Acknowledgements
We thank Hao Xu for helpful correspondence on the work of Liu and Xu, the uniformity ranges of the two deformation arguments, and the determination of Zograf’s constant. We acknowledge GeWu SciLab for providing the collaborative research environment in which this project was developed.
Kun Chen is supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDB1680102). Renqiu Xia is supported by the National Natural Science Foundation of China (Grant No. 6260070460).
2020 Mathematics Subject Classification. Primary 32G15; Secondary 11F72, 14H10, 14H81, 41A60, 58J50, 60D05, 60F05.
Key words and phrases. Weil–Petersson volume, moduli space of curves, intersection number, large genus, random hyperbolic surface, spectral gap, local limit theorem.
School of Mathematical Sciences, Peking University, Beijing 100871, China.
Department of Modern Physics, University of Science and Technology of China, Hefei, Anhui 230026, China.
CAS Key Laboratory of Theoretical Physics, Institute of Theoretical Physics, Chinese Academy of Sciences, Beijing 100190, China.
School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China.
Wuhu Qiheng Co., Ltd., Wuhu, China.
Institute of Theoretical Physics, Chinese Academy of Sciences, Beijing 100190, China.
Introduction
Let M g , n ( L ) \mathcal{M}_{g,n}(\mathbf{L}) M g , n ( L ) be the moduli space of hyperbolic surfaces of genus g g g with n n n labeled geodesic boundary components of lengths L = ( L 1 , … , L n ) \mathbf{L}=(L_1,\ldots,L_n) L = ( L 1 , … , L n ) , where 2 g − 2 + n > 0 2g-2+n>0 2 g − 2 + n > 0 and a boundary component of length zero is a cusp. Its Weil–Peterson volume, in the orbifold convention, is
V g , n ( L ) = 1 ( 3 g − 3 + n ) ! ∫ M g , n ( L ) ω W P 3 g − 3 + n , V g , n = V g , n ( 0 , … , 0 ) . V_{g,n}(\mathbf{L})=\frac{1}{(3g-3+n)!}\int_{\mathcal{M}_{g,n}(\mathbf{L})}\omega_{\mathrm{WP}}^{3g-3+n},\qquad V_{g,n}=V_{g,n}(0,\ldots,0). V g , n ( L ) = ( 3 g − 3 + n )! 1 ∫ M g , n ( L ) ω WP 3 g − 3 + n , V g , n = V g , n ( 0 , … , 0 ) .
These volumes connect intersection theory on moduli space with the geometry of random hyperbolic surfaces: their coefficients in the boundary lengths are intersection numbers, and through Mirzakhani’s integration formula their ratios give expected counts of geometric configurations [141, 142, 144]. See [57, 145, 187, 45] for surveys.
Two classical asymptotic theories cover opposite ends of the problem: Mirzakhani and Zograf obtained an expansion as g → ∞ g\to\infty g → ∞ with n n n fixed [147], and Kaufmann–Manin–Zagier and Manin–Zograf obtained the large-n n n asymptotics at fixed genus [117, 136]. In the first expansion the coefficient of order g − i g^{-i} g − i is a polynomial in n n n of degree 2 i 2i 2 i , with leading term ( − 1 ) i n 2 i / ( i ! ( 2 π 2 ) i ) (-1)^i n^{2i}/(i!(2\pi^2)^i) ( − 1 ) i n 2 i / ( i ! ( 2 π 2 ) i ) . These leading terms alone sum to exp ( − n 2 / ( 2 π 2 g ) ) \exp(-n^2/(2\pi^2g)) exp ( − n 2 / ( 2 π 2 g )) , so the effective parameter is n 2 / g n^2/g n 2 / g : once n ≍ g n\asymp\sqrt{g} n ≍ g , terms of every fixed order contribute, and powers of 1 / g 1/g 1/ g no longer order the corrections asymptotically. The two endpoint theories alone therefore do not describe arbitrary simultaneous growth of genus and cusp count.
We extend the large-genus theory across the full range of cusp densities. A single real parameter Y Y Y , the cusp-density parameter , runs over a compact interval [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] as n / g n/g n / g runs over [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] and locates the saddle point of the cusp generating function. Three ideas make the expansion uniform in n n n . Moving along the string flow replaces the alternating sum that expresses Weil–Petersson volumes through ψ \psi ψ intersection numbers by a positive recursion, whose kernels form a compact family over [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] . The dilaton and KdV equations, controlled by this positivity, determine the dependence on Y Y Y ; the remaining normalization is fixed at the endpoint Y = y 0 Y=y_0 Y = y 0 , where only pure ψ \psi ψ intersection numbers enter and Painlevé I supplies their asymptotics. Finally, each generating function is a series with nonnegative coefficients times an explicit function raised to a power proportional to the genus; this factorization converts its asymptotics into asymptotics of individual coefficients, uniformly in n n n . The same parameter connects the volume problem to the geometry of random surfaces: the scales n ≍ g n\asymp\sqrt{g} n ≍ g and n ≍ g n\asymp g n ≍ g mark changes in the low spectrum of these surfaces and in their geometry near a typical point.
Volumes (Part 1). Theorem 1.1 is a single large-genus expansion
V ^ g , n = V ^ g , n ( 1 + A 1 ( g , n ) 2 g − 1 + A 2 ( g , n ) ( 2 g − 1 ) 2 + ⋯ ) \widehat{V}_{g,n}=\widehat{V}_{g,n}\left(1+\frac{A_1(g,n)}{2g-1}+\frac{A_2(g,n)}{(2g-1)^2}+\cdots\right) V g , n = V g , n ( 1 + 2 g − 1 A 1 ( g , n ) + ( 2 g − 1 ) 2 A 2 ( g , n ) + ⋯ )
valid uniformly over every cusp count , with an explicit leading term V ^ g , n \widehat{V}_{g,n} V g , n , coefficients bounded uniformly in n n n , and single-factorial remainder bounds. In the fixed-cusp limit it identifies the constant conjectured by Zograf; at every density it gives the volume-ratio profiles, and for the one-boundary polynomial it gives the adjacent-coefficient asymptotics in all degrees. These results resolve the two questions posed by Mirzakhani [144, Section 1.4]: the asymptotics of V g , n / V g − 1 , n + 2 V_{g,n}/V_{g-1,n+2} V g , n / V g − 1 , n + 2 when n = n ( g ) → ∞ n=n(g)\to\infty n = n ( g ) → ∞ , and the ratios of adjacent coefficients of V g , 1 ( L ) V_{g,1}(L) V g , 1 ( L ) when the degree grows with the genus. Theorems 1.3 and 1.4 give the respective answers uniformly over all cusp counts and all adjacent degrees. Liu and Xu subsequently determined Zograf’s constant by a different deformation [131]; the two approaches are compared at the beginning of Part 1.
Spectrum (Part 2). The threshold n ≍ g n\asymp\sqrt{g} n ≍ g , at which the fixed-cusp expansion breaks down, is also a spectral phase transition. Both are governed by n 2 / g n^2/g n 2 / g : the expected number of geodesics of bounded length that cut off two cusps is of this order (Subsection 1.6). Huang showed that the spectral gap of a Weil–Petersson random surface tends to 1 / 4 1/4 1/4 when n = o ( g ) n=o(\sqrt{g}) n = o ( g ) [101]. At the critical scale the full positive spectral point process below 1 / 4 1/4 1/4 converges to a Poisson process of explicit model energies, one for each short two-cusp pair of pants (Theorem 15.1); the gap therefore has a nondegenerate random limit, and tends to zero above the scale. Above the critical scale the lowest eigenvalues follow rescaled Poisson laws in specified ranges, and fixed-energy windows carry deterministic limiting spectral measures. The volume theorem supplies the statistics of short curves; the spectral work is to show that the states these curves create exhaust the low spectrum.
Local geometry (Part 3). At positive density n / g → λ ∈ ( 0 , ∞ ) n/g \to\lambda\in(0,\infty) n / g → λ ∈ ( 0 , ∞ ) the same volume ratios, with based holonomies retained in the integration formula, determine the surface seen from a uniformly chosen point and the conformal type of its cusp-filling. As λ → ∞ \lambda\to\infty λ → ∞ this family reaches the area-rooted infinite WP sphere of Budd and Curien [31]. We prove that this endpoint is parabolic, with recurrent Brownian motion and ergodic geodesic flow, as they conjectured [31, Section 7.3], whereas at every finite density the cusp-filled surface is conformally a disc. A growing-radius coupling of their tree models shows that the endpoint surface, rescaled and with its deep cusps trimmed, converges to the Brownian plane, answering another of their questions.
The common step is geometric integration: a cut expresses the law of a finite configuration through volumes of its complementary surfaces. For spectra, these configuration counts must be matched to the entire low spectral subspace; for local geometry, the relative positions of the based loops must also be retained. Figure 1 shows how these two uses of the volume theorem lead to the applications; Table 2 lists the questions from the literature that are answered here.
Figure 1. One volume theory, two geometric uses. Both branches use volume ratios through geometric integration and then the additional arguments inside their boxes. The labelled arrow is the proved fixed-radius local limit as λ → ∞ \lambda\to\infty λ → ∞ . The two arrows into the bottom box show the endpoint surface and the separate coupling argument used for its Brownian-plane scaling limit. Part 3 uses no spectral theorem from Part 2. The all-degree coefficient theorem in Part 1 also uses harmonic comparison directly.
Table 2. Questions from the literature answered or treated in this paper.
The cusp-density parameter
All our formulas are expressed through a single real parameter. We define it first and then explain heuristically why it is the natural parameter; the proofs come later.
The genus-zero relation. Let
X ( y ) = y J 1 ( 2 y ) = y − y 2 2 + y 3 12 − ⋯ , y 0 = j 0 , 1 2 4 , x c = X ( y 0 ) , X(y)=\sqrt{y}J_1(2\sqrt{y})=y-\frac{y^2}{2}+\frac{y^3}{12}-\cdots,\qquad y_0=\frac{j_{0,1}^2}{4},\qquad x_c=X(y_0), X ( y ) = y J 1 ( 2 y ) = y − 2 y 2 + 12 y 3 − ⋯ , y 0 = 4 j 0 , 1 2 , x c = X ( y 0 ) ,
where j 0 , 1 j_{0,1} j 0 , 1 is the first positive zero of J 0 J_0 J 0 . Since X ′ ( y ) = J 0 ( 2 y ) X'(y)=J_0(2\sqrt{y}) X ′ ( y ) = J 0 ( 2 y ) , the function X X X increases from 0 0 0 to x c x_c x c on [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] ; write Y ( x ) Y(x) Y ( x ) for the inverse function. Consider the normalized cusp generating functions
Φ g ( x ) = ∑ n ≥ 0 V g , n ( 2 π 2 ) 3 g − 3 + n x n n ! . (1) \Phi_g(x)=\sum_{n\geq0}\frac{V_{g,n}}{(2\pi^2)^{3g-3+n}}\frac{x^n}{n!}. \tag*{(1)} Φ g ( x ) = n ≥ 0 ∑ ( 2 π 2 ) 3 g − 3 + n V g , n n ! x n . (1)
with the convention that terms with 2 g − 2 + n ≤ 0 2g-2+n\leq0 2 g − 2 + n ≤ 0 are omitted. In genus zero, Φ 0 ′ ′ ( x ) = Y ( x ) \Phi_0''(x)=Y(x) Φ 0 ′′ ( x ) = Y ( x ) ; this is the classical genus-zero computation (cf. [117, 136]), and in our normalization it follows from the string equation (see Step 1 of the proof of Theorem 14.2 in Appendix § 10.1 ). We call x x x the cusp weight: it weights each cusp in {(1.1)}. The cusp weight x x x and the parameter Y = Y ( x ) ∈ [ 0 , y 0 ) Y=Y(x)\in[0,y_0) Y = Y ( x ) ∈ [ 0 , y 0 ) determine each other, and we use whichever is more convenient.
The parameter functions. For 0 ≤ Y < y 0 0\leq Y<y_0 0 ≤ Y < y 0 , the entire function
K Y ( z ) = ∫ 0 1 X ′ ( Y + z 2 ( 1 − s 2 ) ) d s K_Y(z)=\int_0^1 X'\left(Y+\frac{z}{2}(1-s^2)\right)\,ds K Y ( z ) = ∫ 0 1 X ′ ( Y + 2 z ( 1 − s 2 ) ) d s
has only positive simple zeros (§ 3.1 ); it is the kernel of the recursion in § 3.4 . Write its first zero as 2 Δ ( Y ) 2\Delta(Y) 2Δ ( Y ) and put T ( Y ) = Y + Δ ( Y ) T(Y)=Y+\Delta(Y) T ( Y ) = Y + Δ ( Y ) . The substitution v = T − Δ s 2 v=T-\Delta s^2 v = T − Δ s 2 shows that T = T ( Y ) T=T(Y) T = T ( Y ) is the unique root in ( y 0 , π 2 / 4 ) (y_0,\pi^2/4) ( y 0 , π 2 /4 ) of
∫ Y T X ′ ( v ) T − v d v = 0. (2) \int_Y^T\frac{X'(v)}{\sqrt{T-v}}\,dv=0. \tag*{(2)} ∫ Y T T − v X ′ ( v ) d v = 0. (2)
Finally define
I ( Y ) = ∫ Y T ( Y ) X ′ ( v ) T ( Y ) − v d v , λ ( Y ) = 2 X ( Y ) Δ ( Y ) I ( Y ) . (3) I(Y)=\int_Y^{T(Y)}X'(v)\sqrt{T(Y)-v}\,dv,\qquad\lambda(Y)=\frac{2X(Y)\sqrt{\Delta(Y)}}{I(Y)}. \tag*{(3)} I ( Y ) = ∫ Y T ( Y ) X ′ ( v ) T ( Y ) − v d v , λ ( Y ) = I ( Y ) 2 X ( Y ) Δ ( Y ) . (3)
The next paragraph motivates these definitions. In § 3.3 we show that I > 0 I>0 I > 0 and that λ \lambda λ is a strictly increasing bijection from [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) onto [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . At Y = 0 Y=0 Y = 0 one has Δ = T = π 2 / 4 \Delta=T=\pi^2/4 Δ = T = π 2 /4 , I = π / 4 I=\pi/4 I = π /4 and λ = 0 \lambda=0 λ = 0 ; as Y ↑ y 0 Y \uparrow y_0 Y ↑ y 0 , Δ → 0 \Delta\to0 Δ → 0 and λ → ∞ \lambda\to\infty λ → ∞ .
Why λ \lambda λ is the cusp density. Let h = 2 g − 1 h=2g-1 h = 2 g − 1 and x = X ( Y ) x=X(Y) x = X ( Y ) . We shall prove (Proposition 7.1) that, uniformly for 0 ≤ Y < y 0 0 \le Y < y_0 0 ≤ Y < y 0 ,
Φ g ′ ( x ) = Γ ( h ) ϖ ( Y ) π 4 g π h ( 2 I ( Y ) ) − h ( 1 + O ( log g g ) ) . (4) \Phi'_g(x)=\frac{\Gamma(h)\varpi(Y)}{\pi4^g\sqrt{\pi h}}\left(\sqrt{2}I(Y)\right)^{-h}\left(1+O\left(\frac{\log g}{g}\right)\right). \tag*{(4)} Φ g ′ ( x ) = π 4 g π h Γ ( h ) ϖ ( Y ) ( 2 I ( Y ) ) − h ( 1 + O ( g log g ) ) . (4)
where ϖ \varpi ϖ is the explicit positive function in (6) below. Thus the cusp weight changes the exponential growth only through the action I ( Y ) I(Y) I ( Y ) .
Now regard the terms of the series Φ g ′ ( x ) = ∑ k ≥ 0 V g , k + 1 x k / ( ( 2 π 2 ) 3 g − 2 + k k ! ) \Phi'_g(x)=\sum_{k\geq0} V_{g,k+1}x^k/((2\pi^2)^{3g-2+k}k!) Φ g ′ ( x ) = ∑ k ≥ 0 V g , k + 1 x k / (( 2 π 2 ) 3 g − 2 + k k !) as the weights of a probability distribution on the number k k k of additional cusps. Its mean is x d log Φ g ′ / d x x\,d\log\Phi'_g/dx x d log Φ g ′ / d x . Since I ′ ( Y ) = − X ′ ( Y ) Δ ( Y ) I'(Y)=-X'(Y)\sqrt{\Delta(Y)} I ′ ( Y ) = − X ′ ( Y ) Δ ( Y ) and d Y / d x = 1 / X ′ ( Y ) dY/dx=1/X'(Y) d Y / d x = 1/ X ′ ( Y ) , the leading term in (4) gives
x d d x log ( 2 I ( Y ) ) − h = h X ( Y ) Δ ( Y ) I ( Y ) = h λ ( Y ) 2 . x\frac{d}{dx}\log\left(\sqrt{2I(Y)}\right)^{-h}
= h\frac{X(Y)\sqrt{\Delta(Y)}}{I(Y)}
= \frac{h\lambda(Y)}{2}. x d x d log ( 2 I ( Y ) ) − h = h I ( Y ) X ( Y ) Δ ( Y ) = 2 hλ ( Y ) .
So, under the cusp weight x x x , the number of additional cusps is about h λ ( Y ) / 2 ≈ g λ ( Y ) h\lambda(Y)/2\approx g\lambda(Y) hλ ( Y ) /2 ≈ g λ ( Y ) . To recover a single volume V g , n V_{g,n} V g , n , one chooses Y Y Y so that the leading approximation to this mean equals n − 1 n-1 n − 1 , that is, λ ( Y ) = 2 ( n − 1 ) / ( 2 g − 1 ) \lambda(Y)=2(n-1)/(2g-1) λ ( Y ) = 2 ( n − 1 ) / ( 2 g − 1 ) , and estimates the probability of the value k = n − 1 k=n-1 k = n − 1 by a local limit theorem. In particular λ ( Y ) \lambda(Y) λ ( Y ) is, to leading order, the cusp density n / g n/g n / g , so the profiles may be expressed in either Y Y Y or λ ( Y ) \lambda(Y) λ ( Y ) . This heuristic produces the leading term of Theorem 1.1. As n / g n/g n / g increases from 0 0 0 to ∞ \infty ∞ , the parameter Y Y Y increases from 0 0 0 to y 0 y_0 y 0 and the cusp weight x x x from 0 0 0 to x c x_c x c . Subsection 1.5 explains how endpoint normalization and Fourier bounds make this heuristic uniform in n n n .
A matrix-model interpretation. The cusp weight is a defect weight in Jackiw–Teitelboim (JT) gravity, and Y Y Y locates the edge of the corresponding spectral density. At the endpoint Y = y 0 Y=y_0 Y = y 0 the square-root edge becomes cubic. Subsection 13.8 gives the precise dictionary with the defect-deformed matrix model of Maxfield–Turiaci and Witten [137, 181, 61]. In this language the endpoint corresponds to the ( 2 , 3 ) (2,3) ( 2 , 3 ) critical point, where our normalized recursion reduces to pure ψ \psi ψ intersection numbers. The interpretation does not by itself control the remainders.
Absolute volumes and Zograf’s constant
For n ≥ 1 n\geq1 n ≥ 1 put h = 2 g − 1 h=2g-1 h = 2 g − 1 and k = n − 1 k=n-1 k = n − 1 , let Y ∈ [ 0 , y 0 ) Y\in[0,y_0) Y ∈ [ 0 , y 0 ) be the solution of λ ( Y ) = 2 k / h \lambda(Y)=2k/h λ ( Y ) = 2 k / h , and put x = X ( Y ) x=X(Y) x = X ( Y ) . Define
V ^ g , n = ( 2 π 2 ) 3 g − 3 + n Γ ( h ) π 4 g π h ϖ ( Y ) ( 2 I ( Y ) ) h ( k e x ) k k h x Ψ ′ ( x ) . (5) \widehat{V}_{g,n}
=
\frac{(2\pi^2)^{3g-3+n}\Gamma(h)}{\pi\,4^g\sqrt{\pi h}}
\frac{\varpi(Y)}{(\sqrt{2}I(Y))^h}
\left(\frac{k}{ex}\right)^k
\sqrt{\frac{k}{h x\Psi'(x)}}. \tag*{(5)} V g , n = π 4 g π h ( 2 π 2 ) 3 g − 3 + n Γ ( h ) ( 2 I ( Y ) ) h ϖ ( Y ) ( e x k ) k h x Ψ ′ ( x ) k . (5)
where Ψ ( X ( y ) ) = λ ( y ) / 2 \Psi(X(y))=\lambda(y)/2 Ψ ( X ( y )) = λ ( y ) /2 for 0 ≤ y < y 0 0\leq y<y_0 0 ≤ y < y 0 , and ϖ ( Y ) > 0 \varpi(Y)>0 ϖ ( Y ) > 0 is given by
ϖ ( Y ) 2 = I ( Y ) Δ ( Y ) 3 / 2 ( ∫ 0 1 ( 1 − s 2 ) X ( v ) v d s ) − 1 , v = T ( Y ) − Δ ( Y ) s 2 ; (6) \varpi(Y)^2
=
\frac{I(Y)}{\Delta(Y)^{3/2}}
\left(\int_0^1(1-s^2)\frac{X(v)}{v}\,ds\right)^{-1},
\qquad
v=T(Y)-\Delta(Y)s^2; \tag*{(6)} ϖ ( Y ) 2 = Δ ( Y ) 3/2 I ( Y ) ( ∫ 0 1 ( 1 − s 2 ) v X ( v ) d s ) − 1 , v = T ( Y ) − Δ ( Y ) s 2 ; (6)
here X ( v ) / v X(v)/v X ( v ) / v is an entire function of v v v , equal to one at v = 0 v=0 v = 0 . For n = 0 n=0 n = 0 we put V ^ g , 0 = Γ ( 2 g − 2 ) ( 4 π 2 ) 2 g − 3 / π g \widehat{V}_{g,0}=\Gamma(2g-2)(4\pi^2)^{2g-3}/\sqrt{\pi g} V g , 0 = Γ ( 2 g − 2 ) ( 4 π 2 ) 2 g − 3 / π g .
The formula (5) is the heuristic of Subsection 1.1 made explicit. Its first two factors are ( 2 π 2 ) 3 g − 3 + n (2\pi^2)^{3g-3+n} ( 2 π 2 ) 3 g − 3 + n times the approximation (4) of Φ g ′ ( x ) \Phi'_g(x) Φ g ′ ( x ) . The coefficient of x k x^k x k is about Φ g ′ ( x ) x − k \Phi'_g(x)x^{-k} Φ g ′ ( x ) x − k times the probability of the center of the distribution. To leading order, k ! k! k ! times that probability is ( k / e ) k k / ( h x Ψ ′ ( x ) ) (k/e)^k\sqrt{k/(h x\Psi'(x))} ( k / e ) k k / ( h x Ψ ′ ( x )) , where h x Ψ ′ ( x ) h x\Psi'(x) h x Ψ ′ ( x ) is the variance scale supplied by the leading action. We use the Poisson point mass e − k k k / k ! e^{-k}k^k/k! e − k k k / k ! at the center rather than its Gaussian approximation ( 2 π k ) − 1 / 2 (2\pi k)^{-1/2} ( 2 π k ) − 1/2 , so that the same formula remains correct when the cusp count stays bounded; there the number of additional cusps is approximately Poisson rather than Gaussian. At k = 0 k=0 k = 0 the last two factors in (5) are interpreted as their limits, which are one; then Y = 0 Y=0 Y = 0 , ϖ ( 0 ) = 1 \varpi(0)=1 ϖ ( 0 ) = 1 and V ^ g , 1 = Γ ( 2 g − 1 ) ( 4 π 2 ) 2 g − 2 / π ( g − 1 / 2 ) \widehat{V}_{g,1}=\Gamma(2g-1)(4\pi^2)^{2g-2}/\sqrt{\pi(g-1/2)} V g , 1 = Γ ( 2 g − 1 ) ( 4 π 2 ) 2 g − 2 / π ( g − 1/2 ) . The case n = 0 n=0 n = 0 is reduced to n = 1 n=1 n = 1 by the dilaton identity (67) at p = 0 p=0 p = 0 , which for volume polynomials is the relation ∂ L V g , 1 ( 2 π i ) = 2 π i ( 2 g − 2 ) V g , 0 \partial_L V_{g,1}(2\pi i)=2\pi i(2g-2)V_{g,0} ∂ L V g , 1 ( 2 π i ) = 2 π i ( 2 g − 2 ) V g , 0 of Do and Norbury [58, Theorem 2] (Section 11).
The correction coefficients A j ( g , n ) A_j(g,n) A j ( g , n ) below, with A 0 = 1 A_0=1 A 0 = 1 , are determined by finite recursive rules from the kernels together with one-dimensional integrals in Y Y Y (see (140) and (149) ); the first is written out in (E.29). For n ≥ 1 n\geq1 n ≥ 1 they are analytic functions of the parameter Y Y Y , including both endpoints of its interval.
Theorem 1.1 (Uniform absolute expansion). There are constants C , B , c , g 0 > 0 C, B, c, g_0 > 0 C , B , c , g 0 > 0 such that, for every g ≥ g 0 g \ge g_0 g ≥ g 0 and every integer 0 ≤ K ≤ c ( 2 g − 1 ) 0 \le K \le c(2g-1) 0 ≤ K ≤ c ( 2 g − 1 ) ,
sup n ≥ 0 ∣ V g , n V ‾ g , n − ∑ j = 0 K A j ( g , n ) ( 2 g − 1 ) j ∣ ≤ C B K + 1 ( K + 1 ) ! ( 2 g − 1 ) K + 1 , \sup_{n \ge0}\left|\frac{V_{g,n}}{\overline{V}_{g,n}}-\sum_{j=0}^{K}\frac{A_j(g,n)}{(2g-1)^j}\right|\le\frac{CB^{K+1}(K+1)!}{(2g-1)^{K+1}}, n ≥ 0 sup V g , n V g , n − j = 0 ∑ K ( 2 g − 1 ) j A j ( g , n ) ≤ ( 2 g − 1 ) K + 1 C B K + 1 ( K + 1 )! , and the coefficients satisfy ∣ A j ( g , n ) ∣ ≤ C B j j ! |A_j(g,n)| \le CB^j j! ∣ A j ( g , n ) ∣ ≤ C B j j ! for all g ≥ g 0 g \ge g_0 g ≥ g 0 and n ≥ 0 n \ge0 n ≥ 0 . In particular,
sup n ≥ 0 ∣ V g , n V ‾ g , n − 1 ∣ = O ( g − 1 ) . \sup_{n \ge0}\left|\frac{V_{g,n}}{\overline{V}_{g,n}}-1\right|=O(g^{-1}). n ≥ 0 sup V g , n V g , n − 1 = O ( g − 1 ) . Taking K K K to be a sufficiently small fixed multiple of g g g , rounded down to an integer, gives relative error O ( e − a g ) O(e^{-ag}) O ( e − a g ) for some a > 0 a > 0 a > 0 , uniformly in n n n .
The uniformity in n n n is the main point: Theorem 1.1 replaces the fixed-n n n series in n 2 / g n^2/g n 2 / g by an expansion in 1 / g 1/g 1/ g alone. The same estimate applies when n n n is fixed, when n n n is proportional to g g g , and when n / g → ∞ n/g \to\infty n / g → ∞ . In the first regime the saddle is near x = 0 x=0 x = 0 ; as n / g → ∞ n/g \to\infty n / g → ∞ it tends to the radius of convergence x c x_c x c of the fixed-genus generating functions of Manin and Zograf. In this way a single formula joins the two classical regimes. The order g − 1 g^{-1} g − 1 is sharp for this choice of leading term (Theorem 11.1 ); including the first correction (E.29), or (149) for n = 0 n=0 n = 0 , gives error O ( g − 2 ) O(g^{-2}) O ( g − 2 ) . The proof is completed in § 3.31 .
For fixed n n n , Mirzakhani and Zograf proved that V g , n ∼ C ( 2 g − 3 + n ) ! ( 4 π 2 ) 2 g − 3 + n / g V_{g,n} \sim C(2g-3+n)!(4\pi^2)^{2g-3+n}/\sqrt{g} V g , n ∼ C ( 2 g − 3 + n )! ( 4 π 2 ) 2 g − 3 + n / g for a universal constant C > 0 C>0 C > 0 [147, Theorem 1.2]. Zograf had conjectured, on the basis of numerical data, that C = 1 / π C=1/\sqrt{\pi} C = 1/ π [195]; see also [147, Remark 1.3]. The same value appears in physical predictions. Saad, Shenker and Stanford obtained the unmarked large-genus prediction from the JT matrix integral under a resurgence assumption [167, Section 5.6]. Okuyama and Sakai extracted the constant 1 / π 1/\sqrt{\pi} 1/ π , together with first corrections for the coefficients of V g , 1 ( L ) V_{g,1}(L) V g , 1 ( L ) at each fixed degree, numerically from the KdV equation [156, Section 2.4, equations (2.87)–(2.89)]. Kimura conjectured the corresponding asymptotics of V g , n ( L ) V_{g,n}(L) V g , n ( L ) from JT gravity [119, equation (28)], and Eynard, Garcia-Failde, Gregori, Lewański and Schiappa derived it from the nonperturbative topological recursion, assuming that the generating series of Weil–Petersson volumes is resurgent [66, Section 6.1, equation (6.7)].
Our argument does not assume that the Weil–Petersson genus expansion is resurgent. We determine the normalization constant at the endpoint Y = y 0 Y=y_0 Y = y 0 of the parameter interval, where the problem reduces to pure ψ \psi ψ intersection numbers and only the single sequence ⟨ τ 3 g − 3 ⟩ g \langle\tau_{3g-3}\rangle_g ⟨ τ 3 g − 3 ⟩ g enters. An identity of Itzykson and Zuber [106] expresses this sequence through the formal power-series solution of Painlevé I, the string equation of two-dimensional pure gravity in the double-scaling limit [28, 60, 79], whose large-order asymptotics follow from the results of Joshi–Kitaev, Kapaev, and Garoufalidis, Its, Kapaev and Mariño [114, 116, 70]. The leading term and the first correction of this sequence also appear as a special case of the uniform asymptotics for ψ \psi ψ intersection numbers obtained by Guo, Yang and Zagier [83]. We derive both from Painlevé I in Subsections 7.3 and D.11 and use [83] only as an independent check. Comparison with the Mirzakhani–Zograf expansion gives the following corollary.
Corollary 1.2 (Zograf’s constant). For every fixed n ≥ 0 n \ge0 n ≥ 0 , as g → ∞ g \to\infty g → ∞ ,
V g , n = ( 2 g − 3 + n ) ! ( 4 π 2 ) 2 g − 3 + n π g ( 1 + O n ( g − 1 ) ) . V_{g,n}=\frac{(2g-3+n)!(4\pi^2)^{2g-3+n}}{\sqrt{\pi g}}\left(1+O_n(g^{-1})\right). V g , n = π g ( 2 g − 3 + n )! ( 4 π 2 ) 2 g − 3 + n ( 1 + O n ( g − 1 ) ) . This is proved at the end of § 3.19 . Together with the bracket estimate (6.15), it also gives the formula conjectured by Kimura [119, equation (28)]: for fixed n ≥ 1 n \ge1 n ≥ 1 and nonnegative lengths
L i = o ( g ) L_i=o(\sqrt{g}) L i = o ( g ) ,
V g , n ( L ) ∼ 2 π ( 4 π 2 ) 2 g − 3 + n Γ ( 2 g + n − 5 2 ) ∏ i = 1 n sinh ( L i / 2 ) L i / 2 . V_{g,n}(\mathbf{L})\sim\sqrt{\frac{2}{\pi}}(4\pi^2)^{2g-3+n}\Gamma\left(2g+n-\frac{5}{2}\right)\prod_{i=1}^{n}\frac{\sinh(L_i/2)}{L_i/2}. V g , n ( L ) ∼ π 2 ( 4 π 2 ) 2 g − 3 + n Γ ( 2 g + n − 2 5 ) i = 1 ∏ n L i /2 sinh ( L i /2 ) .
The factors at L i = 0 L_i=0 L i = 0 are interpreted by continuity. A direct derivation is given in (104) below.
Bounded geodesic boundaries. Proposition 9.4 extends the leading comparison to a fixed number of nonzero boundary lengths in a compact set, uniformly over all cusp counts. Its explicit factor interpolates, as n / g n/g n / g runs from zero to infinity, between the fixed-cusp factor sinh ( L / 2 ) / ( L / 2 ) \sinh(L/2)/(L/2) sinh ( L /2 ) / ( L /2 ) of [146] and the fixed-genus, large-cusp factor I 0 ( j 0 , 1 L / ( 2 π ) ) I_0(j_{0,1}L/(2\pi)) I 0 ( j 0 , 1 L / ( 2 π )) of [98]. The interpolation is given in (130) ; in this extension both the number and the lengths of the nonzero boundaries remain bounded.
Volume ratios
For sufficiently large g g g , define
B g , n = V g , n V g − 1 , n + 2 , C g , n = V g , n + 1 2 g V g , n , D g , n = V g , n 2 V g , n − 1 V g , n + 1 . B_{g,n}=\frac{V_{g,n}}{V_{g-1,n+2}},\qquad C_{g,n}=\frac{V_{g,n+1}}{2gV_{g,n}},\qquad D_{g,n}=\frac{V_{g,n}^{2}}{V_{g,n-1}V_{g,n+1}}. B g , n = V g − 1 , n + 2 V g , n , C g , n = 2 g V g , n V g , n + 1 , D g , n = V g , n − 1 V g , n + 1 V g , n 2 .
The first two quantities are defined for n ≥ 0 n\ge0 n ≥ 0 , and the third for n ≥ 1 n\ge1 n ≥ 1 . For each fixed n n n , Mirzakhani proved that B g , n → 1 B_{g,n}\to1 B g , n → 1 and C g , n → 4 π 2 C_{g,n}\to4\pi^2 C g , n → 4 π 2 [144, Theorem 3.5], and she asked for the asymptotics of B g , n B_{g,n} B g , n when n = n ( g ) → ∞ n=n(g)\to\infty n = n ( g ) → ∞ [144, Section 1.4]. An estimate of Mirzakhani and Zograf states that ∣ ( 2 g − 2 + n ) V g , n / V g , n + 1 − 1 / ( 4 π 2 ) ∣ ≤ c n / ( 2 g − 2 + n ) \left|(2g-2+n)V_{g,n}/V_{g,n+1}-1/(4\pi^2)\right|\le cn/(2g-2+n) ( 2 g − 2 + n ) V g , n / V g , n + 1 − 1/ ( 4 π 2 ) ≤ c n / ( 2 g − 2 + n ) for stable ( g , n ) (g,n) ( g , n ) with n ≥ 1 n\ge1 n ≥ 1 [147, Lemma 5.1]; together with the fixed-n n n result, this gives C g , n → 4 π 2 C_{g,n}\to4\pi^2 C g , n → 4 π 2 whenever n = o ( g ) n=o(g) n = o ( g ) . When n n n is comparable to or larger than g g g , the asymptotic profiles depend on the density n / g n/g n / g , through the functions
B ( λ ( Y ) ) = π 2 4 Δ ( Y ) , C ( λ ( Y ) ) = λ ( Y ) π 2 X ( Y ) . (7) \mathcal{B}(\lambda(Y))=\frac{\pi^2}{4\Delta(Y)},\qquad\mathcal{C}(\lambda(Y))=\frac{\lambda(Y)\pi^2}{X(Y)}. \tag*{(7)} B ( λ ( Y )) = 4Δ ( Y ) π 2 , C ( λ ( Y )) = X ( Y ) λ ( Y ) π 2 . (7)
They are well defined because λ \lambda λ is a bijection, and they take the continuous values B ( 0 ) = 1 \mathcal{B}(0)=1 B ( 0 ) = 1 and C ( 0 ) = 4 π 2 \mathcal{C}(0)=4\pi^2 C ( 0 ) = 4 π 2 . Heuristically, C \mathcal{C} C appears because x Φ g ′ ′ ( x ) / Φ g ′ ( x ) x\Phi_g''(x)/\Phi_g'(x) x Φ g ′′ ( x ) / Φ g ′ ( x ) is the exact mean of the additional-cusp distribution and the saddle is chosen so that its leading approximation equals n − 1 n-1 n − 1 ; B \mathcal{B} B appears because Φ g − 1 ′ ′ ( x ) / Φ g ′ ( x ) \Phi_{g-1}''(x)/\Phi_g'(x) Φ g − 1 ′′ ( x ) / Φ g ′ ( x ) is approximately 8 Δ ( Y ) 8\Delta(Y) 8Δ ( Y ) . See (83) and (84) .
Theorem 1.3 (Uniform volume ratios). There are constants C 0 , g 0 > 0 C_0,g_0>0 C 0 , g 0 > 0 such that, for every g ≥ g 0 g\ge g_0 g ≥ g 0 ,
sup n ≥ 0 ∣ B g , n B ( n / g ) − 1 ∣ ≤ C 0 g , sup n ≥ 0 ∣ C g , n C ( n / g ) − 1 ∣ ≤ C 0 g , \sup_{n\ge0}\left|\frac{B_{g,n}}{\mathcal{B}(n/g)}-1\right|\le\frac{C_0}{g},\qquad
\sup_{n\ge0}\left|\frac{C_{g,n}}{\mathcal{C}(n/g)}-1\right|\le\frac{C_0}{g}, n ≥ 0 sup B ( n / g ) B g , n − 1 ≤ g C 0 , n ≥ 0 sup C ( n / g ) C g , n − 1 ≤ g C 0 , and
sup n ≥ 1 ∣ D g , n − 1 ∣ ≤ C 0 g . \sup_{n\ge1}|D_{g,n}-1|\le\frac{C_0}{g}. n ≥ 1 sup ∣ D g , n − 1∣ ≤ g C 0 . In particular, along every path n / g → ∞ n/g\to\infty n / g → ∞ with g → ∞ g\to\infty g → ∞ ,
B g , n ∼ π 2 3 5 j 0 , 1 n / g , C g , n ∼ π 2 x c n g . B_{g,n}\sim\frac{\pi^2}{3\sqrt{5}\,j_{0,1}}\sqrt{n/g},\qquad C_{g,n}\sim\frac{\pi^2}{x_c}\frac{n}{g}. B g , n ∼ 3 5 j 0 , 1 π 2 n / g , C g , n ∼ x c π 2 g n .
Section 9 proves the uniform limits directly from a local limit theorem. The inverse-genus error follows from the sharper coefficient estimates, as shown in Section 11.1. The limit of C g , n C_{g,n} C g , n was known for n = o ( g ) n=o(g) n = o ( g ) , as described above. For B g , n B_{g,n} B g , n , the absolute estimate of Mirzakhani and Zograf [147, Theorem 1.8] gives the limit one when n = o ( g ) n=o(\sqrt{g}) n = o ( g ) ; Gatignol and Thomas [71, Theorem 1.5] extend it to n = o ( g ) n=o(g) n = o ( g ) . Theorem 1.3 also covers positive and unbounded cusp density and gives relative error O ( g − 1 ) O(g^{-1}) O ( g − 1 ) uniformly in n n n .
The statement on D g , n D_{g,n} D g , n answers a question of Shen and Wu [172, Question 3], which asks whether D g , n → 1 D_{g,n}\to1 D g , n → 1 as g + n → ∞ g+n\to\infty g + n → ∞ . As they observe, the answer is affirmative when n = o ( g ) n=o(g) n = o ( g ) , by [147, Lemma 5.1], and when g g g is fixed, by [136, Theorem 6.1]; they also proved 1 / 2 ≤ lim inf D g , n ≤ lim sup D g , n ≤ 1 1/2\le\liminf D_{g,n}\le\limsup D_{g,n}\le1 1/2 ≤ lim inf D g , n ≤ lim sup D g , n ≤ 1 for every cusp sequence [172, Proposition 36]. Theorem 1.3 supplies the remaining range, in which g → ∞ g \to\infty g → ∞ and n / g ↛ 0 n/g \not\to0 n / g → 0 . Combined with the fixed-genus case, it gives D g , n → 1 D_{g,n} \to1 D g , n → 1 whenever g + n → ∞ g+n \to\infty g + n → ∞ .
Adjacent coefficients
Mirzakhani’s second question [144, Section 1.4] concerns adjacent coefficients when the degree grows with the genus. We write the one-boundary volume polynomial in the normalization of [144],
V g , 1 ( L ) = ∑ k = 0 N a g , k ( 2 k + 1 ) ! L 2 k , N = 3 g − 2. V_{g,1}(L)=\sum_{k=0}^{N}\frac{a_{g,k}}{(2k+1)!}L^{2k},\qquad N=3g-2. V g , 1 ( L ) = k = 0 ∑ N ( 2 k + 1 )! a g , k L 2 k , N = 3 g − 2.
The normalized intersection bracket b g , k = [ τ k ] g , 1 b_{g,k}=[\tau_k]_{g,1} b g , k = [ τ k ] g , 1 is the corresponding coefficient of V g , 1 ( 2 L ) V_{g,1}(2L) V g , 1 ( 2 L ) , and hence b g , k = 4 k a g , k b_{g,k}=4^k a_{g,k} b g , k = 4 k a g , k (§ 3.4 ). For fixed k k k , its adjacent ratio tends to one [144, Theorem 3.5]; the corresponding ratio a g , k / a g , k + 1 a_{g,k}/a_{g,k+1} a g , k / a g , k + 1 tends to four. Thus the limit one stated in [144, Section 1.4] refers to the bracket normalization, which is also the one used in [129].
Theorem 1.4 (All adjacent degrees). For 0 ≤ k < N 0 \le k < N 0 ≤ k < N , let t g , k ∈ ( 0 , π ) t_{g,k} \in(0,\pi) t g , k ∈ ( 0 , π ) be determined by
2 k + 1 2 g − 1 = t g , k 2 sin t g , k sin t g , k − t g , k cos t g , k . \frac{2k+1}{2g-1}=\frac{t_{g,k}^{2}\sin t_{g,k}}{\sin t_{g,k}-t_{g,k}\cos t_{g,k}}. 2 g − 1 2 k + 1 = sin t g , k − t g , k cos t g , k t g , k 2 sin t g , k . Then
sup 0 ≤ k < 3 g − 2 ∣ a g , k / a g , k + 1 4 π 2 / t g , k 2 − 1 ∣ = O ( g − 1 / 8 ) . \sup_{0\le k<3g-2}\left|\frac{a_{g,k}/a_{g,k+1}}{4\pi^{2}/t_{g,k}^{2}}-1\right|=O\left(g^{-1/8}\right). 0 ≤ k < 3 g − 2 sup 4 π 2 / t g , k 2 a g , k / a g , k + 1 − 1 = O ( g − 1/8 ) . In particular, a g , k / a g , k + 1 → 4 a_{g,k}/a_{g,k+1}\to4 a g , k / a g , k + 1 → 4 for every k = o ( g ) k=o(g) k = o ( g ) , and, with m = N − k ≥ 1 m=N-k\ge1 m = N − k ≥ 1 ,
a g , N − m a g , N − m + 1 ∼ 4 π 2 g 5 m if m = o ( g ) . \frac{a_{g,N-m}}{a_{g,N-m+1}}\sim\frac{4\pi^{2}g}{5m}\qquad\text{if }m=o(g). a g , N − m + 1 a g , N − m ∼ 5 m 4 π 2 g if m = o ( g ) .
The density ( 2 k + 1 ) / ( 2 g − 1 ) (2k+1)/(2g-1) ( 2 k + 1 ) / ( 2 g − 1 ) retains the exact codimension m m m in the saddle equation, which is why the last formula also covers bounded positive m m m . § 3.35 proves the uniform equivalent, and Appendix F proves the stated error. On the range k ≥ ε g k\ge\varepsilon g k ≥ ε g , for each fixed ε ∈ ( 0 , 3 ) \varepsilon\in(0,3) ε ∈ ( 0 , 3 ) , the stronger error O ε ( g − 1 / 4 ) O_\varepsilon(g^{-1/4}) O ε ( g − 1/4 ) holds, again including bounded positive codimension.
For each fixed codimension, Liu and Xu obtained a complete expansion of the coefficients relative to the top one [129, Theorem 1.2], a partial answer to Mirzakhani’s question. The bounded-codimension case of Theorem 1.4 agrees with their result after the change of normalization. What is new is uniformity over all degrees, including degrees and codimensions that grow with the genus.
The trigonometric saddle equation is the stationary equation of an instanton prediction from physics; see Subsection 13.7 for the change of conventions and the distinction from the coefficient theorem.
The diagonal chain. Along the chain V G , 0 , V G − 1 , 2 , … , V 0 , 2 G V_{G,0},V_{G-1,2},\ldots,V_{0,2G} V G , 0 , V G − 1 , 2 , … , V 0 , 2 G the Euler characteristic is constant and genus is traded for pairs of cusps. Mirzakhani noted that the adjacent ratios at the two ends of this chain follow from the fixed-n n n and fixed-g g g asymptotics, but that little was known about the whole sequence [144, Section 1.4]. Theorem 12.1 in § 3.34 determines every entry of the chain up to a factor exp ( o ( G ) ) \exp(o(G)) exp ( o ( G )) , uniformly in its position and including the planar endpoint, through an explicit rate function obtained by integrating the ratio profile B \mathcal{B} B . Unlike Theorem 1.1 , it also covers bounded genus.
Strategy of the proof
The heuristic of Subsection 1.1 has two parts: the approximation (4) of Φ g ′ \Phi'_g Φ g ′ , uniform in the cusp weight, and the passage from Φ g ′ \Phi'_g Φ g ′ to one of its coefficients. Three ideas carry the proof, each removing one obstruction to uniformity in n n n : cancellation in the volume formula, the unknown constant of integration, and the passage from generating functions to coefficients. The first obstruction is present from the outset. Mirzakhani’s volumes are intersection numbers of κ 1 \kappa_1 κ 1 and ψ \psi ψ classes, and κ 1 \kappa_1 κ 1 can be expressed through pure ψ \psi ψ intersections [117, 136]. The resulting formula (24) is an alternating sum over the number of additional ψ \psi ψ insertions, and asymptotics for its individual terms do not control the cancellation.
(i) Positivity from the string flow. Moving along the string flow by a time Y Y Y replaces the alternating weights − X ( k ) ( 0 ) -X^{(k)}(0) − X ( k ) ( 0 ) by the derivatives − X ( k ) ( Y ) -X^{(k)}(Y) − X ( k ) ( Y ) ; see (25) . At Y = 0 Y=0 Y = 0 the resulting shifted amplitudes are the Weil–Petersson brackets; after a natural rescaling, at Y = y 0 Y=y_0 Y = y 0 they become pure ψ \psi ψ intersection numbers with insertions of τ 2 \tau_2 τ 2 . The Dijkgraaf–Verlinde–Verlinde equations give the normalized recursion (28) , whose kernel coefficients are the Taylor coefficients of 1 / K ‾ Y 1/\overline{K}_Y 1/ K Y , where K ‾ Y \overline{K}_Y K Y is the rescaled kernel in (12) . These coefficients are nonnegative throughout [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] , including for the limiting kernel K ‾ y 0 ( w ) = 1 − w \overline{K}_{y_0}(w)=1-w K y 0 ( w ) = 1 − w (§ 3.1 ). The alternating sum is thereby replaced by a positive recursion whose kernels form a compact family. Comparison with an explicit solvable model, the harmonic model (31) , gives two-sided bounds that are uniform in Y Y Y and in all degrees but lose a factor e O ( g ) e^{O(\sqrt{g})} e O ( g ) (Proposition 5.1). This step uses no information about volumes.
(ii) Ratios, and normalization at the pure-ψ \psi ψ endpoint. For a fixed number of distinguished markings, the dilaton equation expresses the Y Y Y -derivative of log Φ g ( p ) ( X ( Y ) ) \log\Phi_g^{(p)}(X(Y)) log Φ g ( p ) ( X ( Y )) through amplitudes with one more marking, and the KdV equation relates genus g g g to genus g − 1 g-1 g − 1 with two more markings. By positivity the amplitudes decrease as a degree increases, and this monotonicity upgrades the coarse bounds of step (i) to relative errors O ( 1 / g ) O(1/g) O ( 1/ g ) for both ratios (Proposition 6.2). Integrating the resulting differential equation in Y Y Y gives (4) up to a factor independent of Y Y Y . This factor is determined at Y = y 0 Y=y_0 Y = y 0 , where only ⟨ τ 2 3 g − 3 ⟩ g \langle\tau_2^{3g-3}\rangle_g ⟨ τ 2 3 g − 3 ⟩ g enters and Painlevé I supplies the asymptotics (§ 3.19 ). The deformation thus carries the normalization problem to a point where it is classical. Iterating the same equations gives all higher corrections, with remainders controlled as the order grows (§ 3.22 , with proofs in Appendix § 8.5 ).
(iii) Extraction through a positive factor. An expansion of a generating function does not by itself approximate its coefficients. With h = 2 g − 2 + p h=2g-2+p h = 2 g − 2 + p , we show that Φ g ( p ) \Phi_g^{(p)} Φ g ( p ) is ( 1 / X ′ ( Y ( x ) ) ) h (1/X'(Y(x)))^h ( 1/ X ′ ( Y ( x )) ) h times a power series with nonnegative coefficients (Lemma 9.1):
Φ g ( p ) = V h R g , p , V = 1 X ′ ∘ Y ′ , [ x j ] R g , p ≥ 0 , hence ∣ Φ g ( p ) ( x e i θ ) ∣ ≤ Φ g ( p ) ( x ) ∣ V ( x e i θ ) V ( x ) ∣ h . \Phi_g^{(p)} = V^h R_{g,p}, \qquad V = \frac{1}{X' \circ Y'}, \qquad[x^j]R_{g,p} \ge0, \qquad\text{hence}\qquad\left|\Phi_g^{(p)}(xe^{i\theta})\right| \le\Phi_g^{(p)}(x)\left|\frac{V(xe^{i\theta})}{V(x)}\right|^h . Φ g ( p ) = V h R g , p , V = X ′ ∘ Y ′ 1 , [ x j ] R g , p ≥ 0 , hence Φ g ( p ) ( x e i θ ) ≤ Φ g ( p ) ( x ) V ( x ) V ( x e i θ ) h .
The decay on the circle ∣ z ∣ = x |z|=x ∣ z ∣ = x is thus controlled by one explicit function raised to the power h h h , whatever the unknown factor R g , p R_{g,p} R g , p is. This gives Fourier bounds uniform in the cusp weight, and hence a Gaussian local limit theorem when both h h h and the model mean h Ψ ( x ) h\Psi(x) h Ψ ( x ) tend to infinity, even as x x x approaches 0 0 0 or x c x_c x c (Proposition 9.2). Bounded cusp counts instead require a Poisson approximation near x = 0 x=0 x = 0 : a change of variables of the type studied by Temme [178] gives an expansion around the Poisson law that remains regular in this range (§ 3.30 , with proofs in Appendix § 8.16 ). Together these estimates yield Theorem 1.1 (§ 3.31 ).
The same positive factor reappears in the applications: it gives the uniform coefficient-ratio bounds that control curves cutting off many cusps (Appendix § 10.1 ) and the boundary factorization of Proposition 9.4.
From volumes to random surfaces. The applications use the volume theorem in one way: through Mirzakhani’s integration formula, ratios of volumes are expected numbers of geometric configurations, so Theorem 1.3 determines the statistics of short separating curves at every cusp density (§ 4.1 ). The spectral applications then require a second passage, from these counts to actual eigenvalues. A short curve produces a trial function, hence an eigenvalue upper bound, but neither a short-curve law nor an upper bound alone determines the low spectrum; one must show that nothing else is there. Three arguments do this in the three regimes. At the critical scale, funnel trial states are matched to the expected spectral count supplied by Huang’s positive trace [101]; on shrinking scales, a comparison on the whole low spectral subspace excludes the modes missed by trial functions; in fixed windows, concentration of orbital statistics and trace approximation identify the limiting measure. The geometric application retains based holonomies in the same integration formula, so that the carrier integrals determine whole pointed balls rather than curve counts.
The spectral transition at the square-root scale
Let n ≥ 1 n \ge1 n ≥ 1 , let X X X have normalized Weil–Petersson law P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP , and let Δ X \Delta_X Δ X be its nonnegative Laplace–Beltrami operator. The spectrum below 1 / 4 1/4 1/4 is discrete; we count its eigenvalues λ j ( X ) \lambda_j(X) λ j ( X ) with multiplicity and write G ( X ) = inf spec ( Δ X ∣ 1 ⊥ ) G(X)=\inf\operatorname{spec}(\Delta_X|_{1^\perp}) G ( X ) = inf spec ( Δ X ∣ 1 ⊥ ) , with G ( X ) = 1 / 4 G(X)=1/4 G ( X ) = 1/4 if there is no positive eigenvalue below 1 / 4 1/4 1/4 . For closed surfaces (n = 0 n=0 n = 0 ), Mirzakhani proved a positive gap with high probability [144, Section 4.5]; the bound was raised to 3 / 16 − ε 3/16-\varepsilon 3/16 − ε by Wu and Xue and by Lipnowski and Wright [189, 128], and to the optimal 1 / 4 − ε 1/4-\varepsilon 1/4 − ε by Anantharaman and Monk [11, Theorem 1.6]. With cusps, Hide, Shen and Wu, and He, Wu and Xue obtained positive lower bounds below the square-root scale [95, Theorem 1.3], [172, Theorem 3], [92, Theorem 1.2], and Huang proved G ( X ) → 1 / 4 G(X)\to1/4 G ( X ) → 1/4 throughout this range [101, Theorem 1.1]. The literature is compared in detail at the beginning of Part 2.
Why g \sqrt{g} g . A short geodesic cutting off two cusps bounds a pair of pants of area 2 π 2\pi 2 π , attached to the rest of the surface through a collar of small conductance, and therefore produces a small eigenvalue. By Mirzakhani’s integration formula, the expected number of geodesics of length at most L L L that cut off a given pair of cusps is ( L 2 / 2 ) V g , n − 1 / V g , n (L^2/2)V_{g,n-1}/V_{g,n} ( L 2 /2 ) V g , n − 1 / V g , n up to a factor 1 + O ( L ) 1+O(L) 1 + O ( L ) . When n = o ( g ) n=o(g) n = o ( g ) , Theorem 1.3 makes this L 2 / ( 16 π 2 g ) L^2/(16\pi^2g) L 2 / ( 16 π 2 g ) , so the expected total over the n 2 / 2 n^2/2 n 2 /2 pairs is about n 2 L 2 / ( 32 π 2 g ) n^2L^2/(32\pi^2g) n 2 L 2 / ( 32 π 2 g ) . Hence such curves of bounded length are absent when n ≪ g n\ll\sqrt{g} n ≪ g , occur in bounded number when n ≍ g n\asymp\sqrt{g} n ≍ g , and have lengths of order g / n → 0 \sqrt{g}/n\to0 g / n → 0 when n ≫ g n\gg\sqrt{g} n ≫ g ; these are the three spectral phases. At positive density the same count requires the full profile C ( n / g ) C(n/g) C ( n / g ) and all planar cusp types, and this is where the uniform theory is indispensable. Table 1 summarizes the resulting picture.
At the critical scale the model spectral object is the complete surface obtained by attaching a funnel to the pants. Let β ( s ) ∈ ( 0 , 1 / 4 ) \beta(s)\in(0,1/4) β ( s ) ∈ ( 0 , 1/4 ) be the unique eigenvalue below 1 / 4 1/4 1/4 of the complete hyperbolic surface with two cusps and one funnel whose convex core has boundary length s s s .
Here Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] is the density parameter, with Y ( λ ) = λ Y(\lambda)=\lambda Y ( λ ) = λ ; Ξ X \Xi_X Ξ X is the counting measure of the spectrum in ( 0 , 1 / 4 ) (0,1/4) ( 0 , 1/4 ) ; G c ∈ ( 0 , 1 / 4 ) G_c\in(0,1/4) G c ∈ ( 0 , 1/4 ) is the random gap limit; κ 0 = 1 / ( 8 π 2 ) \kappa_0=1/(8\pi^2) κ 0 = 1/ ( 8 π 2 ) ; and A g A_g A g is the factor n / ( 2 g ) \sqrt{n}/(2\sqrt{g}) n / ( 2 g ) when g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g . The last column gives the numbers of the theorems and corollaries; the gap limit for n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ is Corollary 14.1, and the empty fixed windows for n = o ( g ) n=o(\sqrt{g}) n = o ( g ) hold with probability tending to one.
For n log g / g → ∞ n\log g/\sqrt{g}\to\infty n log g / g → ∞ . a Proved under n 9 / 4 log 9 / 4 ( n + 2 ) / g 11 / 4 → 0 n^{9/4}\log^{9/4}(n+2)/g^{11/4}\to0 n 9/4 log 9/4 ( n + 2 ) / g 11/4 → 0 , which holds when n ≤ g 11 / 9 − δ n\le g^{11/9-\delta} n ≤ g 11/9 − δ for a fixed δ > 0 \delta>0 δ > 0 . Between this range and the collapse range the microscopic law is open.
Theorem 15.1 and Corollary 15.9 (The critical spectrum). Suppose g → ∞ g \to\infty g → ∞ and n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g} \to c \in(0,\infty) n / g → c ∈ ( 0 , ∞ ) , and let Q c \mathcal{Q}_c Q c be a Poisson process on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) with intensity c 2 sinh ( s / 2 ) d s / ( 8 π 2 ) c^2 \sinh(s/2)\,ds/(8\pi^2) c 2 sinh ( s /2 ) d s / ( 8 π 2 ) . Then
∑ 0 < λ j ( X ) < 1 / 4 δ λ j ( X ) ⟹ ∑ s ∈ Q c δ β ( s ) \sum_{0<\lambda_j(X)<1/4}\delta_{\lambda_j(X)} \Longrightarrow\sum_{s\in\mathcal{Q}_c}\delta_{\beta(s)} 0 < λ j ( X ) < 1/4 ∑ δ λ j ( X ) ⟹ s ∈ Q c ∑ δ β ( s ) in distribution, for the vague topology on [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) , and the mean measures converge. The cuspidal part of the left side tends to zero, so the limiting spectrum is residual. In particular G ( X ) G(X) G ( X ) converges in distribution to the smallest point G c ∈ ( 0 , 1 / 4 ) G_c \in(0,1/4) G c ∈ ( 0 , 1/4 ) of the limit. Consequently, for every sequence n = n ( g ) ≥ 1 n=n(g)\geq1 n = n ( g ) ≥ 1 ,
G ( X ) → P 1 4 ⟺ n g → 0 , G ( X ) → P 0 ⟺ n g → ∞ , G(X) \xrightarrow{\mathbb{P}} \frac{1}{4} \Longleftrightarrow\frac{n}{\sqrt{g}} \to0,\qquad G(X) \xrightarrow{\mathbb{P}} 0 \Longleftrightarrow\frac{n}{\sqrt{g}} \to\infty, G ( X ) P 4 1 ⟺ g n → 0 , G ( X ) P 0 ⟺ g n → ∞ , and G ( X ) G(X) G ( X ) has no deterministic limit in ( 0 , 1 / 4 ) (0,1/4) ( 0 , 1/4 ) along any sequence.
The limit is a random point configuration rather than a deterministic value for each eigenvalue: one model eigenvalue for each short two-cusp pair of pants, and nothing else. The volume input determines the pants process; the substantive spectral step is to show that these localized states exhaust the spectrum below 1 / 4 1/4 1/4 . This is done by matching the count of funnel trial states with the expected spectral count supplied by Huang’s positive trace [101]. Of the two equivalences, the first is Huang’s theorem [101], Theorem 1.1 and Corollary 6.15; the second uses the adjacent-volume limit of Theorem 1.3 and completes the negative answer to Shen–Wu’s Question 2 [172], which asks whether a uniform spectral gap persists when n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ .
Away from the critical scale the lowest eigenvalues obey different laws. Slightly below criticality, if c = n / g → 0 c=n/\sqrt{g}\to0 c = n / g → 0 and c log g → ∞ c\log g\to\infty c log g → ∞ , the defects ( 1 / 4 − λ j ( X ) ) / c 2 (1/4-\lambda_j(X))/c^2 ( 1/4 − λ j ( X )) / c 2 converge to a Poisson process with intensity d x / ( 8 π 2 x 2 ) dx/(8\pi^2x^2) d x / ( 8 π 2 x 2 ) ; in particular ( 1 / 4 − G ( X ) ) / c 2 (1/4-G(X))/c^2 ( 1/4 − G ( X )) / c 2 has the Fréchet law exp ( − 1 / ( 8 π 2 x ) ) \exp(-1/(8\pi^2x)) exp ( − 1/ ( 8 π 2 x )) (Theorem 16.1 ). Above criticality the collar waists shrink, and the eigenvalues are governed by the conductances of short collars weighted by the areas they enclose.
Theorems 17.1, 18.1 and 18.3 (Supercritical spectra). Suppose g → ∞ g\to\infty g → ∞ and n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ , and let A g A_g A g be the explicit scale (17.1). If either n / g n/g n / g stays bounded, or n / g → ∞ n/g\to\infty n / g → ∞ and n 9 / 4 log 10 ( n + 2 ) / g 11 / 4 → 0 n^{9/4}\log^{10}(n+2)/g^{11/4}\to0 n 9/4 log 10 ( n + 2 ) / g 11/4 → 0 , then
∑ 0 < λ j ( X ) < 1 / 4 δ A g λ j ( X ) ⟹ PPP ( 2 t d t ) on [ 0 , ∞ ) . \sum_{0<\lambda_j(X)<1/4}\delta_{A_g\lambda_j(X)} \Longrightarrow\operatorname{PPP}(2t\,dt)\quad\text{on }[0,\infty). 0 < λ j ( X ) < 1/4 ∑ δ A g λ j ( X ) ⟹ PPP ( 2 t d t ) on [ 0 , ∞ ) . Thus P { A g G ( X ) > t } → e − t 2 \mathbb{P}\{A_gG(X)>t\}\to e^{-t^2} P { A g G ( X ) > t } → e − t 2 , and G ( X ) / h ( X ) → 1 / π G(X)/h(X)\to1/\pi G ( X ) / h ( X ) → 1/ π in probability, where h h h is the Cheeger constant. When g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g , one has A g ∼ π n / ( 2 2 g ) A_g\sim\pi n/(2\sqrt{2}g) A g ∼ π n / ( 2 2 g ) . If instead n ≫ g 3 / ( log g ) 4 n\gg g^3/(\log g)^4 n ≫ g 3 / ( log g ) 4 , then A g λ r + ( X ) → 0 A_g\lambda_r^+(X)\to0 A g λ r + ( X ) → 0 in probability for every fixed r r r , where λ r + \lambda_r^+ λ r + is the r r r th positive eigenvalue.
In (17.1), A g = 2 π 2 ( q g M 2 ( x g ) ) 1 / 2 A_g=2\pi^2(q_gM_2(x_g))^{1/2} A g = 2 π 2 ( q g M 2 ( x g ) ) 1/2 , where q g w k ( x g ) q_gw_k(x_g) q g w k ( x g ) is the intensity of short curves cutting off k k k cusps and M 2 M_2 M 2 weights type k k k by ( k − 1 ) 2 (k-1)^2 ( k − 1 ) 2 , which up to a constant is the square of the area 2 π ( k − 1 ) 2\pi(k-1) 2 π ( k − 1 ) of the cap cut off by such a curve. The proof is a two-sided comparison on the whole low spectral subspace: trial functions give upper bounds, and the comparison excludes the low modes they miss, including modes supported on a large component separated by several curves. The collapse at large density has a geometric cause: large planar caps then support low-energy functions of their own, and the collar picture fails. The correct microscopic normalization between g 11 / 9 g^{11/9} g 11/9 and g 3 / ( log g ) 4 g^3/(\log g)^4 g 3 / ( log g ) 4 remains open.
Fixed energy windows behave differently. Let Ξ X \Xi_X Ξ X be the counting measure of the spectrum in ( 0 , 1 / 4 ) (0,1/4) ( 0 , 1/4 ) . Fixed windows contain many eigenvalues, and the normalized counting measures converge to deterministic limits: ( g / n 2 ) Ξ X → ν ∗ (g/n^2)\Xi_X\to\nu_* ( g / n 2 ) Ξ X → ν ∗ when g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g , g − 1 Ξ X → ν λ g^{-1}\Xi_X\to\nu_\lambda g − 1 Ξ X → ν λ when n / g → λ ∈ ( 0 , ∞ ) n/g\to\lambda\in(0,\infty) n / g → λ ∈ ( 0 , ∞ ) , and n − 1 Ξ X → ν ∞ n^{-1}\Xi_X\to\nu_\infty n − 1 Ξ X → ν ∞ when g = o ( n ) g=o(n) g = o ( n ) , including every fixed genus (Theorems 19.1 –19.3 ). Theorem 19.4 includes the continuous spectrum through the classical regularized trace, and Corollary 19.5 derives the limit of the log determinant per unit area. These results rest on concentration of orbital statistics rather than on individual short curves, and the measures ν λ \nu_\lambda ν λ are characterized by planar orbital integrals.
Local geometry at positive cusp density
For every λ ∈ ( 0 , ∞ ) \lambda\in(0,\infty) λ ∈ ( 0 , ∞ ) , Theorem 20.1 constructs a unimodular random surface ( S λ , o λ ) (S_\lambda,o_\lambda) ( S λ , o λ ) as the fixed-radius local limit when n / g → λ n/g \to\lambda n / g → λ , rooted by area, and proves quenched concentration of local statistics. Its law tends to the hyperbolic plane as λ ↓ 0 \lambda\downarrow0 λ ↓ 0 and to the area-rooted infinite WP sphere of Budd and Curien [31] as λ ↑ ∞ \lambda\uparrow\infty λ ↑ ∞ . After filling the cusps, these surfaces are conformally discs at finite density and the plane at the endpoint; Brownian motion on the original hyperbolic surfaces is respectively transient and recurrent (Theorem 20.2 ). By the theorem of Hopf, Tsuji and Sullivan, the geodesic flow is accordingly not ergodic at finite density and ergodic at the endpoint (Corollary 20.3); the endpoint statements confirm the conjecture of Budd and Curien [31, Section 7.3]. A separate growing-radius tree coupling gives the endpoint’s Brownian-plane scaling limit: local pointed Gromov–Hausdorff–Prokhorov convergence after cusp trimming, with the full area projected to the trimmed space, and local pointed Gromov–Prokhorov convergence for the untrimmed surface with its full area (Theorem 21.1 ).
Questions from the literature
Table 2 lists the named questions treated here and the extent of each answer; other earlier work is compared at the beginning of each part.
The near-critical scaling question of Budd–Curien and the dense microscopic spectrum between g 11 / 9 g^{11/9} g 11/9 and g 3 / ( log g ) 4 g^3/(\log g)^4 g 3 / ( log g ) 4 remain open; see § 6.1 .
Organization of the paper
Part 1 follows the three steps of Subsection 1.5: positive recursion, endpoint normalization and coefficient extraction, followed by the volume-ratio and adjacent-coefficient consequences. Part 2 passes from short geometry to the critical spectrum, its moving edge and the supercritical microscopic laws, and then to fixed-energy windows; its opening guide compares the observables, density ranges and earlier work. Part 3 treats local limits, conformal type and the Brownian plane.
Main-text sections contain the statements and principal arguments; long proofs are in the appendices, grouped by part and ordered by first use, and Appendix A gives a proof-location table and a glossary. A reader interested first in volume asymptotics may read Part 1 with Appendices B and C; the other appendices to Part 1 give the inverse-genus rates, the higher-order expansions and the all-marking envelope (D.35) used in Part 2. A reader interested in spectra may start with Part 2, taking Theorem 1.3 , Proposition 9.4 and the positive convolution factor of § 3.25.1 as inputs, which its opening page lists precisely.
We compare the uniform theorem with the existing volume asymptotics. The proof begins in § 3.1 .
Moduli spaces, volumes and curve counting. Weil–Peterson volumes play two roles in this paper: their boundary-length polynomials encode intersection numbers, and ratios of these volumes enter expectations of curve counts on random hyperbolic surfaces. At fixed genus, the large-cusp prediction of Itzykson is recorded in [194, Remark 3]. Zograf’s genus-zero recursion [193] was followed by the proof of the genus-zero large-cusp asymptotics by Kaufmann–Manin–Zagier [117], and Manin–Zograf [136, Theorem 6.1] treated every fixed genus. Wolpert computed the volume of M 1 , 1 \mathcal{M}_{1,1} M 1 , 1 [183], and Penner developed a combinatorial method for computing these volumes [159]. Building on McShane’s identity [138], Mirzakhani derived a recursion that computes V g , n ( L ) V_{g,n}(\mathbf{L}) V g , n ( L ) as a polynomial in the squared boundary lengths [141]; its coefficients are intersection numbers of tautological classes on M g , n \mathcal{M}_{g,n} M g , n [142], with the symplectic normalization established by Wolpert [182]. Mulase and Safnuk recast the recursion as Virasoro constraints and showed that the generating function of mixed κ 1 \kappa_1 κ 1 –ψ \psi ψ intersections is a one-parameter solution of the Korteweg–de Vries (KdV) hierarchy [152]; Liu and Xu showed directly that Mirzakhani’s recursion is equivalent to the Witten–Kontsevich theorem [132], and Liu and Xu [133] and Eynard [64] extended the recursion to volumes weighted by higher Mumford classes; Eynard and Orantin identified it with the loop equations of a matrix model [67]. The identities expressing mixed κ 1 \kappa_1 κ 1 –ψ \psi ψ intersections through pure ψ \psi ψ intersections, which are the starting point of § 3.4 , are due to Arbarello and Cornalba [14], Kaufmann, Manin and Zagier [117] and Manin and Zograf [136]. Related exact formulas include the tight-volume recursions of Budd and Zonneveld [34], the KdV boundary-addition identities and fixed-genus Bessel singularity studied by Johnson [110], and the Virasoro-theoretic proof by Xiong and Zhou of the intersection-number and cohomological-field-theory descriptions of topological recursion, including variation of the spectral-curve function [191, Theorem 1.1 and Section 5.7]. These are exact identities; obtaining uniform remainder estimates from them is a separate analytic problem.
Mirzakhani’s integration formula [141] expresses the average number of simple multicurves of a prescribed type in terms of volumes of the surfaces obtained by cutting along them. After division by V g , n V_{g,n} V g , n , these formulas become expectations under Weil–Peterson probability, so relative volume estimates control geometric probabilities. Mirzakhani used this connection to study random surfaces at large genus [144], and Mirzakhani and Petri used joint curve counts to obtain Poisson length statistics for closed surfaces [146]. A different curve-counting limit fixes the surface and lets the length bound tend to infinity. Mirzakhani’s counting theorem assigns frequencies to simple-curve topological types in that limit [143]. Delecroix, Goujard, Zograf and Zorich related these frequencies to intersection numbers and Masur–Veech volumes [52] and studied random multicurves at large genus [53]; Ren obtained the large-genus ratio of separating to nonseparating frequencies with many cusps [163]. Those frequencies are topological quantities, whereas the processes studied here count short curves on a random surface as its topology varies.
Masur–Veech volumes are the flat counterpart of Weil–Peterson volumes, and their large-genus theory has developed in parallel. Eskin and Zorich conjectured the large-genus asymptotics of Masur–Veech volumes of strata of Abelian differentials [63]; Aggarwal proved them for all strata [4], and Chen, Möller, Sauvaget and Zagier expressed these volumes through intersection numbers on the strata and evaluated their large-genus limits [43]; Chen, Möller and Sauvaget expressed the volumes of the principal strata of quadratic differentials through intersection numbers [44]. Rafi and Zorich recently determined the large-n n n asymptotics of Masur–Veech volumes of strata of quadratic differentials with many simple poles at fixed genus [162].
Asymptotics of volumes and intersection numbers. The fixed-genus regime was recalled at the beginning of the introduction. Here we discuss large-genus estimates at fixed cusp count, coefficients of growing degree, pure ψ \psi ψ intersection numbers and predictions from physics.
Fixed cusp counts. Grushevsky [80] and, for closed surfaces, Schumacher and Trapani [171] showed that for fixed n n n the volumes grow like ( 2 g ) ! (2g)! ( 2 g )! up to factors c g c^{g} c g ; see the account in [144, Section 1.4]. After Mirzakhani’s estimates [144], Mirzakhani and Zograf obtained the expansion to every fixed order [147, Theorem 1.2]. They showed that its coefficients are polynomials in n n n with coefficients in Q [ π − 2 , π 2 ] \mathbb{Q}[\pi^{-2},\pi^{2}] Q [ π − 2 , π 2 ] and conjectured, on numerical evidence, that only Q [ π − 2 ] \mathbb{Q}[\pi^{-2}] Q [ π − 2 ] occurs [147, Theorem 1.2 and Remark 1.3]; to our knowledge this conjecture remains open. Their leading term also holds when n 2 / g → 0 n^{2}/g\to0 n 2 / g → 0 [147, Theorem 1.8], and their cusp-ratio estimate has error O ( n / g ) O(n/g) O ( n / g ) for n ≥ 1 n\ge1 n ≥ 1 [147, Lemma 5.1]. Their first unmarked correction enters (11.5). He, Wu and Xue obtained first-order volume-ratio expansions with error O ( ( 1 + n 2 ) / g 2 ) O((1+n^{2})/g^{2}) O (( 1 + n 2 ) / g 2 ) for n = o ( g ) n=o(\sqrt{g}) n = o ( g ) [92, Theorem 4.11]. Anantharaman and Monk expanded the volume polynomials at large genus [9]. Guillermit sharpened the coefficient bounds and treated boundary lengths of order g \sqrt{g} g for fixed n n n [81, Theorems 2, 4, 5 and Remark 3]. For fixed n n n , her Theorem 2 gives coefficient bounds of order ( C n j ) j (C_nj)^j ( C n j ) j , and her Theorem 4 bounds the remainders of the normalized volumes with boundary themselves, with the same single-factorial dependence on the order. Our estimates retain this order dependence uniformly in the density parameter Y Y Y , including its critical endpoint, and after extraction uniformly over all cusp counts. Hide, Macera and Thomas gave bounds explicit in the expansion order, cusp count and degrees [96, Theorem 4.2, statements A1(k)–A4(k)]. Their genus must exceed an order-dependent multiple of the cusp count, and the errors grow polynomially in the cusp count or degrees, so these bounds do not yield the relative remainder of Theorem 1.1 at positive cusp density.
Sublinear cusp counts. Gatignol and Thomas [71, Theorem 1.5] obtain first-order expansions for normalized intersection brackets and for the cusp and genus volume ratios throughout n = o ( g ) n=o(g) n = o ( g ) , with ratio errors O ( n 2 / g 2 + 1 / g ) O(n^{2}/g^{2}+1/g) O ( n 2 / g 2 + 1/ g ) . In particular, their genus-ratio expansion supplies the two-sided limit needed at n ≍ g n\asymp\sqrt{g} n ≍ g . These estimates overlap with the zero-density consequences of Theorem 1.3; the present theorem also treats positive and unbounded n / g n/g n / g , and the absolute expansion has remainders uniform over all n n n .
Growing degrees. Wu and Xue established effective upper bounds for normalized intersection numbers when some degrees are large compared with 2 g − 2 + n \sqrt{2g-2+n} 2 g − 2 + n , together with bounds for volumes with large boundary lengths [188]. These bounds control individual coefficients and volumes, whereas the adjacent-ratio asymptotics of Theorem 1.4 require relative control across neighboring degrees. Moreover, a bracket may itself be small, and then an additive error polynomial in the degree gives no relative control. Liu and Xu recently proved general balancing inequalities for mixed tautological intersection numbers [130]; the monotonicity statements used here (Lemma 6.1 and the bracket monotonicity in § 3.35 ) are of a simpler kind. We use the Mirzakhani–Zograf bracket-difference estimates near the lowest degrees; the full degree range requires the harmonic comparison and the log-concavity argument of § 3.35 .
Pure ψ \psi ψ intersection numbers. Motivated by Masur–Veech volumes of quadratic differentials, Delecroix, Goujard, Zograf and Zorich conjectured uniform large-genus asymptotics for pure ψ \psi ψ intersections (Appendix § 8.16 of the arXiv version of [52]). Aggarwal proved them for o ( g ) o(\sqrt{g}) o ( g ) insertions [5], and Guo and Yang gave a different proof for at most C log g C\log g C log g insertions, for every fixed C C C , based on an explicit formula for the n n n -point generating series [82, Theorem 1]. For a fixed number of insertions, Eynard, Garcia-Failde, Giacchetto, Gregori and Lewański obtained resurgent expansions to every fixed order, uniformly in the degrees [65, Theorem 3.13].
Our comparison needs only fixed positive upper and lower bounds for the normalized pure-ψ \psi ψ intersections. Appendix § 8.4 proves these bounds by a scalar induction, adapting the recursion estimates of Guo, Yang and Zagier [83] and using the two-point bound of Delecroix, Goujard, Zograf and Zorich [52], as quoted in [51, equation (6)]. The absolute constant is determined by the single sequence ⟨ τ 2 3 g − 3 ⟩ g \langle\tau_{2}^{3g-3}\rangle_{g} ⟨ τ 2 3 g − 3 ⟩ g through the Itzykson–Zuber identity and Painlevé I. To match the expansions at the pure-ψ \psi ψ endpoint when the truncation order grows with g g g , we use the exact identity [106, equation (6.10)], the zero-instanton contour formula of Garoufalidis, Its, Kapaev and Mariño [70], and the Borel representation of Costin, Costin and Huang [48]. Alameddine, Marchal, Nikolaev and Orantin proved the resurgence of the tritronquée solutions of the deformed Painlevé I equation [6] and recently solved the direct monodromy problem for the Borel-summed zero-parameter solution, identifying it with the tritronquée solutions [7]; neither result is used in our proof, which needs only the large-order asymptotics of the formal solution. Borot, Eynard and Giacchetto use comparison with Painlevé I amplitudes to bound topological-recursion correlators, uniformly in genus and number of markings [23, Main Theorem and Section 1.3]. Their bounds concern the growth of the correlators themselves, whereas the factorial bounds in Theorem 1.1 concern the order of the inverse-genus expansion after division by an explicit leading term.
Instanton predictions. The factorial growth of genus expansions in two-dimensional gravity and its nonperturbative origin go back to Shenker [173]; for minimal strings approaching Jackiw–Teitelboim gravity, the resurgent structure was analyzed by Gregori and Schiappa [73]. Eynard, Garcia-Failde, Gregori, Lewański and Schiappa derived large-genus asymptotics for ordinary Weil–Petersson volumes under a resurgence assumption [66]. Griguolo, Papalíni, Russo and Seminara predicted the coefficients of ordinary Weil–Petersson volumes from instanton contributions [77, Section 3.3]. For super Weil–Petersson volumes, X. Huang proved their two conjectural formulas up to a universal multiplicative constant, for fixed n n n with L i = o ( g ) L_i=o(\sqrt{g}) L i = o ( g ) and for descendant degrees ∣ d ∣ = o ( g ) |\mathbf{d}|=o(g) ∣ d ∣ = o ( g ) [102, Theorem 1.7]; Huang, Liu and Xu subsequently determined that constant [103]. Johnson and Rodrigues wrote down the corresponding one-boundary contour integral [111, equation (4.22)]. Hatsuda, Matsumoto and Okuyama constructed transseries solutions of the KdV equation that governs the genus expansion of the JT partition function, whose leading nonperturbative sector reproduces the instanton terms [88]. After a change of conventions, expansion of the Johnson–Rodrigues integral gives our harmonic coefficient model at Y = 0 Y=0 Y = 0 , and its stationary equation is the saddle equation of Theorem 1.4 (Subsection 13.7). Unlike the leading factor sinh ( L / 2 ) / ( L / 2 ) \sinh(L/2)/(L/2) sinh ( L /2 ) / ( L /2 ) at bounded boundary length, this integral retains the dependence of the coefficients on their degree up to the top degree. Theorem 1.4 establishes the resulting asymptotic formula for the true adjacent coefficient ratios, uniformly over all degrees, from exact intersection identities and estimates; no resurgence assumption on the Weil–Petersson genus expansion is used.
Zograf’s constant and the work of Liu and Xu. Our determination of Zograf’s constant C = 1 / π C=1/\sqrt{\pi} C = 1/ π appeared in the first version of this paper [127, Corollary 1.2 and Section 7]. Liu and Xu subsequently gave another proof using a different deformation [131]. Our first version also contained the uniform volume-ratio limits, the diagonal chain, the adjacent-coefficient asymptotics over all degrees, the absolute expansion uniform in the cusp count (then with Gevrey-2 remainder bounds), the vanishing spectral gap with many small eigenvalues, and the short-curve processes; the present version sharpens the remainders to single-factorial order and adds the remaining spectral theory of Part 2 and the geometric theory of Part 3.
The deformation of Liu and Xu joins the cubic curve y = z − 4 z 3 y=z-4z^3 y = z − 4 z 3 to the sine curve y = sin ( 2 π z ) / ( 2 π ) y=\sin(2\pi z)/(2\pi) y = sin ( 2 π z ) / ( 2 π ) while keeping the first positive zero fixed. For each fixed number n ≥ 2 n\geq2 n ≥ 2 of markings, their uniform-variation theorem [131, Theorem 1.1] controls the logarithmic derivative of the normalized volumes uniformly in the curve parameter t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] by O n ( g − 1 ) O_n(g^{-1}) O n ( g − 1 ) . Transport from the cubic endpoint determines the constant [131, Corollary 1.2], and their first-correction theorem [131, Theorem 1.3] adds the 1 / g 1/g 1/ g term with an O n ( g − 2 ) O_n(g^{-2}) O n ( g − 2 ) remainder, uniformly in t t t . Huang, Liu and Xu subsequently applied the same transport strategy to super Weil–Petersson volumes, where the simpler endpoint is governed by the Brézin–Gross–Witten numbers [103].
Our deformation is instead the string flow associated with the cusp weight x = X ( Y ) x=X(Y) x = X ( Y ) . Its amplitudes with a fixed number p p p of distinguished markings encode the derivatives Φ q ( p ) ( x ) \Phi_q^{(p)}(x) Φ q ( p ) ( x ) of the cusp generating function. Both arguments use positivity, descendant recursions, the dilaton and KdV equations, and a Painlevé I normalization, but along different deformation families. In the present paper, § 3.19 already determines the constant by specializing to p = 1 p=1 p = 1 and Y = 0 Y=0 Y = 0 , hence to V g , 1 V_{g,1} V g , 1 ; uniformity in the total cusp count is not needed for this identification. That uniformity is needed for the stronger conclusion, an expansion valid for every n ≥ 0 n\ge0 n ≥ 0 , including n n n growing with g g g . This requires the coefficient extraction in § 3.25 –§ 3.30 and Appendix § 8.16 of the present paper, beyond the estimates for a fixed number of distinguished markings.
Positive kernels
This section supplies the analytic input for the first of the three ideas of Subsection 1.5, positivity from the string flow. Two entire functions of a variable z z z , depending on the density parameter Y Y Y of Subsection 1.1, control everything that follows. The first, K Y K_Y K Y , arises in § 3.4 when the marking singled out by the Dijkgraaf–Verlinde–Verlinde equation is merged with the κ 1 \kappa_1 κ 1 insertions: these mergers act on the amplitudes as K Y ( E ) K_Y(E) K Y ( E ) , where E E E raises a degree by one, so the Taylor coefficients of 1 / K Y 1/K_Y 1/ K Y become the coefficients of the recursion. The second, H Y H_Y H Y , plays the same role for the dilaton equation and for the explicit comparison model. We prove three properties: all zeros of K Y K_Y K Y are positive and simple (Lemma 2.1 ); 1 / K Y 1/K_Y 1/ K Y and 1 / H Y 1/H_Y 1/ H Y have nonnegative Taylor coefficients, even as functions of Y Y Y (Lemma 2.2 ); and after rescaling by the first zero of K Y K_Y K Y both kernels form a compact family, including the endpoint Y = y 0 Y=y_0 Y = y 0 , which we call the critical endpoint or, since the amplitudes there reduce to pure ψ \psi ψ intersection numbers, the pure-ψ \psi ψ endpoint (Subsection 2.2). The first property makes the recursion of § 3.4 positive, the second supplies the convolution factor of § 3.25.1 , and the third makes all later constants uniform in Y Y Y . The first two are special-function statements, proved in Appendix B; the third is proved here, together with the local limit theorem ((16) ) for the comparison model.
Recall from Subsection 1.1 that
X ( y ) = ∑ j ≥ 0 ( − 1 ) j y j + 1 j ! ( j + 1 ) ! = y J 1 ( 2 y ) , X(y)=\sum_{j\ge0}\frac{(-1)^j y^{j+1}}{j!(j+1)!}=\sqrt{y}J_1(2\sqrt{y}), X ( y ) = j ≥ 0 ∑ j ! ( j + 1 )! ( − 1 ) j y j + 1 = y J 1 ( 2 y ) ,
and put a ( y ) = X ′ ( y ) = J 0 ( 2 y ) a(y)=X'(y)=J_0(2\sqrt{y}) a ( y ) = X ′ ( y ) = J 0 ( 2 y ) ; recall also that y 0 = j 1 2 / 4 y_0=j_{1}^{2}/4 y 0 = j 1 2 /4 is the first positive zero of a a a , and that x c = X ( y 0 ) x_c=X(y_0) x c = X ( y 0 ) . Thus a > 0 a>0 a > 0 on [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) , and the inverse branch Y ( x ) Y(x) Y ( x ) , with Y ( 0 ) = 0 Y(0)=0 Y ( 0 ) = 0 , maps [ 0 , x c ) [0,x_c) [ 0 , x c ) onto [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) .
We shall use
K Y ( z ) = ∫ 0 1 X ′ ( Y + z 2 ( 1 − s 2 ) ) d s = ∑ ℓ ≥ 0 X ( ℓ + 1 ) ( Y ) ( 2 ℓ + 1 ) ! ! z ℓ . (8) K_Y(z)=\int_0^1 X'\left(Y+\frac{z}{2}(1-s^2)\right)\,\mathrm{d}s
=\sum_{\ell\ge0}\frac{X^{(\ell+1)}(Y)}{(2\ell+1)!!}z^\ell. \tag*{(8)} K Y ( z ) = ∫ 0 1 X ′ ( Y + 2 z ( 1 − s 2 ) ) d s = ℓ ≥ 0 ∑ ( 2 ℓ + 1 )!! X ( ℓ + 1 ) ( Y ) z ℓ . (8)
H Y ( z ) = 3 ∫ 0 1 s 2 X ′ ( Y + z 2 ( 1 − s 2 ) ) d s = ∑ ℓ ≥ 0 3 X ( ℓ + 1 ) ( Y ) ( 2 ℓ + 3 ) ! ! z ℓ . H_Y(z)=3\int_0^1s^2X'\left(Y+\frac{z}{2}(1-s^2)\right)\,\mathrm{d}s
=\sum_{\ell\ge0}\frac{3X^{(\ell+1)}(Y)}{(2\ell+3)!!}z^\ell. H Y ( z ) = 3 ∫ 0 1 s 2 X ′ ( Y + 2 z ( 1 − s 2 ) ) d s = ℓ ≥ 0 ∑ ( 2 ℓ + 3 )!! 3 X ( ℓ + 1 ) ( Y ) z ℓ .
Beta integration gives the series and hence
K Y = H Y + 2 3 z ∂ z H Y . (9) K_Y=H_Y+\frac{2}{3}z\partial_zH_Y . \tag*{(9)} K Y = H Y + 3 2 z ∂ z H Y . (9)
We write r ( Y ) r(Y) r ( Y ) for the first positive zero of K Y K_Y K Y ; it equals 2 Δ ( Y ) 2\Delta(Y) 2Δ ( Y ) .
