betaFeedback

WorksMathematics

A local limit theorem for singularity visits on translation surfaces

Abstract

Let (X,ω)(X,ω) be a translation surface with at least one singularity. We investigate the distribution of the number of singularity encounters along prime closed geodesics on XX. Writing ℓ(γ)\ell(γ) for the geometric length of a closed geodesic and m(γ)m(γ) for the number of singularities it encounters, we establish a uniform counting local limit theorem for the values of m(γ)m(γ) sampled over ℓ(γ)<T\ell(γ)<T. The key technical step is to introduce and study a trace-class operator whose spectral data encodes the Fourier transform of our counting distribution. Our results hold for every translation surface, without genericity assumptions.

Introduction

A translation surface is a compact Riemann surface XX equipped with a (non-trivial) holomorphic one-form ω\omega. In other words, XX is a compact surface with a flat metric except at a finite set X={x1,…,xn}\mathcal{X}=\{x_1,\ldots,x_n\} of singular points with cone angles 2π(k(xi)+1)2\pi(k(x_i)+1), where k(xi)∈Nk(x_i)\in\mathbb{N}, for i=1,…,ni=1,\ldots,n.

Saddle connections, namely straight line (geodesic) segments joining singularities without any other singularities in their interiors, are fundamental objects in the study of translation surfaces; see the survey [Wri15]. Their counting theory goes back to Masur’s quadratic bounds [17] and includes the asymptotic results of Eskin-Masur [11], Eskin-Masur-Zorich [12] and many other beautiful works [1, 19, 2, 21]. These results concern counting problems for individual saddle connections. In this work we instead consider closed saddle geodesics: locally length-minimising closed paths, up to cyclic reordering, that pass through singularities.

For such a geodesic γ\gamma there are two natural and quite different notions of length. The first is its geometric length ℓ(γ)\ell(\gamma), measured in the flat metric. The second is the number m(γ)m(\gamma) of singularity encounters in one period, counted with multiplicity. Comparing combinatorial and geometric lengths is a central theme in various geometric contexts. In the classical setting of negatively curved surfaces (and more generally manifolds), counting results comparing such lengths have been obtained by many authors, [18, 16, 24 BL98, 22, 6]. The singularities on (X,ω)(X,\omega) can be thought of as concentrated points of negative curvature and it is therefore very natural to try to obtain statistical results, analogous to those for negatively curved surfaces, for translation surfaces.

Cantrell-Pollicott [7] recently established counting limit laws for saddle-connection paths on translation surfaces (extending the work of Colenese-Pollicott [CP21]). To recall their result, let S\mathcal{S} denote the saddle connections of (X,ω)(X,\omega). For s∈Ss\in\mathcal{S} let i(s),t(s)∈Xi(s),t(s)\in\mathcal{X} denote the initial and terminal singularity of ss. A saddle connection path is a string (s1,…,sn)(s_1,\ldots,s_n) of saddle connections forms a locally length minimising path in (X,ω)(X,\omega) if t(sj)=i(sj+1)t(s_j)=i(s_{j+1}) for each j=1,…,n−1j=1,\ldots,n-1 and the angle between sj,sj+1s_j,s_{j+1} is at least π\pi (see § 4 of [7]). The geometric length of p=(s1,…,sn)p=(s_1,\ldots,s_n) is its distance traversed in (X,ω)(X,\omega): ℓ(p)=ℓ(s1)+⋯+ℓ(sn)\ell(p)=\ell(s_1)+\cdots+\ell(s_n). Whereas m(p)=nm(p)=n is the number of saddle connections making up pp. For x∈Xx\in\mathcal{X}, let Px\mathcal{P}_x denote the collection of saddle connection paths starting at xx. Let νT\nu_T denote the uniform counting measure on {p∈Px:ℓ(p)<T}\{p\in\mathcal{P}_x:\ell(p)<T\} for each T>0T>0.

Theorem 1.1 ([7] Theorem 1.2). There exist Λ>0\Lambda>0 and σ2>0\sigma^2>0 such that, for every singularity x∈Xx\in\mathcal{X} and every a<ba<b,

lim⁡T→∞νT{p∈Px:m(γ)−ΛTT∈[a,b]}=12πσ2∫abexp⁡(−t22σ2) dt.\lim_{T\to\infty}\nu_T\left\{p\in\mathcal{P}_x:\frac{m(\gamma)-\Lambda T}{\sqrt{T}}\in[a,b]\right\} = \frac{1}{\sqrt{2\pi\sigma^2}}\int_a^b\exp\left(-\frac{t^2}{2\sigma^2}\right)\,dt.

The original statement in [7] has m(p)+1m(p)+1 instead of m(p)m(p) (i.e. it counts the number of singularities visited) but subtracting one leaves the central limit theorem unchanged.

Main result

The purpose of this note is to obtain a corresponding local limit theorem but for counting over prime closed geodesics. We decided to pursue counting over closed geodesics as it feels more in keeping with the aforementioned results for negatively curved surfaces. Let P\mathcal{P} be the set of oriented prime closed geodesics. These are closed, saddle connection paths, considered up to cyclic permutation, that are not iterates of a shorter closed geodesic (see Definition 2.2). We set P(T)={γ∈P:ℓ(γ)<T}\mathcal{P}(T)=\{\gamma\in\mathcal{P}:\ell(\gamma)<T\} and recall that, by Theorem C in [CP21] we have the asymptotic

#P(T)∼ehThTas T→∞\#\mathcal{P}(T)\sim\frac{e^{hT}}{hT}\quad\text{as }T\to\infty

where h>0h>0 is a constant. For γ∈P\gamma\in\mathcal{P}, the integer m(γ)m(\gamma) counts the saddle connections in one prime period (equivalently the number of singularities visited along the closed geodesic). For T>0T>0 let μT\mu_T denote the uniform measure on P(T)\mathcal{P}(T), so that

μT(E)=#(E∩P(T))#P(T)\mu_T(E)=\frac{\#(E\cap\mathcal{P}(T))}{\#\mathcal{P}(T)}

for E⊂PE\subset\mathcal{P}. Our main theorem is the following uniform local limit theorem.

Theorem 1.2. For every translation surface (X,ω)(X,\omega) with at least one singularity, there exist Λ>0\Lambda>0 and σ2>0\sigma^2>0 such that

lim⁡T→∞sup⁡n∈N∣TμT{γ∈P:m(γ)=n}−12πσ2exp⁡(−(n−ΛT)22σ2T)∣=0.\lim_{T\to\infty}\sup_{n\in\mathbb{N}}\left|\sqrt{T}\mu_T\{\gamma\in\mathcal{P}:m(\gamma)=n\}-\frac{1}{\sqrt{2\pi\sigma^2}}\exp\left(-\frac{(n-\Lambda T)^2}{2\sigma^2T}\right)\right|=0.

The constants Λ\Lambda and σ2\sigma^2 are the same as in Theorem 1.1.

This result shows that the probability of seeing any fixed integer number of singularity encounters has a Gaussian asymptotic at the scale T−1/2T^{-1/2}. We immediately deduce the following corollaries.

Corollary 1.3. Fix a translation surface (X,ω)(X,\omega) with at least one singularity. Then, for every integer q≥2q\ge2 and 0≤l≤q−10\le l\le q-1 we have that

#{γ∈P(T):m(γ)=l(modq)}#P(T)⟶1q\frac{\#\{\gamma\in\mathcal{P}(T):m(\gamma)=l\pmod q\}}{\#\mathcal{P}(T)}\longrightarrow\frac{1}{q}

as T→∞T\to\infty. That is, m(γ)m(\gamma) is uniformly distributed modulo qq (for any qq), when sampling γ∈P\gamma\in\mathcal{P} with respect to the sequence of uniform counting measures μT\mu_T.

This is easy to deduce from Theorem 1.2 (see for example Theorem 1.3 in [DH22]). We also have the following result, which follows from [CP21, Theorem C] and setting T=n/ΛT=n/\Lambda in Theorem 1.2.

Corollary 1.4. Let (X,ω)(X,\omega) be a translation surface with at least one singularity. Write Λ>0,σ2>0\Lambda>0,\sigma^2>0 for the constants in Theorem 1.2. Then

#{γ∈P:m(γ)=n and ℓ(γ)<nΛ}∼Λ3/2h2πσ2ehn/Λn3/2\#\left\{\gamma\in\mathcal{P}:m(\gamma)=n\text{ and }\ell(\gamma)<\frac{n}{\Lambda}\right\}\sim\frac{\Lambda^{3/2}}{h\sqrt{2\pi\sigma^2}}\frac{e^{hn/\Lambda}}{n^{3/2}}

as n→∞n\to\infty (through the integers). Here hh is the exponential growth rate of #P(T)\#\mathcal{P}(T) as T→∞T\to\infty.

This provides a precise asymptotic for the number of prime closed geodesics γ\gamma with m(γ)=nm(\gamma)=n (as long as we scale the geometric length appropriately).

Outline of the proof

Following [CP21, CP25] we encode geodesics by cyclic paths in a countable directed graph whose vertices are oriented saddle connections. The results in [7] provides key structural results, such as connectivity properties, for this graph. We first show, using geometric arguments, that the period of the graph is always one (Proposition 2.4) independent of (X,ω)(X,\omega). This removes possible congruence restrictions on m(γ)m(\gamma).

The key step is then to further develop the functional analytic methods used in [7]. We define a complex parameter family of linear operators ArA_r on ℓ2(S)\ell^2(\mathcal{S}) and prove that they are

  1. trace class; and,

  2. their traces can be used to enumerate closed geodesics.

We then apply the appropriate positive operator theory and analytic perturbation theory to these operators to deduce that they have a spectral gap and an analytically varying, lead eigenvalue, We then obtain non-lattice spectral estimates for this operator using work from [7]. Together with the usual Fourier inversion and smoothing tricks (as in the classical local limit theorems of Stone [25], Lalley [15], etc) this gives precise counts for closed geodesics. Along the way, we identify the mean and variance constants in Theorem 1.2 with those appearing in [7] (i.e. in Theorem 1.1). This identification avoids the need to establish the positivity of the constants directly from the definition of the trace class operator ArA_r which appears difficult to do.

