3FK

Let \alpha be a stable Hodge bundle.

  1. The Lagrangian leaf \alpha^+_0 is closed in \mathcal{M}_H if and only if \alpha is very stable [PPe19, HH21]. For \alpha = [E_M, \phi_D] \in N_d, this holds if and only if D is reduced [HH21].

  2. The Lagrangian leaf \alpha^+_1 is closed in \mathcal{M}_{dR} [DS24], but the leaves do not form a C^1 foliation [Huang26].

Definition

Definition 3.1

GRIS

The Baker–Akhiezer divisor (In local frames adapted to \pi^\ast L, the line bundle \pi^\ast(K^{-1}) \mathcal{L}_{(E,\phi)} of eigenvectors of \pi^\ast(\phi) has sections of the form \left(\begin{smallmatrix} v + a(z) \\ c(z) \end{smallmatrix} \right), reminiscent of the Baker–Akhiezer function in the theory of integrable systems [DT23], [Dinh24wobbly, Remark 2.3].) \widetilde{D}_i(\phi) of the data (L \overset{i}{\rightarrow} E, \phi) is the zero divisor of the morphism of line bundles on \widetilde{C}

\pi^\ast(L) \overset{\pi^\ast i}{\longrightarrow} \pi^\ast(E) \longrightarrow \mathcal{L}_{(E, \phi)}.
Proposition

Proposition 3.2

X3SC

[DT23] Let (E, \phi) be a Higgs bundle with smooth spectral curve and L \overset{i}{\rightarrow} E an injection. Then \pi^\ast(p) < \widetilde{D}_i(\phi) if and only if i vanishes at p and

D_i(\phi) = \mathrm{Nm}(\widetilde{D}_i(\phi)).
Proposition

Proposition 3.3

P68L

[DT23] Let \widetilde{D} be an effective divisor on a non-degenerate spectral curve \widetilde{C}_q \overset{\pi}{\rightarrow} C. Let L and \Lambda be line bundles on C satisfying \mathcal{O}_C(D) \simeq K L^{-2} \Lambda where D = \mathrm{Nm}(\widetilde{D}). Then there exists a triple (L \overset{i}{\rightarrow} E, \phi) unique up to scaling of i such that \widetilde{D}_i(\phi) = \widetilde{D} and \det(E) \cong \Lambda. In particular, i is an embedding if and only if \widetilde{D} contains no summand of the form \pi^\ast(p) for p \in C.

Theorem

Theorem 3.4

HDLC

Let \widetilde{C}_q \overset{\pi}{\rightarrow} C be a non-degenerate spectral curve and \mathcal{L} a line bundle on \widetilde{C}_q. Assume that E = \pi_\ast \mathcal{L} is semistable, and let \det(E) = \Lambda. Then E is wobbly if and only if

\mathcal{L} \simeq \pi^\ast(L) \otimes \mathcal{O}_{\widetilde{C}_q}(\widetilde{D})

where L is a line bundle on C and \widetilde{D} is an effective divisor on \widetilde{C}_q with Q_{\mathrm{Nm}(\widetilde{D})} \neq 0. Furthermore, if \widetilde{D} has no summand of the form \pi^\ast(P) for some effective divisor P on C, then E admits an embedding L \overset{i}{\hookrightarrow} E and there is a canonical bijection between the corresponding space of nilpotent Higgs fields

H^0\left( C, K L^2 \Lambda^{-1} \right) \simeq \left\{ \phi \in H^0(C, \mathrm{End}(E)\otimes K) \text{ nilpotent}, \ker(\phi) = i(L) \right\}

and Q_{\mathrm{Nm}(\widetilde{D})}.

Definition

Definition 3.5

HC9C

A projective connection on C is a pair (P, s) where P is a flat \mathrm{PSL}_2(\mathbb C)-bundle on C with \mathbb{P}^1-fibers and s: C \rightarrow P is a global holomorphic section which is not parallel w.r.t. the flat structure of P. The points where ds = 0 are called apparent singularities; the order of an apparent singularity is the order of the zero of ds at that point.

Definition

Definition 3.6

IPBI

A projective structure is a maximal atlas \left\{ U_\alpha \right\}_{\alpha \in \mathcal{I}} of C together with local holomorphic functions \{ w_\alpha: U_\alpha \rightarrow \mathbb C\} the values of which are related by Möbius transformations, i.e. for all x\in U_\alpha \cap U_\beta,

\begin{aligned} &w_\beta (x) = \frac{a_{\beta \alpha} w_\alpha (x) + b_{\beta \alpha}}{c_{\beta \alpha} w_\alpha (x) + d_{\beta \alpha}}, &\begin{pmatrix} a_{\beta \alpha} & b_{\beta \alpha} \\ c_{\beta \alpha} & d_{\beta \alpha} \end{pmatrix} \in \operatorname{SL}_2(\mathbb C). \end{aligned}

Apparent singularities are where d w_\alpha = 0; the order of an apparent singularity is the order of the zero of d w_\alpha at that point.

Definition

Definition 3.7

QBIL

Let N be a line bundle on C, defined via transition functions (N)_{\alpha \beta} over a coordinate covering \mathcal{U} = \{ (U_\alpha, z_\alpha) \}_{\alpha \in \mathcal{I}} of C. An oper \mathcal{D} on N for the coordinate covering \mathcal{U} is a collection of meromorphic differential operators \{ \mathcal{D}_\alpha = \partial_{z_\alpha}^2 + q_\alpha(z_\alpha)\}_{\alpha \in \mathcal{I}} such that f_\alpha(z_\alpha) is a solution to \mathcal{D}_\alpha if and only if (N)_{\beta \alpha} f_\alpha is a solution to \mathcal{D}_\beta \mid_{U_\alpha \cap U_\beta}. The points at which \left( \frac{f_{\alpha,1}}{f_{\alpha,2}} \right)' = 0, where f_{\alpha,1} and f_{\alpha,2} are two linearly independent solutions of \mathcal{D}_\alpha, are called apparent singularities, with the order of the apparent singularity defined to be the order of the zero of \left( \frac{f_{\alpha,1}}{f_{\alpha,2}} \right)'. Two opers on N for two coordinate coverings are considered equivalent if their union is also an oper. An oper on N is an equivalence class of opers on N for different coverings.

Remark

Remark 3.8

BX9S

It was shown in [iwasaki91] that an oper exists on a line bundle N if and only if \deg(N) = 1-g, such as N \cong K^{-1/2}. Such an oper then can be regarded as a differential operator N \rightarrow N K^2 whose principal symbol is 1 and subprincipal symbol is 0 [Fr07].

Example

Example 3.9

FBRS

On a compact Riemann surface C of genus g\geq 2, the most distinguished projective connection is induced by the uniformization theorem, which realizes C as a quotient of the upper-half plane and equips on it a distinguished maximal coordinate atlas \{\left( U_\alpha, x_\alpha \right) \} from the upper-half plane. We call this the uniformizing projective structure of C. Since nowhere dx_\alpha = 0, the projective structure \{\left( U_\alpha, x_\alpha \right) \} has no apparent singularity. The corresponding oper takes the form \partial_{x_\alpha}^2 in each U_\alpha. This uniformizing projective structure is distinguished in the sense that the projective coordinates x_\alpha all take values in the upper half-plane.

It follows from (?) that if q is a holomorphic quadratic differential, then \mathcal{D}_\alpha \coloneqq \partial_{x_\alpha}^2 + q(x_\alpha) glue into an oper that has no apparent singularity. These form the space of holomorphic projective connections, i.e. those that have no apparent singularities, which is an affine space modeled over H^0(K^2) \cong \mathbb{C}^{3g-3}. The isomorphism class of each such oper is equivalent to a maximal coordinate atlas that is also a projective structure.

Lemma

Lemma 3.10

2PF4

Two \operatorname{SL}_2(\mathbb C)-holomorphic connections (F_1, \nabla_1), (F_2, \nabla_2) have monodromy representations \check{\rho}_1, \check{\rho}_2: \pi_1 \rightarrow \operatorname{SL}_2(\mathbb C) that are lifts of the same projective monodromy representation up to conjugation if and only if F_1 \cong F_2 \otimes N for some line bundle N with N^2 \cong \mathcal{O}_C.

Proof of Lemma 3.10

The condition \check{\rho}_1([\gamma]) = \pm \check{\rho}_2([\gamma]) for any closed path \gamma is equivalent to the fact that the two flat bundles F_1^{\nabla_1} and F_2^{\nabla_2} are such that F_1^{\nabla_1} \cong F_2^{\nabla_2} \otimes N where N is a flat line bundle with constant transition functions valued in \{\pm 1 \}. This occurs if and only if N^2 \cong \mathcal{O}_C.

Proposition

Proposition 3.11

H1G5

Suppose (F, \nabla) is a holomorphic connection and L \hookrightarrow F a subbundle not invariant by \nabla. Then the divisor of apparent singularities \operatorname{div}( (\mathbb{P}(F^\nabla), \mathbb{P}(L^\nabla))) of the projective connection (\mathbb{P}(F^\nabla), \mathbb{P}(L^\nabla)) coincides with the zero divisor of s_{L}(\nabla).

Proof of Proposition 3.11

See [Dinh24, Section 4.1] and [dinh-thesis, Chapter 5].

Corollary

Corollary 3.12

W6FD

Two projective connections with the same irreducible projective monodromy representations up to conjugation that have lifts to \operatorname{SL}_2(\mathbb C) and the same divisors of apparent singularities of degree < 2g-2 are isomorphic.

Proof of Corollary 3.12

Two such projective connections are the projectivization of a holomorphic connection (F,\nabla) and subbundles L_1, L_2 of F such that s_{L_1}(\nabla), s_{L_2}(\nabla) both vanish at \mathbf{x} and are identified via the isomorphism KL_1^{-2} \cong \mathcal{O}_C(\mathbf{x}) \cong KL_2^{-2}. But \deg(\mathbf{x}) < 2g-2, therefore L_1 and L_2 destabilize F and must be the same subbundle of F.