Real zeros and coefficient positivity
Lemma 2.1 (Real zeros). On [ 0 , π 2 / 4 ] [0,\pi^2/4] [ 0 , π 2 /4 ] , X ( y ) > 0 X(y)>0 X ( y ) > 0 for y > 0 y>0 y > 0 , X ′ ′ < 0 X''<0 X ′′ < 0 , and X ′ ′ ′ > 0 X'''>0 X ′′′ > 0 . Moreover y 0 < 3 / 2 y_0<3/2 y 0 < 3/2 ; in particular X ( Y ) ≤ Y < 3 / 2 X(Y)\le Y<3/2 X ( Y ) ≤ Y < 3/2 for 0 ≤ Y ≤ y 0 0\le Y\le y_0 0 ≤ Y ≤ y 0 . All zeros of a ( y ) a(y) a ( y ) are positive real; in particular a a a has no zeros in ∣ y ∣ < y 0 |y|<y_0 ∣ y ∣ < y 0 . For 0 ≤ Y < y 0 0\le Y<y_0 0 ≤ Y < y 0 , every zero of K Y K_Y K Y is positive and simple, and
K Y ( z ) = a ( Y ) ∏ j ≥ 1 ( 1 − z r j ( Y ) ) , 0 < r 1 ( Y ) < r 2 ( Y ) < ⋯ . (10) K_Y(z)=a(Y)\prod_{j\ge1}\left(1-\frac{z}{r_j(Y)}\right),\qquad0<r_1(Y)<r_2(Y)<\cdots. \tag*{(10)} K Y ( z ) = a ( Y ) j ≥ 1 ∏ ( 1 − r j ( Y ) z ) , 0 < r 1 ( Y ) < r 2 ( Y ) < ⋯ . (10)
At Y = 0 Y=0 Y = 0 one has K 0 ( z ) = sin 2 z / 2 z K_0(z)=\sin\sqrt{2}z/\sqrt{2}z K 0 ( z ) = sin 2 z / 2 z , and the lemma is the classical description of the zeros of the sine function. For Y > 0 Y>0 Y > 0 , integration along the characteristics of a first-order equation for K Y K_Y K Y writes K Y K_Y K Y as sin t \sin t sin t minus a small positive correction; concavity then locates the real zeros, and Rouché’s theorem shows that there are no others.
Reality of the zeros does not carry over from K Y K_Y K Y to H Y H_Y H Y : some higher zeros of H Y H_Y H Y can be nonreal. The required substitute is the following stronger coefficient assertion.
Lemma 2.2 (Coefficient positivity). For 0 ≤ Y < y 0 0\le Y<y_0 0 ≤ Y < y 0 ,
log a ( Y ) H Y ( z ) = ∑ m ≥ 1 t m ( Y ) z m , t m ( q ) ( 0 ) ≥ 0 ( m ≥ 1 , q ≥ 0 ) , t m ( 0 ) > 0 , (11) \log\frac{a(Y)}{H_Y(z)}=\sum_{m\ge1}t_m(Y)z^m,\qquad t_m^{(q)}(0)\ge0\ (m\ge1,q\ge0),\qquad t_m(0)>0, \tag*{(11)} log H Y ( z ) a ( Y ) = m ≥ 1 ∑ t m ( Y ) z m , t m ( q ) ( 0 ) ≥ 0 ( m ≥ 1 , q ≥ 0 ) , t m ( 0 ) > 0 , (11) where each t m t_m t m is analytic on ∣ Y ∣ < y 0 |Y|<y_0 ∣ Y ∣ < y 0 and its Taylor series at zero converges to t m ( Y ) t_m(Y) t m ( Y ) there. Consequently H Y − 1 H_Y^{-1} H Y − 1 has strictly positive Taylor coefficients for every real h > 0 h>0 h > 0 , and − H Y ′ / H Y -H_Y'/H_Y − H Y ′ / H Y has nonnegative coefficients. Moreover every Taylor coefficient in Y Y Y of
ϑ ^ ℓ ( Y ) : = a ( Y ) [ z ℓ ] K Y ( z ) − 1 \widehat{\vartheta}_{\ell}(Y):=a(Y)[z^\ell]K_Y(z)^{-1} ϑ ℓ ( Y ) := a ( Y ) [ z ℓ ] K Y ( z ) − 1 is nonnegative, and the same holds for a − 1 a^{-1} a − 1 and for ϑ ℓ = [ z ℓ ] K Y − 1 \vartheta_\ell=[z^\ell]K_Y^{-1} ϑ ℓ = [ z ℓ ] K Y − 1 .
The coefficients t m t_m t m satisfy a system of ordinary differential equations in Y Y Y whose right-hand sides are polynomials with nonnegative coefficients, and they are positive at Y = 0 Y=0 Y = 0 , where H 0 H_0 H 0 has only positive real zeros; induction on the order of differentiation gives the lemma. These are statements about Taylor coefficients in the parameter Y Y Y , which is stronger than positivity at each real Y Y Y ; they supply the positive convolution factor of § 3.25.1 .
Rescaled kernels and a local limit theorem
Set r = r 1 ( Y ) r=r_1(Y) r = r 1 ( Y ) ,
K ‾ Y ( w ) = K Y ( r w ) / a ( Y ) , H ‾ Y ( w ) = H Y ( r w ) / a ( Y ) , A ( Y ) = H ‾ Y ( 1 ) / 3. (12) \overline{K}_Y(w)=K_Y(rw)/a(Y),\qquad\overline{H}_Y(w)=H_Y(rw)/a(Y),\qquad A(Y)=\overline{H}_Y(1)/3. \tag*{(12)} K Y ( w ) = K Y ( r w ) / a ( Y ) , H Y ( w ) = H Y ( r w ) / a ( Y ) , A ( Y ) = H Y ( 1 ) /3. (12)
The first zero r r r is simple and varies analytically. Integration of (9) gives
H Y ( z ) = 3 2 z 3 / 2 ∫ 0 z v 1 / 2 K Y ( v ) d v , H_Y(z)=\frac{3}{2z^{3/2}}\int_0^z v^{1/2}K_Y(v)\,\mathrm{d}v, H Y ( z ) = 2 z 3/2 3 ∫ 0 z v 1/2 K Y ( v ) d v ,
so H Y ( z ) > 0 H_Y(z)>0 H Y ( z ) > 0 on [ 0 , r ] [0,r] [ 0 , r ] . Since H Y − 1 H_Y^{-1} H Y − 1 has nonnegative Taylor coefficients, its radius of convergence exceeds r r r . Indeed, by Pringsheim’s theorem a power series with nonnegative coefficients is singular at the positive point of its circle of convergence; if the radius were at most r r r , that point would be a zero of H Y H_Y H Y in ( 0 , r ] (0,r] ( 0 , r ] . In particular H Y H_Y H Y has no complex zero of modulus at most r r r .
At the critical endpoint put b = − X ′ ′ ( y 0 ) > 0 b=-X''(y_0)>0 b = − X ′′ ( y 0 ) > 0 . The equation K Y ( r ) = 0 K_Y(r)=0 K Y ( r ) = 0 and Taylor expansion at Y = y 0 , r = 0 Y=y_0,r=0 Y = y 0 , r = 0 give
r ( Y ) ∼ 3 a ( Y ) / b , K ‾ Y ( w ) ⟶ 1 − w , H ‾ Y ( w ) ⟶ 1 − 3 w / 5. (13) r(Y)\sim3a(Y)/b,\qquad\overline{K}_Y(w)\longrightarrow1-w,\qquad\overline{H}_Y(w)\longrightarrow1-3w/5. \tag*{(13)} r ( Y ) ∼ 3 a ( Y ) / b , K Y ( w ) ⟶ 1 − w , H Y ( w ) ⟶ 1 − 3 w /5. (13)
These convergences are locally uniform in complex w w w . The family therefore extends continuously to Y = y 0 Y=y_0 Y = y 0 . Compactness and nonvanishing on the closed unit disk give one η > 0 \eta>0 η > 0 such that all H ‾ Y − 1 \overline{H}_Y^{-1} H Y − 1 are analytic and uniformly bounded on ∣ w ∣ ≤ 1 + η |w|\le1+\eta ∣ w ∣ ≤ 1 + η , including this endpoint. Likewise K ‾ Y / ( 1 − w ) \overline{K}_Y/(1-w) K Y / ( 1 − w ) is nonvanishing on a common disk of radius greater than one: its zeros are the higher positive roots from (10) , and (13) treats the endpoint. Consequently, with θ − 1 = 0 \theta_{-1}=0 θ − 1 = 0 ,
1 = θ 0 ≤ θ 1 ≤ ⋯ ≤ C , θ ℓ = [ w ℓ ] K ‾ Y ( w ) − 1 , w ℓ : = θ ℓ − θ ℓ − 1 ≥ 0 , sup Y ∑ ℓ ≥ 0 ( 1 + ℓ ) q w ℓ < ∞ . (14) 1=\theta_0\le\theta_1\le\cdots\le C,\qquad\theta_\ell=[w^\ell]\overline{K}_Y(w)^{-1},\qquad w_\ell:=\theta_\ell-\theta_{\ell-1}\ge0,\qquad\sup_Y\sum_{\ell\ge0}(1+\ell)^q w_\ell<\infty. \tag*{(14)} 1 = θ 0 ≤ θ 1 ≤ ⋯ ≤ C , θ ℓ = [ w ℓ ] K Y ( w ) − 1 , w ℓ := θ ℓ − θ ℓ − 1 ≥ 0 , Y sup ℓ ≥ 0 ∑ ( 1 + ℓ ) q w ℓ < ∞. (14)
Indeed ∑ w ℓ w ℓ = ( 1 − w ) / K ‾ Y ( w ) \sum w_\ell w^\ell=(1-w)/\overline{K}_Y(w) ∑ w ℓ w ℓ = ( 1 − w ) / K Y ( w ) is the inverse product over the higher roots. It has nonnegative coefficients and a common analytic disk. Also A ( Y ) A(Y) A ( Y ) is bounded above and away from zero.
Finally,
P Y ( w ) = H ‾ Y ( 1 ) H ‾ Y ( w ) (15) P_Y(w)=\frac{\overline{H}_Y(1)}{\overline{H}_Y(w)} \tag*{(15)} P Y ( w ) = H Y ( w ) H Y ( 1 ) (15)
is the probability generating function of a nonnegative integer variable. Its probabilities at zero and one have uniform positive lower bounds, its exponential moments are uniformly bounded, and (9) at the first root gives mean 3 / 2 3/2 3/2 . Its variance σ Y 2 \sigma_Y^2 σ Y 2 is bounded above and away from zero. At Y = y 0 Y=y_0 Y = y 0 the law is geometric, with generating function ( 2 / 5 ) / ( 1 − 3 w / 5 ) (2/5)/(1-3w/5) ( 2/5 ) / ( 1 − 3 w /5 ) , so its variance is 15 / 4 15/4 15/4 .
For a sum S h S_h S h of h h h independent copies, Fourier inversion now gives, uniformly in Y Y Y and in integers m m m with ∣ m − 3 h / 2 ∣ ≤ L h \lvert m-3h/2\rvert\le L\sqrt{h} ∣ m − 3 h /2 ∣ ≤ L h , for each fixed L L L ,
Pr ( S h = m ) = 1 2 π h σ Y 2 exp ( − ( m − 3 h / 2 ) 2 2 h σ Y 2 ) + o ( h − 1 / 2 ) . (16) \Pr(S_h=m)=\frac{1}{\sqrt{2\pi h\sigma_Y^2}}\exp\left(-\frac{(m-3h/2)^2}{2h\sigma_Y^2}\right)+o(h^{-1/2}). \tag*{(16)} Pr ( S h = m ) = 2 π h σ Y 2 1 exp ( − 2 h σ Y 2 ( m − 3 h /2 ) 2 ) + o ( h − 1/2 ) . (16)
Indeed, on a fixed neighborhood of zero the centered characteristic logarithm is − σ Y 2 t 2 / 2 + O ( ∣ t ∣ 3 ) -\sigma_Y^2t^2/2+O(\lvert t\rvert^3) − σ Y 2 t 2 /2 + O (∣ t ∣ 3 ) , uniformly by the exponential moments. Away from zero its modulus is uniformly less than one, because the first two probabilities have uniform positive lower bounds. Scaling t t t by h − 1 / 2 h^{-1/2} h − 1/2 , these facts give an integrable Gaussian bound and pointwise convergence; Fourier inversion proves (16) . At a bounded displacement from the mean the error is relative O ( h − 1 ) O(h^{-1}) O ( h − 1 ) , by expansion through fourth order and cancellation of the odd cubic integral. Thus
[ w m ] H ˉ Y ( w ) − h = H ˉ Y ( 1 ) − h Pr ( S h = m ) [w^m]\bar H_Y(w)^{-h}=\bar H_Y(1)^{-h}\Pr(S_h=m) [ w m ] H ˉ Y ( w ) − h = H ˉ Y ( 1 ) − h Pr ( S h = m )
has a uniform central lower bound c L H ˉ Y ( 1 ) − h / h c_L\bar H_Y(1)^{-h}/\sqrt{h} c L H ˉ Y ( 1 ) − h / h , and its logarithm at a bounded displacement is − h log H ˉ Y ( 1 ) − 1 2 log h + O ( 1 ) -h\log\bar H_Y(1)-\frac{1}{2}\log h+O(1) − h log H ˉ Y ( 1 ) − 2 1 log h + O ( 1 ) . These coefficient estimates enter the harmonic model through (36) –(40) in § 3.4 , and the central estimate at bounded displacement fixes the normalization of the leading term in (102) .
The density parameter
This section establishes the properties of the functions T T T , Δ = T − Y \Delta=T-Y Δ = T − Y , I I I and λ \lambda λ defined in (2) and (3) ; its single result is Lemma 3.1 3.1 3.1 , whose proof, which uses the Sonine integral and elementary calculus, is given in Appendix B.§ 3.3 . The meaning of the four functions is as follows. The first positive zero of K Y K_Y K Y is r ( Y ) = 2 Δ ( Y ) r(Y)=2\Delta(Y) r ( Y ) = 2Δ ( Y ) ; equation (2) is the condition K Y ( 2 Δ ) = 0 K_Y(2\Delta)=0 K Y ( 2Δ ) = 0 written in the variable v = T − Δ s 2 v=T-\Delta s^2 v = T − Δ s 2 . The integral I ( Y ) I(Y) I ( Y ) is the action that governs the exponential growth in (4) , and λ ( Y ) / 2 \lambda(Y)/2 λ ( Y ) /2 is the logarithmic derivative of I ( Y ) − 1 I(Y)^{-1} I ( Y ) − 1 with respect to log x \log x log x , that is, the leading cusp-count scale per unit complexity. Lemma 3.1 3.1 3.1 shows that T T T is well defined and that λ \lambda λ is an analytic increasing bijection from [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) onto [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , so that the profiles B , C B,C B , C of (7) are functions of the density λ ≥ 0 \lambda\ge0 λ ≥ 0 ; it also records the behavior of all four functions at the two ends of the interval, which is what the extraction arguments of § 3.25 and § 3.31 need near x = 0 x=0 x = 0 and near x = x c x=x_c x = x c .
For the local-limit arguments we view λ \lambda λ as a function of the cusp weight and put
Ψ ( x ) = λ ( Y ( x ) ) / 2 , x = X ( Y ) , D = x d d x . \Psi(x)=\lambda(Y(x))/2,\qquad x=X(Y),\qquad\mathcal{D}=x\frac{\mathrm{d}}{\mathrm{d}x}. Ψ ( x ) = λ ( Y ( x )) /2 , x = X ( Y ) , D = x d x d .
Lemma 3.1 (The parameter). (i) The root T ( Y ) T(Y) T ( Y ) of (2) in ( y 0 , π 2 / 4 ] (y_0,\pi^2/4] ( y 0 , π 2 /4 ] exists, is unique and depends analytically on Y ∈ [ 0 , y 0 ) Y\in[0,y_0) Y ∈ [ 0 , y 0 ) . One has T ( 0 ) = π 2 / 4 T(0)=\pi^2/4 T ( 0 ) = π 2 /4 , − 1 ≤ T ′ < 0 -1\le T'<0 − 1 ≤ T ′ < 0 , T ′ ′ > 0 T''>0 T ′′ > 0 , and T → y 0 T\to y_0 T → y 0 as Y → y 0 Y\to y_0 Y → y 0 . Moreover I > 0 I>0 I > 0 .
(ii) Put
J ( Y ) = ∫ 0 1 X ( T − Δ s 2 ) d s = X ( Y ) + I ( Y ) Δ , β ( Y ) = X ( Y ) J ( Y ) = λ ( Y ) 2 + λ ( Y ) . J(Y)=\int_0^1 X(T-\Delta s^2)\,\mathrm{d}s=X(Y)+\frac{I(Y)}{\sqrt{\Delta}},\qquad\beta(Y)=\frac{X(Y)}{J(Y)}=\frac{\lambda(Y)}{2+\lambda(Y)}. J ( Y ) = ∫ 0 1 X ( T − Δ s 2 ) d s = X ( Y ) + Δ I ( Y ) , β ( Y ) = J ( Y ) X ( Y ) = 2 + λ ( Y ) λ ( Y ) . Then β ′ > 0 \beta'>0 β ′ > 0 , and λ \lambda λ is an analytic strictly increasing bijection from [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) onto [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . At Y = 0 Y=0 Y = 0 , Δ = T = π 2 / 4 \Delta=T=\pi^2/4 Δ = T = π 2 /4 , I ( 0 ) = π / 4 I(0)=\pi/4 I ( 0 ) = π /4 , J ( 0 ) = 1 / 2 J(0)=1/2 J ( 0 ) = 1/2 , and λ ( Y ) = 4 Y + O ( Y 2 ) \lambda(Y)=4Y+O(Y^2) λ ( Y ) = 4 Y + O ( Y 2 ) .
(iii) Put δ = y 0 − Y \delta=y_0-Y δ = y 0 − Y and b = − X ′ ′ ( y 0 ) = x c / y 0 > 0 b=-X''(y_0)=x_c/y_0>0 b = − X ′′ ( y 0 ) = x c / y 0 > 0 . As δ → 0 \delta\to0 δ → 0 ,
Δ = 3 2 δ + O ( δ 2 ) , I ( Y ) = 8 b 45 Δ 5 / 2 ( 1 + O ( Δ ) ) , (17) \Delta=\frac{3}{2}\delta+O(\delta^2),\qquad I(Y)=\frac{8b}{45}\Delta^{5/2}(1+O(\Delta)), \tag*{(17)} Δ = 2 3 δ + O ( δ 2 ) , I ( Y ) = 45 8 b Δ 5/2 ( 1 + O ( Δ )) , (17)
λ ∼ 45 x c 4 b Δ − 2 = 45 j 0 , 1 2 16 Δ − 2 , (18) \lambda\sim\frac{45x_c}{4b}\Delta^{-2}=\frac{45j_{0,1}^{2}}{16}\Delta^{-2}, \tag*{(18)} λ ∼ 4 b 45 x c Δ − 2 = 16 45 j 0 , 1 2 Δ − 2 , (18)
and consequently
B ( λ ) ∼ π 2 3 5 j 0 , 1 λ , C ( λ ) ∼ π 2 x c λ . (19) \mathcal{B}(\lambda)\sim\frac{\pi^{2}}{3\sqrt{5}j_{0,1}}\sqrt{\lambda},\qquad\mathcal{C}(\lambda)\sim\frac{\pi^{2}}{x_c}\lambda. \tag*{(19)} B ( λ ) ∼ 3 5 j 0 , 1 π 2 λ , C ( λ ) ∼ x c π 2 λ . (19)
Also Δ ≍ 1 − β \Delta\asymp\sqrt{1-\beta} Δ ≍ 1 − β .
(iv) The function Ψ \Psi Ψ is analytic and strictly increasing for 0 < x < x c 0<x<x_c 0 < x < x c . With ε = 1 − x / x c \varepsilon=1-x/x_c ε = 1 − x / x c ,
x ↓ 0 : Ψ = 2 x + O ( x 2 ) , D Ψ = 2 x + O ( x 2 ) , D 2 Ψ = 2 x + O ( x 2 ) , x ↑ x c : Ψ = 5 4 ε ( 1 + O ( ε ) ) , D Ψ = 5 4 ε 2 ( 1 + O ( ε ) ) , D 2 Ψ = O ( ε − 3 ) . (20) \begin{aligned}
x\downarrow0:\quad&\Psi=2x+O(x^{2}),\quad\mathcal{D}\Psi=2x+O(x^{2}),\quad\mathcal{D}^{2}\Psi=2x+O(x^{2}),\\
x\uparrow x_c:\quad&\Psi=\frac{5}{4\varepsilon}(1+O(\sqrt{\varepsilon})),\quad\mathcal{D}\Psi=\frac{5}{4\varepsilon^{2}}(1+O(\sqrt{\varepsilon})),\quad\mathcal{D}^{2}\Psi=O(\varepsilon^{-3}). \tag*{(20)}
\end{aligned} x ↓ 0 : x ↑ x c : Ψ = 2 x + O ( x 2 ) , D Ψ = 2 x + O ( x 2 ) , D 2 Ψ = 2 x + O ( x 2 ) , Ψ = 4 ε 5 ( 1 + O ( ε )) , D Ψ = 4 ε 2 5 ( 1 + O ( ε )) , D 2 Ψ = O ( ε − 3 ) . (20)
These expansions arise from convergent analytic expansions in x x x or ε \sqrt{\varepsilon} ε . In particular Ψ 2 / ( D Ψ ) \Psi^{2}/(\mathcal{D}\Psi) Ψ 2 / ( D Ψ ) has finite endpoint limits 0 0 0 and 5 / 4 5/4 5/4 , and is bounded on the full interval.
The estimate Δ ≍ 1 − β \Delta\asymp\sqrt{1-\beta} Δ ≍ 1 − β in (iii) implies that the logarithmic chain profile of § 3.34 is integrable at one. Part (iv) is restated as (110) in § 3.25 , where it provides the center and variance scales of the local limit theorem, and (85) identifies Ψ \Psi Ψ with the scalar x / ( A a r ) x/(Aar) x / ( A a r ) that arises from the marked cusp ratio.
Intersection amplitudes and the comparison model
This section carries out the first idea of Subsection 1.5. Starting from the alternating formula (24) for the Weil–Petersson brackets, we move along the string flow by a time Y Y Y and obtain the shifted amplitudes (25) , which interpolate between the brackets at Y = 0 Y=0 Y = 0 and pure ψ \psi ψ intersection numbers at Y = y 0 Y=y_0 Y = y 0 . Their normalized form (26) satisfies the recursion (28) , whose coefficients are the Taylor coefficients θ ℓ \theta_\ell θ ℓ of 1 / K ‾ Y 1/\overline{K}_Y 1/ K Y and hence nonnegative by § 3.1 . We then introduce the harmonic model (31) , an explicit solvable comparison for this recursion, and record the mass identity (32) and the coefficient estimates (36) –(41) that § 3.7 uses to compare the two, uniformly in the degrees and in the density parameter. The external inputs are the intersection identities of [117, 136, 129] and the fixed-genus radius of convergence from [136]; we also use the coarse bounds on normalized pure ψ \psi ψ intersections proved in Appendix § 8.4 .
Intersection identities and normalization
We use the following notation for a stable pair ( g , n ) (g,n) ( g , n ) and a degree tuple d \mathbf{d} d : h h h is the complexity, N N N the dimension of M g , n \mathcal{M}_{g,n} M g , n , and the deficit m m m is the number of κ 1 \kappa_1 κ 1 factors needed to fill up the dimension. Write
h = 2 g − 2 + n , N = 3 g − 3 + n , D = ∑ i = 1 n d i , m = N − D . h=2g-2+n,\qquad N=3g-3+n,\qquad D=\sum_{i=1}^{n}d_i,\qquad m=N-D. h = 2 g − 2 + n , N = 3 g − 3 + n , D = i = 1 ∑ n d i , m = N − D .
Throughout the comparison of § 3.7 , n ≥ 1 n\geq1 n ≥ 1 , the pair is stable, and 0 ≤ D ≤ N 0\leq D\leq N 0 ≤ D ≤ N . A tuple with a negative degree or an unstable pair has amplitude zero.
Recall X X X , a = X ′ > 0 a=X'>0 a = X ′ > 0 on [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) and the kernels K Y K_Y K Y and H Y H_Y H Y of (8) , and write r Y = r ( Y ) r_Y=r(Y) r Y = r ( Y ) for the first positive zero of K Y K_Y K Y . The properties of these kernels used below were established in § 3.1 . Put
H ‾ Y ( z ) = H Y ( r Y z ) a , K ‾ Y ( z ) = K Y ( r Y z ) a = H ‾ Y ( z ) + 2 3 z H ‾ Y ′ ( z ) , θ ℓ ( Y ) = [ z ℓ ] K ‾ Y ( z ) − 1 . (21) \overline{H}_Y(z)=\frac{H_Y(r_Yz)}{a},\qquad\overline{K}_Y(z)=\frac{K_Y(r_Yz)}{a}=\overline{H}_Y(z)+\frac{2}{3}z\overline{H}'_Y(z),\qquad\theta_\ell(Y)=[z^\ell]\overline{K}_Y(z)^{-1}. \tag*{(21)} H Y ( z ) = a H Y ( r Y z ) , K Y ( z ) = a K Y ( r Y z ) = H Y ( z ) + 3 2 z H Y ′ ( z ) , θ ℓ ( Y ) = [ z ℓ ] K Y ( z ) − 1 . (21)
At Y = y 0 Y=y_0 Y = y 0 these have the continuous definitions
H ‾ y 0 ( z ) = 1 − 3 5 z , K ‾ y 0 ( z ) = 1 − z , θ ℓ ( y 0 ) = 1. (22) \overline{H}_{y_0}(z)=1-\frac{3}{5}z,\qquad\overline{K}_{y_0}(z)=1-z,\qquad\theta_\ell(y_0)=1. \tag*{(22)} H y 0 ( z ) = 1 − 5 3 z , K y 0 ( z ) = 1 − z , θ ℓ ( y 0 ) = 1. (22)
In particular θ 0 = 1 \theta_0=1 θ 0 = 1 . These are the rescaled kernels and coefficients of (12) and (14) , recalled here together with their endpoint values.
We use the following facts from intersection theory. The ordinary string, dilaton, and Dijkgraaf–Verlinde–Verlinde (DVV) identities hold for Witten’s psi intersections [180, 120, 56], with ⟨ τ 0 ⟩ 0 = 1 \langle\tau_0\rangle_0=1 ⟨ τ 0 ⟩ 0 = 1 and ⟨ τ 1 ⟩ 1 = 1 / 24 \langle\tau_1\rangle_1=1/24 ⟨ τ 1 ⟩ 1 = 1/24 . The normalized forms used here are written in Appendix C. For pure ψ \psi ψ intersection numbers we use the normalization
C g ( e ) = 4 g ∏ i ( 2 e i + 1 ) ! ! 3 2 g − 2 + q Γ ( 2 g − 2 + q ) ⟨ ∏ i = 1 q τ e i ⟩ g , ∑ i e i = 3 g − 3 + q . (23) C_g(\mathbf{e})=\frac{4^g\prod_{i}(2e_i+1)!!}{3^{2g-2+q}\Gamma(2g-2+q)}\left\langle\prod_{i=1}^{q}\tau_{e_i}\right\rangle_g,\qquad\sum_i e_i=3g-3+q. \tag*{(23)} C g ( e ) = 3 2 g − 2 + q Γ ( 2 g − 2 + q ) 4 g ∏ i ( 2 e i + 1 )!! ⟨ i = 1 ∏ q τ e i ⟩ g , i ∑ e i = 3 g − 3 + q . (23)
For positive degrees we use only
1 6 ≤ C g ( e ) ≤ 1. \frac{1}{6}\leq C_g(\mathbf{e})\leq1. 6 1 ≤ C g ( e ) ≤ 1.
Appendix C proves these coarse bounds directly from DVV and the published two-point estimate. The argument adapts the recursion estimates of Guo, Yang and Zagier [83]; their stronger lower bound C g ( 3 g − 2 ) ≤ C g ( e ) C_g(3g-2)\leq C_g(\mathbf{e}) C g ( 3 g − 2 ) ≤ C g ( e ) is not needed here.
With m = 3 g − 3 + n − ∣ d ∣ m=3g-3+n-|\mathbf{d}| m = 3 g − 3 + n − ∣ d ∣ , our mixed brackets are
[ τ d 1 ⋯ τ d n ] g , n = ∏ i ( 2 d i + 1 ) ! ! 4 ∣ d ∣ ( 2 π 2 ) m m ! ∫ M g , n κ 1 m ∏ i ψ i d i . [\tau_{d_1}\cdots\tau_{d_n}]_{g,n}=\prod_i(2d_i+1)!!\,4^{|\mathbf{d}|}\frac{(2\pi^2)^m}{m!}\int_{\mathcal{M}_{g,n}}\kappa_1^m\prod_i\psi_i^{d_i}. [ τ d 1 ⋯ τ d n ] g , n = i ∏ ( 2 d i + 1 )!! 4 ∣ d ∣ m ! ( 2 π 2 ) m ∫ M g , n κ 1 m i ∏ ψ i d i .
Here κ 1 \kappa_1 κ 1 is the logarithmic Mumford class, ψ i \psi_i ψ i is the cotangent class, and [ ω W P ] = 2 π 2 κ 1 [\omega_{\mathrm{WP}}]=2\pi^2\kappa_1 [ ω WP ] = 2 π 2 κ 1 [182]; see also [147, Section 2]. Thus
V g , n ( 2 L ) = ∑ ∣ d ∣ ≤ 3 g − 3 + n [ τ d 1 ⋯ τ d n ] g , n ∏ i L i 2 d i ( 2 d i + 1 ) ! . V_{g,n}(2\mathbf{L})=\sum_{|\mathbf{d}|\leq3g-3+n}[\tau_{d_1}\cdots\tau_{d_n}]_{g,n}\prod_i\frac{L_i^{2d_i}}{(2d_i+1)!}. V g , n ( 2 L ) = ∣ d ∣ ≤ 3 g − 3 + n ∑ [ τ d 1 ⋯ τ d n ] g , n i ∏ ( 2 d i + 1 )! L i 2 d i .
Angle brackets below denote unweighted pure-ψ \psi ψ integrals. Intersections of the wrong dimension and negative-degree amplitudes are zero. Unstable factors in the positive recursion are zero; explicitly specified genus-zero initial derivatives are kept separate.
The mixed-kappa/psi transform [117, 129] gives, at Y = 0 Y=0 Y = 0 ,
\frac{[\tau_{d_1\cdots\tau_{d_n}]_{g,n}}{4^D(2\pi^2)^m}=\prod_i(2d_i+1)!!\sum_{\substack{\ell_k\geq0,\ k\geq2\\ \sum_{k\geq2}(k-1)\ell_k=m}}\left\langle\tau_{\mathbf{d}}\prod_{k\geq2}\tau_k^{\ell_k}\right\rangle_g\prod_{k\geq2}\frac{\left((-1)^k/(k-1)!\right)^{\ell_k}}{\ell_k!}.} \tag*{(24)}
Since X ′ ( 0 ) = 1 X'(0)=1 X ′ ( 0 ) = 1 and − X ( k ) ( 0 ) = ( − 1 ) k / ( k − 1 ) ! -X^{(k)}(0)=(-1)^k/(k-1)! − X ( k ) ( 0 ) = ( − 1 ) k / ( k − 1 )! , the weights in (24) are the derivatives − X ( k ) ( 0 ) -X^{(k)}(0) − X ( k ) ( 0 ) , k ≥ 2 k\geq2 k ≥ 2 . The deformation replaces them by the derivatives − X ( k ) ( Y ) -X^{(k)}(Y) − X ( k ) ( Y ) at a point Y Y Y : for general Y < y 0 Y<y_0 Y < y 0 , define the shifted amplitudes by the finite sum
U g , n Y ( d ) = ∏ i ( 2 d i + 1 ) ! ! ∑ ℓ k ≥ 0 , k ≥ 2 ∑ k ≥ 2 ( k − 1 ) ℓ k = m ⟨ τ d ∏ k ≥ 2 τ k ℓ k ⟩ g ∏ k ≥ 2 ( − X ( k ) ( Y ) ) ℓ k / ℓ k ! a h + ∑ ℓ k . (25) U^Y_{g,n}(\mathbf{d})=\prod_i(2d_i+1)!!\sum_{\substack{\ell_k\geq0,\ k\geq2\\ \sum_{k\geq2}(k-1)\ell_k=m}}\left\langle\tau_{\mathbf{d}}\prod_{k\geq2}\tau_k^{\ell_k}\right\rangle_g\frac{\prod_{k\geq2}(-X^{(k)}(Y))^{\ell_k}/\ell_k!}{a^{h+\sum\ell_k}}. \tag*{(25)} U g , n Y ( d ) = i ∏ ( 2 d i + 1 )!! ℓ k ≥ 0 , k ≥ 2 ∑ k ≥ 2 ( k − 1 ) ℓ k = m ∑ ⟨ τ d k ≥ 2 ∏ τ k ℓ k ⟩ g a h + ∑ ℓ k ∏ k ≥ 2 ( − X ( k ) ( Y ) ) ℓ k / ℓ k ! . (25)
The denominator is obtained by summing degree-one background insertions with time t 1 = 1 − a t_1=1-a t 1 = 1 − a : the dilaton equation gives the factor ∑ j ≥ 0 ( h + ∑ ℓ k ) j t 1 j / j ! = a − h − ∑ ℓ k \sum_{j\geq0}(h+\sum\ell_k)_j t_1^j/j!=a^{-h-\sum\ell_k} ∑ j ≥ 0 ( h + ∑ ℓ k ) j t 1 j / j ! = a − h − ∑ ℓ k , where ( c ) j = c ( c + 1 ) ⋯ ( c + j − 1 ) (c)_j=c(c+1)\cdots(c+j-1) ( c ) j = c ( c + 1 ) ⋯ ( c + j − 1 ) is the rising factorial. Define the normalized amplitudes
u g , n ( d ; Y ) = a h r Y m U g , n Y ( d ) . (26) u_{g,n}(\mathbf{d};Y)=a^{h}r_Y^mU^Y_{g,n}(\mathbf{d}). \tag*{(26)} u g , n ( d ; Y ) = a h r Y m U g , n Y ( d ) . (26)
They extend to Y = y 0 Y=y_0 Y = y 0 through their positive recursion below.
When all degrees vanish, the shifted amplitudes are derivatives of the cusp generating function. This is the familiar shift of the Kontsevich–Witten (KW) times along the string flow [106, 117]; the identification of Φ g \Phi_g Φ g with the KW free energy at the shifted times t 0 = x t_0=x t 0 = x , t 1 = 0 t_1=0 t 1 = 0 , t k = ( − 1 ) k / ( k − 1 ) ! t_k=(-1)^k/(k-1)! t k = ( − 1 ) k / ( k − 1 )! is recorded in [194, Remark 1], where it is attributed to Dubrovin, and the corresponding deformation in ( t 0 , t 1 ) (t_0,t_1) ( t 0 , t 1 ) is used in [156, Section 2.4]. We recall it in the present normalization. Start with KW times t 0 = x t_0=x t 0 = x , t 1 = 0 t_1=0 t 1 = 0 , t k = − X ( k ) ( 0 ) t_k=-X^{(k)}(0) t k = − X ( k ) ( 0 ) , k ≥ 2 k\geq2 k ≥ 2 . The finite formal string flow sends them to
t 0 ′ = x − X ( Y ) , t 1 ′ = 1 − X ′ ( Y ) , t k ′ = − X ( k ) ( Y ) ( k ≥ 2 ) . t'_0=x-X(Y),\qquad t'_1=1-X'(Y),\qquad t'_k=-X^{(k)}(Y)\quad(k\geq2). t 0 ′ = x − X ( Y ) , t 1 ′ = 1 − X ′ ( Y ) , t k ′ = − X ( k ) ( Y ) ( k ≥ 2 ) .
Its vector field is ∑ k ≥ 0 t k + 1 ′ ∂ t k − ∂ t 0 ′ \sum_{k\geq0}t'_{k+1}\partial_{t_k}-\partial_{t'_0} ∑ k ≥ 0 t k + 1 ′ ∂ t k − ∂ t 0 ′ . The string equation makes the genus-g g g free energy invariant for g ≥ 1 g\geq1 g ≥ 1 ; its genus-zero anomaly disappears after three t 0 t_0 t 0 derivatives. Since ∂ t 0 \partial_{t_0} ∂ t 0 commutes with this flow, setting x = X ( Y ) x=X(Y) x = X ( Y ) gives
U g , n Y ( 0 n ) = Φ g ( n ) ( x ) , Φ g ( x ) = ∑ n ≥ 0 V g , n ( 2 π 2 ) 3 g − 3 + n x n n ! . (27) U^Y_{g,n}(0^n)=\Phi_g^{(n)}(x),\qquad\Phi_g(x)=\sum_{n\geq0}\frac{V_{g,n}}{(2\pi^2)^{3g-3+n}}\frac{x^n}{n!}. \tag*{(27)} U g , n Y ( 0 n ) = Φ g ( n ) ( x ) , Φ g ( x ) = n ≥ 0 ∑ ( 2 π 2 ) 3 g − 3 + n V g , n n ! x n . (27)
whenever the amplitude is stable. Since the flow is evaluated at a finite time, we also give a direct argument. Here a = X ′ ( Y ) a=X'(Y) a = X ′ ( Y ) . Applying the string and dilaton equations to the background insertions in (25) , and using ∂ Y ( − X ( k ) ( Y ) ) = − X ( k + 1 ) ( Y ) \partial_Y(-X^{(k)}(Y))=-X^{(k+1)}(Y) ∂ Y ( − X ( k ) ( Y )) = − X ( k + 1 ) ( Y ) , gives ∂ Y U g , n Y ( 0 n ) = a U g , n + 1 Y ( 0 n + 1 ) \partial_Y U^Y_{g,n}(0^n)=aU^Y_{g,n+1}(0^{n+1}) ∂ Y U g , n Y ( 0 n ) = a U g , n + 1 Y ( 0 n + 1 ) , while ∂ Y Φ g ( n ) ( X ( Y ) ) = a Φ g ( n + 1 ) ( X ( Y ) ) \partial_Y\Phi_g^{(n)}(X(Y))=a\Phi_g^{(n+1)}(X(Y)) ∂ Y Φ g ( n ) ( X ( Y )) = a Φ g ( n + 1 ) ( X ( Y )) . At Y = 0 Y=0 Y = 0 the two families agree for every n n n by (24) ; by induction on the order, all their Y Y Y -derivatives at Y = 0 Y=0 Y = 0 agree. For fixed ( g , n ) (g,n) ( g , n ) the sum (25) is finite, so U g , n Y ( 0 n ) U^Y_{g,n}(0^n) U g , n Y ( 0 n ) is real analytic on a neighborhood of [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) , where a > 0 a>0 a > 0 ; so is Φ g ( n ) ( X ( Y ) ) \Phi_g^{(n)}(X(Y)) Φ g ( n ) ( X ( Y )) , because Φ g \Phi_g Φ g has radius of convergence x c x_c x c [136] and 0 ≤ X ( Y ) < x c 0\leq X(Y)<x_c 0 ≤ X ( Y ) < x c . Hence (27) holds on [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) ; compare the proof of [194, Theorem 1]. Unmarked Φ g \Phi_g Φ g is used only for g ≥ 2 g\geq2 g ≥ 2 . The unstable genus-zero derivatives are separate initial terms and are not substituted into the positive recursion.
The positive recursion
We single out the first marking and call it the pivot . Write ⟨ ⟨ τ d 1 ⋯ τ d q ⟩ ⟩ g = ∏ i ( 2 d i + 1 ) ! ! ⟨ τ d 1 ⋯ τ d q ⟩ g \langle\!\langle\tau_{d_1}\cdots\tau_{d_q}\rangle\!\rangle_g=\prod_i(2d_i+1)!!\langle\tau_{d_1}\cdots\tau_{d_q}\rangle_g ⟨ ⟨ τ d 1 ⋯ τ d q ⟩ ⟩ g = ∏ i ( 2 d i + 1 )!! ⟨ τ d 1 ⋯ τ d q ⟩ g . In this normalization the Dijkgraaf–Verlinde–Verlinde equation for the pivot reads, with ( − 1 ) ! ! = 1 (-1)!!=1 ( − 1 )!! = 1 ,
⟨ ⟨ τ d 1 τ d 2 ⋯ τ d q ⟩ ⟩ g = ∑ j = 2 q ( 2 d j + 1 ) ⟨ ⟨ τ d 1 + d j − 1 ∏ i ≠ 1 , j τ d i ⟩ ⟩ g + 1 2 ∑ u + v = d 1 − 2 ⟨ ⟨ τ u τ v τ d 2 ⋯ τ d q ⟩ ⟩ g − 1 + 1 2 ∑ u + v = d 1 − 2 ∑ g 1 + g 2 = g I ⊔ J = { 2 , … , q } ⟨ ⟨ τ u τ I ⟩ ⟩ g 1 ⟨ ⟨ τ v τ J ⟩ ⟩ g 2 . \begin{aligned}
\langle\!\langle\tau_{d_1}\tau_{d_2}\cdots\tau_{d_q}\rangle\!\rangle_g
={}&\sum_{j=2}^{q}(2d_j+1)\langle\!\langle\tau_{d_1+d_j-1}\prod_{i\neq1,j}\tau_{d_i}\rangle\!\rangle_g\\
&+\frac{1}{2}\sum_{u+v=d_1-2}\langle\!\langle\tau_u\tau_v\tau_{d_2}\cdots\tau_{d_q}\rangle\!\rangle_{g-1}\\
&+\frac{1}{2}\sum_{u+v=d_1-2}\sum_{\substack{g_1+g_2=g\\I\sqcup J=\{2,\ldots,q\}}}\langle\!\langle\tau_u\tau_I\rangle\!\rangle_{g_1}\langle\!\langle\tau_v\tau_J\rangle\!\rangle_{g_2}.
\end{aligned} ⟨ ⟨ τ d 1 τ d 2 ⋯ τ d q ⟩ ⟩ g = j = 2 ∑ q ( 2 d j + 1 ) ⟨ ⟨ τ d 1 + d j − 1 i = 1 , j ∏ τ d i ⟩ ⟩ g + 2 1 u + v = d 1 − 2 ∑ ⟨ ⟨ τ u τ v τ d 2 ⋯ τ d q ⟩ ⟩ g − 1 + 2 1 u + v = d 1 − 2 ∑ g 1 + g 2 = g I ⊔ J = { 2 , … , q } ∑ ⟨ ⟨ τ u τ I ⟩ ⟩ g 1 ⟨ ⟨ τ v τ J ⟩ ⟩ g 2 .
For d 1 = 0 d_1=0 d 1 = 0 and d 1 = 1 d_1=1 d 1 = 1 this is the string and the dilaton equation. In the amplitudes (25) , besides the n n n external markings, the intersection numbers contain background insertions: τ k \tau_k τ k with weight t k = − X ( k ) ( Y ) t_k=-X^{(k)}(Y) t k = − X ( k ) ( Y ) for k ≥ 2 k\geq2 k ≥ 2 , and, after the dilaton sum that produced the factor a − ∑ ℓ k a^{-\sum\ell_k} a − ∑ ℓ k , also τ 1 \tau_1 τ 1 with weight t 1 = 1 − a t_1=1-a t 1 = 1 − a . Apply the displayed equation with the pivot τ d 1 \tau_{d_1} τ d 1 and sum over the background. Three kinds of terms appear.
(a) The pivot merges with an external marking j ≥ 2 j\geq2 j ≥ 2 . By the first term above, this gives ( 2 d j + 1 ) U g , n − 1 Y ( d 1 + d j − 1 , … ) (2d_j+1)U^Y_{g,n-1}(d_1+d_j-1,\ldots) ( 2 d j + 1 ) U g , n − 1 Y ( d 1 + d j − 1 , … ) , with the background unchanged.
(b) The pivot merges with a background insertion τ k \tau_k τ k , k ≥ 1 k\geq1 k ≥ 1 . Background insertions carry no double factorial, so the factor 2 k + 1 2k+1 2 k + 1 of the first term is divided by ( 2 k + 1 ) ! ! (2k+1)!! ( 2 k + 1 )!! and becomes 1 / ( 2 k − 1 ) ! ! 1/(2k-1)!! 1/ ( 2 k − 1 )!! . Choosing which of the ℓ k \ell_k ℓ k copies is used turns t k ℓ k / ℓ k ! t_k^{\ell_k}/\ell_k! t k ℓ k / ℓ k ! into t k ⋅ t k ℓ k − 1 / ( ℓ k − 1 ) ! t_k\cdot t_k^{\ell_k-1}/(\ell_k-1)! t k ⋅ t k ℓ k − 1 / ( ℓ k − 1 )! . The result is
t k ( 2 k − 1 ) ! ! U g , n Y ( d 1 + k − 1 , d 2 , … , d n ) . \frac{t_k}{(2k-1)!!}U^Y_{g,n}(d_1+k-1,d_2,\ldots,d_n). ( 2 k − 1 )!! t k U g , n Y ( d 1 + k − 1 , d 2 , … , d n ) .
(c) Contractions. The two remaining terms give the nonseparating and separating contributions with coefficient 1 / 2 1/2 1/2 . In the separating case the background splits between the two components with the binomial factors of exponential generating functions, so the term is a product of two amplitudes.
Let E E E be the operator raising the pivot degree by one. The terms of type (b) form ( 1 − a ) U Y + ∑ k ≥ 2 t k E k − 1 U Y / ( 2 k − 1 ) ! ! (1-a)U^{Y}+\sum_{k\geq2}t_kE^{k-1}U^{Y}/(2k-1)!! ( 1 − a ) U Y + ∑ k ≥ 2 t k E k − 1 U Y / ( 2 k − 1 )!! . Moving them to the left-hand side and recalling t k = − X ( k ) ( Y ) t_k=-X^{(k)}(Y) t k = − X ( k ) ( Y ) gives
( a + ∑ k ≥ 2 X ( k ) ( Y ) ( 2 k − 1 ) ! ! E k − 1 ) U Y = K Y ( E ) U Y = (terms of types (a) and (c)) . \left(a+\sum_{k\geq2}\frac{X^{(k)}(Y)}{(2k-1)!!}E^{k-1}\right)U^{Y}=K_Y(E)U^{Y}=\text{(terms of types (a) and (c))}. ( a + k ≥ 2 ∑ ( 2 k − 1 )!! X ( k ) ( Y ) E k − 1 ) U Y = K Y ( E ) U Y = (terms of types (a) and (c)) .
where we compared with the series (8) for K Y K_Y K Y . For a fixed parent, raising the pivot degree more than m m m times leaves the dimension range, so E m + 1 U Y = 0 E^{m+1}U^{Y}=0 E m + 1 U Y = 0 and K Y ( E ) K_Y(E) K Y ( E ) has the finite inverse ∑ ℓ ≤ m [ z ℓ ] K Y ( z ) − 1 E ℓ \sum_{\ell\leq m}[z^\ell]K_Y(z)^{-1}E^\ell ∑ ℓ ≤ m [ z ℓ ] K Y ( z ) − 1 E ℓ . This is how the Taylor coefficients of 1 / K Y 1/K_Y 1/ K Y enter. Components of the wrong dimension vanish, and for external complexity h ≥ 2 h\geq2 h ≥ 2 no exceptional initial tuple occurs.
After (26) , the exact recursion is
U g , n ( d ) = ∑ j = 2 n ( 2 d j + 1 ) ∑ ℓ = 0 m θ ℓ U g , n − 1 ( d 1 + d j + ℓ − 1 , d 1 , j ^ ) + 1 2 ∑ ℓ = 0 m θ ℓ ∑ u + v = d 1 + ℓ − 2 U g − 1 , n + 1 ( u , v , d 1 ^ ) + 1 2 ∑ ℓ = 0 m θ ℓ ∑ u + v = d 1 + ℓ − 2 ∑ g 1 + g 2 = g I ⊔ J = { 2 , … , n } U g 1 , ∣ I ∣ + 1 ( u , d I ) U g 2 , ∣ J ∣ + 1 ( v , d J ) . (28) \begin{aligned}
U_{g,n}(\mathbf d)={}&\sum_{j=2}^{n}(2d_j+1)\sum_{\ell=0}^{m}\theta_\ell U_{g,n-1}(d_1+d_j+\ell-1,\mathbf d_{\widehat{1,j}})\\
&+\frac{1}{2}\sum_{\ell=0}^{m}\theta_\ell\sum_{u+v=d_1+\ell-2}U_{g-1,n+1}(u,v,\mathbf d_{\widehat{1}})\\
&+\frac{1}{2}\sum_{\ell=0}^{m}\theta_\ell\sum_{u+v=d_1+\ell-2}\sum_{\substack{g_1+g_2=g\\ I\sqcup J=\{2,\ldots,n\}}}U_{g_1,|I|+1}(u,\mathbf d_I)U_{g_2,|J|+1}(v,\mathbf d_J). \tag*{(28)}
\end{aligned} U g , n ( d ) = j = 2 ∑ n ( 2 d j + 1 ) ℓ = 0 ∑ m θ ℓ U g , n − 1 ( d 1 + d j + ℓ − 1 , d 1 , j ) + 2 1 ℓ = 0 ∑ m θ ℓ u + v = d 1 + ℓ − 2 ∑ U g − 1 , n + 1 ( u , v , d 1 ) + 2 1 ℓ = 0 ∑ m θ ℓ u + v = d 1 + ℓ − 2 ∑ g 1 + g 2 = g I ⊔ J = { 2 , … , n } ∑ U g 1 , ∣ I ∣ + 1 ( u , d I ) U g 2 , ∣ J ∣ + 1 ( v , d J ) . (28)
Here d 1 , j ^ \mathbf d_{\widehat{1,j}} d 1 , j denotes the tuple d \mathbf d d with the entries d 1 d_1 d 1 and d j d_j d j removed, and d 1 ^ \mathbf d_{\widehat{1}} d 1 the tuple with d 1 d_1 d 1 removed. The indices u u u and v v v are the degrees at the two new markings corresponding to the branches of the node. The last split is ordered. We call the amplitudes on the right of (28) the children of the amplitude on the left: a merging or nonseparating child has complexity h − 1 h-1 h − 1 , and the two factors of a separating term have complexities summing to h − 1 h-1 h − 1 . The initial values of the recursion, computed directly from (25) , are
U 0 , 3 ( 0 , 0 , 0 ) = 1 , U 1 , 1 ( 1 ) = 1 8 , U 1 , 1 ( 0 ) = θ 1 8 . (29) U_{0,3}(0,0,0)=1,\qquad U_{1,1}(1)=\frac{1}{8},\qquad U_{1,1}(0)=\frac{\theta_1}{8}. \tag*{(29)} U 0 , 3 ( 0 , 0 , 0 ) = 1 , U 1 , 1 ( 1 ) = 8 1 , U 1 , 1 ( 0 ) = 8 θ 1 . (29)
Equations (28) –(29) , with negative degrees and unstable factors set to zero, determine all amplitudes by increasing h h h .
By (14) ,
1 = θ 0 ≤ θ 1 ≤ ⋯ ≤ C θ (30) 1=\theta_0\leq\theta_1\leq\cdots\leq C_\theta \tag*{(30)} 1 = θ 0 ≤ θ 1 ≤ ⋯ ≤ C θ (30)
uniformly in Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] . Thus every term of (28) is nonnegative, and the continuous endpoint (22) defines a positive limiting theory.
The harmonic model
We now introduce the explicit model with which the amplitudes will be compared. It depends on the degrees only through the deficit m m m , and it is chosen so that the linear part of (28) reproduces it exactly on average; this is the mass identity (32) below. In this sense the model is harmonic for the linear part of the recursion, which explains its name. At the endpoint Y = y 0 Y=y_0 Y = y 0 the model is the denominator in the normalization (23) of pure ψ \psi ψ intersection numbers (see (99) ), so Proposition 5.1 extends the coarse bounds on pure ψ \psi ψ intersections to the whole parameter interval.
Suppress Y Y Y temporarily and put
A m ( h ) = [ z m ] H ‾ ( z ) − h , M g , n ( d ) = 3 h Γ ( h ) 4 − g A m ( h ) , R = U / M . (31) A_m(h)=[z^m]\overline{H}(z)^{-h},\qquad\mathcal{M}_{g,n}(\mathbf d)=3^h\Gamma(h)4^{-g}A_m(h),\qquad R=\mathcal{U}/\mathcal{M}. \tag*{(31)} A m ( h ) = [ z m ] H ( z ) − h , M g , n ( d ) = 3 h Γ ( h ) 4 − g A m ( h ) , R = U / M . (31)
All A m ( h ) A_m(h) A m ( h ) are positive. Define
q ℓ = θ ℓ A m − ℓ ( h − 1 ) A m ( h ) , S 0 = ∑ ℓ = 0 m q ℓ , L = 2 D + n = 3 h − 2 m . q_\ell=\theta_\ell\frac{A_{m-\ell}(h-1)}{A_m(h)},\qquad S_0=\sum_{\ell=0}^{m}q_\ell,\qquad L=2D+n=3h-2m. q ℓ = θ ℓ A m ( h ) A m − ℓ ( h − 1 ) , S 0 = ℓ = 0 ∑ m q ℓ , L = 2 D + n = 3 h − 2 m .
For h ≥ 2 h \ge2 h ≥ 2 , the identity K ˉ = H ˉ + ( 2 / 3 ) z H ˉ ′ \bar{K}=\bar{H}+(2/3)z\bar{H}' K ˉ = H ˉ + ( 2/3 ) z H ˉ ′ gives exactly
∑ ℓ = 0 m q ℓ ( 1 − 2 ( m − ℓ ) 3 ( h − 1 ) ) = 1. (32) \sum_{\ell=0}^{m}q_{\ell}\left(1-\frac{2(m-\ell)}{3(h-1)}\right)=1. \tag*{(32)} ℓ = 0 ∑ m q ℓ ( 1 − 3 ( h − 1 ) 2 ( m − ℓ ) ) = 1. (32)
Indeed, the generating series of the left side, before division by A m ( h ) A_m(h) A m ( h ) , is
K ˉ − 1 ( H ˉ − ( h − 1 ) − 2 z 3 ( h − 1 ) ( H ˉ − ( h − 1 ) ) ′ ) = K ˉ − 1 K ˉ H ˉ − h = H ˉ − h . \bar{K}^{-1}\left(\bar{H}^{-(h-1)}-\frac{2z}{3(h-1)}\left(\bar{H}^{-(h-1)}\right)'\right)=\bar{K}^{-1}\bar{K}\bar{H}^{-h}=\bar{H}^{-h}. K ˉ − 1 ( H ˉ − ( h − 1 ) − 3 ( h − 1 ) 2 z ( H ˉ − ( h − 1 ) ) ′ ) = K ˉ − 1 K ˉ H ˉ − h = H ˉ − h .
Dividing (28) by the model turns its linear terms into a weighted sum of the ratios R R R at the children. We check that these weights have total mass one when the pivot degree d 1 d_1 d 1 is positive; this is the meaning of (32) . Assume g ≥ 2 g \ge2 g ≥ 2 ; states of genus at most one are terminal in the argument of § 3.7 . A merging child has type ( g , n − 1 ) (g,n-1) ( g , n − 1 ) and a nonseparating child type ( g − 1 , n + 1 ) (g-1,n+1) ( g − 1 , n + 1 ) ; both have complexity h − 1 h-1 h − 1 and deficit m − ℓ m-\ell m − ℓ . Hence the model ratio of a merging child to its parent is A m − ℓ ( h − 1 ) / ( 3 ( h − 1 ) A m ( h ) ) A_{m-\ell}(h-1)/(3(h-1)A_m(h)) A m − ℓ ( h − 1 ) / ( 3 ( h − 1 ) A m ( h )) , and that of a nonseparating child is four times larger, because of the factor 4 g 4^g 4 g in (31) . The merging children carry the coefficients ( 2 d j + 1 ) θ ℓ (2d_j+1)\theta_\ell ( 2 d j + 1 ) θ ℓ , j ≥ 2 j \ge2 j ≥ 2 , and there are d 1 + ℓ − 1 d_1+\ell-1 d 1 + ℓ − 1 ordered nonseparating pairs, each with coefficient θ ℓ / 2 \theta_\ell/2 θ ℓ /2 . Since 2 D + n = 3 h − 2 m 2D+n=3h-2m 2 D + n = 3 h − 2 m , the total weight is
∑ ℓ = 0 m q ℓ ∑ j ≥ 2 ( 2 d j + 1 ) + 2 ( d 1 + ℓ − 1 ) 3 ( h − 1 ) = ∑ ℓ = 0 m q ℓ 3 ( h − 1 ) − 2 ( m − ℓ ) 3 ( h − 1 ) = 1. \sum_{\ell=0}^{m}q_{\ell}\frac{\sum_{j\ge2}(2d_j+1)+2(d_1+\ell-1)}{3(h-1)}
=\sum_{\ell=0}^{m}q_{\ell}\frac{3(h-1)-2(m-\ell)}{3(h-1)}=1. ℓ = 0 ∑ m q ℓ 3 ( h − 1 ) ∑ j ≥ 2 ( 2 d j + 1 ) + 2 ( d 1 + ℓ − 1 ) = ℓ = 0 ∑ m q ℓ 3 ( h − 1 ) 3 ( h − 1 ) − 2 ( m − ℓ ) = 1.
by (32) . Thus, for a positive pivot, the linear model weights form a probability kernel P P P , and positivity of the separating terms gives R ≥ P R R \ge PR R ≥ P R .
At an all-zero state, where every degree and hence the pivot vanishes, the formal ℓ = 0 \ell=0 ℓ = 0 count includes one nonseparating pair with multiplicity − 1 -1 − 1 , and also includes n − 1 n-1 n − 1 invalid mergers. The true linear mass is instead
s = 1 + 3 − n 3 ( h − 1 ) A m ( h − 1 ) A m ( h ) . (33) s=1+\frac{3-n}{3(h-1)}\frac{A_m(h-1)}{A_m(h)}. \tag*{(33)} s = 1 + 3 ( h − 1 ) 3 − n A m ( h ) A m ( h − 1 ) . (33)
Since H ˉ ( 0 ) = 1 \bar{H}(0)=1 H ˉ ( 0 ) = 1 and H ˉ − 1 \bar{H}^{-1} H ˉ − 1 has positive coefficients, A m ( h − 1 ) ≤ A m ( h ) A_m(h-1)\le A_m(h) A m ( h − 1 ) ≤ A m ( h ) . In the range g ≥ 2 g \ge2 g ≥ 2 , therefore,
s ≥ 2 3 , s ≤ 1 ( n ≥ 3 ) , s ≤ 1 + O ( h − 1 ) ( n = 1 , 2 ) . (34) s\ge\frac{2}{3},\qquad s\le1\ (n\ge3),\qquad s\le1+O(h^{-1})\ (n=1,2). \tag*{(34)} s ≥ 3 2 , s ≤ 1 ( n ≥ 3 ) , s ≤ 1 + O ( h − 1 ) ( n = 1 , 2 ) . (34)
Thus for a zero pivot the linear terms need not have mass exactly one.
We now record the analytic information needed below. The function
P Y ( w ) = H ˉ Y ( 1 ) H ˉ Y ( w ) (35) P_Y(w)=\frac{\bar{H}_Y(1)}{\bar{H}_Y(w)} \tag*{(35)} P Y ( w ) = H ˉ Y ( w ) H ˉ Y ( 1 ) (35)
is the probability generating function of (15) ; its mean 3 / 2 3/2 3/2 , its analyticity on a disk of radius 1 + η > 1 1+\eta>1 1 + η > 1 , the positive lower bounds on its first two coefficients and the two-sided bounds on its variance, all uniform in Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] , were established in Subsection 2.2. Let S h S_h S h be the sum of h h h independent variables with probability generating function (35) . The uniform local limit theorem (16) gives
Pr ( S h = m ) ≥ c B h − 1 / 2 when ∣ m − 3 h / 2 ∣ ≤ B h , (36) \Pr(S_h=m)\ge c_Bh^{-1/2}\quad\text{when }|m-3h/2|\le B\sqrt{h}, \tag*{(36)} Pr ( S h = m ) ≥ c B h − 1/2 when ∣ m − 3 h /2∣ ≤ B h , (36)
and at bounded displacement, where the odd cubic term integrates to zero,
Pr ( S h = 3 h / 2 + d ) = e − d 2 / ( 2 h σ Y 2 ) 2 π h σ Y 2 ( 1 + O B ( h − 1 ) ) , ∣ d ∣ ≤ B , (37) \Pr(S_h=3h/2+d)=\frac{e^{-d^2/(2h\sigma_Y^2)}}{\sqrt{2\pi h\sigma_Y^2}}\left(1+O_B(h^{-1})\right),\qquad|d|\le B, \tag*{(37)} Pr ( S h = 3 h /2 + d ) = 2 π h σ Y 2 e − d 2 / ( 2 h σ Y 2 ) ( 1 + O B ( h − 1 ) ) , ∣ d ∣ ≤ B , (37)
at integer arguments. The same Fourier argument applies after tilting at a positive radius ρ \rho ρ in a compact subset of the analytic interval; here ρ \rho ρ is a radius in the normalized z z z -coordinate. If μ Y ( ρ ) = − ρ H ˉ Y ′ ( ρ ) / H ˉ Y ( ρ ) = m / h \mu_Y(\rho)=-\rho\bar{H}_Y'(\rho)/\bar{H}_Y(\rho)=m/h μ Y ( ρ ) = − ρ H ˉ Y ′ ( ρ ) / H ˉ Y ( ρ ) = m / h , it gives
A m ( h ) = H ˉ Y ( ρ ) − h ρ − m 2 π h ρ μ Y ′ ( ρ ) ( 1 + O ( h − 1 ) ) (38) A_m(h)=\frac{\bar{H}_Y(\rho)^{-h}\rho^{-m}}{\sqrt{2\pi h\rho\mu_Y'(\rho)}}(1+O(h^{-1})) \tag*{(38)} A m ( h ) = 2 π h ρ μ Y ′ ( ρ ) H ˉ Y ( ρ ) − h ρ − m ( 1 + O ( h − 1 )) (38)
uniformly on such compact tilted families. Near the mean, the tilted rate function has bounded second derivative and vanishing value and first derivative; hence, for a fixed sufficiently small c 0 > 0 c_0>0 c 0 > 0 ,
Pr ( S h = 3 h / 2 − u ) ≥ c h − 1 / 2 e − C u 2 / h , 0 ≤ u ≤ c 0 h . (39) \Pr(S_h=3h/2-u)\ge ch^{-1/2}e^{-Cu^2/h},\qquad0\le u\le c_0h. \tag*{(39)} Pr ( S h = 3 h /2 − u ) ≥ c h − 1/2 e − C u 2 / h , 0 ≤ u ≤ c 0 h . (39)
The coefficient/probability relation is
A m ( h ) = H ˉ ( 1 ) − h Pr ( S h = m ) . (40) A_m(h)=\bar{H}(1)^{-h}\Pr(S_h=m). \tag*{(40)} A m ( h ) = H ˉ ( 1 ) − h Pr ( S h = m ) . (40)
Consequently, when L ≤ h L\le\sqrt{h} L ≤ h , (30) , (36) , and the sum of the numerator probabilities give S 0 ≤ C h S_0\le C\sqrt{h} S 0 ≤ C h . When L ≥ 6 L\ge6 L ≥ 6 , the minimum factor in (32) is ( L − 3 ) / ( 3 ( h − 1 ) ) (L-3)/(3(h-1)) ( L − 3 ) / ( 3 ( h − 1 )) , so S 0 ≤ C h / L S_0\le Ch/L S 0 ≤ C h / L . Combining the two estimates,
S 0 ≤ C h , S 0 ≤ C h / L ( L ≥ 6 ) . (41) S_0\le C\sqrt{h},\qquad S_0\le Ch/L\quad(L\ge6). \tag*{(41)} S 0 ≤ C h , S 0 ≤ C h / L ( L ≥ 6 ) . (41)
All constants here are uniform in Y Y Y , and the finitely many states of bounded complexity are absorbed into them. The mass identity (32) , the zero-pivot correction (33) –(34) and the bound (41) on S 0 S_0 S 0 are the inputs of the random-walk argument in Section 5; the central estimate (37) is used again in (102) to normalize the leading term.
Comparison with the harmonic model
This section completes the first idea of Subsection 1.5 by proving Proposition 5.1, the coarse two-sided comparison between the normalized amplitudes of (28) and the harmonic model (31) on which all later estimates rest. It loses a factor e O ( h ) e^{O(\sqrt{h})} e O ( h ) but holds uniformly in the degrees and in the density parameter, including the pure-ψ \psi ψ endpoint, and it uses no information about volumes: its inputs are the positive recursion, the model estimates (32) –(41) of Section 4 and the coarse bounds on pure ψ \psi ψ intersections from Appendix C. The proof is given in full here. Section 6 upgrades the comparison to relative errors for a fixed number of markings, and Section 13 uses its Y = 0 Y=0 Y = 0 case, recorded in (63) , for the coefficient ratios.
Proposition 5.1 (Uniform harmonic comparison). For every fixed n ≥ 1 n\ge1 n ≥ 1 , there is a constant C n C_n C n such that
− C n log 2 ( h + 1 ) ≤ log U g , n ( d ; Y ) h 3 Γ ( h ) 4 − g [ z m ] H ˉ Y ( z ) − h ≤ C n h -C_n\log^2(h+1)\le\log\frac{U_{g,n}(d;Y)}{h^3\Gamma(h)4^{-g}[z^m]\bar{H}_Y(z)^{-h}}\le C_n\sqrt{h} − C n log 2 ( h + 1 ) ≤ log h 3 Γ ( h ) 4 − g [ z m ] H ˉ Y ( z ) − h U g , n ( d ; Y ) ≤ C n h for every stable pair ( g , n ) (g,n) ( g , n ) , every d ∈ Z ≥ 0 n d\in\mathbb{Z}_{\ge0}^n d ∈ Z ≥ 0 n with ∣ d ∣ ≤ 3 g − 3 + n |d|\le3g-3+n ∣ d ∣ ≤ 3 g − 3 + n , and every Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] . Here h = 2 g − 2 + n h=2g-2+n h = 2 g − 2 + n and m = 3 g − 3 + n − ∣ d ∣ m=3g-3+n-|d| m = 3 g − 3 + n − ∣ d ∣ .
The idea is to read the recursion as a random walk. Write R = U / M R=U/M R = U / M for the ratio of a true amplitude to the model amplitude (31) , and divide the recursion (28) by the model. Because the model satisfies the linear part of the recursion with total mass one at a positive pivot (identity (32) ), the linear terms there express R R R as an average of R R R over the states of complexity h − 1 h-1 h − 1 reached in one step, and the separating terms add a nonnegative amount. At all-zero states the linear mass need not equal one, and it is carried along in the averaging. Thus R R R is bounded below by a terminal expectation weighted by the accumulated linear masses, and it is bounded above as soon as the separating contribution is small.
Comparing intersection recursions with random walks has an important precedent in Aggarwal’s proof of uniform asymptotics for pure ψ \psi ψ intersections 6 . Here, in addition, the recursion and its comparison model vary with Y Y Y , and the estimates must remain uniform as this parameter approaches the pure-ψ \psi ψ endpoint.
Throughout this section a state is a stable pair ( g , n ) (g,n) ( g , n ) , n ≥ 1 n \ge1 n ≥ 1 , with a degree tuple d ∈ Z ≥ 0 n \mathbf{d} \in\mathbb{Z}_{\ge0}^{n} d ∈ Z ≥ 0 n , ∣ d ∣ ≤ 3 g − 3 + n |\mathbf{d}| \le3g-3+n ∣ d ∣ ≤ 3 g − 3 + n , and a parameter Y ∈ [ 0 , y 0 ] Y \in[0,y_0] Y ∈ [ 0 , y 0 ] ; constants are uniform in Y Y Y and in the state unless stated otherwise. We write Z 0 Z_0 Z 0 for the number of entries of d \mathbf{d} d equal to zero. A state is terminal if m = 0 m=0 m = 0 or g ≤ 1 g \le1 g ≤ 1 ; the random walks below stop there. The proof has three steps.
Lemma 5.2 bounds R R R at terminal states, using only the string and dilaton equations and the coarse bounds of Appendix § 8.4 .
Lemma 5.3 proves the lower bound. The only defect of the averaging is at entries of degree zero, where some linear terms are missing; we show that the random walk does not create many such entries.
Lemmas 5.4–5.6 prove the upper bound. We divide R R R by a weight that penalizes entries of degree zero and one, show that the weighted separating terms have relative size O ( h − 1 / 2 ) O(h^{-1/2}) O ( h − 1/2 ) , and conclude by examining, at each complexity, the state at which the weighted ratio is largest.
Bounds at terminal states
Lemma 5.2 (Terminal bounds). There are constants c ψ c_\psi c ψ , C ψ C_\psi C ψ , c c c , C > 0 C>0 C > 0 with the following properties.
(i) For every pure ψ \psi ψ tuple e \mathbf{e} e with Z 0 Z_0 Z 0 zero entries,
c ψ ( 2 / 3 ) Z 0 ≤ C g ( e ) ≤ C ψ 2 Z 0 , c_\psi(2/3)^{Z_0} \le C_g(\mathbf{e}) \le C_\psi2^{Z_0}, c ψ ( 2/3 ) Z 0 ≤ C g ( e ) ≤ C ψ 2 Z 0 , in the normalization (23) .
(ii) At every terminal state, R ≥ c e − C Z 0 R \ge c e^{-CZ_0} R ≥ c e − C Z 0 .
(iii) At every state, R ≤ e C h R \le e^{Ch} R ≤ e C h .
Proof. For (i), the standard normalized string and dilaton identities imply
c ψ ( 2 / 3 ) Z 0 ≤ C g ( e ) ≤ C ψ 2 Z 0 , (42) c_\psi(2/3)^{Z_0} \le C_g(\mathbf{e}) \le C_\psi2^{Z_0}, \tag*{(42)} c ψ ( 2/3 ) Z 0 ≤ C g ( e ) ≤ C ψ 2 Z 0 , (42)
where Z 0 Z_0 Z 0 is the number of zero entries. To prove this, first remove the entries equal to one by the normalized dilaton equation, retaining the final genus-one tuple ( 1 ) (1) ( 1 ) . At a noninitial tuple with a zero, let p + p_+ p + denote the number of positive entries; normalized string has total positive coefficient
2 ( 3 g − 3 + q ) + p + 3 ( 2 g − 3 + q ) \frac{2(3g-3+q)+p_+}{3(2g-3+q)} 3 ( 2 g − 3 + q ) 2 ( 3 g − 3 + q ) + p +
between 2 / 3 2/3 2/3 and 2 2 2 . Once the ones have been removed, each string child has one fewer zero; any newly created one can again be removed. With no zeros use the primitive bounds stated after (23) ; the remaining bases are C 0 ( 0 , 0 , 0 ) = 1 / 3 C_0(0,0,0)=1/3 C 0 ( 0 , 0 , 0 ) = 1/3 and C 1 ( 1 ) = 1 / 6 C_1(1)=1/6 C 1 ( 1 ) = 1/6 . Induction proves (42) , which is (i).
For (ii) we use a lower comparison that requires no asymptotic information on volumes. Set all kernel coefficients in (28) equal to one. The corresponding background of KW times has t 1 = 0 , t 2 = 3 , t k = 0 t_1=0,t_2=3,t_k=0 t 1 = 0 , t 2 = 3 , t k = 0 for k ≥ 3 k \ge3 k ≥ 3 , so its amplitudes are
U g , n P I ( d ) = 3 m m ! ∏ i ( 2 d i + 1 ) ! ! ⟨ τ d τ 2 m ⟩ g . (43) \mathcal{U}^{\mathrm{PI}}_{g,n}(\mathbf{d})=\frac{3^m}{m!}\prod_i(2d_i+1)!!\langle\tau_{d}\tau_2^m\rangle_g. \tag*{(43)} U g , n PI ( d ) = m ! 3 m i ∏ ( 2 d i + 1 )!! ⟨ τ d τ 2 m ⟩ g . (43)
They have the same initial ( 0 , 3 ) (0,3) ( 0 , 3 ) and degree-one ( 1 , 1 ) (1,1) ( 1 , 1 ) values, and their degree-zero ( 1 , 1 ) (1,1) ( 1 , 1 ) value 1 / 8 1/8 1/8 is at most U 1 , 1 ( 0 ) = θ 1 / 8 \mathcal{U}_{1,1}(0)=\theta_1/8 U 1 , 1 ( 0 ) = θ 1 /8 . By (30) , induction in the positive recursion gives U ≥ U P I \mathcal{U}\ge\mathcal{U}^{\mathrm{PI}} U ≥ U PI .
Uniform analyticity of H ˉ Y − 1 \bar{H}_Y^{-1} H ˉ Y − 1 at zero gives a fixed B ≥ 1 B \ge1 B ≥ 1 such that [ z j ] H ˉ Y − 1 ≤ B j [z^j]\bar{H}_Y^{-1}\le B^j [ z j ] H ˉ Y − 1 ≤ B j . Thus
A m ( h ) ≤ B m ( h ) m m ! . A_m(h)\le\frac{B^m(h)_m}{m!}. A m ( h ) ≤ m ! B m ( h ) m .
Dividing (43) by (31) and using (23) gives
R ≥ C g ( d , 2 m ) ( 3 5 B ) m . (44) R\ge C_g(\mathbf{d},2^m)\left(\frac{3}{5B}\right)^m. \tag*{(44)} R ≥ C g ( d , 2 m ) ( 5 B 3 ) m . (44)
If g ≤ 1 g\le1 g ≤ 1 , dimension and D ≥ n − Z 0 D\ge n-Z_0 D ≥ n − Z 0 imply m ≤ Z 0 m\le Z_0 m ≤ Z 0 . Hence R ≥ c e − C Z 0 R\ge ce^{-CZ_0} R ≥ c e − C Z 0 in this terminal range. At m = 0 m=0 m = 0 , exactly R = C g ( d ) R=C_g(\mathbf{d}) R = C g ( d ) , so (5.1) applies directly. This proves (ii).
Finally we prove the global bound (iii),
R ≤ e C h . (45) R \le e^{Ch}. \tag*{(45)} R ≤ e C h . (45)
Note that in the normalized finite sum (25) –(26) a background insertion of degree k k k has weight
v k ( Y ) = − r Y k − 1 X ( k ) ( Y ) / a . v_k(\mathrm{Y})=-r_{\mathrm{Y}}^{k-1}X^{(k)}(\mathrm{Y})/a. v k ( Y ) = − r Y k − 1 X ( k ) ( Y ) / a .
The corresponding absolute majorant is
B Y ( z ) = ∑ k ≥ 2 3 ∣ v k ( Y ) ∣ ( 2 k + 1 ) ! ! z k − 1 . B_{\mathrm{Y}}(z)=\sum_{k\ge2}\frac{3|v_k(\mathrm{Y})|}{(2k+1)!!}z^{k-1}. B Y ( z ) = k ≥ 2 ∑ ( 2 k + 1 )!! 3∣ v k ( Y ) ∣ z k − 1 .
This family has a uniformly bounded analytic germ and vanishes at zero. Indeed it is the coefficientwise absolute version of 1 − H ˉ Y 1-\bar{H}_{\mathrm{Y}} 1 − H ˉ Y ; the normalized family in (21) –(22) is compact and analytic on a common disk. Thus a fixed z ∗ > 0 z_* > 0 z ∗ > 0 makes B Y ( z ∗ ) ≤ 1 / 2 B_{\mathrm{Y}}(z_*)\le1/2 B Y ( z ∗ ) ≤ 1/2 . Using the upper bound in (42) on every primitive term of (25) gives
U ≤ C 2 Z 0 3 h Γ ( h ) 4 − g [ z m ] ( 1 − B Y ( z ) ) − h ≤ C 2 Z 0 3 h Γ ( h ) 4 − g z ∗ − m 2 h . U\le C2^{Z_0}3^h\Gamma(h)4^{-g}[z^m](1-B_{\mathrm{Y}}(z))^{-h}\le C2^{Z_0}3^h\Gamma(h)4^{-g}z_*^{-m}2^h. U ≤ C 2 Z 0 3 h Γ ( h ) 4 − g [ z m ] ( 1 − B Y ( z ) ) − h ≤ C 2 Z 0 3 h Γ ( h ) 4 − g z ∗ − m 2 h .
The linear coefficient of − log H ˉ Y -\log\bar{H}_{\mathrm{Y}} − log H ˉ Y is uniformly positive. Therefore
A m ( h ) ≥ ( c h ) m / m ! ≥ e − C ′ h , A_m(h)\ge(ch)^m/m!\ge e^{-C'h}, A m ( h ) ≥ ( c h ) m / m ! ≥ e − C ′ h ,
because m ≤ 3 h / 2 m\le3h/2 m ≤ 3 h /2 , and A 0 ( h ) = 1 A_0(h)=1 A 0 ( h ) = 1 since H ˉ Y ( 0 ) = 1 \bar{H}_{\mathrm{Y}}(0)=1 H ˉ Y ( 0 ) = 1 . Since Z 0 ≤ n ≤ h + 2 Z_0\le n\le h+2 Z 0 ≤ n ≤ h + 2 , division proves (45) . □
The lower bound: control of zero degrees
Lemma 5.3 (Lower bound). There are constants c , C > 0 c,C>0 c , C > 0 such that, at every state,
R ≥ c exp { − C ( Z 0 + log ( h + 1 ) ) log ( h + 1 ) } . R\ge c\exp\{-C(Z_0+\log(h+1))\log(h+1)\}. R ≥ c exp { − C ( Z 0 + log ( h + 1 )) log ( h + 1 )} .
Proof. Since the separating terms of (28) are nonnegative, R R R is at least the total linear mass times the average of R R R over its linear children under the normalized transition probabilities. Iterating gives a lower bound by the terminal value weighted by the product of these masses along a walk that lowers complexity by one at each step. Two losses occur: the terminal bounds (42) and (44) lose a factor exponential in the number of zero entries, and the averaging may have mass s < 1 s<1 s < 1 at states with all degrees zero, again with a loss controlled by the number of zeros. It therefore suffices to show that the walk rarely creates many zero entries.
At a nonterminal state choose any positive pivot if one exists; otherwise choose a zero. Stop only at m = 0 m=0 m = 0 or g ≤ 1 g\le1 g ≤ 1 . Use the linear model weights, divided by their mass s s s , as transition probabilities. For a positive pivot s = 1 s=1 s = 1 ; at an all-zero state use (33) –(34) .
Let Z Z Z be the current zero count. At a positive pivot, a merger with a marking of degree zero lowers Z Z Z by one, except when the pivot has degree one and ℓ = 0 \ell=0 ℓ = 0 ; that merger leaves Z Z Z unchanged. A nonseparating ordered pair can increase Z Z Z by at most two, and at each ℓ \ell ℓ there are at most two pairs containing a zero. Each pair has twice the model weight of a merger with one zero marking. Thus
p b i r t h ≤ 4 p d e a t h / Z ( Z ≥ 1 ) . (46) p_{\mathrm{birth}}\le4p_{\mathrm{death}}/Z\qquad(Z\ge1). \tag*{(46)} p birth ≤ 4 p death / Z ( Z ≥ 1 ) . (46)
That exceptional term contributes neither a birth nor a death. At an all-zero state, a birth occurs only at ℓ = 2 \ell=2 ℓ = 2 , raises the count by one, and its weight is at most 2 / ( Z − 1 ) 2/(Z-1) 2/ ( Z − 1 ) times the mass of mergers with a zero marking. Those mergers have ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 and lower the count by at least one. Hence (46) still holds for Z ≥ 2 Z\ge2 Z ≥ 2 . Normalization by s s s preserves these comparisons.
For Z ≥ 24 Z\ge24 Z ≥ 24 , the conditional drift obeys
E ( 2 Z n e x t ) 2 Z − 1 ≤ 3 p b i r t h − 1 2 p d e a t h ≤ 0. \frac{\mathbb{E}(2^{Z_{\mathrm{next}}})}{2^Z}-1\le3p_{\mathrm{birth}}-\frac{1}{2}p_{\mathrm{death}}\le0. 2 Z E ( 2 Z next ) − 1 ≤ 3 p birth − 2 1 p death ≤ 0.
During each excursion above 24, optional stopping bounds the probability of reaching M M M by 2 Z s t a r t − M 2^{Z_{\mathrm{start}}-M} 2 Z start − M . An excursion entered from below starts at a value at most 26. Since complexity decreases by one, there are at most h i n i t i a l + 1 h_{\mathrm{initial}}+1 h initial + 1 excursions. A union bound therefore shows that, with probability at least 1 / 2 1/2 1/2 ,
max Z ≤ Z i n i t i a l + C + C log ( h i n i t i a l + 1 ) . (47) \max Z \le Z_{\mathrm{initial}}+C+C\log(h_{\mathrm{initial}}+1). \tag*{(47)} max Z ≤ Z initial + C + C log ( h initial + 1 ) . (47)
All stopping times are bounded, so optional stopping applies without further integrability conditions.
Only all-zero steps can have s < 1 s<1 s < 1 . By (33) –(34) ,
− log min ( s , 1 ) ≤ C Z / h . -\log\min(s,1)\le CZ/h. − log min ( s , 1 ) ≤ C Z / h .
On the event (47) , write its upper bound as M M M . Summing over the decreasing complexities costs at most
C M + C M ∑ j = M + 1 h i n i t i a l 1 j ≤ C M log ( h i n i t i a l + 1 ) . CM+CM\sum_{j=M+1}^{h_{\mathrm{initial}}}\frac{1}{j}\le CM\log(h_{\mathrm{initial}}+1). C M + C M j = M + 1 ∑ h initial j 1 ≤ C M log ( h initial + 1 ) .
Iterated positivity gives
R i n i t i a l ≥ E [ R t e r m i n a l ∏ s t e p s s ] . R_{\mathrm{initial}}\ge\mathbb{E}\left[R_{\mathrm{terminal}}\prod_{\mathrm{steps}}s\right]. R initial ≥ E [ R terminal steps ∏ s ] .
Here the initial zero count is Z 0 Z_0 Z 0 . The terminal bounds of Lemma 5.2(ii), applied at terminal zero counts at most M M M , and the probability in (47) now prove
R ≥ c exp { − C ( Z 0 + log ( h + 1 ) ) log ( h + 1 ) } . (48) R\ge c\exp\{-C(Z_0+\log(h+1))\log(h+1)\}. \tag*{(48)} R ≥ c exp { − C ( Z 0 + log ( h + 1 )) log ( h + 1 )} . (48)
No asymptotics for volumes enter at the terminal states. □
The upper bound: a weight for degrees zero and one
The upper bound requires the separating terms to be small, and for general states they are not: a state with many entries of degree zero or one can distribute these markings between the two components in many ways. We therefore divide R R R by a weight that grows with the number of such entries, chosen so that the linear terms still average R / W R/W R / W up to a factor 1 + O ( h − 1 / 2 ) 1+O(h^{-1/2}) 1 + O ( h − 1/2 ) (this is (50) ), while the weighted separating terms become small (Subsection 5.4).
Call degrees zero or one inactive. Let ι \iota ι be their number, Z 0 Z_0 Z 0 the number of zeros, and R a c t R_{\mathrm{act}} R act the number of degrees at least two. Fix a sufficiently small constant c ∗ > 0 c_* > 0 c ∗ > 0 , and set
ϕ ( ι ) = ∏ j = 1 ι max ( 1 , c ∗ j ) , W ( d ) = 4 Z 0 ϕ ( ι ) . (49) \phi(\iota)=\prod_{j=1}^{\iota}\max(1,c_*\sqrt{j}),\qquad W(\mathbf{d})=4^{Z_0}\phi(\iota). \tag*{(49)} ϕ ( ι ) = j = 1 ∏ ι max ( 1 , c ∗ j ) , W ( d ) = 4 Z 0 ϕ ( ι ) . (49)
We call the truncation max ( 1 , ⋅ ) \max(1,\cdot) max ( 1 , ⋅ ) clipping. Since every factor is then at least one, ϕ \phi ϕ is nondecreasing; this monotonicity is essential when an inactive pivot splits into two active degrees. Up to fixed multiplicative constants, ϕ ( ι ) \phi(\iota) ϕ ( ι ) equals c ∗ ι ι ! c_*^\iota\sqrt{\iota!} c ∗ ι ι ! .
In the upper bound we choose the pivot as follows: an active entry when one exists, an entry equal to one if no active entry exists, and a zero only at an all-zero state. Again the walk stops only at terminal states. The linear model weights p c h i l d p_{\mathrm{child}} p child of a state are the weights that appear when (28) is divided by the model; their mass is given by (32) –(34) , and may exceed one at an all-zero state with one or two markings.
Lemma 5.4 (Weighted linear terms). If c ∗ c_* c ∗ is sufficiently small, then at every nonterminal state
∑ l i n e a r p c h i l d W c h i l d W ≤ 1 + C S 0 h ≤ 1 + C h − 1 / 2 . (50) \sum_{\mathrm{linear}}p_{\mathrm{child}}\frac{W_{\mathrm{child}}}{W}\le1+\frac{CS_0}{h}\le1+Ch^{-1/2}. \tag*{(50)} linear ∑ p child W W child ≤ 1 + h C S 0 ≤ 1 + C h − 1/2 . (50)
Proof. The second inequality is (41) . We consider the three possible pivots in turn.
For an active pivot, an ordinary inactive merger reduces ι \iota ι by at least one and does not increase Z 0 Z_0 Z 0 . For all sufficiently large ι \iota ι , its weight ratio is at most 1 / 2 1/2 1/2 . The exception is a pivot of degree two merged with a zero at ι = 0 \iota=0 ι = 0 : it turns that zero into a one, leaves ι \iota ι fixed, and divides the weight by four. The favorable mass is therefore at least a fixed positive multiple of ι S 0 / h \iota S_0/h ι S 0 / h . At each ℓ \ell ℓ , at most four nonseparating pairs contain an inactive degree. A pair creating two inactive degrees has weight ratio, for large ℓ \ell ℓ , at most
16 c ∗ 2 ( ℓ + 1 ) ( ℓ + 2 ) . 16c_*^2\sqrt{(\ell+1)(\ell+2)}. 16 c ∗ 2 ( ℓ + 1 ) ( ℓ + 2 ) .
Its total model mass per ℓ \ell ℓ is bounded by a fixed multiple of q ℓ / h q\ell/h q ℓ / h . Choose c ∗ c_* c ∗ so small that the coefficient of ℓ \ell ℓ in this upper bound is less than half the favorable merger coefficient. Pairs creating only one inactive degree cost O ( ℓ ) O(\sqrt{\ell}) O ( ℓ ) , and all other pairs leave the weight unchanged. The remaining expression is bounded above by C S 0 / h CS_0/h C S 0 / h . Bounded ℓ \ell ℓ , including the clipping threshold, also costs only C S 0 / h CS_0/h C S 0 / h , because there are at most four exceptional pairs and boundedly many inactive mergers.
If the pivot has degree one and all markings are inactive, every merger decreases ℓ \ell ℓ , while a nonseparating step increases it by at most one. Only four pairs per ℓ \ell ℓ can increase the weight; a pair with two active new degrees decreases ℓ \ell ℓ and cannot increase the clipped weight. The same argument applies, with ℓ − 1 \ell-1 ℓ − 1 favorable merging labels.