Section 2 introduces the saddle connection coding and proves preliminary structural results about it. Section 3 develops the trace-class operator and its spectral properties, while Section 4 establishes the local limit theorem and passes to prime geodesics. The appendix identifies the constants with those in [7].

AI disclosure

AI disclosure. This project was the author’s first attempt at using AI assistance in research. The author had considered this local limit problem before and had a full outline for the proof but was struggling to find an appropriate Perron-Frobenius result to apply to the operator defined in Section 3. The author described the outline of their proof to ChatGPT and asked it to ‘find an appropriate Perron-Frobenius theorem’ to apply to the operator defined in Section 3. After a short discussion, the author found an appropriate result ([8], Theorem 1.5) and using it, was able to complete the proof. AI was not used in the write-up of this work, nor to assist with the mathematical arguments (other than the search for a useful Perron-Frobenius theorem). Lastly, AI was used for a few literature searches.

Preliminaries

Translation surfaces

Translation surfaces. We briefly recap some introductory material concerning translation surfaces, see [Wri15] for more details. Let X\mathcal{X} be the nonempty set of zeros of ω\omega and S\mathcal{S} the countable set of oriented saddle connections. For s∈Ss \in\mathcal{S}, let i(s),t(s)∈Xi(s), t(s) \in\mathcal{X} denote its endpoints (initial and terminal respectively) and ℓ(s)>0\ell(s) > 0 its geometric length. We write

ℓ∗≔inf⁡s∈Sℓ(s)>0\ell_* \coloneqq\inf_{s \in\mathcal{S}} \ell(s) > 0

for the length of shortest saddle connection. A concatenation of saddle connections is locally length-minimising (i.e. a local geodesic) precisely when, at each singularity visited, the angle between the incoming and outgoing saddle connections are at least π\pi; see [7], Definition 4.2.

Following Colognese-Pilloni [CP21] we encode possible closed geodesics using an infinite graph.

Definition 2.1. Write s1∼s2s_1 \sim s_2 if t(s1)=i(s2)t(s_1) = i(s_2) and s1s2s_1s_2 is a local geodesic. We define a directed graph GX\mathcal{G}_{\mathcal{X}} with vertex set S\mathcal{S} and an edge joining s1s_1 to s2s_2 precisely when s1∼s2s_1 \sim s_2. A closed saddle path w=(s1,…,sn)w = (s_1, \ldots, s_n) is a path with sj∼sj+1s_j \sim s_{j+1} for 1≤j≤n−11 \le j \le n - 1 and sn∼s1s_n \sim s_1. Its geometric length is ℓ(w)=∑j=1nℓ(sj)\ell(w) = \sum_{j=1}^{n} \ell(s_j) and its saddle connection count is m(w)=nm(w) = n. Two closed paths represent the same oriented closed geodesic if they differ by a cyclic rotation. The functions ℓ\ell and mm extend naturally to closed geodesics.

We define the graph period of GX\mathcal{G}_{\mathcal{X}} to be

d(GX)=gcd⁡{n≥1:GX has an admissible closed path with n edges}d(\mathcal{G}_{\mathcal{X}}) = \gcd\{n \ge1 : \mathcal{G}_{\mathcal{X}}\text{ has an admissible closed path with }n\text{ edges}\}

where gcd denotes the greatest common divisor.

Definition 2.2. A closed geodesic γ\gamma corresponding to a closed saddle connection path (s1,…,sn)(s_1, \ldots, s_n) is prime if it is not the iterate of any closed geodesic consisting of strictly fewer than nn saddle connections.

We require the following three conditions established in Section 4 of [7].

Proposition 2.3. For every (X,ω)(X,\omega) as above:

(i) For every t>0t>0, ∑s∈Se−tℓ(s)<∞\sum_{s\in\mathcal{S}}e^{-t\ell(s)}<\infty;

(ii) There is C>0C>0 such that for any s1,s2∈Ss_1,s_2\in\mathcal{S} there is a saddle connection path beginning with s1s_1, ending with s2s_2, of length at most ℓ(s1)+ℓ(s2)+C\ell(s_1)+\ell(s_2)+C; and,

(iii) The Q\mathbb{Q}-vector space spanned by {ℓ(s):s∈S}\{\ell(s):s\in\mathcal{S}\} has infinite dimension. Likewise, the Q\mathbb{Q}-span of the lengths of closed geodesics is infinite-dimensional.

Proof. See [7] for (i) and (ii). Point (iii) is [7], Proposition 4.8.

The following result shows that the period of GX\mathcal{G}_X is always 1, that is, there are no arithmetic constraints on how many saddle connections make up a closed geodesic.

Proposition 2.4. For every translation surface with at least one singularity, d(GX)=1d(\mathcal{G}_X)=1, i.e. GX\mathcal{G}_X has period one.

Proof. Let d=d(GX)d=d(\mathcal{G}_X) and fix a base vertex s0∈Ss_0\in\mathcal{S}. For each s∈Ss\in\mathcal{S}, define j(s)∈Z/dZj(s)\in\mathbb{Z}/d\mathbb{Z} to be the number of edges in a directed path from s0s_0 to ss, taken modulo dd. This is well defined because if two such paths have lengths n1,n2n_1,n_2, that we can, using Proposition 2.3 (ii), append both paths by the same directed path from ss to s0s_0. The resulting closed paths have lengths n1+kn_1+k and n2+kn_2+k which are both divisible by dd. It follows that

j(s′)=j(s)+1(modd)whenever s⇝s′.(1)j(s')=j(s)+1\pmod d\qquad\text{whenever }s\leadsto s'. \tag*{(1)}

We claim that j(s)j(s) depends only on the initial singularity i(s)i(s). Fix a singularity x∈Xx\in\mathcal{X} and choose a triangulation of XX by saddle connections, [Wri15, Lemma 1.11]. Write L(x)=2π(k(x)+1)≥4πL(x)=2\pi(k(x)+1)\geq4\pi for the total cone angle at x∈Xx\in\mathcal{X}. The outgoing directions of the triangulation edges at x∈Xx\in\mathcal{X} divide the possible angles R/L(x)Z\mathbb{R}/L(x)\mathbb{Z} into finitely many arcs, each of length strictly less than π\pi. In particular, every open angle arc of length greater than π\pi contains an outgoing triangulation-edge angle.

Suppose first that s,s′∈Ss,s'\in\mathcal{S} start at xx and that the angle measured them as they leave xx is less than π\pi. When this is the case, it is easy to find a saddle connection s~\widetilde{s} with t(s~)=xt(\widetilde{s})=x such that both s~⇝s\widetilde{s}\leadsto s and s~⇝s′\widetilde{s}\leadsto s'. By (1) we then have that

j(s)=j(s~)+1=j(s′)(modd).j(s)=j(\widetilde{s})+1=j(s')\pmod d.

Now take arbitrary s,s′s,s' starting at xx and consider their outgoing directions among the finitely many triangulation-edge directions at xx. Walking around the circle from ss to s′s' gives a finite sequence of outgoing saddle connections for which their successive angle gaps are strictly less than π\pi. By the preceding paragraph, jj is constant along this sequence. This proves the claim and we may therefore write j(s)=ax∈Z/dZj(s)=a_x\in\mathbb{Z}/d\mathbb{Z} whenever i(s)=xi(s)=x.

We now note that for any saddle connection ss from x1∈Xx_1\in\mathcal{X} to x2∈Xx_2\in\mathcal{X}. Equation (1) then gives

ax2=j(x2)=j(x1)+1=ax1+1(modd).a_{x_2}=j(x_2)=j(x_1)+1=a_{x_1}+1\pmod d.

Applying the same argument to the reversed saddle connection from x2x_2 to x1x_1 shows that ax1=ax2+1(modd)a_{x_1}=a_{x_2}+1\pmod d. Hence dd divides 2.

Finally, consider any triangle in the triangulation. If two of its vertices represent the same singularity xx, then one side is a saddle connection from xx to itself and d=1d=1. Otherwise its vertices are three distinct singularities x1,x2,x3∈Xx_1,x_2,x_3\in\mathcal{X}. Applying the relation to its three oriented sides x1→x2x_1\to x_2, x2→x3x_2\to x_3 and x3→x1x_3\to x_1 implies that dd divides 3. Since dd divides 2 and 3 we conclude that d=1d=1 and the proof is finished.

The following result will be key to establish the required non-arithmeticity condition for the local limit theorem.

Lemma 2.5. There does not exist τ∈R∖{0}\tau\in\mathbb{R}\setminus\{0\} and β∈R\beta\in\mathbb{R} such that

τℓ(γ)−βm(γ)∈2πZ\tau\ell(\gamma)-\beta m(\gamma)\in2\pi\mathbb{Z}

for all closed geodesics γ\gamma.

Proof. If the conclusion failed then the length of every closed geodesic would belong to τ−1(βZ+2πZ)\tau^{-1}(\beta\mathbb{Z}+2\pi\mathbb{Z}), a rational vector space generated by two real numbers. This contradicts Proposition 2.3(iii).

Functional analysis

As discussed in the introduction, we will define a natural operator on ℓ2(S)\ell^2(\mathcal{S}) and study its spectral properties. Here we briefly recall some preliminary material. See Section 7.1 of [10] for more details.

The space ℓ2(S)\ell^2(\mathcal{S}) is a Hilbert space when equipped with the standard inner product. It can also be equipped with a Banach lattice structure:

For real-valued sequences u,v∈ℓ2(S)u,v\in\ell^2(\mathcal{S}), we write u≤vu\leq v if u(s)≤v(s)u(s)\leq v(s) for every s∈Ss\in\mathcal{S} (here u(s),v(s)u(s),v(s) are the entries of u,vu,v at index ss). For any such uu and vv, their supremum u∨vu\vee v and infimum u∧vu\wedge v are defined coordinate-wise by

(u∨v)(s)=max⁡{u(s),v(s)}and(u∧v)(s)=min⁡{u(s),v(s)},(u\vee v)(s)=\max\{u(s),v(s)\}\quad\text{and}\quad(u\wedge v)(s)=\min\{u(s),v(s)\},

for every s∈Ss\in\mathcal{S}. Both sequences belong to ℓ2(S)\ell^2(\mathcal{S}). The lattice modulus is given by ∣u∣(s)=∣u(s)∣|u|(s)=|u(s)|, which also extends to complex-valued sequences. This norm is monotone in the sense that ∣u∣≤∣v∣|u|\leq|v| implies ∥u∥2≤∥v∥2\|u\|_2\leq\|v\|_2. Together with completeness, these properties make ℓ2(S)\ell^2(\mathcal{S}) a complex Banach lattice (see [10 Definition 7.2]).