Definition

Definition 3.13

5E5Z

Given an irreducible \operatorname{SL}_2(\mathbb C)-holomorphic connection (F,\nabla) and a subbundle L \hookrightarrow F, the associated oper \mathcal{D}_{(L\hookrightarrow F, \nabla)} takes the local form \partial_z^2 + q(z) where q(z) is defined in (?). The divisor of apparent singularities \operatorname{div}(\mathcal{D}_{(L\hookrightarrow F, \nabla)}) coincides with the zero divisor of s_L(\nabla).

Proposition

Proposition 3.14

Y9HP

Let (F,\nabla) be an irreducible \operatorname{SL}_2(\mathbb C)-holomorphic connection and L \hookrightarrow F a subbundle. Then the associated oper \mathcal{D}_{(L \hookrightarrow F, \nabla)} is equivalent to the isomorphism class of the projective connection (\mathbb{P}(F^\nabla), \mathbb{P}(L^\nabla)).

Uses Definition 3.7 · Used by Terminology
Proof of Proposition 3.14

See [Dinh24, Proposition 4.1] and [dinh-thesis, Chapter 5].

Remark

Remark 3.15

VVB6

Twisting the data (L \hookrightarrow F, \nabla) by a square-root of \mathcal{O}_C induces equivalent opers. There exist in total 2^{2g} such data, which differ from each other by a twist by a square-root of \mathcal{O}_C, that define the same oper up to equivalence.

Terminology

Terminology

NIEC

By ? and the proposition above, a projective structure on C possibly with apparent singularities, an isomorphism class of projective connections, and an oper are equivalent descriptions of the same object, a \mathrm{PSL}_2(\mathbb C)-oper with apparent singularities in the terminology of [Dinh24]; we simply call it an oper, or an oper with apparent singularities for emphasis. When its monodromy lifts to \operatorname{SL}_2(\mathbb C), it is given by data (L \hookrightarrow F, \nabla), up to twisting by a square root of \mathcal{O}_C, which we call an \operatorname{SL}_2(\mathbb C)-oper with apparent singularities. We denote it by \mathcal{D} and use its local form \partial_z^2 + q(z) when computing in a local coordinate z.

Proposition

Proposition 4.1

6E8Q

Let (E, \phi) \in \mathcal{M}_H with associated non-degenerate spectral curve \widetilde{C} \overset{\pi}{\rightarrow} C, and L \hookrightarrow E a line subbundle with Baker–Akhiezer divisor \widetilde{D}. Let U \subset C be a simply connected open set that contains no branch point of \widetilde{C} and no point of \mathrm{Nm}(\widetilde{D}), and let U' \coloneqq \left\{ (x_+, x_-) \in U^2 \mid x_+ \neq x_- \right\}. Then there exist a holomorphic family \{(E_{x_+, x_-}, \phi_{x_+, x_-})\}_{(x_+, x_-) \in U'} in \mathcal{M}_H and line subbundles L_{x_+, x_-} \hookrightarrow E_{x_+, x_-} of degree \deg(L) - 1, depending holomorphically on (x_+, x_-), such that

  1. for all (x_+, x_-) \in U', the Baker–Akhiezer divisor of (L_{x_+, x_-} \hookrightarrow E_{x_+, x_-}, \phi_{x_+, x_-} ) is \widetilde{D} + \tilde{x}_+ + \tilde{x}_-, where \tilde{x}_\pm \in \pi^{-1}(x_\pm) lie in two fixed distinct components of \pi^{-1}(U);

  2. for all x_0 \in U, \underset{x_+ \rightarrow x_0 \leftarrow x_-}{\lim} [L_{x_+, x_-}] = [ L(-x_0)] in \mathrm{Pic}^{\deg(L)-1}(C) and \underset{x_+ \rightarrow x_0 \leftarrow x_-}{\lim}(E_{x_+, x_-}, \phi_{x_+, x_-}) = (E, \phi) in \mathcal{M}_H.

Uses Definition 3.1 · Used by Proof
Proposition

Proposition 4.2

IAZ8

Let (E, \phi) \in \mathcal{M}_H with associated non-degenerate spectral curve \widetilde{C} \overset{\pi}{\rightarrow} C, and let L \hookrightarrow E be a line subbundle with Baker–Akhiezer divisor \widetilde{D}. Given x_0 \in C, neither a branch point of \widetilde{C} nor a point of \mathrm{Nm}(\widetilde{D}), let (U, z) be a coordinate neighborhood of x_0, where z(x_0)=0, U is simply connected and contains no branch point of \widetilde{C} and no point of \mathrm{Nm}(\widetilde{D}). For u \in z(U), let x_\pm(u) be the points with z(x_\pm(u)) = \pm u. Then there exist a holomorphic family \{ (E_u, \phi_u)\}_{u \in z(U)} in \mathcal{M}_H with spectral curve \widetilde{C} and (E_0, \phi_0) = (E, \phi), and a holomorphic family of line bundles \{L_u\}_{u \in z(U)} of degree \deg(L) - 1 with L_0 \cong L(-x_0), such that, for all u \neq 0, L_u \hookrightarrow E_u is a line subbundle and the Baker–Akhiezer divisor of (L_u \hookrightarrow E_u, \phi_u) is \widetilde{D} + \tilde{x}_+(u) + \tilde{x}_-(u), where \tilde{x}_\pm(u) \in \pi^{-1}(x_\pm(u)) lie in two fixed distinct components of \pi^{-1}(U).

Proposition

Proposition 4.3

QL2I

Let (E, \phi) \in \mathcal{M}_H with the associated non-degenerate spectral curve \widetilde{C} \overset{\pi}{\rightarrow} C and let L \hookrightarrow E be a line subbundle such that s_L(\phi) has a double zero at x_0 \in C which is not a branch point of \widetilde{C}. Let \widetilde{D} be the Baker–Akhiezer divisor of (L \hookrightarrow E, \phi) and \tilde{x}_0 be the point with multiplicity 2 in \widetilde{D} with \pi(\tilde{x}_0) = x_0. Let (U,z) be a coordinate neighborhood of x_0, where z(x_0) = 0, U is simply connected and contains no branch point of \widetilde{C}. For u \in z(U), let x_\pm(u) be the points with z(x_\pm(u)) = \pm u. Then there exist a family of Higgs bundles \{ (E_u, \phi_u)\}_{u \in z(U)} and a family of line bundles \{L_u\}_{u \in z(U)} of the same degree as L such that

  1. [L_0] = [ L ] in \mathrm{Pic}^{\deg(L)}(C) and (E_0, \phi_0) = (E, \phi) in \mathcal{M}_H;

  2. for all u \in z(U), E_u admits L_u as a subbundle;

  3. for all u \neq 0, the Baker–Akhiezer divisor of (L_u \hookrightarrow E_u, \phi_u) is \widetilde{D} - 2\tilde{x}_0 + \tilde{x}_+(u) + \tilde{x}_-(u), where \tilde{x}_\pm(u) \in \pi^{-1}(x_\pm(u)) lie in the component of \pi^{-1}(U) containing \tilde{x}_0.

Proof of Proposition 4.3

The proposition follows from the discussion preceding Proposition 4.1 by requiring \tilde{x}_\pm now to be contained in the same component of \pi^{-1}(U) containing \tilde{x}_0.

Proposition

Proposition 5.1

F4NA

Let U \subset C be an open subset, i\in \{1, 2\}, d = 3g-4 + i, and \mathbf{x}^d: U \rightarrow C^{[d]}, x \mapsto \mathbf{x}^d(x) = x_1(x) + ... + x_d(x), be a holomorphic map such that, for all x\in U and r \in \{1 , ..., d\}, x_r(x) \neq x and \mathbf{x}^d(x) is Q-generic. Then there exists a family of meromorphic quadratic differentials \left\{ q^{(i)}_{x, \mathbf{x}^d} \right\}_{x \in U} holomorphically parameterized by U such that

  1. q^{(i)}_{x, \mathbf{x}^d} has a simple (double) pole at x if i=1 (respectively, i=2) and is holomorphic on C\setminus \{x\};

  2. \mathbf{x}^d(x) is contained in the support of the zero divisor of q^{(i)}_{x, \mathbf{x}^d}.

Proof of Proposition 5.1

See [dinh-thesis, Proposition 6.1].

Proposition

Proposition 5.2

TJLC

Let (U, z) be an open coordinated subset of C.
1. Let \mathbf{x}^{3g-3}: U \rightarrow C^{[3g-3]}, x \mapsto \mathbf{x}^{3g-3}(x) = x_1(x) + ... + x_{3g-3}, be a holomorphic map such that for all x\in U and r \in \{1, ..., 3g-3\}, x_r(x) \neq x and \mathbf{x}^{3g-3}(x) is Q-generic. Then there exists a unique family of meromorphic quadratic differentials \left\{ q^{(i)}_{z, x, \mathbf{x}} \right\}_{x \in U} holomorphically parameterized by U such that, for each x \in U, q^{(i)}_{z, x, \mathbf{x}^{3g-3}} satisfies conditions (a), (b) in Proposition 5.1, and furthermore takes the local form

\begin{aligned} \frac{dz^2}{( z - z(x))^i} + R^{(i)}_{z, x, \mathbf{x}^{3g-3}} (z) dz^2, \end{aligned}

where R^{(i)}_{z, x, \mathbf{x}^{3g-3}} (z) is the restriction of a holomorphic function R^{(i)}(z_1, z_2) defined on (U, z) \times (U, z) to the slice \{ z_1 = z(x), z_2 \equiv z \}.
2. Let \mathbf{x}^{3g-4}: U \rightarrow C^{[3g-4]}, x \mapsto \mathbf{x}^{3g-4}(x) = x_1(x) + ... + x_{3g-4}, be a holomorphic map such that for all x\in U and r \in \{1, ..., 3g-4\}, x_r(x) \neq x and x + \mathbf{x}^{3g-4}(x) is Q-generic. Then there exists a unique family of meromorphic quadratic differentials \left\{ q^{(i)}_{0, z, x, \mathbf{x}^{3g-4}} \right\}_{x \in U} holomorphically parameterized by U such that, for each x \in U, q^{(i)}_{0, z, x, \mathbf{x}^{3g-4}} satisfies conditions (a), (b) in Proposition 5.1, and furthermore takes the local form

\begin{aligned} \frac{dz^2}{\left( z - z(x) \right)^i} + R^{(i)}_{0, z, x, \mathbf{x}^{3g-4}}(z) dz^2, \end{aligned}

where R^{(i)}_{0, z, x, \mathbf{x}^{3g-4}} (z) is the restriction of a function R^{(i)}_0 (z_1, z_2), which is defined on (U, z_1) \times (U, z_2) and vanishes at the diagonal \{ z_1 = z_2 \}, to the slice \{ z_1 = z(x), z_2 \equiv z \}.