At an all-zero state, all valid mergers have ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 and decrease both the zero count and the inactive weight. Increasing pairs have ℓ ≥ 2 \ell\ge2 ℓ ≥ 2 and increase ℓ \ell ℓ by at most one. Use S 0 − q 0 S_0-q_0 S 0 − q 0 in both birth and merger estimates. The unweighted mass correction (33) is nonpositive for n ≥ 3 n \ge3 n ≥ 3 and O ( h − 1 ) O(h^{-1}) O ( h − 1 ) otherwise. This proves (50) in every case. □
Bounding the separating terms
We now show that, after weighting, the separating terms of (28) have total relative size O ( h − 1 / 2 ) O(h^{-1/2}) O ( h − 1/2 ) , uniformly in the state. Splitting off a component of complexity h 1 ≤ h 2 h_1 \le h_2 h 1 ≤ h 2 costs the factor Γ ( h 1 ) Γ ( h 2 ) / Γ ( h ) \Gamma(h_1)\Gamma(h_2)/\Gamma(h) Γ ( h 1 ) Γ ( h 2 ) /Γ ( h ) , which is at most ( h − 3 h 1 − 1 ) − 1 \binom{h-3}{h_1-1}^{-1} ( h 1 − 1 h − 3 ) − 1 . The key point is that this factor beats the number of ways of distributing the marking labels once the weight has absorbed the inactive ones. For a separating term with node degrees u , v u,v u , v and kernel index ℓ \ell ℓ , write M 1 , M 2 \mathcal{M}_1,\mathcal{M}_2 M 1 , M 2 and W 1 , W 2 \mathcal{W}_1,\mathcal{W}_2 W 1 , W 2 for the model amplitudes and weights of its two factors.
Lemma 5.5 (Weighted separating terms). Let η = 1 / 12 \eta=1/12 η = 1/12 . At every nonterminal state,
∑ s e p a r a t i n g θ ℓ 2 M 1 M 2 M W 1 W 2 W e η min ( h 1 , h 2 ) ≤ C S 0 h − 3 / 2 ( 1 + ι ) ≤ C h − 1 / 2 , \sum_{\mathrm{separating}} \frac{\theta_\ell}{2}\frac{\mathcal{M}_1\mathcal{M}_2}{\mathcal{M}}\frac{\mathcal{W}_1\mathcal{W}_2}{\mathcal{W}}e^{\eta\min(h_1,h_2)}
\le CS_0h^{-3/2}(1+\iota)\le Ch^{-1/2}, separating ∑ 2 θ ℓ M M 1 M 2 W W 1 W 2 e η m i n ( h 1 , h 2 ) ≤ C S 0 h − 3/2 ( 1 + ι ) ≤ C h − 1/2 , where the sum runs over all separating terms of (28) .
Proof. The degree-dimension constraint supplies
4 R a c t + n ≤ 3 h , R a c t ≤ 3 h / 5. (51) 4R_{\mathrm{act}}+n\le3h,\qquad R_{\mathrm{act}}\le3h/5. \tag*{(51)} 4 R act + n ≤ 3 h , R act ≤ 3 h /5. (51)
Remove the pivot, leaving ι ∗ \iota_* ι ∗ inactive and R ∗ R_* R ∗ active labeled markings, with ι ∗ + R ∗ = n − 1 \iota_*+R_*=n-1 ι ∗ + R ∗ = n − 1 , ι ∗ ≤ ι \iota_*\le\iota ι ∗ ≤ ι , and R ∗ ≤ R a c t R_*\le R_{\mathrm{act}} R ∗ ≤ R act . In one separating component put u u u inactive and j j j active original labels, and write s = u + j s=u+j s = u + j . Each child receives one new marking. Comparing (49) with 4 Z 0 c ∗ ι ι ! 4Z_0c_*^\iota\sqrt{\iota!} 4 Z 0 c ∗ ι ι ! , uniformly over the two new degrees,
W 1 W 2 W ≤ c ( s + 2 ) h ( ι ∗ u ) . (52) \frac{\mathcal{W}_1\mathcal{W}_2}{\mathcal{W}}
\le c\frac{\sqrt{(s+2)h}}{\sqrt{\binom{\iota_*}{u}}}. \tag*{(52)} W W 1 W 2 ≤ c ( u ι ∗ ) ( s + 2 ) h . (52)
Indeed, when the ι ∗ \iota_* ι ∗ inactive original labels split as u , ι ∗ − u u,\iota_*-u u , ι ∗ − u , the factorial part of this estimate is bounded by
( u + 1 ) ! ( ι ∗ − u + 1 ) ! ι ∗ ! = ( u + 1 ) ( ι ∗ − u + 1 ) ( ι ∗ u ) . \frac{\sqrt{(u+1)!(\iota_*-u+1)!}}{\sqrt{\iota_*!}}
=
\frac{\sqrt{(u+1)(\iota_*-u+1)}}{\sqrt{\binom{\iota_*}{u}}}. ι ∗ ! ( u + 1 )! ( ι ∗ − u + 1 )! = ( u ι ∗ ) ( u + 1 ) ( ι ∗ − u + 1 ) .
Each child receives at most one inactive new marking. Since u ≤ s u\le s u ≤ s and ι ∗ ≤ h + 2 \iota_*\le h+2 ι ∗ ≤ h + 2 , the numerator is at most C ( s + 2 ) h C\sqrt{(s+2)h} C ( s + 2 ) h . The powers of c ∗ c_* c ∗ , the factor for zeros, and the clipping comparison change by bounded integer exponents. An inactive pivot supplies an additional factor ι − 1 / 2 \iota^{-1/2} ι − 1/2 , which was discarded. The exponential factors and clipping comparisons contribute only a fixed constant, since the total inactive and zero counts change by bounded amounts. Orient h 1 ≤ h 2 h_1 \le h_2 h 1 ≤ h 2 , and put
h ′ = h − 3 , k = h 1 − 1 ≤ h ′ / 2 , s = k + 2 − 2 g 1 . h' = h - 3,\qquad k = h_1 - 1 \le h'/2,\qquad s = k + 2 - 2g_1. h ′ = h − 3 , k = h 1 − 1 ≤ h ′ /2 , s = k + 2 − 2 g 1 .
Then
Γ ( h 1 ) Γ ( h 2 ) Γ ( h ) = 1 ( h − 1 ) ( h − 2 ) ( h ′ k ) (53) \frac{\Gamma(h_1)\Gamma(h_2)}{\Gamma(h)}
= \frac{1}{(h-1)(h-2)\binom{h'}{k}} \tag*{(53)} Γ ( h ) Γ ( h 1 ) Γ ( h 2 ) = ( h − 1 ) ( h − 2 ) ( k h ′ ) 1 (53)
For all sufficiently large h h h , (51) gives R ∗ ≤ 2 h ′ / 3 R_* \le2h'/3 R ∗ ≤ 2 h ′ /3 .
The required label sum is
Λ s = ∑ u + j = s ( ι ∗ u ) 1 / 2 ( R ∗ j ) . \Lambda_s=\sum_{u+j=s}\binom{\iota_*}{u}^{1/2}\binom{R_*}{j}. Λ s = u + j = s ∑ ( u ι ∗ ) 1/2 ( j R ∗ ) .
Cauchy–Schwarz and Vandermonde give
Λ s ≤ ( ι ∗ + R ∗ s ) 1 / 2 ( ∑ j ≤ s ( R ∗ j ) ) 1 / 2 (54) \Lambda_s \le\binom{\iota_*+R_*}{s}^{1/2}
\left(\sum_{j\le s}\binom{R_*}{j}\right)^{1/2} \tag*{(54)} Λ s ≤ ( s ι ∗ + R ∗ ) 1/2 ( j ≤ s ∑ ( j R ∗ ) ) 1/2 (54)
For k ≥ 1 k \ge1 k ≥ 1 , take ξ = k / ( h ′ − k ) ≤ 1 \xi= k/(h'-k) \le1 ξ = k / ( h ′ − k ) ≤ 1 . The two factors before square roots are bounded by ξ − s ( 1 + ξ ) ι ∗ + R ∗ \xi^{-s}(1+\xi)^{\iota_*+R_*} ξ − s ( 1 + ξ ) ι ∗ + R ∗ and ξ − s ( 1 + ξ ) R ∗ \xi^{-s}(1+\xi)^{R_*} ξ − s ( 1 + ξ ) R ∗ , respectively. The binomial mass at its mean gives
( h ′ k ) ≥ c ( k + 1 ) − 1 / 2 ξ − k ( 1 + ξ ) h ′ . \binom{h'}{k} \ge c(k+1)^{-1/2}\xi^{-k}(1+\xi)^{h'}. ( k h ′ ) ≥ c ( k + 1 ) − 1/2 ξ − k ( 1 + ξ ) h ′ .
Thus, if g 1 ≥ 1 g_1 \ge1 g 1 ≥ 1 , so s ≤ k s \le k s ≤ k ,
Λ s ( h ′ k ) ≤ C k + 1 ξ k − s ( 1 + ξ ) − h ′ / 6 ≤ C k + 1 e − k / 6 . (55) \frac{\Lambda_s}{\binom{h'}{k}}
\le C\sqrt{k+1}\,\xi^{k-s}(1+\xi)^{-h'/6}
\le C\sqrt{k+1}e^{-k/6}. \tag*{(55)} ( k h ′ ) Λ s ≤ C k + 1 ξ k − s ( 1 + ξ ) − h ′ /6 ≤ C k + 1 e − k /6 . (55)
Here ι ∗ + R ∗ = n − 1 ≤ h ′ \iota_*+R_* = n-1 \le h' ι ∗ + R ∗ = n − 1 ≤ h ′ , R ∗ ≤ 2 h ′ / 3 R_* \le2h'/3 R ∗ ≤ 2 h ′ /3 , and h ′ log ( 1 + ξ ) ≥ k h'\log(1+\xi) \ge k h ′ log ( 1 + ξ ) ≥ k . The case k = 0 k=0 k = 0 is immediate.
If g 1 = 0 g_1=0 g 1 = 0 , then s = k + 2 s=k+2 s = k + 2 . Validity of this component forces 2 j ≤ s − 2 = u + j − 2 2j \le s-2=u+j-2 2 j ≤ s − 2 = u + j − 2 , hence u ≥ j + 2 ≥ 2 u \ge j+2 \ge2 u ≥ j + 2 ≥ 2 . Use
( ι ∗ u ) = ι ∗ ( ι ∗ − 1 ) u ( u − 1 ) ( ι ∗ − 2 u − 2 ) . \binom{\iota_*}{u}
=\frac{\iota_*(\iota_*-1)}{u(u-1)}
\binom{\iota_*-2}{u-2}. ( u ι ∗ ) = u ( u − 1 ) ι ∗ ( ι ∗ − 1 ) ( u − 2 ι ∗ − 2 ) .
After taking the square root, discard the denominator and extract a factor ι ∗ \iota_* ι ∗ . The remaining sum has u − 2 + j = k u-2+j=k u − 2 + j = k , so (54) –(55) apply with ι ∗ − 2 \iota_*-2 ι ∗ − 2 in place of ι ∗ \iota_* ι ∗ . This use of the degree constraint is what replaces a factorial penalty on active labels.
Combining these estimates with (52) –(53) , and summing the polynomially many genus choices at fixed k k k , proves, for example with η = 1 / 12 \eta=1/12 η = 1/12 ,
∑ genus/label splits Γ ( h 1 ) Γ ( h 2 ) Γ ( h ) sup valid node degrees W 1 W 2 W e η min ( h 1 , h 2 ) ≤ C h − 3 / 2 ( 1 + ι ) . (56) \sum_{\text{genus/label splits}}
\frac{\Gamma(h_1)\Gamma(h_2)}{\Gamma(h)}
\sup_{\text{valid node degrees}}
\frac{\mathcal{W}_1\mathcal{W}_2}{\mathcal{W}}
e^{\eta\min(h_1,h_2)}
\le Ch^{-3/2}(1+\iota). \tag*{(56)} genus/label splits ∑ Γ ( h ) Γ ( h 1 ) Γ ( h 2 ) valid node degrees sup W W 1 W 2 e η m i n ( h 1 , h 2 ) ≤ C h − 3/2 ( 1 + ι ) . (56)
An empty supremum contributes zero. Thus the bound is uniform in the possible node degrees; their sum, with its model coefficients, is handled next.
For a separating recursion term, exact model normalization gives
θ ℓ 2 M 1 M 2 M = θ ℓ 6 Γ ( h 1 ) Γ ( h 2 ) Γ ( h ) A m 1 ( h 1 ) A m 2 ( h 2 ) A m ( h ) . (57) \frac{\theta_\ell}{2}\frac{\mathcal{M}_1\mathcal{M}_2}{\mathcal{M}}
=\frac{\theta_\ell}{6}\frac{\Gamma(h_1)\Gamma(h_2)}{\Gamma(h)}
\frac{A_{m_1}(h_1)A_{m_2}(h_2)}{A_m(h)}. \tag*{(57)} 2 θ ℓ M M 1 M 2 = 6 θ ℓ Γ ( h ) Γ ( h 1 ) Γ ( h 2 ) A m ( h ) A m 1 ( h 1 ) A m 2 ( h 2 ) . (57)
Since h 1 + h 2 = h − 1 h_1+h_2=h-1 h 1 + h 2 = h − 1 and m 1 + m 2 = m − ℓ m_1+m_2=m-\ell m 1 + m 2 = m − ℓ , positivity and coefficient convolution bound the sum over node degrees by A m − ℓ ( h − 1 ) A_{m-\ell}(h-1) A m − ℓ ( h − 1 ) . More explicitly, for each genus/label partition first apply the node-independent envelope (52) , then extend the allowed nonnegative model-coefficient sum to the full convolution. The resulting label envelope is exactly the one summed in (54) –(56) , so variable degree supports cause no interchange issue. Consequently the entire weighted separating mass, including the exponential factor e η min ( h 1 , h 2 ) e^{\eta\min(h_1,h_2)} e η m i n ( h 1 , h 2 ) of (56) , is at most
C S 0 h − 3 / 2 ( 1 + ι ) = O ( h − 1 / 2 ) . (58) CS_0h^{-3/2}(1+\iota)=O(h^{-1/2}). \tag*{(58)} C S 0 h − 3/2 ( 1 + ι ) = O ( h − 1/2 ) . (58)
For L ≥ 6 L\ge6 L ≥ 6 , use S 0 ≤ C h / L S_0\le Ch/L S 0 ≤ C h / L and ι ≤ n ≤ L \iota\le n\le L ι ≤ n ≤ L . For L ≤ 5 L\le5 L ≤ 5 , both ι \iota ι and L L L are bounded, and S 0 ≤ C h S_0\le C\sqrt{h} S 0 ≤ C h gives the smaller O ( h − 1 ) O(h^{-1}) O ( h − 1 ) bound.
The upper bound from successive maxima.
Lemma 5.6 (Upper bound). There is a constant C C C such that R ≤ C W e C h R\le C\mathcal{W}e^{C\sqrt{h}} R ≤ C W e C h at every state.
Proof. By Lemmas 5.4 and 5.5 , the weighted ratio satisfies an approximate maximum principle: at each nonterminal state, R / W R/\mathcal{W} R / W is at most 1 + O ( h − 1 / 2 ) 1+O(h^{-1/2}) 1 + O ( h − 1/2 ) times the largest value at its linear children, plus a separating contribution that is small. We apply this at complexities where the maximum of R / W R/\mathcal{W} R / W reaches a record, first to exclude exponential growth (59) , and then to obtain the bound in (60) .
At m = 0 m=0 m = 0 , (42) and (49) give R / W ≤ C R/\mathcal{W}\le C R / W ≤ C . At g ≤ 1 g\le1 g ≤ 1 , one has D ≤ n D\le n D ≤ n , so R a c t ≤ n / 2 R_{\mathrm{act}}\le n/2 R act ≤ n /2 and ι ≥ n / 2 \iota\ge n/2 ι ≥ n /2 . Thus log W ≥ ( n / 4 ) log n − O ( n ) \log\mathcal{W}\ge(n/4)\log n-O(n) log W ≥ ( n /4 ) log n − O ( n ) , which dominates (45) , since h = n + O ( 1 ) h=n+O(1) h = n + O ( 1 ) . Hence R / W R/\mathcal{W} R / W is bounded at every terminal state.
Let
M h = max Y ∈ [ 0 , y 0 ] valid states of complexity h R W . M_h=\max_{\substack{Y\in[0,y_0]\\ \text{valid states of complexity }h}}\frac{R}{\mathcal{W}}. M h = Y ∈ [ 0 , y 0 ] valid states of complexity h max W R .
The maximum exists: the tuple set is finite and the normalized parameter interval is compact. Equation (45) gives γ = lim sup h − 1 log M h < ∞ \gamma=\limsup h^{-1}\log M_h<\infty γ = lim sup h − 1 log M h < ∞ . Suppose γ > 0 \gamma>0 γ > 0 , and choose 0 < γ ′ < γ 0<\gamma'<\gamma 0 < γ ′ < γ , δ > 0 \delta>0 δ > 0 , so close that γ + δ − γ ′ < η \gamma+\delta-\gamma'<\eta γ + δ − γ ′ < η from (56) . Call h h h a record for a sequence a h a_h a h if a h > a j a_h>a_j a h > a j for all j < h j<h j < h . There are arbitrarily large records of M h e − γ ′ h M_he^{-\gamma'h} M h e − γ ′ h , and their maximizing states are nonterminal. At such a record,
M j / M h ≤ e − γ ′ ( h − j ) ( j < h ) , M j ≤ C δ e ( γ + δ ) j . M_j/M_h\le e^{-\gamma'(h-j)}\quad(j<h),\qquad M_j\le C_\delta e^{(\gamma+\delta)j}. M j / M h ≤ e − γ ′ ( h − j ) ( j < h ) , M j ≤ C δ e ( γ + δ ) j .
The linear part, after division by the maximizing value, is bounded by ( 1 + C h − 1 / 2 ) e − γ ′ (1+Ch^{-1/2})e^{-\gamma'} ( 1 + C h − 1/2 ) e − γ ′ . For a separating term use the record inequality on the larger child and the global bound on the smaller. Their product divided by M h M_h M h is at most a fixed constant times e ( γ + δ − γ ′ ) h s m a l l e^{(\gamma+\delta-\gamma')h_{\mathrm{small}}} e ( γ + δ − γ ′ ) h small . Equation (58) , which includes the factor e η min ( h 1 , h 2 ) e^{\eta\min(h_1,h_2)} e η m i n ( h 1 , h 2 ) , makes the total separating contribution o ( 1 ) o(1) o ( 1 ) . The recursion would therefore give 1 ≤ e − γ ′ + o ( 1 ) 1\le e^{-\gamma'}+o(1) 1 ≤ e − γ ′ + o ( 1 ) , a contradiction. Thus
lim sup h − 1 log M h ≤ 0. (59) \limsup h^{-1}\log M_h\le0. \tag*{(59)} lim sup h − 1 log M h ≤ 0. (59)
To obtain the quantitative envelope, fix 0 < δ < η 0<\delta<\eta 0 < δ < η , so (59) gives M j ≤ C δ e δ j M_j\le C_\delta e^{\delta j} M j ≤ C δ e δ j . Consider record maxima of M h / e A h M_h/e^{A\sqrt{h}} M h / e A h . If these were unbounded, their large maximizing states would again be nonterminal. For the larger separating child, the record bound supplies M l a r g e / M h ≤ e A ( h l a r g e − h ) ≤ 1 M_{\mathrm{large}}/M_h\le e^{A(\sqrt{h_{\mathrm{large}}}-\sqrt{h})}\le1 M large / M h ≤ e A ( h large − h ) ≤ 1 ; bound the smaller child by C δ e δ h s m a l l C_\delta e^{\delta h_{\mathrm{small}}} C δ e δ h small . The separating contribution is therefore C ′ h − 1 / 2 C'h^{-1/2} C ′ h − 1/2 , with C ′ C' C ′ independent of A A A . The linear record ratio is
e A ( h − 1 − h ) = 1 − A 2 h + O A ( h − 1 ) . e^{A(\sqrt{h-1}-\sqrt{h})}=1-\frac{A}{2\sqrt{h}}+O_A(h^{-1}). e A ( h − 1 − h ) = 1 − 2 h A + O A ( h − 1 ) .
Using (50) , the recursion would give
1 ≤ 1 + C + C ′ − A / 2 h + O A ( h − 1 ) . 1\le1+\frac{C+C'-A/2}{\sqrt{h}}+O_A(h^{-1}). 1 ≤ 1 + h C + C ′ − A /2 + O A ( h − 1 ) .
Choosing A > 2 ( C + C ′ + 1 ) A>2(C+C'+1) A > 2 ( C + C ′ + 1 ) excludes unbounded records. We have proved
R / W ≤ C e C h , (60) R/\mathcal{W}\le Ce^{C\sqrt{h}}, \tag*{(60)} R / W ≤ C e C h , (60)
uniformly for Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] .
Completion of the comparison.
Completion of the proof of Proposition 5.1. For fixed n n n , the zero count and the inactive count are at most n n n , so W ≤ 4 n ϕ ( n ) \mathcal{W} \leq4^n\phi(n) W ≤ 4 n ϕ ( n ) is bounded. Lemmas 5.3 and 5.6 therefore give, for every fixed n ≥ 1 n \geq1 n ≥ 1 ,
− C n log 2 ( h + 1 ) ≤ log U g , n ( d ; Y ) 3 h Γ ( h ) 4 − g [ z m ] H ‾ Y ( z ) − h ≤ C n h , (61) -C_n\log^2(h+1) \leq\log\frac{\mathcal{U}_{g,n}(\mathbf{d};Y)}{3^h\Gamma(h)4^{-g}[z^m]\overline{H}_Y(z)^{-h}} \leq C_n\sqrt{h}, \tag*{(61)} − C n log 2 ( h + 1 ) ≤ log 3 h Γ ( h ) 4 − g [ z m ] H Y ( z ) − h U g , n ( d ; Y ) ≤ C n h , (61)
uniformly for Y ∈ [ 0 , y 0 ] Y \in[0,y_0] Y ∈ [ 0 , y 0 ] and every degree tuple in the range of Proposition 5.1. This proves the proposition.
In particular, for each fixed n n n the logarithmic error in (61) is o ( g ) o(g) o ( g ) , uniformly in the degrees and in Y Y Y ; this is the form in which the comparison enters § 3.14 .
We record the case Y = 0 Y=0 Y = 0 , which compares the Weil–Petersson brackets themselves with an explicit model and is the input for the coefficient ratios in § 3.35 . At Y = 0 Y=0 Y = 0 , a = 1 a=1 a = 1 , r 0 = π 2 / 2 r_0=\pi^2/2 r 0 = π 2 /2 , and
H 0 ( z ) = 3 ( sin 2 z − 2 z cos 2 z ) ( 2 z ) 3 . H_0(z)=\frac{3(\sin\sqrt{2z}-\sqrt{2z}\cos\sqrt{2z})}{(\sqrt{2z})^3}. H 0 ( z ) = ( 2 z ) 3 3 ( sin 2 z − 2 z cos 2 z ) .
By (24) and (26) , U = b / 4 N \mathcal{U}=b/4^N U = b / 4 N , where b = [ τ d 1 ⋯ τ d n ] g , n b=[\tau_{d_1}\cdots\tau_{d_n}]_{g,n} b = [ τ d 1 ⋯ τ d n ] g , n . Undoing this same scaling in the model yields
M g , n W P ( d ) = 4 D 3 h Γ ( h ) 4 − g ( 2 π 2 ) m [ z m ] H 0 ( z ) − h . (62) M^{\mathrm{WP}}_{g,n}(\mathbf{d})=4^D3^h\Gamma(h)4^{-g}(2\pi^2)^m[z^m]H_0(z)^{-h}. \tag*{(62)} M g , n WP ( d ) = 4 D 3 h Γ ( h ) 4 − g ( 2 π 2 ) m [ z m ] H 0 ( z ) − h . (62)
Thus (61) applies unchanged to b / M W P b/M^{\mathrm{WP}} b / M WP . For n = 1 n=1 n = 1 , h = 2 g − 1 h=2g-1 h = 2 g − 1 , m = 3 g − 2 − k m=3g-2-k m = 3 g − 2 − k , it gives exactly
e − C log 2 ( g + 1 ) ≤ [ τ k ] g , 1 4 k 3 2 g − 1 Γ ( 2 g − 1 ) 4 − g ( 2 π 2 ) 3 g − 2 − k [ z 3 g − 2 − k ] H 0 ( z ) − ( 2 g − 1 ) ≤ e C g , (63) e^{-C\log^2(g+1)} \leq\frac{[\tau_k]_{g,1}}{4^k3^{2g-1}\Gamma(2g-1)4^{-g}(2\pi^2)^{3g-2-k}[z^{3g-2-k}]H_0(z)^{-(2g-1)}} \leq e^{C\sqrt{g}}, \tag*{(63)} e − C l o g 2 ( g + 1 ) ≤ 4 k 3 2 g − 1 Γ ( 2 g − 1 ) 4 − g ( 2 π 2 ) 3 g − 2 − k [ z 3 g − 2 − k ] H 0 ( z ) − ( 2 g − 1 ) [ τ k ] g , 1 ≤ e C g , (63)
uniformly over 0 ≤ k ≤ 3 g − 2 0 \leq k \leq3g-2 0 ≤ k ≤ 3 g − 2 .
Ratios of marked generating functions
The harmonic comparison of Proposition 5.1 controls logarithms, whereas the volume-ratio problem requires relative errors tending to zero. This section begins the second idea of Subsection 1.5 and obtains that stronger control for two ratios of generating functions with a fixed number p p p of distinguished markings: the cusp ratio, corresponding to adding a cusp, and the genus ratio, corresponding to replacing one unit of genus by two cusps. The result is Proposition 6.2 : both ratios have explicit leading terms with errors O p ( 1 / g ) O_p(1/g) O p ( 1/ g ) , uniformly in the density parameter. The inputs are the kernel bounds (14) , the positive recursion (28) and the o ( g ) o(g) o ( g ) form of Proposition 5.1; all proofs are given here. The passage to volumes with a prescribed cusp count is made in § 3.25 .
The argument combines positivity with two exact identities. Positivity of the recursion implies that a normalized amplitude is nonincreasing when one of its degrees is raised (70) . The dilaton equation (67) writes the cusp ratio as a weighted sum, over degrees d ≥ 1 d \geq1 d ≥ 1 , of the normalized amplitudes c g , p + 1 ( d , 0 p ) c_{g,p+1}(d,0^{p}) c g , p + 1 ( d , 0 p ) . These weights can have either sign; their absolute moments are uniformly bounded by (68) . The KdV equation (69) writes the genus ratio through c g , p ( 1 , 0 p − 1 ) c_{g,p}(1,0^{p-1}) c g , p ( 1 , 0 p − 1 ) and separating terms. Both ratios therefore admit explicit leading terms with relative error O p ( 1 / g ) O_p(1/g) O p ( 1/ g ) , uniformly in Y Y Y , once we know that c g , p ( d ) = 1 − O ( ∣ d ∣ + 1 ) / g c_{g,p}(\mathbf{d})=1-O(|\mathbf{d}|+1)/g c g , p ( d ) = 1 − O ( ∣ d ∣ + 1 ) / g . This last estimate, (78) , follows from the difference of the recursion at adjacent degrees; the coarse bounds of § 3.7 are used only to show that the separating terms are small.
Shifted amplitudes and normalization
Use p p p for the number of distinguished markings, and keep n n n for the number of cusps. Retain the kernels and shifted amplitudes of Sections 2 and 4; write U g , p ( d ; Y ) = U g , p ( d ) U_{g,p}(\mathbf d;Y)=U_{g,p}(\mathbf d) U g , p ( d ; Y ) = U g , p ( d ) , and suppress Y Y Y when convenient. Put h = 2 g − 2 + p h=2g-2+p h = 2 g − 2 + p and m = 3 g − 3 + p − ∣ d ∣ m=3g-3+p-|\mathbf d| m = 3 g − 3 + p − ∣ d ∣ . Define
U ‾ g , p ( d ) = a h r m U g , p ( d ) , W g , p = 4 g U ‾ g , p ( 0 ) , c g , p ( d ) = U ‾ g , p ( d ) U ‾ g , p ( 0 ) . (64) \overline{U}_{g,p}(\mathbf d)=a^{h}r^{m}U_{g,p}(\mathbf d),\qquad W_{g,p}=4^{g}\overline{U}_{g,p}(0),\qquad c_{g,p}(\mathbf d)=\frac{\overline{U}_{g,p}(\mathbf d)}{\overline{U}_{g,p}(0)}. \tag*{(64)} U g , p ( d ) = a h r m U g , p ( d ) , W g , p = 4 g U g , p ( 0 ) , c g , p ( d ) = U g , p ( 0 ) U g , p ( d ) . (64)
Thus U ‾ g , p ( d ) \overline{U}_{g,p}(\mathbf d) U g , p ( d ) is the normalized amplitude U g , p ( d ; Y ) U_{g,p}(\mathbf d;Y) U g , p ( d ; Y ) of (26) ; the bar indicates that the markings are now the distinguished ones. The barred amplitudes obey the exact positive recursion (28) .
We use two consequences of Sections 2–5.
1 = θ 0 ≤ θ 1 ≤ ⋯ , w ℓ : = θ ℓ − θ ℓ − 1 ≥ 0 , θ − 1 = 0 , sup Y ∑ ℓ ≥ 0 ( 1 + ℓ ) k w ℓ < ∞ . (65) 1=\theta_0\leq\theta_1\leq\cdots,\qquad w_\ell:=\theta_\ell-\theta_{\ell-1}\geq0,\qquad\theta_{-1}=0,\qquad\sup_Y\sum_{\ell\geq0}(1+\ell)^k w_\ell<\infty. \tag*{(65)} 1 = θ 0 ≤ θ 1 ≤ ⋯ , w ℓ := θ ℓ − θ ℓ − 1 ≥ 0 , θ − 1 = 0 , Y sup ℓ ≥ 0 ∑ ( 1 + ℓ ) k w ℓ < ∞. (65)
Also A ( Y ) A(Y) A ( Y ) is bounded above and away from zero; see Subsection 2.2.
log W g , p = log Γ ( 2 g − 2 + p ) − ( 2 g − 2 + p ) log A ( Y ) + o ( g ) , (66) \log W_{g,p}=\log\Gamma(2g-2+p)-(2g-2+p)\log A(Y)+o(g), \tag*{(66)} log W g , p = log Γ ( 2 g − 2 + p ) − ( 2 g − 2 + p ) log A ( Y ) + o ( g ) , (66)
uniformly for 0 ≤ Y ≤ y 0 0\leq Y\leq y_0 0 ≤ Y ≤ y 0 . The corresponding crude bound W g , p ≤ C p g + 1 ( 2 g + p ) ! W_{g,p}\leq C_p^{g+1}(2g+p)! W g , p ≤ C p g + 1 ( 2 g + p )! holds uniformly, including bounded stable genera. Proposition 5.1, together with the central local limit (36) for the model, proves (66) without any input on volume ratios. Only its o ( g ) o(g) o ( g ) error is needed below, and only for a positive number of markings.
Dilaton and KdV identities
The shifted weighted dilaton identity is
h U ‾ g , p ( 0 ) = ∑ d ≥ 1 b d ( Y ) U ‾ g , p + 1 ( d , 0 p ) , b d ( Y ) = X ( d ) ( Y ) r d − 1 a ( 2 d + 1 ) ! ! = 1 3 [ z d − 1 ] H ‾ Y ( z ) . (67) h\overline{U}_{g,p}(0)=\sum_{d\geq1}b_d(Y)\overline{U}_{g,p+1}(d,0^{p}),\qquad b_d(Y)=\frac{X^{(d)}(Y)r^{d-1}}{a(2d+1)!!}=\frac{1}{3}[z^{d-1}]\overline{H}_Y(z). \tag*{(67)} h U g , p ( 0 ) = d ≥ 1 ∑ b d ( Y ) U g , p + 1 ( d , 0 p ) , b d ( Y ) = a ( 2 d + 1 )!! X ( d ) ( Y ) r d − 1 = 3 1 [ z d − 1 ] H Y ( z ) . (67)
It holds also at p = 0 p=0 p = 0 , g ≥ 2 g\geq2 g ≥ 2 . To track the shift, the background insertion of degree one contributes 1 − a 1-a 1 − a ; moving it and the higher background insertions to the left leaves the coefficients X ( d ) / ( 2 d + 1 ) ! ! X^{(d)}/(2d+1)!! X ( d ) / ( 2 d + 1 )!! . The factor in (64) then produces exactly r d − 1 / a r^{d-1}/a r d − 1 / a in (67) . The compact analytic family gives
∑ d ≥ 1 b d = A ( Y ) , sup Y ∑ d ≥ 1 ( 1 + d ) k ∣ b d ( Y ) ∣ < ∞ (68) \sum_{d\geq1}b_d=A(Y),\qquad\sup_Y\sum_{d\geq1}(1+d)^k|b_d(Y)|<\infty \tag*{(68)} d ≥ 1 ∑ b d = A ( Y ) , Y sup d ≥ 1 ∑ ( 1 + d ) k ∣ b d ( Y ) ∣ < ∞ (68)
for every fixed k k k .
The genus ratio is governed instead by the KdV equation. Write F F F for the Kontsevich–Witten free energy, with subscripts indicating differentiation with respect to the corresponding times. The integrated KdV identity F 01 = F 00 2 / 2 + F 0000 / 12 F_{01}=F_{00}^2/2+F_{0000}/12 F 01 = F 00 2 /2 + F 0000 /12 , differentiated p − 2 p-2 p − 2 times in the string variable, gives for p ≥ 2 p\geq2 p ≥ 2
U ‾ g , p ( 1 , 0 p − 1 ) = 1 4 U ‾ g − 1 , p + 2 ( 0 ) + 3 2 ∑ j = 0 g ∑ s = 0 p − 2 ( p − 2 s ) U ‾ j , s + 2 ( 0 ) U ‾ g − j , p − s ( 0 ) . (69) \overline{U}_{g,p}(1,0^{p-1})=\frac{1}{4}\overline{U}_{g-1,p+2}(0)+\frac{3}{2}\sum_{j=0}^{g}\sum_{s=0}^{p-2}\binom{p-2}{s}\overline{U}_{j,s+2}(0)\overline{U}_{g-j,p-s}(0). \tag*{(69)} U g , p ( 1 , 0 p − 1 ) = 4 1 U g − 1 , p + 2 ( 0 ) + 2 3 j = 0 ∑ g s = 0 ∑ p − 2 ( s p − 2 ) U j , s + 2 ( 0 ) U g − j , p − s ( 0 ) . (69)
The split is ordered. The factor three from the weighted degree-one insertion turns the original KdV coefficients into 1 / 4 1/4 1/4 and 3 / 2 3/2 3/2 . In (69) , every term has the same total powers of a a a and r r r : the two separating complexities sum to h h h , and their deficits sum to 3 g − 4 + p 3g-4+p 3 g − 4 + p . Thus no omitted parameter factor is present.
The key consequence of positivity, used throughout this section, is the following monotonicity.
Lemma 6.1 (Monotonicity in the degrees). *For every stable ( g , p ) (g,p) ( g , p ) with p ≥ 1 p\geq1 p ≥ 1 , every nonnegative degree tuple d \mathbf d d , every 1 ≤ i ≤ p 1\leq i\leq p 1 ≤ i ≤ p and every Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] ,
U ‾ g , p ( d ) ≥ U ‾ g , p ( d + e i ) , 0 ≤ c g , p ( d ) ≤ 1. (70) \overline{U}_{g,p}(\mathbf d)\geq\overline{U}_{g,p}(\mathbf d+\mathbf e_i),\qquad0\leq c_{g,p}(\mathbf d)\leq1. \tag*{(70)} U g , p ( d ) ≥ U g , p ( d + e i ) , 0 ≤ c g , p ( d ) ≤ 1. (70) :::`
Amplitudes outside the dimension range are zero.
Proof. We subtract the recursion at adjacent degrees. With d ′ = ( d 2 , … , d p ) \mathbf{d}'=(d_2,\ldots,d_p) d ′ = ( d 2 , … , d p ) fixed, let B d ′ ( t ) B_{\mathbf{d}'}(t) B d ′ ( t ) denote the merging and contraction sums in (28) , with the kernel weight removed and d 1 + ℓ = t d_1+\ell=t d 1 + ℓ = t . These sums are nonnegative and vanish for t > 3 g − 3 + p − ∣ d ′ ∣ t>3g-3+p-|\mathbf{d}'| t > 3 g − 3 + p − ∣ d ′ ∣ by dimension. If m = 3 g − 3 + p − ∣ d ∣ ≥ 0 m=3g-3+p-|\mathbf{d}|\ge0 m = 3 g − 3 + p − ∣ d ∣ ≥ 0 , shifting the kernel index gives
U ‾ g , p ( d ) = ∑ ℓ = 0 m θ ℓ B d ′ ( d 1 + ℓ ) , \overline{U}_{g,p}(\mathbf{d})=\sum_{\ell=0}^{m}\theta_{\ell}B_{\mathbf{d}'}(d_1+\ell), U g , p ( d ) = ℓ = 0 ∑ m θ ℓ B d ′ ( d 1 + ℓ ) ,
U ‾ g , p ( d + e 1 ) = ∑ ℓ = 1 m θ ℓ − 1 B d ′ ( d 1 + ℓ ) . \overline{U}_{g,p}(\mathbf{d}+\mathbf{e}_1)=\sum_{\ell=1}^{m}\theta_{\ell-1}B_{\mathbf{d}'}(d_1+\ell). U g , p ( d + e 1 ) = ℓ = 1 ∑ m θ ℓ − 1 B d ′ ( d 1 + ℓ ) .
Their difference is therefore ∑ ℓ = 0 m w ℓ B d ′ ( d 1 + ℓ ) ≥ 0 \sum_{\ell=0}^{m}w_{\ell}B_{\mathbf{d}'}(d_1+\ell)\ge0 ∑ ℓ = 0 m w ℓ B d ′ ( d 1 + ℓ ) ≥ 0 , which proves (70) for i = 1 i=1 i = 1 , and by symmetry for every i i i . This includes the dimension boundary by the zero convention. The difference identity is used for h ≥ 2 h\ge2 h ≥ 2 ; the finite bases satisfy the inequality directly because θ 1 ≥ θ 0 = 1 \theta_1\ge\theta_0=1 θ 1 ≥ θ 0 = 1 . □ \square □
From (67) , (68) , and (70) ,
W g , p W g , p + 1 ≤ C 2 g − 2 + p (71) \frac{W_{g,p}}{W_{g,p+1}}\le\frac{C}{2g-2+p} \tag*{(71)} W g , p + 1 W g , p ≤ 2 g − 2 + p C (71)
in all stable cases used below, with a constant independent of p p p : the absolute sum in (68) is uniform and the ratios in (70) are at most one. Finite stable bases are included in the constant. Multiplying (69) by 4 g 4^g 4 g and using (70) gives
W g − 1 , p + 2 ≤ W g , p , p ≥ 2. (72) W_{g-1,p+2}\le W_{g,p},\qquad p\ge2. \tag*{(72)} W g − 1 , p + 2 ≤ W g , p , p ≥ 2. (72)
The factor 4 g 4^g 4 g in (64) is essential: it converts the coefficient 1 / 4 1/4 1/4 into the comparison constant one.
Separating terms
The separating terms in the KdV identity (69) , and in the degree difference above, are controlled by the convolution sums S g , p S_{g,p} S g , p and T g , p T_{g,p} T g , p below. They are small for two reasons. First, splitting off a component of bounded genus j j j removes 2 j + s 2j+s 2 j + s units of complexity, and the corresponding cusp insertions cost a factor O ( 1 / g ) O(1/g) O ( 1/ g ) each, by (71) –(72) . Second, a split into two components of comparable genus is exponentially small by the binomial entropy in (76) .
For fixed p ≥ 2 p\ge2 p ≥ 2 , define
S g , p = 1 W g , p ∑ j = 0 g ∑ s = 0 p − 1 ( p − 1 s ) W j , s + 1 W g − j , p − s , S_{g,p}=\frac{1}{W_{g,p}}\sum_{j=0}^{g}\sum_{s=0}^{p-1}\binom{p-1}{s}W_{j,s+1}W_{g-j,p-s}, S g , p = W g , p 1 j = 0 ∑ g s = 0 ∑ p − 1 ( s p − 1 ) W j , s + 1 W g − j , p − s ,
T g , p = 1 W g , p ∑ j = 0 g ∑ s = 0 p − 2 ( p − 2 s ) W j , s + 2 W g − j , p − s . (73) T_{g,p}=\frac{1}{W_{g,p}}\sum_{j=0}^{g}\sum_{s=0}^{p-2}\binom{p-2}{s}W_{j,s+2}W_{g-j,p-s}. \tag*{(73)} T g , p = W g , p 1 j = 0 ∑ g s = 0 ∑ p − 2 ( s p − 2 ) W j , s + 2 W g − j , p − s . (73)
We claim, uniformly in Y Y Y ,
S g , p = O p ( g − 2 ) , T g , p = O p ( g − 1 ) , (74) S_{g,p}=O_p(g^{-2}),\qquad T_{g,p}=O_p(g^{-1}), \tag*{(74)} S g , p = O p ( g − 2 ) , T g , p = O p ( g − 1 ) , (74)
with the positive-genus portion of T T T bounded by O p ( g − 2 ) O_p(g^{-2}) O p ( g − 2 ) . In particular T g , 2 = O ( g − 2 ) T_{g,2}=O(g^{-2}) T g , 2 = O ( g − 2 ) because its genus-zero terms are unstable.
By symmetry orient the smaller genus as j ≤ g / 2 j\le g/2 j ≤ g /2 . For 1 ≤ j ≤ ε g 1\le j\le\varepsilon g 1 ≤ j ≤ ε g , insert 2 j 2j 2 j cusps into the larger-genus component using (71) , apply (72) j j j times, and insert the remaining s s s cusps. This gives
W g − j , p − s W g , p ≤ ( C p g ) 2 j + s . (75) \frac{W_{g-j,p-s}}{W_{g,p}}\le\left(\frac{C_p}{g}\right)^{2j+s}. \tag*{(75)} W g , p W g − j , p − s ≤ ( g C p ) 2 j + s . (75)
If p − s = 1 p-s=1 p − s = 1 , instead insert 2 j + 1 2j+1 2 j + 1 cusps before raising the genus, so every genus-comparison step ends with at least two markings; afterward insert s − 1 s-1 s − 1 cusps. This is possible because p ≥ 2 p\ge2 p ≥ 2 , and the same exponent results. In the T T T convolution both relevant marking counts are already at least two. The crude bound following (66) makes the smaller-component contribution a fixed polynomial in j j j times
( 2 j ) ! ( C p / g 2 ) j g − s . (2j)!(C_p/g^2)j^{g-s}. ( 2 j )! ( C p / g 2 ) j g − s .
Choose ε > 0 \varepsilon>0 ε > 0 small enough that successive factorial terms have a ratio bounded by a fixed number below one. Their positive-genus sum is O p ( g − 2 ) O_p(g^{-2}) O p ( g − 2 ) , also with any fixed polynomial in j j j . For genus zero, stability gives s ≥ 2 s\ge2 s ≥ 2 in S S S and s ≥ 1 s\ge1 s ≥ 1 in T T T ; direct use of (71) gives the respective orders.
For ε g ≤ j ≤ g / 2 \varepsilon g\le j\le g/2 ε g ≤ j ≤ g /2 , apply (66) to both factors and the parent. The marking counts belong to a fixed finite set. The powers of A ( Y ) A(Y) A ( Y ) cancel up to bounded exponents, and Stirling’s formula gives
log W j , s + c W g − j , p − s W g , p ≤ − 2 g H ( j / g ) + o ( g ) , c ∈ { 1 , 2 } . (76) \log\frac{W_{j,s+c}W_{g-j,p-s}}{W_{g,p}}\le-2gH(j/g)+o(g),\qquad c\in\{1,2\}. \tag*{(76)} log W g , p W j , s + c W g − j , p − s ≤ − 2 g H ( j / g ) + o ( g ) , c ∈ { 1 , 2 } . (76)
where H ( t ) = − t log t − ( 1 − t ) log ( 1 − t ) H(t)=-t\log t-(1-t)\log(1-t) H ( t ) = − t log t − ( 1 − t ) log ( 1 − t ) . Its positive minimum on [ ε , 1 / 2 ] [\varepsilon,1/2] [ ε , 1/2 ] makes the macroscopic sum exponentially small. This proves (74) . Every use of (66) here has a positive fixed marking count.
Cusp and genus ratios
For p ≥ 2 p\ge2 p ≥ 2 , the adjacent-degree difference obtained from (28) has a merging prefactor O p ( 1 / g ) O_p(1/g) O p ( 1/ g ) by (71) . Its nonseparating prefactor is bounded by
W g − 1 , p + 1 W g , p ≤ C g W g − 1 , p + 2 W g , p = O p ( g − 1 ) . (77) \frac{W_{g-1,p+1}}{W_{g,p}}\le\frac{C}{g}\frac{W_{g-1,p+2}}{W_{g,p}}=O_p(g^{-1}). \tag*{(77)} W g , p W g − 1 , p + 1 ≤ g C W g , p W g − 1 , p + 2 = O p ( g − 1 ) . (77)
using (71) –(72) . To see all three contributions, put D = ∣ d ∣ D=|\mathbf d| D = ∣ d ∣ , Ω 0 = ∑ ℓ ≥ 0 w ℓ \Omega_0=\sum_{\ell\ge0}w_\ell Ω 0 = ∑ ℓ ≥ 0 w ℓ and Ω 1 = ∑ ℓ ≥ 0 ℓ w ℓ \Omega_1=\sum_{\ell\ge0}\ell w_\ell Ω 1 = ∑ ℓ ≥ 0 ℓ w ℓ . Bound each child amplitude by its zero-degree value in the difference identity above. Since there are at most d 1 + ℓ + 1 d_1+\ell+1 d 1 + ℓ + 1 new-degree pairs, this yields
c g , p ( d ) − c g , p ( d + e 1 ) ≤ Ω 0 ( 2 D + p − 1 ) W g , p − 1 W g , p + ( d 1 + 1 ) Ω 0 + Ω 1 ( 2 W g − 1 , p + 1 W g , p + 1 2 S g , p ) . c_{g,p}(\mathbf d)-c_{g,p}(\mathbf d+\mathbf e_1)\le\Omega_0(2D+p-1)\frac{W_{g,p-1}}{W_{g,p}}+(d_1+1)\Omega_0+\Omega_1\left(2\frac{W_{g-1,p+1}}{W_{g,p}}+\frac{1}{2}S_{g,p}\right). c g , p ( d ) − c g , p ( d + e 1 ) ≤ Ω 0 ( 2 D + p − 1 ) W g , p W g , p − 1 + ( d 1 + 1 ) Ω 0 + Ω 1 ( 2 W g , p W g − 1 , p + 1 + 2 1 S g , p ) .
The factor 2 in the nonseparating term includes the factor 4 from the genus normalization of W W W . The two kernel moments are uniformly bounded by (65) . Equations (71) , (74) and (77) now give
0 ≤ c g , p ( d ) − c g , p ( d + e i ) ≤ C p ( ∣ d ∣ + 1 ) g . 0\le c_{g,p}(\mathbf d)-c_{g,p}(\mathbf d+\mathbf e_i)\le\frac{C_p(|\mathbf d|+1)}{g}. 0 ≤ c g , p ( d ) − c g , p ( d + e i ) ≤ g C p ( ∣ d ∣ + 1 ) .
Summing degree increments proves
0 ≤ 1 − c g , p ( d ) ≤ C p ∣ d ∣ ( ∣ d ∣ + 1 ) g , p ≥ 2. (78) 0\le1-c_{g,p}(\mathbf d)\le C_p\frac{|\mathbf d|(|\mathbf d|+1)}{g},\qquad p\ge2. \tag*{(78)} 0 ≤ 1 − c g , p ( d ) ≤ C p g ∣ d ∣ ( ∣ d ∣ + 1 ) , p ≥ 2. (78)
When the right side exceeds one the bound follows from 0 ≤ c g , p ≤ 1 0\le c_{g,p}\le1 0 ≤ c g , p ≤ 1 in (70) . It also bounds the terms outside the dimension range when they are set to zero.
Substitute (78) into (67) . The errors sum absolutely by (68) , giving
( 2 g − 2 + p ) W g , p W g , p + 1 = A ( Y ) + O p ( g − 1 ) , p ≥ 1. (79) (2g-2+p)\frac{W_{g,p}}{W_{g,p+1}}=A(Y)+O_p(g^{-1}),\qquad p\ge1. \tag*{(79)} ( 2 g − 2 + p ) W g , p + 1 W g , p = A ( Y ) + O p ( g − 1 ) , p ≥ 1. (79)
Since A A A is uniformly positive, the error is relative O p ( 1 / g ) O_p(1/g) O p ( 1/ g ) .
To recover the one-marking profile without circularity, first use (71) twice, (72) at parent marking count two, and the now two-sided ratio (79) at marking count one:
W g − 1 , 2 W g , 1 ≤ C g 2 W g − 1 , 4 W g , 2 W g , 2 W g , 1 = O ( g − 1 ) . (80) \frac{W_{g-1,2}}{W_{g,1}}\le\frac{C}{g^2}\frac{W_{g-1,4}}{W_{g,2}}\frac{W_{g,2}}{W_{g,1}}=O(g^{-1}). \tag*{(80)} W g , 1 W g − 1 , 2 ≤ g 2 C W g , 2 W g − 1 , 4 W g , 1 W g , 2 = O ( g − 1 ) . (80)
The separating terms at one marking have two positive genera. Equation (71) , the inequality 1 / [ j ( g − j ) ] ≤ C / g 1/[j(g-j)] \le C/g 1/ [ j ( g − j )] ≤ C / g , and T g , 2 = O ( g − 2 ) T_{g,2}=O(g^{-2}) T g , 2 = O ( g − 2 ) give
1 W g , 1 ∑ j = 1 g − 1 W j , 1 W g − j , 1 ≤ C g W g , 2 W g , 1 T g , 2 = O ( g − 2 ) . (81) \frac{1}{W_{g,1}}\sum_{j=1}^{g-1}W_{j,1}W_{g-j,1}\leq\frac{C}{g}\frac{W_{g,2}}{W_{g,1}}T_{g,2}=O(g^{-2}). \tag*{(81)} W g , 1 1 j = 1 ∑ g − 1 W j , 1 W g − j , 1 ≤ g C W g , 1 W g , 2 T g , 2 = O ( g − 2 ) . (81)
The same degree-difference argument now proves (78) for p = 1 p=1 p = 1 . Thus (79) holds also for p = 0 p=0 p = 0 .
Finally (69) , (74) , and (78) yield
W g − 1 , p + 2 W g , p = 1 + O p ( g − 1 ) , p ≥ 2. (82) \frac{W_{g-1,p+2}}{W_{g,p}}=1+O_p(g^{-1}),\qquad p\geq2. \tag*{(82)} W g , p W g − 1 , p + 2 = 1 + O p ( g − 1 ) , p ≥ 2. (82)
For p = 0 , 1 p=0,1 p = 0 , 1 , add two markings in numerator and denominator and use (79) ; the two cusp products have the same complexity factors and cancel. Hence (82) holds for every fixed p ≥ 0 p\geq0 p ≥ 0 .
We now return to volumes. Recall the cusp generating function Φ g \Phi_g Φ g of (1) and the string-shift identity (4.7): $U_{g,p}(0^p;Y)=\Phi_g^{(p)}(x)$ with $x=X(Y)$, whenever the amplitude is stable. By (6.1)`, W g , p = 4 g a h r m Φ g ( p ) ( x ) W_{g,p}=4^g a^h r^m\Phi_g^{(p)}(x) W g , p = 4 g a h r m Φ g ( p ) ( x ) with m = 3 g − 3 + p m=3g-3+p m = 3 g − 3 + p . Adding a marking raises both h h h and m m m by one, so
x Φ g ( p + 1 ) ( x ) Φ g ( p ) ( x ) = x a r W g , p + 1 W g , p = h x A a r ( 1 + O p ( g − 1 ) ) \frac{x\Phi_g^{(p+1)}(x)}{\Phi_g^{(p)}(x)}=\frac{x}{ar}\frac{W_{g,p+1}}{W_{g,p}}=\frac{hx}{Aar}\left(1+O_p(g^{-1})\right) Φ g ( p ) ( x ) x Φ g ( p + 1 ) ( x ) = a r x W g , p W g , p + 1 = A a r h x ( 1 + O p ( g − 1 ) )
by (79) . Replacing ( g , p ) (g,p) ( g , p ) by ( g − 1 , p + 2 ) (g-1,p+2) ( g − 1 , p + 2 ) leaves h h h unchanged and lowers m m m by one, so Φ g − 1 ( p + 2 ) / Φ g ( p ) = 4 r W g − 1 , p + 2 / W g , p \Phi_{g-1}^{(p+2)}/\Phi_g^{(p)}=4rW_{g-1,p+2}/W_{g,p} Φ g − 1 ( p + 2 ) / Φ g ( p ) = 4 r W g − 1 , p + 2 / W g , p , and (82) applies. This proves, for every fixed positive p p p ,
x ( Φ g ( p ) ) ′ ( x ) Φ g ( p ) ( x ) = h Ψ ( x ) + O p ( Ψ ( x ) ) , h = 2 g − 2 + p , (83) x\frac{(\Phi_g^{(p)})'(x)}{\Phi_g^{(p)}(x)}=h\Psi(x)+O_p(\Psi(x)),\qquad h=2g-2+p, \tag*{(83)} x Φ g ( p ) ( x ) ( Φ g ( p ) ) ′ ( x ) = h Ψ ( x ) + O p ( Ψ ( x )) , h = 2 g − 2 + p , (83)
Φ g − 1 ( p + 2 ) ( x ) Φ g ( p ) ( x ) = 4 r ( Y ) ( 1 + O p ( g − 1 ) ) , Ψ ( x ) = x A ( Y ) a ( Y ) r ( Y ) . (84) \frac{\Phi_{g-1}^{(p+2)}(x)}{\Phi_g^{(p)}(x)}=4r(Y)(1+O_p(g^{-1})),\qquad\Psi(x)=\frac{x}{A(Y)a(Y)r(Y)}. \tag*{(84)} Φ g ( p ) ( x ) Φ g − 1 ( p + 2 ) ( x ) = 4 r ( Y ) ( 1 + O p ( g − 1 )) , Ψ ( x ) = A ( Y ) a ( Y ) r ( Y ) x . (84)
The same genus ratio holds at p = 0 p=0 p = 0 . The estimates are uniform throughout 0 ≤ Y < y 0 0\leq Y<y_0 0 ≤ Y < y 0 ; at x = 0 x=0 x = 0 , (83) is interpreted by continuity.
The scalar defined in the last display is given exactly by the density parameter:
H Y ( r ) = 3 I ( Y ) 2 Δ 3 / 2 , A ( Y ) a ( Y ) r ( Y ) = I ( Y ) Δ , Ψ ( x ) = λ ( Y ) 2 . (85) H_Y(r)=\frac{3I(Y)}{2\Delta^{3/2}},\qquad A(Y)a(Y)r(Y)=\frac{I(Y)}{\sqrt{\Delta}},\qquad\Psi(x)=\frac{\lambda(Y)}{2}. \tag*{(85)} H Y ( r ) = 2 Δ 3/2 3 I ( Y ) , A ( Y ) a ( Y ) r ( Y ) = Δ I ( Y ) , Ψ ( x ) = 2 λ ( Y ) . (85)
The first equality follows by substituting v = T − Δ s 2 v=T-\Delta s^2 v = T − Δ s 2 in I I I , and the other two follow by direct substitution. We summarize the output of this section used in the extraction arguments.
Proposition 6.2 (Marked cusp and genus ratios). For every fixed p ≥ 1 p\geq1 p ≥ 1 , uniformly for 0 ≤ Y < y 0 0\leq Y<y_0 0 ≤ Y < y 0 and x = X ( Y ) x=X(Y) x = X ( Y ) , the cusp ratio of Φ g ( p ) \Phi_g^{(p)} Φ g ( p ) satisfies (83) and the genus ratio satisfies (84) , with Ψ ( x ) = λ ( Y ) / 2 \Psi(x)=\lambda(Y)/2 Ψ ( x ) = λ ( Y ) /2 . The genus ratio holds also at p = 0 p=0 p = 0 .
Proposition 6.2 is used in two ways. The normalized forms (79) and (82) enter the differential equation for the leading term in § 3.19 , and the ratios (83) and (84) supply the center of the local limit theorem and the ratio of generating functions in § 3.25 . The degree estimate (78) is used again for bounded boundary lengths, in (104) and in the proof of Proposition 9.4.
The leading term and its normalization
The marked ratios of Proposition 6.2 determine a differential equation in Y Y Y for log Φ g ( p ) \log\Phi_g^{(p)} log Φ g ( p ) but leave a multiplicative constant undetermined. The equation comes from the dilaton identity (67) , which expresses the Y Y Y -derivative of log Φ g ( p ) \log\Phi_g^{(p)} log Φ g ( p ) through the degree profiles c g , p + 1 ( d , 0 p ) c_{g,p+1}(\mathbf d,0^p) c g , p + 1 ( d , 0 p ) ; to obtain it with error O ( h − 2 ) O(h^{-2}) O ( h − 2 ) we need the 1 / h 1/h 1/ h term of these profiles (Lemma 7.2). We then integrate the equation and fix the constant at Y = y 0 Y=y_0 Y = y 0 , where only the single sequence C g ( 2 3 g − 3 ) C_g(2^{3g-3}) C g ( 2 3 g − 3 ) of pure ψ \psi ψ intersection numbers enters and Painlevé I supplies its asymptotics (97) ; the quantitative form of that input is proved in Appendix § 8.5 , Subsection D.11. The result is Proposition 7.1, whose fixed-cusp specialization proves Corollary 1.2. Matching at a point that approaches the critical endpoint polynomially in the genus keeps the integrated error uniform on the whole interval.
Write K ˉ = K ˉ Y \bar K=\bar K_Y K ˉ = K ˉ Y and H ˉ = H ˉ Y \bar H=\bar H_Y H ˉ = H ˉ Y , and introduce the kernel moments
Θ = ∑ ℓ ≥ 0 w ℓ = − 1 K ˉ ′ ( 1 ) , M 1 = ∑ ℓ ≥ 0 ℓ w ℓ = K ˉ ′ ′ ( 1 ) 2 K ˉ ′ ( 1 ) 2 , Z = A a r 3 / 2 = 2 I . (86) \Theta=\sum_{\ell\geq0}w_\ell=-\frac{1}{\bar K'(1)},\qquad M_1=\sum_{\ell\geq0}\ell w_\ell=\frac{\bar K''(1)}{2\bar K'(1)^2},\qquad\mathcal Z=Aar^{3/2}=\sqrt{2}I. \tag*{(86)} Θ = ℓ ≥ 0 ∑ w ℓ = − K ˉ ′ ( 1 ) 1 , M 1 = ℓ ≥ 0 ∑ ℓ w ℓ = 2 K ˉ ′ ( 1 ) 2 K ˉ ′′ ( 1 ) , Z = A a r 3/2 = 2 I . (86)
Primes on K ˉ , H ˉ \bar K,\bar H K ˉ , H ˉ mean differentiation in their argument; derivatives of parameter functions below are with respect to Y Y Y . All the moments in (86) are uniformly bounded on the closed interval 0 ≤ Y ≤ y 0 0\leq Y\leq y_0 0 ≤ Y ≤ y 0 . The variance v Y v_Y v Y of the law in (15) , written σ Y 2 \sigma_Y^2 σ Y 2 in Section 2, satisfies
v Y = 15 4 − H ˉ ′ ′ ( 1 ) H ˉ ( 1 ) = 1 2 A Θ . (87) v_Y=\frac{15}{4}-\frac{\bar H''(1)}{\bar H(1)}=\frac{1}{2A\Theta}. \tag*{(87)} v Y = 4 15 − H ˉ ( 1 ) H ˉ ′′ ( 1 ) = 2 A Θ 1 . (87)
Indeed H ˉ ′ ( 1 ) = − 3 H ˉ ( 1 ) / 2 \bar H'(1)=-3\bar H(1)/2 H ˉ ′ ( 1 ) = − 3 H ˉ ( 1 ) /2 and K ˉ ′ ( 1 ) = − 5 H ˉ ( 1 ) / 2 + 2 H ˉ ′ ′ ( 1 ) / 3 \bar K'(1)=-5\bar H(1)/2+2\bar H''(1)/3 K ˉ ′ ( 1 ) = − 5 H ˉ ( 1 ) /2 + 2 H ˉ ′′ ( 1 ) /3 .
For fixed p ≥ 1 p\geq1 p ≥ 1 , put h = 2 g − 2 + p h=2g-2+p h = 2 g − 2 + p and define
J g , p ( Y ) = Γ ( h ) 4 − g π h Z ( Y ) h r ( Y ) p / 2 A ( Y ) Θ ( Y ) . (88) J_{g,p}(Y)=\frac{\Gamma(h)4^{-g}}{\sqrt{\pi h}}\mathcal Z(Y)h_r(Y)^{p/2}\sqrt{A(Y)}\Theta(Y). \tag*{(88)} J g , p ( Y ) = π h Γ ( h ) 4 − g Z ( Y ) h r ( Y ) p /2 A ( Y ) Θ ( Y ) . (88)
Proposition 7.1 (Absolute marked prefactor). For each fixed p ≥ 1 p\geq1 p ≥ 1 , uniformly for 0 ≤ Y < y 0 0\leq Y<y_0 0 ≤ Y < y 0 ,
Φ g ( p ) ( X ( Y ) ) = J g , p ( Y ) π ( 1 + O p ( log ( g + 1 ) g ) ) . (89) \Phi_g^{(p)}(X(Y))=\frac{J_{g,p}(Y)}{\pi}\left(1+O_p\left(\frac{\log(g+1)}{g}\right)\right). \tag*{(89)} Φ g ( p ) ( X ( Y )) = π J g , p ( Y ) ( 1 + O p ( g log ( g + 1 ) ) ) . (89)
The first degree correction
7.1. The first degree correction. We compute the first degree correction directly from the positive recursion. The remainder must be uniform in Y Y Y and in a growing degree range, because an expansion at each fixed degree would not suffice in the dilaton sum. Set D = ∣ d ∣ D=|\mathbf d| D = ∣ d ∣ and q = # { i : d i > 0 } q=\#\{i:d_i>0\} q = # { i : d i > 0 } .
Lemma 7.2 (First degree correction). For each fixed p ≥ 1 p\geq1 p ≥ 1 , uniformly for Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] and for degree tuples with D ≤ h 1 / 4 D\leq h^{1/4} D ≤ h 1/4 ,
c g , p ( d ) = 1 − δ p ( d ) h + O p ( ( D + 1 ) 4 h 2 ) , (90) c_{g,p}(\mathbf d)=1-\frac{\delta_p(\mathbf d)}{h}+O_p\left(\frac{(D+1)^4}{h^2}\right), \tag*{(90)} c g , p ( d ) = 1 − h δ p ( d ) + O p ( h 2 ( D + 1 ) 4 ) , (90)
δ p ( d ) = A { Θ ( D 2 + ( p − 4 ) D ) + 2 M 1 D + q ( q + 5 − 2 p ) 2 } . \delta_p(\mathbf d)=A\left\{\Theta\left(D^2+(p-4)D\right)+2M_1D+\frac{q(q+5-2p)}{2}\right\}. δ p ( d ) = A { Θ ( D 2 + ( p − 4 ) D ) + 2 M 1 D + 2 q ( q + 5 − 2 p ) } .
Proof. Subtract the exact recursion (28) at pivot degree t = d i t=d_i t = d i and t + 1 t+1 t + 1 , shifting the second kernel index. The difference has coefficients w ℓ w_\ell w ℓ . By (79) and (82) , its merging prefactor is A / h + O p ( h − 2 ) A/h+O_p(h^{-2}) A / h + O p ( h − 2 ) , and its nonseparating prefactor, including the factor 1 / 2 1/2 1/2 and the genus normalization, is 2 A / h + O p ( h − 2 ) 2A/h+O_p(h^{-2}) 2 A / h + O p ( h − 2 ) . For p = 1 p=1 p = 1 there is no merging term. All child degrees can be replaced by zero with error at most C p ( D + ℓ + 1 ) 2 / h C_p(D+\ell+1)^2/h C p ( D + ℓ + 1 ) 2 / h , by (78) . The coefficients and pair counts are O p ( D + ℓ + 1 ) O_p(D+\ell+1) O p ( D + ℓ + 1 ) , so the moments of w ℓ w_\ell w ℓ make the resulting error O p ( ( D + 1 ) 3 / h 2 ) O_p((D+1)^3/h^2) O p (( D + 1 ) 3 / h 2 ) . The separating difference is O p ( ( D + 1 ) / h 2 ) O_p((D+1)/h^2) O p (( D + 1 ) / h 2 ) by (74) ; at p = 1 p=1 p = 1 use (81) . These bounds also cover dimension-invalid children through the zero convention in (78) . Let z z z count zeros among the other p − 1 p-1 p − 1 entries. A merger is admissible when t + d j + ℓ ≥ 1 t+d_j+\ell\ge1 t + d j + ℓ ≥ 1 ; the nonseparating pair count is ( t + ℓ − 1 ) + (t+\ell-1)_+ ( t + ℓ − 1 ) + . Thus the leading adjacent difference is
A h { Θ ( 2 D + p − 3 ) + 2 M 1 + ( 2 − z ) 1 t = 0 } . (91) \frac{A}{h}\{\Theta(2D+p-3)+2M_1+(2-z)\mathbf{1}_{t=0}\}. \tag*{(91)} h A { Θ ( 2 D + p − 3 ) + 2 M 1 + ( 2 − z ) 1 t = 0 } . (91)
Here the zero-pivot term records both the missing zero-zero merger and the truncated pair count. Since the parent deficit is of order h h h in the indicated degree range, extending the kernel sum to infinity costs O p ( h − 2 ) O_p(h^{-2}) O p ( h − 2 ) , using sufficiently high moments of w ℓ w_\ell w ℓ . Summing increments from zero to d d d proves (90) : the exceptional increments at the creation of successive positive entries are 3 − p 3-p 3 − p , 4 − p 4-p 4 − p , … \ldots … , q + 2 − p q+2-p q + 2 − p . The sum of the remainders is bounded by C p ( D + 1 ) 4 / h 2 C_p(D+1)^4/h^2 C p ( D + 1 ) 4 / h 2 . □
For d > h 1 / 4 d>h^{1/4} d > h 1/4 in the dilaton identity, we use 0 ≤ c ≤ 1 0\le c\le1 0 ≤ c ≤ 1 , the quadratic growth of δ p \delta_p δ p , and sufficiently high uniform absolute moments of b d b_d b d ; moments of order twelve already bound the tail of 1 + d 2 / h 1+d^2/h 1 + d 2 / h by O p ( h − 2 ) O_p(h^{-2}) O p ( h − 2 ) . This justifies summing the expansion, although it gives no useful pointwise remainder at large degree.
The differential equation for the leading term
The three signed dilaton moments are exact:
∑ b d = A , ∑ d b d = − A / 2 , ∑ d 2 b d = H ‾ ′ ′ ( 1 ) / 3 − 7 A / 2. \sum b_d=A,\qquad\sum d b_d=-A/2,\qquad\sum d^2b_d=\overline{H}''(1)/3-7A/2. ∑ b d = A , ∑ d b d = − A /2 , ∑ d 2 b d = H ′′ ( 1 ) /3 − 7 A /2.
Apply (7.5) to $(d,0^p)$ in (6.4)`. Changing the child complexity h + 1 h+1 h + 1 to h h h costs O p ( h − 2 ) O_p(h^{-2}) O p ( h − 2 ) . We obtain
h W g , p W g , p + 1 = A ( 1 − β p h + O p ( h − 2 ) ) , β p = A { 7 − 2 p 4 Θ − M 1 + 2 − p } − 1 2 . (92) h\frac{W_{g,p}}{W_{g,p+1}}=A\left(1-\frac{\beta_p}{h}+O_p(h^{-2})\right),\qquad\beta_p=A\left\{\frac{7-2p}{4}\Theta-M_1+2-p\right\}-\frac12. \tag*{(92)} h W g , p + 1 W g , p = A ( 1 − h β p + O p ( h − 2 ) ) , β p = A { 4 7 − 2 p Θ − M 1 + 2 − p } − 2 1 . (92)
This holds for each fixed p ≥ 0 p\ge0 p ≥ 0 : at p = 0 p=0 p = 0 its only profile input has one marking. Inverting and undoing the normalization gives
d d Y log Φ g ( p ) ( X ( Y ) ) = h + β p + O p ( h − 1 ) A r . (93) \frac{\mathrm{d}}{\mathrm{d}Y}\log\Phi_g^{(p)}(X(Y))=\frac{h+\beta_p+O_p(h^{-1})}{Ar}. \tag*{(93)} d Y d log Φ g ( p ) ( X ( Y )) = A r h + β p + O p ( h − 1 ) . (93)
The correction in (93) has an explicit primitive. Differentiate K Y ( r ) = 0 K_Y(r)=0 K Y ( r ) = 0 , using ∂ Y K = 2 ∂ z K + ( K − a ) / z \partial_YK=2\partial_zK+(K-a)/z ∂ Y K = 2 ∂ z K + ( K − a ) / z , to get
r ′ = − ( Θ + 2 ) , ( log Z ) ′ = − 1 A r . (94) r'=-(\Theta+2),\qquad(\log\mathcal{Z})'=-\frac{1}{Ar}. \tag*{(94)} r ′ = − ( Θ + 2 ) , ( log Z ) ′ = − A r 1 . (94)
The second identity also follows from I ′ = − a Δ I'=-a\sqrt{\Delta} I ′ = − a Δ . Differentiating the root derivative and the corresponding equation for H H H gives
( log Θ ) ′ = a ′ a + 2 Θ + 1 − 2 M 1 r , ( log A ) ′ = 3 Θ / 2 + 3 − 1 / A r − a ′ a . (95) (\log\Theta)'=\frac{a'}{a}+\frac{2\Theta+1-2M_1}{r},
\qquad
(\log A)'=\frac{3\Theta/2+3-1/A}{r}-\frac{a'}{a}. \tag*{(95)} ( log Θ ) ′ = a a ′ + r 2Θ + 1 − 2 M 1 , ( log A ) ′ = r 3Θ/2 + 3 − 1/ A − a a ′ . (95)
For the first formula use Θ = − a / ( r ∂ z K Y ( r ) ) \Theta=-a/(r\partial_zK_Y(r)) Θ = − a / ( r ∂ z K Y ( r )) and ∂ Y ∂ z K Y = 2 ∂ z 2 K Y + ∂ z K Y / r + a / r 2 \partial_Y\partial_zK_Y=2\partial_z^2K_Y+\partial_zK_Y/r+a/r^2 ∂ Y ∂ z K Y = 2 ∂ z 2 K Y + ∂ z K Y / r + a / r 2 at the root z = r z=r z = r . Equations (92) and (95) yield
β p A r = 1 2 ( log ( A Θ ) ) ′ + p 2 ( log r ) ′ , ∣ d d Y log Φ g ( p ) ( X ( Y ) ) J g , p ( Y ) ∣ ≤ C p h A r . (96) \frac{\beta_p}{Ar}=\frac12(\log(A\Theta))'+\frac{p}{2}(\log r)',\qquad
\left|\frac{\mathrm{d}}{\mathrm{d}Y}\log\frac{\Phi_g^{(p)}(X(Y))}{\mathcal{J}_{g,p}(Y)}\right|\le\frac{C_p}{hAr}. \tag*{(96)} A r β p = 2 1 ( log ( A Θ ) ) ′ + 2 p ( log r ) ′ , d Y d log J g , p ( Y ) Φ g ( p ) ( X ( Y )) ≤ h A r C p . (96)
Only exact identities have been differentiated here; the error term in (93) will be integrated, never differentiated.
Fixing the constant at the critical endpoint
Here we need the value at one pure-ψ \psi ψ tuple, not only two-sided bounds. In the normalization (23) ,
C g ( 2 3 g − 3 ) = 1 π + O ( g − 1 ) . (97) C_g(2^{3g-3}) = \frac{1}{\pi} + O(g^{-1}). \tag*{(97)} C g ( 2 3 g − 3 ) = π 1 + O ( g − 1 ) . (97)
Indeed, the identity of Itzykson and Zuber [106, equation (6.10)] expresses C g ( 2 3 g − 3 ) C_g(2^{3g-3}) C g ( 2 3 g − 3 ) through the coefficients of the formal power-series solution of Painlevé I. Their large-order asymptotics follow from the Stokes data of Joshi–Kitaev and Kapaev [114, 116]; see [70, Section 2.3]. The constant itself is evaluated by Stirling’s formula in Lemma D.11 (iii), whose leading case uses only these two inputs. Subsection D.11 proves a quantitative version, (D.48), without using (97) . Estimate (97) is also the case e = ( 2 3 g − 3 ) e=(2^{3g-3}) e = ( 2 3 g − 3 ) of the uniform theorem of Guo, Yang and Zagier [83, Theorem 1]; no other case of that theorem is used. For m = 3 g − 3 + p m=3g-3+p m = 3 g − 3 + p , string followed by dilaton gives the exact reduction
C g ( 0 p , 2 m ) = ∏ i = 0 p − 1 5 ( m − i ) 3 ( 2 g − 3 + p + m − 2 i ) C g ( 2 3 g − 3 ) = 1 π + O p ( g − 1 ) . (98) C_g(0^p,2^m)=\prod_{i=0}^{p-1}\frac{5(m-i)}{3(2g-3+p+m-2i)}C_g(2^{3g-3})=\frac{1}{\pi}+O_p(g^{-1}). \tag*{(98)} C g ( 0 p , 2 m ) = i = 0 ∏ p − 1 3 ( 2 g − 3 + p + m − 2 i ) 5 ( m − i ) C g ( 2 3 g − 3 ) = π 1 + O p ( g − 1 ) . (98)
Each string step chooses a degree-two marking with normalized coefficient 5 / [ 3 ( 2 g − 3 + p + m ) ] 5/[3(2g-3+p+m)] 5/ [ 3 ( 2 g − 3 + p + m )] ; the resulting degree-one marking is removed by normalized dilaton with coefficient one. Since p p p is fixed, the displayed finite product is 1 + O p ( g − 1 ) 1+O_p(g^{-1}) 1 + O p ( g − 1 ) , and (97) gives the final equality.
At the endpoint, (43) gives exactly
U ‾ g , p ( 0 p ; y 0 ) M g , p ( 0 p ; y 0 ) = C g ( 0 p , 2 m ) . (99) \frac{\overline{U}_{g,p}(0^p;y_0)}{\mathcal{M}_{g,p}(0^p;y_0)}=C_g(0^p,2^m). \tag*{(99)} M g , p ( 0 p ; y 0 ) U g , p ( 0 p ; y 0 ) = C g ( 0 p , 2 m ) . (99)
Indeed the numerator is 3 m ⟨ τ 0 p τ 2 m ⟩ g / m ! 3^m\langle\tau_0^p\tau_2^m\rangle_g/m! 3 m ⟨ τ 0 p τ 2 m ⟩ g / m ! and the denominator is 3 h + m Γ ( h + m ) / ( 4 g 5 m m ! ) 3^{h+m}\Gamma(h+m)/(4^g5^m m!) 3 h + m Γ ( h + m ) / ( 4 g 5 m m !) .
We now show that the endpoint value persists on a short interval 0 ≤ y 0 − Y ≤ h − 2 0\le y_0-Y\le h^{-2} 0 ≤ y 0 − Y ≤ h − 2 . The reason is that the amplitudes are polynomials with nonnegative coefficients in the kernel coefficients, and these are within a factor 1 + O ( y 0 − Y ) 1+O(y_0-Y) 1 + O ( y 0 − Y ) of their endpoint value one. Here θ ℓ = [ z ℓ ] K ‾ Y ( z ) \theta_\ell=[z^\ell]\overline{K}_Y(z) θ ℓ = [ z ℓ ] K Y ( z ) are the rescaled kernel coefficients. Put δ = y 0 − Y \delta=y_0-Y δ = y 0 − Y . The normalized kernels are analytic in δ \delta δ near zero on a common disk of radius greater than one. To see the removable parameter singularity, apply the implicit-function theorem to K y 0 − δ ( δ ρ ) / δ K_{y_0-\delta}(\delta\rho)/\delta K y 0 − δ ( δ ρ ) / δ at ( δ , ρ ) = ( 0 , 3 ) (\delta,\rho)=(0,3) ( δ , ρ ) = ( 0 , 3 ) ; its ρ \rho ρ derivative is nonzero. Division by the persistent root 1 − z 1-z 1 − z then gives K ‾ Y ( z ) / ( 1 − z ) = 1 + O ( δ ) \overline{K}_Y(z)/(1-z)=1+O(\delta) K Y ( z ) / ( 1 − z ) = 1 + O ( δ ) on a smaller common disk. Together with coefficient positivity this proves
1 ≤ θ ℓ ( Y ) ≤ 1 + C δ for all ℓ ≥ 0. (100) 1\le\theta_\ell(Y)\le1+C\delta\qquad\text{for all }\ell\ge0. \tag*{(100)} 1 ≤ θ ℓ ( Y ) ≤ 1 + C δ for all ℓ ≥ 0. (100)
By the positive recursion, U ‾ \overline{U} U is a polynomial with nonnegative coefficients in the variables θ ℓ \theta_\ell θ ℓ , homogeneous of degree h h h . At complexity one the bases have degree one, with θ 0 = 1 \theta_0=1 θ 0 = 1 ; each recursion adds one factor, while child complexities sum to h − 1 h-1 h − 1 . Thus
1 ≤ U ‾ g , p ( 0 p ; Y ) U ‾ g , p ( 0 p ; y 0 ) ≤ ( 1 + C δ ) h . (101) 1\le\frac{\overline{U}_{g,p}(0^p;Y)}{\overline{U}_{g,p}(0^p;y_0)}\le(1+C\delta)^h. \tag*{(101)} 1 ≤ U g , p ( 0 p ; y 0 ) U g , p ( 0 p ; Y ) ≤ ( 1 + C δ ) h . (101)
Next, for every Y Y Y , we compare J g , p \mathcal{J}_{g,p} J g , p with the harmonic model at all degrees zero. Here m = 3 g − 3 + p = 3 h / 2 − p / 2 m=3g-3+p=3h/2-p/2 m = 3 g − 3 + p = 3 h /2 − p /2 . Since Z = A r 3 / 2 \mathcal{Z}=A r^{3/2} Z = A r 3/2 , the powers of r r r in a h , m J g , p a^{h,m}\mathcal{J}_{g,p} a h , m J g , p cancel exactly. On the other hand, (31) and (40) give M g , p ( 0 p ; Y ) = 3 h Γ ( h ) 4 − g ( 3 A ) − h Pr ( S h = m ) \mathcal{M}_{g,p}(0^p;Y)=3^h\Gamma(h)4^{-g}(3A)^{-h}\Pr(S_h=m) M g , p ( 0 p ; Y ) = 3 h Γ ( h ) 4 − g ( 3 A ) − h Pr ( S h = m ) , and the central estimate (37) at displacement − p / 2 -p/2 − p /2 , with variance v Y = 1 / ( 2 A Θ ) v_Y=1/(2A\Theta) v Y = 1/ ( 2 A Θ ) from (87) , gives Pr ( S h = m ) = A Θ / ( π h ) ( 1 + O p ( h − 1 ) ) \Pr(S_h=m)=\sqrt{A\Theta/(\pi h)}(1+O_p(h^{-1})) Pr ( S h = m ) = A Θ/ ( π h ) ( 1 + O p ( h − 1 )) . Hence
a h , m J g , p ( Y ) = Γ ( h ) 4 − g A − h A Θ π h = M g , p ( 0 p ; Y ) ( 1 + O p ( h − 1 ) ) , (102) a^{h,m}\mathcal{J}_{g,p}(Y)=\Gamma(h)4^{-g}A^{-h}\sqrt{\frac{A\Theta}{\pi h}}=\mathcal{M}_{g,p}(0^p;Y)(1+O_p(h^{-1})), \tag*{(102)} a h , m J g , p ( Y ) = Γ ( h ) 4 − g A − h π h A Θ = M g , p ( 0 p ; Y ) ( 1 + O p ( h − 1 )) , (102)
where the first equality is exact. Its ratio to its endpoint value is 1 + O p ( h δ ) 1+O_p(h\delta) 1 + O p ( h δ ) when 0 ≤ δ ≤ h − 2 0\le\delta\le h^{-2} 0 ≤ δ ≤ h − 2 , by analyticity and uniform positivity of A , Θ A,\Theta A , Θ . Hence (98) –(102) prove Φ g ( p ) ( X ( Y ) ) / J g , p ( Y ) = 1 / π + O p ( h − 1 ) \Phi_g^{(p)}(X(Y))/\mathcal{J}_{g,p}(Y)=1/\pi+O_p(h^{-1}) Φ g ( p ) ( X ( Y )) / J g , p ( Y ) = 1/ π + O p ( h − 1 ) throughout that interval.
Proof of Proposition 7.1. Take Y ∗ = y 0 − h − 2 Y_\ast=y_0-h^{-2} Y ∗ = y 0 − h − 2 . For Y ≤ Y ∗ Y\le Y_\ast Y ≤ Y ∗ integrate (96) : by (94) , its error is at most
C p h log Z ( Y ) Z ( Y ∗ ) ≤ C p log h h , \frac{C_p}{h}\log\frac{\mathcal{Z}(Y)}{\mathcal{Z}(Y_\ast)}\le C_p\frac{\log h}{h}, h C p log Z ( Y ∗ ) Z ( Y ) ≤ C p h log h ,
since Z \mathcal{Z} Z is decreasing, so that Z ( Y ) ≤ Z ( 0 ) \mathcal{Z}(Y)\le\mathcal{Z}(0) Z ( Y ) ≤ Z ( 0 ) , and Z ( Y ∗ ) ≍ h − 5 \mathcal{Z}(Y_\ast)\asymp h^{-5} Z ( Y ∗ ) ≍ h − 5 . The value at Y ∗ Y_\ast Y ∗ was fixed above. Exponentiation gives (89) for Y ≤ Y ∗ Y\le Y_\ast Y ≤ Y ∗ ; the direct endpoint comparison gives the stronger O p ( h − 1 ) O_p(h^{-1}) O p ( h − 1 ) error for Y ≥ Y ∗ Y\ge Y_\ast Y ≥ Y ∗ . □ \square □
At Y = 0 Y=0 Y = 0 , r = π 2 / 2 r=\pi^2/2 r = π 2 /2 , A = 1 / π 2 A=1/\pi^2 A = 1/ π 2 , Θ = 2 \Theta=2 Θ = 2 , and Z = π / ( 2 2 ) \mathcal{Z}=\pi/(2\sqrt{2}) Z = π / ( 2 2 ) . Taking p = 1 p=1 p = 1 in (89) gives
V g , 1 = Γ ( 2 g − 1 ) ( 4 π 2 ) 2 g − 2 π g ( 1 + O ( log ( g + 1 ) g ) ) . (103) V_{g,1}=\frac{\Gamma(2g-1)(4\pi^2)^{2g-2}}{\sqrt{\pi g}}\left(1+O\left(\frac{\log(g+1)}{g}\right)\right). \tag*{(103)} V g , 1 = π g Γ ( 2 g − 1 ) ( 4 π 2 ) 2 g − 2 ( 1 + O ( g log ( g + 1 ) ) ) . (103)
Proof of Corollary 1.2. Comparing (103) with the expansion of Mirzakhani–Zograf [147, Theorem 1.2] at n = 1 n=1 n = 1 identifies their universal constant as 1 / π 1/\sqrt{\pi} 1/ π . Since the constant does not depend on n n n , their expansion then gives the asserted formula, with relative error O n ( g − 1 ) O_n(g^{-1}) O n ( g − 1 ) , for every fixed n ≥ 0 n\ge0 n ≥ 0 . □ \square □
For completeness, we derive the boundary-length consequence stated in the introduction. Fix n ≥ 1 n\ge1 n ≥ 1 , put x i = L i / 2 x_i=L_i/2 x i = L i /2 , and set S ( x ) = sinh ( x ) / x S(x)=\sinh(x)/x S ( x ) = sinh ( x ) / x , with S ( 0 ) = 1 S(0)=1 S ( 0 ) = 1 . At Y = 0 Y=0 Y = 0 , the normalization (64) gives c g , n ( d ) = [ τ d 1 ⋯ τ d n ] g , n / V g , n c_{g,n}(\mathbf{d})=[\tau_{d_1}\cdots\tau_{d_n}]_{g,n}/V_{g,n} c g , n ( d ) = [ τ d 1 ⋯ τ d n ] g , n / V g , n . Thus the volume polynomial, extended by zero outside its degree range, is
V g , n ( L ) V g , n = ∑ d ≥ 0 c g , n ( d ) ∏ i = 1 n x i 2 d i ( 2 d i + 1 ) ! . \frac{V_{g,n}(\mathbf{L})}{V_{g,n}}=\sum_{\mathbf{d}\ge0}c_{g,n}(\mathbf{d})\prod_{i=1}^{n}\frac{x_i^{2d_i}}{(2d_i+1)!}. V g , n V g , n ( L ) = d ≥ 0 ∑ c g , n ( d ) i = 1 ∏ n ( 2 d i + 1 )! x i 2 d i .
Divide by ∏ i S ( x i ) \prod_iS(x_i) ∏ i S ( x i ) and regard the product weights as the law of independent nonnegative integers d i d_i d i . The identity ∑ d ≥ 0 ( 2 d ) ( 2 d + 1 ) x 2 d / ( 2 d + 1 ) ! = x 2 S ( x ) \sum_{d\ge0}(2d)(2d+1)x^{2d}/(2d+1)!=x^2S(x) ∑ d ≥ 0 ( 2 d ) ( 2 d + 1 ) x 2 d / ( 2 d + 1 )! = x 2 S ( x ) gives E [ d i ( d i + 1 ) ] ≤ x i 2 / 2 \mathbb{E}[d_i(d_i+1)]\le x_i^2/2 E [ d i ( d i + 1 )] ≤ x i 2 /2 . Since D ( D + 1 ) ≤ n ∑ i d i ( d i + 1 ) D(D+1)\le n\sum_i d_i(d_i+1) D ( D + 1 ) ≤ n ∑ i d i ( d i + 1 ) , where D = ∑ i d i D=\sum_i d_i D = ∑ i d i , (78) , including its one-marking case, yields
0 ≤ 1 − V g , n ( L ) V g , n ∏ i S ( L i / 2 ) ≤ C n g ∑ i = 1 n L i 2 . (104) 0\le1-\frac{V_{g,n}(\mathbf{L})}{V_{g,n}\prod_iS(L_i/2)}\le\frac{C_n}{g}\sum_{i=1}^{n}L_i^2. \tag*{(104)} 0 ≤ 1 − V g , n ∏ i S ( L i /2 ) V g , n ( L ) ≤ g C n i = 1 ∑ n L i 2 . (104)
For L i = o ( g ) L_i=o(\sqrt{g}) L i = o ( g ) the right side tends to zero. Corollary 1.2 and Stirling’s formula then give the asserted asymptotic formula.
The leading term J g , p \mathcal{J}_{g,p} J g , p of Proposition 7.1 is carried forward in two directions. § 3.22 refines the estimate (89) to an expansion in 1 / h 1/h 1/ h , and § 3.25 –§ 3.31 extract from J g , 1 \mathcal{J}_{g,1} J g , 1 the leading term V ^ g , n \widehat{V}_{g,n} V g , n of Theorem 1.1; the identities (92) , (94) and (96) are used again there, in (136) and in the saddle-displacement estimate of Subsection 11.1.
Higher-order corrections for marked generating functions
The estimate of Proposition 7.1 has relative error O p ( log g / g ) O_p(\log g/g) O p ( log g / g ) . The logarithm comes from integrating the error O p ( 1 / ( h A r ) ) O_p(1/(hAr)) O p ( 1/ ( h A r )) in (96) towards the critical endpoint, where A r Ar A r has a simple zero. This section states two refinements, still for a fixed number p p p of distinguished markings, obtained by iterating the dilaton and KdV equations of § 3.19 to higher order: an explicit first correction (Theorem 8.1 ), which also gives the leading term with relative error O p ( g − 1 ) O_p(g^{-1}) O p ( g − 1 ) , and a complete expansion whose truncation order may grow linearly with $g ({xref}Theorem 8.2). Both results are proved in {xref}Appendix D, with the endpoint constants again supplied by Painlevé I. They are the input of the extraction theorems of {xref}Section 10`, which make the passage to a prescribed cusp count.