A linear subspace I⊆ℓ2(S)I\subseteq\ell^2(\mathcal{S}) is called a lattice ideal if

u∈I,∣v∣≤∣u∣⟹v∈I,u\in I,\quad|v|\leq|u|\quad\Longrightarrow\quad v\in I,

see [10 Definition 7.9]. A positive, bounded linear operator T:ℓ2(S)→ℓ2(S)T:\ell^2(\mathcal{S})\to\ell^2(\mathcal{S}) is said to be Banach lattice irreducible if its only invariant, closed lattice ideals are {0}\{0\} and ℓ2(S)\ell^2(\mathcal{S}) [10 Definition 7.13].

Every closed lattice ideal of ℓ2(S)\ell^2(\mathcal{S}) is of the form

IJ={u∈ℓ2(S):u(s)=0 for every s∉J}I_J=\{u\in\ell^2(\mathcal{S}):u(s)=0\text{ for every }s\notin J\}

for some J⊆SJ\subseteq\mathcal{S}. Indeed, if II is a closed lattice ideal and JJ is the set of coordinates at which some element of II is nonzero, then the ideal property implies that II contains every standard basis vector ese_s with s∈Js\in J. Since II is closed it therefore contains its closed linear span, which is IJI_J. The reverse inclusion follows from the definition of JJ, and conversely every IJI_J is a closed lattice ideal.

We are now ready to proceed with the proof of our local limit theorem.

A trace-class operator and its spectral data

The proof is based around understanding the properties of the following family of linear operators. Let (us)s∈S(u_s)_{s\in\mathcal{S}} be the standard orthonormal basis of ℓ2(S)\ell^2(\mathcal{S}).

Definition 3.1. For r∈Cr\in\mathbb{C} with Re⁡(r)>0\operatorname{Re}(r)>0, define the following operator on ℓ2(S)\ell^2(\mathcal{S}): for s,s′∈Ss,s'\in\mathcal{S} set

Ar(s,s′)=id⁡{s⇝s′}exp⁡(−r2(ℓ(s)+ℓ(s′)))A_r(s,s')=\operatorname{id}_{\{s\leadsto s'\}}\exp\left(-\frac{r}{2}\bigl(\ell(s)+\ell(s')\bigr)\right)

where id⁡{s⇝s′}=1\operatorname{id}_{\{s\leadsto s'\}}=1 if s⇝s′s\leadsto s' and 00 otherwise. Then for v∈ℓ(S)v\in\ell(\mathcal{S}) we set

Ar(v)=∑s,s′∈SAr(s,s′)Ess′(v)(2)A_r(v)=\sum_{s,s'\in\mathcal{S}}A_r(s,s')E_{ss'}(v) \tag*{(2)}

where Ess′(v)=⟨v,us′⟩usE_{ss'}(v)=\langle v,u_{s'}\rangle u_s.

It is easy to see that this gives a well defined linear operator on ℓ2(S)\ell^2(\mathcal{S}) using Proposition 2.3(i). This operator can be thought of as an infinite matrix. The operator is defined precisely so that the product of the weights around a closed path ww equals exactly e−rℓ(w)e^{-r\ell(w)}. The definition of this operator is inspired by the operator considered in Section 3 of [7]. We now turn our attention to understanding the spectral properties of this operator and how it can be used to encode closed geodesics.

For the following result, recall that the trace norm ∥A∥1\|A\|_1 of a compact operator A:H→HA:H\to H (from a Hilbert space HH to itself) is the sum of the singular values of AA.

Lemma 3.2. For every r∈Cr \in\mathbb{C} with Re⁡(r)>0\operatorname{Re}(r)>0, the infinite matrix ArA_r defines a trace-class operator on ℓ2(S)\ell^2(\mathcal{S}). More precisely, writing ∥⋅∥1\|\cdot\|_1 for the trace norm and Z(t)=∑s∈Se−tℓ(s)Z(t)=\sum_{s\in\mathcal{S}}e^{-t\ell(s)} for t>0t>0, we have that

∥Ar∥1≤Z(Re⁡(r)/2)2<∞.(3)\|A_r\|_1 \le Z(\operatorname{Re}(r)/2)^2<\infty. \tag*{(3)}

The map r↦Arr\mapsto A_r is holomorphic with values in the space of trace-class operators on ℓ2(S)\ell^2(\mathcal{S}) throughout Re⁡(r)>0\operatorname{Re}(r)>0. Moreover, for every integer n≥1n\ge1,

tr⁡Arn=∑w a closed pathm(w)=ne−rℓ(w),(4)\operatorname{tr}A_r^n=\sum_{\substack{w\ \text{a closed path}\\m(w)=n}}e^{-r\ell(w)}, \tag*{(4)}

and the sum is absolutely convergent.

Proof. Let FR={s∈S:ℓ(s)≤R}F_R=\{s\in\mathcal{S}:\ell(s)\le R\} for R>0R>0. By Proposition 2.3(i), FRF_R is finite for every R>0R>0, and

Z(t)=∑s∈Se−tℓ(s)<∞Z(t)=\sum_{s\in\mathcal{S}}e^{-t\ell(s)}<\infty

for $t>0.

Now note that the operators Ess′E_{ss'} in (2) have exactly one nonzero singular value, equal to 1. Hence ∥Ess′∥1=1\|E_{ss'}\|_1=1. Writing α=Re⁡(r)>0\alpha=\operatorname{Re}(r)>0, we then have

∑s,s′∈S∥Ar(s,s′)Ess′(v)∥1≤∑s,s′∈Se−αℓ(s)/2e−αℓ(s′)/2=(∑s∈Se−αℓ(s)/2)2=Z(α/2)2<∞.\begin{aligned} \sum_{s,s'\in\mathcal{S}}\|A_r(s,s')E_{ss'}(v)\|_1\le\sum_{s,s'\in\mathcal{S}}e^{-\alpha\ell(s)/2}e^{-\alpha\ell(s')/2} \\ &=\left(\sum_{s\in\mathcal{S}}e^{-\alpha\ell(s)/2}\right)^2=Z(\alpha/2)^2<\infty. \end{aligned}

For N∈NN\in\mathbb{N}, set

Ar(N)=∑s,s′∈FNAr(s,s′)Ess′.A_r^{(N)}=\sum_{s,s'\in F_N}A_r(s,s')E_{ss'}.

Note that Ar(N)A_r^{(N)} is finite rank and hence trace class. For M>NM>N, the triangle inequality gives

∥Ar(M)−Ar(N)∥1≤∑(s,s′)∈FM2∖FN2e−α(ℓ(s)+ℓ(s′))/2≤∑(s,s′)∈S2∖FN2e−α(ℓ(s)+ℓ(s′))/2=Z(α/2)2−(∑s∈FNe−αℓ(s)/2)2⟶0\begin{aligned} \|A_r^{(M)}-A_r^{(N)}\|_1\le\sum_{(s,s')\in F_M^2\setminus F_N^2}e^{-\alpha(\ell(s)+\ell(s'))/2} \\ &\le\sum_{(s,s')\in\mathcal{S}^2\setminus F_N^2}e^{-\alpha(\ell(s)+\ell(s'))/2} =Z(\alpha/2)^2-\left(\sum_{s\in F_N}e^{-\alpha\ell(s)/2}\right)^2\longrightarrow0 \end{aligned}

where the last limit follows from the convergence of Z(α/2)Z(\alpha/2) above. Therefore (Ar(N))N≥1(A_r^{(N)})_{N\ge1} is a Cauchy sequence in the trace norm. We then use that (S1(ℓ2(S)),∥⋅∥1)(\mathcal{S}_1(\ell^2(\mathcal{S})),\|\cdot\|_1), the space of trace class operators on ℓ2(S)\ell^2(\mathcal{S}), is a Banach space and is hence complete in the trace norm; see Peterson [20 Proposition 2.1.6, p. 21], as well as Simon [23 Chapters 2-3]. Therefore Ar(N)A_r^{(N)} converges to a trace-class operator. It is easy to check that Ar(N)A_r^{(N)} tends to ArA_r as N→∞N\to\infty. Hence each ArA_r is trace class as required.

Analyticity is easy to check from the above. Indeed, fix r0r_0 with α0=Re⁡(r0)>0\alpha_0=\operatorname{Re}(r_0)>0, and work in the disk ∣r−r0∣<α0/2|r-r_0|<\alpha_0/2. There Re⁡(r)>α0/2\operatorname{Re}(r)>\alpha_0/2, and the trace-norm sum of the terms in (2) is bounded above by Z(α0/4)2Z(\alpha_0/4)^2. Every individual operator that is summed in (2) is entire in rr. The Weierstrass theorem for Banach-space-valued holomorphic functions (Theorem 3.1 in [3]) therefore proves local uniform convergence and analyticity in S1(ℓ2(S))\mathcal{S}_1(\ell^2(\mathcal{S})).

We now verify the trace formula. Note that for any fixed n≥1n\ge1 and j≥1j\ge1,

tr⁡((Ar(j))n)=∑s1,…,sn∈Fj∏k=1nid⁡{sk⇝sk+1}exp⁡(−r2∑k=1n(ℓ(sk)+ℓ(sk+1)))=∑(s1,…,sn)∈Fjnsk⇝sk+1 (closed)e−r∑k=1nℓ(sk),(5)\operatorname{tr}\left((A_r^{(j)})^n\right)=\sum_{s_1,\ldots,s_n\in F_j}\prod_{k=1}^{n}\operatorname{id}_{\{s_k\leadsto s_{k+1}\}}\exp\left(-\frac{r}{2}\sum_{k=1}^{n}(\ell(s_k)+\ell(s_{k+1}))\right)=\sum_{\substack{(s_1,\ldots,s_n)\in F_j^n\\s_k\leadsto s_{k+1}\ \mathrm{(closed)}}}e^{-r\sum_{k=1}^{n}\ell(s_k)}, \tag*{(5)}

where sn+1=s1s_{n+1}=s_1. The second equality follows because each ℓ(sk)\ell(s_k) occurs in two neighbouring factors with coefficient $-r/2.