Proof of Proposition 5.2

See [dinh-thesis, Proposition 6.2].

Lemma

Lemma 5.3

DO2R

Fix i \in \{1,2\}.
1. The functions \frac{1}{u} R^{(i)}_{u, +}(-u) and -\frac{1}{u} R^{(i)}_{u,-}(u) are holomorphic on U with variable u = z(x_+). Furthermore, their evaluations at u = 0 are equal.
2. For each fixed z, the limit

\begin{aligned} \check{R}^{(i)} (z) \coloneqq \underset{u \rightarrow 0}{\lim} \frac{1}{u}\left( R^{(i)}_{u,+} (z) - R^{(i)}_{u,-}(z) \right) \end{aligned}

exists. These limits define a holomorphic function on U with variable z.

Lemma

Lemma 5.4

EGI1

Let x_0 and \mathbf{x'} = x_3 + ... + x_{3g-3} satisfy (?), and (U,z), x_1(u), x_2(u) be constructed as above. Then there exists a family \mathbf{q}_{u} \coloneqq \{ (q_{k, u})_{k = 1}^{3g-3} \}_{u \in z(U)} of basis of H^0(K^2) that is parameterized by U and is such that the function \det (\mathbf{q_{r,k}})(u), where \mathbf{q_{r,k}} = q_{k, u}(z_r(x_r)), has a simple zero at u = 0.

Used by Proof
Remark

Remark 5.5

2LUD

We have defined q^{(i)}_{u, \pm} upon choosing a divisor \textbf{w} = w_0 + w_{d+1} + ... + w_{3g-3}. Choosing a different choice of \mathbf{w} amounts to translating the definitions of q^{(i)}_{u, \pm} and q^{(i)}_{u, x_r} by a holomorphic quadratic differential that vanishes at \mathbf{x'} and x_+ + x_- + \mathbf{x_r'} respectively.

Lemma

Lemma 5.6

8NZK

Suppose \nu_3(u), ..., \nu_d(u): U \cong z(U) \rightarrow \mathbb C are holomorphic functions, and \nu_\pm(u): U \cong z(U) \rightarrow \mathbb C are meromorphic functions with simple poles and Laurent expansions at u = 0 of either the form (?) or (?). Then the induced non-homogeneous linear system (?) limits to a degenerate d \times (3g-3) linear system of rank d-1.

Used by Proof
Lemma

Lemma 5.7

XYBK

Let x_0, x_3, ..., x_{d} satisfy condition (?), \nu_0\in \mathbb C, and q \in H^0(K^2) a holomorphic quadratic differential. Then there exist holomorphic functions \nu_3(u), ..., \nu_{d}(u): z(U) \rightarrow \mathbb C and \nu_\pm(u): z(U\setminus \{x_0\}) \rightarrow \mathbb C with Laurent expansions at u =0 of the forms (?) or (?) (with \nu_0 as the coefficient of order u^0), and a family of holomorphic quadratic differentials \{q^{(0)}_u\}_{u\in z(U)} such that

  1. q^{(0)}_u solves the linear system (?) defined by \nu_\pm(u), \nu_3(u), ..., \nu_d(u) for u\neq 0;

  2. q^{(0)}_0 = q solves the corresponding limit linear system (Proof).

Used by Proof, Proof
Lemma

Lemma 5.8

K10M

Let x_0, x_3, ..., x_d be distinct points on C, (U,z) a coordinate neighborhood of x_0 with z(x_0) = 0, \nu_{\pm}(u) = \mp \frac{1}{4u} + \nu_0 \pm \nu' u + \mathcal{O}(u^2) holomorphic functions on U\setminus \{x_0 \}, and q^{(i)}_{u, \pm} defined as in (?) for u \in z\left( U \right), u \neq 0. Then the family of meromorphic quadratic differentials

\begin{aligned} &\sum_{s\in \{\pm\}} \nu_s(u) q^{(1)}_{u, s} - \frac{3}{4} \sum_{s\in \{\pm\}} q^{(2)}_{u, s}, \end{aligned}

which is parameterized by U \setminus \{x_0\}, extends to a family parameterized by U. The meromorphic quadratic differential q^{(2)}_{x_0} corresponding to the extension to x_0 is holomorphic on C \setminus \{x_0\}, vanishes at x_3, ..., x_{d}, and on U takes the form

\begin{aligned} q^{(2)}_{x_0} \equiv q^{(2)}_{x_0}(\nu_0) = \left[ - \frac{2 }{z^2} + \frac{2 \nu_0 }{z} + 2\nu_0 R^{(1)}_{0}(z) -\frac{1}{4} \check{R}^{(1)}(z) - \frac{3}{2}R^{(2)}_0(z) \right] dz^2, \end{aligned}

where \check{R}^{(1)}(z) is defined in part 2 of Lemma 5.3 and R^{(i)}_0(z) defined in (?).

Used by Proof, Proof
Proposition

Proposition 5.9

9AHG

Let x_0, x_3, ..., x_{d} be distinct points on C such that 2x_0 + x_3 + ... + x_d is Q-generic. Let z be a coordinate on U subordinate to the chosen holomorphic projective structure, with z(x_0) = 0. Let \nu_3(u), ..., \nu_d(u) be holomorphic functions on U, \nu_{\pm}(u) = \mp \frac{1}{4u} + \nu_0 \pm \nu' u + \mathcal{O}(u^2) holomorphic functions on U\setminus \{x_0 \}, and \{\mathcal{D}_u = \mathcal{D}_{pr} + q_u \}_{u \in z(U), u \neq 0} the corresponding family of opers parameterized by U\setminus \{x_0\} where q_u is defined as in (?), with \{ q^{(0)}_u \}_{u \in z(U), u \neq 0} the corresponding family of holomorphic quadratic differentials.

If there exists a holomorphic quadratic differential q^{(0)}_0 with q^{(0)}_0 = \underset{u \rightarrow 0}{\lim} q^{(0)}_u, then \{ \mathcal{D}_u \}_{u \in z(U), u\neq 0} extends to a family of opers parameterized by U. The oper corresponding to the extension to x_0 is \mathcal{D}_0 = \mathcal{D}_{pr} + q_0, where

\begin{aligned} q_0 \overset{\text{on } (U,z)}{=} {q^{(0)}_0} + q^{(2)}_{x_0}(\nu_0) + \sum_{r=3}^{d} \nu_r(0) q_{0, x_r}^{(1)} - \frac{3}{4} \sum_{r=3}^{d} q^{(2)}_{0, x_r}. \end{aligned}

(Recall the definition of q^{(2)}_{0, x_r} in (?).) In particular, the apparent singularities of \mathcal{D}_0 consist of a double apparent singularity at x_0 with accessory parameter \nu_0, and simple apparent singularities at x_3, ..., x_d with respective accessory parameters \nu_3(0), ..., \nu_d(0).

Uses Terminology · Used by Proof, Proof
Proof of Proposition 5.9

The existence of \mathcal{D}_0 = \mathcal{D}_{pr} + q_0 with q_0 of the form (?) follows from the definition (?) of q_u and Lemma 5.8. Let q^{reg}_0(z) be the regular part on U of q_0(z), i.e.

\begin{aligned} &q^{reg}_0(z) \coloneqq q_0(z) - \left(- \frac{2 }{z^2} + \frac{2 \nu_0 }{z} \right) dz^2 \\ = &\hspace{3pt} q^{(0)}_0(z) + \left[ 2\nu_0 R^{(1)}_{0}(z) -\frac{1}{4} \check{R}^{(1)}(z) - \frac{3}{2}R^{(2)}_0(z) \right] dz^2 + \sum_{r=3}^{d} \nu_r(0) q_{0, x_r}^{(1)} (z) - \frac{3}{4} \sum_{r=3}^{d} q^{(2)}_{0, x_r} (z). \end{aligned}

To prove the proposition, it remains to show that the coefficients of the Laurent tail of q_0(z) satisfy the condition making x_0 a double apparent singularity, i.e. for q^{reg}_0(z) = \sum_{k\geq 0} q_{0, k} z^k we have

\begin{aligned} \nu_0^3 + q_{0, 0} \nu_0 + \frac{q_{0,1}}{2} = 0. \end{aligned}

To this end, for each u\neq 0, let q^{reg}_u(z) be the regular part on U of q_u(z), i.e.

\begin{aligned} q^{reg}_u(z) &\coloneqq q_u(z) - \left( - \frac{3}{4(z-u)^2} - \frac{3}{4(z+u)^2} + \frac{\nu_+(u)}{z-u} + \frac{\nu_-(u)}{z+u} \right) \\ &= q^{(0)}_u (z) + \sum_{s \in \{\pm\}} \nu_s(u) R^{(1)}_{u,s}(z) - \frac{3}{4} \sum_{s \in \{\pm\}} R^{(2)}_{u,s} (z) + \sum_{r=3}^{d} \nu_r(u) q_{u, x_r}^{(1)}(z) - \frac{3}{4} \sum_{r=3}^{d} q^{(2)}_{u, x_r} (z). \end{aligned}

As a function of both u and z, the leading term in u of q^{reg}_u(z) comes from the leading terms in \nu_\pm(u) R^{(1)}_{u, \pm}(z) and is equal to -(R^{(1)}_{u,+}(z) - R^{(1)}_{u,-}(z) )/( 4u). By part 2 of Lemma 5.3, this has a well-defined u\rightarrow 0 limit. Hence q^{reg}_u(z) has a well-defined u\rightarrow 0 limit, and in fact q^{reg}_u(z) \overset{ u \rightarrow 0}{\longrightarrow} q^{reg}_0(z).