The first correction
The mechanism is that of § 3.19 , carried one order further. Lemma D.1 extends the degree expansion (90) of the normalized profiles by a term e p ( d ) / h 2 e_p(\mathbf{d})/h^2 e p ( d ) / h 2 , uniformly for ∣ d ∣ ≤ h 1 / 8 |\mathbf{d}|\le h^{1/8} ∣ d ∣ ≤ h 1/8 . Inserted into the dilaton identity, it adds a term γ p / h \gamma_p/h γ p / h to the numerator of the differential equation (93) ; see (D.8). The coefficient γ p \gamma_p γ p , defined in (D.7), is a finite expression in kernel moments (see (D.11)), and it vanishes at the pure-ψ \psi ψ endpoint (see (D.12)). Its primitive against 1 / ( A r ) 1/(A r) 1/ ( A r ) is therefore bounded, although A r A r A r vanishes there. The integration constant is fixed at the endpoint by two inputs: the first correction to C g ( 23 g − 3 ) C_g(23^{g-3}) C g ( 2 3 g − 3 ) , which we derive from Painlevé I in Subsection D.11, and the first correction to an explicit negative-binomial coefficient. Accordingly, define
t p = − 2 p 2 + 12 p − 11 20 , L p ( Y ) = t p − ∫ Y y 0 γ p ( v ) A ( v ) r ( v ) d v . (105) t_p=\frac{-2p^2+12p-11}{20},\qquad\mathcal{L}_p(Y)=t_p-\int_Y^{y_0}\frac{\gamma_p(v)}{A(v)r(v)}\,\mathrm{d}v. \tag*{(105)} t p = 20 − 2 p 2 + 12 p − 11 , L p ( Y ) = t p − ∫ Y y 0 A ( v ) r ( v ) γ p ( v ) d v . (105)
Theorem 8.1 (Uniform marked first correction). For every fixed p ≥ 1 p\ge1 p ≥ 1 , uniformly for 0 ≤ Y < y 0 0\le Y<y_0 0 ≤ Y < y 0 ,
Φ g ( p ) ( X ( Y ) ) = J g , p ( Y ) π { 1 + L p ( Y ) h + O p ( log ( g + 1 ) g 2 ) } . (106) \Phi_g^{(p)}(X(Y))=\frac{\mathcal{J}_{g,p}(Y)}{\pi}\left\{1+\frac{\mathcal{L}_p(Y)}{h}+O_p\left(\frac{\log(g+1)}{g^2}\right)\right\}. \tag*{(106)} Φ g ( p ) ( X ( Y )) = π J g , p ( Y ) { 1 + h L p ( Y ) + O p ( g 2 log ( g + 1 ) ) } . (106) In particular the leading relative error is O p ( g − 1 ) O_p(g^{-1}) O p ( g − 1 ) .
At Y = 0 Y=0 Y = 0 the first-order coefficients of the cusp ratio, the genus ratio and the degree profile that enter this expansion reduce to the formulas of Mirzakhani and Zograf [147, Theorems 1.4 and 3.2(i)]; this check is recorded at the end of Subsection D.1.
All orders
The same scheme continues to every order. The recursion is triangular: each new degree coefficient of the profiles uses earlier cusp and genus coefficients, and the dilaton and KdV equations then give the new cusp and genus coefficients. The difficulty is to control the constants as the order grows.
Controlling the expansion order. The aim is to lose only one factorial per order. Polynomial degree estimates alone lead to two factorials, because each recursion step both counts possible degree splits and sums along a coordinate path. Proposition D.9 uses an exponential degree norm instead. A small loss of degree exponent absorbs the path factor at a cost proportional to the order, and in separating terms the dimension of the bounded component controls its node degree independently of the parent degree. Together these estimates give one factorial per expansion order. A uniform Gamma envelope controls the omitted splittings through a linear order range, so the same bound applies to the actual remainders and not only to the formal coefficients. Endpoint transport and coefficient extraction preserve this bound.
Appendix D constructs, by the finite triangular rules (D.13)–(D.18), the formal coefficients of the profiles and of the cusp and genus ratios, and from them functions a p , j ( Y ) a_{p,j}(Y) a p , j ( Y ) and
L p , j ( Y ) = [ t j ] exp ( ∑ i = 1 j a p , i ( Y ) t i ) , L p , 1 = L p ; \mathcal{L}_{p,j}(Y)=[t^j]\exp\left(\sum_{i=1}^{j}a_{p,i}(Y)t^i\right),\qquad\mathcal{L}_{p,1}=\mathcal{L}_p; L p , j ( Y ) = [ t j ] exp ( i = 1 ∑ j a p , i ( Y ) t i ) , L p , 1 = L p ;
see (D.50). These functions are holomorphic on one complex neighborhood of [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] and satisfy ∥ a p , j ∥ + ∥ L p , j ∥ ≤ C p B p j j ! \|a_{p,j}\|+\|\mathcal{L}_{p,j}\|\le C_p\mathfrak{B}_p^j j! ∥ a p , j ∥ + ∥ L p , j ∥ ≤ C p B p j j ! .
Theorem 8.2 (Complete uniform marked expansion). For each fixed p ≥ 1 p\ge1 p ≥ 1 there are C p , B p , c p , h p > 0 C_p,\mathfrak{B}_p,c_p,h_p>0 C p , B p , c p , h p > 0 such that, for h = 2 g − 2 + p ≥ h p h=2g-2+p\ge h_p h = 2 g − 2 + p ≥ h p and 0 ≤ K ≤ c p h 0\le K\le c_ph 0 ≤ K ≤ c p h ,
sup 0 ≤ Y < y 0 ∣ π Φ g ( p ) ( X ( Y ) ) J g , p ( Y ) − ∑ j = 0 K L p , j ( Y ) h j ∣ ≤ C p B p K + 1 ( K + 1 ) ! h K + 1 , L p , 0 = 1. (107) \sup_{0\le Y<y_0}\left|\frac{\pi\Phi_g^{(p)}(X(Y))}{\mathcal{J}_{g,p}(Y)}-\sum_{j=0}^{K}\frac{\mathcal{L}_{p,j}(Y)}{h^j}\right|
\le\frac{C_p\mathfrak{B}_p^{K+1}(K+1)!}{h^{K+1}},\qquad\mathcal{L}_{p,0}=1. \tag*{(107)} 0 ≤ Y < y 0 sup J g , p ( Y ) π Φ g ( p ) ( X ( Y )) − j = 0 ∑ K h j L p , j ( Y ) ≤ h K + 1 C p B p K + 1 ( K + 1 )! , L p , 0 = 1. (107) The coefficients are specified by the finite rules and integrals of Appendix D, and have the Gevrey-1 norms stated above on one complex neighborhood of [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] . The same error bound on the real parameter interval holds for log ( π Φ g ( p ) / J g , p ) \log(\pi\Phi_g^{(p)}/\mathcal{J}_{g,p}) log ( π Φ g ( p ) / J g , p ) , with coefficients a p , j a_{p,j} a p , j and zero constant term. A sufficiently small linear-order truncation has error O p ( e − a p h ) O_p(e^{-a_ph}) O p ( e − a p h ) ; no convergence of the infinite series is required.
The proof has three steps. First, Proposition D.9 normalizes the profiles by ( p + M j ) j e σ D (p+Mj)^{j}e^{\sigma D} ( p + M j ) j e σ D and the scalar ratios by ( p + M j ) j (p+Mj)^{j} ( p + M j ) j . After this normalization the coefficient rules become coefficientwise inequalities between power series with nonnegative coefficients, and a system of analytic majorants (Lemma D.2) bounds the normalized coefficients and the actual remainders, uniformly in the order. Next, at the pure-ψ \psi ψ endpoint, the integration constants are computed from the Itzykson–Zuber identity and the large-order behavior of Painlevé I; the Borel representation of Costin, Costin and Huang [48] gives these constants with a Gevrey-1 remainder. Finally, the logarithmic derivative is integrated from a matching point at distance h − K − 2 h^{-K-2} h − K − 2 from the endpoint, where the positive homogeneous structure of the amplitudes controls the change from the endpoint value.
At Y = 0 Y=0 Y = 0 this is the fixed-cusp expansion mechanism of [147], for which effective order-dependent bounds are given by Hide–Macera–Thomas [96, Theorem 4.2] and single-factorial bounds by Guillermit [81, Theorems 2 and 4]; the comparison with these bounds was made in the survey at the beginning of this part. What is used below is the uniformity in the density parameter through the pure-ψ \psi ψ endpoint: Theorem 8.2 is applied at a fixed order in the proof of Theorem 10.1 and over the whole linear order range in that of Theorem 10.2 , and the linear-order truncation in Theorem 1.1 is inherited from (107) .
Extracting a prescribed cusp count
A generating-function estimate must be localized to recover a volume with a prescribed cusp count. Positivity in Y Y Y (Lemma 2.2) yields the exact convolution factorization below and its Fourier bound (9.2). Combined with the marked ratios of Proposition 6.2 and the endpoint expansions of Lemma 3.1, it gives Proposition 9.2, uniformly across small, interior and near-critical cusp weights.
The volume-ratio limits and bounded-boundary factorization that follow use only the leading estimates of Sections 6–7. Higher-order marked expansions enter only when we seek inverse-genus errors, in Sections 10 and 11.
A positive convolution factor.
Lemma 9.1 (Positive convolution factor). For each stable marked pair ( g , p ) (g,p) ( g , p ) with p ≥ 1 p \ge1 p ≥ 1 , the marked generating function has the exact factorization
Φ g ( p ) ( x ) = V ( x ) h R g , p ( x ) , V ( x ) = Y ′ ( x ) = 1 / a ( Y ( x ) ) , h = 2 g − 2 + p , (108) \Phi_g^{(p)}(x)=V(x)^hR_{g,p}(x),\qquad V(x)=Y'(x)=1/a(Y(x)),\qquad h=2g-2+p, \tag*{(108)} Φ g ( p ) ( x ) = V ( x ) h R g , p ( x ) , V ( x ) = Y ′ ( x ) = 1/ a ( Y ( x )) , h = 2 g − 2 + p , (108)
where R g , p R_{g,p} R g , p has nonnegative Taylor coefficients and converges for 0 ≤ x < x c 0 \le x < x_c 0 ≤ x < x c .
A relative approximation of Φ g ( p ) \Phi_g^{(p)} Φ g ( p ) on the real axis does not, in general, give a relative approximation of its individual coefficients. The factorization (108) supplies the missing structure: under the cusp weight x x x , the number of additional cusps is distributed as the sum of h h h independent variables with generating function V ( x z ) / V ( x ) V(xz)/V(x) V ( x z ) / V ( x ) and one further independent nonnegative integer variable. This yields the Fourier bound (9.2) below, which is the input a local limit theorem needs.
Proof. The proof uses positivity in the parameter: by Lemma 2.2, the functions ϑ ^ ℓ ( Y ) = a ( Y ) [ z ℓ ] K Y ( z ) − 1 \widehat{\vartheta}_\ell(Y)=a(Y)[z^\ell]K_Y(z)^{-1} ϑ ℓ ( Y ) = a ( Y ) [ z ℓ ] K Y ( z ) − 1 and 1 / a ( Y ) 1/a(Y) 1/ a ( Y ) have nonnegative Taylor coefficients in Y Y Y . Since this positivity in Y Y Y is available for the unrescaled coefficients, we do not use the rescaling by r ( Y ) r(Y) r ( Y ) of (4.6) here.
Undo the normalization in (4.8) and regard the unrescaled coefficients ϑ ℓ = [ z ℓ ] K Y ( z ) − 1 \vartheta_\ell=[z^\ell]K_Y(z)^{-1} ϑ ℓ = [ z ℓ ] K Y ( z ) − 1 as independent variables. In this convention the bases are U 0 , 3 = ϑ 0 U_{0,3}=\vartheta_0 U 0 , 3 = ϑ 0 , U 1 , 1 ( 1 ) = ϑ 0 / 8 U_{1,1}(1)=\vartheta_0/8 U 1 , 1 ( 1 ) = ϑ 0 /8 , and U 1 , 1 ( 0 ) = ϑ 1 / 8 U_{1,1}(0)=\vartheta_1/8 U 1 , 1 ( 0 ) = ϑ 1 /8 . Induction from these bases shows that every amplitude is a finite polynomial with nonnegative coefficients in the variable ϑ ℓ \vartheta_\ell ϑ ℓ ; each monomial has total degree h h h , and weighted degree m m m when ϑ ℓ \vartheta_\ell ϑ ℓ is given weight ℓ \ell ℓ . A linear recursion term adds one factor ϑ ℓ \vartheta_\ell ϑ ℓ to a child of complexity h − 1 h-1 h − 1 ; a separating term adds one to children whose complexities sum to h − 1 h-1 h − 1 . Thus factoring a − 1 a^{-1} a − 1 from each of the h h h factors ϑ ℓ = a − 1 ϑ ^ ℓ \vartheta_\ell=a^{-1}\widehat{\vartheta}_\ell ϑ ℓ = a − 1 ϑ ℓ leaves a positive polynomial in the ϑ ^ ℓ \widehat{\vartheta}_\ell ϑ ℓ .
Next, Y ( x ) Y(x) Y ( x ) has nonnegative coefficients and Y ( 0 ) = 0 Y(0)=0 Y ( 0 ) = 0 . To justify the product used here, note that a nonzero zero y y y of X X X gives u ( s ) = J 1 ( 2 y s ) u(s)=J_1(2\sqrt{ys}) u ( s ) = J 1 ( 2 y s ) , with u ( 1 ) = 0 u(1)=0 u ( 1 ) = 0 and regular behavior at zero. Integration of ( s u ′ ) ′ + ( 4 y s − 1 / s ) u = 0 (su')' + (4ys - 1/s)u = 0 ( s u ′ ) ′ + ( 4 y s − 1/ s ) u = 0 against its conjugate gives 4 y ∫ 0 1 s ∣ u ∣ 2 d s = ∫ 0 1 ( s ∣ u ′ ∣ 2 + ∣ u ∣ 2 / s ) d s > 0 4y\int_0^1 s|u|^2\,ds = \int_0^1 (s|u'|^2 + |u|^2/s)\,ds > 0 4 y ∫ 0 1 s ∣ u ∣ 2 d s = ∫ 0 1 ( s ∣ u ′ ∣ 2 + ∣ u ∣ 2 / s ) d s > 0 . Thus all nonzero zeros are positive real. Since X X X has order one half and a simple zero at zero with derivative one, its Hadamard product is X ( Y ) = Y ∏ k ( 1 − Y / β k ) X(Y) = Y\prod_k(1 - Y/\beta_k) X ( Y ) = Y ∏ k ( 1 − Y / β k ) , β k > 0 \beta_k > 0 β k > 0 . Lagrange inversion now gives
[ x k ] Y ( x ) = 1 k [ Y k − 1 ] ( Y / X ( Y ) ) k > 0. [x^k]Y(x) = \frac{1}{k}[Y^{k-1}](Y/X(Y))^k > 0. [ x k ] Y ( x ) = k 1 [ Y k − 1 ] ( Y / X ( Y ) ) k > 0.
The inverse series has radius exactly x c x_c x c : a smaller radius would, by positivity and Pringsheim’s theorem, give a positive real singularity before x c x_c x c , contradicting the analytic real inverse there; the quadratic critical point at x c x_c x c supplies the opposite inequality. Since K Y ( z ) / a ( Y ) = 1 + ∑ ℓ ≥ 1 X ( ℓ + 1 ) ( Y ) z ℓ / ( a ( Y ) ( 2 ℓ + 1 ) ! ! ) K_Y(z)/a(Y) = 1 + \sum_{\ell\ge1} X^{(\ell+1)}(Y)z^\ell/(a(Y)(2\ell+1)!!) K Y ( z ) / a ( Y ) = 1 + ∑ ℓ ≥ 1 X ( ℓ + 1 ) ( Y ) z ℓ / ( a ( Y ) ( 2 ℓ + 1 )!!) , each ϑ b \vartheta_b ϑ b is a polynomial in the ratios X ( i ) ( Y ) / a ( Y ) X^{(i)}(Y)/a(Y) X ( i ) ( Y ) / a ( Y ) ; by Lemma 2.1 these ratios and 1 / a 1/a 1/ a are analytic on ∣ Y ∣ < y 0 |Y| < y_0 ∣ Y ∣ < y 0 , so their nonnegative Taylor series converge on [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) . Composition therefore preserves coefficient positivity and converges for 0 ≤ x < x c 0 \le x < x_c 0 ≤ x < x c , since Y ( x ) < y 0 Y(x) < y_0 Y ( x ) < y 0 . Finally, by (4.7), Φ g ( p ) ( x ) = U g , p Y ( 0 p p ) \Phi_g^{(p)}(x) = U_{g,p}^Y(0_p^p) Φ g ( p ) ( x ) = U g , p Y ( 0 p p ) with Y = Y ( x ) Y = Y(x) Y = Y ( x ) , and the factor a ( Y ) − h a(Y)^{-h} a ( Y ) − h removed above equals V ( x ) h V(x)^h V ( x ) h . This proves (9.1). □
For the coefficient probability law with generating function Φ g ( p ) ( x z ) / Φ g ( p ) ( x ) \Phi_g^{(p)}(xz)/\Phi_g^{(p)}(x) Φ g ( p ) ( x z ) / Φ g ( p ) ( x ) , positivity in (9.1) gives
∣ Φ g ( p ) ( x e i θ ) Φ g ( p ) ( x ) ∣ ≤ ∣ V ( x e i θ ) V ( x ) ∣ h . (109) \left|\frac{\Phi_g^{(p)}(xe^{i\theta})}{\Phi_g^{(p)}(x)}\right|
\le\left|\frac{V(xe^{i\theta})}{V(x)}\right|^h . \tag*{(109)} Φ g ( p ) ( x ) Φ g ( p ) ( x e i θ ) ≤ V ( x ) V ( x e i θ ) h . (109)
This Fourier estimate prevents a ratio of generating functions on the real axis from concealing different local coefficient behavior.
Mean, variance and a central limit theorem
We first identify the leading center and variance scale of the cusp count under the weight x x x and prove a Gaussian limit when both the complexity and the model mean tend to infinity. The next subsection upgrades this to a local limit theorem. Recall D = x d / d x \mathcal{D} = x\,d/dx D = x d / d x and the function Ψ \Psi Ψ of Lemma 3.1 . For convenience we repeat its endpoint expansions (3.5), together with that of V = 1 / a ( Y ( x ) ) V = 1/a(Y(x)) V = 1/ a ( Y ( x )) , which follows from a ( Y ) = 1 − Y + O ( Y 2 ) a(Y) = 1 - Y + O(Y^2) a ( Y ) = 1 − Y + O ( Y 2 ) and Y ( x ) = x + O ( x 2 ) Y(x) = x + O(x^2) Y ( x ) = x + O ( x 2 ) :
x ↓ 0 : V ( x ) = 1 + x + O ( x 2 ) , Ψ ( x ) = 2 x + O ( x 2 ) , D Ψ ( x ) = 2 x + O ( x 2 ) , D 2 Ψ ( x ) = 2 x + O ( x 2 ) ; x ↑ x c : Ψ ( x ) = 5 4 ε ( 1 + O ( ε ) ) , D Ψ ( x ) = 5 4 ε 2 ( 1 + O ( ε ) ) , D 2 Ψ ( x ) = O ( ε − 3 ) , ε = 1 − x / x c . (110) \begin{aligned}
x \downarrow0:\quad&V(x) = 1 + x + O(x^2),\qquad\Psi(x) = 2x + O(x^2),\\
&\mathcal{D}\Psi(x) = 2x + O(x^2),\qquad\mathcal{D}^2\Psi(x) = 2x + O(x^2);\\
x \uparrow x_c:\quad&\Psi(x) = \frac{5}{4\varepsilon}(1 + O(\sqrt{\varepsilon})),\\
&\mathcal{D}\Psi(x) = \frac{5}{4\varepsilon^2}(1 + O(\sqrt{\varepsilon})),\qquad\mathcal{D}^2\Psi(x) = O(\varepsilon^{-3}),\qquad\varepsilon= 1 - x/x_c. \tag*{(110)}
\end{aligned} x ↓ 0 : x ↑ x c : V ( x ) = 1 + x + O ( x 2 ) , Ψ ( x ) = 2 x + O ( x 2 ) , D Ψ ( x ) = 2 x + O ( x 2 ) , D 2 Ψ ( x ) = 2 x + O ( x 2 ) ; Ψ ( x ) = 4 ε 5 ( 1 + O ( ε )) , D Ψ ( x ) = 4 ε 2 5 ( 1 + O ( ε )) , D 2 Ψ ( x ) = O ( ε − 3 ) , ε = 1 − x / x c . (110)
On compact subintervals of ( 0 , x c ) (0,x_c) ( 0 , x c ) , all needed derivatives are bounded and D Ψ \mathcal{D}\Psi D Ψ is bounded away from zero.
For F = Φ g ( p ) F = \Phi_g^{(p)} F = Φ g ( p ) at a positive cusp weight x x x , let
P x ( N = k ) = [ z k ] F ( z ) x k F ( x ) , μ = h Ψ ( x ) , σ 2 = h D Ψ ( x ) . (111) \mathbb{P}_x(N = k) = \frac{[z^k]F(z)x^k}{F(x)},\qquad\mu= h\Psi(x),\qquad\sigma^2 = h\mathcal{D}\Psi(x). \tag*{(111)} P x ( N = k ) = F ( x ) [ z k ] F ( z ) x k , μ = h Ψ ( x ) , σ 2 = h D Ψ ( x ) . (111)
The last two quantities are a model center and a model variance; they are not claimed to be the exact finite-genus mean and variance. Suppose h → ∞ h \to\infty h → ∞ and μ → ∞ \mu\to\infty μ → ∞ . For every fixed real u u u , integrating (6.20) gives
log E x e u ( N − μ ) / σ = ∫ 0 u / σ { h Ψ ( x e s ) − h Ψ ( x ) + O p ( Ψ ( x e s ) ) } d s . (112) \log\mathbb{E}_x e^{u(N-\mu)/\sigma}
= \int_0^{u/\sigma}\{h\Psi(xe^s) - h\Psi(x) + O_p(\Psi(xe^s))\}\,ds. \tag*{(112)} log E x e u ( N − μ ) / σ = ∫ 0 u / σ { h Ψ ( x e s ) − h Ψ ( x ) + O p ( Ψ ( x e s ))} d s . (112)
The interval of integration stays inside ( 0 , x c ) (0,x_c) ( 0 , x c ) . Expanding h Ψ ( x e s ) − h Ψ ( x ) = s h D Ψ ( x ) + O ( s 2 h D 2 Ψ ) h\Psi(xe^s)-h\Psi(x) = sh\mathcal{D}\Psi(x)+O(s^2h\mathcal{D}^2\Psi) h Ψ ( x e s ) − h Ψ ( x ) = s h D Ψ ( x ) + O ( s 2 h D 2 Ψ ) , the first part of the integral is u 2 / 2 u^2/2 u 2 /2 plus a Taylor remainder, and the second part is an integrated error. Both errors are o ( 1 ) o(1) o ( 1 ) . In the critical regime and on an interior compact interval, the Taylor remainder and the integrated error are O u ( h − 1 / 2 ) O_u(h^{-1/2}) O u ( h − 1/2 ) . At x → 0 x \to0 x → 0 , where σ 2 ∼ μ \sigma^2 \sim\mu σ 2 ∼ μ , the Taylor remainder is O u ( μ − 1 / 2 ) O_u(\mu^{-1/2}) O u ( μ − 1/2 ) and the integrated error is O u ( x / μ ) = O u ( μ / h ) = o ( 1 ) O_u(x/\sqrt{\mu}) = O_u(\sqrt{\mu}/h) = o(1) O u ( x / μ ) = O u ( μ / h ) = o ( 1 ) . These assertions follow directly from (110) ; no estimate for the derivative of the error in (83) is required.
Thus the real moment generating functions converge to exp ( u 2 / 2 ) \exp(u^2/2) exp ( u 2 /2 ) along every such sequence. Exponential Markov bounds give tightness; bounds at slightly larger real arguments give uniform integrability along weakly convergent subsequences, whose limiting moment generating functions then identify the standard Gaussian. Therefore
( N − μ ) / σ ⟹ N ( 0 , 1 ) . (113) (N-\mu)/\sigma\Longrightarrow\mathcal{N}(0,1). \tag*{(113)} ( N − μ ) / σ ⟹ N ( 0 , 1 ) . (113)
Fourier bounds and the local limit theorem
A central limit theorem alone does not determine ratios of individual coefficients; for these we prove the corresponding local limit theorem. Central and local limit theorems for coefficients of generating functions go back to Bender [21]; see Hwang [105] for rates in the quasi-powers setting and [69, Chapter IX] for an account. Here the base function, the power and the evaluation point vary together, up to both ends of the parameter interval, so the bounds are proved directly. The proof dominates the characteristic function by an integrable function separately at small cusp weight, at interior cusp weights, and near the critical point.
Proposition 9.2 (Uniform local limit theorem). Fix p ≥ 1 p \ge1 p ≥ 1 and use the notation (111) . Along every sequence with h → ∞ h \to\infty h → ∞ , μ → ∞ \mu\to\infty μ → ∞ and x ∈ ( 0 , x c ) x \in(0,x_c) x ∈ ( 0 , x c ) ,
P x ( N = k ) = 1 + o ( 1 ) 2 π σ e − u 2 / 2 whenever ( k − μ ) / σ ⟶ u . (114) \mathbb{P}_x(N=k) = \frac{1+o(1)}{\sqrt{2\pi}\sigma}e^{-u^2/2}\qquad\text{whenever }(k-\mu)/\sigma\longrightarrow u. \tag*{(114)} P x ( N = k ) = 2 π σ 1 + o ( 1 ) e − u 2 /2 whenever ( k − μ ) / σ ⟶ u . (114) No relation between h h h and the distance from x x x to x c x_c x c is required.
Proof. Passing to a subsequence, we may assume that x x x tends to zero, to an interior point, or to x c x_c x c . In each case we bound the characteristic function of ( N − μ ) / σ (N-\mu)/\sigma ( N − μ ) / σ by an integrable function, uniformly along the sequence; Fourier inversion and the pointwise convergence implied by (113) then give (114) .
On a compact subinterval of ( 0 , x c ) (0,x_c) ( 0 , x c ) , the coefficient probabilities at zero and one of V ( x z ) / V ( x ) V(xz)/V(x) V ( x z ) / V ( x ) have positive uniform lower bounds. If these probabilities are p 0 , p 1 p_0,p_1 p 0 , p 1 , then
∣ ∑ j p j e i j θ ∣ 2 ≤ 1 − 2 p 0 p 1 ( 1 − cos θ ) . (115) \left|\sum_j p_j e^{ij\theta}\right|^2 \le1-2p_0p_1(1-\cos\theta). \tag*{(115)} j ∑ p j e ij θ 2 ≤ 1 − 2 p 0 p 1 ( 1 − cos θ ) . (115) Equations (109) and (115) give an upper bound exp ( − c h θ 2 ) \exp(-ch\theta^2) exp ( − c h θ 2 ) on [ − π , π ] [-\pi,\pi] [ − π , π ] . Since σ \sigma σ is comparable to h \sqrt{h} h , this is integrable Gaussian domination on the rescaled Fourier axis.
At x → 0 x \to0 x → 0 , one has p 0 ≍ 1 p_0 \asymp1 p 0 ≍ 1 , p 1 ≍ x p_1 \asymp x p 1 ≍ x . The same argument gives exp ( − c h x θ 2 ) \exp(-chx\theta^2) exp ( − c h x θ 2 ) . Since h x ≍ μ hx \asymp\mu h x ≍ μ and σ 2 ∼ μ → ∞ \sigma^2 \sim\mu\to\infty σ 2 ∼ μ → ∞ , it again gives integrable domination after rescaling.
Near the critical point a different bound is needed, because the law with generating function V ( x z ) / V ( x ) V(xz)/V(x) V ( x z ) / V ( x ) then has variance of order ϵ − 2 \epsilon^{-2} ϵ − 2 and its probabilities at zero and one tend to zero. Quadratic analytic inversion at x c x_c x c gives
V ( z ) = A 0 ( 1 − z / x c ) − 1 / 2 Ξ ( 1 − z / x c ) , A 0 > 0 , Ξ ( 0 ) = 1. (116) V(z)=A_0(1-z/x_c)^{-1/2}\Xi\left(\sqrt{1-z/x_c}\right),\qquad A_0>0,\quad\Xi(0)=1. \tag*{(116)} V ( z ) = A 0 ( 1 − z / x c ) − 1/2 Ξ ( 1 − z / x c ) , A 0 > 0 , Ξ ( 0 ) = 1. (116) where Ξ \Xi Ξ is analytic and nonzero near zero and has real coefficients. Put Λ ( w ) = log Ξ ( w ) \Lambda(w)=\log\Xi(\sqrt{w}) Λ ( w ) = log Ξ ( w ) and w ( θ ) = 1 − ( 1 − ϵ ) e i θ w(\theta)=1-(1-\epsilon)e^{i\theta} w ( θ ) = 1 − ( 1 − ϵ ) e i θ . Then
log ∣ V ( x e i θ ) V ( x ) ∣ = − 1 4 log ( 1 + 2 ( 1 − ϵ ) ( 1 − cos θ ) ϵ 2 ) + ℜ Λ ( w ( θ ) ) − Λ ( ϵ ) . (117) \log\left|\frac{V(xe^{i\theta})}{V(x)}\right|=-\frac{1}{4}\log\left(1+\frac{2(1-\epsilon)(1-\cos\theta)}{\epsilon^2}\right)+\Re\Lambda(w(\theta))-\Lambda(\epsilon). \tag*{(117)} log V ( x ) V ( x e i θ ) = − 4 1 log ( 1 + ϵ 2 2 ( 1 − ϵ ) ( 1 − cos θ ) ) + ℜΛ ( w ( θ )) − Λ ( ϵ ) . (117) There are fixed δ , ϵ 0 > 0 \delta,\epsilon_0>0 δ , ϵ 0 > 0 such that
∣ V ( x e i θ ) V ( x ) ∣ ≤ ( 1 + θ 2 4 ϵ 2 ) − 1 / 8 ( ∣ θ ∣ ≤ δ , 0 < ϵ ≤ ϵ 0 ) . (118) \left|\frac{V(xe^{i\theta})}{V(x)}\right|\leq\left(1+\frac{\theta^2}{4\epsilon^2}\right)^{-1/8}\qquad(|\theta|\leq\delta,\ 0<\epsilon\leq\epsilon_0). \tag*{(118)} V ( x ) V ( x e i θ ) ≤ ( 1 + 4 ϵ 2 θ 2 ) − 1/8 ( ∣ θ ∣ ≤ δ , 0 < ϵ ≤ ϵ 0 ) . (118)
Indeed, for ∣ θ ∣ ≤ ε |\theta| \le\varepsilon ∣ θ ∣ ≤ ε , the first derivative at zero of the real part of Λ ( w ( θ ) ) \Lambda(w(\theta)) Λ ( w ( θ )) vanishes, while its second derivative is O ( ε − 3 / 2 ) O(\varepsilon^{-3/2}) O ( ε − 3/2 ) . Its change is therefore O ( θ 2 ε − 3 / 2 ) O(\theta^2\varepsilon^{-3/2}) O ( θ 2 ε − 3/2 ) , absorbed by half of the leading logarithm in (117) when ε \varepsilon ε is small. For ε ≤ ∣ θ ∣ ≤ δ \varepsilon\le|\theta| \le\delta ε ≤ ∣ θ ∣ ≤ δ , its change is bounded by C δ C\sqrt{\delta} C δ , while the logarithm is at least log ( 5 / 4 ) \log(5/4) log ( 5/4 ) . Choose δ \delta δ small enough for the same absorption. We used the elementary bound 2 ( 1 − ε ) ( 1 − cos θ ) ≥ θ 2 / 4 2(1-\varepsilon)(1-\cos\theta) \ge\theta^2/4 2 ( 1 − ε ) ( 1 − cos θ ) ≥ θ 2 /4 on the small arc.
On the outer arc δ ≤ ∣ θ ∣ ≤ π \delta\le|\theta| \le\pi δ ≤ ∣ θ ∣ ≤ π ,
∣ V ( x e i θ ) V ( x ) ∣ ≤ C δ ε . (119) \left| \frac{V(xe^{i\theta})}{V(x)} \right| \le C_{\delta}\sqrt{\varepsilon}. \tag*{(119)} V ( x ) V ( x e i θ ) ≤ C δ ε . (119)
To see that V V V is uniformly bounded there, note that nonnegative coefficients and Y ( x ) → y 0 Y(x) \to y_0 Y ( x ) → y 0 give absolute convergence of Y Y Y on ∣ x ∣ ≤ x c |x| \le x_c ∣ x ∣ ≤ x c . Its first two coefficients are positive, so strict triangle inequality gives ∣ Y ( x c e i θ ) ∣ < y 0 |Y(x_ce^{i\theta})| < y_0 ∣ Y ( x c e i θ ) ∣ < y 0 away from θ = 0 \theta=0 θ = 0 , while ∣ Y ( z ) ∣ ≤ Y ( ∣ z ∣ ) < y 0 |Y(z)| \le Y(|z|) < y_0 ∣ Y ( z ) ∣ ≤ Y ( ∣ z ∣ ) < y 0 for ∣ z ∣ < x c |z| < x_c ∣ z ∣ < x c . By continuity of Y Y Y on the closed disk, ∣ Y ∣ ≤ y 1 |Y| \le y_1 ∣ Y ∣ ≤ y 1 for some y 1 < y 0 y_1 < y_0 y 1 < y 0 on the closed sector ∣ z ∣ ≤ x c |z| \le x_c ∣ z ∣ ≤ x c , δ ≤ ∣ arg z ∣ ≤ π \delta\le|\arg z| \le\pi δ ≤ ∣ arg z ∣ ≤ π . By Lemma 2.1 , X ′ X' X ′ has no zero in ∣ Y ∣ < y 0 |Y| < y_0 ∣ Y ∣ < y 0 , so V = 1 / a ( Y ) V=1/a(Y) V = 1/ a ( Y ) is uniformly bounded on this sector. Since (116) gives V ( x ) ≍ ε − 1 / 2 V(x) \asymp\varepsilon^{-1/2} V ( x ) ≍ ε − 1/2 , (119) follows.
On the critical Fourier scale θ = ε t / h \theta=\varepsilon t/\sqrt{h} θ = εt / h , (109) and (118) are bounded by
( 1 + t 2 4 h ) − h / 8 ≤ ( 1 + t 2 64 ) − 2 , h ≥ 16. (120) \left(1+\frac{t^2}{4h}\right)^{-h/8} \le\left(1+\frac{t^2}{64}\right)^{-2}, \qquad h \ge16. \tag*{(120)} ( 1 + 4 h t 2 ) − h /8 ≤ ( 1 + 64 t 2 ) − 2 , h ≥ 16. (120)
The inequality follows because ( 1 + a / h ) h (1+a/h)^h ( 1 + a / h ) h increases with h h h . The right side is integrable. The outer contribution after multiplication by h / ε \sqrt{h}/\varepsilon h / ε is at most
2 π h C δ h ε h / 2 − 1 ⟶ 0. (121) 2\pi\sqrt{h}C_{\delta}^{h}\varepsilon^{h/2-1} \longrightarrow0. \tag*{(121)} 2 π h C δ h ε h /2 − 1 ⟶ 0. (121)
This convergence is uniform for arbitrarily small ε \varepsilon ε : fix sufficiently small ε 0 \varepsilon_0 ε 0 with C δ ε 0 < 1 C_{\delta}\sqrt{\varepsilon_0}<1 C δ ε 0 < 1 , and bound (121) by $2\pi\varepsilon_0^{-1}\sqrt{h}(C_{\delta}\sqrt{\varepsilon_0})^hf o r for f or h \ge 2.
These integrable bounds, in the three regimes, complete the proof. □
The same Fourier bounds also bound every point mass, not only those near the center. We record this for later use in Appendices G, H, L and S.
Lemma 9.3 (Uniform point-mass bound). For every c 0 > 0 c_0>0 c 0 > 0 there is a constant C C C with the following property. Let x ∈ ( 0 , x c ) x\in(0,x_c) x ∈ ( 0 , x c ) , let h , h ′ h,h' h , h ′ be positive integers with h ′ ≥ max ( 16 , c 0 h ) h' \ge\max(16,c_0h) h ′ ≥ max ( 16 , c 0 h ) , and let G ( z ) = V ( z ) h ′ R ( z ) G(z)=V(z)^{h'}R(z) G ( z ) = V ( z ) h ′ R ( z ) , where R R R has nonnegative Taylor coefficients, converges for ∣ z ∣ < x c |z|<x_c ∣ z ∣ < x c and satisfies R ( x ) > 0 R(x)>0 R ( x ) > 0 . Then, with σ 2 = h D Ψ ( x ) \sigma^2=h\mathcal{D}\Psi(x) σ 2 = h D Ψ ( x ) ,
sup j ≥ 0 [ z j ] G ( z ) x j G ( x ) ≤ C σ . \sup_{j\ge0}\frac{[z^j]G(z)x^j}{G(x)} \le\frac{C}{\sigma}. j ≥ 0 sup G ( x ) [ z j ] G ( z ) x j ≤ σ C .
Proof. By Fourier inversion and positivity of the coefficients of R R R ,
[ z j ] G ( z ) x j G ( x ) ≤ 1 2 π ∫ − π π ∣ V ( x e i θ ) V ( x ) ∣ h ′ d θ . \frac{[z^j]G(z)x^j}{G(x)} \le\frac{1}{2\pi}\int_{-\pi}^{\pi}\left|\frac{V(xe^{i\theta})}{V(x)}\right|^{h'}\,d\theta. G ( x ) [ z j ] G ( z ) x j ≤ 2 π 1 ∫ − π π V ( x ) V ( x e i θ ) h ′ d θ .
On a compact subinterval of ( 0 , x c ) (0,x_c) ( 0 , x c ) , (115) bounds the integrand by e − c h ′ θ 2 e^{-ch'\theta^2} e − c h ′ θ 2 , so the integral is O ( h − 1 / 2 ) O(h^{-1/2}) O ( h − 1/2 ) , while σ ≍ h \sigma\asymp\sqrt{h} σ ≍ h . Near zero the same argument with p 1 ≍ x p_1\asymp x p 1 ≍ x gives e − c h ′ x θ 2 e^{-ch'x\theta^2} e − c h ′ x θ 2 and the bound O ( ( h x ) − 1 / 2 ) O((hx)^{-1/2}) O (( h x ) − 1/2 ) , while σ 2 ≍ h x \sigma^2\asymp hx σ 2 ≍ h x . Near x c x_c x c , put θ = ε t / h ′ \theta=\varepsilon t/\sqrt{h'} θ = εt / h ′ on ∣ θ ∣ ≤ δ |\theta|\le\delta ∣ θ ∣ ≤ δ ; (118) and the argument of (120) bound this part by ( ε / h ′ ) ∫ ( 1 + t 2 / 64 ) − 2 d t (\varepsilon/\sqrt{h'})\int(1+t^2/64)^{-2}\,dt ( ε / h ′ ) ∫ ( 1 + t 2 /64 ) − 2 d t . On the outer arc (119) gives at most 2 π ( C δ ε ) h ′ ≤ C ε h − 1 / 2 2\pi(C_{\delta}\sqrt{\varepsilon})^{h'}\le C\varepsilon h^{-1/2} 2 π ( C δ ε ) h ′ ≤ C ε h − 1/2 once C δ ε 0 < 1 C_{\delta}\sqrt{\varepsilon_0}<1 C δ ε 0 < 1 . Since σ ≍ h / ε \sigma\asymp\sqrt{h}/\varepsilon σ ≍ h / ε there, all three regimes give O ( 1 / σ ) O(1/\sigma) O ( 1/ σ ) .
Ratios at a prescribed number of cusps
We now prove the limits in Theorem 1.3 . Each volume ratio is a ratio of coefficients, at the same index, of two marked generating functions. We evaluate both at the same cusp weight, chosen so that the index is, up to a negligible shift, the model center of both. The local limit theorem (114) shows that the two local probabilities are asymptotically equal, so the coefficient ratio is asymptotic to the ratio of the generating functions themselves, which Section 6 determines.
Uniform convergence in Theorem 1.3 . First suppose g , n → ∞ g,n \to\infty g , n → ∞ . Put
k = n − 1 , h = 2 g − 1 , h Ψ ( x ) = k . (122) k=n-1,\qquad h=2g-1,\qquad h\Psi(x)=k. \tag*{(122)} k = n − 1 , h = 2 g − 1 , h Ψ ( x ) = k . (122)
The last equation has a unique solution in ( 0 , x c ) (0,x_c) ( 0 , x c ) by (85) and the parameter bijection. The two families Φ g ′ \Phi'_g Φ g ′ and Φ g − 1 ′ ′ ′ \Phi'''_{g-1} Φ g − 1 ′′′ have exactly the same complexity h h h , model center, and model variance. Thus their two local probabilities in (114) , at the common coefficient index k k k , have ratio tending to one. Equations (84) and (114) give
[ x k ] Φ g − 1 ′ ′ ′ ( x ) [ x k ] Φ g ′ ( x ) = 4 r ( Y ) ( 1 + o ( 1 ) ) . (123) \frac{[x^k]\Phi'''_{g-1}(x)}{[x^k]\Phi'_g(x)}=4r(Y)(1+o(1)). \tag*{(123)} [ x k ] Φ g ′ ( x ) [ x k ] Φ g − 1 ′′′ ( x ) = 4 r ( Y ) ( 1 + o ( 1 )) . (123)
By the definition (1) , the coefficients are
[ x k ] Φ g ′ ( x ) = V g , n ( 2 π 2 ) 3 g − 3 + n k ! , [ x k ] Φ g − 1 ′ ′ ′ ( x ) = V g − 1 , n + 2 ( 2 π 2 ) 3 g − 4 + n k ! . (124) [x^k]\Phi'_g(x)=\frac{V_{g,n}}{(2\pi^2)^{3g-3+n}k!},\qquad[x^k]\Phi'''_{g-1}(x)=\frac{V_{g-1,n+2}}{(2\pi^2)^{3g-4+n}k!}. \tag*{(124)} [ x k ] Φ g ′ ( x ) = ( 2 π 2 ) 3 g − 3 + n k ! V g , n , [ x k ] Φ g − 1 ′′′ ( x ) = ( 2 π 2 ) 3 g − 4 + n k ! V g − 1 , n + 2 . (124)
Consequently
B g , n = π 2 2 r ( Y ) ( 1 + o ( 1 ) ) = B ( 2 ( n − 1 ) 2 g − 1 ) ( 1 + o ( 1 ) ) . (125) B_{g,n}=\frac{\pi^2}{2r(Y)}(1+o(1))=\mathcal{B}\left(\frac{2(n-1)}{2g-1}\right)(1+o(1)). \tag*{(125)} B g , n = 2 r ( Y ) π 2 ( 1 + o ( 1 )) = B ( 2 g − 1 2 ( n − 1 ) ) ( 1 + o ( 1 )) . (125)
The factorial k ! k! k ! cancels, and the powers of 2 π 2 2\pi^2 2 π 2 differ by exactly one; this explains the factor in (125) .
For the cusp ratio, compare Φ g ′ ′ \Phi''_g Φ g ′′ with Φ g ′ \Phi'_g Φ g ′ at the same index k k k and the same cusp weight (122) . Their complexities are h + 1 h+1 h + 1 and h h h . Their centers differ by Ψ ( x ) \Psi(x) Ψ ( x ) , and their variances have ratio ( h + 1 ) / h (h+1)/h ( h + 1 ) / h . Moreover
Ψ ( x ) h D Ψ ( x ) = O ( h − 1 / 2 ) (126) \frac{\Psi(x)}{\sqrt{h\mathcal{D}\Psi(x)}}=O(h^{-1/2}) \tag*{(126)} h D Ψ ( x ) Ψ ( x ) = O ( h − 1/2 ) (126)
uniformly in x x x , because Ψ 2 / ( D Ψ ) \Psi^2/(\mathcal{D}\Psi) Ψ 2 / ( D Ψ ) is continuous in the interior and has finite endpoint limits zero and 5 / 4 5/4 5/4 , respectively. Thus the local probability ratio again tends to one. Undoing (79) ,
C g , n = π 2 g [ x k ] Φ g ′ ′ [ x k ] Φ g ′ = π 2 g h A a r ( 1 + o ( 1 ) ) = π 2 ( n − 1 ) g x ( 1 + o ( 1 ) ) . (127) C_{g,n}=\frac{\pi^2}{g}\frac{[x^k]\Phi''_g}{[x^k]\Phi'_g}=\frac{\pi^2}{g}\frac{h}{\mathop{\mathrm{Aar}}}(1+o(1))=\frac{\pi^2(n-1)}{gx}(1+o(1)). \tag*{(127)} C g , n = g π 2 [ x k ] Φ g ′ [ x k ] Φ g ′′ = g π 2 Aar h ( 1 + o ( 1 )) = g x π 2 ( n − 1 ) ( 1 + o ( 1 )) . (127)
Here, as in (124) , the coefficients [ x k ] Φ g ′ ′ [x^k]\Phi''_g [ x k ] Φ g ′′ and [ x k ] Φ g ′ [x^k]\Phi'_g [ x k ] Φ g ′ carry the normalizing powers ( 2 π 2 ) 3 g − 2 + n (2\pi^2)^{3g-2+n} ( 2 π 2 ) 3 g − 2 + n and ( 2 π 2 ) 3 g − 3 + n (2\pi^2)^{3g-3+n} ( 2 π 2 ) 3 g − 3 + n ; division by 2 g 2g 2 g gives the factor π 2 / g \pi^2/g π 2 / g . With λ e f f = 2 ( n − 1 ) / ( 2 g − 1 ) \lambda_{\mathrm{eff}}=2(n-1)/(2g-1) λ eff = 2 ( n − 1 ) / ( 2 g − 1 ) , the last expression equals
( 1 − 1 2 g ) C ( λ e f f ) ( 1 + o ( 1 ) ) . (128) \left(1-\frac{1}{2g}\right)\mathcal{C}(\lambda_{\mathrm{eff}})(1+o(1)). \tag*{(128)} ( 1 − 2 g 1 ) C ( λ eff ) ( 1 + o ( 1 )) . (128)
One may replace λ e f f \lambda_{\mathrm{eff}} λ eff by n / g n/g n / g in both (125) and (128) . At a zero limiting slope, use the continuous values B ( 0 ) = 1 \mathcal{B}(0)=1 B ( 0 ) = 1 and C ( 0 ) = 4 π 2 \mathcal{C}(0)=4\pi^2 C ( 0 ) = 4 π 2 of (7) ; at a finite positive slope use continuity; and at an infinite slope use the endpoint equivalents (19) of Lemma 3.1 . Since λ e f f / ( n / g ) → 1 \lambda_{\mathrm{eff}}/(n/g)\to1 λ eff / ( n / g ) → 1 whenever n → ∞ n\to\infty n → ∞ , these give the desired relative replacement in all cases.
Finally the exact identity
D g , n = C g , n − 1 C g , n (129) D_{g,n}=\frac{C_{g,n-1}}{C_{g,n}} \tag*{(129)} D g , n = C g , n C g , n − 1 (129)
and the adjacent continuity of the profile C \mathcal{C} C give D g , n → 1 D_{g,n}\to1 D g , n → 1 for every growing-cusp sequence. For bounded cusp counts, the classical fixed-cusp theorems give B g , n → 1 B_{g,n}\to1 B g , n → 1 , C g , n → 4 π 2 C_{g,n}\to4\pi^2 C g , n → 4 π 2 , and hence D g , n → 1 D_{g,n}\to1 D g , n → 1 ; these are the only bounded-cusp inputs used here.
If any supremum in Theorem 1.3 did not tend to zero, select a violating sequence of cusp counts. A subsequence either has a fixed bounded cusp count or tends to infinity. The preceding conclusions contradict a fixed violation in both cases. This proves the uniform convergence assertions. Subsection 11.1 supplies the stated inverse-genus rate using Theorem 10.1 .
Bounded geodesic boundaries
The positive convolution factor also allows a fixed number of cusps to be replaced by geodesic boundaries of bounded length. Each boundary then contributes a factor that depends on the cusp density and connects the familiar large-genus and large-cusp boundary profiles. For 0 ≤ Y ≤ y 0 0 \le Y \le y_0 0 ≤ Y ≤ y 0 and L ≥ 0 L \ge0 L ≥ 0 , put
b Y ( L ) = ∫ 0 1 I 0 ( L π T ( Y ) − Δ ( Y ) s 2 ) d s , (130) b_Y(L)=\int_0^1 I_0\left(\frac{L}{\pi}\sqrt{T(Y)-\Delta(Y)s^2}\right)\,ds, \tag*{(130)} b Y ( L ) = ∫ 0 1 I 0 ( π L T ( Y ) − Δ ( Y ) s 2 ) d s , (130)
where I 0 I_0 I 0 is the modified Bessel function of order zero, T = Y + Δ T=Y+\Delta T = Y + Δ and r = 2 Δ r=2\Delta r = 2Δ . At Y = y 0 Y=y_0 Y = y 0 use T = y 0 T=y_0 T = y 0 and Δ = 0 \Delta=0 Δ = 0 . Write Y λ Y_\lambda Y λ for the inverse of λ ( Y ) \lambda(Y) λ ( Y ) ; thus Y 0 = 0 Y_0=0 Y 0 = 0 and Y ∞ = y 0 Y_\infty=y_0 Y ∞ = y 0 .
Proposition 9.4 (Boundary lengths at arbitrary cusp density). Fix an integer p ≥ 1 p \ge1 p ≥ 1 and L 0 < ∞ L_0 < \infty L 0 < ∞ . As g → ∞ g \to\infty g → ∞ ,
sup n ≥ p L ∈ [ 0 , L 0 ] p ∣ V g , n ( L 1 , … , L p , 0 n − p ) V g , n ∏ i = 1 p b Y n / g ( L i ) − 1 ∣ ⟶ 0. (131) \sup_{\substack{n\ge p\\ \mathbf{L}\in[0,L_0]^p}}\left|\frac{V_{g,n}(L_1,\ldots,L_p,0^{n-p})}{V_{g,n}\prod_{i=1}^{p}b_{Y_{n/g}}(L_i)}-1\right|\longrightarrow0. \tag*{(131)} n ≥ p L ∈ [ 0 , L 0 ] p sup V g , n ∏ i = 1 p b Y n / g ( L i ) V g , n ( L 1 , … , L p , 0 n − p ) − 1 ⟶ 0. (131)
The endpoint factors are
b 0 ( L ) = sinh ( L / 2 ) L / 2 , b y 0 ( L ) = I 0 ( j 0 , 1 L 2 π ) , (132) b_0(L)=\frac{\sinh(L/2)}{L/2},\qquad b_{y_0}(L)=I_0\left(\frac{j_{0,1}L}{2\pi}\right), \tag*{(132)} b 0 ( L ) = L /2 sinh ( L /2 ) , b y 0 ( L ) = I 0 ( 2 π j 0 , 1 L ) , (132)
with value one at L = 0 L=0 L = 0 .
The first endpoint factor occurs in the fixed-cusp estimate of Mirzakhani and Petri [146, Proposition 3.1]; the second is the fixed-genus, large-cusp factor of Hide and Thomas [98, Theorem 1.5]. Here the genus tends to infinity and the cusp count is unrestricted, while the number and lengths of the nonzero boundaries remain bounded. For comparison, Budd and Lions obtain a boundary factor for tight volumes with a random cusp count in a near-critical regime [32, Proposition 4.2]. Exact transformations between ordinary and tight boundary generating functions were established in [34].
Proof. Put h = 2 g − 2 + p h=2g-2+p h = 2 g − 2 + p , F ( x ) = Φ g ( p ) ( x ) F(x)=\Phi_g^{(p)}(x) F ( x ) = Φ g ( p ) ( x ) and
G ( x ; L ) = ∑ k ≥ 0 V g , p + k ( L , 0 k ) ( 2 π 2 ) 3 g − 3 + p + k x k k ! . G(x;\mathbf{L})=\sum_{k\ge0}\frac{V_{g,p+k}(\mathbf{L},0^k)}{(2\pi^2)^{3g-3+p+k}}\frac{x^k}{k!}. G ( x ; L ) = k ≥ 0 ∑ ( 2 π 2 ) 3 g − 3 + p + k V g , p + k ( L , 0 k ) k ! x k .
We first compare G G G with F F F on the real axis, and then show that the same comparison holds for their coefficients at the saddle.
The boundary insertion. Use the string flow of (27) , holding its parameter Y Y Y fixed while differentiating the KW times. Its linear part gives
∂ t d = ∑ b = 0 d Y d − b ( d − b ) ! ∂ t b ′ . \partial_{t_d}=\sum_{b=0}^{d}\frac{Y^{d-b}}{(d-b)!}\partial_{t_b}'. ∂ t d = b = 0 ∑ d ( d − b )! Y d − b ∂ t b ′ .
For g ≥ 2 g\ge2 g ≥ 2 the free energy is invariant under the flow. Apply the displayed identity to each external insertion before setting x = X ( Y ) x=X(Y) x = X ( Y ) . In the volume normalization of (24) , an external degree d d d has coefficient ( L 2 / ( 4 π 2 ) ) d / d ! (L^2/(4\pi^2))^d/d! ( L 2 / ( 4 π 2 ) ) d / d ! times the unweighted derivative. Consequently, with t i = L i 2 / ( 4 π 2 ) t_i=L_i^2/(4\pi^2) t i = L i 2 / ( 4 π 2 ) , the exact identity is
G ( X ( Y ) ; L ) F ( X ( Y ) ) = ∑ a , b ≥ 0 c g , p ( b ; Y ) ∏ i = 1 p t i a i + b i Y a i r b i a i ! ( a i + b i ) ! ( 2 b i + 1 ) ! ! . (133) \frac{G(X(Y);\mathbf{L})}{F(X(Y))}=\sum_{\mathbf{a},\mathbf{b}\ge0}c_{g,p}(\mathbf{b};Y)\prod_{i=1}^{p}\frac{t_i^{a_i+b_i}Y^{a_i}r^{b_i}}{a_i!(a_i+b_i)!(2b_i+1)!!}. \tag*{(133)} F ( X ( Y )) G ( X ( Y ) ; L ) = a , b ≥ 0 ∑ c g , p ( b ; Y ) i = 1 ∏ p a i ! ( a i + b i )! ( 2 b i + 1 )!! t i a i + b i Y a i r b i . (133)
Indeed, (64) gives U g , p Y ( b ) / U g , p Y ( 0 ) = r ∣ b ∣ c g , p ( b ; Y ) U_{g,p}^{Y}(\mathbf{b})/U_{g,p}^{Y}(0)=r^{|\mathbf{b}|}c_{g,p}(\mathbf{b};Y) U g , p Y ( b ) / U g , p Y ( 0 ) = r ∣ b ∣ c g , p ( b ; Y ) . Terms with invalid b \mathbf{b} b are zero; there is no bound on a \mathbf{a} a .
The identity
∫ 0 1 ( 1 − s 2 ) b d s = 2 b b ! ( 2 b + 1 ) ! ! \int_0^1 (1-s^2)^b\,\mathrm{d}s=\frac{2^b b!}{(2b+1)!!} ∫ 0 1 ( 1 − s 2 ) b d s = ( 2 b + 1 )!! 2 b b !
and the series for I 0 I_0 I 0 show that replacing c g , p c_{g,p} c g , p by one in (133) gives ∏ i b Y ( L i ) \prod_i b_Y(L_i) ∏ i b Y ( L i ) . The same double series with an extra factor ∣ b ∣ ( ∣ b ∣ + 1 ) |b|(|b|+1) ∣ b ∣ ( ∣ b ∣ + 1 ) is bounded uniformly for Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] and L ∈ [ 0 , L 0 ] p L\in[0,L_0]^p L ∈ [ 0 , L 0 ] p : the inequality ( a + b ) ! ≥ a ! b ! (a+b)!\geq a!b! ( a + b )! ≥ a ! b ! supplies a summable factorial majorant for every fixed polynomial moment. Thus (78) , including its one-marking case, gives
G ( X ( Y ) ; L ) F ( X ( Y ) ) = ∏ i b Y ( L i ) ( 1 + O p , L 0 ( g − 1 ) ) . (134) \frac{G(X(Y);\mathbf{L})}{F(X(Y))}=\prod_i b_Y(L_i)(1+O_{p,L_0}(g^{-1})). \tag*{(134)} F ( X ( Y )) G ( X ( Y ) ; L ) = i ∏ b Y ( L i ) ( 1 + O p , L 0 ( g − 1 )) . (134)
All terms in (133) are nonnegative. The majorant also justifies convergence of the formal string-flow identity for 0 ≤ x < x c 0\leq x<x_c 0 ≤ x < x c . The factors b Y ( L ) b_Y(L) b Y ( L ) are at least one; they and their first Y Y Y derivatives are uniformly bounded on the compact set under consideration.
Extraction at a prescribed cusp count. The proof of (108) applies to every external degree tuple: a ( Y ) h U g , p Y ( b ) a(Y)^h U^Y_{g,p}(\mathbf{b}) a ( Y ) h U g , p Y ( b ) has nonnegative Taylor coefficients in Y Y Y . Before division by F F F , the insertion identity is a sum of these amplitudes with nonnegative powers of Y Y Y and nonnegative constants. Since Y ( x ) Y(x) Y ( x ) also has nonnegative coefficients, it follows that
G ( x ; L ) = V ( x ) h R g , p , L ( x ) , [ x j ] R g , p , L ( x ) ≥ 0. G(x;\mathbf{L})=V(x)^hR_{g,p,\mathbf{L}}(x),\qquad[x^j]R_{g,p,\mathbf{L}}(x)\geq0. G ( x ; L ) = V ( x ) h R g , p , L ( x ) , [ x j ] R g , p , L ( x ) ≥ 0.
Convergence follows from (134) . Hence
∣ G ( x e i θ ; L ) G ( x ; L ) ∣ ≤ ∣ V ( x e i θ ) V ( x ) ∣ h . (135) \left|\frac{G(xe^{i\theta};\mathbf{L})}{G(x;\mathbf{L})}\right|\leq\left|\frac{V(xe^{i\theta})}{V(x)}\right|^h. \tag*{(135)} G ( x ; L ) G ( x e i θ ; L ) ≤ V ( x ) V ( x e i θ ) h . (135)
Suppose first that k = n − p → ∞ k=n-p\to\infty k = n − p → ∞ , and choose x x x by h Ψ ( x ) = k h\Psi(x)=k h Ψ ( x ) = k . Put σ 2 = h D Ψ ( x ) \sigma^2=h\mathcal{D}\Psi(x) σ 2 = h D Ψ ( x ) . For fixed real u u u , the ratio of the centered moment generating functions associated with G G G and F F F is
( G / F ) ( x e u / σ ; L ) ( G / F ) ( x ; L ) . \frac{(G/F)(xe^{u/\sigma};\mathbf{L})}{(G/F)(x;\mathbf{L})}. ( G / F ) ( x ; L ) ( G / F ) ( x e u / σ ; L ) .
It tends to one, uniformly in L \mathbf{L} L . To see this, note that the endpoint expansions (110) and the quadratic inverse at x c x_c x c show that
sup 0 < x < x c x Y ′ ( x ) D Ψ ( x ) < ∞ . \sup_{0<x<x_c}\frac{xY'(x)}{\sqrt{\mathcal{D}\Psi(x)}}<\infty. 0 < x < x c sup D Ψ ( x ) x Y ′ ( x ) < ∞.
On the tilted interval x e u / σ xe^{u/\sigma} x e u / σ the model variance is comparable to its value at x x x , and the interval remains subcritical. Therefore Y ( x e u / σ ) − Y ( x ) = O u ( h − 1 / 2 ) Y(xe^{u/\sigma})-Y(x)=O_u(h^{-1/2}) Y ( x e u / σ ) − Y ( x ) = O u ( h − 1/2 ) . Now use (134) and the bounded derivative of b Y b_Y b Y .
The real Gaussian limit (113) thus holds for G G G with the same center k k k and scale σ \sigma σ . Its Fourier bound (135) is identical to the one used in (115) –(121) . Fourier inversion consequently gives the local limit (114) for G G G as well. At the common index k k k , both local probabilities are asymptotic to ( 2 π σ ) − 1 (\sqrt{2\pi}\sigma)^{-1} ( 2 π σ ) − 1 . Their ratio tends to one, and hence
[ x k ] G ( x ; L ) [ x k ] F ( x ) = ∏ i b Y ( L i ) ( 1 + o ( 1 ) ) , λ ( Y ) = 2 ( n − p ) 2 g − 2 + p . \frac{[x^k]G(x;\mathbf{L})}{[x^k]F(x)}=\prod_i b_Y(L_i)(1+o(1)),\qquad\lambda(Y)=\frac{2(n-p)}{2g-2+p}. [ x k ] F ( x ) [ x k ] G ( x ; L ) = i ∏ b Y ( L i ) ( 1 + o ( 1 )) , λ ( Y ) = 2 g − 2 + p 2 ( n − p ) .
All errors are uniform on the fixed length box, including along sequences tending to either endpoint of the parameter interval.
For bounded k k k , pass to a subsequence on which k k k is fixed. Equation (78) at Y = 0 Y=0 Y = 0 , with p + k p+k p + k markings, gives the fixed-cusp boundary estimate (104) . This yields the same conclusion since the saddle then tends to zero.
Finally, replacing 2 ( n − p ) / ( 2 g − 2 + p ) 2(n-p)/(2g-2+p) 2 ( n − p ) / ( 2 g − 2 + p ) by n / g n/g n / g changes the corresponding Y Y Y by o ( 1 ) o(1) o ( 1 ) uniformly in n ≥ p n\geq p n ≥ p . Indeed, along a subsequence on which n / g n/g n / g tends to zero, a positive finite limit, or infinity, both parameters have the same limit in [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] . Continuity of b Y ( L ) b_Y(L) b Y ( L ) , uniformly on the length box, proves (131) . At Y = 0 Y=0 Y = 0 , T = Δ = π 2 / 4 T=\Delta=\pi^2/4 T = Δ = π 2 /4 ; expanding I 0 I_0 I 0 and integrating gives the first formula in (132) . At Y = y 0 Y=y_0 Y = y 0 , the integrand is constant, which gives the second. □
Extraction with an inverse-genus error
This section completes the third idea of Subsection 1.5 with quantitative extraction estimates. The local limit theorem (114) gives no rate. Its quantitative form, Proposition E.1 in Appendix § 8.16 , gives only a power rate in min ( h , μ ) \min(h,\mu) min ( h , μ ) , and its Gaussian approximation degenerates when the coefficient index stays bounded. To obtain an error of order g − 1 g^{-1} g − 1 uniformly in the index, we retain an exact Poisson factor rather than its Gaussian approximation. Theorem 10.1 supplies the estimate used for the leading absolute-volume formula and the volume-ratio rates in § 3.31 , and Theorem 10.2 extends it to all orders. Their inputs are the marked expansions of § 3.22 , the convolution factor of Lemma 9.1 and the endpoint expansions (110) ; both theorems are proved in Appendix § 8.16 .
Why a Poisson factor. Near zero cusp weight the leading term of Φ g ( p ) \Phi_g^{(p)} Φ g ( p ) behaves like e h S ( x ) e^{hS(x)} e h S ( x ) with S ( x ) = 2 x + O ( x 2 ) S(x)=2x+O(x^2) S ( x ) = 2 x + O ( x 2 ) ; see (136) below. At a positive weight x x x , the normalized terms e − 2 h x ( 2 h x ) k / k ! e^{-2hx}(2hx)^k/k! e − 2 h x ( 2 h x ) k / k ! are exactly Poisson probabilities with mean 2 h x 2hx 2 h x . Hence, when the coefficient index k k k stays bounded, the number of additional cusps under the saddle weight is approximately Poisson, as anticipated in Subsection 1.1. A Gaussian approximation of its central point mass then has a relative error that does not tend to zero for fixed positive k k k , and it is undefined at k = 0 k=0 k = 0 . If the leading term is instead written with the exact Poisson mass p k = e − k k k / k ! p_k=e^{-k}k^k/k! p k = e − k k k / k ! at the mean, the ratio of the true coefficient to this model is 1 + O ( 1 / h ) 1+O(1/h) 1 + O ( 1/ h ) for every k k k . For large k k k one has 2 π k p k → 1 \sqrt{2\pi k}p_k\to1 2 π k p k → 1 , which recovers the Gaussian formula.
Fix p ≥ 1 p\ge1 p ≥ 1 , put F h = Φ g ( p ) F_h=\Phi_g^{(p)} F h = Φ g ( p ) , h = 2 g − 2 + p h=2g-2+p h = 2 g − 2 + p , and write
J g , p ( x ) = J g , p ( 0 ) e h S ( x ) B p ( x ) , S = log Z ( 0 ) Z ( Y ( x ) ) , B p = r p / 2 A Θ r ( 0 ) p / 2 A ( 0 ) Θ ( 0 ) . (136) \mathcal{J}_{g,p}(x)=\mathcal{J}_{g,p}(0)e^{hS(x)}B_p(x),\qquad S=\log\frac{\mathcal{Z}(0)}{\mathcal{Z}(Y(x))},\qquad B_p=\frac{r^{p/2}\sqrt{A\Theta}}{r(0)^{p/2}\sqrt{A(0)\Theta(0)}}. \tag*{(136)} J g , p ( x ) = J g , p ( 0 ) e h S ( x ) B p ( x ) , S = log Z ( Y ( x )) Z ( 0 ) , B p = r ( 0 ) p /2 A ( 0 ) Θ ( 0 ) r p /2 A Θ . (136)
We view J \mathcal{J} J as a function of x = X ( Y ) x=X(Y) x = X ( Y ) . Then D S = Ψ \mathcal{D}S=\Psi D S = Ψ and D log B p = β p Ψ \mathcal{D}\log B_p=\beta_p\Psi D log B p = β p Ψ . For each integer k > 1 k>1 k > 1 , choose x x x by h Ψ ( x ) = k h\Psi(x)=k h Ψ ( x ) = k , and put v = D Ψ v=\mathcal{D}\Psi v = D Ψ , σ 2 = h v \sigma^2=hv σ 2 = h v , M = min ( h , k ) M=\min(h,k) M = min ( h , k ) . Let p k = e − k k k / k ! p_k=e^{-k}k^k/k! p k = e − k k k / k ! , with p 0 = 1 p_0=1 p 0 = 1 .
Theorem 10.1 (Uniform Poisson-corrected extraction). For every fixed p ≥ 1 p\ge1 p ≥ 1 there is h p h_p h p such that, for h ≥ h p h\ge h_p h ≥ h p and uniformly over integers k ≥ 0 k\ge0 k ≥ 0 ,
[ z k ] F h ( z ) = J g , p ( x ) π x − k p k k h D Ψ ( x ) ( 1 + O p ( h − 1 ) ) . (137) [z^k]F_h(z)=\frac{\mathcal{J}_{g,p}(x)}{\pi}x^{-k}p_k\sqrt{\frac{k}{h\mathcal{D}\Psi(x)}}\left(1+O_p(h^{-1})\right). \tag*{(137)} [ z k ] F h ( z ) = π J g , p ( x ) x − k p k h D Ψ ( x ) k ( 1 + O p ( h − 1 ) ) . (137) At k = 0 k=0 k = 0 take x = 0 x=0 x = 0 , x − k = 1 x^{-k}=1 x − k = 1 and p 0 = 1 p_0=1 p 0 = 1 ; the variance factor k / ( h D Ψ ( x ) ) \sqrt{k/(h\mathcal{D}\Psi(x))} k / ( h D Ψ ( x )) is one by continuity along k = h Ψ ( x ) k=h\Psi(x) k = h Ψ ( x ) .
Theorem 10.1 is the case K = 0 K=0 K = 0 of Theorem 10.2 below; it is proved separately because its proof uses Theorem 8.2 only at a fixed order, whereas the proof of Theorem 10.2 uses the full linear-order range. The proof, given in Subsections E.2–E.4, has three parts. First, a sufficiently long marked expansion, namely Theorem 8.2 at a fixed finite order, is transferred to small complex neighborhoods of the real saddle point. Second, the Poisson correction is identified. Third, the small indices are estimated directly. The argument uses no estimate for a growing number of markings and never differentiates a remainder; away from zero, the model is continued analytically only to small neighborhoods of the real axis.
The full expansion at a prescribed cusp count
Theorem 10.1 recovers the leading term with an inverse-genus error. We now extract the whole marked expansion, keeping the remainder uniform in both the cusp count and the truncation order. The result is Theorem 10.2, which provides the remaining input for the full absolute-volume expansion in § 3.31 .
The idea. Away from zero cusp weight, including near the critical point, the coefficient integral has a nondegenerate saddle point, and the Gaussian method of steepest descent gives an expansion in 1 / h 1/h 1/ h . Near zero this fails for bounded k k k : the saddle point of e h S ( z ) z − k e^{hS(z)}z^{-k} e h S ( z ) z − k approaches the singularity of z − k z^{-k} z − k at the origin. Following Temme, we change variables so that the phase S ( z ) − ( k / h ) log z S(z)-(k/h)\log z S ( z ) − ( k / h ) log z becomes exactly u − ( k / h ) log u u-(k/h)\log u u − ( k / h ) log u . The coefficient is then a Poisson coefficient [ u k ] e h u f ( u ) [u^k]e^{hu}f(u) [ u k ] e h u f ( u ) with an analytic amplitude f f f . Expanding f f f about u = k / h u=k/h u = k / h produces the signed central polynomials T m ( k ) T_m(k) T m ( k ) of the coefficient operator in (E.24), divided by powers of h h h . These are not the central moments of a probability distribution: for example, T 2 ( k ) = − k T_2(k)=-k T 2 ( k ) = − k . The resulting expansion in 1 / h 1/h 1/ h is uniform in k k k in this small-weight range, including k = 0 k=0 k = 0 . The two coefficient formulas agree on an overlap of their ranges, which identifies them.
The proof is given in Subsections E.5–E.7: a Poisson integral is used near zero cusp weight and a Gaussian integral away from zero, including near criticality. Near zero we follow Temme’s method for uniform Laplace and loop integrals [178, Sections 2.3 and 3.7].
Retain h = 2 g − 2 + p h=2g-2+p h = 2 g − 2 + p , k = h Ψ ( x ) k=h\Psi(x) k = h Ψ ( x ) , v = D Ψ ( x ) v=\mathcal{D}\Psi(x) v = D Ψ ( x ) and the functions S S S , B p B_p B p in (136) . Regard L p , j \mathcal{L}_{p,j} L p , j as functions of x x x via Y ( x ) Y(x) Y ( x ) . Write
Q g , p , k = J g , p ( x ) π x − k p k k h v . (138) \mathcal{Q}_{g,p,k}=\frac{\mathcal{J}_{g,p}(x)}{\pi}x^{-k}p_k\sqrt{\frac{k}{hv}}. \tag*{(138)} Q g , p , k = π J g , p ( x ) x − k p k h v k . (138)
The conventions at k = 0 k=0 k = 0 are those of Theorem 10.1.
Put L p , 0 = 1 \mathcal{L}_{p,0}=1 L p , 0 = 1 . Let G w \mathcal{G}_w G w be the linear functional on polynomials which maps w 2 a w^{2a} w 2 a to ( − 1 ) a ( 2 a − 1 ) ! ! (-1)^a(2a-1)!! ( − 1 ) a ( 2 a − 1 )!! and odd powers to zero. Define the formal even series
N p ( τ , x ) = G w [ exp { ∑ m ≥ 3 τ m − 2 D m S m ! v m / 2 w m + ∑ m ≥ 1 τ m D m log B p m ! v m / 2 w m } ⋅ ∑ s ≥ 0 τ 2 s L p , s ( x e τ w / v ) ] . (139) \mathcal{N}_p(\tau,x)=\mathcal{G}_w\left[\exp\left\{\sum_{m\geq3}\frac{\tau^{m-2}\mathcal{D}^mS}{m!v^{m/2}}w^m+\sum_{m\geq1}\frac{\tau^m\mathcal{D}^m\log B_p}{m!v^{m/2}}w^m\right\}\cdot\sum_{s\geq0}\tau^{2s}\mathcal{L}_{p,s}\left(xe^{\tau w/\sqrt v}\right)\right]. \tag*{(139)} N p ( τ , x ) = G w [ exp { m ≥ 3 ∑ m ! v m /2 τ m − 2 D m S w m + m ≥ 1 ∑ m ! v m /2 τ m D m log B p w m } ⋅ s ≥ 0 ∑ τ 2 s L p , s ( x e τ w / v ) ] . (139)
Formal composition in the last line means Taylor expansion at x x x . At every order only finitely many derivatives are used. If 2 π k p k ∼ ∑ j ≥ 0 s j k − j \sqrt{2\pi k}\,p_k\sim\sum_{j\geq0}s_jk^{-j} 2 π k p k ∼ ∑ j ≥ 0 s j k − j , s 0 = 1 s_0=1 s 0 = 1 , set
A p , j ( Y ) = [ τ 2 j ] N p ( τ , X ( Y ) ) ∑ i ≥ 0 s i Ψ ( X ( Y ) ) − i τ 2 i . (140) \mathcal{A}_{p,j}(Y)= [\tau^{2j}]\frac{\mathcal{N}_p(\tau,X(Y))}{\sum_{i\geq0}s_i\Psi(X(Y))^{-i}\tau^{2i}}. \tag*{(140)} A p , j ( Y ) = [ τ 2 j ] ∑ i ≥ 0 s i Ψ ( X ( Y ) ) − i τ 2 i N p ( τ , X ( Y )) . (140)
This initially defines the functions for 0 < Y < y 0 0<Y<y_0 0 < Y < y 0 .
Theorem 10.2 (Uniform expansion at a prescribed cusp count). *For every fixed p ≥ 1 p\geq1 p ≥ 1 there are constants C p , B p , c p , h p > 0 C_p,\mathcal{B}_p,c_p,h_p>0 C p , B p , c p , h p > 0 such that, for h = 2 g − 2 + p ≥ h p h=2g-2+p\geq h_p h = 2 g − 2 + p ≥ h p , every integer k ≥ 0 k\geq0 k ≥ 0 and 0 ≤ K ≤ c p h 0\leq K\leq c_ph 0 ≤ K ≤ c p h ,
∣ [ z k ] Φ g ( p ) Q g , p , k − ∑ j = 0 K A p , j ( Y ) h j ∣ ≤ C p B p K + 1 ( K + 1 ) ! h K + 1 , h Ψ ( X ( Y ) ) = k , (141) \left|\frac{[z^k]\Phi_g^{(p)}}{\mathcal{Q}_{g,p,k}}-\sum_{j=0}^{K}\frac{\mathcal{A}_{p,j}(Y)}{h^j}\right|\leq\frac{C_p\mathcal{B}_p^{K+1}(K+1)!}{h^{K+1}},\qquad h\Psi(X(Y))=k, \tag*{(141)} Q g , p , k [ z k ] Φ g ( p ) − j = 0 ∑ K h j A p , j ( Y ) ≤ h K + 1 C p B p K + 1 ( K + 1 )! , h Ψ ( X ( Y )) = k , (141) where A p , 0 = 1 \mathcal{A}_{p,0}=1 A p , 0 = 1 . The coefficients are analytic on one complex neighborhood Ω p \Omega_p Ω p of [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] , with ∥ A p , j ∥ Ω p ≤ C p B p j j ! \|\mathcal{A}_{p,j}\|_{\Omega_p}\leq C_p\mathcal{B}_p^j j! ∥ A p , j ∥ Ω p ≤ C p B p j j ! , and A p , j ( 0 ) = L p , j ( 0 ) \mathcal{A}_{p,j}(0)=\mathcal{L}_{p,j}(0) A p , j ( 0 ) = L p , j ( 0 ) . For a sufficiently small fixed c p ′ > 0 c'_p>0 c p ′ > 0 , truncation at K = ⌊ c p ′ h ⌋ K=\lfloor c'_ph\rfloor K = ⌊ c p ′ h ⌋ has relative error O p ( e − a p h ) O_p(e^{-a_ph}) O p ( e − a p h ) , uniformly in k k k , for some a p > 0 a_p>0 a p > 0 . The marking count is fixed. The infinite series need not converge.
The first coefficient A p , 1 \mathcal{A}_{p,1} A p , 1 is given explicitly in (E.29), and its value at the critical endpoint in (E.30). Retaining it improves the relative remainder to O p ( h − 2 ) O_p(h^{-2}) O p ( h − 2 ) , uniformly over all coefficient indices. Section 11 applies Theorem 10.1 with p = 1 p=1 p = 1 , and with the pairs ( g , 1 ) , ( g − 1 , 3 ) (g,1),(g-1,3) ( g , 1 ) , ( g − 1 , 3 ) and ( g , 1 ) , ( g , 2 ) (g,1),(g,2) ( g , 1 ) , ( g , 2 ) for the two volume ratios, and applies Theorem 10.2 with p = 1 p=1 p = 1 to obtain the coefficients A j ( g , n ) A_j(g,n) A j ( g , n ) of Theorem 1.1.
Absolute volumes at every cusp count
Specializing the extraction theorems to one distinguished marking, p = 1 p=1 p = 1 , gives the absolute-volume expansion and the inverse-genus volume-ratio estimates. Theorem 10.1 10.1 10.1 supplies the explicit leading term with relative error O ( g − 1 ) O(g^{-1}) O ( g − 1 ) ; Theorem 10.2 10.2 10.2 supplies the higher corrections and their uniform remainder. For n = 0 n=0 n = 0 , the leading estimate also follows from the fixed-cusp expansion of [147], and the full expansion is recovered from n = 1 n=1 n = 1 by the exact dilaton relation.
For g ≥ 2 g\geq2 g ≥ 2 and n ≥ 1 n\geq1 n ≥ 1 , set h = 2 g − 1 h=2g-1 h = 2 g − 1 , k = n − 1 k=n-1 k = n − 1 , and choose x = X ( Y ) x=X(Y) x = X ( Y ) by h Ψ ( x ) = k h\Psi(x)=k h Ψ ( x ) = k . At k = 0 k=0 k = 0 take x = Y = 0 x=Y=0 x = Y = 0 . Define
V ^ g , n = ( 2 π 2 ) 3 g − 3 + n J g , 1 ( Y ) π ( k e x ) k k h D Ψ ( x ) ( n ≥ 1 ) , \widehat{V}_{g,n}=(2\pi^2)^{3g-3+n}\frac{J_{g,1}(Y)}{\pi}\left(\frac{k}{e x}\right)^k\sqrt{\frac{k}{h\mathcal{D}\Psi(x)}}\qquad(n\geq1), V g , n = ( 2 π 2 ) 3 g − 3 + n π J g , 1 ( Y ) ( e x k ) k h D Ψ ( x ) k ( n ≥ 1 ) ,
V ^ g , 0 = Γ ( 2 g − 2 ) ( 4 π 2 ) 2 g − 3 π g . (142) \widehat{V}_{g,0}=\frac{\Gamma(2g-2)(4\pi^2)^{2g-3}}{\sqrt{\pi g}}. \tag*{(142)} V g , 0 = π g Γ ( 2 g − 2 ) ( 4 π 2 ) 2 g − 3 . (142)
At k = 0 k=0 k = 0 , both k k k -dependent factors on the first line are defined by their limits, which are one. All functions in this formula were defined in (85) , (86) , and (88) . We have r = 2 Δ r=2\Delta r = 2Δ , A α = I / Δ A\alpha=I/\sqrt{\Delta} A α = I / Δ by (85) , and Θ = − 1 / K ‾ ′ ( 1 ) = − a / ( r ∂ z K Y ( r ) ) \Theta=-1/\overline{K}'(1)=-a/(r\partial_zK_Y(r)) Θ = − 1/ K ′ ( 1 ) = − a / ( r ∂ z K Y ( r )) by (86) . Since X ′ ′ ( v ) = − X ( v ) / v X''(v)=-X(v)/v X ′′ ( v ) = − X ( v ) / v , differentiating K Y K_Y K Y gives − 2 ∂ z K Y ( r ) = ∫ 0 1 ( 1 − s 2 ) X ( v ) v − 1 d s -2\partial_zK_Y(r)=\int_0^1(1-s^2)X(v)v^{-1}\,ds − 2 ∂ z K Y ( r ) = ∫ 0 1 ( 1 − s 2 ) X ( v ) v − 1 d s with v = T − Δ s 2 v=T-\Delta s^2 v = T − Δ s 2 . Hence r A Θ = ϖ ( Y ) 2 rA\Theta=\varpi(Y)^2 r A Θ = ϖ ( Y ) 2 , that is, r 1 / 2 A Θ = ϖ ( Y ) r^{1/2}\sqrt{A\Theta}=\varpi(Y) r 1/2 A Θ = ϖ ( Y ) , and Z = 2 I \mathcal{Z}=\sqrt{2}I Z = 2 I , with ϖ \varpi ϖ as in the introduction. Hence (142) is the formula (5) , with D Ψ ( x ) = x Ψ ′ ( x ) \mathcal{D}\Psi(x)=x\Psi'(x) D Ψ ( x ) = x Ψ ′ ( x ) .
Theorem 11.1 (The leading term). There is a constant C > 0 C>0 C > 0 such that, for all sufficiently large g g g ,
sup n ≥ 0 ∣ V g , n V ^ g , n − 1 ∣ ≤ C g . (143) \sup_{n\geq0}\left|\frac{V_{g,n}}{\widehat{V}_{g,n}}-1\right|\leq\frac{C}{g}. \tag*{(143)} n ≥ 0 sup V g , n V g , n − 1 ≤ g C . (143) The order g − 1 g^{-1} g − 1 is sharp for the leading model (142) :
lim inf g → ∞ g sup n ≥ 0 ∣ V g , n V ^ g , n − 1 ∣ ≥ 7 12 − 17 6 π 2 > 0. (144) \liminf_{g\to\infty}g\sup_{n\geq0}\left|\frac{V_{g,n}}{\widehat{V}_{g,n}}-1\right|\geq\frac{7}{12}-\frac{17}{6\pi^2}>0. \tag*{(144)} g → ∞ lim inf g n ≥ 0 sup V g , n V g , n − 1 ≥ 12 7 − 6 π 2 17 > 0. (144)
Proof. For n ≥ 1 n\geq1 n ≥ 1 , apply Theorem 10.1 10.1 10.1 with p = 1 p=1 p = 1 . The coefficient identity (124) gives
V g , n = ( 2 π 2 ) 3 g − 3 + n ( n − 1 ) ! x − k J g , 1 ( Y ) π p k k h D Ψ ( x ) ( 1 + O ( g − 1 ) ) , (145) V_{g,n}=(2\pi^2)^{3g-3+n}(n-1)!x^{-k}\frac{J_{g,1}(Y)}{\pi}p_k\sqrt{\frac{k}{h\mathcal{D}\Psi(x)}}(1+O(g^{-1})), \tag*{(145)} V g , n = ( 2 π 2 ) 3 g − 3 + n ( n − 1 )! x − k π J g , 1 ( Y ) p k h D Ψ ( x ) k ( 1 + O ( g − 1 )) , (145)
uniformly, including k = 0 k=0 k = 0 by continuity. Since k ! p k = ( k / e ) k k!p_k=(k/e)^k k ! p k = ( k / e ) k , this is precisely the first line of (142) with uniform relative error O ( g − 1 ) O(g^{-1}) O ( g − 1 ) . For n = 0 n=0 n = 0 , the constant was independently identified in Section 7 7 7 , so the fixed-cusp expansion of Mirzakhani–Zograf [147, Theorem 1.2] gives V g , 0 / V ^ g , 0 = 1 + c 0 ( 1 ) g − 1 + O ( g − 2 ) V_{g,0}/\widehat{V}_{g,0}=1+c_0^{(1)}g^{-1}+O(g^{-2}) V g , 0 / V g , 0 = 1 + c 0 ( 1 ) g − 1 + O ( g − 2 ) , where c n ( 1 ) c_n^{(1)} c n ( 1 ) denotes their first coefficient. Its value is stated in [147, Remark 1.3], and it also follows from two proved statements. In the normalization of [147], the first-correction theorem of Liu and Xu [131, Theorem 1.3, at $t=1$] gives c 2 ( 1 ) = 1 / 12 + 1 / ( 6 π 2 ) c_2^{(1)}=1/12+1/(6\pi^2) c 2 ( 1 ) = 1/12 + 1/ ( 6 π 2 ) . Indeed, the conversion between the two leading terms and the corresponding first coefficients is
2 g Γ ( 2 g − 1 2 ) Γ ( 2 g ) = 1 + 3 16 g + O ( g − 2 ) , \sqrt{2g}\frac{\Gamma\left(2g-\frac{1}{2}\right)}{\Gamma(2g)}=1+\frac{3}{16g}+O(g^{-2}), 2 g Γ ( 2 g ) Γ ( 2 g − 2 1 ) = 1 + 16 g 3 + O ( g − 2 ) ,
c 2 ( 1 ) = b 2 ( 1 ) + 3 16 = − 5 48 + 1 6 π 2 + 3 16 = 1 12 + 1 6 π 2 , (146) c_2^{(1)}=b_2(1)+\frac{3}{16}=-\frac{5}{48}+\frac{1}{6\pi^2}+\frac{3}{16}=\frac{1}{12}+\frac{1}{6\pi^2}, \tag*{(146)} c 2 ( 1 ) = b 2 ( 1 ) + 16 3 = − 48 5 + 6 π 2 1 + 16 3 = 12 1 + 6 π 2 1 , (146)
where b 2 b_2 b 2 is the coefficient in [131]. The cusp-ratio expansion [147, Theorem 1.4 gives c n + 1 ( 1 ) − c n ( 1 ) = 2 / π 2 − 1 / 4 − n / π 2 c_{n+1}^{(1)}-c_n^{(1)}=2/\pi^2-1/4-n/\pi^2 c n + 1 ( 1 ) − c n ( 1 ) = 2/ π 2 − 1/4 − n / π 2 for n = 0 , 1 n=0,1 n = 0 , 1 . Hence
V g , 0 V ^ g , 0 = 1 + 1 g ( 7 12 − 17 6 π 2 ) + O ( g − 2 ) . (147) \frac{V_{g,0}}{\widehat{V}_{g,0}}=1+\frac{1}{g}\left(\frac{7}{12}-\frac{17}{6\pi^2}\right)+O(g^{-2}). \tag*{(147)} V g , 0 V g , 0 = 1 + g 1 ( 12 7 − 6 π 2 17 ) + O ( g − 2 ) . (147)
This closes (143) and proves (144) .