The right-hand sums converge absolutely as j→∞j \to\infty, since

∑s1,…,sn∈Ssk⇝sk+1(closed)∣e−r∑k=1nℓ(sk)∣≤∑s1,…,sn∈S∏k=1ne−αℓ(sk)=Z(α)n<∞.(6)\sum_{\substack{s_1,\ldots,s_n\in\mathcal{S}\\s_k\leadsto s_{k+1}\,(\mathrm{closed})}}\left|e^{-r\sum_{k=1}^{n}\ell(s_k)}\right|\leq\sum_{s_1,\ldots,s_n\in\mathcal{S}}\prod_{k=1}^{n}e^{-\alpha\ell(s_k)}=Z(\alpha)^n<\infty. \tag*{(6)}

It remains to justify passing to the limit on the left hand side of (5). Put BN=Ar(N)B_N=A_r^{(N)} and B=ArB=A_r so that ∥BN−B∥1→0\lVert B_N-B\rVert_1\to0 as N→∞N\to\infty. For every fixed positive integer nn, we will use the following elementary algebraic identity:

BNn−Bn=∑j=0n−1BNn−1−j(BN−B)Bj.(7)B_N^n-B^n=\sum_{j=0}^{n-1}B_N^{n-1-j}(B_N-B)B^j. \tag*{(7)}

Recall that the trace-class operators satisfy the following: if TT is trace class and S,RS,R are bounded, then

∥STR∥1≤∥S∥op∥T∥1∥R∥op,(8)\lVert STR\rVert_1\leq\lVert S\rVert_{\mathrm{op}}\lVert T\rVert_1\lVert R\rVert_{\mathrm{op}}, \tag*{(8)}

see Simon [23 Chapters 2-3]. Here ∥⋅∥op\lVert\cdot\rVert_{\mathrm{op}} is the operator norm. Applying (8) to each term of (7) gives

∥BNn−Bn∥1≤∑j=0n−1∥BN∥opn−1−j∥BN−B∥1∥B∥opj≤nMn−1∥BN−B∥1,\begin{aligned} \lVert B_N^n-B^n\rVert_1\leq\sum_{j=0}^{n-1}\lVert B_N\rVert_{\mathrm{op}}^{n-1-j}\lVert B_N-B\rVert_1\lVert B\rVert_{\mathrm{op}}^j \\ &\leq nM^{n-1}\lVert B_N-B\rVert_1, \end{aligned}

where M=Z(α/2)2M=Z(\alpha/2)^2. The last inequality uses ∥BN∥op≤∥BN∥1≤M\lVert B_N\rVert_{\mathrm{op}}\leq\lVert B_N\rVert_1\leq M and ∥B∥op≤∥B∥1≤M\lVert B\rVert_{\mathrm{op}}\leq\lVert B\rVert_1\leq M, which follow from (3). Because MM and nn are fixed, this proves that BNn→BnB_N^n\to B^n in trace norm.

For a trace-class operator TT, the trace is a continuous linear functional satisfying ∣tr⁡T∣≤∥T∥1\lvert\operatorname{tr}T\rvert\leq\lVert T\rVert_1; see Simon [23 Chapters 2-3]. Hence

∣tr⁡((Ar(N))n)−tr⁡(Arn)∣≤∥(Ar(N))n−Arn∥1⟶0.(9)\left|\operatorname{tr}\left((A_r^{(N)})^n\right)-\operatorname{tr}(A_r^n)\right|\leq\left\lVert(A_r^{(N)})^n-A_r^n\right\rVert_1\longrightarrow0. \tag*{(9)}

On the other hand, each fixed closed path (s1,…,sn)(s_1,\ldots,s_n) lies in FNn\mathcal{F}_N^n for all sufficiently large NN. Since (6) dominates the absolute values of all the terms by a summable function on Sn\mathcal{S}^n, the dominated convergence theorem gives that

lim⁡N→∞∑(s1,…,sn)∈FNnsk⇝sk+1(closed)e−r∑kℓ(sk)=∑(s1,…,sn)∈Snsk⇝sk+1(closed)e−r∑kℓ(sk).\lim_{N\to\infty}\sum_{\substack{(s_1,\ldots,s_n)\in\mathcal{F}_N^n\\s_k\leadsto s_{k+1}\,(\mathrm{closed})}}e^{-r\sum_k\ell(s_k)} = \sum_{\substack{(s_1,\ldots,s_n)\in\mathcal{S}^n\\s_k\leadsto s_{k+1}\,(\mathrm{closed})}}e^{-r\sum_k\ell(s_k)}.

Taking limits in (5), using (9) for its left side and the previous display equation for its right side, proves (4). □

We now show that the operators ArA_r are Banach lattice irreducible as defined in Section 2.2. For J⊂SJ\subset\mathcal{S} let IJI_J denote the ideal defined there.

Proposition 3.3. For every real r>0r>0, the operator Ar:ℓ2(S)→ℓ2(S)A_r:\ell^2(\mathcal{S})\to\ell^2(\mathcal{S}) is Banach lattice irreducible.

Proof. Since r>0r>0, all matrix coefficients of ArA_r are nonnegative, so ArA_r is a positive operator. Suppose, for a contradiction, that there exists a nontrivial closed lattice ideal IJI_J invariant under ArA_r (for some J⊂SJ\subset\mathcal{S}). Choose s′∈Js'\in J and s∈S∖Js\in\mathcal{S}\setminus J. By Proposition 2.3(ii), there is a directed path s=s0⇝s1⇝⋯⇝sn=s′s=s_0\leadsto s_1\leadsto\cdots\leadsto s_n=s'. Since each transition along this path has strictly positive weight, we obtain

(Arn)(s,s′)≥∏k=0n−1Ar(sk,sk+1)>0.(A_r^n)(s,s')\geq\prod_{k=0}^{n-1}A_r(s_k,s_{k+1})>0.

Let us′u_{s'} denote the standard coordinate vector corresponding to s′s'. Since s′∈Js'\in J, we have us′∈IJu_{s'}\in I_J. By the invariance of IJI_J, Arnus′∈IJA_r^n u_{s'}\in I_J. Then, as s∉Js\notin J, (Arnus′)(s)=0(A_r^n u_{s'})(s)=0. On the other hand,

(Arnus′)(s)=(Arn)(s,s′)>0,(A_r^n u_{s'})(s)=(A_r^n)(s,s')>0,

which is a contradiction. Therefore no nontrivial closed lattice ideal is invariant under ArA_r, and hence ArA_r is Banach lattice irreducible. □

:::

We can now deduce good properties of the spectrum of ArA_r using the aforementioned Perron-Frobenius theorem [8 Theorem 1.5].

Proposition 3.4. There is a unique h>0h > 0 such that ρ(Ah)=1\rho(A_h) = 1. On a real neighbourhood II of hh, ArA_r has a simple positive spectral radius λ(r)\lambda(r), depending real-analytically on rr, and

spec⁡(Ar)∩{z:∣z∣=λ(r)}={λ(r)},(10)\operatorname{spec}(A_r) \cap\{z : \lvert z\rvert= \lambda(r)\} = \{\lambda(r)\}, \tag*{(10)}

is simple. Lastly,

a=−ddslog⁡λ(s)∣s=handv=d2ds2log⁡λ(s)∣s=ha = -\left.\frac{d}{d s} \log\lambda(s)\right|_{s=h} \qquad\text{and}\qquad v = \left.\frac{d^2}{d s^2} \log\lambda(s)\right|_{s=h}

are both strictly positive.

Proof. Perron-Frobenius theory and the spectral gap. By Lemma 3.2 and Proposition 3.3, ArA_r is a positive, compact and Banach lattice irreducible operator. Therefore by [8 Theorem 1.5], its spectral radius λ(r):=ρ(Ar)>0\lambda(r) := \rho(A_r) > 0 is an algebraically simple eigenvalue, with strictly positive right and left eigenvectors ur,vr∈ℓ2(S)u_r, v_r \in\ell^2(\mathcal{S}):

Arur=λ(r)ur,Ar∗vr=λ(r)vr,ur(s),vr(s)>0A_r u_r = \lambda(r)u_r,\qquad A_r^* v_r = \lambda(r)v_r,\qquad u_r(s),v_r(s)>0

for s∈Ss \in\mathcal{S}. By compactness, the spectrum of ArA_r, other than those with modulus λ(r)\lambda(r) is contained in a disk of radius strictly less that λ(r)\lambda(r).

We now verify that GX\mathcal{G}_X having period one excludes other eigenvalues of modulus λ(r)\lambda(r). Suppose Arz=μzA_r z = \mu z with z≠0z \ne0 and ∣μ∣=λ(r)\lvert\mu\rvert= \lambda(r). The coordinate-wise triangle inequality gives

λ(r)∣z∣=∣Arz∣≤Ar∣z∣.\lambda(r)\lvert z\rvert= \lvert A_r z\rvert\le A_r\lvert z\rvert.

We then take the product of this inequality with the strictly positive vector vrv_r. Using Ar∗vr=λ(r)vrA_r^*v_r = \lambda(r)v_r, we see that the nonnegative vector Ar∣z∣−λ(r)∣z∣A_r\lvert z\rvert-\lambda(r)\lvert z\rvert has zero inner product with vrv_r. It must therefore vanish coordinate-wise. By simplicity of the eigenvalue, ∣z∣=cur\lvert z\rvert= c u_r for some c>0c > 0. After rescaling, write z(s)=ur(s)ζ(s)z(s) = u_r(s)\zeta(s) with ∣ζ(s)∣=1\lvert\zeta(s)\rvert= 1. With η:=μ/λ(r)\eta:= \mu/\lambda(r), the eigenvector equations imply

∑s′∈SAr(s,s′)ur(s′)ζ(s′)=ηζ(s)∑s′∈SAr(s,s′)ur(s′).\sum_{s'\in\mathcal{S}} A_r(s,s')u_r(s')\zeta(s')=\eta\zeta(s)\sum_{s'\in\mathcal{S}} A_r(s,s')u_r(s').