Now, this enables us to expand q^{reg}_u(z) = \sum_{m, n \geq 0} q_{m,n} u^m z^n, where letting m=0 gives the coefficients q_{0, n} of the Taylor expansion q^{reg}_0(z) = \sum_{n \geq 0} q_{0, n} z^n. Plugging this expansions of q^{reg}_0(z) and \nu_{\pm}(u) into condition (?) for z = \pm u,

\begin{aligned} &\nu_\pm(u)^2 - \frac{3}{4} \frac{1}{(2u)^2} + \frac{\nu_\mp(u)}{(\pm 2 u)} + q^{reg}_u(\pm u) = 0, &u \neq 0, \end{aligned}

we obtain a series of constraints by order of u starting from the order u^{-2}. The constraints up to order u^{-1} are automatically satisfied by the leading terms of \nu_\pm(u). Solving the constraints at order u^0 and u^1, we obtain (Proof).

Proposition

Proposition 5.10

2INS

Let \mathcal{D} be a branched projective structure with \mathrm{div}(\mathcal{D}) = 2x_0 + x_3 + ... + x_d being Q-generic and d \leq 3g-3. Then there exists a coordinate neighborhood (U, z) centred at x_0 and a family of branched projective structures \{\mathcal{D}_u\}_{u\in z(U)} parametrized by U such that

  1. \mathcal{D}_0 = \mathcal{D};

  2. for u\neq 0, \mathrm{div}(\mathcal{D}_u) = x_+ + x_- + x_3 + \dots x_d with x_\pm \in U and z(x_\pm) = \pm u;

  3. for u\neq 0, the accessory parameters \nu_\pm(u) of x_\pm w.r.t. the local coordinate z have Laurent expansions as functions of u of Form (D) (?), where 2\nu_0^\mathcal{D} is the accessory parameter of the double apparent singularity x_0 of \mathcal{D}.

Furthermore, this family defines via taking monodromy a holomorphic map U \rightarrow \mathrm{Hom}(\pi_1, \mathrm{PSL}_2), which is injective for d < 2g-2.

Proof of Proposition 5.10

Fix a holomorphic projective structure with corresponding oper \mathcal{D}_{pr}, and choose coordinate neighborhoods (U_r, z_r) of x_r for 3 \leq r \leq d and (U,z) of x_0 subordinate to this holomorphic projective structure, where U satisfies condition (?). We can write \mathcal{D} = \mathcal{D}_{pr} + q_{\mathcal{D}}, where q_\mathcal{D} is a meromorphic quadratic differential with local expression -\frac{3}{4(z-z_r(x_r))^2} dz_r^2 + \frac{\nu^\mathcal{D}_r}{z-z_r(x_r)} dz_r^2 + ... on U_{r}, 3 \leq r \leq d, and -\frac{2}{z^2} dz^2 + \frac{2 \nu^\mathcal{D}_0}{z} dz^2 + ... on U. Then there exists a unique holomorphic quadratic differential q^{(0)}_0 (\mathcal{D}) such that q_{\mathcal{D}} can be written in the form (?), i.e.

\begin{aligned} q_{\mathcal{D}} = q^{(0)}_0 (\mathcal{D}) + q^{(2)}_{x_0}(\nu^\mathcal{D}_0) + \sum_{r=3}^{d} \nu^\mathcal{D}_r q_{0, x_r}^{(1)} - \frac{3}{4} \sum_{r=3}^{d} q^{(2)}_{0, x_r}. \end{aligned}

By Lemma 5.7, there exist holomorphic functions \nu_3(u), ..., \nu_{d}(u): z(U) \rightarrow \mathbb C, \nu_\pm(u): z(U\setminus \{x_0\}) \rightarrow \mathbb C with Laurent expansions at u =0 of the forms (?) (in particular, with \nu_0^\mathcal{D} as the coefficient of order u^0), and a family \{q^{(0)}_u\}_{u\in z(U)} of holomorphic quadratic differentials with \underset{u \rightarrow 0}{\lim} q^{(0)}_u = q^{(0)}_0 (\mathcal{D}) such that

\begin{aligned} &q_u = q^{(0)}_u + \sum_{s\in\{ \pm\}} \nu_{s}(u) q^{(1)}_{u, s} - \frac{3}{4} \sum_{s\in \{\pm\}} q^{(2)}_{u, s} + \sum_{r=3}^{d} \nu_r(u) q_{u, x_r}^{(1)} - \frac{3}{4} \sum_{r=3}^{d} q^{(2)}_{u, x_r}, &u \neq 0, \end{aligned}

is a meromorphic quadratic differential with simple apparent singularities at x_3, ..., x_d and x_\pm \in U with z(x_\pm) = \pm u. By Proposition 5.9, q_0 \coloneqq \underset{u \rightarrow 0}{\lim} q_u = q_\mathcal{D}, and hence \{\mathcal{D}_u \coloneqq \mathcal{D}_{pr} + q_u \}_{u\in z(U)} defines a family we seek.

To see that taking the projective monodromy representation of \mathcal{D}_u defines a holomorphic map U \rightarrow \operatorname{Hom}(\pi_1, \mathrm{PSL}_2), consider a coordinate neighborhood (V, w), where V is distinct from U\cup \{x_3, ..., x_d\} and w is subordinate to the chosen holomorphic projective structure. On (V, w), \mathcal{D}_u takes the form \partial_w^2 + q_u(w) where q_u(w) is a function holomorphic on U \times V. By standard results on differential equations that vary holomorphically with respect to deformation parameters [ilyashenko-yakovenko], the local solutions to \partial_w^2 + q_u(w) are holomorphic functions on U \times V. Analytically continuing the ratio of two such linearly independent solutions defines the projective monodromy representation of \mathcal{D}_u, which is now holomorphic in u. By [dinh-thesis, Corollary 5.4] , if the number of apparent singularities is less than 2g-2 to start with, this holomorphic map is injective.

Lemma

Lemma 5.11

1DZA

Let x_0, x_3, ..., x_d be distinct points on C, (U,z) a coordinate neighborhood of x_0 with z(x_0) = 0, \nu_{\pm}(u) = \pm \frac{3}{4u} \pm \nu' u + \mathcal{O}(u^2) holomorphic functions on U \setminus \{x_0\}, and q^{(i)}_{u, \pm} defined as in (?) for u \in z\left( U \right), u \neq 0. Then the family of meromorphic quadratic differentials

\begin{aligned} \sum_{s\in \{\pm\}} \nu_s(u) q^{(1)}_{u, s} - \frac{3}{4} \sum_{s\in \{\pm\}} q^{(2)}_{u, s}, \end{aligned}

which is parameterized by U \setminus \{x_0\}, extends to a family parameterized by U. The quadratic differential \Delta q^{(0)}_{x_0} corresponding to the extension to x_0 is holomorphic on C, vanishes at x_3, ..., x_{d}, and on U takes the form

\begin{aligned} \Delta q^{(0)}_{x_0} = \left[ \frac{3}{4} \check{R}^{(1)}(z) - \frac{3}{2} R^{(2)}_0(z) \right] dz^2. \end{aligned}
Used by Proof, Proof
Proposition

Proposition 5.12

JTBF

Let x_0, x_3, ..., x_{d} be distinct points on C such that 2x_0 + x_3 + ... + x_d is Q-generic. Let z be a coordinate on U subordinate to the chosen holomorphic projective structure, with z(x_0) = 0. Let \nu_3(u), ..., \nu_d(u) be holomorphic functions on U, \nu_{\pm}(u) = \pm \frac{3}{4u} \pm \nu' u + \mathcal{O}(u^2) holomorphic functions on U\setminus \{x_0\}, and \{\mathcal{D}_u = \mathcal{D}_{pr} + q_u \}_{u \in z(U), u \neq 0} the corresponding family of opers parameterized by U\setminus \{x_0\} where q_u is defined as in (?), with \{ q^{(0)}_u \}_{u \in z(U), u \neq 0} the corresponding family of holomorphic quadratic differentials.

If there exists a holomorphic quadratic differential q^{(0)}_0 with q^{(0)}_0 = \underset{u \rightarrow 0}{\lim} q^{(0)}_u, then \{ \mathcal{D}_u \}_{u \in z(U), u\neq 0} extends to a family of opers parameterized by U. The oper corresponding to the extension to x_0 is \mathcal{D}_0 = \mathcal{D}_{pr} + q_0, where

\begin{aligned} q_0 \overset{\text{on } (U,z)}{=} {q^{(0)}_0} + \Delta q^{(0)}_{x_0} + \sum_{r=3}^{d} \nu_r(0) q_{0, x_r}^{(1)} - \frac{3}{4} \sum_{r=3}^{d} q^{(2)}_{0, x_r}, \end{aligned}

where \Delta q^{(0)}_{x_0} is defined as in (5.11). In particular, the apparent singularities of \mathcal{D}_0 are simple and located at x_3, ..., x_d with accessory parameters \nu_3(0), ..., \nu_d(0).