The order g − 1 g^{-1} g − 1 is also attained at large cusp density, without the value of c 0 ( 1 ) c_0^{(1)} c 0 ( 1 ) . By Theorem 10.2 with p = 1 p=1 p = 1 and K = 1 K=1 K = 1 , together with the coefficient identity used in the proof of Theorem 1.1 below, V g , n / V ^ g , n = 1 + A 1 , 1 ( Y g , n ) / ( 2 g − 1 ) + O ( g − 2 ) V_{g,n}/\widehat{V}_{g,n}=1+A_{1,1}(Y_{g,n})/(2g-1)+O(g^{-2}) V g , n / V g , n = 1 + A 1 , 1 ( Y g , n ) / ( 2 g − 1 ) + O ( g − 2 ) uniformly in n ≥ 1 n\geq1 n ≥ 1 , where ( 2 g − 1 ) Ψ ( X ( Y g , n ) ) = n − 1 (2g-1)\Psi(X(Y_{g,n}))=n-1 ( 2 g − 1 ) Ψ ( X ( Y g , n )) = n − 1 . For fixed g g g , Y g , n → y 0 Y_{g,n}\to y_0 Y g , n → y 0 as n → ∞ n\to\infty n → ∞ , and A 1 , 1 ( y 0 ) = − 29 / 120 A_{1,1}(y_0)=-29/120 A 1 , 1 ( y 0 ) = − 29/120 by (E.30). Hence the left side of (144) is also at least 29 / 240 29/240 29/240 . The latter concerns the uncorrected model chosen here, rather than an obstruction to adding further asymptotic corrections. □
The model (142) contains no unknown constant: its normalization was fixed at the pure-ψ \psi ψ endpoint in § 3.19 , and the fixed-cusp expansion of [147] enters only afterwards, for n = 0 n=0 n = 0 in the proof above. The matrix-model reading of its action I ( Y ) I(Y) I ( Y ) and of the parameter Y Y Y is given in Subsection 13.8.
An inverse-genus error for the volume ratios
The sharper extraction estimate also gives a uniform rate in the volume-ratio theorem. For the genus-exchange ratio, the two generating functions have the same saddle. For the cusp ratio, adding a cusp changes the complexity by one, and the argument below controls the resulting saddle displacement even near the endpoints.
Completion of the proof of Theorem 1.3 . For n ≥ 1 n\geq1 n ≥ 1 put h = 2 g − 1 h=2g-1 h = 2 g − 1 , k = n − 1 k=n-1 k = n − 1 , and h Ψ ( x ) = k h\Psi(x)=k h Ψ ( x ) = k . Apply Theorem 10.1 at ( g , p ) = ( g , 1 ) (g,p)=(g,1) ( g , p ) = ( g , 1 ) and ( g − 1 , 3 ) (g-1,3) ( g − 1 , 3 ) . Their Poisson and variance factors cancel exactly, and (88) gives J g − 1 , 3 / J g , 1 = 4 r \mathcal{J}_{g-1,3}/\mathcal{J}_{g,1}=4r J g − 1 , 3 / J g , 1 = 4 r . Consequently (124) yields
B g , n = B ( 2 k / h ) ( 1 + O ( h − 1 ) ) . B_{g,n}=\mathcal{B}(2k/h)(1+O(h^{-1})). B g , n = B ( 2 k / h ) ( 1 + O ( h − 1 )) .
For the cusp ratio assume first that k ≥ 1 k\geq1 k ≥ 1 , and let x ′ x' x ′ satisfy ( h + 1 ) Ψ ( x ′ ) = k (h+1)\Psi(x')=k ( h + 1 ) Ψ ( x ′ ) = k . Write q = log x q=\log x q = log x , q ′ = log x ′ q'=\log x' q ′ = log x ′ and v = D Ψ v=\mathcal{D}\Psi v = D Ψ . The convergent endpoint expansions give
sup 0 < x < x c { Ψ ( x ) 2 v ( x ) + ∣ Ψ ( x ) v ( x ) D log v ( x ) ∣ + ∣ β 2 ( x ) ∣ } < ∞ . \sup_{0<x<x_c}\left\{\frac{\Psi(x)^2}{v(x)}+\left|\frac{\Psi(x)}{v(x)}\mathcal{D}\log v(x)\right|+|\beta_2(x)|\right\}<\infty. 0 < x < x c sup { v ( x ) Ψ ( x ) 2 + v ( x ) Ψ ( x ) D log v ( x ) + ∣ β 2 ( x ) ∣ } < ∞.
Indeed, near zero Ψ ∼ v ∼ 2 x \Psi\sim v\sim2x Ψ ∼ v ∼ 2 x ; near criticality Ψ ∼ 5 / ( 4 ε ) \Psi\sim5/(4\varepsilon) Ψ ∼ 5/ ( 4 ε ) and v ∼ 5 / ( 4 ε 2 ) v\sim5/(4\varepsilon^2) v ∼ 5/ ( 4 ε 2 ) . The function β 2 \beta_2 β 2 is bounded by (92) . Since Ψ ( x ′ ) / Ψ ( x ) = h / ( h + 1 ) \Psi(x')/\Psi(x)=h/(h+1) Ψ ( x ′ ) /Ψ ( x ) = h / ( h + 1 ) , integration with respect to log Ψ \log\Psi log Ψ shows that
v ( x ′ ) v ( x ) = 1 + O ( h − 1 ) , ∣ q ′ − q ∣ ≤ C Ψ ( x ) h v ( x ) . \frac{v(x')}{v(x)}=1+O(h^{-1}),\qquad|q'-q|\leq\frac{C\Psi(x)}{hv(x)}. v ( x ) v ( x ′ ) = 1 + O ( h − 1 ) , ∣ q ′ − q ∣ ≤ h v ( x ) C Ψ ( x ) .
For L ( q ) = log J g , 2 ( e q ) − k q L(q)=\log\mathcal{J}_{g,2}(e^q)-kq L ( q ) = log J g , 2 ( e q ) − k q , the exact model identities give
L ′ ( q ) = ( h + 1 ) Ψ ( e q ) − k + β 2 ( e q ) Ψ ( e q ) . L'(q)=(h+1)\Psi(e^q)-k+\beta_2(e^q)\Psi(e^q). L ′ ( q ) = ( h + 1 ) Ψ ( e q ) − k + β 2 ( e q ) Ψ ( e q ) .
On [ q ′ , q ] [q',q] [ q ′ , q ] its absolute value is at most C Ψ ( x ) C\Psi(x) C Ψ ( x ) , so ∣ L ( q ′ ) − L ( q ) ∣ ≤ C Ψ ( x ) 2 / ( h v ( x ) ) = O ( h − 1 ) |L(q')-L(q)|\leq C\Psi(x)^2/(hv(x))=O(h^{-1}) ∣ L ( q ′ ) − L ( q ) ∣ ≤ C Ψ ( x ) 2 / ( h v ( x )) = O ( h − 1 ) . Thus the saddle displacement and the variance factor in (137) each change the coefficient ratio by 1 + O ( h − 1 ) 1+O(h^{-1}) 1 + O ( h − 1 ) . At the common point x x x , (88) gives exactly
J g , 2 ( x ) J g , 1 ( x ) = h A r h h + 1 . \frac{\mathcal{J}_{g,2}(x)}{\mathcal{J}_{g,1}(x)}=\frac{h}{Ar}\sqrt{\frac{h}{h+1}}. J g , 1 ( x ) J g , 2 ( x ) = A r h h + 1 h .
Using the coefficient normalization in (127) , we obtain
C g , n = h 2 g C ( 2 k / h ) ( 1 + O ( h − 1 ) ) . C_{g,n}=\frac{h}{2g}\mathcal{C}(2k/h)(1+O(h^{-1})). C g , n = 2 g h C ( 2 k / h ) ( 1 + O ( h − 1 )) .
At k = 0 k = 0 k = 0 both saddles are zero. The conventions in Theorem 10.1 give the same two ratio formulas directly, without logarithmic coordinates.
The positive profiles are analytic at zero, and their expansions at infinity in (17) –(19) are convergent in the inverse square root of the density. Hence
sup λ ≥ 0 ( 1 + λ ) ( ∣ ( log B ) ′ ( λ ) ∣ + ∣ ( log C ) ′ ( λ ) ∣ ) < ∞ . \sup_{\lambda\ge0}(1+\lambda)\left(\left|(\log\mathcal{B})'(\lambda)\right|+\left|(\log\mathcal{C})'(\lambda)\right|\right)<\infty. λ ≥ 0 sup ( 1 + λ ) ( ∣ ( log B ) ′ ( λ ) ∣ + ∣ ( log C ) ′ ( λ ) ∣ ) < ∞.
For λ ∗ = 2 ( n − 1 ) / ( 2 g − 1 ) \lambda_* = 2(n-1)/(2g-1) λ ∗ = 2 ( n − 1 ) / ( 2 g − 1 ) and λ = n / g \lambda= n/g λ = n / g , ∣ λ ∗ − λ ∣ / ( 1 + min ( λ ∗ , λ ) ) = O ( g − 1 ) |\lambda_*-\lambda|/(1+\min(\lambda_*,\lambda))=O(g^{-1}) ∣ λ ∗ − λ ∣/ ( 1 + min ( λ ∗ , λ )) = O ( g − 1 ) uniformly in n ≥ 1 n\ge1 n ≥ 1 . Integrating the preceding bound replaces λ ∗ \lambda_* λ ∗ by λ \lambda λ with relative error O ( g − 1 ) O(g^{-1}) O ( g − 1 ) . For n = 0 n=0 n = 0 , the fixed-cusp expansion of [147] at the finitely many cusp counts 0 , 1 , 2 0,1,2 0 , 1 , 2 gives the required rate directly. Finally D g , n = C g , n − 1 / C g , n D_{g,n}=C_{g,n-1}/C_{g,n} D g , n = C g , n − 1 / C g , n and the same logarithmic derivative bound prove sup n ≥ 1 ∣ D g , n − 1 ∣ = O ( g − 1 ) \sup_{n\ge1}|D_{g,n}-1|=O(g^{-1}) sup n ≥ 1 ∣ D g , n − 1∣ = O ( g − 1 ) . □
Proof of Theorem 1.1
For n ≥ 1 n\ge1 n ≥ 1 the expansion follows from the extraction theorem, Theorem 10.2 , with p = 1 p=1 p = 1 and the same leading term. The unmarked case n = 0 n=0 n = 0 is recovered from n = 1 n=1 n = 1 by the exact dilaton relation, which for volume polynomials is the identity ∂ L V g , 1 ( 2 π i ) = 2 π i ( 2 g − 2 ) V g , 0 \partial_LV_{g,1}(2\pi i)=2\pi i(2g-2)V_{g,0} ∂ L V g , 1 ( 2 π i ) = 2 π i ( 2 g − 2 ) V g , 0 of Do and Norbury [58, Theorem 2]; see also [144, equation (3.5)]. Its higher coefficients therefore need no separate asymptotic normalization. For h = 2 g − 1 h=2g-1 h = 2 g − 1 , the scalar C 0 C_0 C 0 of (D.13) obeys
V g , 0 = V g , 1 C 0 ( h − 1 ; 0 ) 4 π 2 ( h − 1 ) , V ^ g , 1 4 π 2 ( h − 1 ) V ^ g , 0 = 1 + 1 h . (148) V_{g,0}=\frac{V_{g,1}C_0(h-1;0)}{4\pi^2(h-1)},\qquad\frac{\widehat{V}_{g,1}}{4\pi^2(h-1)\widehat{V}_{g,0}}=\sqrt{1+\frac{1}{h}}. \tag*{(148)} V g , 0 = 4 π 2 ( h − 1 ) V g , 1 C 0 ( h − 1 ; 0 ) , 4 π 2 ( h − 1 ) V g , 0 V g , 1 = 1 + h 1 . (148)
Indeed (67) holds at p = 0 p=0 p = 0 ; at Y = 0 Y=0 Y = 0 , undo its normalizations with r = π 2 / 2 r=\pi^2/2 r = π 2 /2 , A = 1 / π 2 A=1/\pi^2 A = 1/ π 2 . Equivalently, the first identity in (148) is the case n = 0 n=0 n = 0 of [58, Theorem 2]. The second identity follows exactly from (142) . The scalar expansion of C 0 C_0 C 0 follows directly from the last line of (D.17), using the proved one-marking profile and its summable remainder; this step does not require an unmarked harmonic comparison.
The coefficients in Theorem 1.1 are defined as follows. For n = 0 n=0 n = 0 , define constants a 0 , j a_{0,j} a 0 , j by the finite-order formal rule
1 + ∑ j ≥ 1 a 0 , j t j = ( 1 + t ) 1 / 2 C 0 ( t 1 − t ; 0 ) ( 1 + ∑ j ≥ 1 L 1 , j ( 0 ) t j ) . (149) 1+\sum_{j\ge1}a_{0,j}t^j=(1+t)^{1/2}C_0\left(\frac{t}{1-t};0\right)\left(1+\sum_{j\ge1}L_{1,j}(0)t^j\right). \tag*{(149)} 1 + j ≥ 1 ∑ a 0 , j t j = ( 1 + t ) 1/2 C 0 ( 1 − t t ; 0 ) ( 1 + j ≥ 1 ∑ L 1 , j ( 0 ) t j ) . (149)
Here the second argument denotes Y = 0 Y=0 Y = 0 , and t = 1 / h t=1/h t = 1/ h ; thus the cusp scalar uses complexity h − 1 h-1 h − 1 . Put A j ( g , 0 ) = a 0 , j A_j(g,0)=a_{0,j} A j ( g , 0 ) = a 0 , j , and, for n ≥ 1 n\ge1 n ≥ 1 , put A j ( g , n ) = A 1 , j ( Y g , n ) A_j(g,n)=A_{1,j}(Y_{g,n}) A j ( g , n ) = A 1 , j ( Y g , n ) with the coefficients (140) of Theorem 10.2 , where ( 2 g − 1 ) Ψ ( X ( Y g , n ) ) = n − 1 (2g-1)\Psi(X(Y_{g,n}))=n-1 ( 2 g − 1 ) Ψ ( X ( Y g , n )) = n − 1 .
Proof of Theorem 1.1. For n ≥ 1 n\ge1 n ≥ 1 , take p = 1 p=1 p = 1 , k = n − 1 k=n-1 k = n − 1 in Theorem 10.2 . The coefficient identity (124) and k ! p k = ( k / e ) k k!p_k=(k/e)^k k ! p k = ( k / e ) k identify [ z k ] Φ g ( 1 ) / Q g , 1 , k [z^k]\Phi_g^{(1)}/Q_{g,1,k} [ z k ] Φ g ( 1 ) / Q g , 1 , k with V g , n / V ^ g , n V_{g,n}/\widehat{V}_{g,n} V g , n / V g , n , including k = 0 k=0 k = 0 . For n = 0 n=0 n = 0 , the dilaton scalar in (148) has a one-marking child of complexity h h h , so exactly
C 0 ( h − 1 ; 0 ) = 1 A ( 0 ) ∑ d ≥ 1 b d ( 0 ) c g , 1 ( d ; 0 ) . C_0(h-1;0)=\frac{1}{A(0)}\sum_{d\ge1}b_d(0)c_{g,1}(d;0). C 0 ( h − 1 ; 0 ) = A ( 0 ) 1 d ≥ 1 ∑ b d ( 0 ) c g , 1 ( d ; 0 ) .
Apply (D.37) at this child with L = K + 1 L=K+1 L = K + 1 . Fix 0 < σ < log ρ 0<\sigma<\log\rho 0 < σ < log ρ within the absolute exponential moment of b d b_d b d . Then
( 1 + M L ) L ∑ d ∣ b d ( 0 ) ∣ A ( 0 ) e σ d ≤ C D L L ! . (1+ML)^L\sum_d\frac{|b_d(0)|}{A(0)}e^{\sigma d}\le CD^LL!. ( 1 + M L ) L d ∑ A ( 0 ) ∣ b d ( 0 ) ∣ e σ d ≤ C D L L ! .
This gives a Gevrey-1 remainder for the whole scalar sum, including invalid degrees. Its coefficient series is precisely C 0 ( t / ( 1 − t ) ; 0 ) C_0(t/(1-t);0) C 0 ( t / ( 1 − t ) ; 0 ) : the child substitution in (D.17) becomes ( t / ( 1 − t ) ) / ( 1 + t / ( 1 − t ) ) = t (t/(1-t))/(1+t/(1-t))=t ( t / ( 1 − t )) / ( 1 + t / ( 1 − t )) = t .
The other two exact factors in (148) are 1 + 1 / h \sqrt{1+1/h} 1 + 1/ h , with a convergent expansion, and the n = 1 n=1 n = 1 ratio, with the marked estimate at zero. All three exact factors are bounded. The finite-product identity used in Proposition D.3 contains one remainder per term; its factorial weights obey j ! ( L − j ) ! ≤ L ! j!(L-j)! \le L! j ! ( L − j )! ≤ L ! . Applied twice, it gives the same bound for their product and exactly the coefficients in (149) . Take common constants with the n ≥ 1 n \ge1 n ≥ 1 case. The coefficient bound ∣ A j ( g , n ) ∣ ≤ C B j j ! |A_j(g,n)| \le CB^j j! ∣ A j ( g , n ) ∣ ≤ C B j j ! follows from those of Theorem 10.2 and Proposition D.9 . Finally, for K = ⌊ c ′ ( 2 g − 1 ) ⌋ K=\lfloor c'(2g-1)\rfloor K = ⌊ c ′ ( 2 g − 1 )⌋ with c ′ ≤ c c' \le c c ′ ≤ c chosen after the final base B B B so that B c ′ < 1 Bc'<1 B c ′ < 1 , the inequality B L L ! h − L ≤ ( B L / h ) L B^L L!h^{-L}\le(BL/h)^L B L L ! h − L ≤ ( B L / h ) L gives the exponentially small error. □
In particular, the explicit first correction (E.29) gives a relative error O ( g − 2 ) O(g^{-2}) O ( g − 2 ) uniformly over all cusp counts. At the unmarked endpoint, by uniqueness of asymptotic expansions, a 0 , 1 = 7 / 6 − 17 / ( 3 π 2 ) a_{0,1}=7/6-17/(3\pi^2) a 0 , 1 = 7/6 − 17/ ( 3 π 2 ) , the first coefficient of (11.5) once h = 2 g − 1 h=2g-1 h = 2 g − 1 is substituted. Equivalently, by (149) and β 0 ( 0 ) = 4 / π 2 − 1 / 2 \beta_0(0)=4/\pi^2-1/2 β 0 ( 0 ) = 4/ π 2 − 1/2 from (92) , L 1 , 1 ( 0 ) = 1 / 6 − 5 / ( 3 π 2 ) \mathcal{L}_{1,1}(0)=1/6-5/(3\pi^2) L 1 , 1 ( 0 ) = 1/6 − 5/ ( 3 π 2 ) , that is,
∫ 0 y 0 γ 1 ( v ) A ( v ) r ( v ) d v = − 13 60 + 5 3 π 2 . \int_0^{y_0}\frac{\gamma_1(v)}{A(v)r(v)}\,\mathrm{d}v=-\frac{13}{60}+\frac{5}{3\pi^2}. ∫ 0 y 0 A ( v ) r ( v ) γ 1 ( v ) d v = − 60 13 + 3 π 2 5 .
a value that is not computed independently here. The existence of an all-order expansion at a fixed cusp count is the theorem of Mirzakhani–Zograf [147]; here the coefficients and remainders remain controlled as the cusp count crosses every density scale. This completes the proofs of Theorems 1.1 and 1.3 . The remaining sections of this part treat the diagonal chain and the coefficient ratios, which use the volume-ratio limits and the harmonic comparison respectively; the volume ratios of Theorem 1.3 , together with Proposition 9.4 , are the inputs that Parts 2 and 3 take from the volume theory, beginning in § 4.1 .
The full diagonal chain
This section proves Theorem 12.1, the logarithmic rate along the diagonal chain announced in the introduction, after Theorem 1.4 . The absolute expansion of Theorem 1.1 assumes g → ∞ g\to\infty g → ∞ . Along the chain V G , 0 , V G − 1 , 2 , … , V 0 , 2 G V_{G,0},V_{G-1,2},\ldots,V_{0,2G} V G , 0 , V G − 1 , 2 , … , V 0 , 2 G , on which the Euler characteristic is constant and genus is traded for pairs of cusps, one also needs the bounded-genus end. To reach it we sum the logarithms of the cusp ratios of Theorem 1.3 , whose profile C C C is expressed through the parameter functions of Lemma 3.1 , and compare the result with the fixed-genus asymptotics of Manin and Zograf [136]. The result has only logarithmic precision even though the cusp ratios are known with relative errors, because summing n n n of their logarithms accumulates an error n ϵ g n\epsilon_g n ϵ g with ϵ g → 0 \epsilon_g\to0 ϵ g → 0 . The proof is given in full here; apart from the limits in Theorem 1.3 and the classical fixed-cusp and fixed-genus asymptotics, it uses only the parameter identities of Lemma 3.1 .
Let Y ( u ) Y(u) Y ( u ) correspond to λ = 2 u / ( 1 − u ) \lambda=2u/(1-u) λ = 2 u / ( 1 − u ) and define
F ( t ) = ∫ 0 t log 4 Δ ( Y ( u ) ) π 2 d u ( 0 ≤ t < 1 ) , F ( 1 ) = − 2 log ( 2 x c ) . (150) F(t)=\int_0^t \log\frac{4\Delta(Y(u))}{\pi^2}\,\mathrm{d}u \quad(0\le t<1), \qquad F(1)=-2\log(2x_c). \tag*{(150)} F ( t ) = ∫ 0 t log π 2 4Δ ( Y ( u )) d u ( 0 ≤ t < 1 ) , F ( 1 ) = − 2 log ( 2 x c ) . (150)
The endpoint value agrees with the limit of the integral.
Theorem 12.1 (A logarithmic rate for the full chain). Uniformly over stable ( g , n ) (g,n) ( g , n ) with 2 g + n → ∞ 2g+n\to\infty 2 g + n → ∞ ,
log V g , n ( 2 g + n − 3 ) ! ( 4 π 2 ) 2 g + n − 3 = 2 g + n 2 F ( n 2 g + n ) + o ( 2 g + n ) . (151) \log\frac{V_{g,n}}{(2g+n-3)!(4\pi^2)^{2g+n-3}}=\frac{2g+n}{2}F\left(\frac{n}{2g+n}\right)+o(2g+n). \tag*{(151)} log ( 2 g + n − 3 )! ( 4 π 2 ) 2 g + n − 3 V g , n = 2 2 g + n F ( 2 g + n n ) + o ( 2 g + n ) . (151) Consequently,
sup 0 ≤ j ≤ G ∣ 1 G log V G − j , 2 j V G , 0 − F ( j / G ) ∣ → 0. (152) \sup_{0\le j\le G}\left|\frac{1}{G}\log\frac{V_{G-j,2j}}{V_{G,0}}-F(j/G)\right|\to0. \tag*{(152)} 0 ≤ j ≤ G sup G 1 log V G , 0 V G − j , 2 j − F ( j / G ) → 0. (152)
The second statement determines each entry of the chain up to a factor exp ( o ( G ) ) \exp(o(G)) exp ( o ( G )) , uniformly in its position and including the planar endpoint V 0 , 2 G V_{0,2G} V 0 , 2 G . When the genus tends to infinity, Theorem 1.1 gives far more precise information; the point of Theorem 12.1 is that it also covers bounded genus.
We use the functions T T T , Δ \Delta Δ , I I I , J J J and the profile C C C of § 3.3 , and the variable t = λ / ( 2 + λ ) ∈ [ 0 , 1 ) t=\lambda/(2+\lambda)\in[0,1) t = λ / ( 2 + λ ) ∈ [ 0 , 1 ) ; thus Y ( u ) Y(u) Y ( u ) in (150) is the parameter with λ = 2 u / ( 1 − u ) \lambda=2u/(1-u) λ = 2 u / ( 1 − u ) .
Proof of Theorem 12.1. We first give a closed form for (150) . From (3.1) and differentiation under the integral,
I ′ = − X ′ ( Y ) Δ , J ′ = − I Δ ′ 2 Δ 3 / 2 . I'=-X'(Y)\sqrt{\Delta},\qquad J'=-\frac{I\Delta'}{2\Delta^{3/2}}. I ′ = − X ′ ( Y ) Δ , J ′ = − 2 Δ 3/2 I Δ ′ .
Writing t = β ( Y ) = X / J t=\beta(Y)=X/J t = β ( Y ) = X / J , these identities show that
F ( t ) = − 2 log ( 2 J ( Y ) ) − ( 1 − t ) log 4 Δ ( Y ) π 2 (153) F(t)=-2\log(2J(Y))-(1-t)\log\frac{4\Delta(Y)}{\pi^2} \tag*{(153)} F ( t ) = − 2 log ( 2 J ( Y )) − ( 1 − t ) log π 2 4Δ ( Y ) (153)
has derivative log ( 4 Δ / π 2 ) \log(4\Delta/\pi^2) log ( 4Δ/ π 2 ) . Its value at zero is zero, because J ( 0 ) = 1 / 2 J(0)=1/2 J ( 0 ) = 1/2 and Δ ( 0 ) = π 2 / 4 \Delta(0)=\pi^2/4 Δ ( 0 ) = π 2 /4 ; hence it is precisely (150) . As t → 1 t\to1 t → 1 , J → x c J\to x_c J → x c and ( 1 − t ) log Δ → 0 (1-t)\log\Delta\to0 ( 1 − t ) log Δ → 0 , proving its stated continuous endpoint.
Set
Z g , n = V g , n ( 2 g + n − 3 ) ! ( 4 π 2 ) 2 g + n − 3 , f ( λ ) = log 2 C ( λ ) 4 π 2 ( 2 + λ ) = − log ( 2 J ( Y ) ) . Z_{g,n}=\frac{V_{g,n}}{(2g+n-3)!(4\pi^2)^{2g+n-3}},\qquad f(\lambda)=\log\frac{2C(\lambda)}{4\pi^2(2+\lambda)}=-\log(2J(Y)). Z g , n = ( 2 g + n − 3 )! ( 4 π 2 ) 2 g + n − 3 V g , n , f ( λ ) = log 4 π 2 ( 2 + λ ) 2 C ( λ ) = − log ( 2 J ( Y )) .
Since J ′ > 0 J'>0 J ′ > 0 , f f f decreases from zero to − log ( 2 x c ) -\log(2x_c) − log ( 2 x c ) and has finite total variation. Equation (153) gives the exact identity
∫ 0 λ f ( v ) d v = ( 1 + λ 2 ) F ( λ 2 + λ ) . (154) \int_0^\lambda f(v)\,\mathrm{d}v=\left(1+\frac{\lambda}{2}\right)F\left(\frac{\lambda}{2+\lambda}\right). \tag*{(154)} ∫ 0 λ f ( v ) d v = ( 1 + 2 λ ) F ( 2 + λ λ ) . (154)
For g ≥ 2 g\ge2 g ≥ 2 , the exact adjacent-cusp relation is
log Z g , n = log Z g , 0 + ∑ j = 0 n − 1 log 2 g C g , j 4 π 2 ( 2 g + j − 2 ) . \log Z_{g,n}=\log Z_{g,0}+\sum_{j=0}^{n-1}\log\frac{2gC_{g,j}}{4\pi^2(2g+j-2)}. log Z g , n = log Z g , 0 + j = 0 ∑ n − 1 log 4 π 2 ( 2 g + j − 2 ) 2 g C g , j .
Let ε g → 0 \varepsilon_g\to0 ε g → 0 bound the logarithmic error in the uniform C C C profile. Replacing C C C by C C C , and replacing 2 g + j − 2 2g+j-2 2 g + j − 2 by 2 g + j 2g+j 2 g + j , gives total error O ( n ε g + 1 + log ( 1 + n / g ) ) O(n\varepsilon_g+1+\log(1+n/g)) O ( n ε g + 1 + log ( 1 + n / g )) . The bounded variation of f f f gives the uniform Riemann-sum estimate
∑ j = 0 n − 1 f ( j / g ) = g ∫ 0 n / g f ( v ) d v + O ( 1 ) , \sum_{j=0}^{n-1}f(j/g)=g\int_0^{n/g}f(v)\,\mathrm{d}v+O(1), j = 0 ∑ n − 1 f ( j / g ) = g ∫ 0 n / g f ( v ) d v + O ( 1 ) ,
with a constant independent of n n n . The fixed-cusp closed-volume asymptotic gives log Z g , 0 = o ( g ) \log Z_{g,0}=o(g) log Z g , 0 = o ( g ) . Using (154) , we obtain
log Z g , n = 2 g + n 2 F ( n 2 g + n ) + o ( 2 g + n ) , (155) \log Z_{g,n}=\frac{2g+n}{2}F\left(\frac{n}{2g+n}\right)+o(2g+n), \tag*{(155)} log Z g , n = 2 2 g + n F ( 2 g + n n ) + o ( 2 g + n ) , (155)
uniformly for all n n n as g → ∞ g\to\infty g → ∞ .
The same statement holds uniformly over all stable pairs with 2 g + n → ∞ 2g+n\to\infty 2 g + n → ∞ . To check the remaining case, keep g g g fixed and use the fixed-genus leading cusp asymptotic [136, Section 6], which gives
log Z g , n = − n log ( 2 x c ) + O g ( log ( n + 1 ) ) . \log Z_{g,n}=-n\log(2x_c)+O_g(\log(n+1)). log Z g , n = − n log ( 2 x c ) + O g ( log ( n + 1 )) .
Meanwhile F ( 1 ) − F ( t ) = O ( ( 1 − t ) ∣ log ( 1 − t ) ∣ ) F(1)-F(t)=O((1-t)|\log(1-t)|) F ( 1 ) − F ( t ) = O (( 1 − t ) ∣ log ( 1 − t ) ∣ ) , because Δ ≍ 1 − t \Delta\asymp\sqrt{1-t} Δ ≍ 1 − t by Lemma 3.1 (iii), so the right side of (151) differs from the same leading term by only O g ( log ( n + 1 ) ) O_g(\log(n+1)) O g ( log ( n + 1 )) . Every sequence of stable pairs has a subsequence on which the genus is fixed or tends to infinity. This proves the full uniformity claimed for (151) .
Along ( g , n ) = ( G − j , 2 j ) (g,n)=(G-j,2j) ( g , n ) = ( G − j , 2 j ) , the value 2 g + n = 2 G 2g+n=2G 2 g + n = 2 G and the factorial normalization are constant. Subtracting the closed value log Z G , 0 = o ( G ) \log Z_{G,0}=o(G) log Z G , 0 = o ( G ) from (151) proves (152) , including j = 0 j=0 j = 0 and j = G j=G j = G . This full-chain statement includes bounded genera, whereas the absolute equivalent in Theorem 11.1 is uniform in cusp count as the genus tends to infinity. □
Uniform one-boundary coefficient ratios
This section proves the uniform equivalent (157) in Theorem 1.4 , which answers Mirzakhani’s question on adjacent coefficients of V g , 1 ( L ) V_{g,1}(L) V g , 1 ( L ) ; the quantitative error stated in the theorem is deferred to Appendix F. The coefficient problem uses the first idea of Subsection 1.5 only: the harmonic comparison of Proposition 5.1 at Y = 0 Y=0 Y = 0 , in the one-marking form (63) , but neither the marked ratios nor the extraction theorems. The comparison determines the coefficients only on the exponential scale; log-concavity of the underlying mixed intersection numbers, which follows from the nefness of the tautological classes, converts this logarithmic information into adjacent ratios in the interior. Near the top degree we use an exact intersection number, and near degree zero the bracket difference estimates of Mirzakhani and Zograf [147]. All proofs are given here.
Put N = 3 g − 2 N=3g-2 N = 3 g − 2 and retain the defining normalization
V g , 1 ( L ) = ∑ k = 0 N a g , k ( 2 k + 1 ) ! L 2 k , b g , k = [ τ k ] g , 1 = 4 k a g , k . V_{g,1}(L)=\sum_{k=0}^{N}\frac{a_{g,k}}{(2k+1)!}L^{2k},\qquad b_{g,k}=[\tau_k]_{g,1}=4^k a_{g,k}. V g , 1 ( L ) = k = 0 ∑ N ( 2 k + 1 )! a g , k L 2 k , b g , k = [ τ k ] g , 1 = 4 k a g , k .
For every valid adjacent index 0 ≤ k < N 0\leq k<N 0 ≤ k < N , define t g , k ∈ ( 0 , π ) t_{g,k}\in(0,\pi) t g , k ∈ ( 0 , π ) by
α g , k : = 2 k + 1 2 g − 1 = t g , k 2 sin t g , k sin t g , k − t g , k cos t g , k . (156) \alpha_{g,k}:=\frac{2k+1}{2g-1}=\frac{t_{g,k}^{2}\sin t_{g,k}}{\sin t_{g,k}-t_{g,k}\cos t_{g,k}}. \tag*{(156)} α g , k := 2 g − 1 2 k + 1 = sin t g , k − t g , k cos t g , k t g , k 2 sin t g , k . (156)
The assertion to prove is
sup 0 ≤ k < 3 g − 2 ∣ a g , k / a g , k + 1 4 π 2 / t g , k 2 − 1 ∣ ⟶ 0. (157) \sup_{0\leq k<3g-2}\left|\frac{a_{g,k}/a_{g,k+1}}{4\pi^2/t_{g,k}^{2}}-1\right|\longrightarrow0. \tag*{(157)} 0 ≤ k < 3 g − 2 sup 4 π 2 / t g , k 2 a g , k / a g , k + 1 − 1 ⟶ 0. (157)
Equivalently, the comparison value for b g , k / b g , k + 1 b_{g,k}/b_{g,k+1} b g , k / b g , k + 1 is π 2 / t g , k 2 \pi^2/t_{g,k}^{2} π 2 / t g , k 2 . The ratio at k = N k=N k = N is undefined and is not included. In the boundary-length normalization of V g , 1 ( L ) V_{g,1}(L) V g , 1 ( L ) , the ratio tends to four when k = o ( g ) k=o(g) k = o ( g ) , while it diverges when 1 ≤ N − k = o ( g ) 1\leq N-k=o(g) 1 ≤ N − k = o ( g ) .
Coefficient identities and further inputs
Let
J k = ∫ M g , 1 κ 1 N − k ψ 1 k , m = N − k , h = 2 g − 1. J_k=\int_{\mathcal{M}_{g,1}}\kappa_1^{N-k}\psi_1^k,\qquad m=N-k,\qquad h=2g-1. J k = ∫ M g , 1 κ 1 N − k ψ 1 k , m = N − k , h = 2 g − 1.
The exact coefficient identities are
a g , k = ( 2 k + 1 ) ! ( 2 π 2 ) m m ! J k , a g , k a g , k + 1 = 2 π 2 ( 2 k + 3 ) ( N − k ) J k J k + 1 . (158) a_{g,k}=(2k+1)!\frac{(2\pi^2)^m}{m!}J_k,\qquad
\frac{a_{g,k}}{a_{g,k+1}}=\frac{2\pi^2}{(2k+3)(N-k)}\frac{J_k}{J_{k+1}}. \tag*{(158)} a g , k = ( 2 k + 1 )! m ! ( 2 π 2 ) m J k , a g , k + 1 a g , k = ( 2 k + 3 ) ( N − k ) 2 π 2 J k + 1 J k . (158)
Here J k J_k J k is a mixed κ 1 \kappa_1 κ 1 –ψ 1 \psi_1 ψ 1 intersection number, not a pure ψ \psi ψ intersection number.
We use the following consequences of § 3.7 . Define
H ( z ) = 3 ( sin t − t cos t ) t 3 , t = 2 z , A ~ m ( h ) = [ z m ] H ( z ) − h , H(z)=\frac{3(\sin t-t\cos t)}{t^3},\qquad t=\sqrt{2z},\qquad\widetilde{A}_m(h)=[z^m]H(z)^{-h}, H ( z ) = t 3 3 ( sin t − t cos t ) , t = 2 z , A m ( h ) = [ z m ] H ( z ) − h ,
with the removable value H ( 0 ) = 1 H(0)=1 H ( 0 ) = 1 , and
M g , k = 4 k 3 h Γ ( h ) 4 g ( 2 π 2 ) m A ~ m ( h ) . (159) M_{g,k}=4^k\frac{3^h\Gamma(h)}{4g}(2\pi^2)^m\widetilde{A}_m(h). \tag*{(159)} M g , k = 4 k 4 g 3 h Γ ( h ) ( 2 π 2 ) m A m ( h ) . (159)
Here H = H 0 H=H_0 H = H 0 is the kernel of § 3.1 at Y = 0 Y=0 Y = 0 . The tilde indicates that A ~ m ( h ) \widetilde{A}_m(h) A m ( h ) is a coefficient of the unscaled kernel H 0 H_0 H 0 ; thus A ~ m ( h ) = r 0 − m A m ( h ) \widetilde{A}_m(h)=r_0^{-m}A_m(h) A m ( h ) = r 0 − m A m ( h ) , with A m ( h ) A_m(h) A m ( h ) from (31) at Y = 0 Y=0 Y = 0 and r 0 = π 2 / 2 r_0=\pi^2/2 r 0 = π 2 /2 .
There is an absolute C C C such that, for every sufficiently large g g g , uniformly in 0 ≤ k ≤ N 0\leq k\leq N 0 ≤ k ≤ N ,
exp [ − C log 2 ( g + 1 ) ] ≤ b g , k M g , k ≤ exp ( C g ) . (160) \exp[-C\log^2(g+1)]\leq\frac{b_{g,k}}{M_{g,k}}\leq\exp(C\sqrt{g}). \tag*{(160)} exp [ − C log 2 ( g + 1 )] ≤ M g , k b g , k ≤ exp ( C g ) . (160)
This is the one-marking case of (61) –(63) . At Y = 0 Y=0 Y = 0 , the unscaled amplitude is U 0 = b / [ 4 k ( 2 π 2 ) m ] U^0=b/[4k(2\pi^2)^m] U 0 = b / [ 4 k ( 2 π 2 ) m ] ; undoing that scaling gives exactly (159) , without an additional genus-dependent factor. In particular
sup 0 ≤ k ≤ N 1 g ∣ log b g , k M g , k ∣ ⟶ 0. (161) \sup_{0\leq k\leq N}\frac{1}{g}\left|\log\frac{b_{g,k}}{M_{g,k}}\right|\longrightarrow0. \tag*{(161)} 0 ≤ k ≤ N sup g 1 log M g , k b g , k ⟶ 0. (161)
The lower estimate in (160) will also control the bottom endpoint.
We also use the following three facts, which are independent of the harmonic comparison.
The cotangent classes are nef on the compactification. Indeed, the nef logarithmic relative dualizing class in [118, Theorem 0.4] is the cotangent class at the additional mark under the identification of the universal curve with the moduli space with one additional mark. The logarithmic kappa class is nef as well: for the forgetful universal curve π : M ‾ g , n + 1 → M ‾ g , n \pi:\overline{\mathcal{M}}_{g,n+1}\to\overline{\mathcal{M}}_{g,n} π : M g , n + 1 → M g , n , one has κ 1 = π ∗ ( ψ n + 1 ) \kappa_1=\pi_*(\psi_{n+1}) κ 1 = π ∗ ( ψ n + 1 ) [14]. Its intersection against any complete curve is the square of the pulled-back nef cotangent class on the pulled-back universal surface, and hence is nonnegative. The argument can be carried out after passing to a finite cover supporting a family, and on its projective coarse surface; passage to the coarse space multiplies all mixed intersection numbers by the same positive factor and does not affect the inequality. Hence these nef classes, as rational divisor classes on the projective coarse moduli space, satisfy the Khovanskii–Teissier inequalities [121, Section 1.6]; see also [26, Theorem 2.15]. This gives
J k 2 ≥ J k − 1 J k + 1 , 1 ≤ k < N . (162) J_k^2\geq J_{k-1}J_{k+1},\qquad1\leq k<N. \tag*{(162)} J k 2 ≥ J k − 1 J k + 1 , 1 ≤ k < N . (162)
Consequently q k : = J k / J k + 1 q_k:=J_k/J_{k+1} q k := J k / J k + 1 is nondecreasing. No log-concavity of the normalized bracket sequence is assumed.
Bracket monotonicity and the fixed-marking absolute difference estimate of Mirzakhani–Zograf [147, Theorem 3.2(ii)] give, with c k = b g , k / V g , 1 c_k=b_{g,k}/V_{g,1} c k = b g , k / V g , 1 ,
c 0 = 1 , 0 ≤ c k − c k + 1 ≤ C k + 1 g , 0 ≤ k < N . (163) c_0=1,\qquad0\leq c_k-c_{k+1}\leq C\frac{k+1}{g},\qquad0\leq k<N. \tag*{(163)} c 0 = 1 , 0 ≤ c k − c k + 1 ≤ C g k + 1 , 0 ≤ k < N . (163)
For k ≥ 1 k\geq1 k ≥ 1 , their explicit bound is b g , k − b g , k + 1 ≤ 16 ( ( k + 1 ) V g , 0 + k V g − 1 , 2 ) b_{g,k}-b_{g,k+1}\leq16((k+1)V_{g,0}+kV_{g-1,2}) b g , k − b g , k + 1 ≤ 16 (( k + 1 ) V g , 0 + k V g − 1 , 2 ) ; the fixed-marking volume ratios turn it into (163) . The k = 0 k=0 k = 0 case is the fixed-degree estimate 1 − c 1 = O ( 1 / g ) 1-c_1=O(1/g) 1 − c 1 = O ( 1/ g ) . Summing (163) gives 1 − c k ≤ C k ( k + 1 ) / g 1-c_k\leq Ck(k+1)/g 1 − c k ≤ C k ( k + 1 ) / g . These are absolute, not relative, estimates, so they cannot be divided by an arbitrarily small c k c_k c k .
The fixed-marking leading volume asymptotic gives
V g , 1 ≍ g − 1 / 2 Γ ( 2 g − 1 ) ( 4 π 2 ) 2 g − 2 , (164) V_{g,1}\asymp g^{-1/2}\Gamma(2g-1)(4\pi^2)^{2g-2}, \tag*{(164)} V g , 1 ≍ g − 1/2 Γ ( 2 g − 1 ) ( 4 π 2 ) 2 g − 2 , (164)
and the one-point pure-ψ \psi ψ identity is J N = 1 / ( 24 g g ! ) J_N=1/(24^g g!) J N = 1/ ( 2 4 g g !) . Only the upper bound in (164) is needed below. The exact value of the top ratio is derived below from the ordinary DVV and string equations.
The saddle equation for the degree
The reciprocal of H H H has strictly positive Taylor coefficients and is analytic on a neighborhood of [ 0 , π 2 / 2 ] [0,\pi^2/2] [ 0 , π 2 /2 ] . These facts also follow directly from 2 z H ′ ′ + 5 H ′ + H = 0 2zH''+5H'+H=0 2 z H ′′ + 5 H ′ + H = 0 : writing − H ′ / H = ∑ j ≥ 0 r j z j -H'/H=\sum_{j\geq0}r_jz^j − H ′ / H = ∑ j ≥ 0 r j z j gives r 0 = 1 / 5 r_0=1/5 r 0 = 1/5 and
( 2 j + 5 ) r j = 2 ∑ u + v = j − 1 r u r v > 0 ( j ≥ 1 ) . (2j+5)r_j=2\sum_{u+v=j-1}r_ur_v>0\qquad(j\geq1). ( 2 j + 5 ) r j = 2 u + v = j − 1 ∑ r u r v > 0 ( j ≥ 1 ) .
Hence − log H -\log H − log H and H − 1 H^{-1} H − 1 have positive nonconstant coefficients. The function H H H has no positive real zero up to and including π 2 / 2 \pi^2/2 π 2 /2 , because sin t − t cos t > 0 \sin t-t\cos t>0 sin t − t cos t > 0 for 0 < t ≤ π 0<t\leq\pi 0 < t ≤ π ; by Pringsheim’s theorem this excludes a smaller complex singularity of H − 1 H^{-1} H − 1 .
For 0 < z ≤ π 2 / 2 0<z\leq\pi^2/2 0 < z ≤ π 2 /2 , tilt these coefficients to a probability law with generating function H ( z ) / H ( z w ) H(z)/H(zw) H ( z ) / H ( z w ) . Its mean is
μ ( z ) = − z H ′ ( z ) H ( z ) = 1 2 ( 3 − t 2 sin t sin t − t cos t ) , t = 2 z . (165) \mu(z)=-\frac{zH'(z)}{H(z)}=\frac{1}{2}\left(3-\frac{t^2\sin t}{\sin t-t\cos t}\right),\qquad t=\sqrt{2z}. \tag*{(165)} μ ( z ) = − H ( z ) z H ′ ( z ) = 2 1 ( 3 − sin t − t cos t t 2 sin t ) , t = 2 z . (165)
Its variance is z μ ′ ( z ) > 0 z\mu'(z)>0 z μ ′ ( z ) > 0 . Thus the right side of (156) , as a function of t t t , decreases strictly from three to zero on ( 0 , π ) (0,\pi) ( 0 , π ) , proving existence and uniqueness in (156) . Note also the exact dimension identity
m h = 3 − α g , k 2 , α g , k = 3 − 2 m h . (166) \frac{m}{h}=\frac{3-\alpha_{g,k}}{2},\qquad\alpha_{g,k}=3-\frac{2m}{h}. \tag*{(166)} h m = 2 3 − α g , k , α g , k = 3 − h 2 m . (166)
For m / h m/h m / h in a compact subset of ( 0 , 3 / 2 ) (0,3/2) ( 0 , 3/2 ) , there is a unique saddle z z z with μ ( z ) = m / h \mu(z)=m/h μ ( z ) = m / h . Fourier inversion for the tilted iid sum, whose laws have uniform exponential moments and positive probabilities at zero and one, gives
log A ~ m ( h ) = − h log H ( z ) − m log z + O ( log h ) . (167) \log\widetilde{A}_m(h)=-h\log H(z)-m\log z+O(\log h). \tag*{(167)} log A m ( h ) = − h log H ( z ) − m log z + O ( log h ) . (167)
The error is uniform on such compact subsets.
Interior degrees: exponential comparison and concavity
The idea is that, for concave functions converging pointwise, slopes also converge wherever the limit is differentiable. Suppose k / g → α ∈ ( 0 , 3 ) k/g\to\alpha\in(0,3) k / g → α ∈ ( 0 , 3 ) and let z = t 2 / 2 z=t^2/2 z = t 2 /2 solve μ ( z ) = ( 3 − α ) / 2 \mu(z)=(3-\alpha)/2 μ ( z ) = ( 3 − α ) /2 . Interpolate linearly on the degree mesh the values
f g ( k / g ) = 1 g log J k J N g 2 ( N − k ) . (168) f_g(k/g)=\frac{1}{g}\log\frac{J_k}{J_N g^{2(N-k)}}. \tag*{(168)} f g ( k / g ) = g 1 log J N g 2 ( N − k ) J k . (168)
By (162) , these functions are concave. Equations (158) –(161) , (167) , and Stirling’s formula give their pointwise limit on ( 0 , 3 ) (0,3) ( 0 , 3 ) :
f ( α ) = − 2 log H ( z ) − ( 3 − α ) log z + ( 3 − α ) log ( 3 − α ) + 3 log 6 − α log ( 2 α ) − 2 ( 3 − α ) . (169) \begin{aligned}
f(\alpha)={}&-2\log H(z)-(3-\alpha)\log z+(3-\alpha)\log(3-\alpha)\\
&+3\log6-\alpha\log(2\alpha)-2(3-\alpha). \tag*{(169)}
\end{aligned} f ( α ) = − 2 log H ( z ) − ( 3 − α ) log z + ( 3 − α ) log ( 3 − α ) + 3 log 6 − α log ( 2 α ) − 2 ( 3 − α ) . (169)
Explicitly, the model for the intersection number J k J_k J k is J k m o d = 3 h Γ ( h ) 4 − g m ! A ~ m ( h ) / ( 2 k + 1 ) ! ! J_k^{\mathrm{mod}}=3^h\Gamma(h)4^{-g}m!\widetilde{A}_m(h)/(2k+1)!! J k mod = 3 h Γ ( h ) 4 − g m ! A m ( h ) / ( 2 k + 1 )!! . Its quotient by the top model value is exactly m ! A ~ m ( h ) ( 2 N + 1 ) ! ! / ( 2 k + 1 ) ! ! m!\widetilde{A}_m(h)(2N+1)!!/(2k+1)!! m ! A m ( h ) ( 2 N + 1 )!! / ( 2 k + 1 )!! ; the two occurrences of ( N − k ) log g (N-k)\log g ( N − k ) log g in its Stirling expansion cancel the affine normalization in (168) . The logarithmic errors in (161) , including the top one, vanish after division by g g g .
Differentiation of (169) gives
f ′ ( α ) = log z 2 α ( 3 − α ) . (170) f'(\alpha)=\log\frac{z}{2\alpha(3-\alpha)}. \tag*{(170)} f ′ ( α ) = log 2 α ( 3 − α ) z . (170)
The terms containing d z / d α dz/d\alpha d z / d α cancel by the saddle equation. Concavity of log J k \log J_k log J k now upgrades the exponential-scale information to an adjacent ratio. For fixed small ε > 0 \varepsilon>0 ε > 0 , the slope on [ k / g , ( k + 1 ) / g ] [k/g,(k+1)/g] [ k / g , ( k + 1 ) / g ] is at most the secant slope on [ α − 2 ε , α − ε ] [\alpha-2\varepsilon,\alpha-\varepsilon] [ α − 2 ε , α − ε ] and at least the one on [ α + ε , α + 2 ε ] [\alpha+\varepsilon,\alpha+2\varepsilon] [ α + ε , α + 2 ε ] , for all sufficiently large g g g . These four endpoints are fixed, so pointwise convergence suffices. First let g g g tend to infinity and then let ε \varepsilon ε decrease to zero. Since (169) is differentiable, the adjacent slope converges to (170) . Consequently
J k g 2 J k + 1 ⟶ 2 α ( 3 − α ) z . (171) \frac{J_k}{g^2J_{k+1}}\longrightarrow\frac{2\alpha(3-\alpha)}{z}. \tag*{(171)} g 2 J k + 1 J k ⟶ z 2 α ( 3 − α ) . (171)
Using the exact identity (158) ,
a g , k a g , k + 1 ⟶ 2 π 2 z = 4 π 2 t 2 . (172) \frac{a_{g,k}}{a_{g,k+1}}\longrightarrow\frac{2\pi^2}{z}=\frac{4\pi^2}{t^2}. \tag*{(172)} a g , k + 1 a g , k ⟶ z 2 π 2 = t 2 4 π 2 . (172)
The top endpoint, including bounded codimension
Near the top degree, log-concavity squeezes the ratios between the interior limit and an exact value in codimension one. The one-kappa pushforward identity gives J N − 1 = ⟨ τ 2 τ 3 g − 3 ⟩ g J_{N-1}=\langle\tau_{2}\tau_{3g-3}\rangle_{g} J N − 1 = ⟨ τ 2 τ 3 g − 3 ⟩ g . Applying DVV at the degree-two insertion and then the string equation twice gives
15 J N − 1 = { ( 6 g − 3 ) ( 6 g − 5 ) + 12 g } J N . 15J_{N-1}=\{(6g-3)(6g-5)+12g\}J_{N}. 15 J N − 1 = {( 6 g − 3 ) ( 6 g − 5 ) + 12 g } J N .
The nonseparating term is ( 1 / 2 ) ⟨ τ 0 2 τ 3 g − 3 ⟩ g − 1 = 12 g J N (1/2)\langle\tau_{0}^{2}\tau_{3g-3}\rangle_{g-1}=12gJ_{N} ( 1/2 ) ⟨ τ 0 2 τ 3 g − 3 ⟩ g − 1 = 12 g J N ; separating terms vanish by dimension. At g = 1 g=1 g = 1 , the same identity follows directly from ⟨ τ 0 3 ⟩ 0 = 1 \langle\tau_{0}^{3}\rangle_{0}=1 ⟨ τ 0 3 ⟩ 0 = 1 . Therefore
q N − 1 = 12 g 2 − 12 g + 5 5 , a g , N − 1 a g , N = 2 π 2 ( 12 g 2 − 12 g + 5 ) 5 ( 6 g − 3 ) . (173) q_{N-1}=\frac{12g^{2}-12g+5}{5},\qquad\frac{a_{g,N-1}}{a_{g,N}}=\frac{2\pi^{2}(12g^{2}-12g+5)}{5(6g-3)}. \tag*{(173)} q N − 1 = 5 12 g 2 − 12 g + 5 , a g , N a g , N − 1 = 5 ( 6 g − 3 ) 2 π 2 ( 12 g 2 − 12 g + 5 ) . (173)
Let m = N − k ≥ 1 m=N-k\geq1 m = N − k ≥ 1 and m = o ( g ) m=o(g) m = o ( g ) . For each fixed α < 3 \alpha<3 α < 3 , eventually q ⌊ α g ⌋ ≤ q N − m ≤ q N − 1 q_{\lfloor\alpha g\rfloor}\leq q_{N-m}\leq q_{N-1} q ⌊ α g ⌋ ≤ q N − m ≤ q N − 1 by (162) . As α \alpha α increases to three, the limit on the right side of (171) tends to 12 / 5 12/5 12/5 , because
t 2 sin t sin t − t cos t = 3 − t 2 5 + O ( t 4 ) ( t → 0 ) . (174) \frac{t^{2}\sin t}{\sin t-t\cos t}=3-\frac{t^{2}}{5}+O(t^{4})\qquad(t\to0). \tag*{(174)} sin t − t cos t t 2 sin t = 3 − 5 t 2 + O ( t 4 ) ( t → 0 ) . (174)
The exact top value (173) gives the same limit. Hence
q N − m g 2 ⟶ 12 5 , a g , N − m a g , N − m + 1 ∼ 4 π 2 g 5 m ( m = o ( g ) , m ≥ 1 ) . (175) \frac{q_{N-m}}{g^{2}}\longrightarrow\frac{12}{5},\qquad\frac{a_{g,N-m}}{a_{g,N-m+1}}\sim\frac{4\pi^{2}g}{5m}\qquad(m=o(g),\ m\geq1). \tag*{(175)} g 2 q N − m ⟶ 5 12 , a g , N − m + 1 a g , N − m ∼ 5 m 4 π 2 g ( m = o ( g ) , m ≥ 1 ) . (175)
In particular the statement includes every fixed positive codimension. For fixed m m m , this leading term agrees, after the length normalization above, with the fixed-codimension expansion of Liu–Xu [129, Theorem 1.2]. The same monotonicity argument also covers unbounded m = o ( g ) m=o(g) m = o ( g ) and joins this endpoint to the interior limit.
The lowest degrees
Near degree zero the interior argument gives no information: the comparison (160) determines log b g , k \log b_{g,k} log b g , k only up to an error o ( g ) o(g) o ( g ) , whereas for k = o ( g ) k=o(g) k = o ( g ) the quantities to be determined vary on a smaller scale. Two other inputs replace it. For degrees up to a power of log g \log g log g , the absolute estimate (163) of Mirzakhani and Zograf gives the ratios directly. For larger k = o ( g ) k=o(g) k = o ( g ) , we argue by contradiction: by the monotonicity of the ratios q k q_k q k , a normalized bracket ratio ρ k = b g , k / b g , k + 1 \rho_k=b_{g,k}/b_{g,k+1} ρ k = b g , k / b g , k + 1 bounded away from one would persist over a range of degrees comparable to k k k , and would force b g , k b_{g,k} b g , k to drop below the lower bound (178) .
Set z 0 = π 2 / 2 z_0=\pi^{2}/2 z 0 = π 2 /2 , and let S h S_h S h be the sum of h h h iid variables with generating function H ( z 0 ) / H ( z 0 w ) H(z_0)/H(z_0w) H ( z 0 ) / H ( z 0 w ) . Since H ( z 0 ) = 3 / π 2 H(z_0)=3/\pi^{2} H ( z 0 ) = 3/ π 2 , (159) becomes the exact identity
M g , k = π 2 Γ ( h ) ( 4 π 2 ) h − 1 Pr ( S h = N − k ) . (176) M_{g,k}=\pi^{2}\Gamma(h)(4\pi^{2})^{h-1}\Pr(S_h=N-k). \tag*{(176)} M g , k = π 2 Γ ( h ) ( 4 π 2 ) h − 1 Pr ( S h = N − k ) . (176)
The summands have mean 3 / 2 3/2 3/2 , nonzero variance, span one, and exponential moments. Uniform exponential tilting in a fixed neighborhood of that mean gives, for 0 ≤ k ≤ c 0 g 0\leq k\leq c_0g 0 ≤ k ≤ c 0 g with one fixed small c 0 > 0 c_0>0 c 0 > 0 ,
Pr ( S h = N − k ) ≥ c g − 1 / 2 exp [ − C ( k + 1 ) 2 g ] . (177) \Pr(S_h=N-k)\geq cg^{-1/2}\exp\left[-C\frac{(k+1)^{2}}{g}\right]. \tag*{(177)} Pr ( S h = N − k ) ≥ c g − 1/2 exp [ − C g ( k + 1 ) 2 ] . (177)
Indeed the local limit theorem for the tilted sum gives the factor g − 1 / 2 g^{-1/2} g − 1/2 , while the rate function has a quadratic zero at the mean; the displacement here is exactly k + 1 / 2 k+1/2 k + 1/2 . Combining (160) , (164) , and (176) –(177) yields
0 ≤ log V g , 1 b g , k ≤ C log 2 ( g + 1 ) + C ( k + 1 ) 2 g . (178) 0\leq\log\frac{V_{g,1}}{b_{g,k}}\leq C\log^{2}(g+1)+C\frac{(k+1)^{2}}{g}. \tag*{(178)} 0 ≤ log b g , k V g , 1 ≤ C log 2 ( g + 1 ) + C g ( k + 1 ) 2 . (178)
Write
ρ k = b g , k b g , k + 1 ≥ 1 , ω k = ( 2 k + 3 ) ( N − k ) , ρ k = π 2 q k 2 ω k . \rho_k=\frac{b_{g,k}}{b_{g,k+1}}\geq1,\qquad\omega_k=(2k+3)(N-k),\qquad\rho_k=\frac{\pi^{2}q_k}{2\omega_k}. ρ k = b g , k + 1 b g , k ≥ 1 , ω k = ( 2 k + 3 ) ( N − k ) , ρ k = 2 ω k π 2 q k .
Suppose k = o ( g ) k=o(g) k = o ( g ) but, along a subsequence, ρ k ≥ 1 + η \rho_k \ge1+\eta ρ k ≥ 1 + η for a fixed η > 0 \eta>0 η > 0 . First consider k > log 3 ( g + 1 ) k>\log^3(g+1) k > log 3 ( g + 1 ) ; then (178) is o ( k ) o(k) o ( k ) at every index comparable to k k k . Put δ = η / [ 2 ( 2 + η ) ] \delta=\eta/[2(2+\eta)] δ = η / [ 2 ( 2 + η )] . For k ≤ j ≤ ⌊ ( 1 + δ ) k ⌋ k\le j\le\lfloor(1+\delta)k\rfloor k ≤ j ≤ ⌊( 1 + δ ) k ⌋ ,
ω j ω k ≤ 2 j + 3 2 k + 3 ≤ 1 + δ , ρ j ≥ ρ k ω k ω j ≥ 1 + η / 2. (179) \frac{\omega_j}{\omega_k}\le\frac{2j+3}{2k+3}\le1+\delta,\qquad\rho_j\ge\rho_k\frac{\omega_k}{\omega_j}\ge1+\eta/2. \tag*{(179)} ω k ω j ≤ 2 k + 3 2 j + 3 ≤ 1 + δ , ρ j ≥ ρ k ω j ω k ≥ 1 + η /2. (179)
These inequalities follow from the monotonicity of the raw ratios q j q_j q j . Multiplying the resulting lower bounds over this range forces log ( V g , 1 / b g , ⌊ ( 1 + δ ) k ⌋ + 1 ) ≥ c η k \log(V_{g,1}/b_{g,\lfloor(1+\delta)k\rfloor+1})\ge c_\eta k log ( V g , 1 / b g , ⌊( 1 + δ ) k ⌋ + 1 ) ≥ c η k , contradicting (178) .
For the remaining indices k ≤ log 3 ( g + 1 ) k\le\log^3(g+1) k ≤ log 3 ( g + 1 ) , summing the absolute estimate (163) gives c k + 1 = 1 − o ( 1 ) c_{k+1}=1-o(1) c k + 1 = 1 − o ( 1 ) uniformly, so division in (163) is legitimate. In fact there is a fixed small δ 0 > 0 \delta_0>0 δ 0 > 0 for which
ρ k = 1 + O ( ( k + 1 ) / g ) ( 0 ≤ k ≤ δ 0 g ) . (180) \rho_k=1+O((k+1)/g)\qquad(0\le k\le\delta_0\sqrt{g}). \tag*{(180)} ρ k = 1 + O (( k + 1 ) / g ) ( 0 ≤ k ≤ δ 0 g ) . (180)
This proves
a g , k a g , k + 1 ⟶ 4 whenever k = o ( g ) . (181) \frac{a_{g,k}}{a_{g,k+1}}\longrightarrow4\qquad\text{whenever }k=o(g). \tag*{(181)} a g , k + 1 a g , k ⟶ 4 whenever k = o ( g ) . (181)
No positive lower bound on c k c_k c k was assumed outside the small square-root window.
Combining the degree ranges
Proof of the uniform equivalent (157) . Take any sequence of valid indices and a subsequence on which k / g k/g k / g converges in [ 0 , 3 ] [0,3] [ 0 , 3 ] . An interior limit is covered by (172) , and a zero limit by (181) , while the comparison in (156) also tends to four. If the limit is three, then m = N − k = o ( g ) m=N-k=o(g) m = N − k = o ( g ) and (175) applies. Equations (166) and (174) give
t g , k 2 ∼ 10 m 2 g − 1 , 4 π 2 t g , k 2 ∼ 4 π 2 g 5 m . t_{g,k}^2\sim\frac{10m}{2g-1},\qquad\frac{4\pi^2}{t_{g,k}^2}\sim\frac{4\pi^2g}{5m}. t g , k 2 ∼ 2 g − 1 10 m , t g , k 2 4 π 2 ∼ 5 m 4 π 2 g .
Thus the true and the proposed ratios are asymptotically equal on that subsequence too. A sequence violating the supremum statement (157) would contradict one of these three cases. □
The effective density ( 2 k + 1 ) / ( 2 g − 1 ) (2k+1)/(2g-1) ( 2 k + 1 ) / ( 2 g − 1 ) in (156) is essential near the top: it retains the exact codimension m m m in 3 − α g , k = 2 m / ( 2 g − 1 ) 3-\alpha_{g,k}=2m/(2g-1) 3 − α g , k = 2 m / ( 2 g − 1 ) . Replacing it by k / g k/g k / g before taking the endpoint limit would change the fixed-codimension denominator.
Relation to the instanton prediction
The denominator H H H in (159) gives the effective potential of the instanton calculus for ordinary Weil–Petersson volumes. To see the correspondence, choose the positive sign convention
A i n s t ( w ) = sin ( 2 π w ) − 2 π w cos ( 2 π w ) 4 π 3 = 2 3 w 3 H ( 2 π 2 w 2 ) . A_{\mathrm{inst}}(w)=\frac{\sin(2\pi w)-2\pi w\cos(2\pi w)}{4\pi^3}=\frac{2}{3}w^3H(2\pi^2w^2). A inst ( w ) = 4 π 3 sin ( 2 π w ) − 2 π w cos ( 2 π w ) = 3 2 w 3 H ( 2 π 2 w 2 ) .
With x = w 2 x=w^2 x = w 2 , this is the effective potential of [66, equation (3.15)]; its values at its stationary points are the ordinary Weil–Petersson instanton actions. The ordinary-volume polynomial prediction in [77, Section 3.3] and the one-boundary term in [111, equation (4.22)] contain, up to an overall factor independent of the coefficient degree, the contour expression
Γ ( h ) L ∮ e L w w A i n s t ( w ) h d w , h = 2 g − 1. \frac{\Gamma(h)}{L}\oint\frac{e^{Lw}}{wA_{\mathrm{inst}}(w)^h}\,\mathrm{d}w,\qquad h=2g-1. L Γ ( h ) ∮ w A inst ( w ) h e L w d w , h = 2 g − 1.
The coefficient of L 2 k L^{2k} L 2 k , multiplied by ( 2 k + 1 ) ! (2k+1)! ( 2 k + 1 )! , is proportional to
Γ ( h ) Res w = 0 w 2 k A i n s t ( w ) − h d w . \Gamma(h)\mathop{\operatorname{Res}}_{w=0}w^{2k}A_{\mathrm{inst}}(w)^{-h}\,\mathrm{d}w. Γ ( h ) Res w = 0 w 2 k A inst ( w ) − h d w .
Since 3 h − 2 k − 1 = 2 m 3h-2k-1=2m 3 h − 2 k − 1 = 2 m , expansion at zero gives the exact identity
Γ ( h ) Res w = 0 w 2 k A i n s t ( w ) − h d w = Γ ( h ) ( 3 / 2 ) h ( 2 π 2 ) m A ~ m ( h ) = 2 4 − k M g , k . \Gamma(h)\mathop{\operatorname{Res}}_{w=0}w^{2k}A_{\mathrm{inst}}(w)^{-h}\,\mathrm{d}w
=\Gamma(h)(3/2)^h(2\pi^2)^m\widetilde{A}_m(h)
=2^{4-k}M_{g,k}. Γ ( h ) Res w = 0 w 2 k A inst ( w ) − h d w = Γ ( h ) ( 3/2 ) h ( 2 π 2 ) m A m ( h ) = 2 4 − k M g , k .
Thus the model for the coefficients a g , k a_{g,k} a g , k has precisely the normalization 4 − k M g , k 4^{-k}M_{g,k} 4 − k M g , k , up to the degree-independent overall factor in the prediction. Moreover, the exponent in the d w / w dw/w d w / w representation has its stationary point at
2 k + 1 h = w A i n s t ′ ( w ) A i n s t ( w ) = t 2 sin t sin t − t cos t , t = 2 π w . \frac{2k+1}{h}=\frac{wA_{\mathrm{inst}}'(w)}{A_{\mathrm{inst}}(w)}=\frac{t^2\sin t}{\sin t-t\cos t},\qquad t=2\pi w. h 2 k + 1 = A inst ( w ) w A inst ′ ( w ) = sin t − t cos t t 2 sin t , t = 2 π w .
Thus both the comparison model and the saddle equation already appear in the instanton calculus. The estimates (160) and the concavity argument above justify the asymptotic formula for adjacent ratios of the true mixed intersection numbers, including both endpoint regimes, without using the contour expression.
Cusp weights and the JT spectral edge
Saad, Shenker and Stanford related the genus expansion of Jackiw–Teitelboim (JT) gravity to a matrix integral, with Weil–Petersson volumes as its geometric input [167]; Stanford and Witten extended this correspondence to the other random-matrix ensembles [175]. See [139] for a review. The parameter Y Y Y describes a spectral edge in its defect deformation, which was introduced by Maxfield–Turiaci and by Witten [137, 181]. We record the normalizations that relate our uniform estimates to that model. Nothing in this subsection is used in the proofs; its purpose is to explain why the endpoint Y = y 0 Y=y_0 Y = y 0 , at which the constant of § 3.19 was fixed, is the point where Painlevé I appears. A defect of opening angle 2 π α 2\pi\alpha 2 π α is a cusp at α = 0 \alpha=0 α = 0 [61, Section 2]. Let λ d e f \lambda_{\mathrm{def}} λ def be its weight, not to be confused with the cusp density λ = n / g \lambda=n/g λ = n / g , and use the convention 2 γ = 1 2\gamma=1 2 γ = 1 of Maxfield–Turiaci [137]. Their string equation at α = 0 \alpha=0 α = 0 and its branch connected to the undeformed edge are
E 0 2 π I 1 ( 2 π E 0 ) + λ d e f = 0 , E 0 = − Y π 2 , x = 2 π 2 λ d e f . \frac{\sqrt{E_0}}{2\pi}I_1(2\pi\sqrt{E_0})+\lambda_{\mathrm{def}}=0,\qquad E_0=-\frac{Y}{\pi^2},\qquad x=2\pi^2\lambda_{\mathrm{def}}. 2 π E 0 I 1 ( 2 π E 0 ) + λ def = 0 , E 0 = − π 2 Y , x = 2 π 2 λ def .
The Bessel expressions are interpreted by their entire power series. Substituting gives exactly X ( Y ) = x X(Y)=x X ( Y ) = x . Moreover, the definition (27) gives, for g ≥ 2 g\ge2 g ≥ 2 ,
∑ n ≥ 0 V g , n λ d e f n n ! = ( 2 π 2 ) 3 g − 3 Φ g ( x ) . \sum_{n\ge0}V_{g,n}\frac{\lambda_{\mathrm{def}}^n}{n!}=(2\pi^2)^{3g-3}\Phi_g(x). n ≥ 0 ∑ V g , n n ! λ def n = ( 2 π 2 ) 3 g − 3 Φ g ( x ) .
Thus a fixed cusp count is a coefficient of the defect-weighted amplitude, whereas Y Y Y parametrizes its spectral edge.
The genus-zero density, with its overall topological factor removed, can be written using our kernel. Specializing the Abel integral in [137, equation (3.50)] to α = 0 \alpha=0 α = 0 gives
ρ 0 ( E ) = 1 4 π ∫ E 0 E I 0 ( 2 π u ) E − u d u = q 2 π K Y ( − 2 π 2 q ) , q = E − E 0 ≥ 0. \rho_0(E)=\frac{1}{4\pi}\int_{E_0}^{E}\frac{I_0(2\pi\sqrt{u})}{\sqrt{E-u}}\,du=\frac{\sqrt{q}}{2\pi}K_Y(-2\pi^2q),\qquad q=E-E_0\ge0. ρ 0 ( E ) = 4 π 1 ∫ E 0 E E − u I 0 ( 2 π u ) d u = 2 π q K Y ( − 2 π 2 q ) , q = E − E 0 ≥ 0.
For the second equality set u = E 0 + q ( 1 − s 2 ) u=E_0+q(1-s^2) u = E 0 + q ( 1 − s 2 ) and use (8) . At Y = 0 Y=0 Y = 0 this is the JT density sinh ( 2 π E ) / ( 4 π 2 ) \sinh(2\pi\sqrt{E})/(4\pi^2) sinh ( 2 π E ) / ( 4 π 2 ) . At Y = y 0 Y=y_0 Y = y 0 , put b = − X ′ ′ ( y 0 ) > 0 b=-X''(y_0)>0 b = − X ′′ ( y 0 ) > 0 . Since a ( y 0 ) = 0 a(y_0)=0 a ( y 0 ) = 0 , expansion of K y 0 K_{y_0} K y 0 gives
ρ 0 ( E 0 + q ) = π b 3 q 3 / 2 + O ( q 5 / 2 ) . \rho_0(E_0+q)=\frac{\pi b}{3}q^{3/2}+O(q^{5/2}). ρ 0 ( E 0 + q ) = 3 π b q 3/2 + O ( q 5/2 ) .
The square-root edge has become a cubic critical edge, the local geometry associated with the ( 2 , 3 ) (2,3) ( 2 , 3 ) critical point, whose string equation is Painlevé I [28, 60, 79]. This gives a physical interpretation of the endpoint at which the normalized intersection recursion reduces to pure-ψ \psi ψ intersections.
The exponent of the large-genus action agrees with this critical geometry. With δ = y 0 − Y \delta=y_0-Y δ = y 0 − Y , the endpoint expansions of § 3.3 give
x c − x = b 2 δ 2 ( 1 + O ( δ ) ) , Z = 2 3 b 5 δ 5 / 2 ( 1 + O ( δ ) ) = 2 9 / 4 3 5 b 1 / 4 ( x c − x ) 5 / 4 ( 1 + O ( x c − x ) ) . x_c-x=\frac{b}{2}\delta^2(1+O(\delta)),\qquad\mathcal{Z}=\frac{2\sqrt{3}b}{5}\delta^{5/2}(1+O(\delta))=\frac{2^{9/4}\sqrt{3}}{5b^{1/4}}(x_c-x)^{5/4}(1+O(\sqrt{x_c-x})). x c − x = 2 b δ 2 ( 1 + O ( δ )) , Z = 5 2 3 b δ 5/2 ( 1 + O ( δ )) = 5 b 1/4 2 9/4 3 ( x c − x ) 5/4 ( 1 + O ( x c − x )) .
Theorem 8.2 retains the full action Z ( Y ) \mathcal{Z}(Y) Z ( Y ) uniformly as Y → y 0 Y\to y_0 Y → y 0 . For a fixed marking count, keeping only its leading critical power would introduce the factor exp ( O ( g x c − x ) ) \exp(O(g\sqrt{x_c-x})) exp ( O ( g x c − x )) , and this replacement need not be uniform in joint limits. Formal genus counting balances ℏ \hbar ℏ against ( x c − x ) 5 / 4 (x_c-x)^{5/4} ( x c − x ) 5/4 ; here, however, g g g is not the inverse string coupling. A double-scaled sum over genera would need a separate summation prescription and estimates beyond the coefficient asymptotics proved here.
Part 2. Spectral laws of random surfaces
Mirzakhani’s integration formula turns volume ratios into expected counts of short curves. The task in this part is to identify which of these configurations produce the low eigenvalues and to exclude other modes. We use the uniform ratios of Theorem 1.3 , the bounded-boundary factorization of Proposition 9.4 , the positive convolution factor of § 3.25.1 and the all-marking envelope (D.35). Huang’s trace expansion [101] supplies the principal external spectral input; its precise hypotheses and the extensions needed here are given in Appendices § 10.10 , S and § 10.15 .
The volume estimates feed two spectral questions: the law of the few lowest eigenvalues, and the density of the many eigenvalues in a fixed-energy window. The two questions require different normalizations and different modes of convergence; Figure 2 shows the two operations on the same counting measure. Throughout this part, eigenvalue counting measures include multiplicities and exclude the constant mode.
Figure 2. Two observables above the square-root cusp scale. The upper branch applies in the ranges of Theorems 17.1 and 18.3 . The lower branch divides by d g = n 2 / g d_g=n^2/g d g = n 2 / g , g g g or n n n in the regimes g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g , n / g → λ ∈ ( 0 , ∞ ) n/g\to\lambda\in(0,\infty) n / g → λ ∈ ( 0 , ∞ ) and g = o ( n ) g=o(n) g = o ( n ) of § 4.19 . A deterministic limit of normalized counts does not determine the law of the individual lowest eigenvalues.
The lowest eigenvalues. § 4.1 supplies the short-curve statistics and their first variational consequences. At n ≍ g n\asymp\sqrt{g} n ≍ g , § 4.7 identifies the complete spectrum below 1 / 4 1/4 1/4 by matching funnel trial states with Huang’s positive trace. § 4.12 refines this description near the quarter threshold, including fluctuations and their moments. Above the critical cusp scale and for bounded n / g n/g n / g , § 4.13 proves the Poisson law at the scale A g = 2 π 2 g M 2 ( x g ) A_g=2\pi^2\sqrt{g}M_2(x_g) A g = 2 π 2 g M 2 ( x g ) . Here q g q_g q g is the two-cusp short-curve intensity and M 2 M_2 M 2 is the area-weighted sum over planar cusp types; see (182) and (237) . For n / g → ∞ n/g\to\infty n / g → ∞ , § 4.14 proves that this law persists in a range containing n ≍ g α n\asymp g^\alpha n ≍ g α for every 1 < α < 11 / 9 1<\alpha<11/9 1 < α < 11/9 , and that its normalization collapses when n ≫ g 3 / ( log g ) 4 n\gg g^3/(\log g)^4 n ≫ g 3 / ( log g ) 4 . Both ranges are only sufficient, so the microscopic description has a genuine gap between them. The residual-type conclusion extends to all sequences g → ∞ g\to\infty g → ∞ , n / g → ∞ n/g\to\infty n / g → ∞ (Corollary 18.4).
Spectral measures in fixed windows. Let Ξ X = ∑ 0 < λ j ( X ) < 1 / 4 δ λ j ( X ) \Xi_X=\sum_{0<\lambda_j(X)<1/4}\delta_{\lambda_j(X)} Ξ X = ∑ 0 < λ j ( X ) < 1/4 δ λ j ( X ) . § 4.19 studies Ξ X \Xi_X Ξ X without rescaling energy. Its three main theorems treat the regimes g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g , n / g → λ ∈ ( 0 , ∞ ) n/g\to\lambda\in(0,\infty) n / g → λ ∈ ( 0 , ∞ ) and g = o ( n ) g=o(n) g = o ( n ) , the last of which includes fixed genus and genus zero. After division by n 2 / g n^2/g n 2 / g , g g g and n n n , respectively, the measure converges in L 1 L^1 L 1 against continuous compactly supported tests on [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) , and hence vaguely in probability. In the sublinear regime the limit is the funnel pushforward ((252) ); at positive density all planar filling types contribute. Theorem 19.4 extends the orbital description to a signed full-energy measure with scattering and threshold corrections; Corollary 19.5 is its consequence for the determinant. The proofs of these results are collected in Appendix V.
The distinction between the two kinds of limit is essential. A fixed-window measure records the proportion of eigenvalues in an energy range, and this proportion can be unchanged by a vanishing fraction of much smaller eigenvalues; the microscopic process resolves those lowest levels individually. The fixed-window theorems therefore coexist with the unresolved dense microscopic law.
Previous work on the spectral transition. This subsection places the results of this part next to earlier work, in the order of the phase diagram of Table 1 : closed surfaces and the subcritical range n = o ( g ) n=o(\sqrt{g}) n = o ( g ) , the critical scale n ≍ g n\asymp\sqrt{g} n ≍ g , the supercritical and dense ranges, counts and spectral measures in fixed windows, and the tight length spectrum. The statements themselves are in § 4.1 to § 4.19 ; only the comparison is made here.
Table 1. The spectral phases of Weil–Petersson random surfaces as g → ∞ g\to\infty g → ∞ .
Random hyperbolic surfaces were first studied in a combinatorial model, by Brooks and Makover [29]; the Weil–Petersson model was first studied by Guth, Parlier and Young [86] and by Mirzakhani [144]. For Weil–Petersson random closed surfaces, Mirzakhani established a positive spectral gap with high probability [144, Section 4.5]. Wu and Xue [189] and, independently, Lipnowski and Wright [128] raised the lower bound to 3 / 16 − ε 3/16-\varepsilon 3/16 − ε . Anantharaman and Monk obtained 2 / 9 − ε 2/9-\varepsilon 2/9 − ε , first in the original arXiv version of the first part of their Friedman–Ramanujan series [10] and now in a separate article [12, Theorem 1.1], and then 1 / 4 − ε 1/4-\varepsilon 1/4 − ε [11, Theorem 1.6]; Hide, Macera and Thomas gave a polynomial rate of approach to 1 / 4 1/4 1/4 [96]. In the different model of uniformly random covers of a fixed surface, Magee, Naud and Puder obtained the relative spectral gap 3 / 16 − ε 3/16-\varepsilon 3/16 − ε for covers of a compact surface [134], and Magee, Puder and van Handel proved strong convergence of the corresponding random permutation representations, which yields the near-optimal relative gap 1 / 4 − ε 1/4-\varepsilon 1/4 − ε [135]. For covers of a fixed cusped surface, Hide and Magee proved that no new eigenvalues appear below 1 / 4 − ε 1/4-\varepsilon 1/4 − ε , and deduced that there are closed surfaces of genus tending to infinity with first eigenvalue tending to 1 / 4 1/4 1/4 [97]. Growing cusp counts introduce the separating pants mechanism studied by Shen and Wu [172]. See [151, 160] for surveys of spectral gaps of random hyperbolic surfaces.
The transition occurs at the scale g \sqrt{g} g . Below it, Hide [95, Theorem 1.3] proved that for n = O ( g α ) n=O(g^\alpha) n = O ( g α ) with 0 ≤ α < 1 / 2 0\leq\alpha<1/2 0 ≤ α < 1/2 there is, with probability tending to one, no nonzero eigenvalue below 1 / 4 − ( ( 2 α + 1 ) / 4 ) 2 − ε 1/4-((2\alpha+1)/4)^2-\varepsilon 1/4 − (( 2 α + 1 ) /4 ) 2 − ε . Shen and Wu proved a uniform positive lower bound for the spectral gap throughout the range n = o ( g ) n=o(\sqrt{g}) n = o ( g ) [172, Theorem 3], and He, Wu and Xue improved the bound for n = O ( g α ) n=O(g^\alpha) n = O ( g α ) to 1 / 4 − ( 6 ( 1 − α ) ) − 2 − ε 1/4-(6(1-\alpha))^{-2}-\varepsilon 1/4 − ( 6 ( 1 − α ) ) − 2 − ε , which tends to 5 / 36 − ε 5/36-\varepsilon 5/36 − ε as α → 1 / 2 \alpha\to1/2 α → 1/2 [92, Theorem 1.2]. Huang subsequently proved convergence to 1 / 4 1/4 1/4 throughout n = o ( g ) n=o(\sqrt{g}) n = o ( g ) , together with matching expected-count bounds and the necessity of this cusp condition [101, Theorem 1.1, Proposition 6.14 and Corollary 6.15]. We use his canonical trace coefficients and positive exceptional measure; these subcritical results are not new here.
When n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g}\to c\in(0,\infty) n / g → c ∈ ( 0 , ∞ ) , Shen and Wu [172, Theorem 2] describe the small-value distribution of the Cheeger constant, not a limiting law for Laplace eigenvalues; for closed surfaces of large genus, Budzinski, Curien and Petri bounded the Cheeger constant asymptotically by 2 / π 2/\pi 2/ π [36]. Theorem 15.1 determines the limiting Laplace-eigenvalue process. Its ingredients are the funnel model of § 4.7 , spectral regularization through pinching [170, 25], and saturation of the expected count by localized ground states. The spectral theory of degenerating hyperbolic surfaces of finite area goes back to Hejhal, Wolpert and Ji [94, 185, 186, 109] and, for the eigenvalues below 1 / 4 1/4 1/4 , to Colbois and Courtois [47]; Schulze’s results cover surfaces with funnels, which is the case needed for the model F s F_s F s . The moving-window Theorem 16.1 makes this comparison uniform for growing waists and energies approaching 1 / 4 1/4 1/4 ; in the stated near-critical range it identifies the scale n 2 / g n^2/g n 2 / g and the limiting inverse-square Poisson intensity. Corollary 16.3 explains why rare surfaces give a logarithmic enhancement of the mean gap defect. Above the critical scale the spectral gap vanishes in probability (Corollary 14.1). This answers negatively the question of Shen and Wu whether a uniform spectral gap persists when n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ [172, Question 2]. Shen–Wu proved the vanishing for g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g and observed that Zograf’s deterministic bound [192] covers n / g → ∞ n/g\to\infty n / g → ∞ ; the adjacent-volume limit supplies the remaining range n ≍ g n \asymp g n ≍ g . The statement concerns the Weil–Petersson distribution: Guo, Hua and Shen construct surfaces with n / g n/g n / g tending to any prescribed positive limit and a uniform spectral gap [84].
Hide and Thomas obtained Poisson limits and small eigenvalues at fixed genus [99, 98]; their deterministic linear count below 1 / 4 1/4 1/4 [100, Theorem 1.3]} applies whenever g < a n g < an g < an , for each fixed a > 0 a > 0 a > 0 , and is linear-optimal by the sharp topological upper bound 2 g + n − 2 2g+n-2 2 g + n − 2 of Otal and Rosas [157]. Below every fixed positive threshold, a deterministic linear total count for g = o ( n ) g = o(n) g = o ( n ) is expected to follow from the conformal capacitor bound of [87] (Remark 14.5); Corollary 14.4 covers the full range g ≤ a n g \le an g ≤ an in Weil–Petersson probability. Closed surfaces behave differently: Monk proved that, with probability tending to one, a Weil–Petersson random closed surface has O ( g ( log g ) − 3 / 4 ) O(g(\log g)^{-3/4}) O ( g ( log g ) − 3/4 ) eigenvalues below 1 / 4 1/4 1/4 [149, Theorem 4]. At high energy, Rudnick showed that the Weil–Petersson variance of a smooth linear statistic counting eigenvalues of closed surfaces in a short energy window has the Gaussian orthogonal ensemble value in the double limit, first of large genus and then of high energy [165]; Rudnick and Wigman proved the corresponding central limit theorem [166]. In the range g ≪ n ≪ g \sqrt{g} \ll n \ll g g ≪ n ≪ g , the second-moment volume inputs for our short-curve and small-eigenvalue results already follow from [172]. The uniform volume and boundary theorems supply a single intensity across all cusp densities and the growing-genus extension at positive and infinite density.
Gatignol and Thomas [71] obtain primitive-geodesic Poisson limits in the range g ≪ n ≪ g \sqrt{g} \ll n \ll g g ≪ n ≪ g and, at the critical scale, for lengths below 2 arcsinh 1 2\operatorname{arcsinh} 1 2 arcsinh 1 (their Theorems 1.2–1.3). In the former range they also obtain fixed-threshold lower eigenvalue counts of order n 2 / g n^2/g n 2 / g [71, Theorem 1.10]. These results overlap our short-curve and counting consequences at zero cusp density. Our critical length process is unrestricted in length and retains the pants clusters, and Theorem 15.1 concerns the full subthreshold Laplace spectrum rather than a short-length window or a lower eigenvalue count.
Budd and Lions studied the tight length spectrum in a near-critical Boltzmann model, in which the cusp count is random [32], and asked whether their limit persists for other cusp weights and under the fixed-count measures [32, Open Question 1.6]. Recovering a prescribed cusp count requires a local coefficient estimate, not merely concentration of the random count. Theorem 14.8 supplies the fixed-count extension: tight volume estimates and a positive Fourier bound allow coefficient extraction at every cusp density. Corollary 14.9 then treats every subcritical cusp weight, including arbitrary approach to the critical endpoint. The planar separating curves of Theorem 14.2 lie outside the tight spectrum (Subsection 14.5), so the two length processes describe disjoint families of curves.
Short curves and small eigenvalues
This section converts the volume theory of Part 1 into statistics of short curves and into the eigenvalue upper bounds they imply. Through Mirzakhani’s integration formula [141], expected numbers of short curves are ratios of volumes of cut surfaces, and these ratios are supplied by Theorem 1.3 and, for fixed boundary lengths, by Proposition 9.4 . We first apply the adjacent-volume limit to Shen–Wu’s vanishing-gap argument [172] (Corollary 14.1, Subsection 14.1). We then determine the short-curve process and the systole when n / g → ∞ n/\sqrt{g} \to\infty n / g → ∞ and n / g n/g n / g has a limit (Theorem 14.2 , Subsection 14.2), deduce many small eigenvalues (Corollary 14.4 , Subsection 14.3) and locate them in the residual spectrum (Theorem 14.7 , Subsection 14.4). The last two subsections state the limits of the tight length spectrum at every cusp density (Theorem 14.8 ) and of the full primitive length spectrum at the critical scale n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g} \to c \in(0,\infty) n / g → c ∈ ( 0 , ∞ ) (Theorem 14.10 ); the longer proofs are in Appendices G, H and I. On the spectral side, everything here gives eigenvalue upper bounds or lower counts; the matching upper counts, which show that the short curves exhaust the low spectrum, are the subject of § 4.7 to § 4.19 .
The first positive eigenvalue
The simplest spectral consequence, Corollary 14.1, uses only the discrete cusp ratio D g , n − 1 D_{g,n-1} D g , n − 1 of Theorem 1.3 and holds throughout n / g → ∞ n/\sqrt{g} \to\infty n / g → ∞ . A short curve cutting off two cusps bounds a two-cusp pair of pants of fixed area 2 π 2\pi 2 π , attached to the rest of the surface along a short boundary; the Cheeger ratio, and hence the Rayleigh quotient, is therefore small. Shen–Wu’s second-moment argument makes such a curve likely. Shen and Wu asked for the adjacent-volume limit needed to extend this argument beyond n = o ( g ) n=o(g) n = o ( g ) [172, Question 3], and Theorem 1.3 provides that limit along every sequence with g → ∞ g\to\infty g → ∞ .
Corollary 14.1 (Vanishing spectral gap). Suppose that g → ∞ g\to\infty g → ∞ and that the integer cusp counts satisfy n ( g ) / g → ∞ n(g)/\sqrt{g}\to\infty n ( g ) / g → ∞ . Then, for every ε > 0 \varepsilon>0 ε > 0 ,
P g , n ( g ) W P { X : Δ X has an eigenvalue in ( 0 , ε ) } ⟶ 1. \mathbb{P}^{\mathrm{WP}}_{g,n(g)}\{X:\Delta_X\text{ has an eigenvalue in }(0,\varepsilon)\}\longrightarrow1. P g , n ( g ) WP { X : Δ X has an eigenvalue in ( 0 , ε )} ⟶ 1.