The coefficients on the right-hand sum are positive, and the weighted average of the unit complex numbers ζ(s′)\zeta(s') again has modulus one. This forces

ζ(s′)=ηζ(s)for every s⇝s′.\zeta(s')=\eta\zeta(s)\quad\text{for every }s\leadsto s'.

Going around any closed path of saddle connection count m(w)m(w) gives ηm(w)=1\eta^{m(w)}=1. The gcd of all these counts is one by Proposition 2.4, hence η=1\eta=1. It follows that

spec⁡(Ar)∩{∣z∣=λ(r)}={λ(r)}.\operatorname{spec}(A_r)\cap\{\lvert z\rvert=\lambda(r)\}=\{\lambda(r)\}.

Since ArA_r is compact and λ(r)\lambda(r) is a simple isolated eigenvalue, analytic perturbation theory [Kat80, Chapter VII, §1.3, Theorem 1.8] shows that r↦λ(r)r\mapsto\lambda(r) is real-analytic on (0,∞)(0,\infty).

Existence and uniqueness of hh: For α>0\alpha>0 we have the entry-wise inequality

0≤Ar+α≤e−αℓ∗Ar0\le A_{r+\alpha}\le e^{-\alpha\ell^*}A_r

therefore

λ(r+α)≤e−αℓ∗λ(r)<λ(r).\lambda(r+\alpha)\le e^{-\alpha\ell^*}\lambda(r)<\lambda(r).

Thus λ\lambda is strictly decreasing. Lemma 3.2 also gives

λ(r)≤∥Ar∥1≤Z(r/2)2→0\lambda(r)\le\lVert A_r\rVert_1\le Z(r/2)^2\to0

as r→∞r\to\infty where the convergence follows by dominated convergence, using Z(t0)<∞Z(t_0)<\infty for any fixed t0>0t_0>0.

It is easy to check that λ(r)>1\lambda(r)>1 for sufficiently small r>0r>0. By continuity, strict monotonicity, and λ(r)→0\lambda(r)\to0 at infinity, there is a unique h>0h>0 for which λ(h)=1\lambda(h)=1.

To prove positivity of the constants we appeal to the work in Cantrell-Pollicott. See Appendix § 5 that identifies aa and vv with the mean and variance corresponding to the counting central limit theorem, Theorem 1.1 above. □

Definition 3.5. We set

Λ=1aandσ2=va3.\Lambda= \frac{1}{a} \qquad\text{and} \qquad\sigma^2 = \frac{v}{a^3}.

Remark 3.6. These will turn out to be the mean and variance of our local limit theorem and they coincide with the mean and variance in Theorem 1.1 above. We also note that it is simple to check that h>0h > 0 in the above proposition is the exponential growth rate of #P(T)\#\mathcal{P}(T) as T→∞T \to\infty.

The following result controls the spectral radius of the operators ArA_r away from r=hr=h.

Lemma 3.7. For all real r∈Ir \in I and τ≠0\tau\ne0,

ρ(Ar+iτ)<λ(r).(11)\rho(A_{r+i\tau}) < \lambda(r). \tag*{(11)}

For any compact K⊂R∖{0}K \subset\mathbb{R}\setminus\{0\} there exist CK>0C_K > 0 and 0<qK<10 < q_K < 1 such that

sup⁡r∈I′, τ∈K∥λ(r)−nAr+iτn∥1≤CKqKn\sup_{r \in I',\,\tau\in K} \left\|\lambda(r)^{-n} A_{r+i\tau}^{n}\right\|_1 \le C_K q_K^n

for all n≥1n \ge1, where I′⊂II' \subset I is any sufficiently small closed interval containing hh.

Proof. Fix r∈Ir \in I and τ≠0\tau\ne0, and abbreviate

B=Ar+iτ,A=Ar,λ=λ(r),g(s,s′)=12(ℓ(s)+ℓ(s′)).B=A_{r+i\tau}, \qquad A=A_r, \qquad\lambda=\lambda(r), \qquad g(s,s')=\frac{1}{2}\bigl(\ell(s)+\ell(s')\bigr).

Then B(s,s′)=A(s,s′)e−iτg(s,s′)B(s,s')=A(s,s')e^{-i\tau g(s,s')} and ∣B(s,s′)∣=A(s,s′)|B(s,s')|=A(s,s') for all s,s′s,s'. It follows that for z∈ℓ2(S)z\in\ell^2(\mathcal{S}) and every n≥1n\ge1 we have that ∣Bnz∣≤An∣z∣|B^n z|\le A^n|z| coordinate wise and so

∥Bn∥op≤∥An∥op.\|B^n\|_{\mathrm{op}}\le\|A^n\|_{\mathrm{op}}.

The spectral-radius formula therefore gives ρ(B)≤ρ(A)=λ\rho(B)\le\rho(A)=\lambda.

We now want to understand the equality case. This is essentially the countable-state analogue of Wielandt’s theorem for an irreducible nonnegative matrix [13 Chapter 8]. Since BB is trace class by Lemma 3.2, it is compact. If ρ(B)=λ>0\rho(B)=\lambda>0, there is an eigenvector z≠0z\ne0 and some β∈R\beta\in\mathbb{R} such that

Bz=λe−iβz.(12)Bz=\lambda e^{-i\beta}z. \tag*{(12)}

Let ur,vr∈ℓ2(S)u_r,v_r\in\ell^2(\mathcal{S}) be the strictly positive right and left lead eigenvectors of AA, so that Aur=λurAu_r=\lambda u_r and A∗vr=λvrA^*v_r=\lambda v_r (Proposition 3.4). Taking absolute values of (12) gives

λ∣z∣=∣Bz∣≤A∣z∣.\lambda|z|=|Bz|\le A|z|.

The difference A∣z∣−λ∣z∣A|z|-\lambda|z| belongs to ℓ2\ell^2, is coordinate-wise nonnegative, and has zero pairing with vrv_r:

⟨vr,A∣z∣−λ∣z∣⟩=⟨A∗vr−λvr,∣z∣⟩=0.\langle v_r,A|z|-\lambda|z|\rangle =\langle A^*v_r-\lambda v_r,|z|\rangle=0.

Because vr(s)>0v_r(s)>0 for every state ss, each coordinate of this difference must vanish. Thus A∣z∣=λ∣z∣A|z|=\lambda|z| and simplicity of the lead eigenvalue implies ∣z∣=cur|z|=cu_r for some c>0c>0. Rescale zz so that c=1c=1, and write zs=ur(s)ζsz_s=u_r(s)\zeta_s, where ∣ζs∣=1|\zeta_s|=1 for every ss.

The ssth coordinate of (12) is now

∑s′:s⇝s′A(s,s′)ur(s′)e−iτg(s,s′)ζs′=λur(s)e−iβζs.(13)\sum_{s':s\rightsquigarrow s'} A(s,s')u_r(s')e^{-i\tau g(s,s')}\zeta_{s'}=\lambda u_r(s)e^{-i\beta}\zeta_s. \tag*{(13)}

whereas the eigenvalue equation yields ∑s′:s⇝s′A(s,s′)ur(s′)=λur(s)\sum_{s':s\rightsquigarrow s'}A(s,s')u_r(s')=\lambda u_r(s). Now, all of the terms in the sum in left hand side of (13) must have the same argument as the right-hand side. Since A(s,s′)ur(s′)>0A(s,s')u_r(s')>0, on every admissible transition s⇝s′s\rightsquigarrow s', it follows that

e−iτg(s,s′)ζs′=e−iβζs.(14)e^{-i\tau g(s,s')}\zeta_{s'}=e^{-i\beta}\zeta_s. \tag*{(14)}

For a closed path w=(s1,…,sm)w=(s_1,\ldots,s_m), multiply (14) along its mm transitions, with sm+1=s1s_{m+1}=s_1. The ζs\zeta_s factors cancel, leaving

e−iτℓ(w)=e−iβm(w),e^{-i\tau\ell(w)}=e^{-i\beta m(w)},

or equivalently τℓ(w)−βm(w)∈2πZ\tau\ell(w)-\beta m(w)\in2\pi\mathbb{Z} for every closed path ww. This contradicts Lemma 2.5. Therefore ρ(B)<λ\rho(B)<\lambda, proving (11).

For the final statement, let K=I′×K\mathcal{K}=I' \times K and set Dr,τ=λ(r)−1Ar+iτD_{r,\tau}=\lambda(r)^{-1}A_{r+i\tau}. The conclusion follows from the facts that the family Dr,τD_{r,\tau} depends continuously (in fact analytically by Lemma 3.2) on (r,τ)(r,\tau) in trace norm on the compact set K\mathcal{K}, and every member has spectral radius less than one.

Deducing the local limit theorem

We are now ready to deduce the local limit theorem. We follow the arguments in [15]. For n≥1n\ge1 and T>0T>0, let An(T)A_n(T) count closed saddle paths:

An(T)=#{w:w a closed saddle path, m(w)=n, ℓ(w)<T}.A_n(T)=\#\{w:w\text{ a closed saddle path},\ m(w)=n,\ \ell(w)<T\}.

By Proposition 2.4, An(T)A_n(T) is positive for every nn and sufficiently large TT depending on nn.

Lemma 4.1. Let I′⊂II'\subset I be a compact interval containing hh. For r∈I′r\in I' put ar=−(log⁡λ)′(r)a_r=-(\log\lambda)'(r) and vr=(log⁡λ)′′(r)>0v_r=(\log\lambda)''(r)>0. Uniformly for r∈I′r\in I', T∈RT\in\mathbb{R},

An(T)=erTλ(r)nr2πvrnexp⁡(−(T−arn)22vrn)+o(erTλ(r)nn)(15)A_n(T)=\frac{e^{rT}\lambda(r)^n}{r\sqrt{2\pi v_r n}}\exp\left(-\frac{(T-a_r n)^2}{2v_r n}\right)+o\left(\frac{e^{rT}\lambda(r)^n}{\sqrt{n}}\right) \tag*{(15)}

as n→∞n\to\infty. In particular, there are constants C>0C>0 and n0∈Nn_0\in\mathbb{N} independent of T,rT,r such that

An(T)≤CerTλ(r)nn(16)A_n(T)\le C\frac{e^{rT}\lambda(r)^n}{\sqrt{n}} \tag*{(16)}

for n≥n0n\ge n_0, r∈I′r\in I'.