Proof of Proposition 5.12

The proof follows directly from Lemma 5.11. The apparency constraints (?) at x_3, \dots, x_d hold for all u \neq 0 and pass to the limit u \rightarrow 0 by continuity. In contrast to Proposition 5.9, there is no constraint to check at x_0, since \Delta q^{(0)}_{x_0} is holomorphic.

Remark

Remark 5.13 (Holomorphic dependence of g_u)

DDRL

For each u\neq 0, let q^{reg}_u(z) be the regular part on U of q_u(z), i.e.

\begin{aligned} q^{reg}_u(z) &\coloneqq q_u(z) - \left( - \frac{3}{4(z-u)^2} - \frac{3}{4(z+u)^2} + \frac{\nu_+(u)}{z-u} + \frac{\nu_-(u)}{z+u} \right) \\ &= q^{(0)}_u (z) + \sum_{s \in \{\pm\}} \nu_s(u) R^{(1)}_{u,s}(z) - \frac{3}{4} \sum_{s \in \{\pm\}} R^{(2)}_{u,s} (z) + \sum_{r=3}^{d} \nu_r(u) q_{u, x_r}^{(1)}(z) - \frac{3}{4} \sum_{r=3}^{d} q^{(2)}_{u, x_r} (z). \end{aligned}

As a function of both u and z, the leading term in u of q^{reg}_u(z) comes from the leading terms in \nu_\pm(u) R^{(1)}_{u, \pm}(z) and is equal to 3(R^{(1)}_{u,+}(z) - R^{(1)}_{u,-}(z) )/( 4u). By part 2 of Lemma 5.3, this has a well-defined u\rightarrow 0 limit; by the expansion in its proof, it is in fact holomorphic in (u,z). Hence q^{reg}_u(z) has a well-defined u\rightarrow 0 limit, and in fact q^{reg}_u(z) \overset{ u \rightarrow 0}{\longrightarrow} q^{reg}_0(z). Since moreover q^{(0)}_u extends holomorphically to u = 0 (the coefficients (Proof) have removable singularities, cf. Lemma 5.7), and \nu_r(u) and the building blocks depend holomorphically on u, the function g_u(z) \coloneqq q^{reg}_u(z) in (?) is holomorphic in both u and z.

Uses Lemma 5.3 · Used by Proof
Proposition

Proposition 5.14

665Y

Let \mathcal{D} be a branched projective structure with \mathrm{div}(\mathcal{D}) = x_3+ ... + x_{d} for d \leq 3g-3, and x_0\in C such that 2x_0 + x_3 + ... x_d is Q-generic. Then there exists a coordinate neighborhood (U, z) centred at x_0 and a family of branched projective structures \{\mathcal{D}_u\}_{u\in z(U)} parametrized by U such that

  1. \mathcal{D}_0 = \mathcal{D};

  2. for u\neq 0, \mathrm{div}(\mathcal{D}_u) = x_+ + x_- + x_3 + \dots x_d with x_\pm \in U and z(x_\pm) = \pm u;

  3. for u\neq 0, the accessory parameters \nu_\pm(u) of x_\pm w.r.t. the local coordinate z have Laurent expansions as functions of u of Form (B) (?).

Furthermore, this family defines via taking monodromy a holomorphic map U \rightarrow \operatorname{Hom}(\pi_1, \mathrm{PSL}_2), which is injective for d < 2g-2.

Proof of Proposition 5.14

The proof is the similar to that of Proposition 5.10.

Proposition

Proposition 5.15

HDEP

Let \{\mathcal{D}_u\}_{u \in z(U)} be a family as in Proposition 5.14 with d even, and suppose that the projective monodromy of \mathcal{D}_0 is irreducible. Then, after shrinking U, there exist a holomorphic family \{[E_u, \nabla_u]\}_{u \in z(U)} in \mathcal{M}_{dR} and line subbundles L_u \hookrightarrow E_u with \mathcal{D}_u = \mathcal{D}_{(L_u \hookrightarrow E_u, \nabla_u)}. In particular, for the closed strata (?),

[E_u, \nabla_u] \in \mathcal{M}^{\leq d}_{dR} \quad \text{for all } u \in z(U), \qquad \qquad [E_0, \nabla_0] \in \mathcal{M}^{\leq d-2}_{dR}.
Proof of Proposition 5.15

As the number of apparent singularities is even throughout the family \{\mathcal{D}_u\}, we can lift the projective monodromy \rho_u of \mathcal{D}_u to a holomorphic family \check{\rho}_u of \operatorname{SL}_2(\mathbb C)-representations. Since \rho_0 is irreducible and irreducibility is an open condition, \check{\rho}_u is irreducible for all u \in z(U) after shrinking U. Since the Riemann–Hilbert correspondence is biholomorphic on irreducible representations [simpson-moduli2], the flat connections [E_u, \nabla_u] with monodromy \check{\rho}_u form a holomorphic family. Since (E_u, \nabla_u) realizes the monodromy of \mathcal{D}_u, the oper \mathcal{D}_u is induced by a line subbundle L_u \hookrightarrow E_u, i.e. \mathcal{D}_u = \mathcal{D}_{(L_u \hookrightarrow E_u, \nabla_u)} (cf. ?). Note that

\deg(L_u^{-2}K) = \deg \mathrm{div}(\mathcal{D}_u) = \begin{cases} d, & u \neq 0, \\ d-2, & u = 0. \end{cases}
Corollary

Corollary 5.16

CGK1

Given d < 2g-2, the family [E_u, \nabla_u] of Proposition 5.15 satisfies

[E_u, \nabla_u] \in (N_d)^+_1 \quad \text{for } u \neq 0, \qquad \qquad \text{ and } \qquad [E_0, \nabla_0] \in (N_{d-2})^+_1 \subset \overline{(N_d)^+_1}.

Here the last inclusion is Simpson’s, recalled in ?.

Definition

Definition 6.1

79B0

Given a projective structure with developing map A: \widetilde{\Sigma} \rightarrow \mathbb{P}^1, we say that two paths \gamma_1 and \gamma_2 on \Sigma, both defined on the same interval [0, T] are twin paths if they overlap once developed, i.e.

  1. \gamma_1(0) = \gamma_2(0) is an apparent singularity;

  2. If \alpha: [-T, T] \rightarrow \Sigma is given by \alpha(t) = \gamma_1(t) for t \geq 0 and \gamma_2(-t) for t \leq 0, then the developed \overline{\alpha} of \alpha is even: \overline{\alpha}(t) = \overline{\alpha}(-t).

In particular, if \gamma_1 and \gamma_2 are embedded and disjoint except at \gamma_1(0) = \gamma_2(0) then we call them embedded twin paths.

Theorem

Theorem 6.2

RL72

Let \sigma' be a branched projective structure on \Sigma, and let \gamma_1, \gamma_2 be twin paths bounding a bubble \Delta. Then \sigma' = \mathrm{Bub}(\sigma, \gamma), where \sigma is obtained from \sigma' by collapsing \Delta and identifying \gamma_1(t) with \gamma_2(t), and \gamma is the resulting segment.

Uses Definition 6.1 · Used by Proof
Proof of Theorem 6.2

Since \gamma_1 and \gamma_2 have the same developed image, identifying \gamma_1(t) with \gamma_2(t) is compatible with the projective charts outside \Delta; at the two endpoints the cone angle drops from 4\pi to 2\pi, so they become regular points of \sigma, and \overline{\gamma} is the developed image of \gamma. Gluing \Delta back, a copy of \mathbb{P}^1 cut along \overline{\gamma}, along the arc \gamma is the bubbling of \sigma along \gamma.

Theorem

Theorem 6.3

XXP6

Let \{\mathcal{D}_u\} be a family of branched projective structures as in Proposition 5.14, with \mathrm{div}(\mathcal{D}_u) = x_+ + x_- + x_3 + \dots + x_d for u \neq 0. Then there exists \epsilon > 0 such that for every u with 0 < |u| < \epsilon, the structure \mathcal{D}_u has twin paths \gamma_{u,1} and \gamma_{u,2} from x_+ to x_- bounding a bubble. Consequently, \mathcal{D}_u = \mathrm{Bub}(\sigma_u, \gamma_u), where \sigma_u is a branched projective structure with apparent singularities x_3, \dots, x_d and the same monodromy as \mathcal{D}_u, on a Riemann surface with the same underlying smooth surface as C.

Uses Definition 6.1 · Used by Proof
Proposition

Proposition 6.4

7WX4

Let \{\mathcal{D}_u\} be the family of projective structures as in Proposition 5.14, with accessory parameters \nu_\pm(u) of Form B (?) and \mathcal{D} = \partial^2 + q_0 as the u \rightarrow 0 limit. Let \mathcal{Q} be the corresponding relative quadratic differential on S. Then for nonzero u and in local chart (u, w),

\beta^\ast \mathcal{Q}(u, w) = \left( \frac{-3/4}{(w-1)^2} + \frac{-3/4}{(w+1)^2} + \frac{u \nu_+(u)}{w-1} + \frac{u \nu_-(u)}{w+1} + u^2 g_u(uw) \right) dw^2, \qquad u\neq 0.

The extension of \beta^\ast \mathcal{Q} to \pi^{-1}(0) \simeq C \cup_{x_0} \mathbb{E} restricts to q_0 on C and

\widetilde{\mathcal{Q}}_0 \mid_{\mathbb{E}} (w) = \frac{-3}{(w-1)^2 (w+1)^2} dw^2

on the exceptional divisor \mathbb{E} \simeq \mathbb{P}^1.