Proof of Corollary 14.1. Put n = n ( g ) n=n(g) n = n ( g ) and
L g = ( g n ) 1 / 2 . L_g=\left(\frac{\sqrt{g}}{n}\right)^{1/2}. L g = ( n g ) 1/2 .
Let N g ( X ) N_g(X) N g ( X ) count the simple closed geodesics of length at most L g L_g L g whose complement has components of topological types ( 0 , 3 ) (0,3) ( 0 , 3 ) and ( g , n − 1 ) (g,n-1) ( g , n − 1 ) . Proposition 35 of [172], in the notation of its arXiv version 4, states under the sole assumption n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ that
lim inf g → ∞ P g , n W P { N g ≥ 1 } ≥ lim inf g → ∞ V g , n − 1 2 V g , n V g , n − 2 . \liminf_{g\to\infty}\mathbb{P}^{\mathrm{WP}}_{g,n}\{N_g\geq1\}\geq\liminf_{g\to\infty}\frac{V_{g,n-1}^{2}}{V_{g,n}V_{g,n-2}}. g → ∞ lim inf P g , n WP { N g ≥ 1 } ≥ g → ∞ lim inf V g , n V g , n − 2 V g , n − 1 2 .
The ratio on the right is D g , n − 1 D_{g,n-1} D g , n − 1 , which tends to one by Theorem 1.3 . Thus P g , n W P { N g ≥ 1 } → 1 \mathbb{P}^{\mathrm{WP}}_{g,n}\{N_g\geq1\}\to1 P g , n WP { N g ≥ 1 } → 1 . The same conclusion also follows within this paper: N g N_g N g is the count N L g N_{L_g} N L g of Subsection 14.2, with q g q_g q g as in (182) , and since L g → 0 L_g\to0 L g → 0 and q g L g 2 ≍ n g / ( g + n ) → ∞ q_gL_g^2\asymp n\sqrt{g}/(g+n)\to\infty q g L g 2 ≍ n g / ( g + n ) → ∞ , Proposition 14.3 gives N g / ( q g L g 2 ) → 1 N_g/(q_gL_g^2)\to1 N g / ( q g L g 2 ) → 1 in probability. Its proof is the same second-moment argument, with D g , n − 1 → 1 D_{g,n-1}\to1 D g , n − 1 → 1 as the volume input.
Write h ( X ) h(X) h ( X ) for the Cheeger constant. On the event N g ≥ 1 N_g\geq1 N g ≥ 1 , a counted geodesic bounds a hyperbolic pair of pants of area 2 π 2\pi 2 π ; the other component has area 2 π ( 2 g + n − 3 ) ≥ 2 π 2\pi(2g+n-3)\geq2\pi 2 π ( 2 g + n − 3 ) ≥ 2 π for all sufficiently large g g g . This separating curve is an admissible Cheeger cut, so
h ( X ) ≤ L g 2 π . h(X)\leq\frac{L_g}{2\pi}. h ( X ) ≤ 2 π L g .
Let R ( X ) \mathcal{R}(X) R ( X ) be the infimum of the Rayleigh quotient over nonzero H 1 ( X ) H^1(X) H 1 ( X ) functions orthogonal to constants. Buser’s inequality [39, Theorem 7.1], in the form recorded in [172, Theorem 11] for complete noncompact hyperbolic surfaces, gives R ( X ) ≤ c h ( X ) \mathcal{R}(X)\leq c h(X) R ( X ) ≤ c h ( X ) with an absolute constant c c c ; no smallness of h h h is needed, because the Cheeger constant of a finite-area noncompact hyperbolic surface is bounded, so the quadratic term of the compact inequality is absorbed into the linear one. Since L g → 0 L_g\to0 L g → 0 , this upper bound is eventually smaller than both ε \varepsilon ε and 1 / 4 1/4 1/4 . The essential spectrum starts at 1 / 4 1/4 1/4 ; the min–max principle, as recalled in [172, Theorem 7], therefore yields a positive discrete eigenvalue λ 1 ( X ) = R ( X ) < ε \lambda_1(X)=\mathcal{R}(X)<\varepsilon λ 1 ( X ) = R ( X ) < ε on N g ≥ 1 N_g\geq1 N g ≥ 1 , for all sufficiently large g g g . This proves the asserted probability limit. □
The use of D g , n − 1 D_{g,n-1} D g , n − 1 above is the only change to the final step of Shen–Wu’s proof; under their condition n = o ( g ) n=o(g) n = o ( g ) this limit follows from [147, Lemma 5.1], as they note. The short-curve estimate, the Cheeger cut and the passage to a discrete eigenvalue are their geometric and spectral inputs. Corollary 14.1 is the supercritical half of the gap phase diagram (Corollary 15.9); the rate at which the gap vanishes, and the law of the lowest eigenvalues, are determined in § 4.13 .
Short-curve processes and the systole
This subsection determines the limiting process of short separating curves and the systole above the critical scale (Theorem 14.2 ). The input is Mirzakhani’s integration formula ((187) ), whose volume ratios are evaluated by Theorem 1.3 in the form ((189) ); the uniformity in the cusp density comes from the coefficient-ratio bounds of § 3.25 , and the proof is in Appendix § 10.1 . The moment estimates are reused in Subsection 14.3 to count many disjoint short pairs of pants. Throughout this and the next subsection, g → ∞ g \to\infty g → ∞ , n / g → ∞ n/\sqrt{g} \to\infty n / g → ∞ , and X X X has law P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP . Let P 2 ( X ) \mathcal{P}_2(X) P 2 ( X ) be the set of unoriented simple closed geodesics bounding a two-cusp pair of pants. Define the deterministic intensity parameter
q g = n ( n − 1 ) 8 g C ( n / g ) . (182) q_g = \frac{n(n-1)}{8gC(n/g)}. \tag*{(182)} q g = 8 g C ( n / g ) n ( n − 1 ) . (182)
The profile endpoints in Theorem 1.3 imply q g ≍ n 2 / ( g + n ) → ∞ q_g \asymp n^2/(g+n) \to\infty q g ≍ n 2 / ( g + n ) → ∞ . In particular q g ∼ n 2 / ( 32 π 2 g ) q_g \sim n^2/(32\pi^2 g) q g ∼ n 2 / ( 32 π 2 g ) when n / g → 0 n/g \to0 n / g → 0 , and q g ∼ x c n / ( 8 π 2 ) q_g \sim x_c n/(8\pi^2) q g ∼ x c n / ( 8 π 2 ) when n / g → ∞ n/g \to\infty n / g → ∞ .
We write PPP ( ν ) \operatorname{PPP}(\nu) PPP ( ν ) for a Poisson point process with intensity measure ν \nu ν .
Theorem 14.2 (Short-curve processes and the systole). Suppose g → ∞ g \to\infty g → ∞ and n / g → ∞ n/\sqrt{g} \to\infty n / g → ∞ , and let X X X have law P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP .
(i) Curves cutting off two cusps. The random locally finite counting measures
∑ γ ∈ P 2 ( X ) δ q g ℓ γ ⟹ PPP ( 2 t d t ) on [ 0 , ∞ ) (183) \sum_{\gamma\in\mathcal{P}_2(X)} \delta_{\sqrt{q_g}\ell_\gamma} \Longrightarrow\operatorname{PPP}(2t\,dt) \text{ on } [0,\infty) \tag*{(183)} γ ∈ P 2 ( X ) ∑ δ q g ℓ γ ⟹ PPP ( 2 t d t ) on [ 0 , ∞ ) (183) converge in distribution for the vague topology. Thus counts in finitely many disjoint bounded intervals [ a i , b i ] [a_i,b_i] [ a i , b i ] converge jointly to independent Poisson variables with means b i 2 − a i 2 b_i^2-a_i^2 b i 2 − a i 2 . If ℓ min ( 2 ) \ell_{\min}^{(2)} ℓ m i n ( 2 ) is the shortest length in P 2 ( X ) \mathcal{P}_2(X) P 2 ( X ) , with value infinity when this set is empty, then
P g , n W P { q g ℓ min ( 2 ) > t } ⟶ e − t 2 ( t ≥ 0 ) . \mathbb{P}^{\mathrm{WP}}_{g,n}\{\sqrt{q_g}\ell_{\min}^{(2)}>t\}\longrightarrow e^{-t^2}\qquad(t\geq0). P g , n WP { q g ℓ m i n ( 2 ) > t } ⟶ e − t 2 ( t ≥ 0 ) . (ii) All planar cusp types. Let P k ( X ) \mathcal{P}_k(X) P k ( X ) consist of simple geodesics whose genus-zero side contains exactly k k k cusps. If n / g → λ ∈ [ 0 , ∞ ] n/g \to\lambda\in[0,\infty] n / g → λ ∈ [ 0 , ∞ ] , put x λ = X ( Y λ ) x_\lambda=X(Y_\lambda) x λ = X ( Y λ ) , where λ ( Y λ ) = λ \lambda(Y_\lambda)=\lambda λ ( Y λ ) = λ , with x 0 = 0 x_0=0 x 0 = 0 and x ∞ = x c x_\infty=x_c x ∞ = x c . Define
w k ( x ) = 2 V 0 , k + 1 k ! ( x 2 π 2 ) k − 2 , W ( x ) = ∑ k ≥ 2 w k ( x ) = 2 Φ ′ ( x ) x 2 . (184) w_k(x)=\frac{2V_{0,k+1}}{k!}\left(\frac{x}{2\pi^2}\right)^{k-2},\qquad W(x)=\sum_{k\geq2}w_k(x)=\frac{2\Phi'(x)}{x^2}. \tag*{(184)} w k ( x ) = k ! 2 V 0 , k + 1 ( 2 π 2 x ) k − 2 , W ( x ) = k ≥ 2 ∑ w k ( x ) = x 2 2 Φ ′ ( x ) . (184) Here w 2 ( 0 ) = W ( 0 ) = 1 w_2(0)=W(0)=1 w 2 ( 0 ) = W ( 0 ) = 1 , w k ( 0 ) = 0 w_k(0)=0 w k ( 0 ) = 0 for k > 2 k>2 k > 2 , and W ( x c ) < ∞ W(x_c)<\infty W ( x c ) < ∞ . Then
∑ k ≥ 2 ∑ γ ∈ P k ( X ) δ ( q g ℓ γ , k ) ⟹ PPP ( 2 t d t ⊗ ∑ k ≥ 2 w k ( x λ ) δ k ) . (185) \sum_{k\geq2}\sum_{\gamma\in\mathcal{P}_k(X)}
\delta_{(\sqrt{q_g}\ell_\gamma,k)}
\Longrightarrow
\operatorname{PPP}\left(2t\,dt\otimes\sum_{k\geq2}w_k(x_\lambda)\delta_k\right). \tag*{(185)} k ≥ 2 ∑ γ ∈ P k ( X ) ∑ δ ( q g ℓ γ , k ) ⟹ PPP ( 2 t d t ⊗ k ≥ 2 ∑ w k ( x λ ) δ k ) . (185) Convergence is for the vague topology on the product with the discrete type space. The projection to lengths also converges, to PPP ( 2 W ( x λ ) t d t ) \operatorname{PPP}(2W(x_\lambda)t\,dt) PPP ( 2 W ( x λ ) t d t ) .
(iii) Primitive geodesics and the systole. Under the density-limit assumption of (ii), its projected process is also the limit of the process of all unoriented primitive closed geodesics, at the same rescaling. Writing sys ( X ) \operatorname{sys}(X) sys ( X ) for the shortest closed geodesic length, we have
P g , n W P { q g sys ( X ) > t } ⟶ e − W ( x λ ) t 2 ( t ≥ 0 ) . \mathbb{P}^{\mathrm{WP}}_{g,n}\{\sqrt{q_g}\operatorname{sys}(X)>t\}\longrightarrow e^{-W(x_\lambda)t^2}\qquad(t\geq0). P g , n WP { q g sys ( X ) > t } ⟶ e − W ( x λ ) t 2 ( t ≥ 0 ) . With probability tending to one every systolic curve is planar separating. Select one by any measurable tie rule, and let K g K_g K g be its planar-side cusp count, with value zero if it is not planar. Then ( q g sys ( X ) , K g ) (\sqrt{q_g}\operatorname{sys}(X),K_g) ( q g sys ( X ) , K g ) converges jointly to a pair ( Z , K ) (Z,K) ( Z , K ) with
P { Z ∈ d t , K = k } = 2 t w k ( x λ ) e − W ( x λ ) t 2 d t , \mathbb{P}\{Z\in dt,K=k\}=2t\,w_k(x_\lambda)e^{-W(x_\lambda)t^2}\,dt, P { Z ∈ d t , K = k } = 2 t w k ( x λ ) e − W ( x λ ) t 2 d t , P { K = k } = w k ( x λ ) W ( x λ ) ( t > 0 , k ≥ 2 ) . (186) \mathbb{P}\{K=k\}=\frac{w_k(x_\lambda)}{W(x_\lambda)}
\qquad(t>0,\ k\geq2). \tag*{(186)} P { K = k } = W ( x λ ) w k ( x λ ) ( t > 0 , k ≥ 2 ) . (186)
The method of factorial moments for curves separating off a prescribed number of cusps has a direct precedent in Hide–Thomas [99, Theorem 1.7] for punctured spheres, where summed types and nested configurations are also treated [99, Remark 1.9 and Proposition 3.6]. Here the uniform volume ratios give the entire family of type intensities when g → ∞ g\to\infty g → ∞ , n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ and n / g → λ ∈ [ 0 , ∞ ] n/g\to\lambda\in[0,\infty] n / g → λ ∈ [ 0 , ∞ ] .
These planar curves lie outside the tight spectrum of [32, Definition 2.1], which is treated separately in Subsection 14.5. The all-primitive process of Mirzakhani–Petri [146, Theorem 4.1] concerns closed surfaces at fixed physical lengths, whereas the full process here is observed on a shrinking scale; a tight-spectrum limit at its longer natural scale is a separate question. The same Poisson law, with the Mirzakhani–Petri intensity, also governs short cycles on random unicellular maps and random metric maps of large genus [108, 18].
Factorial moments. For sufficiently small L L L , the collar lemma forces all curves counted by N L = # { γ ∈ P 2 ( X ) : ℓ γ ≤ L } N_L = \#\{\gamma\in P_2(X) : \ell_\gamma\le L\} N L = # { γ ∈ P 2 ( X ) : ℓ γ ≤ L } to be disjoint; for closed surfaces, the tangle-free hypothesis of Monk and Thomas [150] gives embedded cylinders around short geodesics up to lengths of order log g \log g log g , which is not needed here. Their pants interiors are also disjoint: otherwise one pants boundary would be an essential geodesic in the interior of the other pants, hence boundary-parallel, contradicting uniqueness of geodesic representatives. Thus their pairs of cusp labels are disjoint, and the complement of r r r such pants has type ( g , n − r ) (g,n-r) ( g , n − r ) . Since g ≥ 2 g \ge2 g ≥ 2 eventually, the exceptional double-pants decomposition of a four-punctured sphere does not arise.
Mirzakhani’s integration formula [141, 144], with the same conventions as [172, Lemma 29], gives, for any fixed r r r and short length windows I 1 , … , I r I_1,\ldots,I_r I 1 , … , I r , the ordered distinct-curve integral
E ∑ γ 1 , … , γ r ∈ P 2 ( X ) d i s t i n c t ∏ j = 1 r 1 ℓ γ j ∈ I j = ( n ) 2 r 2 r V g , n ∫ I 1 × ⋯ × I r V g , n − r ( x 1 , … , x r , 0 n − 2 r ) ∏ j = 1 r x j d x j . (187) \mathbb{E}\sum_{\substack{\gamma_1,\ldots,\gamma_r\in P_2(X)\\\mathrm{distinct}}}\prod_{j=1}^{r}\mathbf{1}_{\ell_{\gamma_j}\in I_j}
=
\frac{(n)_{2r}}{2^r V_{g,n}}\int_{I_1\times\cdots\times I_r}V_{g,n-r}(x_1,\ldots,x_r,0^{n-2r})\prod_{j=1}^{r}x_j\,dx_j . \tag*{(187)} E γ 1 , … , γ r ∈ P 2 ( X ) distinct ∑ j = 1 ∏ r 1 ℓ γ j ∈ I j = 2 r V g , n ( n ) 2 r ∫ I 1 × ⋯ × I r V g , n − r ( x 1 , … , x r , 0 n − 2 r ) j = 1 ∏ r x j d x j . (187)
Here ( n ) 2 r = n ( n − 1 ) ⋯ ( n − 2 r + 1 ) (n)_{2r}=n(n-1)\cdots(n-2r+1) ( n ) 2 r = n ( n − 1 ) ⋯ ( n − 2 r + 1 ) . The factor 2 r 2^r 2 r forgets the order of the two labels within each pair; since the tuple of curves is ordered, there is no r ! r! r ! divisor. Each pants volume is V 0 , 3 = 1 V_{0,3}=1 V 0 , 3 = 1 . Mirzakhani’s boundary estimate [144, Lemma 3.2], in the form of [172, Lemma 14],
1 ≤ V g , m ( x ) V g , m ≤ exp ( 1 2 ∑ j x j ) (188) 1\le\frac{V_{g,m}(\mathbf{x})}{V_{g,m}}\le\exp\left(\frac{1}{2}\sum_j x_j\right) \tag*{(188)} 1 ≤ V g , m V g , m ( x ) ≤ exp ( 2 1 j ∑ x j ) (188)
is uniform in the topological type. For fixed r r r , it replaces the boundary volume in (187) by V g , n − r ( 1 + O r ( L ) ) V_{g,n-r}(1+O_r(L)) V g , n − r ( 1 + O r ( L )) when all I j ⊂ [ 0 , L ] I_j\subset[0,L] I j ⊂ [ 0 , L ] .
Theorem 1.3 also implies, for every fixed r r r ,
V g , n − r V g , n = ( 2 g C ( n / g ) ) − r ( 1 + o ( 1 ) ) . (189) \frac{V_{g,n-r}}{V_{g,n}}=(2g\mathcal{C}(n/g))^{-r}(1+o(1)). \tag*{(189)} V g , n V g , n − r = ( 2 g C ( n / g ) ) − r ( 1 + o ( 1 )) . (189)
Indeed, the left side is a product of r r r adjacent cusp ratios. Replacing ( n − j ) / g (n-j)/g ( n − j ) / g by n / g n/g n / g , for 1 ≤ j ≤ r 1\le j\le r 1 ≤ j ≤ r , changes each profile factor by 1 + o ( 1 ) 1+o(1) 1 + o ( 1 ) : along any subsequence with n / g → λ ∗ ∈ [ 0 , ∞ ] n/g\to\lambda_*\in[0,\infty] n / g → λ ∗ ∈ [ 0 , ∞ ] this follows from the continuity and positivity of C \mathcal{C} C when λ ∗ < ∞ \lambda_*<\infty λ ∗ < ∞ , and from C ( λ ) ∼ π 2 λ x c \mathcal{C}(\lambda)\sim\frac{\pi^2\lambda}{x_c} C ( λ ) ∼ x c π 2 λ together with ( n − j ) / n → 1 (n-j)/n\to1 ( n − j ) / n → 1 when λ ∗ = ∞ \lambda_*=\infty λ ∗ = ∞ .
Idea of the proof. The proof, given in Appendix § 10.1 , computes mixed factorial moments from (187) . Three inputs make the computation uniform in the cusp density. First, a coefficient-ratio bound obtained from the positive convolution factor of § 3.25 shows that curves cutting off many cusps contribute a summable tail, uniformly along the sequence. Second, for finitely many fixed types the disjoint ( s t a r ) (\mathit{star}) ( star ) configurations give the Poisson moments ∏ w k ( x λ ) ( b i 2 − a i 2 ) \prod w_k(x_\lambda)(b_i^2-a_i^2) ∏ w k ( x λ ) ( b i 2 − a i 2 ) , while nested configurations lose a factor n − 1 n^{-1} n − 1 by a count of the root degree in the cut tree. Third, the same positive factor bounds the total contribution of all separating curves with positive genus on both sides by O ( 1 / g ) O(1/g) O ( 1/ g ) , and nonseparating curves are rare on the scale q ˉ g − 1 / 2 \bar q_g^{-1/2} q ˉ g − 1/2 ; thus the systole is planar separating with probability tending to one.
The intensity q g q_g q g and the type weights w k w_k w k are carried through the rest of this part: they define the microscopic spectral scale (237) of § 4.13 , the type process (185) supplies the limiting rates in the proof of Theorem 17.1, and the same weights w k w_k w k govern the cap process of § 4.14 at growing density.
Many small eigenvalues
This subsection turns the short-curve counts into eigenvalue bounds. Proposition 14.3 is a law of large numbers for the number N L N_L N L of short two-cusp pairs of pants, both at thresholds L → 0 L \to0 L → 0 , where (187) and the boundary estimate (188) suffice, and at fixed L L L , where the boundary factorization of Proposition 9.4 is needed. Corollary 14.4 converts the count into many small eigenvalues by half-collar trial functions, throughout n / g → ∞ n/\sqrt{g} \to\infty n / g → ∞ and, below every fixed threshold, throughout g ≤ a n g \le an g ≤ an . These are lower counts; the matching upper counts and the limiting densities are proved in Section 19.
Proposition 14.3 (Many short pants). If L = L g → 0 L = L_g \to0 L = L g → 0 and q g L g 2 → ∞ q_g L_g^2 \to\infty q g L g 2 → ∞ , then
N L g q g L g 2 ⟶ 1 in probability . (190) \frac{N_{L_g}}{q_g L_g^2} \longrightarrow1 \quad\text{in probability}. \tag*{(190)} q g L g 2 N L g ⟶ 1 in probability . (190) Under these assumptions, for each fixed ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) , with probability tending to one there are at least ⌊ ( 1 − ε ) q g L g 2 ⌋ − 1 \lfloor(1-\varepsilon)q_g L_g^2 \rfloor- 1 ⌊( 1 − ε ) q g L g 2 ⌋ − 1 positive discrete eigenvalues in ( 0 , 6 L g ] (0,6L_g] ( 0 , 6 L g ] , counted with multiplicity. For each fixed 0 < L < 2 arcsinh 1 0 < L < 2\operatorname{arcsinh} 1 0 < L < 2 arcsinh 1 , one also has
N L q g B Y n / g ( L ) ⟶ 1 in probability , B Y ( L ) = 2 ∫ 0 L t b Y ( t ) d t , (191) \frac{N_L}{q_g \mathcal{B}_{Y_{n/g}}(L)} \longrightarrow1 \quad\text{in probability}, \qquad\mathcal{B}_Y(L) = 2\int_0^L t b_Y(t)\,\mathrm{d}t, \tag*{(191)} q g B Y n / g ( L ) N L ⟶ 1 in probability , B Y ( L ) = 2 ∫ 0 L t b Y ( t ) d t , (191) where b Y b_Y b Y is the boundary profile in (130) .
Proof. The first two cases of (187) give
m L : = E N L = ( n ) 2 L 2 4 V g , n − 1 V g , n ( 1 + O ( L ) ) ∼ q g L 2 , m_L := \mathbb{E}N_L = \frac{(n)_2L^2}{4}\frac{V_{g,n-1}}{V_{g,n}}(1+O(L)) \sim q_gL^2, m L := E N L = 4 ( n ) 2 L 2 V g , n V g , n − 1 ( 1 + O ( L )) ∼ q g L 2 ,
and
E ( N L ) 2 m L 2 = ( n − 2 ) ( n − 3 ) n ( n − 1 ) D g , n − 1 − 1 ( 1 + O ( L ) ) ⟶ 1. \frac{\mathbb{E}(N_L)_2}{m_L^2} = \frac{(n-2)(n-3)}{n(n-1)}D_{g,n-1}^{-1}(1+O(L)) \longrightarrow1. m L 2 E ( N L ) 2 = n ( n − 1 ) ( n − 2 ) ( n − 3 ) D g , n − 1 − 1 ( 1 + O ( L )) ⟶ 1.
Since m L → ∞ m_L \to\infty m L → ∞ , the relative variance tends to zero, proving (190) . This second-moment argument does not use the point-process limit at a window growing with g g g .
Hide–Thomas [99, Lemma 5.3] prove that N N N such pants boundaries of length at most L L L give N − 1 N-1 N − 1 positive eigenvalues below 6 L 6L 6 L , provided 6 L < 1 / 4 6L < 1/4 6 L < 1/4 . For the reader’s convenience we recall the variational argument. Inside each pants choose a function equal to inward distance ρ \rho ρ on the half-collar 0 ≤ ρ ≤ 1 0 \le\rho\le1 0 ≤ ρ ≤ 1 , equal to one beyond it, and zero outside the pants. For all sufficiently small boundary lengths ℓ \ell ℓ , the collar lemma supplies this embedded half-collar. The function lies in H 1 ( X ) H^1(X) H 1 ( X ) , and
∫ ∣ ∇ f ∣ 2 = ℓ sinh 1 , ∫ ∣ f ∣ 2 ≥ 2 π − ℓ sinh 1. \int|\nabla f|^2 = \ell\sinh1, \qquad\int|f|^2 \ge2\pi-\ell\sinh1. ∫ ∣∇ f ∣ 2 = ℓ sinh 1 , ∫ ∣ f ∣ 2 ≥ 2 π − ℓ sinh 1.
Different pants have disjoint supports up to their boundaries, so the span of these functions has dimension N N N and Rayleigh quotient at most 6 L 6L 6 L for small L L L . Its intersection with the orthogonal complement of constants has dimension at least N − 1 N-1 N − 1 . Since the essential spectrum starts at 1 / 4 1/4 1/4 , the form version of min–max gives the claimed positive discrete eigenvalues. Apply (190) and let g g g grow.
For fixed L L L , the O ( L ) O(L) O ( L ) boundary error above is insufficient. Instead apply Proposition 9.4 with one and two nonzero boundaries to (187) . Uniformly on [ 0 , L ] r [0,L]^r [ 0 , L ] r , for r = 1 , 2 r=1,2 r = 1 , 2 ,
V g , n − r ( t 1 , … , t r , 0 n − 2 r ) V g , n − r = ( 1 + o ( 1 ) ) ∏ i = 1 r b Y n / g ( t i ) . \frac{V_{g,n-r}(t_1,\ldots,t_r,0^{n-2r})}{V_{g,n-r}} = (1+o(1))\prod_{i=1}^r b_{Y_{n/g}}(t_i). V g , n − r V g , n − r ( t 1 , … , t r , 0 n − 2 r ) = ( 1 + o ( 1 )) i = 1 ∏ r b Y n / g ( t i ) .
The total end count here is n − r n-r n − r . Replacing Y ( n − r ) / g Y_{(n-r)/g} Y ( n − r ) / g by Y n / g Y_{n/g} Y n / g costs o ( 1 ) o(1) o ( 1 ) uniformly: both have the same subsequential limit, including either endpoint in [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] , and the boundary profile is uniformly continuous on the fixed length box. Together with (189) , this yields
E N L = ( 1 + o ( 1 ) ) q g B Y n / g ( L ) , E ( N L ) 2 = ( 1 + o ( 1 ) ) q g 2 B Y n / g ( L ) 2 . \mathbb{E}N_L = (1+o(1))q_g\mathcal{B}_{Y_{n/g}}(L), \qquad\mathbb{E}(N_L)_2 = (1+o(1))q_g^2\mathcal{B}_{Y_{n/g}}(L)^2. E N L = ( 1 + o ( 1 )) q g B Y n / g ( L ) , E ( N L ) 2 = ( 1 + o ( 1 )) q g 2 B Y n / g ( L ) 2 .
The combinatorial ratio for the second identity is ( n − 2 ) ( n − 3 ) / ( n ( n − 1 ) ) → 1 (n-2)(n-3)/(n(n-1)) \to1 ( n − 2 ) ( n − 3 ) / ( n ( n − 1 )) → 1 . Since b Y ≥ 1 b_Y \ge1 b Y ≥ 1 , the mean is at least ( 1 + o ( 1 ) ) q g L 2 → ∞ (1+o(1))q_gL^2 \to\infty ( 1 + o ( 1 )) q g L 2 → ∞ . The relative variance tends to zero, proving (191) without a density-limit assumption. □
Corollary 14.4 (Small eigenvalues at shrinking and fixed thresholds). Suppose g → ∞ g \to\infty g → ∞ and n / g → ∞ n/\sqrt{g} \to\infty n / g → ∞ . If j = j ( g ) → ∞ j=j(g) \to\infty j = j ( g ) → ∞ and j = o ( q g ) j=o(q_g) j = o ( q g ) , then
P g , n W P { λ j ( X ) exists and λ j ( X ) ≤ 12 j / q g } → 1 , \mathbb{P}^{\mathrm{WP}}_{g,n}\left\{\lambda_j(X)\text{ exists and }\lambda_j(X)\le12\sqrt{j/q_g}\right\}\to1, P g , n WP { λ j ( X ) exists and λ j ( X ) ≤ 12 j / q g } → 1 , where positive discrete eigenvalues are ordered with multiplicity. For each fixed 0 < ε < 1 / 4 0<\varepsilon<1/4 0 < ε < 1/4 , let Z ε ( X ) = # { j : 0 < λ j ( X ) < ε } Z_\varepsilon(X)=\#\{j:0<\lambda_j(X)<\varepsilon\} Z ε ( X ) = # { j : 0 < λ j ( X ) < ε } , with multiplicity. Then also
P g , n W P { Z ε ( X ) ≥ ε 2 576 q g } → 1. (192) \mathbb{P}^{\mathrm{WP}}_{g,n}\left\{Z_\varepsilon(X)\ge\frac{\varepsilon^2}{576}q_g\right\}\to1. \tag*{(192)} P g , n WP { Z ε ( X ) ≥ 576 ε 2 q g } → 1. (192) Moreover, for every a > 0 a>0 a > 0 and 0 < ε < 1 / 4 0<\varepsilon<1/4 0 < ε < 1/4 there is c ε , a > 0 c_{\varepsilon,a}>0 c ε , a > 0 such that, along every sequence n → ∞ n\to\infty n → ∞ and 0 ≤ g ≤ a n 0\le g\le an 0 ≤ g ≤ an ,
P g , n W P { Z ε ( X ) ≥ c ε , a ( 2 g + n − 2 ) } → 1. (193) \mathbb{P}^{\mathrm{WP}}_{g,n}\left\{Z_\varepsilon(X)\ge c_{\varepsilon,a}(2g+n-2)\right\}\to1. \tag*{(193)} P g , n WP { Z ε ( X ) ≥ c ε , a ( 2 g + n − 2 ) } → 1. (193)
Proof of Corollary 14.4. For j → ∞ j\to\infty j → ∞ with j = o ( q g ) j=o(q_g) j = o ( q g ) , choose L g = 2 j / q g L_g=2\sqrt{j/q_g} L g = 2 j / q g . Then L g → 0 L_g\to0 L g → 0 , q g L g 2 = 4 j q_gL_g^2=4j q g L g 2 = 4 j , and (190) implies N L g ≥ j + 1 N_{L_g}\ge j+1 N L g ≥ j + 1 with probability tending to one. The preceding variational estimate gives λ j ≤ 6 L g = 12 j / q g \lambda_j\le6L_g=12\sqrt{j/q_g} λ j ≤ 6 L g = 12 j / q g .
For the fixed threshold choose L = ε / 12 < 1 / 48 L=\varepsilon/12<1/48 L = ε /12 < 1/48 . The half-collar trial functions above have Rayleigh quotient at most 6 L = ε / 2 < ε 6L=\varepsilon/2<\varepsilon 6 L = ε /2 < ε , so Z ε ≥ N L − 1 Z_\varepsilon\ge N_L-1 Z ε ≥ N L − 1 . By (191) and b Y ( L ) ≥ L 2 b_Y(L)\ge L^2 b Y ( L ) ≥ L 2 , with probability tending to one N L ≥ q g L 2 / 2 N_L\ge q_gL^2/2 N L ≥ q g L 2 /2 . Since q g L 2 → ∞ q_gL^2\to\infty q g L 2 → ∞ , eventually N L − 1 ≥ q g L 2 / 4 = ε 2 q g / 576 N_L-1\ge q_gL^2/4=\varepsilon^2q_g/576 N L − 1 ≥ q g L 2 /4 = ε 2 q g /576 . This proves (192) .
If g → ∞ g\to\infty g → ∞ and g ≤ a n g\le an g ≤ an , put λ = n / g ≥ 1 / a \lambda=n/g\ge1/a λ = n / g ≥ 1/ a . Uniformly in this range,
q g 2 g + n − 2 = ( 1 + o ( 1 ) ) λ 2 8 C ( λ ) ( 2 + λ ) . \frac{q_g}{2g+n-2}=(1+o(1))\frac{\lambda^2}{8C(\lambda)(2+\lambda)}. 2 g + n − 2 q g = ( 1 + o ( 1 )) 8 C ( λ ) ( 2 + λ ) λ 2 .
The last expression is positive and continuous on [ 1 / a , ∞ ) [1/a,\infty) [ 1/ a , ∞ ) , with positive limit x c / ( 8 π 2 ) x_c/(8\pi^2) x c / ( 8 π 2 ) at infinity. Its infimum is therefore positive, and (192) implies (193) in the growing-genus case. For fixed genus, Hide–Thomas [98, Theorem 1.1] give C ( ε ) n C(\varepsilon)n C ( ε ) n eigenvalues below ε \varepsilon ε with probability tending to one, with C ( ε ) > 0 C(\varepsilon)>0 C ( ε ) > 0 independent of the fixed genus. The order of quantifiers in [98, Theorem 1.1] is the one needed here: the constant is chosen first, for every ε > 0 \varepsilon>0 ε > 0 , and the statement then holds for every genus g ≥ 0 g\ge0 g ≥ 0 as n → ∞ n\to\infty n → ∞ . A constant depending on the genus would not suffice, since the subsequence argument below must produce a single c ε , a c_{\varepsilon,a} c ε , a for all bounded genera. Removing the zero eigenvalue and reducing the constant gives (193) in that case as well. Choose c ε , a c_{\varepsilon,a} c ε , a smaller than both constants. Every subsequence of a general genus sequence has a further subsequence with g → ∞ g\to\infty g → ∞ or g g g fixed, completing the proof. □
The spectral mechanism is established in the fixed-genus work of Hide–Thomas [99, Theorem 1.10 and Remark 5.5]; their later bordered-volume asymptotics give linear counts below each fixed positive threshold at fixed genus [98, Theorem 1.1], and in the range g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g the volume inputs of Shen–Wu [172] and the count of Gatignol–Thomas [71, Theorem 1.10] already give the order n 2 / g n^2/g n 2 / g . Here the pants mechanism is combined with the uniform boundary factorization to cover positive and unbounded cusp density. This argument yields neither a matching spectral upper count nor a limiting spectral density; both are obtained in § 4.19 , where the count of order n 2 / g n^2/g n 2 / g is identified with the measure ν ∗ \nu_* ν ∗ of Theorem 19.1 .
Remark 14.5 (A deterministic total count via conformal capacitors). This remark sketches an alternative, deterministic route to a total count of small eigenvalues when g = o ( n ) g=o(n) g = o ( n ) ; it is not used anywhere else in this paper, and the details indicated below are not carried out. A deterministic form of the total-count statement for g = o ( n ) g=o(n) g = o ( n ) expected in [100, Section 1.1] should follow from the conformal capacitor argument of Hassannezhad [87]: for a finite-area surface X X X , its min–max levels, indexed from zero, would satisfy
λ j var ( X ) area ( X ) ≤ C ( g + j + 1 ) . \lambda_j^{\operatorname{var}}(X)\operatorname{area}(X)\le C(g+j+1). λ j var ( X ) area ( X ) ≤ C ( g + j + 1 ) .
To apply the classical argument one would compactify the cusps and choose a smooth conformal reference metric h h h with curvature at least − 1 -1 − 1 and area O ( g + 1 ) O(g+1) O ( g + 1 ) . The original area extends to a finite nonatomic measure on the compactification. For m = j + 1 m=j+1 m = j + 1 , the capacitor construction supplies m m m disjoint plateau functions with mass at least c area ( X ) / m c\operatorname{area}(X)/m c area ( X ) / m and energy at most C ( 1 + area h / m ) C(1+\operatorname{area}_h/m) C ( 1 + area h / m ) . One then has to check that these bounded functions belong to H 1 ( X ) H^1(X) H 1 ( X ) , using that the Dirichlet energy is conformally invariant in dimension two while the cusp ends are noncompact, and that cutting off cusp tails justifies the form domain. Min–max would give the bound. Since levels below 1 / 4 1/4 1/4 are discrete, taking j j j a sufficiently small multiple of ε n \varepsilon n ε n would prove the asserted total count when g = o ( n ) g=o(n) g = o ( n ) .
Equation (193) also covers arbitrary g ≤ a n g\le an g ≤ an in Weil–Petersson probability, below every prescribed ε < 1 / 4 \varepsilon<1/4 ε < 1/4 ; the conformal bound need not do so at a fixed positive genus-to-cusp ratio. Hide–Thomas [100, Theorem 1.3] give a deterministic linear count below 1 / 4 1/4 1/4 throughout that larger range. The residual strengthening in Theorem 14.7 does not follow from a total-count bound.
Many small residual eigenvalues
The count of Corollary 14.4 can be obtained within the residual spectrum (Theorem 14.7 ), in the same regimes n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ and g ≤ a n g\le an g ≤ an . Residual eigenfunctions are the square-integrable residues of Eisenstein series, orthogonal to the cuspidal spectrum [93, 107]; the constant zero mode is excluded from our count. Whether a generic cofinite surface carries cusp forms at all is a classical question going back to Phillips and Sarnak [161]; see also Wolpert [184]. Let R ε ( X ) R_\varepsilon(X) R ε ( X ) be the total residual multiplicity in ( 0 , ε ) (0,\varepsilon) ( 0 , ε ) . Hide–Thomas [98, Section 1.1, after Remark 1.2] ask whether their many small eigenvalues are cusp forms and anticipate a linear residual count. The deterministic Lemma 14.6 gives that count from the short-pants law of Proposition 14.3, without assuming the conjecture of Otal and Rosas [157, Section 6] that a surface of genus g ≥ 2 g\ge2 g ≥ 2 with cusps has at most 2 g − 3 2g-3 2 g − 3 small cuspidal eigenvalues (see [17, (1.1)]); its main external input is the topological eigenvalue bound of Ballmann–Matthiesen–Mondal [16]. The idea is that cusp forms have zero angular mean on cusp horocycles, so on a short two-cusp pair of pants their mass is bounded by a fixed multiple of their energy. Low-energy trial functions supported on the pants are therefore far from the cuspidal spectral subspace, and project injectively onto the residual eigenspace.
Lemma 14.6 (Short pants force residual eigenvalues). Put C ∗ = 4 ( 1 + 8 π ) C_* = 4(1+8\pi) C ∗ = 4 ( 1 + 8 π ) and, for 0 < ε < 1 / 4 0<\varepsilon<1/4 0 < ε < 1/4 , set
L ε = ε 24 ( 1 + C ∗ ε ) . (194) L_\varepsilon=\frac{\varepsilon}{24(1+C_*\varepsilon)}. \tag*{(194)} L ε = 24 ( 1 + C ∗ ε ) ε . (194) If a complete connected finite-area hyperbolic surface contains N N N embedded two-cusp pairs of pants with disjoint interiors and geodesic boundary lengths at most L ε L_\varepsilon L ε , then R ε ( X ) ≥ N − 1 R_\varepsilon(X)\ge N-1 R ε ( X ) ≥ N − 1 .
Proof. We first establish a uniform inequality on a two-cusp pants P s P_s P s with waist s ≤ 1 / 48 s\le1/48 s ≤ 1/48 . Double it across the geodesic boundary to obtain a four-cusped sphere. The distance cutoff used in the proof of Proposition 14.3 vanishes at the waist and has Rayleigh quotient at most 6 s ≤ 1 / 8 6s\le1/8 6 s ≤ 1/8 . Its odd reflection therefore produces a positive eigenvalue below $1/4 in the odd sector of the double. The even sector contains the constant zero mode. By [16, Theorem 1.5 and equation (1.4)], the double has at most two eigenvalues at or below $1/4, counting zero and multiplicity. Hence its even sector has no positive spectrum below 1 / 4 1/4 1/4 . Even reflection identifies this sector with the Neumann form on P s P_s P s ; both its norm and energy are doubled. Consequently
∫ P s ∣ v − v ‾ P s ∣ 2 ≤ 4 E P s ( v ) , v ∈ H 1 ( P s ) . (195) \int_{P_s}\left|v-\overline{v}_{P_s}\right|^2\le4\mathcal{E}_{P_s}(v),\qquad v\in H^1(P_s). \tag*{(195)} ∫ P s ∣ v − v P s ∣ 2 ≤ 4 E P s ( v ) , v ∈ H 1 ( P s ) . (195)
This applies the topological eigenvalue bound to the boundaryless double, not to a Neumann problem with boundary.
The standard cusp neighborhoods bounded by length-one horocycles are embedded [100, Section 4.1]. In either pants cusp, the subtail bounded by a horocycle of length b = 1 / 4 b=1/4 b = 1/4 has area b b b and lies inside P s P_s P s . Indeed, a closed geodesic meeting this subtail must leave the length-one cusp neighborhood and so has length at least 2 log 4 > s 2\log4>s 2 log 4 > s . If v v v has zero angular mean on the horocycles of this subtail H H H , then ∫ H v = 0 \int_H v=0 ∫ H v = 0 , and
b ∣ v ‾ P s ∣ 2 ≤ ∫ H ∣ v − v ‾ P s ∣ 2 ≤ 4 E P s ( v ) . b\lvert\overline{v}_{P_s}\rvert^2\leq\int_H\lvert v-\overline{v}_{P_s}\rvert^2\leq4\mathcal{E}_{P_s}(v). b ∣ v P s ∣ 2 ≤ ∫ H ∣ v − v P s ∣ 2 ≤ 4 E P s ( v ) .
Since area ( P s ) = 2 π \operatorname{area}(P_s)=2\pi area ( P s ) = 2 π , this and (195) imply
∥ v ∥ L 2 ( P s ) 2 ≤ C ∗ E P s ( v ) . (196) \lVert v\rVert^2_{L^2(P_s)}\leq C_*\mathcal{E}_{P_s}(v). \tag*{(196)} ∥ v ∥ L 2 ( P s ) 2 ≤ C ∗ E P s ( v ) . (196)
Every finite linear combination of cusp-form eigenfunctions has the required zero angular means. Their constant Fourier coefficients, zero near infinity, vanish throughout the embedded cusp cylinder by the Fourier equation. Summing (196) over the pants union U U U gives
∥ u ∥ L 2 ( U ) 2 ≤ C ∗ E X ( u ) (197) \lVert u\rVert^2_{L^2(U)}\leq C_*\mathcal{E}_X(u) \tag*{(197)} ∥ u ∥ L 2 ( U ) 2 ≤ C ∗ E X ( u ) (197)
for every such combination, with no factor depending on N N N .
The N N N distance cutoffs span a space supported in U U U with Rayleigh quotient at most η = 6 L ε = ε / [ 4 ( 1 + C ∗ ε ) ] \eta=6L_\varepsilon=\varepsilon/[4(1+C_*\varepsilon)] η = 6 L ε = ε / [ 4 ( 1 + C ∗ ε )] . Its intersection W W W with the mean-zero space has dimension at least N − 1 N-1 N − 1 and is still supported in U U U . If the residual spectral subspace in ( 0 , ε ) (0,\varepsilon) ( 0 , ε ) had dimension smaller than dim W \dim W dim W , there would be a unit f ∈ W f\in W f ∈ W orthogonal to it. The exact spectral decomposition gives f = u + v f=u+v f = u + v , where u u u belongs to the cuspidal spectral subspace in ( 0 , ε ) (0,\varepsilon) ( 0 , ε ) and v v v has spectrum in [ ε , ∞ ) [\varepsilon,\infty) [ ε , ∞ ) . The constant has already been removed; the continuous spectrum starts at 1 / 4 1/4 1/4 . Spectral projections contract the quadratic form, so
E X ( u ) ≤ η , ∥ v ∥ 2 ≤ η / ε . \mathcal{E}_X(u)\leq\eta,\qquad\lVert v\rVert^2\leq\eta/\varepsilon. E X ( u ) ≤ η , ∥ v ∥ 2 ≤ η / ε .
As f f f is supported in U U U , (197) yields
1 ≤ ∥ u ∥ L 2 ( U ) + ∥ v ∥ L 2 ( U ) ≤ C ∗ η + η / ε ≤ 2 ( C ∗ + 1 / ε ) η = 1 / 2 , 1\leq\lVert u\rVert_{L^2(U)}+\lVert v\rVert_{L^2(U)}\leq\sqrt{C_*\eta}+\sqrt{\eta/\varepsilon}\leq\sqrt{2(C_*+1/\varepsilon)\eta}=1/\sqrt{2}, 1 ≤ ∥ u ∥ L 2 ( U ) + ∥ v ∥ L 2 ( U ) ≤ C ∗ η + η / ε ≤ 2 ( C ∗ + 1/ ε ) η = 1/ 2 ,
a contradiction. Thus projection onto the actual residual eigenspace is injective on W W W , proving the dimension bound. The argument allows an eigenvalue to have both cuspidal and residual multiplicity and counts only the latter. □
Theorem 14.7 (Many small residual eigenvalues). For g → ∞ g\to\infty g → ∞ , n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ and each fixed 0 < ε < 1 / 4 0<\varepsilon<1/4 0 < ε < 1/4 ,
P g , n W P { R ε ( X ) ≥ L ε 2 q g / 4 } ⟶ 1. (198) \mathbb{P}^{\mathrm{WP}}_{g,n}\{R_\varepsilon(X)\geq L_\varepsilon^2q_g/4\}\longrightarrow1. \tag*{(198)} P g , n WP { R ε ( X ) ≥ L ε 2 q g /4 } ⟶ 1. (198) For every a > 0 a>0 a > 0 and 0 < ε < 1 / 4 0<\varepsilon<1/4 0 < ε < 1/4 there is d ε , a > 0 d_{\varepsilon,a}>0 d ε , a > 0 such that, along every sequence n → ∞ n\to\infty n → ∞ and 0 ≤ g ≤ a n 0\leq g\leq an 0 ≤ g ≤ an ,
P g , n W P { R ε ( X ) ≥ d ε , a ( 2 g + n − 2 ) } ⟶ 1. (199) \mathbb{P}^{\mathrm{WP}}_{g,n}\{R_\varepsilon(X)\geq d_{\varepsilon,a}(2g+n-2)\}\longrightarrow1. \tag*{(199)} P g , n WP { R ε ( X ) ≥ d ε , a ( 2 g + n − 2 )} ⟶ 1. (199) Proof. Since L ε < 1 / 96 L_\varepsilon<1/96 L ε < 1/96 , the counted pants are interior-disjoint. Proposition 14.3 and B Y ( L ) ≥ L 2 \mathcal{B}_Y(L)\geq L^2 B Y ( L ) ≥ L 2 give, with probability tending to one, N L ε − 1 ≥ q g L ε 2 / 4 N_{L_\varepsilon}-1\geq q_gL_\varepsilon^2/4 N L ε − 1 ≥ q g L ε 2 /4 . Lemma 14.6 proves (198) . The positive lower bound for q g / ( 2 g + n − 2 ) q_g/(2g+n-2) q g / ( 2 g + n − 2 ) when g ≤ a n g\leq an g ≤ an , established in the proof of Corollary 14.4 , gives (199) when g → ∞ g\to\infty g → ∞ . For each fixed genus, use the short-pants count of Hide–Thomas [98, Section 5, equation (5.1)]: N L ε ≥ c ( L ε ) n N_{L_\varepsilon}\geq c(L_\varepsilon)n N L ε ≥ c ( L ε ) n with probability tending to one, where the positive constant is independent of the fixed genus. Applying Lemma 14.6 gives the residual count in this case as well. Reduce d ε , a d_{\varepsilon,a} d ε , a to a common constant and use the same fixed-or-diverging genus subsequence argument as in Corollary 14.4 . □
This establishes the linear residual count anticipated in [98, Section 1.1], including every fixed genus, and extends it to growing genus through the uniform volume theorem. The count is far below the Otal–Rosas bound 2 g + n − 2 2g+n-2 2 g + n − 2 on the number of eigenvalues in [ 0 , 1 / 4 ] [0,1/4] [ 0 , 1/4 ] , counting the eigenvalue 0 0 0 [157]: q g / ( 2 g + n − 2 ) q_g/(2g+n-2) q g / ( 2 g + n − 2 ) is bounded by (182) , and L ε 2 / 4 < 10 − 4 L_{\varepsilon}^{2}/4<10^{-4} L ε 2 /4 < 1 0 − 4 . It does not assert that all small eigenvalues are residual, and it does not bound the total number of small cuspidal eigenvalues. The mechanism behind Lemma 14.6, that cusp forms have vanishing cusp averages while localized pants states do not, returns in sharper form at the critical scale (Lemma 15.8) and in the dense regime (Corollary 18.4), where it identifies the whole low eigenspace as residual.
The tight length spectrum
We now consider curves that remain noncontractible after all cusps are filled, and prove their Poisson limit at every cusp density (Theorem 14.8) and for every subcritical cusp weight (Corollary 14.9). Following Budd and Lions [32], a closed geodesic is tight if its class on the filled surface is noncontractible and it is the unique shortest geodesic on the original surface among those in that free homotopy class after filling the cusps. The minimality condition matters, since several ordinary geodesics can belong to the same class after the cusps are filled. The corresponding tight Weil–Peterson volumes were introduced by Budd and Zonneveld [34]; see also [30] for the related irreducible metric maps. The tight integration formula and the geometric argument used in the proof are due to [32]; the new input is a volume estimate at a prescribed cusp count, obtained from the shifted amplitudes and the coefficient extraction of Part 1, and the proofs are in Appendix H.
Let Y λ Y_\lambda Y λ satisfy λ ( Y λ ) = λ \lambda(Y_\lambda)=\lambda λ ( Y λ ) = λ , as in Proposition 9.4, and set
α ( λ ) = 2 r ( Y λ ) π , ν ( d t ) = cosh t − 1 t d t . (200) \alpha(\lambda)=\frac{\sqrt{2r(Y_\lambda)}}{\pi},\qquad\nu(\mathrm{d}t)=\frac{\cosh t-1}{t}\,\mathrm{d}t. \tag*{(200)} α ( λ ) = π 2 r ( Y λ ) , ν ( d t ) = t cosh t − 1 d t . (200)
Here α ( 0 ) = 1 \alpha(0)=1 α ( 0 ) = 1 , and the density of ν \nu ν is defined to be zero at the origin.
Theorem 14.8 (Tight lengths at arbitrary cusp density). Let n = n ( g ) ≥ 0 n=n(g)\ge0 n = n ( g ) ≥ 0 be any integer sequence, and let X X X have law P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP . As g → ∞ g\to\infty g → ∞ , the point process
∑ γ p r i m i t i v e , u n o r i e n t e d γ t i g h t o n X δ α ( n / g ) ℓ X ( γ ) ⟹ PPP ( ν ) (201) \sum_{\substack{\gamma\ \mathrm{primitive,\ unoriented}\\ \gamma\ \mathrm{tight\ on}\ X}}\delta_{\alpha(n/g)\ell_X(\gamma)}\Longrightarrow\operatorname{PPP}(\nu) \tag*{(201)} γ primitive , unoriented γ tight on X ∑ δ α ( n / g ) ℓ X ( γ ) ⟹ PPP ( ν ) (201) converges in distribution for the vague topology on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . For every fixed L < ∞ L<\infty L < ∞ , with probability tending to one all the curves counted in [ 0 , L ] [0,L] [ 0 , L ] are simple, pairwise disjoint, and have connected complement after filling the cusps.
In particular, n / g n/g n / g need not converge. At large density, the endpoint expansion of r r r gives
α ( λ ) ∼ 3 j 0 , 1 5 π λ − 1 / 4 . \alpha(\lambda)\sim\frac{\sqrt{3j_{0,1}\sqrt{5}}}{\pi}\lambda^{-1/4}. α ( λ ) ∼ π 3 j 0 , 1 5 λ − 1/4 .
This agrees with the scale in [32, Corollary 1.3]. Budd and Lions asked whether their limit persists for other cusp weights and under the fixed-count measures P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP [32, Open Question 1.6]. Theorem 14.8 treats the fixed-count part of this question, and Corollary 14.9 below treats the remaining cusp-weight regimes. The planar curves in Theorem 14.2 are contractible after the cusps are filled, and therefore do not occur in (201) .
The proof follows the moment method of Budd and Lions: their tight integration formula [32, Proposition 3.2] expresses factorial moments of tight curve counts through tight volumes of the cut surfaces. In Appendix H we write the generating functions of tight volumes through the shifted amplitudes of Section 4 (Lemma H.1), extract the coefficient at the prescribed cusp count as in Section 9 (Lemma H.2), show that configurations with disconnected complement are negligible (Lemma H.3), and identify the Poisson limit (Subsection H.4).
Random cusp counts. For 0 ≤ μ < μ c r = x c / ( 2 π 2 ) 0 \leq\mu< \mu_{\mathrm{cr}} = x_{\mathrm{c}}/(2\pi^2) 0 ≤ μ < μ cr = x c / ( 2 π 2 ) , the Boltzmann measure P g , μ W P \mathbb{P}^{\mathrm{WP}}_{g,\mu} P g , μ WP first chooses the cusp count N N N with probabilities proportional to μ n V g , n / n ! \mu^n V_{g,n}/n! μ n V g , n / n ! , then samples P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP conditionally on N = n N=n N = n . The probability generating function of the cusp count is Φ g ( 2 π 2 μ z ) / Φ g ( 2 π 2 μ ) \Phi_g(2\pi^2\mu z)/\Phi_g(2\pi^2\mu) Φ g ( 2 π 2 μ z ) / Φ g ( 2 π 2 μ ) , as in [32].
Corollary 14.9 (Arbitrary subcritical cusp weights). For any sequence 0 ≤ μ g < μ c r 0 \leq\mu_g < \mu_{\mathrm{cr}} 0 ≤ μ g < μ cr , put
a g = 2 r ( Y ( 2 π 2 μ g ) ) π . a_g = \frac{\sqrt{2r(Y(2\pi^2\mu_g))}}{\pi}. a g = π 2 r ( Y ( 2 π 2 μ g )) . As g → ∞ g \to\infty g → ∞ , under P g , μ g W P \mathbb{P}^{\mathrm{WP}}_{g,\mu_g} P g , μ g WP , the counting measure of primitive unoriented tight lengths rescaled by a g a_g a g converges in distribution to PPP ( ν ) \operatorname{PPP}(\nu) PPP ( ν ) for the vague topology on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . In particular, if μ g → μ < μ c r \mu_g \to\mu< \mu_{\mathrm{cr}} μ g → μ < μ cr , the unscaled lengths have limiting intensity ( cosh ( a μ t ) − 1 ) d t / t (\cosh(a_\mu t)-1)\,\mathrm{d}t/t ( cosh ( a μ t ) − 1 ) d t / t , where a μ = 2 r ( Y ( 2 π 2 μ ) ) / π a_\mu= \sqrt{2r(Y(2\pi^2\mu))}/\pi a μ = 2 r ( Y ( 2 π 2 μ )) / π . If μ g → μ c r \mu_g \to\mu_{\mathrm{cr}} μ g → μ cr , no condition on the speed of convergence is required.
The proof is given in Subsection H.5. The tight spectrum is not used in the spectral sections that follow; its role here is to complete the description of short curves at every cusp density, the planar separating curves having been treated in Subsection 14.2. The next subsection returns to those planar curves at the critical scale, where both families live on the same length scale.
The full length spectrum at the critical scale
This subsection determines the full primitive length spectrum at the critical scale n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g} \to c \in(0,\infty) n / g → c ∈ ( 0 , ∞ ) (Theorem 14.10 ). The tight spectrum omits curves that become contractible when the cusps are filled; at this scale, these curves contribute on the same length scale as the nonseparating curves, and their limiting contribution is a cluster process: each two-cusp pair of pants contributes its whole primitive length spectrum, not only its boundary. The volume inputs are the bounds of Mirzakhani–Zograf and Shen–Wu [147, 172] together with the zero-density case of Theorem 1.3 and Proposition 9.4 ; the proof is in Appendix § 10.6 . The waist process Q c Q_c Q c obtained here is the geometric input of the critical spectral law in § 4.7 .
For s > 0 s > 0 s > 0 , let P s P_s P s be the hyperbolic two-cusp pair of pants with one geodesic boundary of length s s s . Write K s K_s K s for its primitive unoriented length counting measure, including the boundary once and excluding parabolic classes. Equal lengths are counted with their geometric multiplicities. Let
μ c ( d s ) = c 2 8 π 2 sinh ( s / 2 ) d s , ν ( d t ) = cosh t − 1 t d t . (202) \mu_c(\mathrm{d}s) = \frac{c^2}{8\pi^2}\sinh(s/2)\,\mathrm{d}s,\qquad\nu(\mathrm{d}t) = \frac{\cosh t - 1}{t}\,\mathrm{d}t. \tag*{(202)} μ c ( d s ) = 8 π 2 c 2 sinh ( s /2 ) d s , ν ( d t ) = t cosh t − 1 d t . (202)
Theorem 14.10 (Critical primitive length spectrum). Suppose g → ∞ g \to\infty g → ∞ and n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g} \to c \in(0,\infty) n / g → c ∈ ( 0 , ∞ ) , and let X X X have law P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP . Then
∑ γ p r i m i t i v e u n o r i e n t e d o n X δ ℓ X ( γ ) ⟹ Λ c : = N + ∑ s ∈ Q c K s (203) \sum_{\substack{\gamma\ \mathrm{primitive}\\ \mathrm{unoriented\ on}\ X}} \delta_{\ell_X(\gamma)}
\ \Longrightarrow\ \Lambda_c := \mathcal{N} + \sum_{s\in Q_c} K_s \tag*{(203)} γ primitive unoriented on X ∑ δ ℓ X ( γ ) ⟹ Λ c := N + s ∈ Q c ∑ K s (203) in distribution in the space of locally finite counting measures on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) with the vague topology. Here N \mathcal{N} N and Q c Q_c Q c are independent Poisson point processes of intensities ν \nu ν and μ c \mu_c μ c . In particular, for every nonnegative continuous compactly supported function f f f ,
E e − ⟨ f , Λ c ⟩ = exp { − ∫ ( 1 − e − f ( t ) ) ν ( d t ) − ∫ ( 1 − e − ⟨ f , K s ⟩ ) μ c ( d s ) } . (204) \mathbb{E}e^{-\langle f,\Lambda_c\rangle}
=
\exp\left\{
-\int(1-e^{-f(t)})\,\nu(\mathrm{d}t)
-\int(1-e^{-\langle f,K_s\rangle})\,\mu_c(\mathrm{d}s)
\right\}. \tag*{(204)} E e − ⟨ f , Λ c ⟩ = exp { − ∫ ( 1 − e − f ( t ) ) ν ( d t ) − ∫ ( 1 − e − ⟨ f , K s ⟩ ) μ c ( d s ) } . (204)
Shen and Wu obtained the factorial-moment asymptotics for short two-cusp pairs of pants and the limiting probability that such a curve occurs at this cusp scale [172, Lemmas 28 and 33]; on their stated short window the corresponding Poisson mean is μ c ( [ 0 , L ] ) = c 2 ( cosh ( L / 2 ) − 1 ) / ( 4 π 2 ) \mu_c([0,L]) = c^2(\cosh(L/2)-1)/(4\pi^2) μ c ([ 0 , L ]) = c 2 ( cosh ( L /2 ) − 1 ) / ( 4 π 2 ) . Theorem 14.10 concerns the full primitive process on every fixed length window, including the correlated lengths inside each pair of pants. The background intensity and the geometric moment method come from Mirzakhani and Petri [146]. The bounds of [147, 172] alone give an upper estimate and a positive lower bound for the genus ratio V g − 1 , n + 2 / V g , n V_{g-1,n+2}/V_{g,n} V g − 1 , n + 2 / V g , n at this scale, but not its limit, which is why the zero-density case of Theorem 1.3 is needed; the first-order expansion of Gatignon and Thomas [71, Theorem 1.5] is another source of its two-sided limit. The full uniform expansion is not used.
Nonsimple curves distinguish this regime from closed surfaces. He, Shen, Wu and Xue proved that the nonsimple systole of a Weil–Petersson random closed surface is asymptotic to log g \log g log g , and more precisely concentrates about log g − log log g \log g-\log\log g log g − log log g [91, Theorem 1.1]; in that model nonsimple curves therefore leave every fixed length window. At the critical scale, by contrast, two-cusp pairs of pants retain nonsimple curves of bounded length.
Outline of the proof. Short multicurves are organized by their cut graphs. Using only the volume bounds of [147, 172], Lemma I.1 shows that a cut graph contributes in the limit only if it consists of one component carrying the genus, possibly with nonseparating loops, and two-cusp pairs of pants attached to it as leaves. A bounded-multiplicity argument (Lemmas I.2 and I.3) reduces intersecting simple curves and nonsimple curves to such cut graphs; hence, with probability tending to one, every short nonsimple curve lies in a two-cusp pair of pants. The mixed factorial moments of nonseparating lengths and pants waists converge to those of independent Poisson processes (Lemma I.4, where the zero-density volume ratios enter), and attaching to each waist the full spectrum of its pair of pants gives Λ c \Lambda_c Λ c . The details are in Appendix I.
The limit is not Poisson. If m L ( s ) = K s ( [ 0 , L ] ) m_L(s)=K_s([0,L]) m L ( s ) = K s ([ 0 , L ]) , then Lemma I.2 and the support restriction s ≤ 2 L s\le2L s ≤ 2 L give finite moments and
Var Λ c ( [ 0 , L ] ) − E Λ c ( [ 0 , L ] ) = ∫ m L ( s ) ( m L ( s ) − 1 ) μ c ( d s ) . (205) \operatorname{Var}\Lambda_c([0,L])-\mathbb{E}\Lambda_c([0,L])=\int m_L(s)(m_L(s)-1)\,\mu_c(ds). \tag*{(205)} Var Λ c ([ 0 , L ]) − E Λ c ([ 0 , L ]) = ∫ m L ( s ) ( m L ( s ) − 1 ) μ c ( d s ) . (205)
Besides its waist, P s P_s P s contains the primitive figure-eight around its two cusps, of length 2 arccosh ( 2 + cosh ( s / 2 ) ) 2\operatorname{arccosh}(2+\cosh(s/2)) 2 arccosh ( 2 + cosh ( s /2 )) by [164, Theorem 1.3]. For L > 2 arccosh 3 L>2\operatorname{arccosh}3 L > 2 arccosh 3 , both lengths lie below L L L on an interval of positive μ c \mu_c μ c -measure. The right side of (205) is therefore strictly positive. The theorem identifies these correlations in the limit; no convergence of the full-spectrum second moments is asserted. Of the cluster structure, only the waist process Q c Q_c Q c , its intensity μ c \mu_c μ c and the factorial-moment bounds behind it are used in Section 15; the cut-graph Lemmas I.1 and I.3 of Appendix I are reused there to isolate the pants.
The critical low spectrum
At the critical scale n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g}\to c\in(0,\infty) n / g → c ∈ ( 0 , ∞ ) , the full positive spectral point process below 1 / 4 1/4 1/4 converges to a Poisson process of model energies (Theorem 15.1). This gives the gap phase diagram (Corollary 15.9). The geometric input is the waist process Q c Q_c Q c of Theorem 14.10; the spectral input is the trace expansion of Huang [101], restated and extended to the critical envelope in Appendix J; the model input is the spectral theory of the funnel surface F s F_s F s defined below, through Borthwick–Judge–Perry [25], Schulze [170] and Ballmann–Mattiessen–Mondal [16]. The argument is carried out in this section, with the cut-graph lemmas of Appendix I as the remaining input.
At the critical scale, a bounded number of two-cusp pairs of pants occur in each bounded waist interval, and their exterior neighborhoods become arbitrarily deep funnels as the genus grows. The spectral model attached to a waist is therefore the complete surface obtained by gluing a funnel to the pair of pants, not the pants with a Dirichlet condition at its boundary. The distinction matters for bounded waist lengths, since a cusped pair of pants need not have a Dirichlet eigenvalue below 1 / 4 1/4 1/4 [100, Remark 2.2].
Let F s F_s F s be the complete hyperbolic surface with two cusps and one funnel whose convex core is P s P_s P s , with boundary length s > 0 s>0 s > 0 ; we call F s F_s F s the funnel model of waist s s s . Write β ( s ) \beta(s) β ( s ) for its lowest eigenvalue; see [24] for the spectral theory of surfaces with funnels. The relation between the bottom of the spectrum and the critical exponent goes back to Patterson [158]. The cusp theorem of Beardon [20] and the spectral relation of Sullivan [176], Theorems 2.17 and 2.21 give
β ( s ) = δ ( s ) ( 1 − δ ( s ) ) ∈ ( 0 , 1 / 4 ) , 1 / 2 < δ ( s ) < 1 , (206) \beta(s)=\delta(s)(1-\delta(s))\in(0,1/4),\qquad1/2<\delta(s)<1, \tag*{(206)} β ( s ) = δ ( s ) ( 1 − δ ( s )) ∈ ( 0 , 1/4 ) , 1/2 < δ ( s ) < 1 , (206)
where δ ( s ) \delta(s) δ ( s ) is the critical exponent of its uniformizing group. By [16, Theorem 1.5 and Corollary 1.6], there is at most one eigenvalue below 1 / 4 1/4 1/4 , counted with multiplicity; since δ ( s ) > 1 / 2 \delta(s)>1/2 δ ( s ) > 1/2 , there is exactly one. The function β \beta β is continuous, including the limit β ( s ) → 0 \beta(s)\to0 β ( s ) → 0 as s ↓ 0 s\downarrow0 s ↓ 0 , by spectral continuity through pinching [170, Theorems 2.14 and 3.21]; the pinched finite-area component is the thrice-punctured sphere, whose constant eigenfunction has eigenvalue zero. Lemma 15.2 below gives the other endpoint, β ( s ) → 1 / 4 \beta(s)\to1/4 β ( s ) → 1/4 as s → ∞ s\to\infty s → ∞ .
Retain the pants process Q c Q_c Q c and its intensity μ c \mu_c μ c from (202) , and define the energy intensity
ν c ( B ) = c 2 8 π 2 ∫ 0 ∞ 1 { β ( s ) ∈ B } sinh ( s / 2 ) d s , B ⊂ [ 0 , 1 / 4 ) . (207) \nu_c(B)=\frac{c^2}{8\pi^2}\int_0^\infty\mathbf{1}_{\{\beta(s)\in B\}}\sinh(s/2)\,\mathrm{d}s,\qquad B\subset[0,1/4). \tag*{(207)} ν c ( B ) = 8 π 2 c 2 ∫ 0 ∞ 1 { β ( s ) ∈ B } sinh ( s /2 ) d s , B ⊂ [ 0 , 1/4 ) . (207)
It is finite on compact subsets of [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) and has no atom at zero. We use G ( X ) = inf spec ( Δ X ∣ 1 ⊥ ) G(X)=\inf\operatorname{spec}(\Delta_X|_{1^\perp}) G ( X ) = inf spec ( Δ X ∣ 1 ⊥ ) for the gap, with G ( X ) = 1 / 4 G(X)=1/4 G ( X ) = 1/4 when no positive eigenvalue lies below 1 / 4 1/4 1/4 .
Theorem 15.1 (Critical exceptional spectrum). Suppose g → ∞ g\to\infty g → ∞ and n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g}\to c\in(0,\infty) n / g → c ∈ ( 0 , ∞ ) . For a Weil–Petersson random surface X ∈ M g , n X\in\mathcal{M}_{g,n} X ∈ M g , n ,
Ξ X : = ∑ 0 < λ j ( X ) < 1 / 4 δ λ j ( X ) ⟹ Ξ c : = ∑ s ∈ Q c δ β ( s ) (208) \Xi_X:=\sum_{0<\lambda_j(X)<1/4}\delta_{\lambda_j(X)}\Longrightarrow\Xi_c:=\sum_{s\in Q_c}\delta_{\beta(s)} \tag*{(208)} Ξ X := 0 < λ j ( X ) < 1/4 ∑ δ λ j ( X ) ⟹ Ξ c := s ∈ Q c ∑ δ β ( s ) (208) in the vague topology on [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) . All discrete eigenvalues, including residual eigenvalues, are counted with multiplicity. The limit is a Poisson random measure of intensity ν c \nu_c ν c , and E Ξ X → ν c \mathbb{E}\Xi_X\to\nu_c E Ξ X → ν c vaguely. In particular G ( X ) G(X) G ( X ) converges in distribution to G c ∈ ( 0 , 1 / 4 ) G_c\in(0,1/4) G c ∈ ( 0 , 1/4 ) with
P { G c > t } = exp { − ν c ( [ 0 , t ] ) } , 0 ≤ t < 1 / 4. (209) \mathbb{P}\{G_c>t\}=\exp\{-\nu_c([0,t])\},\qquad0\le t<1/4. \tag*{(209)} P { G c > t } = exp { − ν c ([ 0 , t ])} , 0 ≤ t < 1/4. (209) Writing Ξ X r e s \Xi_X^{\mathrm{res}} Ξ X res and Ξ X c u s p \Xi_X^{\mathrm{cusp}} Ξ X cusp for the residual and cuspidal parts, respectively, we have jointly
( Ξ X r e s , Ξ X c u s p ) ⟹ ( Ξ c , 0 ) , E Ξ X r e s ⟶ ν c , E Ξ X c u s p ⟶ 0 (210) (\Xi_X^{\mathrm{res}},\Xi_X^{\mathrm{cusp}})\Longrightarrow(\Xi_c,0),\qquad\mathbb{E}\Xi_X^{\mathrm{res}}\longrightarrow\nu_c,\qquad\mathbb{E}\Xi_X^{\mathrm{cusp}}\longrightarrow0 \tag*{(210)} ( Ξ X res , Ξ X cusp ) ⟹ ( Ξ c , 0 ) , E Ξ X res ⟶ ν c , E Ξ X cusp ⟶ 0 (210) in the same vague topologies. More generally, let C ε ( X ) C_\varepsilon(X) C ε ( X ) be the cuspidal multiplicity in ( 0 , 1 / 4 − ε ] (0,1/4-\varepsilon] ( 0 , 1/4 − ε ] . For every fixed 0 < M < ∞ 0<M<\infty 0 < M < ∞ and 0 < ε < 1 / 4 0<\varepsilon<1/4 0 < ε < 1/4 ,
sup 1 ≤ n ≤ M g E g , n W P C ε ( X ) ⟶ 0. (211) \sup_{1\le n\le M\sqrt{g}}\mathbb{E}_{g,n}^{\mathrm{WP}}C_\varepsilon(X)\longrightarrow0. \tag*{(211)} 1 ≤ n ≤ M g sup E g , n WP C ε ( X ) ⟶ 0. (211)
The primitive length process in Theorem 14.10 retains all geodesics inside each pants and is a cluster process. In contrast, each funnel model contributes exactly one subthreshold eigenvalue to (208) . Neither the pushforward description nor the gap law requires monotonicity of β \beta β .
The proof combines the waist process Q c Q_c Q c with Huang’s orbital expansion and positive exceptional measure [101, Sections 3.7 and 6.8], and adds three steps, each treated in its own subsection after a preliminary lemma on large waists. First, we extend the trace estimate to each bounded critical envelope and identify its quadratic correction with the model energies; a finite contour deformation handles the continuous-spectrum contribution. The resulting Proposition 15.3, together with Proposition 15.5, is a two-sided limit of the mean measure; in the count matching it is used as the upper half, and the lower half is supplied by the localized lower count of Lemma 15.7 . Second, disjoint local ground states give lower spectral counts. Since their mean exhausts the limiting expected count, convergence follows without expanding joint eigenvalue traces. Third, compact model states with small operator residuals have independent nonzero cusp averages. Spectral projection and the matching dimension make the cusp-average map injective on the actual low eigenspace, so this space is residual (Lemma 15.8 ). These arguments concern fixed compact windows below 1 / 4 1/4 1/4 , not windows approaching the threshold; the latter are the subject of § 4.12 .
The model at large waist lengths
The following lemma supplies the endpoint β ( s ) → 1 / 4 \beta(s)\to1/4 β ( s ) → 1/4 and the integrability of the large-waist tail of the orbital sum. Both are used in the identification of the mean measure (Proposition 15.5 ); the endpoint also gives the local finiteness of the fixed-window measure (252) in § 4.19 , and the generators introduced in the proof reappear in § 4.12 .
Lemma 15.2. For every z > 1 / 2 z > 1/2 z > 1/2 , sufficiently large S S S satisfies
∫ S ∞ sinh ( s / 2 ) ∑ γ p r i m i t i v e o r i e n t e d k ≥ 1 ℓ γ e − z k ℓ γ d s < ∞ . (212) \int_S^\infty\sinh(s/2)\sum_{\substack{\gamma\ \mathrm{primitive\ oriented}\\ k\geq1}}\ell_\gamma e^{-zk\ell_\gamma}\,ds < \infty. \tag*{(212)} ∫ S ∞ sinh ( s /2 ) γ primitive oriented k ≥ 1 ∑ ℓ γ e − z k ℓ γ d s < ∞. (212) Moreover δ ( s ) → 1 / 2 \delta(s)\to1/2 δ ( s ) → 1/2 as s → ∞ s\to\infty s → ∞ . Thus for each t < 1 / 4 t<1/4 t < 1/4 there is S t < ∞ S_t<\infty S t < ∞ such that β ( s ) > t \beta(s)>t β ( s ) > t for s > S t s>S_t s > S t .
Proof. Use the two-parabolic marking
A = ( 1 a 0 1 ) , B = ( 1 0 − a 1 ) , a = 2 cosh ( s / 4 ) = κ + κ − 1 , κ = e s / 4 . A=\begin{pmatrix}1&a\\0&1\end{pmatrix},\qquad B=\begin{pmatrix}1&0\\-a&1\end{pmatrix},\qquad a=2\cosh(s/4)=\kappa+\kappa^{-1},\qquad\kappa=e^{s/4}. A = ( 1 0 a 1 ) , B = ( 1 − a 0 1 ) , a = 2 cosh ( s /4 ) = κ + κ − 1 , κ = e s /4 . These generate the free group of F s F_s F s , and A B AB A B represents the waist. Every hyperbolic conjugacy class has an alternating cyclic representative W = A m 1 B n 1 ⋯ A m q B n q W=A^{m_1}B^{n_1}\cdots A^{m_q}B^{n_q} W = A m 1 B n 1 ⋯ A m q B n q with nonzero integer exponents. For C ( u ) = ( u − 1 1 0 ) C(u)=\begin{pmatrix}u&-1\\1&0\end{pmatrix} C ( u ) = ( u 1 − 1 0 ) , A m B n = − C ( m a ) C ( n a ) A^mB^n=-C(ma)C(na) A m B n = − C ( ma ) C ( na ) . The projective map of C ( m a ) C(ma) C ( ma ) sends r ∈ [ − κ − 1 , κ − 1 ] r\in[-\kappa^{-1},\kappa^{-1}] r ∈ [ − κ − 1 , κ − 1 ] to ( m a − r ) − 1 (ma-r)^{-1} ( ma − r ) − 1 , where r = x 2 / x 1 r=x_2/x_1 r = x 2 / x 1 is the ratio of the coordinates on which C ( m a ) C(ma) C ( ma ) acts; it preserves this interval and contracts it by at most κ − 2 \kappa^{-2} κ − 2 . Its first-coordinate expansion is at least ∣ m ∣ κ |m|\kappa ∣ m ∣ κ . At the fixed direction of the product, the expanding eigenvalue therefore gives
ℓ ( W ) ≥ q s + 2 ∑ j = 1 q log ∣ m j n j ∣ . (213) \ell(W)\geq qs+2\sum_{j=1}^{q}\log|m_jn_j|. \tag*{(213)} ℓ ( W ) ≥ q s + 2 j = 1 ∑ q log ∣ m j n j ∣. (213) Consequently, for z 0 > 1 / 2 z_0>1/2 z 0 > 1/2 , the sum of e − z 0 ℓ ( W ) e^{-z_0\ell(W)} e − z 0 ℓ ( W ) over all such words is bounded by
∑ q ≥ 1 { 4 ζ ( 2 z 0 ) 2 e − z 0 s } q . \sum_{q\geq1}\left\{4\zeta(2z_0)^2e^{-z_0s}\right\}^{q}. q ≥ 1 ∑ { 4 ζ ( 2 z 0 ) 2 e − z 0 s } q . This overcounts every oriented conjugacy class, including powers. Choose 1 / 2 < z 0 < z 1/2<z_0<z 1/2 < z 0 < z and then S S S so that the ratio is at most 1 / 2 1/2 1/2 . The inequality u e − ( z − z 0 ) u ≤ [ e ( z − z 0 ) ] − 1 ue^{-(z-z_0)u}\leq[e(z-z_0)]^{-1} u e − ( z − z 0 ) u ≤ [ e ( z − z 0 ) ] − 1 bounds the sum in (212) by C z , z 0 e − z 0 s C_{z,z_0}e^{-z_0s} C z , z 0 e − z 0 s , proving integrability. The same word estimate makes the Selberg Euler product nonzero in ℜ u ≥ z 0 \Re u\geq z_0 ℜ u ≥ z 0 for large s s s . Its eigenvalue zero at δ ( s ) \delta(s) δ ( s ) [25, Theorem 1.1] implies δ ( s ) ≤ z 0 \delta(s)\leq z_0 δ ( s ) ≤ z 0 there. Let z 0 ↓ 1 / 2 z_0\downarrow1/2 z 0 ↓ 1/2 and use (206) . □
The mean measure and its model identification
This subsection proves the two-sided limit of the mean spectral measure at the critical scale. It has two parts: a trace estimate that expresses the mean spectrum through Huang’s exceptional measure (Proposition 15.3), and an identification of that measure with the model energies (Proposition 15.5). The first part rests on the extension of Huang’s orbital estimate to the critical envelope n ≤ M g n\leq M\sqrt{g} n ≤ M g , proved in Appendix § 10.10 ; the second rests on the regularized Selberg zeta function of the funnel model through the pinching limit (Lemma 15.4 ), which uses Schulze [170] and Borthwick–Judge–Perry [25]. We use the admissible multiplier of Huang [101, Section 1.2]: h = b 0 ∗ b 0 h=b_0*b_0 h = b 0 ∗ b 0 is even, nonnegative, smooth, supported in [ − L , L ] [-L,L] [ − L , L ] for a fixed L > 0 L>0 L > 0 , and has integral one, f ( r ) = ∫ h ( u ) e − i r u d u f(r)=\int h(u)e^{-iru}\,du f ( r ) = ∫ h ( u ) e − i r u d u , and a = f ( i / 2 ) > 1 a=f(i/2)>1 a = f ( i /2 ) > 1 . Thus 0 ≤ f ( r ) ≤ 1 0\leq f(r)\leq1 0 ≤ f ( r ) ≤ 1 for real r r r , while τ ↦ f ( i τ ) \tau\mapsto f(i\tau) τ ↦ f ( i τ ) is strictly increasing for τ > 0 \tau>0 τ > 0 . The support length L L L enters the main text only through the cusp truncation height Y Q = 3 e L Q Y_Q=3e^{LQ} Y Q = 3 e L Q below. Subsection J.1 of Appendix § 10.10 restates in our notation the results of [101] used in this subsection; in particular Definition J.1 introduces the orbital functional O X ( p ) \mathcal{O}_X(p) O X ( p ) and the canonical coefficients ω 1 , ω j , n \omega_1,\omega_{j,n} ω 1 , ω j , n and W i W_i W i of the orbital expansion.
Write T 2 \mathcal{T}_2 T 2 for Huang’s positive exceptional measure and T 2 \mathcal{T}_2 T 2 for his weighted distribution. Their construction (Lemma J.11) is as follows. For each n ≥ 1 n\geq1 n ≥ 1 the functional K n ( q ) = ω 2 , n ( x q ) \mathcal{K}_n(q)=\omega_{2,n}(xq) K n ( q ) = ω 2 , n ( x q ) on polynomials q q q extends uniquely to a distribution supported in [ 0 , a ] [0,a] [ 0 , a ] of order at most m = 2 D 0 + 2 m=2D_0+2 m = 2 D 0 + 2 , where D 0 D_0 D 0 is the exponent in the coefficient bound of Lemma J.8 , and ∣ K n ( φ ) ∣ ≤ K f ( n + 1 ) 2 ∥ φ ∥ C m ( [ 0 , a ] ) |\mathcal{K}_n(\varphi)|\leq K_f(n+1)^2\|\varphi\|_{C^m([0,a])} ∣ K n ( φ ) ∣ ≤ K f ( n + 1 ) 2 ∥ φ ∥ C m ([ 0 , a ]) . Taking the second difference in n n n of the quadratic identity ((J.5)) gives
T 2 = K 3 − 2 K 2 + K 1 , T 2 ( q ) = W 2 ( x q ) , ∣ T 2 ( φ ) ∣ ≤ 38 K f ∥ φ ∥ C m ( [ 0 , a ] ) . \mathcal{T}_2=\mathcal{K}_3-2\mathcal{K}_2+\mathcal{K}_1,\qquad\mathcal{T}_2(q)=W_2(xq),\qquad|\mathcal{T}_2(\varphi)|\leq38K_f\|\varphi\|_{C^m([0,a])}. T 2 = K 3 − 2 K 2 + K 1 , T 2 ( q ) = W 2 ( x q ) , ∣ T 2 ( φ ) ∣ ≤ 38 K f ∥ φ ∥ C m ([ 0 , a ]) .
The restriction of T 2 \mathcal{T}_2 T 2 to ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) , divided by the coordinate x x x , is a positive Radon measure η 2 \eta_2 η 2 on ( 1 , a ] (1,a] ( 1 , a ] :
T 2 ( x j − 1 ) = Λ 2 ( h ∗ j ) , j ≥ 1 , T 2 = x η 2 on ( 1 , a ] . (214) \mathcal{T}_2(x^{j-1})=\Lambda_2(h^{*j}),\qquad j\geq1,\qquad\mathcal{T}_2=x\eta_2\quad\text{on }(1,a]. \tag*{(214)} T 2 ( x j − 1 ) = Λ 2 ( h ∗ j ) , j ≥ 1 , T 2 = x η 2 on ( 1 , a ] . (214)
Here Λ 2 \Lambda_2 Λ 2 is the coefficient of ( n 2 ) \binom{n}{2} ( 2 n ) in the first orbital correction (J.3), so that W 2 ( p ) = Λ 2 ( H p ) W_2(p)=\Lambda_2(Hp) W 2 ( p ) = Λ 2 ( H p ) and H x j = h ∗ j Hx_j=h^{*j} H x j = h ∗ j . Huang’s construction allows η 2 \eta_2 η 2 to have an atom at a a a , and his [101, Theorem 6.11] is stated for tests that may be nonzero at a a a ; Proposition 15.5 below shows that no such atom is present; Huang’s argument does not by itself exclude it.
Proposition 15.3 (Critical mean spectrum). If n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g}\to c\in(0,\infty) n / g → c ∈ ( 0 , ∞ ) , then, for every ψ ∈ C c ∞ ( ( 1 , ∞ ) ) \psi\in C^\infty_c((1,\infty)) ψ ∈ C c ∞ (( 1 , ∞ )) ,
E ∑ 0 < λ j < 1 / 4 ψ ( f j 1 / 4 − λ j ) ⟶ c 2 2 ∫ ψ d η 2 . (215) \mathbb{E}\sum_{0<\lambda_j<1/4}\psi\left(f_j\sqrt{1/4-\lambda_j}\right)\longrightarrow\frac{c^2}{2}\int\psi\,d\eta_2. \tag*{(215)} E 0 < λ j < 1/4 ∑ ψ ( f j 1/4 − λ j ) ⟶ 2 c 2 ∫ ψ d η 2 . (215)
The tests may be nonzero at a a a ; the constant eigenfunction is excluded before taking the limit. The assertion includes residual eigenvalues.
Proof. Choose M > c + 1 M>c+1 M > c + 1 , so n + 1 ≤ M g n+1\leq M\sqrt{g} n + 1 ≤ M g eventually. Appendix J extends Huang’s orbital estimate to this envelope, with constants depending on M M M . In particular, for Q = ⌊ g σ ⌋ Q=\lfloor g^\sigma\rfloor Q = ⌊ g σ ⌋ with sufficiently small σ > 0 \sigma>0 σ > 0 , and p = x p ~ p=x\widetilde p p = x p of degree at most Q Q Q , it gives
∣ E O X ( p ) − ω 1 ( p ) − g − 1 ω 2 , n ( p ) ∣ ≤ { 2 ( n + 1 ) 3 A Q 3 g 2 + ( n + 1 ) 2 g 2 − Q } ∥ p ~ ∥ [ 0 , a ] = o ( 1 ) ∥ p ~ ∥ [ 0 , a ] , (216) \left|\mathbb{E}\mathcal{O}_X(p)-\omega_1(p)-g^{-1}\omega_{2,n}(p)\right|
\leq\left\{\frac{2(n+1)^3A_Q^3}{g^2}+\frac{(n+1)^2}{g}2^{-Q}\right\}\|\widetilde p\|_{[0,a]}=o(1)\|\widetilde p\|_{[0,a]}, \tag*{(216)} E O X ( p ) − ω 1 ( p ) − g − 1 ω 2 , n ( p ) ≤ { g 2 2 ( n + 1 ) 3 A Q 3 + g ( n + 1 ) 2 2 − Q } ∥ p ∥ [ 0 , a ] = o ( 1 ) ∥ p ∥ [ 0 , a ] , (216)
where A Q = ( 3 C 0 Q ) D 0 A_Q=(3C_0Q)^{D_0} A Q = ( 3 C 0 Q ) D 0 and the coefficients are exactly those of Lemmas J.8 and J.10 (the coefficients of [101, Lemma 6.7]; the transfer is the deterministic part of the proof of [101, Proposition 6.9]). The same appendix gives E N X ( 1 / 4 − ε ) ≤ C M , ε ( n + 1 ) 2 / g \mathbb{E}N_X(1/4-\varepsilon)\leq C_{M,\varepsilon}(n+1)^2/g E N X ( 1/4 − ε ) ≤ C M , ε ( n + 1 ) 2 / g .