Proof. Fix a compact interval I′⊂II'\subset I containing hh. Throughout the proof, all constants and error terms are uniform for r∈I′r\in I'. Set, for r∈I′r\in I', T>0T>0 and n≥1n\ge1,

L(r)=log⁡λ(r),ar=−L′(r),vr=L′′(r)andGn,r(T)=12πvrnexp⁡(−(T−arn)22vrn).L(r)=\log\lambda(r),\quad a_r=-L'(r),\quad v_r=L''(r)\quad\text{and}\quad G_{n,r}(T)=\frac{1}{\sqrt{2\pi v_r n}}\exp\left(-\frac{(T-a_r n)^2}{2v_r n}\right).

By Proposition 3.4 and compactness of I′I', there are constants r−>0r_->0 and v−>0v_->0 with r≥r−r\ge r_- and vr≥v−v_r\ge v_- for r∈I′r\in I'. In particular 0≤Gn,r(T)≤(2πv−n)−1/20\le G_{n,r}(T)\le(2\pi v_-n)^{-1/2}.

Let νn,r\nu_{n,r} be the measure that puts mass λ(r)−ne−rℓ(w)\lambda(r)^{-n}e^{-r\ell(w)} at ℓ(w)\ell(w) for each closed path ww with m(w)=nm(w)=n. Lemma 3.2 gives the identity

ν^n,r(τ):=∫Reiτt νn,r(dt)=λ(r)−ntr⁡Ar−iτn\widehat{\nu}_{n,r}(\tau):=\int_{\mathbb{R}}e^{i\tau t}\,\nu_{n,r}(dt)=\lambda(r)^{-n}\operatorname{tr}A_{r-i\tau}^{n}

for the Fourier transform ν^n,r\widehat{\nu}_{n,r}.

Now, by Proposition 3.4, the simple leading eigenvalue admits an analytic continuation λ(z)\lambda(z) to a complex neighbourhood of I′I'. After reducing ϵ>0\epsilon>0 if necessary, there are analytic spectral projections Πr,τ\Pi_{r,\tau} and remainders Nr,τN_{r,\tau} such that, for r∈I′r\in I' and ∣τ∣≤ϵ|\tau|\le\epsilon,

Ar−iτ=λ(r−iτ)Πr,τ+Nr,τ,Πr,τNr,τ=Nr,τΠr,τ=0A_{r-i\tau}=\lambda(r-i\tau)\Pi_{r,\tau}+N_{r,\tau},\quad\Pi_{r,\tau}N_{r,\tau}=N_{r,\tau}\Pi_{r,\tau}=0

see [Kat80], Theorem 1.7, pp. 368–369. The remainder spectra lie, uniformly, in a disk of radius q0λ(r)q_0\lambda(r), with some q0<1q_0<1. Since the operators depend continuously in trace norm we have that

ν^n,r(τ)=(λ(r−iτ)λ(r))n+O(q0n)(17)\widehat{\nu}_{n,r}(\tau)=\left(\frac{\lambda(r-i\tau)}{\lambda(r)}\right)^n+O(q_0^n) \tag*{(17)}

uniformly in ∣τ∣≤ϵ|\tau|\le\epsilon. Using Taylor’s theorem, uniformly in rr, gives

log⁡λ(r−iτ)λ(r)=iarτ−12vrτ2+O(∣τ∣3)(18)\log\frac{\lambda(r-i\tau)}{\lambda(r)}=ia_r\tau-\frac{1}{2}v_r\tau^2+O(|\tau|^3) \tag*{(18)}

and because vr≥v−>0v_r\ge v_->0, we can reduce ϵ\epsilon further so that

Re⁡log⁡λ(r−iτ)λ(r)≤−cτ2(19)\operatorname{Re}\log\frac{\lambda(r-i\tau)}{\lambda(r)}\le-c\tau^2 \tag*{(19)}

for all ∣τ∣≤ϵ|\tau|\le\epsilon and c>0c>0 independent of τ\tau. For every 0<ϵ≤B0<\epsilon\le B, Lemma 3.7 provides constants CB>0C_B>0 and 0<qB<10<q_B<1 such that

sup⁡r∈I′ϵ≤∣τ∣≤B∣ν^n,r(τ)∣≤CBqBn.(20)\sup_{\substack{r\in I'\\ \epsilon\le|\tau|\le B}}|\widehat{\nu}_{n,r}(\tau)|\le C_Bq_B^n. \tag*{(20)}

Let ff be a Schwartz function whose Fourier transform f^(τ)=∫e−iτuf(u) du\widehat{f}(\tau)=\int e^{-i\tau u}f(u)\,du has compact support. Fourier inversion gives

∫f(t−T) νn,r(dt)=12π∫Rf^(τ)e−iτTν^n,r(τ) dτ.(21)\int f(t-T)\,\nu_{n,r}(dt)=\frac{1}{2\pi}\int_{\mathbb{R}}\widehat{f}(\tau)e^{-i\tau T}\widehat{\nu}_{n,r}(\tau)\,d\tau. \tag*{(21)}

The part of the integral with ∣τ∣≥ϵ|\tau|\ge\epsilon is O(qfn)O(q_f^n) by (20). When ∣τ∣≤ϵ|\tau|\le\epsilon, (17)–(19) and the substitution τ=u/n\tau=u/\sqrt{n} give, uniformly for r∈I′r\in I',

∫−ϵϵ∣(λ(r−iτ)λ(r))n−eiarnτ−vrnτ2/2∣ dτ=o(n−1/2).\int_{-\epsilon}^{\epsilon}\left|\left(\frac{\lambda(r-i\tau)}{\lambda(r)}\right)^n-e^{iar n\tau-v_rn\tau^2/2}\right|\,d\tau=o(n^{-1/2}).

Indeed, after substitution the integrands converge to zero uniformly in rr for each fixed uu, and are dominated by 2e−c′u22e^{-c'u^2} for some c′>0c'>0. Also, continuity of f^\widehat{f} at 00 allows us to replace f^(τ)\widehat{f}(\tau) by f^(0)=∫f\widehat{f}(0)=\int f with an o(n−1/2)o(n^{-1/2}) error. The factor e−iτTe^{-i\tau T} has absolute value one, so every error is uniform in T∈RT\in\mathbb{R}. Substituting into (21) gives

sup⁡r∈I′T∈Rn∣∫f(t−T) νn,r(dt)−Gn,r(T)∫Rf(u) du∣→0.(22)\sup_{\substack{r\in I'\\T\in\mathbb{R}}}\sqrt{n}\left|\int f(t-T)\,\nu_{n,r}(dt)-G_{n,r}(T)\int_{\mathbb{R}}f(u)\,du\right|\to0. \tag*{(22)}

A standard smoothing step (which we omit) in the proof of local limit theorems (see [15 Section 7] for example) extends this statement (first to indicator functions and then) to the Riemann-integrable function fr(u)=eru1{u<0}f_r(u)=e^{ru}\mathbf{1}_{\{u<0\}}, uniformly for r∈I′r\in I'. With this choice, we have the exact identity

∫fr(t−T) νn,r(dt)=e−rTλ(r)−nAn(T).\int f_r(t-T)\,\nu_{n,r}(dt)=e^{-rT}\lambda(r)^{-n}A_n(T).

Applying this extension of (22) and multiplying by erTλ(r)ne^{rT}\lambda(r)^n proves

An(T)=erTλ(r)nr2πvrnexp⁡(−(T−arn)22vrn)+o(erTλ(r)nn),A_n(T)=\frac{e^{rT}\lambda(r)^n}{r\sqrt{2\pi v_rn}}\exp\left(-\frac{(T-a_rn)^2}{2v_rn}\right)+o\left(\frac{e^{rT}\lambda(r)^n}{\sqrt{n}}\right),

as n→∞n\to\infty where the little-oo is uniform in r∈I′r\in I' and T∈RT\in\mathbb{R}. This also gives (16) for all sufficiently large nn.

Lemma 4.2. There exist δ,c,C>0\delta,c,C>0 such that, whenever n∈Nn\in\mathbb{N} and ∣n−T/a∣≤δT|n-T/a|\le\delta T,

An(T)≤CehTTexp⁡(−c(n−T/a)2T).(23)A_n(T)\le\frac{Ce^{hT}}{\sqrt{T}}\exp\left(-c\frac{(n-T/a)^2}{T}\right). \tag*{(23)}

Furthermore, for some cδ>0c_\delta>0,

∑n≥1∣n−T/a∣>δTAn(T)=O(e(h−cδ)T)(24)\sum_{\substack{n\ge1\\|n-T/a|>\delta T}}A_n(T)=O(e^{(h-c_\delta)T}) \tag*{(24)}

as T→∞T\to\infty.

Proof. As before, write L(r)=log⁡λ(r)L(r)=\log\lambda(r), ar=−L′(r)a_r=-L'(r) and vr=L′′(r)v_r=L''(r) so that L(h)=0L(h)=0, ah=a>0a_h=a>0, and vh=v>0v_h=v>0. Shrinking the interval II from Proposition 3.4 if necessary, choose a compact interval J⊂IJ\subset I, with hh in its interior, and constants 0<v−≤v+0<v_-\le v_+ such that

v−≤L′′(r)≤v+(25)v_-\le L''(r)\le v_+ \tag*{(25)}

for r∈Jr\in J. In particular r↦arr\mapsto a_r is strictly decreasing on JJ.