Proof of Proposition 6.4

The local form (6.4) follows by plugging (?) into (?) and substituting z = uw. Here g_u(z) is holomorphic in (u, z) by Remark 5.13. Since Form (B) implies

u \nu_\pm(u) = \pm \frac{3}{4} \pm u^2 \nu' + \mathcal{O}(u^3),

we can compute the restriction of limit \underset{u \rightarrow 0}{\lim} \mathcal{Q}(u, w) to the open dense subset \{ w \in \mathbb{C} \} \subset \mathbb{E} directly by taking the u \rightarrow 0 limit of the expression in (6.4). This yields

\frac{-3}{4} \left[ \frac{1}{(w-1)^2} + \frac{1}{(w+1)^2} \right] + \frac{3}{4} \left[ \frac{1}{w-1} - \frac{1}{w+1} \right]

which is equal to (6.4)

Proposition

Proposition 6.5

H8GE
  1. J: \mathbb{P}^1 \rightarrow \mathbb{P}^1 has degree 2: for \xi \in \mathbb{C}, J(w) = \xi if and only if w^2 - \xi w + 1 = 0, and J^{-1}(\infty) = \{0, \infty\}.

  2. The critical points of J(w) are w = \pm 1, and the corresponding critical values are J(\pm 1) = \pm 2.

  3. J(\mathbb{S}^1) = [-2, 2] where \mathbb{S}^1 is the unit circle.

  4. The restriction of J to \mathrm{int} (\mathbb{D}_1) is a biholomorphism to \mathbb{P}^1 \setminus [-2, 2].

Used by Proof
Proposition

Proposition 6.6

8G8W

J(w) is a developing map of the differential equation defined by (6.4),

\left( \partial_w^2 - \frac{3}{(w-1)^2 (w+1)^2} \right) f(w) = 0.

In other words, there exists a basis of solutions f_1(w) and f_2(w) with J = f_1/f_2.

Proof of Proposition 6.6

We have J' = 1 - w^{-2} = \frac{w^2-1}{w^2}, J'' = \frac{2}{w^3} and J''' = -\frac{6}{w^4}, so that

\frac{J''}{J'} = \frac{2}{w(w^2-1)}, \qquad \frac{J'''}{J'} = -\frac{6}{w^2(w^2-1)}.

Hence

\{ J, w \} = \frac{J'''}{J'} - \frac{3}{2} \left( \frac{J''}{J'} \right)^2 = -\frac{6}{w^2(w^2-1)} - \frac{6}{w^2(w^2-1)^2} = -\frac{6}{(w-1)^2(w+1)^2},

i.e. \frac{1}{2}\{J, w\} is the potential of (6.4).

Proposition

Proposition 6.7

FEKG

Let \{\mathcal{D}_u\} be the family of projective structures as in Proposition 5.14. Fix R > 1, and consider \mathbb{D}_R = \{ \lvert w \rvert < R \} \subset \mathbb{E}. Then there exist \epsilon > 0 and a holomorphic map

\mathbf{A}: \mathbb{D}_\epsilon \times \mathbb{D}_R \longrightarrow \mathbb{P}^1, \qquad \qquad (u, w) \longmapsto A_u(w)

such that for all u \in \mathbb{D}_\epsilon, A_u is a developing map of the restriction of \beta^\ast \mathcal{D}_u to \mathbb{D}_R. In particular, A_u simply ramifies at w = \pm 1 and nowhere else on \mathbb{D}_R, and A_0 is Möbius equivalent to the Joukowsky function J = w + w^{-1}.

Used by Proof, Proof
Lemma

Lemma 6.8

UK28

For R > 0, if R - R^{-1} > M then \overline{\mathcal{B}} \subset \mathbb{D}_R.

Used by Proof
Proof of Lemma 6.8

Note that R > 2 since R - R^{-1} > M > 2. For |w| \geq R we have

|J(w)| = |w + w^{-1}| \geq |w| - |w|^{-1} \geq R - R^{-1} > M,

since t \mapsto t - t^{-1} is increasing. Hence J^{-1}(\overline{\mathbb{D}}_M) \subset \mathbb{D}_R, and in particular \overline{\mathcal{A}} \subset \mathbb{D}_R. Since c(\theta) \geq M^2 - 2 > 2, we have \rho > 1, and so \overline{\mathcal{H}} \subset \overline{\mathbb{D}}_1 \subset \mathbb{D}_R. Therefore \overline{\mathcal{B}} = \overline{\mathcal{A}} \cup \overline{\mathcal{H}} \subset \mathbb{D}_R.

Proposition

Proposition 6.9

8661

There exists a normalized family \mathbf{A} as in Proposition 6.7 with A_0 = J and such that A_u converges uniformly to J as u \rightarrow 0 on \overline{\mathcal{B}} with respect to the chordal metric.

Used by Proof, Proof
Proof of Proposition 6.9

Choose R in Proposition 6.7 with R - R^{-1} > M, so that \overline{\mathcal{B}} \subset \mathbb{D}_R by Lemma 6.8. Let T be the Möbius transformation with T \circ A_0 = J. Since developing maps are defined up to post-composition with Möbius transformations, (u, w) \mapsto T \circ A_u(w) is again a family as in Proposition 6.7; replacing A_u by T \circ A_u, we may assume A_0 = J. The map \mathbf{A} is continuous on the compact set \overline{\mathbb{D}}_{\epsilon/2} \times \overline{\mathcal{B}}, hence uniformly continuous there with respect to d by the Heine–Cantor theorem [Rudin76, Theorem 4.19]. In other words, for every \eta > 0 there is \tau > 0 with d(A_u(w), J(w)) = d(A_u(w), A_0(w)) < \eta for all |u| < \tau and w \in \overline{\mathcal{B}}.

Lemma

Lemma 6.10

GSK0

Let 0 \leq s < r, and a, b \in \mathbb{P}^1 with |a| \leq s and |b| \geq r. Then d(a, b) \geq d(s, r).

Used by Proof
Proof

Proof

7RXY

Put x = |a| and y = |b|, so that x \leq s < r \leq y. If b = \infty, then d(a, b) = 2/\sqrt{1+x^2} \geq 2/\sqrt{1+s^2} > d(s, r), since r - s < \sqrt{1+r^2}. Otherwise |a - b| \geq y - x > 0, and (1+x^2)(1+y^2) = (1+xy)^2 + (y-x)^2 gives

d(a, b) \geq \frac{2(y-x)}{\sqrt{(1+xy)^2 + (y-x)^2}} = h\big(\tau(x,y)\big), \qquad \tau(x, y) \coloneqq \frac{y-x}{1+xy}, \qquad h(\tau) \coloneqq \frac{2}{\sqrt{1+\tau^{-2}}}.

Since h is increasing and d(s, r) = h(\tau(s, r)), it remains to note that \tau decreases in x and increases in y:

\tau(x, y) - \tau(s, y) = \frac{(s-x)(1+y^2)}{(1+xy)(1+sy)} \geq 0, \qquad \tau(s, y) - \tau(s, r) = \frac{(y-r)(1+s^2)}{(1+sy)(1+sr)} \geq 0.
Theorem

Theorem 6.11

Q0YY

(Rouché) Let \Omega \subset \mathbb{C} be a regular domain, and f, g holomorphic on an open neighborhood of \overline{\Omega} with |f-g| < |f| on \partial \Omega. Then the number of zeroes of f on \Omega, counted with multiplicity, is equal to that of g.

Used by Proof
Lemma

Lemma 6.12

3OHV

Let \mathbf{A} be as in Proposition 6.9. Then there exists \epsilon such that for all u \in \mathbb{D}_\epsilon,

A_u: \mathcal{W}_u \rightarrow \mathbb{D}_{M'}

is a branched double cover ramified exactly at w = \pm 1.

Used by Proof
Proof of Lemma 6.12

Let E(u) \coloneqq \underset{w \in \overline{\mathcal{B}}}{\sup} d\left( A_u(w), J(w) \right). By uniform convergence on \overline{\mathcal{B}} (cf. Proposition 6.9), there exists \epsilon such that for all u \in \mathbb{D}_\epsilon we have

E(u) < \min \left\{d(M', M), d(M, M+1), \frac{M-M'}{\kappa} \right\}, \qquad \kappa = \frac{1}{2}\left( 1 + (M+1)^2 \right).

For a fixed u \in \mathbb{D}_\epsilon, we claim that

  1. |A_u| > M' on \overline{\mathcal{B}} \setminus \mathcal{A} = \mathcal{H} \sqcup \partial \mathcal{A};

  2. |A_u| < M+1 on \overline{\mathcal{A}};

  3. \underset{\overline{\mathcal{A}}}{\sup}|A_u - J| < M-M'.

To see (a), suppose that there exists w \in \overline{\mathcal{B}} \setminus \mathcal{A} with |A_u(w)| \leq M'. Since |J(w)| \geq M, it follows from Lemma 6.10 that d\left( A_u(w), J(w) \right) \geq d(M', M), which contradicts the bound (Proof). To see (b), suppose that there exists w \in \overline{\mathcal{A}} with |A_u(w)| \geq M+1. Since |J(w)| \leq M, it follows again from Lemma 6.10 that d\left( A_u(w), J(w) \right) \geq d(M, M+1), which contradicts the bound (Proof). To see (c), note that one can write

|A_u - J| = \frac{1}{2} \sqrt{(1+|A_u|^2) (1+ |J|^2)} \; d(A_u, J).

It follows from (b) that |A_u - J| \leq \kappa \, d(A_u, J) on \overline{\mathcal{A}}. Part (c) now follows from the fact that we have chosen \epsilon such that d(A_u(w), J(w)) \leq E(u) < (M-M')/\kappa for all w \in \overline{\mathcal{A}}.