Approximate ξ = ψ / x \xi=\psi/x ξ = ψ / x , extended by zero near zero, by its shifted Chebyshev truncation r Q r_Q r Q of degree Q − 1 Q-1 Q − 1 , and put v Q = x r Q v_Q=xr_Q v Q = x r Q . For every fixed K K K , ∥ r Q − ξ ∥ ∞ = O ( Q − K ) \|r_Q-\xi\|_\infty=O(Q^{-K}) ∥ r Q − ξ ∥ ∞ = O ( Q − K ) , and r Q → ξ r_Q\to\xi r Q → ξ in the fixed C m C^m C m norm controlling T 2 \mathcal{T}_2 T 2 and the three distributions K 1 , K 2 , K 3 \mathcal{K}_1,\mathcal{K}_2,\mathcal{K}_3 K 1 , K 2 , K 3 of Lemma J.11. Let τ X , Q ( p ) = Tr ( χ Y Q p ( A X ) χ Y Q ) \tau_{X,Q}(p)=\operatorname{Tr}(\chi_{Y_Q}p(A_X)\chi_{Y_Q}) τ X , Q ( p ) = Tr ( χ Y Q p ( A X ) χ Y Q ) be the positive trace of Lemma J.6, compressed to the complement of the width-one cusp regions above the common height Y Q = 3 e L Q Y_Q=3e^{LQ} Y Q = 3 e L Q of (J.16). Its expected first moment is O M ( g + n Q ) O_M(g+nQ) O M ( g + n Q ) by Lemma J.9, whose hypothesis is the base case of (J.37). Since A X ≥ 0 A_X\geq0 A X ≥ 0 and ∣ v Q − ψ ∣ ≤ x ∥ r Q − ξ ∥ ∞ |v_Q-\psi|\leq x\|r_Q-\xi\|_\infty ∣ v Q − ψ ∣ ≤ x ∥ r Q − ξ ∥ ∞ on [ 0 , a ] [0,a] [ 0 , a ] , positivity, in the form of Lemma J.6(ii) applied to v Q − ψ = x ( r Q − ξ ) v_Q-\psi=x(r_Q-\xi) v Q − ψ = x ( r Q − ξ ) , bounds the difference between the constant-removed traces τ X , Q ∘ ( v Q ) \tau^\circ_{X,Q}(v_Q) τ X , Q ∘ ( v Q ) and τ X , Q ∘ ( ψ ) = Tr ( χ Y Q ψ ( A X ) ( 1 − P 0 ) χ Y Q ) \tau^\circ_{X,Q}(\psi)=\operatorname{Tr}(\chi_{Y_Q}\psi(A_X)(1-P_0)\chi_{Y_Q}) τ X , Q ∘ ( ψ ) = Tr ( χ Y Q ψ ( A X ) ( 1 − P 0 ) χ Y Q ) of Lemma J.6(v) by ∥ r Q − ξ ∥ ∞ τ X , Q ( x ) \|r_Q-\xi\|_\infty\tau_{X,Q}(x) ∥ r Q − ξ ∥ ∞ τ X , Q ( x ) , whose expectation is O M ( ( g + n Q ) Q − K ) O_M((g+nQ)Q^{-K}) O M (( g + n Q ) Q − K ) . As ψ \psi ψ vanishes on [ 0 , 1 ] [0,1] [ 0 , 1 ] , τ X , Q ∘ ( ψ ) \tau^\circ_{X,Q}(\psi) τ X , Q ∘ ( ψ ) involves only the eigenvalues in ( 0 , 1 / 4 ) (0,1/4) ( 0 , 1/4 ) , each weighted by its retained mass. In the exact trace identity (J.17), the identity, parabolic and leading real-spectrum terms, bounded by (J.18) and by ∣ μ 0 ( p ) ∣ ≤ 5 N R ( p ) |\mu_0(p)|\leq5N R(p) ∣ μ 0 ( p ) ∣ ≤ 5 N R ( p ) , cost O M ( ( g + n Q 2 ) Q − K ) O_M((g+nQ^2)Q^{-K}) O M (( g + n Q 2 ) Q − K ) for this approximation, since ξ \xi ξ vanishes on [ 0 , 1 ] [0,1] [ 0 , 1 ] ; choose K K K large. Remove the constant projection as in (J.19): the removed mass of the constant eigenfunction contributes exactly n v Q ( a ) / ( area ( X ) Y Q ) = o ( 1 ) nv_Q(a)/(\operatorname{area}(X)Y_Q)=o(1) n v Q ( a ) / ( area ( X ) Y Q ) = o ( 1 ) , because v Q ( a ) v_Q(a) v Q ( a ) is bounded and Y Q Y_Q Y Q grows exponentially in Q Q Q . Consequently the expected compressed exceptional trace E τ X , Q ∘ ( v Q ) \mathbb{E}\tau^\circ_{X,Q}(v_Q) E τ X , Q ∘ ( v Q ) equals g − 1 ω 2 , n ( v Q ) + o ( 1 ) g^{-1}\omega_{2,n}(v_Q)+o(1) g − 1 ω 2 , n ( v Q ) + o ( 1 ) by (216) .
The exact quadratic identity (J.5) is ω 2 , n = W 0 + n W 1 + ( n 2 ) W 2 \omega_{2,n}=W_0+nW_1+\binom{n}{2}W_2 ω 2 , n = W 0 + n W 1 + ( 2 n ) W 2 , where, on tests x q xq x q , each W i ( x q ) W_i(xq) W i ( x q ) is a fixed linear combination of K 1 ( q ) , K 2 ( q ) , K 3 ( q ) \mathcal{K}_1(q),\mathcal{K}_2(q),\mathcal{K}_3(q) K 1 ( q ) , K 2 ( q ) , K 3 ( q ) . The C m C^m C m convergence just noted makes each W i ( v Q ) W_i(v_Q) W i ( v Q ) converge; only ( n 2 ) W 2 ( v Q ) / g \binom{n}{2}W_2(v_Q)/g ( 2 n ) W 2 ( v Q ) / g survives. By (214) , its limit is the right side of (215) . Finally, the removed mass of an eigenfunction with λ ≤ 1 / 4 − ε \lambda\leq1/4-\varepsilon λ ≤ 1/4 − ε is at most e − 2 ε L Q e^{-2\sqrt{\varepsilon}LQ} e − 2 ε L Q by Lemma J.7. The expected count bound removes these weights in expectation. Smooth majorants and the same local count bound also extend the conclusion to continuous compact tests on the physical interval [ 0 , 1 / 4 ] [0,1/4] [ 0 , 1/4 ] .
It remains to identify η 2 \eta_2 η 2 . Put w ( s ) = sinh ( s / 2 ) / ( 4 π 2 ) w(s)=\sinh(s/2)/(4\pi^2) w ( s ) = sinh ( s /2 ) / ( 4 π 2 ) . Huang’s explicit quadratic coefficient [101, Section 3.7] gives the exact identity
Λ 2 ( H ) = ∫ 0 ∞ w ( s ) O F s ( H ) d s , O F s ( H ) = ∑ γ p r i m i t i v e o r i e n t e d ∑ k ≥ 1 ℓ γ H ( k ℓ γ ) 2 sinh ( k ℓ γ / 2 ) . (217) \Lambda_2(H)=\int_0^\infty w(s)\mathcal{O}_{F_s}(H)\,ds,\qquad\mathcal{O}_{F_s}(H)=\sum_{\gamma\ \mathrm{primitive\ oriented}}\sum_{k\geq1}\frac{\ell_\gamma H(k\ell_\gamma)}{2\sinh(k\ell_\gamma/2)}. \tag*{(217)} Λ 2 ( H ) = ∫ 0 ∞ w ( s ) O F s ( H ) d s , O F s ( H ) = γ primitive oriented ∑ k ≥ 1 ∑ 2 sinh ( k ℓ γ /2 ) ℓ γ H ( k ℓ γ ) . (217)
The sum includes the two oriented waist classes and all nonsimple classes in the pants. The former give ∑ k ≥ 1 s H ( k s ) / sinh ( k s / 2 ) \sum_{k\geq1}sH(ks)/\sinh(ks/2) ∑ k ≥ 1 sH ( k s ) / sinh ( k s /2 ) , and the latter give Huang’s Z 2 = ( 8 π 2 ) − 1 ∫ 2 sinh ( s / 2 ) I 2 ( s ) d s Z_2=(8\pi^2)^{-1}\int2\sinh(s/2)I_2(s)\,ds Z 2 = ( 8 π 2 ) − 1 ∫ 2 sinh ( s /2 ) I 2 ( s ) d s ; this verifies the normalization in (217) .
The analytic input for the identification is a regularized Selberg zeta function of F s F_s F s that is uniformly controlled through the pinching limit s ↓ 0 s\downarrow0 s ↓ 0 . Let Z s ( z ) Z_s(z) Z s ( z ) be the Selberg zeta function of F s F_s F s in the oriented primitive-class convention of [25], that is, the Euler product used in Lemma 15.2, and set
D s ( z ) = ∏ j ≥ 0 ( 1 − e − ( z + j ) s ) 2 , N s ( z ) = β ( s ) − z ( 1 − z ) . D_s(z)=\prod_{j\geq0}(1-e^{-(z+j)s})^2,\qquad N_s(z)=\beta(s)-z(1-z). D s ( z ) = j ≥ 0 ∏ ( 1 − e − ( z + j ) s ) 2 , N s ( z ) = β ( s ) − z ( 1 − z ) .
Thus D s D_s D s is the contribution of the two oriented waist classes to the Euler product, and N s N_s N s vanishes exactly at z = δ ( s ) z=\delta(s) z = δ ( s ) and z = 1 − δ ( s ) z=1-\delta(s) z = 1 − δ ( s ) .
Lemma 15.4 (Regularized zeta function through pinching). For every s > 0 s>0 s > 0 the function Z s / ( D s N s ) Z_s/(D_sN_s) Z s / ( D s N s ) is holomorphic and zero-free on { ℜ z > 1 / 2 } \{\Re z>1/2\} { ℜ z > 1/2 } . For every S > 0 S>0 S > 0 and every compact K ⊂ { ℜ z > 1 / 2 } K\subset\{\Re z>1/2\} K ⊂ { ℜ z > 1/2 } there are constants 0 < c K , S ≤ C K , S < ∞ 0<c_{K,S}\leq C_{K,S}<\infty 0 < c K , S ≤ C K , S < ∞ such that
c K , S ≤ ∣ Z s ( z ) D s ( z ) N s ( z ) ∣ ≤ C K , S ( z ∈ K , 0 < s ≤ S ) . c_{K,S}\leq\left|\frac{Z_s(z)}{D_s(z)N_s(z)}\right|\leq C_{K,S}\qquad(z\in K,\ 0<s\leq S). c K , S ≤ D s ( z ) N s ( z ) Z s ( z ) ≤ C K , S ( z ∈ K , 0 < s ≤ S ) .
Proof. We place F s F_s F s in the framework of Schulze [170]. There a geometrically finite hyperbolic surface is assembled from pairs of pants along an augmented admissible graph G ∗ G^* G ∗ [170, Section 1.2]: the proper edges carry the gluing cylinders, and the remaining phantom edges carry the infinite ends, a phantom edge of positive length ℓ \ell ℓ giving a funnel and one of length zero a cusp. The parameter space is B = { ( ℓ , τ ) : ℓ : G 1 ∗ → [ 0 , ∞ ) , τ : G 1 ∗ → R } \mathcal{B}=\{(\ell,\tau):\ell:G^*_1\to[0,\infty),\ \tau:G^*_1\to\mathbb{R}\} B = {( ℓ , τ ) : ℓ : G 1 ∗ → [ 0 , ∞ ) , τ : G 1 ∗ → R } , in which length zero is admitted on every edge [170, Section 2.2]. The surface F s F_s F s is the non-elementary component of Y G ( λ s ) Y_G(\lambda_s) Y G ( λ s ) for the graph with one vertex and three phantom edges, two of length zero (the cusps) and one, the funnel edge d d d , of length ℓ ( d ) = s \ell(d)=s ℓ ( d ) = s ; the twist parameters play no role on phantom edges. Each zero-length phantom edge also contributes an isolated elementary parabolic component, so Y G ( λ s ) Y_G(\lambda_s) Y G ( λ s ) has two such components for s > 0 s>0 s > 0 [170, Definition 1.5 and the discussion following it]. These components have no L 2 L^2 L 2 eigenvalues and contribute the factor 1 1 1 to the Selberg zeta function [170, Section 1.4]. Shrinking the funnel edge, ℓ ( d ) = s ↓ 0 \ell(d)=s\downarrow0 ℓ ( d ) = s ↓ 0 , is therefore a degeneration inside B \mathcal{B} B , and U = { λ s : 0 < s ≤ S } \mathcal{U}=\{\lambda_s:0<s\leq S\} U = { λ s : 0 < s ≤ S } has compact closure { λ s : 0 ≤ s ≤ S } \{\lambda_s:0\leq s\leq S\} { λ s : 0 ≤ s ≤ S } in B \mathcal{B} B . The limit surface Y G ( λ 0 ) Y_G(\lambda_0) Y G ( λ 0 ) is the disjoint union of a thrice-punctured sphere and three elementary parabolic components, each isometric to H / ⟨ z ↦ z + 1 ⟩ \mathbb{H}/\langle z\mapsto z+1\rangle H / ⟨ z ↦ z + 1 ⟩ .
In this notation Schulze’s regularizing factor for the edge d d d is Z d , λ s ( z ) = ∏ k ≥ 0 ( 1 − e − ( z + k ) s ) 2 = D s ( z ) Z_{d,\lambda_s}(z)=\prod_{k\geq0}(1-e^{-(z+k)s})^2=D_s(z) Z d , λ s ( z ) = ∏ k ≥ 0 ( 1 − e − ( z + k ) s ) 2 = D s ( z ) [170, Section 1.4 and Lemma 3.22], and N ( λ s , z ) = ∏ t < 1 / 4 ( t − z ( 1 − z ) ) \mathcal{N}(\lambda_s,z)=\prod_{t<1/4}(t-z(1-z)) N ( λ s , z ) = ∏ t < 1/4 ( t − z ( 1 − z )) , the product over the eigenvalues t t t of Y G ( λ s ) Y_G(\lambda_s) Y G ( λ s ) below 1 / 4 1/4 1/4 , equals N s ( z ) N_s(z) N s ( z ) : for s > 0 s>0 s > 0 the only such eigenvalue is β ( s ) \beta(s) β ( s ) , by [16, Theorem 1.5 and Corollary 1.6 together with δ ( s ) > 1 / 2 \delta(s)>1/2 δ ( s ) > 1/2 , while at s = 0 s=0 s = 0 it is the eigenvalue 0 0 0 of the constant function on the finite-area component, the elementary components having no L 2 L^2 L 2 eigenvalues; this agrees with the limit β ( s ) → 0 \beta(s)\to0 β ( s ) → 0 as s ↓ 0 s\downarrow0 s ↓ 0 recorded after (206) , that is, with the continuity of β \beta β through pinching [170, Theorem 2.14]. Theorem 3.21 of [170], applied to the relatively compact sets K K K and U \mathcal{U} U , gives the two-sided bound. Theorem 3.18 there gives continuity of λ ↦ Z λ / ∏ d Z d , λ \lambda\mapsto\mathcal{Z}_\lambda/\prod_d\mathcal{Z}_{d,\lambda} λ ↦ Z λ / ∏ d Z d , λ on B \mathcal{B} B for the topology of locally uniform convergence on { ℜ z > 1 / 2 } \{\Re z>1/2\} { ℜ z > 1/2 } ; in particular Z s / D s Z_s/D_s Z s / D s is holomorphic on { ℜ z > 1 / 2 } \{\Re z>1/2\} { ℜ z > 1/2 } for each s > 0 s>0 s > 0 . Finally, by [25, Theorem 1.1], the zeros of Z s Z_s Z s in { ℜ z > 1 / 2 } \{\Re z>1/2\} { ℜ z > 1/2 } are exactly the points z z z for which z ( 1 − z ) z(1-z) z ( 1 − z ) is an L 2 L^2 L 2 eigenvalue of F s F_s F s , with multiplicity equal to that of the eigenvalue. By the uniqueness statement just quoted, the only such zero is a simple zero at z = δ ( s ) z=\delta(s) z = δ ( s ) , which is cancelled by the simple zero of N s N_s N s ; and D s D_s D s has no zeros in { ℜ z > 1 / 2 } \{\Re z>1/2\} { ℜ z > 1/2 } . Hence Z s / ( D s N s ) Z_s/(D_sN_s) Z s / ( D s N s ) is holomorphic and zero-free there. □
Proposition 15.5 (Identification of the exceptional measure). With x s = f ( i ( δ ( s ) − 1 / 2 ) ) x_s=f(i(\delta(s)-1/2)) x s = f ( i ( δ ( s ) − 1/2 )) ,
η 2 = ∫ 0 ∞ w ( s ) δ x s d s on ( 1 , a ] . (218) \eta_2=\int_0^\infty w(s)\delta_{x_s}\,ds\quad\text{on }(1,a]. \tag*{(218)} η 2 = ∫ 0 ∞ w ( s ) δ x s d s on ( 1 , a ] . (218) In particular it has no atom at a a a , and the physical mean measure in Proposition Proposition 15.3 is ν c \nu_c ν c .
Proof. We keep the infinite contour tails on an absolutely convergent line. This avoids a uniform high-frequency scattering estimate for pinching surfaces. By Lemma 15.4 , for every S > 0 S>0 S > 0 the function Z s / ( D s N s ) Z_s/(D_sN_s) Z s / ( D s N s ) is holomorphic and zero-free in ℜ z > 1 / 2 \Re z>1/2 ℜ z > 1/2 and bounded above and away from zero on each compact subset, uniformly for 0 < s ≤ S 0<s\leq S 0 < s ≤ S . Cauchy’s estimate for its derivative on slightly larger compact sets, divided by the lower bound, therefore gives
Z s ′ Z s ( z ) = 2 z − 1 β ( s ) − z ( 1 − z ) + D s ′ D s ( z ) + E s ( z ) , (219) \frac{Z_s'}{Z_s}(z)=\frac{2z-1}{\beta(s)-z(1-z)}+\frac{D_s'}{D_s}(z)+E_s(z), \tag*{(219)} Z s Z s ′ ( z ) = β ( s ) − z ( 1 − z ) 2 z − 1 + D s D s ′ ( z ) + E s ( z ) , (219)
where E s E_s E s is locally uniformly bounded and holomorphic, including the pinching limit. The elementary series D s ′ / D s = 2 ∑ k ≥ 1 s e − z k s / ( 1 − e − k s ) D_s'/D_s=2\sum_{k\geq1}se^{-zks}/(1-e^{-ks}) D s ′ / D s = 2 ∑ k ≥ 1 s e − z k s / ( 1 − e − k s ) is O K ( 1 + ∣ log s ∣ ) O_K(1+|\log s|) O K ( 1 + ∣ log s ∣ ) on each compact K ⊂ { ℜ z > 1 / 2 } K\subset\{\Re z>1/2\} K ⊂ { ℜ z > 1/2 } : split at k s = 1 ks=1 k s = 1 and bound the first part by a harmonic sum. This is integrable against w ( s ) d s w(s)\,ds w ( s ) d s at zero.
For ℜ z > 1 \Re z>1 ℜ z > 1 define
A ( z ) = ∫ 0 ∞ w ( s ) Z s ′ Z s ( z ) d s . A(z)=\int_0^\infty w(s)\frac{Z_s'}{Z_s}(z)\,ds. A ( z ) = ∫ 0 ∞ w ( s ) Z s Z s ′ ( z ) d s .
The preceding bound and Lemma 15.2 give absolute convergence. By the Euler product and (217) , A A A is the Laplace transform ∫ 0 ∞ e − ( z − 1 / 2 ) u M ( d u ) \int_0^\infty e^{-(z-1/2)u}\mathcal{M}(du) ∫ 0 ∞ e − ( z − 1/2 ) u M ( d u ) of the positive orbital measure defining Λ 2 \Lambda_2 Λ 2 on positive lengths. Hence ∣ A ( σ + i y ) ∣ ≤ A ( σ ) |A(\sigma+iy)|\leq A(\sigma) ∣ A ( σ + i y ) ∣ ≤ A ( σ ) for σ > 1 \sigma>1 σ > 1 . For any σ 0 > 1 / 2 \sigma_0>1/2 σ 0 > 1/2 , choose S S S large using Lemma 15.2. Then
A ( z ) = M S ( z ) + H ( z ) , M S ( z ) = ∫ 0 S w ( s ) 2 z − 1 β ( s ) − z ( 1 − z ) d s , (220) A(z)=M_S(z)+H(z),\qquad M_S(z)=\int_0^S w(s)\frac{2z-1}{\beta(s)-z(1-z)}\,ds, \tag*{(220)} A ( z ) = M S ( z ) + H ( z ) , M S ( z ) = ∫ 0 S w ( s ) β ( s ) − z ( 1 − z ) 2 z − 1 d s , (220)
where H H H is holomorphic on a half-plane slightly larger than ℜ z ≥ σ 0 \Re z\geq\sigma_0 ℜ z ≥ σ 0 . This follows from (219) near zero, ordinary compact family continuity for positive bounded waists, and the absolutely convergent large-waist tail. Initially the equality holds for ℜ z > 1 \Re z>1 ℜ z > 1 ; elsewhere it specifies the regularized continuation. No growth assertion for H H H in a whole vertical strip is needed.
Put F ( z ) = f ( i ( z − 1 / 2 ) ) F(z)=f(i(z-1/2)) F ( z ) = f ( i ( z − 1/2 )) and let B ( v ) = T 2 ( ( v − x ) − 1 ) B(v)=\mathcal{T}_2((v-x)^{-1}) B ( v ) = T 2 (( v − x ) − 1 ) , for v ∉ [ 0 , a ] v\notin[0,a] v ∈ / [ 0 , a ] . The moment identity (214) gives B ( v ) = ∑ j ≥ 1 Λ 2 ( h ∗ j ) v − j B(v)=\sum_{j\geq1}\Lambda_2(h^{*j})v^{-j} B ( v ) = ∑ j ≥ 1 Λ 2 ( h ∗ j ) v − j near infinity. Fourier inversion shifted to ℑ r = σ − 1 / 2 \Im r=\sigma-1/2 ℑ r = σ − 1/2 , followed by integration against M \mathcal{M} M , yields, for real v > F ( σ ) v>F(\sigma) v > F ( σ ) ,
B ( v ) = 1 2 π i ∫ σ − i ∞ σ + i ∞ F ( z ) v − F ( z ) A ( z ) d z . (221) B(v)=\frac{1}{2\pi i}\int_{\sigma-i\infty}^{\sigma+i\infty}\frac{F(z)}{v-F(z)}A(z)\,dz. \tag*{(221)} B ( v ) = 2 π i 1 ∫ σ − i ∞ σ + i ∞ v − F ( z ) F ( z ) A ( z ) d z . (221)
Indeed h ∗ j ( u ) = ( 2 π ) − 1 ∫ f ( r + i ( σ − 1 / 2 ) ) j e i r u − ( σ − 1 / 2 ) u d r h^{*j}(u)=(2\pi)^{-1}\int f(r+i(\sigma-1/2))^j e^{iru-(\sigma-1/2)u}\,dr h ∗ j ( u ) = ( 2 π ) − 1 ∫ f ( r + i ( σ − 1/2 ) ) j e i r u − ( σ − 1/2 ) u d r ; set z = σ − i r z=\sigma-ir z = σ − i r and reverse orientation. The bound ∣ F ( σ + i y ) ∣ ≤ F ( σ ) < v |F(\sigma+iy)|\leq F(\sigma)<v ∣ F ( σ + i y ) ∣ ≤ F ( σ ) < v and rapid vertical decay justify every interchange and the geometric sum.
The M S M_S M S contribution to (221) is
∫ 0 S w ( s ) x s v − x s d s . (222) \int_0^S w(s)\frac{x_s}{v-x_s}\,ds. \tag*{(222)} ∫ 0 S w ( s ) v − x s x s d s . (222)
For fixed s s s , its resolvent factor is 1 / ( z − δ ( s ) ) + 1 / ( z − ( 1 − δ ( s ) ) ) 1/(z-\delta(s))+1/(z-(1-\delta(s))) 1/ ( z − δ ( s )) + 1/ ( z − ( 1 − δ ( s ))) . Shift to ℜ z = 1 / 2 \Re z=1/2 ℜ z = 1/2 . Only the pole δ ( s ) \delta(s) δ ( s ) is crossed, with residue x s / ( v − x s ) x_s/(v-x_s) x s / ( v − x s ) ; the new vertical integral is zero because F ( 1 / 2 + i y ) = f ( y ) F(1/2+iy)=f(y) F ( 1/2 + i y ) = f ( y ) is even, while the resolvent factor is odd in y y y . Horizontal contributions vanish by rapid decay, and v − F v-F v − F has no zero in the strip. Integrating over 0 ≤ s ≤ S 0\leq s\leq S 0 ≤ s ≤ S is legitimate since δ ( s ) \delta(s) δ ( s ) stays uniformly above 1 / 2 1/2 1/2 there, including its limit 1 1 1 at zero. This also checks the residue and orientation factors.
Fix 1 < b < a 1<b<a 1 < b < a and choose σ 0 > 1 / 2 \sigma_0>1/2 σ 0 > 1/2 so F ( σ 0 ) < b F(\sigma_0)<b F ( σ 0 ) < b . Choose S S S as in (220) , also ensuring δ ( s ) < σ 0 \delta(s)<\sigma_0 δ ( s ) < σ 0 for s > S s>S s > S . Choose R R R so large that ∣ F ∣ < b |F|<b ∣ F ∣ < b on the horizontal segments at heights ± R \pm R ± R in σ 0 ≤ ℜ z ≤ σ \sigma_0\le\Re z\le\sigma σ 0 ≤ ℜ z ≤ σ , and on the two tails ℜ z = σ \Re z=\sigma ℜ z = σ , ∣ ℑ z ∣ ≥ R |\Im z|\ge R ∣ℑ z ∣ ≥ R . Deform only the central segment of the H H H integral to ℜ z = σ 0 \Re z=\sigma_0 ℜ z = σ 0 , retaining those two original infinite tails and connecting them by the horizontal segments. On this contour ∣ F ∣ < b |F|<b ∣ F ∣ < b . The resulting integral D ( v ) D(v) D ( v ) is holomorphic for ∣ v ∣ > b |v|>b ∣ v ∣ > b : the finite pieces are compact, and on the tails H = A − M S H=A-M_S H = A − M S is bounded and F F F decays rapidly. Finite contour deformation for large real v v v and then analytic continuation give
B ( v ) − ∫ 0 S w ( s ) x s v − x s d s = D ( v ) . B(v)-\int_0^S w(s)\frac{x_s}{v-x_s}\,ds=D(v). B ( v ) − ∫ 0 S w ( s ) v − x s x s d s = D ( v ) .
Thus the Cauchy transform of T 2 − ∫ 0 S w ( s ) x s δ x s d s \mathcal{T}_2-\int_0^S w(s)x_s\delta_{x_s}\,ds T 2 − ∫ 0 S w ( s ) x s δ x s d s continues holomorphically through ( b , ∞ ) (b,\infty) ( b , ∞ ) . The distributional jump formula implies that this distribution vanishes there: the difference of its upper and lower boundary values is − 2 π i -2\pi i − 2 π i times the distribution, by convolution with the Poisson approximate identity. This argument includes a neighborhood of a a a , excluding endpoint atoms and their derivatives. Divide by the nonzero coordinate x x x . Since s > S s>S s > S implies x s < b x_s<b x s < b , and b > 1 b>1 b > 1 was arbitrary, (218) follows. Finally β ( s ) > 0 \beta(s)>0 β ( s ) > 0 for s > 0 s>0 s > 0 , so there is no mass at zero energy; the factor c 2 / 2 c^2/2 c 2 /2 in Proposition 15.3 gives exactly ν c \nu_c ν c . □
The identification of η 2 \eta_2 η 2 is independent of the cusp scale and is reused twice: in § 4.19 , where the same pushforward gives the fixed-window density ν ∗ \nu_* ν ∗ when g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g , and in Appendix § 10.15 , where the model energies near 1 / 4 1/4 1/4 enter the moving-edge law of § 4.12 .
Localization and exhaustion of the mean
This subsection supplies the lower half of the count matching and completes the proof of the point-process and gap assertions of Theorem 15.1. Lemma 15.6 shows that short two-cusp pairs of pants have deep, disjoint funnel neighborhoods in X X X , using the cut-graph Lemmas I.1 and I.3 of Appendix § 10.6 ; Lemma 15.7 transplants model ground states into these neighborhoods to bound the spectral count from below by the waist count. Together with the mean limit of Propositions 15.3 and 15.5, this forces the actual count and the waist count to agree in L 1 L^1 L 1 .
Lemma 15.6 (Deep disjoint model neighborhoods). Fix 0 < d < S < ∞ 0<d<S<\infty 0 < d < S < ∞ and R < ∞ R<\infty R < ∞ . If n / g → c > 0 n/\sqrt{g}\to c>0 n / g → c > 0 , with probability tending to one every two-cusp pants of waist length in [ d , S ] [d,S] [ d , S ] has an embedded exterior funnel neighborhood of radius R R R , and all these enlarged pants are pairwise disjoint.
Proof. Lemma I.3 first excludes intersecting waists. Expand the exterior normal collars simultaneously. Their metric is d r 2 + s 2 cosh 2 r d θ 2 dr^2+s^2\cosh^2r\,d\theta^2 d r 2 + s 2 cosh 2 r d θ 2 with θ ∈ R / Z \theta\in\mathbb{R}/\mathbb{Z} θ ∈ R / Z , and there are no focal points. A first failure before radius R R R supplies an embedded essential arc of length at most 2 R 2R 2 R in the complement, joining one waist to itself or two distinct waists. A boundary-parallel arc cannot cause a first collision, since the corresponding cyclic funnel cover has embedded normal coordinates.
A regular neighborhood of one pants and a returning arc has genus zero, two cusps and two boundary components. Tighten its essential boundaries. Filling a peripheral punctured disk instead gives a genus-zero piece with at least three cusps and one essential boundary. An unpunctured disk would make the arc boundary-parallel; both boundaries cannot disappear when the ambient genus tends to infinity. If two tightened boundaries coincide, retain the two incidences of the same cut edge. An arc between distinct pants produces a genus-zero piece with four cusps and one boundary. In all cases the total cut length is bounded, for example by 4 S + 8 R + 1 4S+8R+1 4 S + 8 R + 1 .
There are finitely many cut-graph shapes with at most two edges. The first type has a specified degree-two genus-zero vertex and therefore v ≥ 1 v\ge1 v ≥ 1 , w ≥ 1 w\ge1 w ≥ 1 in Lemma I.1; its expected count is O ( g − 1 / 2 ) O(g^{-1/2}) O ( g − 1/2 ) . The degree-one types have at least three cusps and receive that lemma’s extra g − 1 / 2 g^{-1/2} g − 1/2 factor. Markov’s inequality excludes all of them. Only existence of such a cut is used; no bound on the number of arcs realizing a cut is needed. □
We now show that the localized model states give a lower bound for the actual eigenvalue count; together with the mean computed above, this lower bound forces count matching. Write N X ( t ) = Ξ X ( [ 0 , t ] ) N_X(t)=\Xi_X([0,t]) N X ( t ) = Ξ X ([ 0 , t ]) and Z g ( t ) = # { s ∈ Q X : β ( s ) ≤ t } Z_g(t)=\#\{s\in Q_X:\beta(s)\le t\} Z g ( t ) = # { s ∈ Q X : β ( s ) ≤ t } , where Q X Q_X Q X is the pants-waist process. At every continuity threshold of ν c \nu_c ν c , the waist-process limit of Lemma I.4 and the factorial-moment bounds in its proof give convergence in distribution and uniform integrability of Z g ( t ) Z_g(t) Z g ( t ) . Indeed the boundary of { s : β ( s ) ≤ t } \{s:\beta(s)\le t\} { s : β ( s ) ≤ t } lies in { 0 } ∪ { β = t } \{0\}\cup\{\beta=t\} { 0 } ∪ { β = t } , which is μ c \mu_c μ c -null; β ( s ) ≤ t \beta(s)\le t β ( s ) ≤ t forces s ≤ S t s\le S_t s ≤ S t , and the second moment is bounded by that of Q X ( [ 0 , S t ] ) Q_X([0,S_t]) Q X ([ 0 , S t ]) . The expected number of waists below d d d is O c ( d 2 ) O_c(d^2) O c ( d 2 ) , uniformly for large g g g , by the same integration formula.
Lemma 15.7 (Localized lower count). Suppose n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g}\to c\in(0,\infty) n / g → c ∈ ( 0 , ∞ ) . Then
P { N X ( t ) ≥ Z g ( t − η ) } ⟶ 1 , 0 < η < t < 1 / 4. (223) \mathbb{P}\{N_X(t)\ge Z_g(t-\eta)\}\longrightarrow1,\qquad0<\eta<t<1/4. \tag*{(223)} P { N X ( t ) ≥ Z g ( t − η )} ⟶ 1 , 0 < η < t < 1/4. (223) Proof. For each compact waist interval [ d , S ] [d,S] [ d , S ] and ϵ > 0 \epsilon>0 ϵ > 0 , choose normalized compactly supported form-domain states v s v_s v s on F s F_s F s with Rayleigh quotient at most β ( s ) + ϵ \beta(s)+\epsilon β ( s ) + ϵ . Their supports may be taken in uniformly bounded funnel radius and cusp heights, and ( ∫ ∣ v s ∣ ) 2 ≤ C (\int|v_s|)^2\le C ( ∫ ∣ v s ∣ ) 2 ≤ C uniformly. To see this, approximate a ground state at each fixed parameter in H 1 H^1 H 1 by a smooth compactly supported function, transport it to neighboring marked metrics, and use continuity of its norm, energy and β \beta β . A finite cover of [ d , S ] [d,S] [ d , S ] supplies common bounds; no uniformity down to s = 0 s=0 s = 0 is required.
On the event of Lemma 15.6, transplant the states of k k k selected pants into X X X ; they have disjoint supports. For u = ∑ a i v i u=\sum a_i v_i u = ∑ a i v i and the projection P P P off the global constant,
∥ P u ∥ 2 2 ≥ ( 1 − k C area ( X ) ) ∑ i ∣ a i ∣ 2 , E ( P u ) = E ( u ) . (224) \|Pu\|_2^2\ge\left(1-\frac{kC}{\operatorname{area}(X)}\right)\sum_i|a_i|^2,\qquad\mathcal{E}(Pu)=\mathcal{E}(u). \tag*{(224)} ∥ P u ∥ 2 2 ≥ ( 1 − area ( X ) k C ) i ∑ ∣ a i ∣ 2 , E ( P u ) = E ( u ) . (224) Thus, if all selected energies are at most t − η t-\eta t − η , use trial error η / 2 \eta/2 η /2 and let g → ∞ g\to\infty g → ∞ . For bounded k k k , the projection is injective and all projected Rayleigh quotients are below t t t , so min–max gives k k k positive eigenvalues, without losing a mode to the constant. Tightness of pants counts removes the bound on k k k . Removing the lower waist cutoff d d d and using the deterministic upper cutoff S t S_t S t of Lemma 15.2 gives (223) . More precisely, the failure probability first has limit superior O c ( d 2 ) O_c(d^2) O c ( d 2 ) ; let d ↓ 0 d\downarrow0 d ↓ 0 to obtain the displayed limit. □
Proof of the total-spectrum assertions in Theorem 15.1. Propositions 15.3 and 15.5 give E Ξ X → ν c \mathbb{E}\Xi_X\to\nu_c E Ξ X → ν c vaguely on [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) . This is a two-sided limit of the mean measure: at every continuity threshold t t t of ν c \nu_c ν c it gives E N X ( t ) → ν c ( [ 0 , t ] ) \mathbb{E}N_X(t)\to\nu_c([0,t]) E N X ( t ) → ν c ([ 0 , t ]) . In the count matching below it is used as the upper half, through the first term on the right of the L 1 L^1 L 1 identity, while the lower half is supplied by the localized lower count of Lemma 15.7 . Take a continuity threshold t t t and choose t − η t-\eta t − η also to be one. On the good event in (223) ,
( Z g ( t ) − N X ( t ) ) + ≤ Z g ( t ) − Z g ( t − η ) . \bigl(Z_g(t)-N_X(t)\bigr)_+\le Z_g(t)-Z_g(t-\eta). ( Z g ( t ) − N X ( t ) ) + ≤ Z g ( t ) − Z g ( t − η ) .
On its complement the left side is at most Z g ( t ) Z_g(t) Z g ( t ) , whose uniform integrability makes that contribution vanish. First let g → ∞ g\to\infty g → ∞ and then η ↓ 0 \eta\downarrow0 η ↓ 0 . The limiting expected increment vanishes, so E ( Z g ( t ) − N X ( t ) ) + → 0 \mathbb{E}(Z_g(t)-N_X(t))_+\to0 E ( Z g ( t ) − N X ( t ) ) + → 0 . Since both expected counts converge to ν c ( [ 0 , t ] ) \nu_c([0,t]) ν c ([ 0 , t ]) , the identity
E ∣ N X ( t ) − Z g ( t ) ∣ = E N X ( t ) − E Z g ( t ) + 2 E ( Z g ( t ) − N X ( t ) ) + \mathbb{E}|N_X(t)-Z_g(t)|=\mathbb{E}N_X(t)-\mathbb{E}Z_g(t)+2\mathbb{E}(Z_g(t)-N_X(t))_+ E ∣ N X ( t ) − Z g ( t ) ∣ = E N X ( t ) − E Z g ( t ) + 2 E ( Z g ( t ) − N X ( t ) ) +
proves L 1 L^1 L 1 matching. Finitely many thresholds give joint count convergence; local mean bounds give tightness on [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) . This proves the point-process assertion. It does not assume uniform integrability of the actual spectral counts.
The pushforward of a Poisson random measure is Poisson, with intensity (207) . Its void probability is (209) . The measure is locally finite at zero and has no atom there. Its total mass is infinite, since every β ( s ) < 1 / 4 \beta(s)<1/4 β ( s ) < 1/4 while ∫ 0 ∞ sinh ( s / 2 ) d s = ∞ \int_0^\infty\sinh(s/2)\,ds=\infty ∫ 0 ∞ sinh ( s /2 ) d s = ∞ . It follows that 0 < G c < 1 / 4 0<G_c<1/4 0 < G c < 1/4 almost surely. Count convergence at continuity thresholds now gives convergence of the gap, including the possible value 1 / 4 1/4 1/4 for a finite surface without small positive eigenvalues. The same argument gives convergence of every fixed ordered small eigenvalue to the corresponding Poisson order statistic; each such eigenvalue exists with probability tending to one.
The residual character of the critical spectrum
The count matching just proved also identifies the spectral type, once cusp averages are controlled on the whole low eigenspace; this gives the residual assertions (15.5)–(15.6) of Theorem 15.1. The mechanism is the one of Lemma 14.6, now applied to the exact low eigenspace rather than to a trial space. For a compact annulus A A A in a cusp, bounded by two horocycles of length at most one, put h A = 1 A / area ( A ) h_A = 1_A/\sqrt{\operatorname{area}(A)} h A = 1 A / area ( A ) . Every cuspidal eigenfunction is orthogonal to h A h_A h A : its constant Fourier coefficient vanishes throughout the cusp cylinder, by the Fourier equation. The same holds for finite linear combinations.
Lemma 15.8 (Injectivity of the cusp-average map). (a) For every compact [ d , S ] ⊂ ( 0 , ∞ ) [d,S] \subset(0,\infty) [ d , S ] ⊂ ( 0 , ∞ ) and ζ > 0 \zeta>0 ζ > 0 , there are smoothly compactly supported functions f s f_s f s on F s F_s F s and compact cusp annuli A s ⊂ P s A_s \subset P_s A s ⊂ P s , s ∈ [ d , S ] s \in[d,S] s ∈ [ d , S ] , such that
⟨ f s , h A s ⟩ = 1 , ∥ ( Δ F s − β ( s ) ) f s ∥ 2 ≤ ζ . (225) \langle f_s,h_{A_s}\rangle= 1,\qquad\lVert(\Delta_{F_s}-\beta(s))f_s\rVert_2 \le\zeta. \tag*{(225)} ⟨ f s , h A s ⟩ = 1 , ∥( Δ F s − β ( s )) f s ∥ 2 ≤ ζ . (225) Their supports have uniformly bounded funnel radius and cusp heights. No unit-norm condition on f s f_s f s is imposed.
(b) Let 0 < b < t < 1 / 4 0<b<t<1/4 0 < b < t < 1/4 . Suppose that k k k models F s 1 , … , F s k F_{s_1},\ldots,F_{s_k} F s 1 , … , F s k with energies β i = β ( s i ) ∈ [ b , t − b ] \beta_i=\beta(s_i)\in[b,t-b] β i = β ( s i ) ∈ [ b , t − b ] , together with the functions f i = f s i f_i=f_{s_i} f i = f s i and annuli A i = A s i A_i=A_{s_i} A i = A s i of (a), have been transplanted isometrically into X X X with pairwise disjoint supports, and that dim E = k \dim E=k dim E = k for E = 1 ( 0 , t ] ( Δ X ) L 2 ( X ) E=\mathbf{1}_{(0,t]}(\Delta_X)L^2(X) E = 1 ( 0 , t ] ( Δ X ) L 2 ( X ) . Let P P P be the orthogonal projection onto E E E , put F a = ∑ i a i f i F a=\sum_i a_i f_i F a = ∑ i a i f i for a ∈ C k a\in\mathbb{C}^k a ∈ C k , and T v = ( ⟨ v , h A i ⟩ ) i ≤ k Tv=(\langle v,h_{A_i}\rangle)_{i\le k} T v = (⟨ v , h A i ⟩ ) i ≤ k . Then T F = I TF=I T F = I , ∥ T ∥ ≤ 1 \lVert T\rVert\le1 ∥ T ∥ ≤ 1 , and
∥ ( I − P ) f i ∥ 2 ≤ ζ / b , ∥ T P F − I ∥ ≤ k ζ / b . (226) \lVert(I-P)f_i\rVert_2\le\zeta/b,\qquad\lVert TPF-I\rVert\le\sqrt{k}\,\zeta/b. \tag*{(226)} ∥( I − P ) f i ∥ 2 ≤ ζ / b , ∥ T P F − I ∥ ≤ k ζ / b . (226) If k ζ / b < 1 \sqrt{k}\,\zeta/b<1 k ζ / b < 1 , then T T T is injective on E E E , and E E E consists of residual eigenfunctions.
Proof. (a) To construct the functions at a fixed parameter, take the positive ground state u u u and divide it by ⟨ u , h A ⟩ > 0 \langle u,h_A\rangle>0 ⟨ u , h A ⟩ > 0 . Multiply by smooth compact cutoffs χ R \chi_R χ R that equal one on A A A and a growing compact core. In logarithmic cusp height and funnel distance, choose the end transitions with ∣ ∇ χ R ∣ + ∣ Δ χ R ∣ ≤ C / R \lvert\nabla\chi_R\rvert+\lvert\Delta\chi_R\rvert\le C/R ∣ ∇ χ R ∣ + ∣ Δ χ R ∣ ≤ C / R . Then
( Δ − β ) ( χ R u ) = ( Δ χ R ) u − 2 ⟨ ∇ χ R , ∇ u ⟩ ⟶ 0 in L 2 , (\Delta-\beta)(\chi_Ru)=(\Delta\chi_R)u-2\langle\nabla\chi_R,\nabla u\rangle\longrightarrow0\quad\text{in }L^2, ( Δ − β ) ( χ R u ) = ( Δ χ R ) u − 2 ⟨ ∇ χ R , ∇ u ⟩ ⟶ 0 in L 2 ,
since u u u has finite norm and energy. The observation remains one. Transport a fixed such compact function to nearby marked metrics, preserving the cusp coordinates on the annulus. The operator residual depends continuously on the parameter, by continuity of β \beta β . Renormalization by the positive observation, if needed, preserves this continuity. A finite cover of [ d , S ] [d,S] [ d , S ] supplies the common support bounds in (225) . This argument requires neither a globally chosen continuous ground state nor a uniform branching limit.
(b) Since the transplanted supports are pairwise disjoint and the transplantation is isometric, ⟨ f j , h A i ⟩ = δ i j \langle f_j,h_{A_i}\rangle=\delta_{ij} ⟨ f j , h A i ⟩ = δ ij by (225) , so T F = I TF=I T F = I ; and ∥ T ∥ ≤ 1 \lVert T\rVert\le1 ∥ T ∥ ≤ 1 because the h A i h_{A_i} h A i are orthonormal. Also ∥ ( Δ X − β i ) f i ∥ 2 ≤ ζ \lVert(\Delta_X-\beta_i)f_i\rVert_2\le\zeta ∥( Δ X − β i ) f i ∥ 2 ≤ ζ . The spectral theorem and (225) give (226) : the complement of P P P includes zero and the continuous spectrum, and its distance from each β i \beta_i β i is at least b b b , so ∥ ( I − P ) f i ∥ 2 ≤ ζ / b \lVert(I-P)f_i\rVert_2\le\zeta/b ∥( I − P ) f i ∥ 2 ≤ ζ / b ; then ∥ T P F − I ∥ = ∥ T ( P − I ) F ∥ ≤ ∥ ( I − P ) F ∥ ≤ k ζ / b \lVert TPF-I\rVert=\lVert T(P-I)F\rVert\le\lVert(I-P)F\rVert\le\sqrt{k}\,\zeta/b ∥ T P F − I ∥ = ∥ T ( P − I ) F ∥ ≤ ∥( I − P ) F ∥ ≤ k ζ / b . If the last bound is less than one, T P F TPF T P F is invertible. Thus P F PF P F maps onto E E E and T T T is injective on E E E . Since cusp-form combinations have zero cusp averages, no nonzero cusp combination can lie in E E E . The orthogonal residual–cuspidal decomposition of each discrete eigenspace therefore makes E E E entirely residual, even when eigenvalues have multiplicity. No separation between the individual β i \beta_i β i is needed.
Proof of (15.5)–(15.6). First suppose n / q → c > 0 n/\sqrt{q}\to c>0 n / q → c > 0 . Fix a continuity threshold t < 1 / 4 t<1/4 t < 1/4 . The preceding total-spectrum proof gives ∣ E ∣ / N X ( t ) − Z g ( t ) ∣ → 0 \lvert E\lvert/N_X(t)-Z_g(t)\rvert\to0 ∣ E ∣ / N X ( t ) − Z g ( t )∣ → 0 , hence P { N X ( t ) = Z g ( t ) } → 1 \mathbb{P}\{N_X(t)=Z_g(t)\}\to1 P { N X ( t ) = Z g ( t )} → 1 . Also N X ( t ) N_X(t) N X ( t ) is uniformly integrable: Z g ( t ) Z_g(t) Z g ( t ) has bounded second moment and their difference tends to zero in L 1 L^1 L 1 .
Fix an arbitrarily small failure allowance α > 0 \alpha> 0 α > 0 . By count tightness choose K K K so that Z g ( t ) > K Z_g(t) > K Z g ( t ) > K has limiting probability at most α \alpha α . Choose d > 0 d > 0 d > 0 so that a waist below d d d occurs with limiting probability at most α \alpha α . Finally choose η > 0 \eta> 0 η > 0 , with t − η t-\eta t − η a continuity point, so that a model energy in ( t − η , t ] (t-\eta,t] ( t − η , t ] occurs with limiting probability at most α \alpha α . This follows from mean convergence and ν c ( { t } ) = 0 \nu_c(\{t\}) = 0 ν c ({ t }) = 0 . All remaining waists lie in [ d , S t ] [d,S_t] [ d , S t ] , and their energies lie in [ β min , t − η ] [\beta_{\min},t-\eta] [ β m i n , t − η ] , where β min = min [ d , S t ] β > 0 \beta_{\min} = \min_{[d,S_t]} \beta> 0 β m i n = min [ d , S t ] β > 0 .
Put b = min ( β min , η ) b = \min(\beta_{\min},\eta) b = min ( β m i n , η ) and choose ζ < b / ( 2 max ( K , 1 ) ) \zeta< b/(2\sqrt{\max(K,1)}) ζ < b / ( 2 max ( K , 1 ) ) . The common support radius in (225) is now fixed before g → ∞ g \to\infty g → ∞ . Lemma 15.6 transplants these states and their annuli into disjoint neighborhoods with probability tending to one. On the count-matching event, k = Z g ( t ) = dim E k = Z_g(t) = \dim E k = Z g ( t ) = dim E , so Lemma 15.8(b) makes E E E residual; k = 0 k = 0 k = 0 is immediate. Let α ↓ 0 \alpha\downarrow0 α ↓ 0 . The cuspidal count in ( 0 , t ] (0,t] ( 0 , t ] vanishes in probability, and its domination by the uniformly integrable N X ( t ) N_X(t) N X ( t ) gives convergence to zero in mean.
For any fixed t 0 < 1 / 4 t_0 < 1/4 t 0 < 1/4 , choose a continuity threshold t ∈ ( t 0 , 1 / 4 ) t \in(t_0,1/4) t ∈ ( t 0 , 1/4 ) . Domination gives the same mean conclusion at t 0 t_0 t 0 , without assuming that ν c \nu_c ν c is atomless. Subtracting the cuspidal measure from the total one proves the joint process and mean assertions in (210) .
If (211) failed, there would be a sequence 1 ≤ n ≤ M g 1 \le n \le M\sqrt{g} 1 ≤ n ≤ M g along which the expected cuspidal count stays bounded away from zero. Pass to n / g → c ∈ [ 0 , M ] n/\sqrt{g} \to c \in[0,M] n / g → c ∈ [ 0 , M ] . For c > 0 c > 0 c > 0 the preceding argument is a contradiction. For c = 0 c = 0 c = 0 , the local count bound of Appendix § 10.10 , applied on the envelope n + 1 ≤ ( M + 1 ) g n + 1 \le(M + 1)\sqrt{g} n + 1 ≤ ( M + 1 ) g , gives E N X ( 1 / 4 − ε ) ≤ C M , ε ( n + 1 ) 2 / g → 0 \mathbb{E}N_X(1/4-\varepsilon) \le C_{M,\varepsilon}(n+1)^2/g \to0 E N X ( 1/4 − ε ) ≤ C M , ε ( n + 1 ) 2 / g → 0 (compare Huang’s total-count bound [101, Theorem 1.1]), again a contradiction. □
15.5. The three cusp regimes. The corollary below assembles Huang’s subcritical theorem, Theorem 15.1 and Corollary 14.1 into the gap phase diagram announced in the introduction. Its two endpoints are refined in the next sections: the approach to 1 / 4 1/4 1/4 from below in § 4.12 , and the rate of vanishing above the critical scale in § 4.13 .
Corollary 15.9 (Gap phase diagram). For n = n ( g ) ≥ 1 n = n(g) \ge1 n = n ( g ) ≥ 1 , as g → ∞ g \to\infty g → ∞ ,
G ( X ) → P 1 / 4 ⟺ n / g → 0 , G ( X ) → P 0 ⟺ n / g → ∞ . G(X) \xrightarrow{\mathbb{P}} 1/4 \Longleftrightarrow n/\sqrt{g} \to0,\qquad G(X) \xrightarrow{\mathbb{P}} 0 \Longleftrightarrow n/\sqrt{g} \to\infty. G ( X ) P 1/4 ⟺ n / g → 0 , G ( X ) P 0 ⟺ n / g → ∞.
At n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g} \to c \in(0,\infty) n / g → c ∈ ( 0 , ∞ ) the limit is the nondegenerate law (209) . In particular there is no deterministic limit in ( 0 , 1 / 4 ) (0,1/4) ( 0 , 1/4 ) , for any sequence of cusp counts.
Proof. The first equivalence is Huang’s result [101, Theorem 1.1 and Corollary 6.15]. The forward cusp condition in the second implies gap convergence by Corollary 14.1. Conversely, if n / g n/\sqrt{g} n / g does not tend to infinity, pass to a subsequence with a finite limit. A zero limit gives gap limit 1 / 4 1/4 1/4 by Huang; a positive limit gives G c > 0 G_c > 0 G c > 0 by Theorem 15.1. Both contradict convergence to zero. For a proposed deterministic interior limit, pass to a subsequence on which n / g n/\sqrt{g} n / g converges in [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] . The endpoints have already been treated. At a positive finite limit, ν c ( [ 0 , t ] ) \nu_c([0,t]) ν c ([ 0 , t ]) is finite and strictly positive for every t > 0 t > 0 t > 0 : positivity follows from β ( s ) → 0 \beta(s) \to0 β ( s ) → 0 . Thus the void probability lies strictly between zero and one, so the limiting law is not a point mass. This also proves the claimed nondegeneracy. □
Fluctuations near the quarter threshold
Huang’s theorem gives G ( X ) → 1 / 4 G(X) \to1/4 G ( X ) → 1/4 whenever n = o ( g ) n = o(\sqrt{g}) n = o ( g ) , with bounds in each fixed spectral window [101, Theorem 1.1]. This section determines the fluctuations of the spectrum near 1 / 4 1/4 1/4 in a range approaching the critical scale from below: the Fréchet law of the gap defect and the Poisson law of all deficits on the scale n 2 / g n^2/g n 2 / g (Theorem 16.1 ), the residual type of the corresponding eigenfunctions (Corollary 16.2 ) and the moments of the defect (Corollary 16.3 ). The inputs are the funnel model of § 4.7 , whose energies near 1 / 4 1/4 1/4 are read off from the critical exponent of Hecke groups [174], and Huang’s trace expansion, restated and made uniform in a moving window in Appendix § 10.15 , where the proofs are given. In this range the relevant pants waists grow and the model eigenvalues approach the continuous spectrum, so neither substituting a varying c c c into Theorem 15.1 nor convergence of the geometric counts alone gives the result below.
Write d ( w ) d(w) d ( w ) for the critical exponent of the Hecke group generated by z ↦ z + w z \mapsto z+w z ↦ z + w and z ↦ − 1 / z z \mapsto-1/z z ↦ − 1/ z , w > 2 w>2 w > 2 . Its index-two torsion-free subgroup generated by z ↦ z + w z \mapsto z+w z ↦ z + w and z ↦ z / ( 1 − w z ) z \mapsto z/(1-wz) z ↦ z / ( 1 − w z ) uniformizes a surface with two cusps and one funnel [174, Sections 1 and 3]; it is the funnel model F s F_s F s when w = 2 cosh ( s / 4 ) w=2\cosh(s/4) w = 2 cosh ( s /4 ) . Indeed, in terms of the generators A , B A,B A , B of Lemma 15.2, with a = 2 cosh ( s / 4 ) a=2\cosh(s/4) a = 2 cosh ( s /4 ) , the waist is represented by A B AB A B and tr ( A B ) = 2 − a 2 \operatorname{tr}(AB)=2-a^2 tr ( A B ) = 2 − a 2 , so ∣ tr ( A B ) ∣ = a 2 − 2 = 2 cosh ( s / 2 ) |\operatorname{tr}(AB)|=a^2-2=2\cosh(s/2) ∣ tr ( A B ) ∣ = a 2 − 2 = 2 cosh ( s /2 ) , the trace of a hyperbolic element of translation length s s s ; the translation z ↦ z + w z\mapsto z+w z ↦ z + w of the Hecke group is A A A with a = w a=w a = w . Hence, with κ 0 = ( 8 π 2 ) − 1 \kappa_0=(8\pi^2)^{-1} κ 0 = ( 8 π 2 ) − 1 ,
q ( w ) = ( d ( w ) − 1 / 2 ) 2 , q ( 2 cosh ( s / 4 ) ) = 1 / 4 − β ( s ) , h ∗ ( y ) = q ( 4 + y ) , y ≥ 0. (227) q(w)=(d(w)-1/2)^2,\qquad q(2\cosh(s/4))=1/4-\beta(s),\qquad h_*(y)=q\left(\sqrt{4+y}\right),\quad y\geq0. \tag*{(227)} q ( w ) = ( d ( w ) − 1/2 ) 2 , q ( 2 cosh ( s /4 )) = 1/4 − β ( s ) , h ∗ ( y ) = q ( 4 + y ) , y ≥ 0. (227)
Continuity at y = 0 y=0 y = 0 is understood. The function h ∗ h_* h ∗ decreases strictly from 1 / 4 1/4 1/4 to zero: d ( 2 ) = 1 d(2)=1 d ( 2 ) = 1 , d ( w ) → 1 / 2 d(w)\to1/2 d ( w ) → 1/2 as w → ∞ w\to\infty w → ∞ , and w ↦ d ( w ) w\mapsto d(w) w ↦ d ( w ) is strictly decreasing on [ 2 , ∞ ) [2,\infty) [ 2 , ∞ ) , see [174, equation (9) and the references therein]. Soares’s asymptotic d ( w ) = 1 / 2 + w − 1 + O ( w − 2 log w ) d(w)=1/2+w^{-1}+O(w^{-2}\log w) d ( w ) = 1/2 + w − 1 + O ( w − 2 log w ) [174, Theorem 1.9] gives
h ∗ ( y ) = y − 1 + O ( y − 3 / 2 log y ) , h ∗ ( y ) ≤ C 1 + y . (228) h_*(y)=y^{-1}+O(y^{-3/2}\log y),\qquad h_*(y)\leq\frac{C}{1+y}. \tag*{(228)} h ∗ ( y ) = y − 1 + O ( y − 3/2 log y ) , h ∗ ( y ) ≤ 1 + y C . (228)
In particular h ∗ − ( 1 + y ) − 1 h_*-(1+y)^{-1} h ∗ − ( 1 + y ) − 1 is integrable. Define H ( t ) = ∣ { y ≥ 0 : h ∗ ( y ) ≥ t } ∣ H(t)=|\{y\geq0:h_*(y)\geq t\}| H ( t ) = ∣ { y ≥ 0 : h ∗ ( y ) ≥ t } ∣ for 0 < t < 1 / 4 0<t<1/4 0 < t < 1/4 .
Theorem 16.1 (The moving spectral edge). Let X X X have normalized Weil–Petersson law on M g , n \mathcal{M}_{g,n} M g , n , and suppose
c = n g ⟶ 0 , c log g ⟶ ∞ . (229) c=\frac{n}{\sqrt{g}}\longrightarrow0,\qquad c\log g\longrightarrow\infty. \tag*{(229)} c = g n ⟶ 0 , c log g ⟶ ∞. (229) Counting every positive discrete eigenvalue with multiplicity, put
D X = ∑ 0 < λ j ( X ) < 1 / 4 δ ( 1 / 4 − λ j ( X ) ) / c 2 . D_X=\sum_{0<\lambda_j(X)<1/4}\delta_{(1/4-\lambda_j(X))/c^2}. D X = 0 < λ j ( X ) < 1/4 ∑ δ ( 1/4 − λ j ( X )) / c 2 . Then D X D_X D X converges in distribution, vaguely on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , to a Poisson random measure with intensity
ν e d g e ( d x ) = κ 0 x − 2 d x . (230) \nu_{\mathrm{edge}}(dx)=\kappa_0x^{-2}\,dx. \tag*{(230)} ν edge ( d x ) = κ 0 x − 2 d x . (230) The convergence also holds jointly for finitely many upper-tail counts [ a , ∞ ) [a,\infty) [ a , ∞ ) , a > 0 a>0 a > 0 , and their means converge to κ 0 / a \kappa_0/a κ 0 / a . For the gap defect D ( X ) = 1 / 4 − G ( X ) D(X)=1/4-G(X) D ( X ) = 1/4 − G ( X ) ,
D ( X ) c 2 ⟹ Z , P ( Z ≤ x ) = exp ( − κ 0 / x ) , x > 0. (231) \frac{D(X)}{c^2}\Longrightarrow Z,\qquad\mathbb{P}(Z\leq x)=\exp(-\kappa_0/x),\quad x>0. \tag*{(231)} c 2 D ( X ) ⟹ Z , P ( Z ≤ x ) = exp ( − κ 0 / x ) , x > 0. (231) For each fixed j j j , the first j j j positive eigenvalues below 1 / 4 1/4 1/4 exist and are simple with probability tending to one. Their ordered deficits, divided by c 2 c^2 c 2 , converge jointly to ( κ 0 / T 1 , … , κ 0 / T j ) (\kappa_0/T_1,\ldots,\kappa_0/T_j) ( κ 0 / T 1 , … , κ 0 / T j ) , where T i T_i T i are the arrival times of a rate-one Poisson process. Their normalized eigenfunctions are, up to sign, o P ( 1 ) o_{\mathbb{P}}(1) o P ( 1 ) in L 2 L^2 L 2 from the ground states of the funnel models F s F_s F s of the corresponding pairs of pants, cut off in the funnel and transplanted to embedded neighborhoods of these pants in X X X ; the neighborhoods are mutually disjoint and have area at most a deterministic g o ( 1 ) g^{o(1)} g o ( 1 ) .
The law is Fréchet with shape parameter one. Its geometric origin is transparent in the coordinate y = 2 ( cosh ( s / 2 ) − 1 ) y=2(\cosh(s/2)-1) y = 2 ( cosh ( s /2 ) − 1 ) : the model intensity is κ 0 c 2 d y \kappa_0c^2\,dy κ 0 c 2 d y , while the deficit is h ∗ ( y ) ∼ y − 1 h_*(y)\sim y^{-1} h ∗ ( y ) ∼ y − 1 . The substantial step is the transfer of this model computation to the entire actual spectral count. Appendix § 10.15 proves the moving mean, conditional isolation and exact count matching needed for this transfer, as well as the eigenfunction assertion. The condition c log g → ∞ c\log g\to\infty c log g → ∞ enters through localization: a model ground state with deficit of order c 2 c^2 c 2 decays on the funnel only over lengths of order 1 / c 1/c 1/ c , so it must be cut off at a radius R R R with c R → ∞ cR\to\infty c R → ∞ , while our estimates for isolation of the embedded funnel neighborhoods require R = o ( log g ) R=o(\log g) R = o ( log g ) . Condition (229) includes n = ⌊ g / ( log g ) α ⌋ n=\lfloor\sqrt{g}/(\log g)^\alpha\rfloor n = ⌊ g / ( log g ) α ⌋ for 0 < α < 1 0<\alpha<1 0 < α < 1 , but does not cover the whole subcritical range. The Poisson assertion does not describe deficits of order o ( c 2 ) o(c^2) o ( c 2 ) . It does compute, for instance, the probability that fewer than j j j modes fall below a prescribed moving energy: for each fixed a > 0 a>0 a > 0 and positive integer j j j ,
P { D X ( [ a , ∞ ) ) < j } ⟶ e − κ 0 / a ∑ r = 0 j − 1 ( κ 0 / a ) r r ! . \mathbb{P}\{D_X([a,\infty))<j\}\longrightarrow e^{-\kappa_0/a}\sum_{r=0}^{j-1}\frac{(\kappa_0/a)^r}{r!}. P { D X ([ a , ∞ )) < j } ⟶ e − κ 0 / a r = 0 ∑ j − 1 r ! ( κ 0 / a ) r .
Corollary 16.2 (Spectral type at the moving edge). Under the stronger condition
c ⟶ 0 , log ( 1 / c ) c log g ⟶ 0 , (232) c\longrightarrow0,\qquad\frac{\log(1/c)}{c\log g}\longrightarrow0, \tag*{(232)} c ⟶ 0 , c log g log ( 1/ c ) ⟶ 0 , (232) the expected cuspidal count in ( 0 , 1 / 4 − a c 2 ] (0,1/4-ac^2] ( 0 , 1/4 − a c 2 ] tends to zero for every fixed a > 0 a>0 a > 0 . The residual and cuspidal deficit processes converge jointly to ( PPP ( ν e d g e ) , 0 ) (\operatorname{PPP}(\nu_{\mathrm{edge}}),0) ( PPP ( ν edge ) , 0 ) , including their upper-tail means. In particular each fixed initial list of small eigenvalues is residual with probability tending to one. If G c u s p G_{\mathrm{cusp}} G cusp is the smallest positive cuspidal eigenvalue below 1 / 4 1/4 1/4 , or 1 / 4 1/4 1/4 if none exists, then ( 1 / 4 − G c u s p ) / c 2 → 0 (1/4-G_{\mathrm{cusp}})/c^2\to0 ( 1/4 − G cusp ) / c 2 → 0 in probability.
The stronger condition controls the error relative to a cusp observation that itself tends to zero; without this relative control, localization does not distinguish residual eigenfunctions from cusp forms. The proof is in Subsection K.4.
Corollary 16.3 (Rare gaps and moments). For every sequence n → ∞ n\to\infty n → ∞ , n = o ( g ) n=o(\sqrt{g}) n = o ( g ) and every fixed 0 < t < 1 / 4 0<t<1/4 0 < t < 1/4 , let N X p h y s ( t ) N_X^{\mathrm{phys}}(t) N X phys ( t ) count the positive eigenvalues at most 1 / 4 − t 1/4-t 1/4 − t . Then
P { D ( X ) ≥ t } ∼ E N X p h y s ( t ) ∼ κ 0 n 2 g H ( t ) . (233) \mathbb{P}\{D(X)\geq t\}\sim\mathbb{E}N_X^{\mathrm{phys}}(t)\sim\kappa_0\frac{n^2}{g}H(t). \tag*{(233)} P { D ( X ) ≥ t } ∼ E N X phys ( t ) ∼ κ 0 g n 2 H ( t ) . (233) Conditioned on this event, N X p h y s ( t ) = 1 N_X^{\mathrm{phys}}(t)=1 N X phys ( t ) = 1 with probability tending to one. Under (16.3) , for each fixed p > 0 p>0 p > 0 the moments of the actual gap defect satisfy
E D p c 2 p ⟶ κ 0 p Γ ( 1 − p ) , 0 < p < 1 , (234) \frac{\mathbb{E}D^p}{c^{2p}}\longrightarrow\kappa_0^p\Gamma(1-p),\qquad0<p<1, \tag*{(234)} c 2 p E D p ⟶ κ 0 p Γ ( 1 − p ) , 0 < p < 1 , (234) E D = c 2 4 π 2 log ( 1 / c ) + O ( c 2 ) , (235) \mathbb{E}D=\frac{c^2}{4\pi^2}\log(1/c)+O(c^2), \tag*{(235)} E D = 4 π 2 c 2 log ( 1/ c ) + O ( c 2 ) , (235) E D p c 2 ⟶ K p : = κ 0 ∫ 0 ∞ h ∗ ( y ) p d y , p > 1. (236) \frac{\mathbb{E}D^p}{c^2}\longrightarrow K_p:=\kappa_0\int_0^\infty h_*(y)^p\,dy,\qquad p>1. \tag*{(236)} c 2 E D p ⟶ K p := κ 0 ∫ 0 ∞ h ∗ ( y ) p d y , p > 1. (236) Consequently Var ( D ) ∼ K 2 c 2 \operatorname{Var}(D)\sim K_2c^2 Var ( D ) ∼ K 2 c 2 and D / E D → 0 D/\mathbb{E}D\to0 D / E D → 0 in probability.
Thus the typical defect has order c 2 c^2 c 2 , the mean carries an additional logarithm, and higher moments are governed by rare surfaces. These conclusions require expectation-level estimates for discarded geometries, since the weak limit in (231) has infinite mean. Subsection K.5 proves the moment statements directly, rather than exchanging a limit and an expectation.
For comparison, Hide–Macera–Thomas discuss a Tracy–Widom prediction for suitable random hyperbolic surface models after their Problem 1.3 [96, Introduction, Motivation]; their main spectral theorem concerns closed Weil–Petersson surfaces. The cusp-driven law above has a different mechanism and parameter range. It resolves neither that prediction nor the existence of closed surfaces of arbitrarily large genus with gap at least 1 / 4 1/4 1/4 .
This completes the description of the lowest eigenvalues at and below the critical scale. The remaining sections of this part treat n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ , where the gap tends to zero (Corollary 14.1) and the questions become the microscopic law of the lowest eigenvalues (§ 4.13 and § 4.14 ) and the density of eigenvalues in fixed windows (§ 4.19 ).
Supercritical spectral laws
When n / g → ∞ n/\sqrt{g} \to\infty n / g → ∞ , the short separating curves of Theorem 14.2 produce eigenvalues tending to zero, and Corollary 14.4 already gives their order. This section identifies the law of the lowest eigenvalues when n / g n/g n / g stays bounded (Theorem 17.1): on the explicit scale A g A_g A g they form a Poisson process, all planar cusp types contributing with weights given by the squares of their areas. The work is to exclude every low mode not produced by a short curve, including modes supported on a large component separated by several curves; this is done by comparing the low spectrum with a weighted star whose rates are collar conductances divided by enclosed areas. The inputs are the type process of Theorem 14.2 , the all-marking volume envelope ((D.35)) of Part 1, and two appendices: the cut lemmas of Appendix § 10.20 and the surface–graph comparison of Appendix M. The proof itself is given in this section.
Retain q g q_g q g and w k w_k w k from ((182) ) and ((184) ). Let x g x_g x g solve ( 2 g − 1 ) Ψ ( x g ) = n − 1 (2g-1)\Psi(x_g)=n-1 ( 2 g − 1 ) Ψ ( x g ) = n − 1 , and put
M 2 ( x ) = ∑ k ≥ 2 ( k − 1 ) 2 w k ( x ) , A g = 2 π 2 q g M 2 ( x g ) . (237) M_2(x)=\sum_{k\geq2}(k-1)^2w_k(x), \qquad A_g=2\pi^2\sqrt{q_gM_2(x_g)}. \tag*{(237)} M 2 ( x ) = k ≥ 2 ∑ ( k − 1 ) 2 w k ( x ) , A g = 2 π 2 q g M 2 ( x g ) . (237)
The function M 2 M_2 M 2 is finite and continuous on [ 0 , x c ) [0,x_c) [ 0 , x c ) , with M 2 ( 0 ) = 1 M_2(0)=1 M 2 ( 0 ) = 1 . Indeed, writing F = Φ 0 ′ F=\Phi'_0 F = Φ 0 ′ and D = x ∂ x D=x\partial_x D = x ∂ x , it equals 2 ( D − 1 ) 2 F / x 2 = 2 / a ( Y ( x ) ) − 2 Y ( x ) / x + W ( x ) 2(D-1)^2F/x^2=2/a(Y(x))-2Y(x)/x+W(x) 2 ( D − 1 ) 2 F / x 2 = 2/ a ( Y ( x )) − 2 Y ( x ) / x + W ( x ) , extended continuously to zero. Here a ( Y ) = X ′ ( Y ) = J 0 ( 2 Y ) a(Y)=X'(Y)=J_0(2\sqrt{Y}) a ( Y ) = X ′ ( Y ) = J 0 ( 2 Y ) is the kernel coefficient of § 3.1 , and Y ( x ) Y(x) Y ( x ) increases from 0 0 0 to the first zero y 0 y_0 y 0 of a a a as x x x increases from 0 0 0 to x c x_c x c . Hence a ( Y ( x ) ) → 0 a(Y(x))\to0 a ( Y ( x )) → 0 and M 2 ( x ) → ∞ M_2(x)\to\infty M 2 ( x ) → ∞ as x ↑ x c x\uparrow x_c x ↑ x c . The hypothesis sup g n / g < ∞ \sup_g n/g<\infty sup g n / g < ∞ in the theorem below keeps x g x_g x g in a compact subinterval of [ 0 , x c ) [0,x_c) [ 0 , x c ) , on which M 2 M_2 M 2 is bounded and the cusp weights w k w_k w k have a uniform exponential tail; this compactness is used throughout the proof.
Theorem 17.1 (Supercritical spectral process). Suppose g → ∞ g\to\infty g → ∞ , n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ and sup g n / g < ∞ \sup_g n/g<\infty sup g n / g < ∞ . For X X X with law P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP ,
∑ 0 < λ j ( X ) < 1 / 4 δ A g λ j ( X ) ⟹ PPP ( 2 t d t ) on [ 0 , ∞ ) (238) \sum_{0<\lambda_j(X)<1/4}\delta_{A_g\lambda_j(X)} \Longrightarrow\operatorname{PPP}(2t\,dt)\ \text{on }[0,\infty) \tag*{(238)} 0 < λ j ( X ) < 1/4 ∑ δ A g λ j ( X ) ⟹ PPP ( 2 t d t ) on [ 0 , ∞ ) (238) in distribution for the vague topology, counting multiplicities. In particular, for every $t\geq 0,
P { A g G ( X ) > t } ⟶ e − t 2 , G ( X ) h ( X ) ⟶ 1 π in probability . (239) \mathbb{P}\{A_gG(X)>t\}\longrightarrow e^{-t^2}, \qquad\frac{G(X)}{h(X)}\longrightarrow\frac{1}{\pi}\quad\text{in probability}. \tag*{(239)} P { A g G ( X ) > t } ⟶ e − t 2 , h ( X ) G ( X ) ⟶ π 1 in probability . (239) Here G G G is the spectral gap, with the convention of § 4.7 . If g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g , then A g ∼ π n / ( 2 2 g ) A_g\sim\pi n/(2\sqrt{2}g) A g ∼ π n / ( 2 2 g ) .
Marked-geodesic and variational methods of this kind have precedents in Hide–Thomas [99] and Gatignol–Thomas [71]. The additional conclusion here is a two-sided spectral comparison on a shrinking energy scale. Two-sided estimates of small eigenvalues in terms of short separating geodesics go back to Schoen, Wolpert and Yau [169] for compact surfaces; see also Dodziuk, Pignataro, Randol and Sullivan [59], and Wu and Xue [190] for lower bounds uniform in the genus. Burger’s surface–graph comparison [38], Theorems 1.1–1.2 concerns degeneration with constants depending on the topological signature; we instead control the pieces, traces and intermediate branches uniformly on events of probability tending to one. The all-marking volume envelope ((D.35)) is essential for excluding competing cuts.
The convergence ((238) ) is asserted only in distribution. Unlike Theorem 16.1 , whose statements about means rest on a positive trace estimate (Lemma K.1 ), no positive trace input is available in the present range, so no convergence of spectral first moments is claimed. Theorem 18.3 extends the law to a specified range with n / g → ∞ n/g\to\infty n / g → ∞ , at the cost of endpoint estimates and a more general contraction argument.
Proof of Theorem 17.1. Fix M M M with n ≤ M g n \le Mg n ≤ M g , and write P = 2 π 2 P = 2\pi^2 P = 2 π 2 and Q = n x g / ( 4 P ) ∼ q g Q = nx_g/(4P) \sim q_g Q = n x g / ( 4 P ) ∼ q g . All constants may depend on M M M . The proof has three steps: obtain uniform gaps on cut pieces, compare the low spectrum with a weighted star, and transfer the marked-curve law.
The cut pieces. Set
δ = Q − 1 / 10 , b = Q − 1 / 20 , K = C M log ( Q + 2 ) , R = Q 1 / 20 , τ = R / Q . (240) \delta= Q^{-1/10}, \qquad b = Q^{-1/20}, \qquad K = C_M \log(Q+2), \qquad R = Q^{1/20}, \qquad\tau= R/\sqrt{Q}. \tag*{(240)} δ = Q − 1/10 , b = Q − 1/20 , K = C M log ( Q + 2 ) , R = Q 1/20 , τ = R / Q . (240)
By Lemma L.2, with probability tending to one, all curves of length at most δ \delta δ are planar separating. Cutting along all of them gives a tree with a genus-g g g root. Each root branch has at most K K K planar vertices and area at most 2 π K 2\pi K 2 π K ; every geodesic piece has Neumann gap at least
γ = c M δ 2 / log 2 ( Q + 2 ) . (241) \gamma= c_M\delta^2/\log^2(Q+2). \tag*{(241)} γ = c M δ 2 / log 2 ( Q + 2 ) . (241)
Trim the half-collars at circumference b b b . Write m v m_v m v for the remaining piece areas, W e W_e W e for their total mass below an edge, and κ e \kappa_e κ e for the full collar conductance, as in Lemma M.2 . Retain edges for which d e = κ e / W e ≤ τ d_e=\kappa_e/W_e\le\tau d e = κ e / W e ≤ τ . If an edge encloses k k k cusps, then W e ≤ 2 π ( k − 1 ) W_e\le2\pi(k-1) W e ≤ 2 π ( k − 1 ) and κ e ≥ ℓ e / π \kappa_e\ge\ell_e/\pi κ e ≥ ℓ e / π , so a retained edge satisfies Q ℓ e / ( k − 1 ) ≤ P R \sqrt{Q}\ell_e/(k-1)\le PR Q ℓ e / ( k − 1 ) ≤ P R . The isolation assertion of Lemma L.2, applied with P R PR P R , makes each retained edge an isolated root leaf. Other branches need not be unnested.
Spectral comparison. Contract all nonretained vertices into the root and use their total trimmed mass as its vertex mass. Let μ j \mu_j μ j be the positive eigenvalues of this weighted star. For fixed t t t , Lemma M.2 applies uniformly for 0 ≤ Λ ≤ t / Q 0\le\Lambda\le t/\sqrt{Q} 0 ≤ Λ ≤ t / Q . Its trace constant satisfies T ≤ C M K / ( b γ ) = O ( Q 1 / 4 log 3 ( Q + 2 ) ) T\le C_MK/(b\gamma)=O(Q^{1/4}\log^3(Q+2)) T ≤ C M K / ( bγ ) = O ( Q 1/4 log 3 ( Q + 2 )) . Thus
Λ γ = O t ( Q − 3 / 10 log 2 ( Q + 2 ) ) , K Λ τ = O t ( Q − 1 / 20 log ( Q + 2 ) ) , \frac{\Lambda}{\gamma}=O_t(Q^{-3/10}\log^2(Q+2)), \qquad\frac{K\Lambda}{\tau}=O_t(Q^{-1/20}\log(Q+2)), γ Λ = O t ( Q − 3/10 log 2 ( Q + 2 )) , τ K Λ = O t ( Q − 1/20 log ( Q + 2 )) ,
and K T Λ = O t ( Q − 1 / 4 log 4 ( Q + 2 ) ) KT\Lambda=O_t(Q^{-1/4}\log^4(Q+2)) K T Λ = O t ( Q − 1/4 log 4 ( Q + 2 )) , while τ T = O ( Q − 1 / 5 log 3 ( Q + 2 ) ) \tau T=O(Q^{-1/5}\log^3(Q+2)) τ T = O ( Q − 1/5 log 3 ( Q + 2 )) . Every error tends to zero. Consequently, with c g , t → 1 c_{g,t}\to1 c g , t → 1 ,
N + ( Λ ) ≤ N X + ( Λ ) ≤ N + ( c g , t Λ ) ( 0 ≤ Λ ≤ t / Q ) . (242) N^+(\Lambda)\le N_X^+(\Lambda)\le N^+(c_{g,t}\Lambda)\qquad(0\le\Lambda\le t/\sqrt{Q}). \tag*{(242)} N + ( Λ ) ≤ N X + ( Λ ) ≤ N + ( c g , t Λ ) ( 0 ≤ Λ ≤ t / Q ) . (242)
Counts are inclusive at the upper endpoint and exclude zero.
For a retained leaf with k k k cusps, the area and conductance estimates give
d e = ( 1 + O ( b + δ / b ) ) ℓ e P ( k − 1 ) . (243) d_e=(1+O(b+\delta/b))\frac{\ell_e}{P(k-1)}. \tag*{(243)} d e = ( 1 + O ( b + δ / b )) P ( k − 1 ) ℓ e . (243)
The one-waist estimate (L.2) implies
E ∑ retained caps area ( cap ) ≤ C Q τ 2 ∑ k ≥ 2 ( k − 1 ) 3 w k ( x g ) e P τ ( k − 1 ) = O ( R 2 ) . \mathbb{E}\sum_{\text{retained caps}}\operatorname{area}(\operatorname{cap})
\le CQ\tau^2\sum_{k\ge2}(k-1)^3w_k(x_g)e^{P\tau(k-1)}
=O(R^2). E retained caps ∑ area ( cap ) ≤ C Q τ 2 k ≥ 2 ∑ ( k − 1 ) 3 w k ( x g ) e P τ ( k − 1 ) = O ( R 2 ) .
It is enough to bound by all eligible waists; nesting only overcounts. Since Q = O ( g ) Q=O(g) Q = O ( g ) , R 2 = o ( g ) R^2=o(g) R 2 = o ( g ) , and the full trimmed mass is ( 1 + O ( b ) ) area ( X ) (1+O(b))\operatorname{area}(X) ( 1 + O ( b )) area ( X ) , the retained leaf mass fraction ρ g \rho_g ρ g tends to zero in probability. Ordering the retained rates,
d j ≤ μ j ≤ d j / ( 1 − ρ g ) . (244) d_j\le\mu_j\le d_j/(1-\rho_g). \tag*{(244)} d j ≤ μ j ≤ d j / ( 1 − ρ g ) . (244)
To see this, quotient the star by constants and use leaf–root differences z i z_i z i . Its mass is ∑ m i z i 2 − ( ∑ m i z i ) 2 / ∑ v m v \sum m_i z_i^2-(\sum m_i z_i)^2/\sum_v m_v ∑ m i z i 2 − ( ∑ m i z i ) 2 / ∑ v m v , between ( 1 − ρ g ) ∑ m i z i 2 (1-\rho_g)\sum m_i z_i^2 ( 1 − ρ g ) ∑ m i z i 2 and ∑ m i z i 2 \sum m_i z_i^2 ∑ m i z i 2 ; its energy is ∑ κ i z i 2 \sum\kappa_i z_i^2 ∑ κ i z i 2 . Min–max gives (17.8).
The limiting counts. First take a subsequence with n / g → λ < ∞ n/g\to\lambda<\infty n / g → λ < ∞ . For any fixed set of cusp types, Theorem 14.2 maps under ( s , k ) ↦ s / [ P ( k − 1 ) ] (s,k)\mapsto s/[P(k-1)] ( s , k ) ↦ s / [ P ( k − 1 )] to a Poisson process. The expected contribution of types k > K 0 k>K_0 k > K 0 on a bounded score interval is at most
C t ∑ k > K 0 ( k − 1 ) 2 w k ( x g e C t / Q ) . C_t\sum_{k>K_0}(k-1)^2w_k\left(x_ge^{C_t/\sqrt{Q}}\right). C t k > K 0 ∑ ( k − 1 ) 2 w k ( x g e C t / Q ) .
by (L.2). This tends uniformly to zero as K 0 → ∞ K_0\to\infty K 0 → ∞ , since the arguments stay below a fixed x 2 < x c x_2<x_c x 2 < x c . The estimate includes waists longer than δ \delta δ . For a fixed type and bounded score the waist is eventually shorter than δ \delta δ , and its rate is retained because R → ∞ R \to\infty R → ∞ . Thus the limiting rate intensity is 2 P 2 M 2 ( x λ ) t d t 2P^2M_2(x_\lambda)t\,dt 2 P 2 M 2 ( x λ ) t d t in the Q \sqrt{Q} Q scale.
Equations (242) –(244) transfer joint counts at any finite collection of thresholds. The limiting intensity is atomless, so counts in intervals whose endpoints differ by factors tending to one vanish in probability. Geometric first-moment bounds and the comparison on events of probability tending to one give spectral count tightness in probability on compact intervals. This proves point-process convergence on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , including control at zero. Continuity of M 2 M_2 M 2 and compactness of the bounded density interval remove the subsequence assumption after the normalization (237) .
The first-point law gives the gap assertion, since the probability of having no positive subthreshold eigenvalue tends to zero by first taking g → ∞ g \to\infty g → ∞ and then t → ∞ t \to\infty t → ∞ . For the ratio assertion, the uniformity in (242) permits evaluation at Λ = λ j ( X ) \Lambda=\lambda_j(X) Λ = λ j ( X ) whenever μ j ≤ t / Q \mu_j \le t/\sqrt{Q} μ j ≤ t / Q ; hence μ j / c g , t ≤ λ j ( X ) ≤ μ j \mu_j/c_{g,t} \le\lambda_j(X) \le\mu_j μ j / c g , t ≤ λ j ( X ) ≤ μ j . The first rate is tight and bounded away from zero in probability in this scale. Lemma L.2 identifies the geodesic Cheeger minimum with the same minimum of ℓ / [ 2 π ( k − 1 ) ] \ell/[2\pi(k-1)] ℓ / [ 2 π ( k − 1 )] . Here the geodesic Cheeger constant H H H of [172] is an infimum over separating geodesic multicurves that cut the surface into two connected pieces, which are the cuts treated by Lemma L.1; allowing disconnected sides does not lower this infimum, by the graph-averaging reduction in the proof of Lemma M.1 (Appendix M), which selects, among the same geodesics, a cut with both sides connected and no larger ratio. Finally H / ( 1 + H ) ≤ h ≤ H H/(1+H) \le h \le H H / ( 1 + H ) ≤ h ≤ H for the geodesic Cheeger constant H H H [172, Proposition 13], and H → 0 H \to0 H → 0 in probability. This proves the coupled ratio G / h → 1 / π G/h \to1/\pi G / h → 1/ π . The stated sparse-density constant follows from q g ∼ n 2 / ( 32 π 2 g ) q_g \sim n^2/(32\pi^2g) q g ∼ n 2 / ( 32 π 2 g ) and M 2 ( x g ) → 1 M_2(x_g) \to1 M 2 ( x g ) → 1 . □
The scale A g A_g A g and the star comparison are carried into Section 18, where the law is extended to a range with n / g → ∞ n/g \to\infty n / g → ∞ and shown to fail at very large density. The same regime g ≪ n ≪ g \sqrt{g} \ll n \ll g g ≪ n ≪ g , observed in a fixed-energy window rather than on the scale A g A_g A g , is treated in Section 19.
The low spectrum at growing cusp density
This section follows the microscopic law of Theorem 17.1 into the dense regime n / g → ∞ n/g \to\infty n / g → ∞ . The law persists in a range specified in Theorem 18.3, which contains n ≍ g α n \asymp g^\alpha n ≍ g α for every 1 < α < 11 / 9 1<\alpha<11/9 1 < α < 11/9 , but not at all densities: once the planar components themselves grow, their internal geometry can produce eigenvalues smaller than the conductance of their attaching collars divided by their areas, and Theorem 18.1 proves that this collapse occurs whenever n ≫ g 3 / ( log g ) 4 n \gg g^3/(\log g)^4 n ≫ g 3 / ( log g ) 4 , with no upper restriction on the cusp count. Independently of either range, Corollary 18.4 shows that the lowest eigenspaces are residual throughout the dense regime. The inputs are the large-density endpoint expansions of Part 1, which give the growing-cap law (Lemma 18.2, Appendix P), the star comparison of Section 17 in the more general form of Appendix Q, and, for the collapse, the low-energy cusp cutoffs of Appendices N and O. Below g 3 g^3 g 3 , the geometric cap process still has its expected Poisson limit.
Theorem 18.1 (Dense spectral collapse). Let g → ∞ g \to\infty g → ∞ and let n = n ( g ) n=n(g) n = n ( g ) be any integer sequence satisfying
g 3 ( log g ) 4 ≪ n . (245) \frac{g^3}{(\log g)^4} \ll n. \tag*{(245)} ( log g ) 4 g 3 ≪ n . (245) For X X X with law P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP ,
A g λ r + ( X ) → 0 in probability for every fixed r ≥ 1. (246) A_g\lambda_r^+(X) \to0 \quad\text{in probability for every fixed } r \ge1. \tag*{(246)} A g λ r + ( X ) → 0 in probability for every fixed r ≥ 1. (246) Here A g A_g A g is defined by (237) , and λ r + \lambda_r^+ λ r + is the r r r th positive eigenvalue below 1 / 4 1/4 1/4 , set to 1 / 4 1/4 1/4 when fewer than r r r such eigenvalues exist. Consequently, the spectral counting measures in (238) are not tight on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) along any sequence satisfying (245) .
This is a shrinking-scale conclusion. Hide–Thomas [100], Theorem 1.3 prove a deterministic linear lower bound on the number of eigenvalues below $1/4$ whenever g < a n g<an g < an , for each fixed $a>0. The classical genus bound of Grigor’yan–Netrusov–Yau [76, Corollary 5.6] already gives (246) when n / g 3 → ∞ n/g^{3} \to\infty n / g 3 → ∞ ; the additional content is collapse below and through the scale g 3 g^{3} g 3 , where that genus bound alone is insufficient. The small-eigenvalue estimates of Gatignol–Thomas [71, Theorem 1.10] apply when n / g → 0 n/g \to0 n / g → 0 .
Theorem 18.1 neither gives a replacement law for all sequences with n / g → ∞ n/g \to\infty n / g → ∞ nor locates a sharp transition. In particular, the collapse argument does not address a fixed power n = g α n = g^{\alpha} n = g α with 1 < α < 3 1 < \alpha< 3 1 < α < 3 ; Theorem 18.3 gives the matching Poisson law for 1 < α < 11 / 9 1 < \alpha< 11/9 1 < α < 11/9 . The logarithmic exponent in (245) is a sufficient condition, not a claimed transition exponent. Finally, non-tightness on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) does not rule out a vague limit on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , since points may escape through zero in the latter space.
The proof places low-energy cutoffs around a typical cusp of a large WP sphere. Appendix § 10.22 constructs them quantitatively, by conformal regularization and the finite planar resistance estimate of Gurel-Gurevich–Nachmias [85, Lemma 3.6], and Appendix O transfers them to the actual bordered cap law using the positive tree encoding of Budd–Meeusen–Zonneveld [33]. We first establish the growing-cap law from the uniform volume estimates.