Choose δ>0\delta>0 sufficiently small that whenever

∣n−Ta∣≤δT,\left|n-\frac{T}{a}\right|\le\delta T,

there are constants c0,C0>0c_0,C_0>0 with

c0T≤n≤C0T(26)c_0T\le n\le C_0T \tag*{(26)}

and furthermore T/nT/n belongs to the interior of {ar:r∈J}\{a_r:r\in J\}. When this is the case there is a unique r=r(n,T)∈Jr=r(n,T)\in J such that ar=T/na_r=T/n or equivalently L′(r)=−T/nL'(r)=-T/n. By the uniform bound (16) in Lemma 4.1,

An(T)≤Cnexp⁡(rT+nL(r)).(27)A_n(T)\le\frac{C}{\sqrt{n}}\exp\left(rT+nL(r)\right). \tag*{(27)}

We estimate the exponent exactly. Since T=nar=−nL′(r)T=na_r=-nL'(r),

rT+nL(r)−hT=n(L(r)−(r−h)L′(r)−L(h))=−n∫hr(t−h)L′′(t) dt≤−ν−n2(r−h)2.(28)\begin{aligned} rT+nL(r)-hT&=n\left(L(r)-(r-h)L'(r)-L(h)\right)\\ &=-n\int_h^r(t-h)L''(t)\,dt\\ &\le-\frac{\nu_-n}{2}(r-h)^2. \tag*{(28)} \end{aligned}

The integral is nonnegative for either r>hr>h or r<hr<h. Also, the mean-value theorem and (25) give

∣Tn−a∣=∣ar−a∣≤ν+∣r−h∣,\left|\frac{T}{n}-a\right|=|a_r-a|\le\nu_+|r-h|,

and therefore

rT+nL(r)≤hT−ν−2ν+2(T−an)2n.(29)rT+nL(r)\le hT-\frac{\nu_-}{2\nu_+^2}\frac{(T-an)^2}{n}. \tag*{(29)}

Now T−an=−a(n−T/a)T-an=-a(n-T/a) and using (26) in (27)–(29), we obtain constants C,c>0C,c>0 such that

An(T)≤CehTTexp⁡(−c(n−T/a)2T)A_n(T)\le\frac{Ce^{hT}}{\sqrt{T}}\exp\left(-c\frac{(n-T/a)^2}{T}\right)

whenever ∣n−T/a∣≤δT|n-T/a|\le\delta T and TT is sufficiently large. This proves (23).

We now show the furthermore statement. First we record an estimate that is uniform over all n≥1n\ge1 for each fixed real r>0r>0:

tr⁡Arn≤Crλ(r)n.(30)\operatorname{tr}A_r^n\le C_r\lambda(r)^n. \tag*{(30)}

This follows as in the proof of Lemma 4.1.

Choose ε>0\varepsilon>0 small enough that r−=h−ε>0r_-=h-\varepsilon>0 and r+=h+εr_+=h+\varepsilon lie in JJ. Strict monotonicity of λ\lambda gives

L−=L(r−)>0,L+=L(r+)<0.L_-=L(r_-)>0,\qquad L_+=L(r_+)<0.

By Taylor’s formula at hh,

L(h±ε)=∓aε+O(ε2).(31)L(h\pm\varepsilon)=\mp a\varepsilon+O(\varepsilon^2). \tag*{(31)}

We may choose ε\varepsilon sufficiently small (depending on δ\delta) so that for some $\kappa>0,

−ε+(a−1−δ)L−≤−κandε+(a−1+δ)L+≤−κ.(32)-\varepsilon+(a^{-1}-\delta)L_-\le-\kappa \qquad\text{and}\qquad \varepsilon+(a^{-1}+\delta)L_+\le-\kappa. \tag*{(32)}

Indeed, both left sides equal −aδε+O(ε2)-a\delta\varepsilon+O(\varepsilon^2), by (31).

For every r>0r>0 the trace formula gives the bound

An(T)≤erTtr⁡Arn≤Crexp⁡(rT+nL(r)).(33)A_n(T)\le e^{rT}\operatorname{tr}A_r^n\le C_r\exp\left(rT+nL(r)\right). \tag*{(33)}

If n/T≤a−1−δn/T\le a^{-1}-\delta, let r=r−r=r_-. Since L−>0L_->0, equations (32) and (33) imply

An(T)≤C−eT(r−+(n/T)L−)≤C−e(h−κ)T.A_n(T)\le C_-e^{T(r_-+(n/T)L_-)}\le C_-e^{(h-\kappa)T}.

If n/T≥a−1+δn/T\ge a^{-1}+\delta instead set r=r+r=r_+. Since L+<0L_+<0, the analogous estimate gives

An(T)≤C+eT[r++(n/T)L+]≤C+e(h−κ)T.A_n(T)\le C_+e^{T[r_++(n/T)L_+]}\le C_+e^{(h-\kappa)T}.

Therefore the same exponential bound holds with constants independent of n,Tn,T satisfying

∣n−Ta∣>δT.\left|n-\frac{T}{a}\right|>\delta T.

Finally, each saddle connection has length at least ℓ∗\ell_*, so An(T)=0A_n(T)=0 for nℓ∗≥Tn\ell_* \ge T. There are at most ⌈T/ℓ∗⌉\lceil T/\ell_* \rceil potentially nonzero terms. Hence

∑n≥1∣n−T/a∣>δTAn(T)≤C(1+T)e(h−κ)T=O(e(h−cδ)T)\sum_{\substack{n\ge1\\ |n-T/a|>\delta T}} A_n(T) \le C(1+T)e^{(h-\kappa)T}=O(e^{(h-c_\delta)T})

for any fixed 0<cδ<κ0<c_\delta<\kappa. This proves (24). Increasing constants if needed covers bounded values of TT, concluding the proof.

Concluding the proof

For a prime orbit consisting of nn saddle connections, there are exactly nn distinct closed paths obtained by choosing the starting saddle connection. Therefore, writing Bn(T)B_n(T) for the number of prime oriented closed geodesics with nn saddle connections, we have

Bn(T)=An(T)−Rn(T)n.B_n(T)=\frac{A_n(T)-R_n(T)}{n}.

where Rn(T)R_n(T) counts non prime closed saddle paths. The following simple result shows that non-prime orbits are negligible in the count.

Lemma 4.3. There is a constant C>0C>0 with

∑n≥1Rn(T)≤CTehT/2\sum_{n\ge1}R_n(T)\le CTe^{hT/2}

for all T>1T>1. Here h>0h>0 denotes the exponential growth rate of #P(T)\#\mathcal{P}(T).

Proof. Every non-prime closed path is a k≥2k\ge2 repetition of a prime closed path of length at most T/kT/k. Therefore

∑n≥1Rn(T)≤∑2≤k≤T/ℓ∗A(T/k).\sum_{n\ge1}R_n(T)\le\sum_{2\le k\le T/\ell_*}A(T/k).

We then recall that there exists C>0C>0 such that A(t)≤CehtA(t)\le Ce^{ht} for all t≥0t\ge0. Hence, in the above, there are O(T)O(T) terms and each is O(ehT/2)O(e^{hT/2}); the result follows.

We can now conclude the proof by combining our above work.

Proof of Theorem 1.2. For n≥1n\ge1, T>0T>0 and u∈Ru\in\mathbb{R} we set

un=n−ΛTT=n−T/aT,σ2=νa3,andφ(u)=12πσ2e−u2/(2σ2).u_n=\frac{n-\Lambda T}{\sqrt{T}}=\frac{n-T/a}{\sqrt{T}},\qquad\sigma^2=\frac{\nu}{a^3},\qquad\text{and}\qquad\varphi(u)=\frac{1}{\sqrt{2\pi\sigma^2}}e^{-u^2/(2\sigma^2)}.

Also fix R>0R>0.

When ∣un∣≤R|u_n|\le R, we have n=T/a+OR(T)n=T/a+O_R(\sqrt{T}) and so nn is comparable with TT. Applying Lemma 4.1 with r=hr=h (the exponential growth rate of #P(T)\#\mathcal{P}(T)) and using L(h)=0L(h)=0, we find, uniformly for these nn,

An(T)=ehTh2πνnexp⁡(−(T−an)22νn)+oR(ehTT).(34)A_n(T)=\frac{e^{hT}}{h\sqrt{2\pi\nu n}}\exp\left(-\frac{(T-an)^2}{2\nu n}\right)+o_R\left(\frac{e^{hT}}{\sqrt{T}}\right). \tag*{(34)}

as T→∞T\to\infty. This is exploiting the uniformity in TT in Lemma 4.1. Since T−an=−aunTT-an=-au_n\sqrt{T} and n=(T/a)(1+aun/T)n=(T/a)(1+au_n/\sqrt{T}) we have that

(T−an)22νn=a3un22ν(1+OR(T−1/2))=un22σ2+OR(T−1/2);\frac{(T-an)^2}{2\nu n}=\frac{a^3u_n^2}{2\nu}\left(1+O_R(T^{-1/2})\right)=\frac{u_n^2}{2\sigma^2}+O_R(T^{-1/2});

and,

1h2πνn=1haT12πσ2(1+OR(T−1/2)).\frac{1}{h\sqrt{2\pi\nu n}}=\frac{1}{ha\sqrt{T}}\frac{1}{\sqrt{2\pi\sigma^2}}\left(1+O_R(T^{-1/2})\right).

Substituting these into (34) shows that

An(T)=ehThaTφ(un)+oR(ehTT)for ∣un∣≤R,(35)A_n(T)=\frac{e^{hT}}{ha\sqrt{T}}\varphi(u_n)+o_R\left(\frac{e^{hT}}{\sqrt{T}}\right)\qquad\text{for } |u_n|\le R, \tag*{(35)}

where the error is uniform over all such integers nn. Dividing (35) by nn, subtracting the non prime contribution (which are uniformly, exponentially negligible by Lemma 4.3), and then dividing by B(T)B(T) gives

TBn(T)B(T)=φ(un)+o(1)\sqrt{T}\frac{B_n(T)}{B(T)}=\varphi(u_n)+o(1)

as T→∞T\to\infty uniformly for all ∣un∣<R|u_n|<R.

To conclude we need to control the n≥1n\ge1 with ∣un∣≥R|u_n|\ge R. To that end, let δ>0\delta>0 be as in Lemma 4.2. When R<∣un∣≤δTR<|u_n|\le\delta\sqrt{T} we have that

TBn(T)B(T)≤Ce−cR2\sqrt{T}\frac{B_n(T)}{B(T)}\le Ce^{-cR^2}

for some C,c>0C,c>0 by (23). For ∣un∣>δT|u_n|>\delta\sqrt{T} we have an exponentially small upper bound by (24).