Observe that it follows from (a) that \mathcal{W}_u \subset \overline{\mathcal{W}_u} \subset \mathcal{A}. Properness of the map A_u: \mathcal{W}_u \rightarrow \mathbb{D}_{M'} follows from (a) and the compactness of \overline{\mathcal{B}}. Indeed, let K \subset \mathbb{D}_{M'} be compact. By (a), we have A_u^{-1}(K) \cap \mathcal{W}_u = A_u^{-1}(K) \cap \overline{\mathcal{B}}, which is a closed subset of \overline{\mathcal{B}} and hence compact. To compute the degree of A_u, for \xi \in \mathbb{D}_{M'}, let f = J - \xi and g = A_u - \xi. On \partial \mathcal{A} where |J| = M, we have

|f - g| = |A_u - J| < M - M' < M - |\xi| = |J| - |\xi| \leq |J - \xi| = |f|.

By (b), A_u has no poles near \overline{\mathcal{A}}, so f and g are holomorphic on a neighborhood of \overline{\mathcal{A}}. We therefore can apply Rouché Theorem 6.11 to conclude that the number of preimages of \xi via A_u is equal to that via J on \mathcal{A}, which is 2 by Proposition 6.5(i). Since \mathcal{W}_u = A_u^{-1}(\mathbb{D}_{M'}) \cap \mathcal{B} = A_u^{-1}(\mathbb{D}_{M'}) \cap \mathcal{A} by (a), the preimages of \xi \in \mathbb{D}_{M'} in \mathcal{A} are contained in \mathcal{W}_u. Finally, by continuity of \mathbf{A} and since the branch locus of A_0 = J is J(\pm 1) = \pm 2, we can further shrink \epsilon if necessary to have |A_u(\pm 1)| < M'. Then the ramification loci \{ w = \pm 1 \} \subset \mathcal{A} are contained in \mathcal{W}_u. By construction of the family (Proposition 6.7), A_u ramifies at \pm 1 and nowhere else.

Proof

Proof (-bubble)

QLKV

We consider the holomorphic family of developing maps in Proposition 6.9 with \epsilon given in Lemma 6.12. Fix nonzero u \in \mathbb{D}_\epsilon, and let y_{u, \pm} \coloneqq A_u(\pm 1) \in \mathbb{D}_{M'} be the branch points of A_u. Let \overline{\gamma}_u be the straight segment connecting y_{u, \pm}: by construction, we have \overline{\gamma}_u \subset \mathbb{D}_{M'}. Then the preimage A_u^{-1}(\overline{\gamma}_u) consists of exactly two segments \gamma_{u,1} and \gamma_{u,2} that connect the critical points w = \pm 1 and develop to the same image A_u(\gamma_{u,1}) = A_u(\gamma_{u,2}) = \overline{\gamma}_u.

Let \Delta_u be the bounded component of \mathbb{C} \setminus (\gamma_{u,1} \cup \gamma_{u,2}). Then \overline{\Delta_u} \subset \mathcal{B} and A_u(\Delta_u) \subset \mathbb{P}^1 \setminus \overline{\gamma}_u. The map A_u: \Delta_u \rightarrow \mathbb{P}^1 \setminus \overline{\gamma}_u is proper and a local homeomorphism, since w = \pm 1 are the only critical points of A_u on \mathbb{D}_R. This means \Delta_u is a covering of \mathbb{P}^1 \setminus \overline{\gamma}_u. Since both of these domains are connected and simply connected, A_u must map \Delta_u biholomorphically onto \mathbb{P}^1 \setminus \overline{\gamma}_u.

We now recall that the coordinates change z \rightarrow w = z/u is a Möbius transform for nonzero u. In other words, A_u(z/u) is a developing map of \mathcal{D}_u. Hence the images of \gamma_{u,1} and \gamma_{u,2} under the inverse Möbius transform are twin paths connecting the apparent singularities x_\pm of \mathcal{D}_u, bounding a disc that develops biholomorphically onto \mathbb{P}^1 \setminus \overline{\gamma}_u. Let \sigma_u be the projective structure obtained from \mathcal{D}_u by collapsing the disc \Delta_u, and \gamma_u be the segment on \sigma_u along which the two twin paths are glued, which develops onto \overline{\gamma}_u. By Theorem 6.2, we have \mathcal{D}_u = \mathrm{Bub}(\sigma_u, \gamma_u).

Proof

Proof (-linear-system-right-side)

3793
Proof.

1. We need to show that the u\rightarrow 0 limit of \frac{1}{u} R^{(i)}_{u,+}(-u) and -\frac{1}{u} R^{(i)}_{u,-}(u) exist and are equal. Consider the Taylor expansion of R^{(i)}_{0}(z_1, z) = \sum_{m, n \geq 0} R^{(i)}_{m,n} z_1^m z^n at its zero (z_1, z) = (0,0). Since R^{(i)}_{0,0} = 0,

\begin{aligned} R^{(i)}_{u,-}(u) &\coloneqq R^{(i)}_{0}(-u, u) = - \left( R_{1,0} - R_{0,1} \right) u + \mathcal{O}(u^2), \\ R^{(i)}_{u,+}(-u) &\coloneqq R^{(i)}_{0}(u, -u) = \left( R_{1,0} - R_{0,1} \right) u + \mathcal{O}(u^2). \end{aligned}

It follows that

\begin{aligned} \underset{u \rightarrow 0}{\lim} \frac{1}{u} R^{(i)}_{u,+}(-u) = - \underset{u \rightarrow 0}{\lim} \frac{1}{u} R^{(i)}_{u,-}(u). \end{aligned}

2. For each fixed z, it follows from the Taylor expansion R^{(i)}_{0}(z_1, z) = \sum_{m, n \geq 0} R^{(i)}_{m,n} z_1^m z^n that,

\begin{aligned} R^{(i)}_{u,+}(z) - R^{(i)}_{u,-}(z) = 2u \sum_{n\geq 0} R^{(i)}_{1,n}z^n + 2u^3 \sum_{n\geq 0} R^{(i)}_{3,n} z^n + 2u^5 \sum_{n\geq 0} R^{(i)}_{5,n} z^n + ... \end{aligned}

is a holomorphic function on (U,u) that vanishes at u = 0; we can calculate

\begin{aligned} \underset{u \rightarrow 0}{\lim} \frac{1}{u}\left( R^{(i)}_{u,+}(z) - R^{(i)}_{u,-}(z) \right) = 2 \sum_{n\geq 0} R^{(i)}_{1,n}z^n. \end{aligned}
Proof

Proof (-basis-of-quadratic-differentials)

OP3A
Proof.

For each k in the range 3 \leq k \leq 3g-3 and each u \in z(U), there exists a unique up to scaling quadratic differential that vanishes at x_1(u), x_2(u) and all x_{r} with r \neq k. By our assumption on x_0 and \mathbf{x'}, this quadratic differential does not vanish at x_k. By scaling, for 3 \leq k \leq 3g-3, we define a unique quadratic differential q_{k,u} with q_{k, u} \left( z_r(x_r) \right) = \delta_{kr}. We now define q_1 \equiv q_{1,u} and q_2 \equiv q_{2,u} independent of u as follows. Let q_1 be a (unique up to scaling) holomorphic quadratic differential that vanishes at x_0 + \mathbf{x'}. By our assumption on x_0 and \mathbf{x'}, its zero at x_0 is simple. For some x_0' \in U with x_0' \neq x_0, let q_2 be a (unique up to scaling) holomorphic quadratic differential that vanishes at x_0' + \mathbf{x'}. By our assumption on x_0 and \mathbf{x'}, q_2 does not vanish at x_0 and hence is linearly independent from q_1. Then for each u \in z(U), let \left( q_{k, u} \right) \coloneqq \left( q_1, q_2, q_{3, u}, ..., q_{3g-3, u} \right): this forms a basis of H^0(K^2) for each u\in z(U). It follows that the function

\begin{aligned} \det (\mathbf{q_{r,k}})(u) &= \begin{vmatrix} &q_1(z(x_1)) & &q_2(z(x_1)) & & 0 \\ &q_1(z(x_2)) & &q_2(z(x_2)) & & 0 \\ & 0 & & 0 & & I_{(3g-5) \times (3g-5)} \end{vmatrix} \\ &= q_1(z(x_1)) q_2(z(x_2)) - q_1(z(x_2)) q_2(z(x_1)) \\ &= 2 q_{1}'(z(x_0)) q_{2}(z(x_0)) u + \mathcal{O}(u^2) \end{aligned}

has a simple zero at u=0.

Proof

Proof (-limit-linear-system)

JYU7
Proof.

This can be checked explicitly by plugging (?) and (?) in (?). Specifically, the coefficients at the orders u^{-1} and u^0 of \nu_\pm(u) ensure that C_\pm(u) are regular at u=0, while the coefficients at the order u^1 of \nu_\pm(u) together with part 1 of Lemma 5.3 ensure that C_+(0) = C_-(0) = C_0, where

\begin{aligned} C_0 \coloneqq \begin{cases} -\nu' - \frac{3}{4} \underset{u \rightarrow 0}{\lim} \frac{1 }{u} R^{(1)}_{u,-}(u) &\text{ for } \nu_\pm(u) = \pm \frac{3}{4u} \pm \nu' u + \mathcal{O}(u^2), \\ -\nu_0^2 + \nu' - \frac{1}{4} \underset{u \rightarrow 0}{\lim} \frac{1}{u} R^{(1)}_{u, -}(u) &\text{ for } \nu_\pm(u) = \mp \frac{1}{4u} + \nu_0 \pm \nu' u + \mathcal{O}(u^2). \end{cases} \end{aligned}

Then the linear system (?) limits to

\begin{aligned} &q_u^{(0)}(0) = C_0, &q_{u}^{(0)} (z_r(x_r)) = - \nu_r(0)^2, \end{aligned}

for 3 \leq r \leq d. Note that the dimension of Q_{x_0 + \mathbf{x'}} must have the minimal value 3g- 2 - d since otherwise condition (?) will be violated. This yields the rank d-1 of the system (Proof).