Combining all of the above with the fact that

sup⁡∣un∣>Rexp⁡(−(n−ΛT)22σ2T)≤C′e−c′R2\sup_{|u_n|>R}\exp\left(-\frac{(n-\Lambda T)^2}{2\sigma^2T}\right)\le C'e^{-c'R^2}

for some C′,c′>0C',c'>0, we deduce that

lim sup⁡T→∞sup⁡n∈N∣TBn(T)B(T)−12πσ2exp⁡(−(n−ΛT)22σ2T)∣≤Ce−cR2+C′e−c′R2.\limsup_{T\to\infty}\sup_{n\in\mathbb{N}}\left|\sqrt{T}\frac{B_n(T)}{B(T)}-\frac{1}{\sqrt{2\pi\sigma^2}}\exp\left(-\frac{(n-\Lambda T)^2}{2\sigma^2T}\right)\right|\le Ce^{-cR^2}+C'e^{-c'R^2}.

Letting R→∞R\to\infty gives the required asymptotic for Bn(T)B_n(T), concluding the proof. □

Identification and positivity of the constants

In this section we identify the constants a,va,v from Proposition 3.4 with the mean and variance appearing in Theorem 1.1.

Let Px\mathcal{P}_x be the collection of nonempty finite saddle-connection paths starting at a fixed singularity xx as defined in the introduction. For p=(s1,…,sn)p=(s_1,\ldots,s_n) write ℓ(p)=∑jℓ(sj)\ell(p)=\sum_j\ell(s_j) and m(p)=nm(p)=n. By Theorem 1.1 there is a central limit theorem with mean Λ\Lambda and σ2\sigma^2 which are both strictly positive.

For real r>0r>0 and tt near zero set Br,t:=etArB_{r,t}:=e^tA_r, where ArA_r is as in Section 3. For r>0r>0, x∈Xx\in\mathcal{X} and s∈Ss\in\mathcal{S} let

ξr,x(s):=1{i(s)=x}e−rℓ(s)/2,ηr(s):=e−rℓ(s)/2.\xi_{r,x}(s):=\mathbf{1}_{\{i(s)=x\}}e^{-r\ell(s)/2},\qquad\eta_r(s):=e^{-r\ell(s)/2}.

These determine points in ℓ2(S)\ell^2(\mathcal{S}) by Proposition 2.3(i). By construction we have the exact formal identity

∑p∈Pxe−rℓ(p)+tm(p)=et∑k≥0⟨ξr,x,Br,tkηr⟩(36)\sum_{p\in\mathcal{P}_x}e^{-r\ell(p)+tm(p)} =e^t\sum_{k\ge0}\langle\xi_{r,x},B_{r,t}^k\eta_r\rangle \tag*{(36)}

where both sides could be infinite.

Write λ(r)=ρ(Ar)\lambda(r)=\rho(A_r) as in Section 3. For real rr near hh, the positive compact operator Br,tB_{r,t} has simple leading eigenvalue etλ(r)e^t\lambda(r) and strictly positive left and right eigenvectors. Its remaining spectrum is contained in a strictly smaller disk (Proposition 3.4). Therefore the terms in the sum on the right of (36) satisfy

⟨ξr,x,Br,tkηr⟩=c(r,t)(etλ(r))k+Or,t(qr,tk),0≤qr,t<etλ(r),\langle\xi_{r,x},B_{r,t}^k\eta_r\rangle=c(r,t)\left(e^t\lambda(r)\right)^k+O_{r,t}(q_{r,t}^k),\qquad0\le q_{r,t}<e^t\lambda(r),

where c(r,t)>0c(r,t)>0. In particular, the series in (36) converges precisely when

etλ(r)<1.(37)e^t\lambda(r)<1. \tag*{(37)}

We next consider the normalisation in [7]. In [7 Section 3.5] the cost is centred before Proposition 3.12 is applied: the centred cost is ψ(p)=m(p)−Λℓ(p)\psi(p)=m(p)-\Lambda\ell(p). Let K(t)K(t) denote the exponential growth exponent for the centred weighted path count as in [7 Proposition 3.12]. The derivative calculation in [7 Theorem 2.4, Lemma 3.10 and the proof of Theorem 2.6] identifies

K(0)=h,K′(0)=0andK′′(0)=σ2.K(0)=h,\quad K'(0)=0\quad\text{and}\quad K''(0)=\sigma^2.

Centring does not change the second derivative. We also have

∑p∈Pxe−rℓ(p)+tψ(p)=∑p∈Pxe−(r+tΛ)ℓ(p)+tm(p).\sum_{p\in\mathcal{P}_x} e^{-r\ell(p)+t\psi(p)} = \sum_{p\in\mathcal{P}_x} e^{-(r+t\Lambda)\ell(p)+tm(p)}.

It follows that the abscissa of convergence of the uncentred series is the analytic function

H(t)=K(t)+tΛ,H(0)=h,H′(0)=Λ,H′′(0)=σ2.H(t)=K(t)+t\Lambda,\quad H(0)=h,\quad H'(0)=\Lambda,\quad H''(0)=\sigma^2.

Here we have undone the length normalisation h=1h=1 used for the calculations in [7 Section 3]. Comparing (36) with (37) now proves that

etλ(H(t))=1andH(0)=h.e^t\lambda\left(H(t)\right)=1\quad\text{and}\quad H(0)=h.

Finally, let a,va,v be as in Proposition 3.4. Then, differentiating t+log⁡λ(H(t))=0t+\log\lambda(H(t))=0 at t=0t=0 gives

1−aH′(0)=0andv(H′(0))2−aH′′(0)=0.1-aH'(0)=0\quad\text{and}\quad v\left(H'(0)\right)^2-aH''(0)=0.

Therefore the two sets of constants are related by

Λ=1a,σ2=va3,\Lambda=\frac{1}{a},\quad\sigma^2=\frac{v}{a^3},

and a,va,v as in Proposition 3.4 are strictly positive.

References

Mathematical Institute, University of St Andrews, North Haugh, St Andrews, KY16 9SS, Scotland

Email address: [email protected]

References

  1. [1]Jayadev S Athreya and Jon Chaika. The distribution of gaps for saddle connection directions. Geometric and Functional Analysis, 22(6):1491–1516, 2012.arxiv.org/abs/1012.4298
  2. [2]Jayadev S Athreya, Samantha Fairchild, and Howard Masur. Counting pairs of saddle connections. Advances in Mathematics, 431:109233, 2023.DOI
  3. [3]Elisabetta Barletta, Sorin Dragomir, and Francesco Esposito. Cauchy formula for vector-valued holomorphic functions and the Cauchy-Kovalevskaja theorem. Analysis and Geometry in Metric Spaces, 13(1), 2025.DOI
  4. [6]Stephen Cantrell and Mark Pollicott. Comparison theorems for closed geodesics on negatively curved surfaces. Groups, Geometry, and Dynamics, 16(2):461–491, 2022.arxiv.org/abs/2002.09767
  5. [7]Stephen Cantrell and Mark Pollicott. Counting statistics for geodesics on flat surfaces, 2025. arXiv:2503.13091.arxiv.org/abs/2503.13091
  6. [8]Kung-Ching Chang, Xuefeng Wang, and Xie Wu. On the spectral theory of positive operators and PDE applications. Discrete and Continuous Dynamical Systems, 40(6):3171–3200, 2020.DOI
  7. [10]Tanja Eisner, Bálint Farkas, Markus Haase, and Rainer Nagel. Operator Theoretic Aspects of Ergodic Theory, volume 272 of Graduate Texts in Mathematics. Springer, 2015.DOI
  8. [11]Alex Eskin and Howard Masur. Asymptotic formulas on flat surfaces. Ergodic Theory and Dynamical Systems, 21(2):443–478, 2001.DOI
  9. [12]Alex Eskin, Howard Masur, and Anton Zorich. Moduli spaces of Abelian differentials: The principal boundary, counting problems, and the Siegel–Veech constants. Publications Mathématiques de l’IHÉS, 97:61–179, 2003.
  10. [13]R. A. Horn and C. R. Johnson. Matrix Analysis. Cambridge University Press, 2nd edition, 2013.
  11. [15]Steven P. Lalley. Distribution of periodic orbits of symbolic and axiom a flows. Advances in Applied Mathematics, 8:154–193, 1987.DOI
  12. [16]Steven P. Lalley. Renewal theorems in symbolic dynamics, with applications to geodesic flows, noneuclidean tessellations and their fractal limits. Acta Mathematica, 163:1–55, 1989.DOI
  13. [17]H. Masur. The growth rate of trajectories of a quadratic differential. Ergodic Theory Dynam. Systems, 10:151–176, 1990.DOI
  14. [18]J. Milnor. A note on curvature and fundamental group. Journal of Differential Geometry, 2(1):1 – 7, 1968.DOI
  15. [19]Amos Nevo, Rene Rühr, and Barak Weiss. Effective counting on translation surfaces. Advances in Mathematics, 360:106890, 2020.arxiv.org/abs/1708.06263
  16. [20]J. Peterson. Notes on von Neumann algebras. Lecture notes, 2013.
  17. [21]Donald Robertson and Benjamin Dozier. Uniform distribution of saddle connection lengths in all sl (2, r) orbits. Geometriae Dedicata, 217(4):65, 2023.
  18. [22]Richard Sharp. A local limit theorem for closed geodesics and homology. Transactions of the American Mathematical Society, 356(12):4897–4908, 2004.DOI
  19. [23]B. Simon. Trace Ideals and Their Applications, volume 120 of Mathematical Surveys and Monographs. American Mathematical Society, 2nd edition, 2005.
  20. [24]R. Sharp and R. Schwartz. The correlation of length spectra of two hyperbolic surfaces. Communications in Mathematical Physics, 153:423–430, 1993.DOI
  21. [25]C. Stone. A local limit theorem for nonlattice multi-dimensional distribution functions. Ann. Math. Statist., 36:546–551, 1965.DOI

Paper details

vlnc:W_4wwbz8vq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

ChatGPT

View statement
Looking for related papers…
Contents