Proof

Proof (-tune-limit-hol-q)

M1WA
Proof.

We will make use of the representation (?) of (?) using a family \mathbf{q}_{u} = \{ (q_{k, u})_{k = 1}^{3g-3} \}_{u \in z(U)} of basis of H^0(K^2) constructed in Lemma 5.4. Let us use the basis corresponding to u=0 and expand q = \sum_{k=1}^{3g-3} E^0_k q_{k, 0}. We need to show that there exist \nu_\pm(u) of the prescribed forms such that the solutions to (?),

\begin{aligned} &E_1(u) = \frac{C_+(u) q_2(-u) - C_-(u) q_2(u)}{q_1(u) q_2(-u) - q_1(-u) q_2(u)}, &E_2(u) = \frac{ q_1(u) C_-(u) - q_1(-u) C_+(u) }{q_1(u) q_2(-u) - q_1(-u) q_2(u)}, \end{aligned}

where C_\pm(u) are defined as in (?), have well-defined u \rightarrow 0 limits which are equal to E_1^0 and E_2^0 respectively. To this end, denote by N_1(u) and N_2(u) the respective numerators in (Proof). Since \underset{u\rightarrow 0}{\lim} C_\pm(u) = C_0 (cf. (Proof)), both N_1(u) and N_2(u) are holomorphic and vanish at u = 0. Since the denominators of (Proof) have simple zero at u=0 by Lemma 5.4, the u\rightarrow 0 limits of E_1(u) and E_2(u) are well-defined by L’Hôpital’s rule. These limits are determined by the coefficient at order u^1 of N_1(u) and N_2(u). With R^{(i)}_{u, \pm}(z) = \sum_{m,n} R^{(i)}_{m,n} (\pm u)^m z^n, one can compute explicitly these coefficients to be

\begin{aligned} N_1(u) &= \frac{q_2(0)}{2}\left[ -4 (\nu_{+,2} + \nu_{-,2}) + 3 \left( R^{(2)}_{0,1} - R^{(2)}_{1,0} + R^{(1)}_{2,0} - R^{(1)}_{1,1} + R^{(1)}_{0,2} \right) \right] u \\ & \hspace{6mm} + \frac{q_2'(0)}{2} \left( 4\nu_1 + 3 R^{(1)}_{1,0} - 3 R^{(1)}_{0,1} \right)u + \mathcal{O}(u^2), \\ N_2(u) &= - \frac{q_1'(0)}{2} \left[ 4\nu_1 + 3 R^{(1)}_{1,0} - 3 R^{(1)}_{0,1} \right] u + \mathcal{O}(u^2) \end{aligned}

for the ansatz \nu_\pm(u) = \pm \frac{3}{4u} \pm \nu_1 u + \nu_{\pm, 2}u^2 + \mathcal{O}(u^3), and

\begin{aligned} N_1(u) &= -\frac{q_2(0)}{2} \left[ 8\nu_0 \nu_1 + 4\nu_0 \left( R^{(1)}_{0,1} - R^{(1)}_{1,0} \right) + R^{(1)}_{2,0} - R^{(1)}_{1,1} + R^{(1)}_{2,0} + 3 R^{(2)}_{1,0} - 3 R^{(2)}_{0,1} \right]u \\ &\hspace{6mm} + \frac{q_2'(0)}{2} \left[ 4\nu_0^2 - 4\nu_1 + R^{(1)}_{0,1} - R^{(1)}_{1,0} \right]u +\mathcal{O}(u^2) \\ N_2(u) &= \frac{q_1'(0)}{2} \left( -4 \nu_0^2 + 4 \nu_1 - R^{(1)}_{0,1} + R^{(1)}_{1,0} \right) u + \mathcal{O}(u^2) \end{aligned}

for the ansatz \nu_\pm(u) = \mp \frac{1}{4u} + \nu_0 \pm \nu_1 u + \nu_{\pm, 2}u^2 + \mathcal{O}(u^3).

Since q_2(0) and q_1'(0) are non-zero, we can tune (\nu_1, \nu_{\pm, 2}) in the first case and (\nu_0, \nu_1) in the second case to tune the coefficients of order u^1 of N_1(u) and N_2(u) as we want. Hence there exist holomorphic functions \nu_\pm(u): U\setminus \{x_0\} \rightarrow \mathbb C with the Laurent expansion at u = 0 of the form (?) or (?), so that \underset{u \rightarrow 0}{\lim} E_1(u) = E_1^0 and \underset{u \rightarrow 0}{\lim} E_2(u) = E_2^0. For 3 \leq r \leq d, let \nu_r \in \mathbb C be such that E^0_{r} = -\nu_r^2, and define \nu_r(u) = \nu_r for all u\in z(U). The tuple (E_1(u), E_2(u), E_3(u), ..., E_{3g-3}(u) ) = (E_1(u), E_2(u), E^0_3, ..., E^0_{3g-3}) then defines for each u \neq 0 a holomorphic quadratic differential q^{(0)}_u \coloneqq \sum_{k=1}^{3g-3} E_k(u) q_{k,u} that solves the linear system (?) defined by (\nu_\pm(u), \nu_3(u), ..., \nu_d(u)), with \underset{u \rightarrow 0}{\lim} q^{(0)}_u = q a solution to the corresponding limit linear system (Proof).

Proof

Proof (-db-limit-1)

TG2Q
Proof.

We will show explicitly that, as a multi-variable function with variables u and z, (5.8) limits to (5.8) as u \rightarrow 0. To this end, it suffices to show that given any z \neq 0, the u \rightarrow 0 limit of the evaluation of (5.8) at z is the evaluation of (5.8) at z. Plugging \nu_{\pm}(u) = \mp \frac{1}{4u} + \nu_0 \pm \nu' u + \mathcal{O}(u^2) into (5.8) yields

\begin{aligned} \frac{-2z^2 - u^2}{(z-u)^2 (z+u)^2} + \frac{2\nu_0 z}{(z-u)(z+u)} - \frac{R^{(1)}_{u,+}(z) - R^{(1)}_{u,-}(z) }{ 4u } + \nu_0 \left( R^{(1)}_{u,+}(z) + R^{(1)}_{u,-}(z) \right) & \\ - \frac{3}{4} \left( R^{(2)}_{u,+}(z) + R^{(2)}_{u,-}(z) \right) + \mathcal{O}(u)(z) &. \end{aligned}

Here we have ignored the factor dz^2 and denoted by \mathcal{O}(u)(z) a function with variables u and z such that, given any z' \neq 0, the function F_{z'}(u) \coloneqq \mathcal{O}(u)(z) \mid_{z = z'} is holomorphic on \{ u \mid 0 \leq |u| \leq |z'|/2 \} with F_{z'}(u) \sim \mathcal{O}(u). It follows that \underset{u \rightarrow 0}{\lim} \mathcal{O}(u)(z) = \mathcal{O}(u)(z)\mid_{u = 0}, which a priori defines a function in z, is the zero function. By a similar argument, we can take the u \rightarrow 0 limit of the other terms in (Proof) explicitly and obtain q^{(2)}_{x_0} as in (5.8) as the limit.

To show that q^{(2)}_{x_0} is holomorphic on C \setminus \{x_0\}, it suffices to show holomorphicity on C\setminus U. Given p\in C\setminus U and a coordinate neighborhood (V,w) of p, let q^{(i)}_{u,\pm}(w) dw^2 be the local form of q^{(i)}_{u,\pm} on V. We want to show that the evaluation of (5.8) at p, which defines the function

\begin{aligned} \sum_{s\in \{\pm\}} \nu_s(u) q^{(1)}_{u, s}(w(p)) - \frac{3}{4} \sum_{s\in \{\pm\}} q^{(2)}_{u, s}(w(p)) = \frac{1}{4u}\left( -q^{(1)}_{u,+}(w(p)) + q^{(1)}_{u,-} (w(p)) \right) + \mathcal{O}(1) \end{aligned}

in u which is holomorphic at u \neq 0 has a well-defined u \rightarrow 0 limit. Since q^{(i)}_{u,\pm} \overset{u \rightarrow 0}{\rightarrow} q^{(i)}_0, which is holomorphic on C \setminus \{x_0\}, it suffices to show that the leading term \frac{1}{4u}\left( -q^{(1)}_{u,+}(w(p)) + q^{(1)}_{u,-} (w(p)) \right) of (?) has a well-defined u \rightarrow 0 limit. This is achieved by an argument similar to the proof of part 2 of Lemma 5.3. Furthermore, for p \in \{x_3, ..., x_d\}, since (?) vanishes for u\neq 0, it is identically zero for all u \in z(U).

Proof

Proof (-bb-limit-1)

EB8I
Proof.

The proof is the same as that of Lemma 5.8, with \nu_\pm(u) of Form (B) in place of Form (D). The only difference is the polar part on U: plugging \nu_\pm(u) = \pm \frac{3}{4u} + \mathcal{O}(u) into (5.11) gives, up to dz^2,

-\frac{3}{4}\left[ \frac{1}{(z-u)^2} + \frac{1}{(z+u)^2} \right] + \frac{3}{4u}\left[ \frac{1}{z-u} - \frac{1}{z+u} \right] + \mathcal{O}(u) = - \frac{3u^2}{(z^2-u^2)^2} + \mathcal{O}(u),

which tends to 0 for every fixed z \neq 0, instead of to -2/z^2 as in (Proof). Hence the limit (5.11) has no pole at x_0, and it is holomorphic on U by Lemma 5.3.

Source and licence

The text on this page is from arXiv:2610.22326v1 by Duong Dinh, Troy Figiel and Joerg Teschner, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.

The record of each item is in analysis/objects/2610/2610.22326.jsonl.