Growing planar caps
At growing density a large planar cap cut off by a short curve contains about ϵ − 1 ≍ n / g \epsilon^{-1} \asymp n/g ϵ − 1 ≍ n / g cusps, and the spectral scale is no longer set by the two-cusp pairs of pants of § 4.1 alone. This subsection records the parameters of this regime and the joint law of cap waists and cap sizes (Lemma 18.2 ), which is the geometric input of both the collapse and the persistence theorems. Put P = 2 π P = 2\pi P = 2 π , let ( 2 g − 1 ) Ψ ( x ) = n − 1 (2g - 1)\Psi(x) = n - 1 ( 2 g − 1 ) Ψ ( x ) = n − 1 , and set
ϵ = 1 − x x c , Q = n x 4 P , M a ( x ) = ∑ k ≥ 2 ( k − 1 ) a w k ( x ) , S = Q M 2 ( x ) , η = 1 n ϵ 3 / 2 . (247) \epsilon= 1 - \frac{x}{x_c}, \qquad Q = \frac{nx}{4P}, \qquad M_a(x) = \sum_{k \ge2}(k - 1)^a w_k(x), \qquad S = \sqrt{Q M_2(x)}, \qquad\eta= \frac{1}{n\epsilon^{3/2}}. \tag*{(247)} ϵ = 1 − x c x , Q = 4 P n x , M a ( x ) = k ≥ 2 ∑ ( k − 1 ) a w k ( x ) , S = Q M 2 ( x ) , η = n ϵ 3/2 1 . (247)
The endpoint expansions, which require only n / g → ∞ n/g \to\infty n / g → ∞ , give
ϵ ∼ 5 g 2 n , M 2 ( x ) ∼ 2 b x c ϵ − 1 / 2 , S ∼ ( y 0 / 5 ) 1 / 4 2 π n 3 / 4 g − 1 / 4 . (248) \epsilon\sim\frac{5g}{2n}, \qquad M_2(x) \sim\sqrt{\frac{2}{b x_c}}\epsilon^{-1/2}, \qquad S \sim\frac{(y_0/5)^{1/4}}{2\pi}n^{3/4}g^{-1/4}. \tag*{(248)} ϵ ∼ 2 n 5 g , M 2 ( x ) ∼ b x c 2 ϵ − 1/2 , S ∼ 2 π ( y 0 /5 ) 1/4 n 3/4 g − 1/4 . (248)
where b = − X ′ ′ ( y 0 ) = x c / y 0 b = -X^{\prime\prime}(y_0) = x_c/y_0 b = − X ′′ ( y 0 ) = x c / y 0 as in Lemma 3.1 (iii). Also Q / q g → 1 Q/q_g \to1 Q / q g → 1 , so A g ∼ P S A_g \sim PS A g ∼ P S , and M a ( x ) = O ( ϵ − ( a − 3 / 2 ) ) M_a(x) = O(\epsilon^{-(a-3/2)}) M a ( x ) = O ( ϵ − ( a − 3/2 ) ) for a ≥ 2 a \ge2 a ≥ 2 .
Lemma 18.2 (Cap process with growing cusp count). Suppose g ≪ n ≪ g 3 g \ll n \ll g^{3} g ≪ n ≪ g 3 . For each simple geodesic bounding a genus-zero cap with k k k cusps, assign score S ℓ / ( k − 1 ) S\ell/(k - 1) S ℓ / ( k − 1 ) and mark ϵ ( k − 1 ) \epsilon(k - 1) ϵ ( k − 1 ) . On bounded score intervals and compact mark intervals in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , their counting measure converges in distribution to a Poisson process with intensity
2 t d t v − 1 / 2 e − v π d v . (249) 2t\,dt\,\frac{v^{-1/2}e^{-v}}{\sqrt{\pi}}\,dv. \tag*{(249)} 2 t d t π v − 1/2 e − v d v . (249) The expected number of nested pairs in any such window is O ( η ) O(\eta) O ( η ) .
The mark records the size of a cap on its natural scale ϵ − 1 \epsilon^{-1} ϵ − 1 , and the score normalizes its waist length by the enclosed area. The nested-pair bound permits simultaneous use of disjoint caps in the spectral comparison. The factorial-moment proof is given in Appendix P.
Cutoffs and the spectral consequence
The new mechanism is that a large planar cap can support a low-energy function whose cost is controlled inside the cap rather than by the conductance of its attaching collar. The argument has three cases. Below g 3 g^{3} g 3 , the cap process above supplies the locations and sizes of the caps, and a cutoff at a typical cusp, transported to the bordered cap, gives the required trial function. There the decisive comparison is between the cutoff energy O ( 1 / log K ) O(1/\log K) O ( 1/ log K ) inside a large planar component and the scale ℓ ∗ \ell_* ℓ ∗ of its attaching waist. Here K = ϵ − 1 K = \epsilon^{-1} K = ϵ − 1 is the typical cusp count of a large cap and ℓ ∗ = K / S \ell_* = K/S ℓ ∗ = K / S is typical waist length, with ϵ \epsilon ϵ and S S S as in (247) . When n / g 3 → 0 n/g^{3} \to0 n / g 3 → 0 in the range of Theorem 18.1, one has ℓ ∗ log K → ∞ \ell_* \log K \to\infty ℓ ∗ log K → ∞ , so the cutoff energy is much smaller than the waist scale. At the cubic scale we count smaller caps to recover a diverging supply of disjoint supports; above it the classical genus bound applies. The complete argument, including the exceptional-cap estimates and the centering of the trial space, is in Appendix P.
This construction is variational. Determining the actual dense spectral scale requires matching lower estimates and control of the other low modes; neither follows from the construction.
Persistence of the Poisson spectrum
There is nevertheless an unbounded-density range in which collar conductances still determine all low eigenvalues. Write L = log ( n + 2 ) L=\log(n+2) L = log ( n + 2 ) and retain the exact parameters (247) .
Theorem 18.3 (Poisson spectrum at growing density). Suppose g → ∞ g\to\infty g → ∞ , n / g → ∞ n/g\to\infty n / g → ∞ and
L 10 ϵ 3 S ⟶ 0. (250) \frac{L^{10}}{\epsilon^3 S}\longrightarrow0. \tag*{(250)} ϵ 3 S L 10 ⟶ 0. (250) For X X X with law P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP ,
∑ 0 < λ j ( X ) < 1 / 4 δ P S λ j ( X ) ⟹ PPP ( 2 t d t ) on [ 0 , ∞ ) (251) \sum_{0<\lambda_j(X)<1/4}\delta_{PS\lambda_j(X)}\Longrightarrow\operatorname{PPP}(2t\,dt)\ \text{on }[0,\infty) \tag*{(251)} 0 < λ j ( X ) < 1/4 ∑ δ P S λ j ( X ) ⟹ PPP ( 2 t d t ) on [ 0 , ∞ ) (251) in distribution for the vague topology, with multiplicities and residual eigenvalues included. The normalization P S PS P S may be replaced by A g A_g A g . In particular, for every t ≥ 0 t\geq0 t ≥ 0 ,
P { A g G ( X ) > t } ⟶ e − t 2 , G ( X ) ℏ ( X ) ⟶ 1 π in probability . \mathbb{P}\{A_gG(X)>t\}\longrightarrow e^{-t^2},\qquad\frac{G(X)}{\hbar(X)}\longrightarrow\frac{1}{\pi}\quad\text{in probability}. P { A g G ( X ) > t } ⟶ e − t 2 , ℏ ( X ) G ( X ) ⟶ π 1 in probability . Every fixed initial list of rescaled positive eigenvalues converges jointly to the corresponding ordered points of this Poisson process.
By (248) , condition (250) is equivalent to n 9 / 4 log 10 ( n + 2 ) / g 11 / 4 → 0 n^{9/4}\log^{10}(n+2)/g^{11/4}\to0 n 9/4 log 10 ( n + 2 ) / g 11/4 → 0 ; in particular, it includes n ≍ g α n\asymp g^\alpha n ≍ g α for every fixed 1 < α < 11 / 9 1<\alpha<11/9 1 < α < 11/9 . The exponent 11 / 9 11/9 11/9 has a simple origin in the proof. The spectral comparison requires the low eigenvalues, of size 1 / S 1/S 1/ S by (248) , to be small relative to the Neumann gaps γ \gamma γ of the pieces cut out by short planar curves; these gaps are obtained from a Cheeger inequality and carry its square loss, γ = c u 2 \gamma=cu^2 γ = c u 2 with u = c δ / K u=c\delta/K u = cδ / K , where δ = 1 / L \delta=1/L δ = 1/ L is the cut length and K = C 1 L / ϵ K=C_1L/\epsilon K = C 1 L / ϵ is the maximal cusp number of a δ \delta δ -short cap, so that γ ≍ ϵ 2 L − 4 \gamma\asymp\epsilon^2L^{-4} γ ≍ ϵ 2 L − 4 . The square accounts for ϵ 2 \epsilon^2 ϵ 2 ; the attachment errors carry one further factor K K K , the area scale of a branch, which accounts for ϵ − 1 \epsilon^{-1} ϵ − 1 ; and S S S enters through Λ ≤ T / S \Lambda\leq T/S Λ ≤ T / S and τ = R / S \tau=R/S τ = R / S . The critical quantity is therefore ϵ 3 S ≍ g 11 / 4 n − 9 / 4 \epsilon^3S\asymp g^{11/4}n^{-9/4} ϵ 3 S ≍ g 11/4 n − 9/4 . A gap estimate for the root piece that is linear rather than quadratic in its Cheeger constant would improve the exponent, which we do not claim to be sharp. Neither this exponent nor the logarithmic power is a proposed phase boundary, and the theorem leaves unresolved a substantial interval below the collapse regime of Theorem 18.1 .
The proof, given in Appendix Q, combines the growing-cap law with a comparison on the whole low spectral subspace. Unlike in the bounded-density case, the retained waists need not be isolated leaves of the original cut tree. We therefore contract the intervening branches and control their joint attachment errors, including the loss incurred in passing from surface energy to piece means. This yields a continuum conclusion beyond the geometric point process, without invoking a fixed-topology degeneration estimate uniformly in the number of cusps.
Residual type at every dense growth rate
Identifying the spectral type of the lowest eigenspaces does not require the persistence condition (250) . The following result concerns these actual eigenspaces throughout the dense regime.
Corollary 18.4 (Residual microscopic spectrum). Suppose g → ∞ g\to\infty g → ∞ and n / g → ∞ n/g\to\infty n / g → ∞ . For every fixed T > 0 T>0 T > 0 ,
P { X has a cuspidal eigenfunction with 0 < λ ≤ T / A g } ⟶ 0. \mathbb{P}\{X\text{ has a cuspidal eigenfunction with }0<\lambda\leq T/A_g\}\longrightarrow0. P { X has a cuspidal eigenfunction with 0 < λ ≤ T / A g } ⟶ 0.
For each fixed j ≥ 1 j \ge1 j ≥ 1 , the first j j j positive eigenvalues, counted with multiplicity, exist below 1 / 4 1/4 1/4 and their full eigenspaces are residual with probability tending to one. If, in addition, (250) holds, the positive residual eigenvalue process rescaled by A g A_g A g converges in distribution to PPP(2 t d t 2t\,dt 2 t d t ) vaguely on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) .
Proof. Let η b \eta_b η b be the infimum of the Rayleigh quotient on the H 1 H^1 H 1 functions whose angular average vanishes throughout every fixed-circumference cusp tail; the precise form domain is given in Appendix R. Proposition R.1 of that appendix proves S η b → ∞ S\eta_b \to\infty S η b → ∞ in probability without an upper restriction on n / g n/g n / g ; here η b \eta_b η b is unrelated to the parameter η \eta η of (247) . Every cusp-form eigenfunction belongs to this form domain, hence its eigenvalue is at least η b \eta_b η b . Since A g ∼ P S A_g \sim PS A g ∼ P S , this proves the first assertion. The cuspidal/residual decomposition below 1 / 4 1/4 1/4 [17, Section 3.3] then makes every eigenspace in this window entirely residual, including in the presence of multiplicities.
We separately check that S λ j + = O P ( 1 ) S\lambda_j^+ = O_{\mathbb{P}}(1) S λ j + = O P ( 1 ) for each fixed j j j . If n / g 3 → 0 n/g^3 \to0 n / g 3 → 0 , fix a mark interval [ a 0 , a 1 ] ⊂ ( 0 , ∞ ) [a_0,a_1] \subset(0,\infty) [ a 0 , a 1 ] ⊂ ( 0 , ∞ ) in Lemma 18.2 . By choosing a large fixed score bound R R R , with probability arbitrarily close to one there are at least j + 1 j+1 j + 1 interior-disjoint caps in this window. Their waist lengths tend to zero; the collar lemma and the vanishing nested-pair bound ensure disjointness. On each cap take a function equal to one off the inward width-one waist half-collar, equal to inward distance on that half-collar, and zero outside. For a cap with k k k cusps and waist ℓ \ell ℓ , its energy is at most ℓ sinh 1 \ell\sinh1 ℓ sinh 1 and its squared norm is at least 2 π ( k − 1 ) − ℓ sinh 1 2\pi(k-1)-\ell\sinh1 2 π ( k − 1 ) − ℓ sinh 1 . Its Rayleigh quotient is therefore at most C ℓ / ( k − 1 ) ≤ C R / S C\ell/(k-1) \le CR/S C ℓ / ( k − 1 ) ≤ C R / S . Disjoint supports give the same bound on the entire ( j + 1 ) (j+1) ( j + 1 ) -dimensional span. Min–max, with the constant eigenvalue counted once, yields λ j + ≤ C R / S \lambda_j^+ \le CR/S λ j + ≤ C R / S .
If n / g 3 n/g^3 n / g 3 stays bounded below by a positive constant, then Theorem 18.1 gives S λ j + → 0 S\lambda_j^+ \to0 S λ j + → 0 in probability. Every subsequence has a further subsequence of one of these two types, which proves the asserted upper tightness for every dense sequence. First choose a large fixed T T T and then apply the first assertion to ( 0 , T / A g ] (0,T/A_g] ( 0 , T / A g ] . Tightness shows that the first j j j positive levels lie in this window with probability arbitrarily close to one, so their full eigenspaces are residual with probability tending to one. Under the additional persistence condition, the residual point-process limit now follows from Theorem 18.3 . □
This settles the spectral type of any fixed initial list of eigenvalues at all dense growth rates, and complements the fixed-genus residual-count question in [98, Section 1.1, after Remark 1.2]. It does not bound the total number of small cusp forms or settle the Otal–Rosas conjecture discussed in [17, Introduction]. The correct dense normalization and the joint law outside the persistence range remain open. The fixed-window counterpart of this regime is deterministic and has no such gap: Theorem 19.3 identifies the limiting spectral measure for every sequence with g = o ( n ) g=o(n) g = o ( n ) .
Spectral measures in fixed-energy windows
The shrinking-scale process of Theorem 17.1 describes only the lowest eigenvalues. This section studies the counting measure Ξ X \Xi_X Ξ X of the spectrum in ( 0 , 1 / 4 ) (0,1/4) ( 0 , 1/4 ) without rescaling energy and proves deterministic limits in the three regimes above the critical scale: ν ∗ \nu_* ν ∗ when g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g (Theorem 19.1 ), ν λ \nu_\lambda ν λ when n / g → λ ∈ ( 0 , ∞ ) n/g \to\lambda\in(0,\infty) n / g → λ ∈ ( 0 , ∞ ) (Theorem 19.2 ) and ν ∞ \nu_\infty ν ∞ when g = o ( n ) g=o(n) g = o ( n ) (Theorem 19.3 ), followed by the regularized full-energy measure (Theorem 19.4 ) and its determinant consequence (Corollary 19.5 ). The common method is the one announced in Subsection 1.5: concentration of orbital statistics, obtained from the volume theory through the integration formula, is transferred to the spectrum by Huang’s trace approximation [101]. The trace and orbital inputs are proved in Appendices S, T and U, and the spectral transfers in Appendix V.
When g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g , a fixed-energy window below 1 / 4 1/4 1/4 contains order n 2 / g n^2/g n 2 / g eigenvalues, and we first identify their deterministic density, including matching expected counts. Retain the funnel model F s F_s F s and its ground-state energy β ( s ) \beta(s) β ( s ) from § 4.7 , and set
μ ∗ ( d s ) = sinh ( s / 2 ) 8 π 2 d s , ν ∗ = β ∗ μ ∗ . (252) \mu_\ast(ds)=\frac{\sinh(s/2)}{8\pi^2}\,ds,\qquad\nu_\ast=\beta_\ast\mu_\ast. \tag*{(252)} μ ∗ ( d s ) = 8 π 2 sinh ( s /2 ) d s , ν ∗ = β ∗ μ ∗ . (252)
This measure is locally finite on [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) by Lemma 15.2, and has no atom at zero.
Theorem 19.1 (Fixed-energy spectral density). Suppose g → ∞ g\to\infty g → ∞ and g ≪ n ≪ g \sqrt{g}\ll n\ll g g ≪ n ≪ g , and put d g = n 2 / g d_g=n^2/g d g = n 2 / g . For every φ ∈ C c ( [ 0 , 1 / 4 ) ) \varphi\in C_c([0,1/4)) φ ∈ C c ([ 0 , 1/4 )) ,
E ∣ 1 d g ∑ 0 < λ j ( X ) < 1 / 4 φ ( λ j ( X ) ) − ∫ φ d ν ∗ ∣ ⟶ 0. (253) \mathbb{E}\left|\frac{1}{d_g}\sum_{0<\lambda_j(X)<1/4}\varphi(\lambda_j(X))-\int\varphi\,d\nu_\ast\right|\longrightarrow0. \tag*{(253)} E d g 1 0 < λ j ( X ) < 1/4 ∑ φ ( λ j ( X )) − ∫ φ d ν ∗ ⟶ 0. (253) In particular the normalized counting measures converge vaguely in probability to ν ∗ \nu_\ast ν ∗ , and their mean measures converge vaguely. Eigenvalues are counted with multiplicity, including residual eigenvalues.
Gatignol–Thomas [71, Theorem 1.10] prove a lower count of order n 2 / g n^2/g n 2 / g below each fixed positive threshold in this regime; Theorem 19.1 supplies the matching density. The upper bound uses Huang’s canonical trace coefficients and positive truncation method [101], together with a new complementary-volume tail, proved in Appendix S, which removes the quadratic cusp restriction on the trace remainder. Positive limiting n / g n/g n / g is treated separately in Theorem 19.2 ; neither result concerns moving windows at 1 / 4 1/4 1/4 .
Structure of the proof. The proof, given in Appendix V, has two halves. The upper half identifies the mean measure: after the cusp labels are summed at the saddle, the trace formula leaves only the term of the two-cusp pairs of pants, whose distribution is ν ∗ \nu_\ast ν ∗ by Proposition 15.5 . The lower half places disjoint funnel trial states on short two-cusp pants. A law of large numbers for their waists, together with the removal of overlapping configurations, shows that these states account for the whole mean. Matching the two halves gives convergence in L 1 L^1 L 1 . In the sublinear regime only two-cusp pairs of pants contribute; the next two subsections treat the densities at which all planar filling types enter.
Positive cusp density
When n / g → λ ∈ ( 0 , ∞ ) n/g\to\lambda\in(0,\infty) n / g → λ ∈ ( 0 , ∞ ) , planar subsurfaces with more than two cusps contribute at the same order as two-cusp pairs of pants, so the limiting low spectrum involves all planar filling types. Theorem 19.2 identifies the limit of g − 1 Ξ X g^{-1}\Xi_X g − 1 Ξ X as a measure ν λ \nu_\lambda ν λ characterized through the Weil–Petersson integrals of these types and the boundary profile of Proposition 9.4 ; this is the regime in which the density parameter Y λ Y_\lambda Y λ at an interior point, rather than at an endpoint of [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] , enters the spectrum. The orbital concentration is proved in Appendix T and the spectral transfer in Appendix V.
Theorem 19.2 (Low spectrum at positive density). There is a vaguely continuous family of Radon measures { ν λ : 0 < λ < ∞ } \{\nu_\lambda:0<\lambda<\infty\} { ν λ : 0 < λ < ∞ } on [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) such that, whenever g → ∞ g\to\infty g → ∞ and n / g → λ ∈ ( 0 , ∞ ) n/g\to\lambda\in(0,\infty) n / g → λ ∈ ( 0 , ∞ ) ,
E ∣ 1 g ∑ 0 < λ j ( X ) < 1 / 4 φ ( λ j ( X ) ) − ∫ φ d ν λ ∣ ⟶ 0 , φ ∈ C c ( [ 0 , 1 / 4 ) ) . (254) \mathbb{E}\left|\frac{1}{g}\sum_{0<\lambda_j(X)<1/4}\varphi(\lambda_j(X))-\int\varphi\,d\nu_\lambda\right|\longrightarrow0,\qquad\varphi\in C_c([0,1/4)). \tag*{(254)} E g 1 0 < λ j ( X ) < 1/4 ∑ φ ( λ j ( X )) − ∫ φ d ν λ ⟶ 0 , φ ∈ C c ([ 0 , 1/4 )) . (254) Eigenvalues are counted with multiplicity, including residual modes. The measure ν λ \nu_\lambda ν λ has no atom at zero. The planar orbital integrals ((2) ) and the limit ((V.8)) determine ν λ \nu_\lambda ν λ uniquely.
Thus the normalized low spectrum converges vaguely in probability, and its mean converges as well. No absolute continuity of ν λ \nu_\lambda ν λ is assumed. Hide–Thomas [98, Remark 1.8] ask about spectral measures in the large-cusp regime; the theorem treats their low-energy part when genus and cusp count grow proportionally. The endpoint g = o ( n ) g=o(n) g = o ( n ) , including fixed genus, is treated in Theorem 19.3 . Neither theorem treats the continuous spectrum above 1 / 4 1/4 1/4 .
The proof differs from the sublinear case in its order of limits. After division by g g g , one may let g g g tend to infinity with the polynomial degree fixed, and only then let the degree tend to infinity. No trace expansion uniform in a growing degree is needed.
Structure of the proof. Appendix T constructs a functional Ω λ \Omega_\lambda Ω λ from planar orbital integrals and proves
E ∣ g − 1 O X ( H ) − Ω λ ( H ) ∣ ⟶ 0 ( H ∈ C c ∞ ( R ) even ) . (255) \mathbb{E}\lvert g^{-1}O_X(H)-\Omega_\lambda(H)\rvert\longrightarrow0 \qquad(H\in C^\infty_{\mathrm{c}}(\mathbb{R})\ \text{even}). \tag*{(255)} E ∣ g − 1 O X ( H ) − Ω λ ( H )∣ ⟶ 0 ( H ∈ C c ∞ ( R ) even ) . (255)
The functional is continuous in λ \lambda λ on compact density intervals. Huang’s positive trace [101, Sections 5–6] then transfers this fixed-support assertion to the actual spectrum, as carried out in Appendix V. The functional Ω λ \Omega_\lambda Ω λ and the measures ν λ \nu_\lambda ν λ are carried forward: their λ → ∞ \lambda\to\infty λ → ∞ limits are identified in the next subsection (257) , and they reappear as the low-energy part of the regularized measure in Theorem 19.4 .
The cusp-dominated limit
The endpoint n / g → ∞ n/g\to\infty n / g → ∞ has a universal fixed-energy limit ν ∞ \nu_\infty ν ∞ , which agrees with the large-cusp limit at every fixed genus, including genus zero (Theorem 19.3 ). The input is the critical-weight endpoint of the volume theory, x = x c x=x_c x = x c and Y = y 0 Y=y_0 Y = y 0 , at which the planar orbital functional of the previous subsection has a limit; the orbital concentration is proved in Appendix U and the spectral transfer in Appendix V. The statement concerns energies held fixed below 1 / 4 1/4 1/4 ; the shrinking scale of the first few eigenvalues in the same regime is the unresolved question of § 4.14 .
Theorem 19.3 (Universal cusp-dominated low spectrum). There is a Radon measure ν ∞ \nu_\infty ν ∞ on [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) such that, whenever n → ∞ n\to\infty n → ∞ and g = g ( n ) ≥ 0 g=g(n)\ge0 g = g ( n ) ≥ 0 is any integer sequence with g = o ( n ) g=o(n) g = o ( n ) ,
E ∣ 1 n ∑ 0 < λ j ( X ) < 1 / 4 φ ( λ j ( X ) ) − ∫ φ d ν ∞ ∣ ⟶ 0 , φ ∈ C c ( [ 0 , 1 / 4 ) ) . (256) \mathbb{E}\left\lvert\frac{1}{n}\sum_{0<\lambda_j(X)<1/4}\varphi(\lambda_j(X))-\int\varphi\,d\nu_\infty\right\rvert\longrightarrow0,\qquad\varphi\in C_{\mathrm{c}}([0,1/4)). \tag*{(256)} E n 1 0 < λ j ( X ) < 1/4 ∑ φ ( λ j ( X )) − ∫ φ d ν ∞ ⟶ 0 , φ ∈ C c ([ 0 , 1/4 )) . (256) This includes multiplicities and residual eigenvalues. For the measures of Theorem 19.2 ,
λ − 1 ν λ ⟶ ν ∞ vaguely as λ → ∞ . (257) \lambda^{-1}\nu_\lambda\longrightarrow\nu_\infty\quad\text{vaguely as }\lambda\to\infty. \tag*{(257)} λ − 1 ν λ ⟶ ν ∞ vaguely as λ → ∞. (257) The measure is characterized by (V.8) with Ω λ \Omega_\lambda Ω λ replaced by Ω ∞ \Omega_\infty Ω ∞ defined below. It has no atom at zero.
Here Ω ∞ \Omega_\infty Ω ∞ is obtained by dividing the right side of (T.2) by λ \lambda λ and setting
x = x c , K x = K x c , b Y ( L ) = b y 0 ( L ) = I 0 ( j 0 , 1 L 2 π ) . x=x_c,\qquad K_x=K_{x_c},\qquad\mathbf{b}_Y(L)=\mathbf{b}_{y_0}(L)=I_0\left(\frac{j_{0,1}L}{2\pi}\right). x = x c , K x = K x c , b Y ( L ) = b y 0 ( L ) = I 0 ( 2 π j 0 , 1 L ) .
The series K x c K_{x_c} K x c converges, since its coefficients at critical weight decay as j − 5 / 2 j^{-5/2} j − 5/2 by the genus-zero case of Manin–Zograf [136, Theorem 6.1]; see also [98, Theorem 2.4]. The boundary bound and fixed-support fiber estimates therefore give
λ − 1 Ω λ ( H ) ⟶ Ω ∞ ( H ) ( H ∈ C c ∞ ( R ) even ) . (258) \lambda^{-1}\Omega_\lambda(H)\longrightarrow\Omega_\infty(H)\qquad(H\in C^\infty_{\mathrm{c}}(\mathbb{R})\ \text{even}). \tag*{(258)} λ − 1 Ω λ ( H ) ⟶ Ω ∞ ( H ) ( H ∈ C c ∞ ( R ) even ) . (258)
Hide–Thomas [98, Theorem 1.1] prove a linear lower count of small eigenvalues at fixed genus; Theorem 19.3 identifies the full limiting low spectral measure in that regime and, more generally, throughout g = o ( n ) g=o(n) g = o ( n ) . This treats the exceptional discrete part of the large-cusp spectral-measure question in [98, Remark 1.8]; the full regularized measure, including scattering, is treated in Subsection 19.3.
Structure of the proof. Appendix U proves
E ∣ n − 1 O X ( H ) − Ω ∞ ( H ) ∣ ⟶ 0 ( g = o ( n ) , n → ∞ ) . (259) \mathbb{E}\lvert n^{-1}O_X(H)-\Omega_\infty(H)\rvert\longrightarrow0\qquad(g=o(n),\ n\to\infty). \tag*{(259)} E ∣ n − 1 O X ( H ) − Ω ∞ ( H )∣ ⟶ 0 ( g = o ( n ) , n → ∞ ) . (259)
It treats diverging genus, fixed positive genus and genus zero separately and obtains the same functional in each case. The distinction is necessary: the positive-convolution local limit requires diverging genus, whereas on a sphere the leading complementary component is selected by a summable coefficient convolution. The transfer to the spectrum is that of Theorem 19.2 , with normalization n n n ; see Appendix V. The functional Ω ∞ \Omega_{\infty} Ω ∞ is the θ = 0 \theta=0 θ = 0 endpoint of the family used in the regularized law of the next subsection.
The full regularized spectral measure
The three preceding theorems concern the discrete spectrum below 1 / 4 1/4 1/4 . To include the continuous spectrum, which starts at 1 / 4 1/4 1/4 , we use the classical geometric regularization of the heat trace [112, 104], and prove that the resulting signed spectral measure, divided by n n n , has a deterministic full-energy limit whenever g / n → θ ∈ [ 0 , ∞ ) g/n\to\theta\in[0,\infty) g / n → θ ∈ [ 0 , ∞ ) (Theorem 19.4 ). For the trace formula, Eisenstein series and the scattering determinant on cofinite surfaces see [93, 107]. The regularized measure combines all discrete eigenvalues with the logarithmic derivative of the scattering determinant, and the parabolic corrections must be retained when the number of cusps is comparable to the area. The inputs are the orbital limits ((255) ) and ((259) ) of the two preceding subsections; the proof is in Appendix V.
In width-one cusp coordinates, let Φ X ( s ) \Phi_X(s) Φ X ( s ) be the scattering matrix and ϕ X ( s ) = det Φ X ( s ) \phi_X(s)=\det\Phi_X(s) ϕ X ( s ) = det Φ X ( s ) . For F ∈ C c ( [ 0 , ∞ ) ) F\in C_c([0,\infty)) F ∈ C c ([ 0 , ∞ )) , define
m X ( F ) = ∑ j F ( λ j ( X ) ) − 1 4 π ∫ R F ( 1 / 4 + r 2 ) ϕ X ′ ϕ X ( 1 / 2 + i r ) d r + n 2 π ∫ R F ( 1 / 4 + r 2 ) ( Re ψ ( 1 + i r ) + log 2 ) d r − n − tr Φ X ( 1 / 2 ) 4 F ( 1 / 4 ) . (260) m_X(F)=\sum_j F(\lambda_j(X))-\frac{1}{4\pi}\int_{\mathbb{R}}F(1/4+r^2)\frac{\phi_X'}{\phi_X}(1/2+ir)\,dr
+\frac{n}{2\pi}\int_{\mathbb{R}}F(1/4+r^2)\left(\operatorname{Re}\psi(1+ir)+\log2\right)\,dr-\frac{n-\operatorname{tr}\Phi_X(1/2)}{4}F(1/4). \tag*{(260)} m X ( F ) = j ∑ F ( λ j ( X )) − 4 π 1 ∫ R F ( 1/4 + r 2 ) ϕ X ϕ X ′ ( 1/2 + i r ) d r + 2 π n ∫ R F ( 1/4 + r 2 ) ( Re ψ ( 1 + i r ) + log 2 ) d r − 4 n − tr Φ X ( 1/2 ) F ( 1/4 ) . (260)
Here ψ = Γ ′ / Γ \psi=\Gamma'/\Gamma ψ = Γ ′ /Γ , the prime on ϕ X \phi_X ϕ X is the complex s s s derivative, and the sum includes every discrete L 2 L^2 L 2 eigenvalue, with multiplicity, including zero and embedded eigenvalues. The scattering matrix is regular and unitary on the critical line; its value at 1 / 2 1/2 1/2 is an involution. Formula (260) defines a signed locally finite measure, not a positive eigenvalue-counting measure, and fixes the regularization used throughout this subsection.
For Re z > 0 \operatorname{Re}z>0 Re z > 0 , set
G z ( u ) = e − z / 4 − u 2 / ( 4 z ) 4 π z , k H ( z ) = 1 4 π ∫ R e − z ( 1 / 4 + r 2 ) r tanh ( π r ) d r , (261) G_z(u)=\frac{e^{-z/4-u^2/(4z)}}{\sqrt{4\pi z}},\qquad k_{\mathrm{H}}(z)=\frac{1}{4\pi}\int_{\mathbb{R}}e^{-z(1/4+r^2)}r\tanh(\pi r)\,dr, \tag*{(261)} G z ( u ) = 4 π z e − z /4 − u 2 / ( 4 z ) , k H ( z ) = 4 π 1 ∫ R e − z ( 1/4 + r 2 ) r tanh ( π r ) d r , (261)
G θ ( H ) = θ Ω 1 / θ ( H ) ( θ > 0 ) , G 0 ( H ) = Ω ∞ ( H ) . G_\theta(H)=\theta\Omega_{1/\theta}(H)\quad(\theta>0),\qquad G_0(H)=\Omega_\infty(H). G θ ( H ) = θ Ω 1/ θ ( H ) ( θ > 0 ) , G 0 ( H ) = Ω ∞ ( H ) .
The square root is positive on the positive real axis. Initially H H H is smooth, even and compactly supported; the proof in Appendix V extends G θ G_\theta G θ to G z G_z G z by length cutoffs.
Theorem 19.4 (Full regularized spectral law). There is a vaguely continuous family of signed Radon measures m θ m_\theta m θ , 0 ≤ θ < ∞ 0\leq\theta<\infty 0 ≤ θ < ∞ , on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) such that whenever n → ∞ n\to\infty n → ∞ and g / n → θ g/n\to\theta g / n → θ ,
E ∣ m X ( F ) n − m θ ( F ) ∣ ⟶ 0 , F ∈ C c ( [ 0 , ∞ ) ) . (262) \mathbb{E}\left|\frac{m_X(F)}{n}-m_\theta(F)\right|\longrightarrow0,\qquad F\in C_c([0,\infty)). \tag*{(262)} E n m X ( F ) − m θ ( F ) ⟶ 0 , F ∈ C c ([ 0 , ∞ )) . (262) The limit is characterized by its heat transform
R θ ( z ) : = ∫ 0 ∞ e − z E d m θ ( E ) = 2 π ( 1 + 2 θ ) k H ( z ) + G θ ( G z ) , Re z > 0. (263) R_\theta(z):=\int_0^\infty e^{-zE}\,dm_\theta(E)=2\pi(1+2\theta)k_{\mathrm{H}}(z)+G_\theta(G_z),\qquad\operatorname{Re}z>0. \tag*{(263)} R θ ( z ) := ∫ 0 ∞ e − z E d m θ ( E ) = 2 π ( 1 + 2 θ ) k H ( z ) + G θ ( G z ) , Re z > 0. (263) Writing R X ( z ) = m X ( e − z E ) R_X(z)=m_X(e^{-zE}) R X ( z ) = m X ( e − z E ) , one also has
E sup z ∈ K ∣ n − 1 R X ( q ) ( z ) − R θ ( q ) ( z ) ∣ ⟶ 0 (264) \mathbb{E}\sup_{z\in K}\left|n^{-1}R_X^{(q)}(z)-R_\theta^{(q)}(z)\right|\longrightarrow0 \tag*{(264)} E z ∈ K sup n − 1 R X ( q ) ( z ) − R θ ( q ) ( z ) ⟶ 0 (264) for every compact K ⊂ { Re z > 0 } K\subset\{\operatorname{Re}z>0\} K ⊂ { Re z > 0 } and fixed q ≥ 0 q\geq0 q ≥ 0 . On [ 0 , 1 / 4 ) [0,1/4) [ 0 , 1/4 ) the limit equals θ ν 1 / θ \theta\nu_{1/\theta} θ ν 1/ θ if θ > 0 \theta>0 θ > 0 and ν ∞ \nu_\infty ν ∞ if θ = 0 \theta=0 θ = 0 . In particular it has no atom at zero.
Thus, for the standard regularization (260) , the large-cusp spectral-measure question in [98, Remark 1.8] has a full-energy limit, including the threshold. For uniformly discrete sequences Benjamini–Schramm converging to H \mathbb{H} H , the normalized spectral measures converge to the Plancherel measure [2]; for hyperbolic surfaces see Le Masson and Shal̈sten [125] and Monk [149]. Here the local limit is not H \mathbb{H} H . Le Masson–Sahlsten [126, Theorem 5.1] obtain spectral convergence when the cusp count is sublinear in the genus, under Benjamini–Schramm convergence to H \mathbb{H} H and a uniform systole bound; the present regime has a different orbital limit and imposes no systole bound.
The theorem does not separate cusp forms from scattering, identify the atom at 1 / 4 1/4 1/4 , or give a moving-window law. Nor does signed weak convergence alone imply sharp interval counts, even at continuity points of the signed limit.
Structure of the proof. The starting point is the exact regularized trace formula [104, Proposition 5.1],
R X ( z ) = area ( X ) k H ( z ) + O X ( G z ) . (265) \mathcal{R}_X(z)=\operatorname{area}(X)k_{\mathbb{H}}(z)+\mathcal{O}_X(G_z). \tag*{(265)} R X ( z ) = area ( X ) k H ( z ) + O X ( G z ) . (265)
This uses the unshifted Laplacian and the oriented convention for O X \mathcal{O}_X O X of Definition J.1 (a sum over oriented primitive classes and their repetitions); every term in the parabolic part retains the factor e − z / 4 e^{-z/4} e − z /4 . The proof then has three steps. A packing bound in the thick part controls long repetitions without any systole bound. After these are cut off, the orbital term converges by (255) and (259) . Finally, continuous energy tests are recovered from heat transforms through a bound on the signed total variation, which uses the residual-pole decomposition of the scattering determinant. The details are in Appendix V. The heat transform (263) is the input of the determinant limit in the next subsection.
A determinant consequence
The full spectral law also controls a singular global statistic, the log determinant of the Laplacian: Corollary 19.5 gives the L 1 L^1 L 1 limit of the normalized log determinant, and of log Z X ′ ( 1 ) \log Z_X'(1) log Z X ′ ( 1 ) , whenever g / n → θ ∈ [ 0 , ∞ ) g/n\to\theta\in[0,\infty) g / n → θ ∈ [ 0 , ∞ ) . Its inputs are the heat transform (263) of Theorem 19.4 on compact time intervals and, at both ends of the time axis, the short-geodesic intensity bound and low-energy tail of Proposition U.1 (Appendix U); the proof is in Appendix V. On compact hyperbolic surfaces the zeta-regularized determinant is expressed through the Selberg zeta function [55, 168], and Efrat treated surfaces of finite area [62]. Naud [154] proved convergence of the normalized log determinant for several closed random-surface models, and He–Wu [89] obtained L 1 L^1 L 1 convergence in the closed Weil–Petersson model. Jorgenson, Smajlović and Spilioti [113, Theorem 5.3] obtain deterministic large-volume limits for weighted Laplacian determinants on compact surfaces, under a weak spectral gap, uniform discreteness and a non-accumulation hypothesis for bounded-length geodesics. Gross–Lachman–Nachmias [78, Theorem 1] treat arbitrary local limits of compact hyperbolic surfaces under uniform integrability of short inverse lengths, without a uniform spectral gap. In the present noncompact setting, the growing-cusp law has a nonzero orbital contribution, and the scattering and parabolic terms in (260) are retained.
Write A X = area ( X ) A_X=\operatorname{area}(X) A X = area ( X ) and a θ = 2 π ( 1 + 2 θ ) a_\theta=2\pi(1+2\theta) a θ = 2 π ( 1 + 2 θ ) , and put
B X ( t ) = O X ( G t ) , B θ ( t ) = G θ ( G t ) , I θ = ∫ 0 ∞ B θ ( t ) d t t . (266) B_X(t)=\mathcal{O}_X(G_t),\qquad B_\theta(t)=\mathcal{G}_\theta(G_t),\qquad I_\theta=\int_0^\infty B_\theta(t)\frac{\mathrm{d}t}{t}. \tag*{(266)} B X ( t ) = O X ( G t ) , B θ ( t ) = G θ ( G t ) , I θ = ∫ 0 ∞ B θ ( t ) t d t . (266)
Use the Selberg product
Z X ( s ) = ∏ γ p r i m i t i v e o r i e n t e d ∏ j ≥ 0 ( 1 − e − ( s + j ) ℓ γ ) , Re s > 1 , Z_X(s)=\prod_{\gamma\ \mathrm{primitive\ oriented}}\prod_{j\ge0}\left(1-e^{-(s+j)\ell_\gamma}\right),\qquad\operatorname{Re}s>1, Z X ( s ) = γ primitive oriented ∏ j ≥ 0 ∏ ( 1 − e − ( s + j ) ℓ γ ) , Re s > 1 ,
with the same oriented convention as (265) . Its continuation has a simple zero at s = 1 s=1 s = 1 ; see [93, 62, 25].7 To specify the determinant on the noncompact surface, define
ζ X g e o m ( w ) = 1 Γ ( w ) ∫ 0 ∞ t w − 1 ( R X ( t ) − 1 ) d t , log det g e o m ′ Δ X = − ( ζ X g e o m ) ′ ( 0 ) . (267) \zeta_X^{\mathrm{geom}}(w)=\frac{1}{\Gamma(w)}\int_0^\infty t^{w-1}\left(\mathcal{R}_X(t)-1\right)\,dt,\qquad\log\det{}'_{\mathrm{geom}}\Delta_X=-\left(\zeta_X^{\mathrm{geom}}\right)'(0). \tag*{(267)} ζ X geom ( w ) = Γ ( w ) 1 ∫ 0 ∞ t w − 1 ( R X ( t ) − 1 ) d t , log det geom ′ Δ X = − ( ζ X geom ) ′ ( 0 ) . (267)
The integral initially converges for Re w > 1 \operatorname{Re} w>1 Re w > 1 ; continuation to zero is justified in the proof. The hyperbolic-plane density is
E H = 4 ζ R ′ ( − 1 ) − 1 / 2 + log ( 2 π ) 4 π , E_{\mathbb{H}}=\frac{4\zeta_{\mathbb{R}}'(-1)-1/2+\log(2\pi)}{4\pi}, E H = 4 π 4 ζ R ′ ( − 1 ) − 1/2 + log ( 2 π ) ,
where ζ R \zeta_{\mathbb{R}} ζ R is the Riemann zeta function [154, Section 2].
Corollary 19.5 (Log determinants at positive cusp density). The function I θ I_\theta I θ in (266) is finite, strictly positive and continuous for 0 ≤ θ < ∞ 0\leq\theta<\infty 0 ≤ θ < ∞ . If n → ∞ n\to\infty n → ∞ and g / n → θ g/n\to\theta g / n → θ under normalized Weil–Petersson law, then
E ∣ log Z X ′ ( 1 ) n + I θ ∣ ⟶ 0 , \mathbb{E}\left|\frac{\log Z_X'(1)}{n}+I_\theta\right|\longrightarrow0, E n log Z X ′ ( 1 ) + I θ ⟶ 0 , E ∣ log det g e o m ′ Δ X A X − ( E H − I θ a θ ) ∣ ⟶ 0. (268) \mathbb{E}\left|\frac{\log\det{}'_{\mathrm{geom}}\Delta_X}{A_X}-\left(E_{\mathbb{H}}-\frac{I_\theta}{a_\theta}\right)\right|\longrightarrow0. \tag*{(268)} E A X log det geom ′ Δ X − ( E H − a θ I θ ) ⟶ 0. (268) In particular the latter limit is strictly below E H E_{\mathbb{H}} E H .
The two limits are equivalent: for every surface X X X one has the exact identity log det g e o m ′ Δ X = A X E H + log Z X ′ ( 1 ) \log\det{}'_{\mathrm{geom}}\Delta_X=A_XE_{\mathbb{H}}+\log Z_X'(1) log det geom ′ Δ X = A X E H + log Z X ′ ( 1 ) (Appendix V), and A X / n → a θ A_X/n\to a_\theta A X / n → a θ . The corollary includes g = o ( n ) g=o(n) g = o ( n ) and every fixed genus. The determinant is defined precisely by (267) ; alternative noncompact regularizations may introduce cusp-dependent constants, whereas the Selberg-derivative statement fixes the normalization independently of this terminology. Neither statement asserts convergence of exponential moments.
Structure of the proof. The classical heat-to-determinant method [154, 89, 78] expresses log Z X ′ ( 1 ) \log Z_X'(1) log Z X ′ ( 1 ) through the integral of the orbital heat term B X ( t ) B_X(t) B X ( t ) against d t / t dt/t d t / t , with the zero mode subtracted at large times. Theorem 19.4 gives convergence on compact time intervals, so the work lies at the two ends, which require averaged bounds uniform in the surface. At small times we use an intensity bound for short simple geodesics of every separating topology; at large times, the low-energy tail of Proposition U.1 replaces a uniform spectral gap. Short two-cusp pairs of pants make I θ I_\theta I θ strictly positive. The details are in Appendix V. This corollary closes the spectral part; the regularized measure and the determinant are not used in Part 3.
Part 3. Local geometry and scaling limits
At positive cusp density n / g → λ ∈ ( 0 , ∞ ) n/g \to\lambda\in(0,\infty) n / g → λ ∈ ( 0 , ∞ ) , the volume-ratio profiles of Theorem 1.3 enter the carrier integrals of Appendix T. Retaining based holonomies in these integrals determines whole pointed balls and constructs the local family. This part uses no spectral theorem of Part 2; it also uses the genus-zero convolution identities of Appendix U and the cusp cutoffs of Appendix § 10.22 . The family’s planar endpoint has a Brownian-plane scaling limit, proved through a separate growing-radius tree coupling.
In genus zero, Budd and Curien showed that the Weil–Petersson random sphere with n + 1 n+1 n + 1 cusps, with its metric rescaled by n − 1 / 4 n^{-1/4} n − 1/4 , converges to the Brownian sphere [31], the universal scaling limit of uniform random planar maps [124, 140]. In the discrete setting, uniform triangulations of high genus converge locally to the planar stochastic hyperbolic triangulations of Curien [49, 37], whose near-critical scaling limit is the hyperbolic Brownian plane of Budzinski [35]. Theorem 20.1 is a continuum Weil–Petersson analogue of the high-genus local-limit statement. For closed Weil–Petersson surfaces, Monk proved quantitative Benjamini–Schramm convergence to the hyperbolic plane [149]; the limit λ ↓ 0 \lambda\downarrow0 λ ↓ 0 in (271) recovers the plane from positive density. Theorems 20.1 and 20.2 construct the local family at every finite positive cusp density and show that the cusp-filled surfaces are conformally discs, with the recurrent infinite WP sphere as the infinite-density endpoint. Budd and Curien asked for a hyperbolic analogue of their infinite sphere, as in [49], and conjectured that the infinite sphere is parabolic, with recurrent Brownian motion and ergodic geodesic flow [31, Section 7.3]. The family S λ S_\lambda S λ provides such analogues, and Theorem 20.2 with Corollary 20.3 proves the conjecture. The conformal-type conclusion concerns cusp-filled surfaces and does not construct a canonical embedding or describe the finer conformal structure; the near-critical scaling question remains open. A related peeling approach by Budd, Budzinski, Curien and Petri was announced in [31, reference 19]; our local construction instead retains based holonomies in the carrier integration formula. Theorem 21.1 answers their question about the Brownian-plane scaling limit of the infinite area-rooted WP surface [31, Section 7.3], with cusp-trimmed measured GHP convergence, full-area GP convergence and a finite mesoscopic version. Its growing-radius coupling, proved in Appendix Y, describes the planar geometry but does not supply the energy comparison required for the dense microscopic spectrum.
Local geometry and conformal type at positive cusp density
We construct the local WP family at n / g → λ ∈ ( 0 , ∞ ) n/g \to\lambda\in(0,\infty) n / g → λ ∈ ( 0 , ∞ ) and determine the conformal type after filling its cusps. Theorem 20.1 is proved in Appendix W; the finite-density and recurrent-endpoint parts of Theorem 20.2 are proved in Appendices X and § 10.22 , respectively.
Throughout this section, B r ( S , o ) B_r(S,o) B r ( S , o ) carries the restricted ambient metric and the hyperbolic area, and r r r is fixed before the large-genus limit. Total variation distances are taken between laws of pointed measured balls up to isometry. Write A g , n = 2 π ( 2 g − 2 + n ) A_{g,n}=2\pi(2g-2+n) A g , n = 2 π ( 2 g − 2 + n ) .
Theorem 20.1 (The local WP family). For every λ ∈ ( 0 , ∞ ) \lambda\in(0,\infty) λ ∈ ( 0 , ∞ ) there is a unique law L λ \mathcal{L}_\lambda L λ of complete pointed hyperbolic surfaces ( S λ , o λ ) (S_\lambda,o_\lambda) ( S λ , o λ ) with the following properties. If g → ∞ g \to\infty g → ∞ , n / g → λ n/g \to\lambda n / g → λ , X X X has normalized WP law on M g , n \mathcal{M}_{g,n} M g , n , and o o o is conditionally uniform for its area, then, for every fixed r < ∞ r<\infty r < ∞ ,
d T V ( Law ( B r ( X , o ) ) , Law ( B r ( S λ , o λ ) ) ) ⟶ 0. (269) d_{\mathrm{TV}}\bigl(\operatorname{Law}(B_r(X,o)),\operatorname{Law}(B_r(S_\lambda,o_\lambda))\bigr) \longrightarrow0. \tag*{(269)} d TV ( Law ( B r ( X , o )) , Law ( B r ( S λ , o λ )) ) ⟶ 0. (269) Moreover, for every bounded Borel function F F F of a pointed measured r r r -ball,
1 A g , n ∫ X F ( B r ( X , p ) ) d A ( p ) ⟶ E L λ F ( B r ) (270) \frac{1}{A_{g,n}}\int_X F(B_r(X,p))\,dA(p) \longrightarrow\mathbb{E}_{\mathcal{L}_\lambda}F(B_r) \tag*{(270)} A g , n 1 ∫ X F ( B r ( X , p )) d A ( p ) ⟶ E L λ F ( B r ) (270) in L 2 ( W P ) L^2(\mathrm{WP}) L 2 ( WP ) .
The law L λ \mathcal{L}_\lambda L λ is unimodular. It depends continuously on λ \lambda λ in fixed-radius total variation, with endpoints
L λ ⟶ ( H 2 , o ) ( λ ↓ 0 ) , L λ ⟶ ( S ∞ ∗ , o ∞ ) ( λ ↑ ∞ ) (271) \mathcal{L}_\lambda\longrightarrow(\mathbb{H}^2,o)\quad(\lambda\downarrow0),\qquad
\mathcal{L}_\lambda\longrightarrow(S^*_\infty,o_\infty)\quad(\lambda\uparrow\infty) \tag*{(271)} L λ ⟶ ( H 2 , o ) ( λ ↓ 0 ) , L λ ⟶ ( S ∞ ∗ , o ∞ ) ( λ ↑ ∞ ) (271) in the same sense, where the second endpoint is area-rooted as in [31, Theorem 34].
Here unimodularity means that, for every nonnegative isometry-invariant Borel function m m m of a surface with two marked points,
E ∫ S m ( S , o , z ) d A ( z ) = E ∫ S m ( S , z , o ) d A ( z ) . (272) \mathbb{E}\int_{S} m(S,o,z)\,dA(z)=\mathbb{E}\int_{S} m(S,z,o)\,dA(z). \tag*{(272)} E ∫ S m ( S , o , z ) d A ( z ) = E ∫ S m ( S , z , o ) d A ( z ) . (272)
This is the manifold mass-transport identity of [1]; for unimodular random networks see [8]. The theorem requires no ergodicity assertion. The family is nonconstant: for every fixed circumference 0 < b ≤ 2 0 < b \le2 0 < b ≤ 2 of the standard cusp horocycles, its cusp occupation probability is
P L λ { o λ lies in a b -cusp tail of S λ } = b λ 2 π ( 2 + λ ) . (273) \mathbb{P}_{\mathcal{L}_{\lambda}}\{o_{\lambda}\text{ lies in a }b\text{-cusp tail of }S_{\lambda}\}=\frac{b\lambda}{2\pi(2+\lambda)}. \tag*{(273)} P L λ { o λ lies in a b -cusp tail of S λ } = 2 π ( 2 + λ ) bλ . (273)
Theorem 20.2 (Conformal transition). For 0 < λ < ∞ 0 < \lambda< \infty 0 < λ < ∞ , almost surely S λ S_{\lambda} S λ has genus zero and infinite area, and every simple closed geodesic bounds a unique finite planar cap. After filling its isolated cusps, the resulting Riemann surface is conformally equivalent to D \mathbb{D} D . Brownian motion for the original complete hyperbolic metric is transient.
At the endpoint λ = ∞ \lambda=\infty λ = ∞ , the Riemann surface obtained from S ∞ S_{\infty} S ∞ by filling its cusps is conformally equivalent to C \mathbb{C} C , and Brownian motion for the hyperbolic metric of S ∞ S_{\infty} S ∞ is recurrent.
The endpoint assertion is the parabolicity conjectured by Budd and Curien [31, Section 7.3], who also conjectured that the geodesic flow is ergodic. The latter follows from the classical equivalence between recurrence of Brownian motion and ergodicity of the geodesic flow.
Corollary 20.3 (Geodesic flow). For 0 < λ < ∞ 0 < \lambda< \infty 0 < λ < ∞ , almost surely the geodesic flow on the unit tangent bundle of S λ S_{\lambda} S λ is not ergodic for the Liouville measure. At the endpoint, almost surely the geodesic flow on the unit tangent bundle of S ∞ S_{\infty} S ∞ is ergodic. In particular, the area-rooted infinite Weil–Petersson sphere is parabolic, has recurrent Brownian motion and has ergodic geodesic flow, as conjectured in [31, Section 7.3].
Proof. Each surface is complete and hyperbolic, hence of the form H 2 / Γ \mathbb{H}^{2}/\Gamma H 2 /Γ for a torsion-free Fuchsian group Γ \Gamma Γ . For such a surface, Brownian motion is recurrent if and only if the Poincaré series ∑ γ ∈ Γ e − d ( z , γ z ) \sum_{\gamma\in\Gamma}e^{-d(z,\gamma z)} ∑ γ ∈ Γ e − d ( z , γ z ) diverges, because the Green function of the quotient is the Γ \Gamma Γ -sum of that of H 2 \mathbb{H}^{2} H 2 , which decays like e − d e^{-d} e − d ; and divergence of this series is equivalent to ergodicity of the geodesic flow for the Liouville measure. This is the theorem of Hopf, Tsuji and Sullivan [177]; see [19, Section 1.1] for this list of equivalent conditions. Theorem 20.2 gives transience for 0 < λ < ∞ 0 < \lambda< \infty 0 < λ < ∞ and recurrence at the endpoint. □
Appendix W proves Theorem 20.1 by retaining the primitive holonomies of short based loops. Their number in a fixed displacement window has a deterministic bound, even in a thin cusp, and a support lemma for generators and their pairwise products permits exact carrier unfolding. The one- and two-carrier coefficient estimates of Appendix T, applied to uniformly bounded rooted fibers, then determine all local statistics and their variances.
The finite-density assertion of Theorem 20.2 rests on the subcritical cap-size tail, which implies that only finitely many caps with small boundary-to-area ratio can contain a fixed non-elementary core. This gives an eventual isoperimetric inequality for connected domains containing a fixed point, and the Green-level argument in Appendix X proves transience directly for the original metric. The recurrent endpoint is the conformal conclusion proved in Appendix § 10.22 . The dichotomy parallels the hyperbolic–parabolic dichotomy for unimodular random planar maps [13], the recurrence of local limits of finite planar graphs [22, 85] and the parabolicity of the Riemann surface of the uniform infinite planar triangulation [72]; we use these only as analogies. For recurrence and transience of Brownian motion on Riemannian manifolds see [75].
The limits of these statements are as follows. The convergence (269) keeps the radius fixed, and its estimates may depend on λ \lambda λ ; the results give neither a growing-radius approximation uniform as λ → ∞ \lambda\to\infty λ → ∞ nor the inverse-energy estimate needed for the dense microscopic spectrum. The near-critical scaling limit of a hyperbolic family requested in [31, Section 7.3] therefore remains a separate question (Section 22). The next section treats a metric scaling limit at the planar endpoint.
The planar endpoint and the Brownian plane
The area-rooted infinite WP sphere, the λ = ∞ \lambda= \infty λ = ∞ endpoint of (271) , has the Brownian plane as its scaling limit. The same limit holds for finite WP spheres at every mesoscopic radius 1 ≪ r n ≪ n 1 / 4 1 \ll r_n \ll n^{1/4} 1 ≪ r n ≪ n 1/4 . We prove measured Gromov–Hausdorff–Prokhorov convergence after trimming the deep cusps and projecting their full area onto the trim boundaries, and Gromov–Prokhorov convergence for the untrimmed surface with its full area. The proof in Appendix Y couples growing balls through the continuous-type trees of Budd and Curien [31], then transfers their finite Brownian-sphere limit [31, Theorem 2] by the intermediate-size strategy of Curien and Le Gall [50].
Write S n S_n S n for a WP sphere with n + 1 n+1 n + 1 cusps, A n = 2 π ( n − 1 ) A_n = 2\pi(n-1) A n = 2 π ( n − 1 ) for its area, and o n o_n o n for a conditionally uniform area point. Let ( S ∞ ∗ , o ∞ ) (S_\infty^\ast,o_\infty) ( S ∞ ∗ , o ∞ ) be the area-rooted infinite WP sphere of [31, Theorem 34]. For either surface, remove the open cusp neighborhoods bounded by horocycles of length one and write S ∘ S^\circ S ∘ for the remaining closed subset. Set
X ( S , o ) = S ∘ ∪ { o } . X(S,o) = S^\circ\cup\{o\}. X ( S , o ) = S ∘ ∪ { o } .
This space carries the original ambient hyperbolic distance, not its intrinsic path distance. The root is retained even when it lies in a removed cusp. Let P : S → S ∘ P:S \to S^\circ P : S → S ∘ be vertical projection in each cusp and the identity elsewhere, and put A ‾ = P ∗ d A \overline{A}=P_\ast dA A = P ∗ d A . In particular,
A ‾ = d A ∣ S ∘ + ∑ C d ℓ ∂ C . (274) \overline{A} = dA|_{S^\circ} + \sum_C d\ell_{\partial C}. \tag*{(274)} A = d A ∣ S ∘ + C ∑ d ℓ ∂ C . (274)
Thus each cusp contributes its entire unit area as boundary arclength.
Let ( P , D ∞ , μ ∞ , ρ ∞ ) (\mathcal{P},D_\infty,\mu_\infty,\rho_\infty) ( P , D ∞ , μ ∞ , ρ ∞ ) be the Brownian plane, with the natural volume obtained by pushing Lebesgue measure through the coding of Curien and Le Gall [50]. With the Brownian-sphere distance convention of [31, 50], put
c W P = 2 π 3 j 0 , 1 , v W P = 2 π c W P 4 . (275) c_{\mathrm{WP}} = \frac{2\pi}{\sqrt{3j_{0,1}}}, \qquad v_{\mathrm{WP}} = \frac{2\pi}{c_{\mathrm{WP}}^{4}}. \tag*{(275)} c WP = 3 j 0 , 1 2 π , v WP = c WP 4 2 π . (275)
We use local pointed Gromov–Hausdorff–Prokhorov convergence on proper spaces, allowing the marked root to lie outside the measure support. Explicitly, the local distance is the integral, against e − r d r e^{-r}\,dr e − r d r , of the compact pointed GHP distance between closed r r r -balls, capped at one. Balls carry restricted ambient distances and finite measures, without renormalization. This is the formula by which Abraham, Delmas and Hoscheit define the Gromov–Hausdorff–Prokhorov distance on locally compact length spaces [3]; we use it on proper spaces. The pointed GP convention omits the Hausdorff term and identifies spaces on the support of the measure together with the marked root, as in the Gromov–Prokhorov metric of Greven, Pfaffelhuber and Winter [74] and its pointed locally finite form, the Gromov-vague topology of [15]. These definitions also apply to the possibly nonlength spaces X ( S , o ) X(S,o) X ( S , o ) , whose root may lie outside the support of the measure. Indeed, for a proper pointed space with a locally finite measure every closed ball is compact and carries a finite restricted measure, so the compact pointed GHP distance between r r r -balls is finite, with the roots compared as marked points whether or not they lie in the supports; as r ′ ↓ r r' \downarrow r r ′ ↓ r the closed r ′ r' r ′ -balls decrease to the closed r r r -ball in the Hausdorff sense and the restricted measures converge weakly by monotone convergence, so the integrand is right-continuous in r r r , hence Borel, and the integral is defined. No length-space property is used, and if two spaces have isometric measured r r r -balls for almost every r r r , then for every r r r (restrict an isometry of a larger ball) and hence, by a diagonal argument on proper spaces, the spaces are isometric. Convergence in this local distance follows from convergence of the compact measured r r r -balls for almost every r r r , by dominated convergence of the capped integrand against e − r d r e^{-r}\,dr e − r d r ; this is the form in which the convergences below are established in Appendix Y, through an almost sure representation for the finite spheres and through bounded Lipschitz test functions of the restricted balls for the infinite surface. The limiting spaces there are length spaces, and only the prelimit spaces lack this property.
Theorem 21.1 (Brownian-plane limit). As L → ∞ L \to\infty L → ∞ , the following convergences hold in distribution:
( X ( S ∞ ∗ , o ∞ ) , L − 1 d , v W P − 1 L − 4 A ‾ , o ∞ ) ⟹ ( P , D ∞ , μ ∞ , ρ ∞ ) , (276) \left(X(S_\infty^*,o_\infty),L^{-1}d,v_{\mathrm{WP}}^{-1}L^{-4}\overline{A},o_\infty\right)\Longrightarrow(\mathcal{P},D_\infty,\mu_\infty,\rho_\infty), \tag*{(276)} ( X ( S ∞ ∗ , o ∞ ) , L − 1 d , v WP − 1 L − 4 A , o ∞ ) ⟹ ( P , D ∞ , μ ∞ , ρ ∞ ) , (276) ( S ∞ ∗ , L − 1 d , v W P − 1 L − 4 d A , o ∞ ) ⟹ ( P , D ∞ , μ ∞ , ρ ∞ ) . (277) \left(S_\infty^*,L^{-1}d,v_{\mathrm{WP}}^{-1}L^{-4}dA,o_\infty\right)\Longrightarrow(\mathcal{P},D_\infty,\mu_\infty,\rho_\infty). \tag*{(277)} ( S ∞ ∗ , L − 1 d , v WP − 1 L − 4 d A , o ∞ ) ⟹ ( P , D ∞ , μ ∞ , ρ ∞ ) . (277) The topology is local pointed GHP in (276) and local pointed GP in (277) . Both statements also hold with ( S n , o n ) (S_n,o_n) ( S n , o n ) and L = r n L=r_n L = r n for every deterministic sequence 1 ≪ r n ≪ n 1 / 4 1\ll r_n\ll n^{1/4} 1 ≪ r n ≪ n 1/4 .
The main step of the proof, given in Appendix Y, is a coupling of finite and infinite WP balls at every radius o ( n 1 / 4 ) o(n^{1/4}) o ( n 1/4 ) . An exact likelihood for a tree prefix marked by a uniform leaf, together with mixing of the continuous angle types, permits the prefix depth to grow. A functional central limit theorem for the spine labels then gives the necessary geometric separation. Passing through a finite surface of intermediate size transfers the Brownian-sphere tangent coupling of [50]; its coding construction preserves the full volume measure as well as distances.
Corollary 21.2 (Radial volume profile). Put V ∞ ( t ) = μ ∞ ( B t ( ρ ∞ ) ) V_\infty(t)=\mu_\infty(B_t(\rho_\infty)) V ∞ ( t ) = μ ∞ ( B t ( ρ ∞ )) . Then
( A ( B L t ( S ∞ ∗ , o ∞ ) ) v W P L 4 ) t ≥ 0 ⟹ ( V ∞ ( t ) ) t ≥ 0 in C l o c ( [ 0 , ∞ ) ) . (278) \left(\frac{A(B_{Lt}(S_\infty^*,o_\infty))}{v_{\mathrm{WP}}L^4}\right)_{t\geq0}\Longrightarrow(V_\infty(t))_{t\geq0}\quad\text{in }C_{\mathrm{loc}}([0,\infty)). \tag*{(278)} ( v WP L 4 A ( B L t ( S ∞ ∗ , o ∞ )) ) t ≥ 0 ⟹ ( V ∞ ( t ) ) t ≥ 0 in C loc ([ 0 , ∞ )) . (278) In particular, log A ( B L ( S ∞ ∗ , o ∞ ) ) / log L → 4 \log A(B_L(S_\infty^*,o_\infty))/\log L\to4 log A ( B L ( S ∞ ∗ , o ∞ )) / log L → 4 in probability. The same conclusions hold for ( S n , o n ) (S_n,o_n) ( S n , o n ) and L = r n L=r_n L = r n in the mesoscopic range of Theorem 21.1 .
Proof. The Brownian-plane radial measure has a continuous positive density on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) [122, Proposition 1]. Thus every positive radius has zero boundary mass, and V ∞ V_\infty V ∞ is continuous and finite, with V ∞ ( t ) > 0 V_\infty(t)>0 V ∞ ( t ) > 0 for t > 0 t>0 t > 0 . Theorem 21.1 gives joint convergence of the ball masses at any finite list of radii. The prelimit profiles are continuous and nondecreasing: Riemannian distance spheres have area zero. On a finite mesh, monotonicity bounds each intermediate value between its two endpoint values. Uniform continuity of the limiting profile therefore upgrades these joint limits to convergence in C l o c C_{\mathrm{loc}} C loc . Finally the logarithm of the normalized volume at radius one is tight, proving the stated exponent. □
The cusp trimming cannot be dropped from the Hausdorff statement: the untrimmed metric spaces do not converge locally in GH. Indeed, choose any K K K distinct cusp tails in S ∞ ∗ S_\infty^* S ∞ ∗ and entrance points at distances at most a K < ∞ a_K<\infty a K < ∞ from the root. Points at depth L / 2 L/2 L /2 in these tails lie in B 3 L / 4 ( o ∞ ) B_{3L/4}(o_\infty) B 3 L /4 ( o ∞ ) when L ≥ 4 a K L\geq4a_K L ≥ 4 a K , and are pairwise at distance at least L L L . The rescaled unit balls thus fail GH tightness. Since the deep cusps carry exponentially small mass, this is compatible with (277) . The theorem describes geometry and area; the energy comparison needed for the dense microscopic spectrum remains a separate problem.
Perspectives and open problems
The uniform expansion separates the changing leading growth of V g , n V_{g,n} V g , n from its inverse-genus corrections. The saddle parameter absorbs the former; the positive string-flow recursion controls the latter through both endpoints. Painlevé I fixes the normalization at the pure-ψ \psi ψ endpoint, while the positive convolution factor and Poisson extraction make the passage to individual volumes uniform from bounded cusp counts to the critical singularity. The resulting volume ratios provide the counting input for the spectral laws and the local geometry.
The two applications use this input differently. When n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g}\to c\in(0,\infty) n / g → c ∈ ( 0 , ∞ ) , rare two-cusp pairs of pants support funnel ground states. A trace comparison proves that they asymptotically exhaust the positive spectrum on each fixed compact energy window below 1 / 4 1/4 1/4 . In the supercritical Poisson regimes, the lowest eigenvalues are controlled by collar conductances divided by enclosed areas. At sufficiently large cusp density, low-energy functions inside growing planar caps invalidate that normalization. Fixed-energy windows still admit deterministic limits, because orbital statistics see the aggregate contribution of the many local pieces. The distinction between these limits leaves the dense microscopic problem below unresolved.
At n ≍ g n \asymp g n ≍ g , all planar cusp types contribute to the same carrier integrals. Retaining based holonomies makes these integrals determine whole pointed neighborhoods, yielding the local WP family. Its cusp-filled surfaces are conformally discs at each finite positive density; the infinite-density endpoint is the recurrent infinite WP sphere, whose cusp-filling is conformally the plane, as Budd and Curien conjectured. A separate growing-prefix coupling gives the Brownian-plane metric and area limit. Thus the common volume estimates support two further arguments: spectral exhaustion for the low eigenvalues and holonomy control for the local family. Passing from the latter’s metric limit to an energy limit remains an additional problem.
Open problems
We list the problems that the present methods leave open. Each lies beyond a specific obstruction identified in the proofs.
The dense microscopic spectrum. Theorem 18.3 proves the Poisson law at the scale A g A_g A g under (250) , which holds for n ≤ g 11 / 9 − δ n \leq g^{11/9-\delta} n ≤ g 11/9 − δ , and Theorem 18.1 shows that this normalization fails for n ≫ g 3 / ( log g ) 4 n \gg g^3/(\log g)^4 n ≫ g 3 / ( log g ) 4 . Neither range is asserted to be sharp. Between them one must determine the correct scale and the limit law of the smallest eigenvalues. Upper bounds come from trial functions on large planar caps; the missing ingredient is a matching lower bound on the whole low spectral subspace, which requires control of the interaction between growing planar caps and the rest of the surface. Corollary 18.4 already identifies the spectral type of these eigenvalues at every dense growth rate.
Edge fluctuations without short pants. Theorem 16.1 gives the Fréchet law of 1 / 4 − G ( X ) 1/4-G(X) 1/4 − G ( X ) when n / g → 0 n/\sqrt{g} \to0 n / g → 0 and ( n / g ) log g → ∞ (n/\sqrt{g})\log g \to\infty ( n / g ) log g → ∞ ; it does not cover the whole subcritical range. When short two-cusp pants are too rare to dominate, and in particular for closed surfaces, the fluctuations of the spectral gap around 1 / 4 1/4 1/4 remain open; Huang’s theorem [101] and the polynomial rate of Hide, Macera and Thomas [96] give convergence but no limit law.
Small cusp forms. With probability tending to one, there are no cusp forms with eigenvalue in ( 0 , 1 / 4 − ε ] (0,1/4-\varepsilon] ( 0 , 1/4 − ε ] when n = O ( g ) n=O(\sqrt{g}) n = O ( g ) (Theorem 15.1), and none below T / A g T/A_g T / A g when n / g → ∞ n/g \to\infty n / g → ∞ (Corollary 18.4). Whether Weil–Petersson random surfaces carry any cusp form below 1 / 4 1/4 1/4 in the other regimes, and how many, is open; this is the probabilistic counterpart of the question of Phillips and Sarnak [161] and of the conjecture of Otal and Rosas [157, Section 6] that, for g ≥ 2 g \geq2 g ≥ 2 , at most 2 g − 3 2g-3 2 g − 3 small eigenvalues are cuspidal; see [17], and [148] for partial results. Relatedly, the regularized measure of Theorem 19.4 does not separate cusp forms from scattering, and the fixed-window limits are compatible with a vanishing proportion of eigenvalues collapsing to zero.
Resurgence uniformly in the cusp density. The single-factorial bounds in Theorem 1.1 give Borel transforms convergent near the origin. Uniform analytic continuation, growth bounds for Borel summation and the corresponding Stokes data remain to be related to the nonperturbative predictions of [66], uniformly in n / g n/g n / g .
Near-critical scaling of the local family. The local family of Theorem 20.1 is obtained at fixed radius and fixed positive density. The near-critical scaling question of Budd and Curien [31, Section 7.3] requires estimates uniform as the radius and the density grow together. In the discrete setting the corresponding limit is Budzinski’s hyperbolic Brownian plane [35]; a continuum Weil–Petersson analogue is not known.
Energy on the Brownian plane. Theorem 21.1 gives a metric and area limit at the planar endpoint, but a limiting Dirichlet form or diffusion clock needs further energy control. This is also the obstacle to deriving a dense microscopic spectral law from the Brownian-plane geometry alone.
7. Mixed intersection numbers. Relative asymptotics for mixed intersection numbers when both the number of nonzero degrees and their sum grow with the genus, uniformly over their allocation, are not supplied by the present methods: the bounded boundary factorization of Proposition 9.4 does not give such individual coefficient estimates.
The uniform volume expansion determines the weights of finite geometric configurations across cusp densities. The spectral and local-limit results show what can be recovered from these weights, while the dense microscopic spectrum and joint geometric scaling require further control of energy and growing neighborhoods.
Proof appendices
Appendix A. Proof locations and notation
A.1. Proof locations. For each main result, the following table gives the place of its proof and the principal results and methods used there. References to § 3.1 –§ 3.7 include the supporting Appendices B and C. In the table, Thm., Cor., Prop., Sec. and App. abbreviate Theorem, Corollary, Proposition, Section and Appendix.
Table 3.
Table 4.
The volume-ratio limits and the short-pants consequences in § 4.1 use the main-text estimates and Appendices B–C. The subsequent length and spectral results use the additional arguments in Appendices G–V, and the geometric results those in Appendices W–Y, as indicated above. Within each part the appendices follow the order in which they are used. Appendices D–F are needed only for the inverse-genus rates and for the expansions to higher order, apart from the all-marking envelope (D.35), which Part 2 uses. They can be read independently of one another: Appendix D after § 3.22 , Appendix E after § 3.30 (it uses Theorem 8.2 only through its statement), and Appendix F after § 3.35 .
Notation
The first table lists the principal symbols of Part 1 and where they are defined; the second does the same for the recurring symbols of Parts 2 and 3.
Table 5.
In Parts 2 2 2 and 3 3 3 , X X X denotes a random surface, not the genus-zero relation of Part 1 1 1 , and λ \lambda λ is the cusp density n / g → λ n/g \to\lambda n / g → λ .
Table 6 (PDF p. 109).
The complexity parameter h h h is specified locally: we use h = 2 g − 2 + n h=2g-2+n h = 2 g − 2 + n for volumes with n n n boundary components and h = 2 g − 2 + p h=2g-2+p h = 2 g − 2 + p for generating functions with p p p distinguished markings. In particular, h = 2 g − 1 h=2g-1 h = 2 g − 1 in (5) , where p = 1 p=1 p = 1 and the remaining cusps are counted by k k k . Positive constants C C C , c c c may change from line to line, and a subscript, as in O p O_{p} O p , indicates the parameters on which an implied constant depends.
A few letters in Part 1 have an additional local meaning, stated where it occurs: D D D is the degree sum ∣ d ∣ |\mathbf{d}| ∣ d ∣ in § 3.4 –§ 3.19 and in Appendix § 8.5 ; h 0 = 2 g − 2 h_{0}=2g-2 h 0 = 2 g − 2 is the complexity of a closed surface (Appendices § 8.5 , § 10.2 and § 10.6 ); the harmonic coefficients A m ( h ) A_{m}(h) A m ( h ) are unrelated to the kernel scalar A ( Y ) A(Y) A ( Y ) and to the coefficients A j ( g , n ) A_{j}(g,n) A j ( g , n ) ; the Laplacian Δ X \Delta_{X} Δ X of a surface X X X is unrelated to Δ ( Y ) \Delta(Y) Δ ( Y ) ; and T T T is a length window in Appendix § 10.1 and an energy window T / A g T/A_{g} T / A g in Corollary 18.4.
In Parts 2 and 3 the following letters also carry more than one meaning. The cusp density λ \lambda λ of (3) is unrelated to the eigenvalues λ j ( X ) \lambda_{j}(X) λ j ( X ) and λ r + ( X ) \lambda_{r}^{+}(X) λ r + ( X ) ; s s s is a waist length in P s P_{s} P s , F s F_{s} F s , β ( s ) \beta(s) β ( s ) and Q c Q_{c} Q c , but the spectral parameter of Φ X ( s ) \Phi_{X}(s) Φ X ( s ) and Z X ( s ) Z_{X}(s) Z X ( s ) in Subsections 19.3 and 19.4; c c c is the critical ratio n / g → c n/\sqrt{g}\to c n / g → c in § 4.7 and § 4.12 and in Q c Q_{c} Q c , μ c \mu_{c} μ c , ν c \nu_{c} ν c and Λ c \Lambda_{c} Λ c , apart from its use as a generic constant and in c W P c_{\mathrm{WP}} c WP .
Q Q Q is the parameter n x / ( 4 P ) nx/(4P) n x / ( 4 P ) in (247) and a polynomial degree in § 4.7 and Appendix § 10.10 , and neither is the pants process Q c Q_{c} Q c ; K K K denotes the kernel K Y K_{Y} K Y in Part 1, the length measure K s K_{s} K s in Subsection 14.6, the systolic cusp count K g K_{g} K g in Theorem 14.2 , the moment constants K p K_{p} K p in Corollary 16.3 and the planar series K x ( L ) K_{x}(L) K x ( L ) of (T.2) in § 4.19 and Appendix T; H H H denotes the kernel H Y H_{Y} H Y in Part 1, an even test function in O X ( H ) \mathcal{O}_{X}(H) O X ( H ) and Ω λ ( H ) \Omega_{\lambda}(H) Ω λ ( H ) , and the level-set function H ( t ) H(t) H ( t ) of § 4.12 .
D D D is also the gap defect D ( X ) = 1 / 4 − G ( X ) D(X)=1/4-G(X) D ( X ) = 1/4 − G ( X ) and the deficit process D X \mathcal{D}_{X} D X in § 4.12 , the operator x ∂ x x\partial_{x} x ∂ x in § 4.13 , and the Brownian-plane distance D ∞ D_{\infty} D ∞ ; P P P is the constant 2 π 2 2\pi^{2} 2 π 2 in (247) , the pants P s P_{s} P s , the cusp projection in § 5.2 and, as P \mathcal{P} P , the Brownian plane; S S S is the dense spectral scale in (247) and a surface in S λ S_{\lambda} S λ , S n S_{n} S n and S ∞ ∗ S_{\infty}^{*} S ∞ ∗ ; and ν \nu ν without subscript is the tight intensity (200) , whereas ν ∗ \nu_{*} ν ∗ , ν λ \nu_{\lambda} ν λ , ν ∞ \nu_{\infty} ν ∞ , ν c \nu_{c} ν c and ν e d g e \nu_{\mathrm{edge}} ν edge are spectral limit measures.
Appendices to Part 1: volume asymptotics
Appendix B. Proofs for the kernels and the parameter
This appendix proves the three lemmas on which the kernel family and the density parameter of Part 1 rest: Lemma 2.1 (real zeros of K Y K_Y K Y ), Lemma 2.2 (positivity of the Taylor coefficients in Y Y Y , which supplies the convolution factor of Lemma 9.1) and Lemma 3.1 (the functions T T T , Δ \Delta Δ , I I I , λ \lambda λ and Ψ \Psi Ψ , whose endpoint expansions are used in Sections 9 and 11). The proofs use only the genus-zero relation X ′ ( y ) = J 0 ( 2 y ) X'(y)=J_0(2\sqrt{y}) X ′ ( y ) = J 0 ( 2 y ) , the explicit kernels (8) and classical facts about entire functions of order 1 / 2 1/2 1/2 ; no volume estimate enters, and the only dependence among them is that the later two proofs use the sign properties of X X X and the genus-zero product for a a a established in the proof of Lemma 2.1. The notation is that of Sections 2 and 3, and the subsections follow the order of the three lemmas.
B.1. Proof of Lemma 2.1. The proof has three stages, following the sketch given after the statement of the lemma in Subsection 2.1. We first verify the sign properties of X X X and its derivatives and locate the zeros of a a a . We then write K Y K_Y K Y along characteristics as sin t \sin t sin t minus a positive correction (B.1) and count its real zeros by concavity. Finally a Rouché comparison with sin t / t \sin t/t sin t / t on large circles shows that the real zeros exhaust all zeros, which gives the product (10) .
We use the elementary identities X ′ ′ ( y ) = − X ( y ) / y X''(y)=-X(y)/y X ′′ ( y ) = − X ( y ) / y and X ′ ′ ′ ( y ) = ∑ j ≥ 0 ( − 1 ) j y j / ( j ! ( j + 2 ) ! ) X'''(y)=\sum_{j\geq0}(-1)^j y^j/(j!(j+2)!) X ′′′ ( y ) = ∑ j ≥ 0 ( − 1 ) j y j / ( j ! ( j + 2 )!) . On [ 0 , π 2 / 4 ] [0,\pi^2/4] [ 0 , π 2 /4 ] , X ( y ) > 0 X(y)>0 X ( y ) > 0 for y > 0 y>0 y > 0 , X ′ ′ < 0 X''<0 X ′′ < 0 , and X ′ ′ ′ > 0 X'''>0 X ′′′ > 0 . For the first assertion we use the alternating-series lower bound
X ( y ) / y ≥ 1 − y / 2 + y 2 / 12 − y 3 / 144 > 0 ( 0 ≤ y ≤ 5 / 2 ) ; X(y)/y \geq1-y/2+y^2/12-y^3/144>0 \qquad(0\leq y\leq5/2); X ( y ) / y ≥ 1 − y /2 + y 2 /12 − y 3 /144 > 0 ( 0 ≤ y ≤ 5/2 ) ;
the cubic on the right is decreasing and is positive at 5 / 2 5/2 5/2 . The other assertions follow from the displayed identities and X ′ ′ ′ ( y ) ≥ 1 / 2 − y / 6 > 0 X'''(y)\geq1/2-y/6>0 X ′′′ ( y ) ≥ 1/2 − y /6 > 0 . Similarly, the partial sum of the series for a ( 3 / 2 ) a(3/2) a ( 3/2 ) through j = 4 j=4 j = 4 is an upper bound by the alternating-series test, and it is negative; this proves y 0 < 3 / 2 y_0<3/2 y 0 < 3/2 . In particular X ( Y ) ≤ Y < 3 / 2 X(Y)\leq Y<3/2 X ( Y ) ≤ Y < 3/2 .
All zeros of a ( y ) a(y) a ( y ) are positive real. Indeed, if a ( y ) = 0 a(y)=0 a ( y ) = 0 , the function u ( s ) = J 0 ( 2 y s ) u(s)=J_0(2\sqrt{y}s) u ( s ) = J 0 ( 2 y s ) solves ( s u ′ ) ′ + 4 y s u = 0 (su')'+4ysu=0 ( s u ′ ) ′ + 4 y s u = 0 , is regular at zero, and vanishes at one. Integration against u ˉ \bar{u} u ˉ gives
4 y ∫ 0 1 s ∣ u ∣ 2 d s = ∫ 0 1 s ∣ u ′ ∣ 2 d s > 0. 4y\int_0^1 s|u|^2\,ds=\int_0^1 s|u'|^2\,ds>0. 4 y ∫ 0 1 s ∣ u ∣ 2 d s = ∫ 0 1 s ∣ u ′ ∣ 2 d s > 0.
Since the zeros are positive real and y 0 y_0 y 0 is the smallest of them, a a a has no zeros in ∣ y ∣ < y 0 |y|<y_0 ∣ y ∣ < y 0 . Moreover the series is entire of order 1 / 2 1/2 1/2 , so its Hadamard product has genus zero and no nonconstant exponential factor; this product is used in the proof of Lemma 2.2 below.
We next prove that every zero of K Y K_Y K Y is positive and simple. At Y = 0 Y=0 Y = 0 ,
K 0 ( z ) = sin 2 z 2 z . K_0(z)=\frac{\sin\sqrt{2z}}{\sqrt{2z}}. K 0 ( z ) = 2 z sin 2 z .
Coefficient comparison in (8) gives
∂ Y K Y = 2 ∂ z K Y + K Y − a ( Y ) z . \partial_Y K_Y=2\partial_z K_Y+\frac{K_Y-a(Y)}{z}. ∂ Y K Y = 2 ∂ z K Y + z K Y − a ( Y ) .
Integration along the characteristics z + 2 Y = constant z+2Y=\text{constant} z + 2 Y = constant yields
t 2 − 4 Y K Y ( ( t 2 − 4 Y ) / 2 ) = sin t − R Y ( t ) , R Y ( t ) = 2 ∫ 0 Y a ( u ) t 2 − 4 u d u . (279) \sqrt{t^2-4Y}\,K_Y\left((t^2-4Y)/2\right)=\sin t-R_Y(t),\qquad R_Y(t)=2\int_0^Y\frac{a(u)}{\sqrt{t^2-4u}}\,du. \tag*{(279)} t 2 − 4 Y K Y ( ( t 2 − 4 Y ) /2 ) = sin t − R Y ( t ) , R Y ( t ) = 2 ∫ 0 Y t 2 − 4 u a ( u ) d u . (279)
For 0 < Y < y 0 0<Y<y_0 0 < Y < y 0 and real t > 2 Y t>2\sqrt{Y} t > 2 Y , R Y > 0 R_Y>0 R Y > 0 , R Y ′ < 0 R_Y'<0 R Y ′ < 0 , R Y ′ ′ > 0 R_Y''>0 R Y ′′ > 0 , and
R Y ( t ) ≤ 2 X ( Y ) t 2 − 4 Y . R_Y(t)\leq\frac{2X(Y)}{\sqrt{t^2-4Y}}. R Y ( t ) ≤ t 2 − 4 Y 2 X ( Y ) .
At Y = 0 Y=0 Y = 0 , R 0 R_0 R 0 vanishes identically, as in the sinc formula above. For Y > 0 Y>0 Y > 0 , sin t − R Y ( t ) \sin t-R_Y(t) sin t − R Y ( t ) is strictly concave on each positive sine interval. On the first interval ( 2 Y , π ) (2\sqrt{Y},\pi) ( 2 Y , π ) , it tends to zero at the left endpoint, is positive immediately afterwards by K Y ( 0 ) = a ( Y ) > 0 K_Y(0)=a(Y)>0 K Y ( 0 ) = a ( Y ) > 0 , and is negative at π \pi π ; hence it has exactly one simple interior zero. On ( 2 j π , ( 2 j + 1 ) π ) (2j\pi,(2j+1)\pi) ( 2 j π , ( 2 j + 1 ) π ) , j ≥ 1 j\ge1 j ≥ 1 , it is negative at both endpoints and positive at the midpoint, since
R Y ( ( 2 j + 1 2 ) π ) < 3 ( 5 π / 2 ) 2 − 6 < 1. R_Y\left((2j+\tfrac{1}{2})\pi\right)<\frac{3}{\sqrt{(5\pi/2)^2-6}}<1. R Y ( ( 2 j + 2 1 ) π ) < ( 5 π /2 ) 2 − 6 3 < 1.
Strict concavity gives exactly two simple zeros there. Negative sine intervals contain no zero. For real z ≤ 0 z\le0 z ≤ 0 , every argument of X ′ X' X ′ in (8) is less than y 0 y_0 y 0 , where X ′ > 0 X'>0 X ′ > 0 ; for negative arguments this follows from the positive modified-Bessel series. Thus K Y K_Y K Y has no nonpositive real zero.
To exclude nonreal zeros, set G Y ( t ) = K Y ( ( t 2 − 4 Y ) / 2 ) G_Y(t)=K_Y((t^2-4Y)/2) G Y ( t ) = K Y (( t 2 − 4 Y ) /2 ) . It is entire and even. Formula (B.1), with the square root asymptotic to t t t , holds outside the segment [ − 2 Y , 2 Y ] [-2\sqrt{Y},2\sqrt{Y}] [ − 2 Y , 2 Y ] . On ∣ t ∣ = ( N + 1 / 2 ) π |t|=(N+1/2)\pi ∣ t ∣ = ( N + 1/2 ) π ,
G Y ( t ) sin t / t = t t 2 − 4 Y ( 1 − R Y ( t ) sin t ) = 1 + O ( N − 1 ) . (280) \frac{G_Y(t)}{\sin t/t}=\frac{t}{\sqrt{t^2-4Y}}\left(1-\frac{R_Y(t)}{\sin t}\right)=1+O(N^{-1}). \tag*{(280)} sin t / t G Y ( t ) = t 2 − 4 Y t ( 1 − sin t R Y ( t ) ) = 1 + O ( N − 1 ) . (280)
The error is uniform: R Y = O ( N − 1 ) R_Y=O(N^{-1}) R Y = O ( N − 1 ) , and ∣ sin t ∣ |\sin t| ∣ sin t ∣ is bounded below on these circles. For the latter assertion use ∣ sin t ∣ ≥ sinh 1 |\sin t|\ge\sinh1 ∣ sin t ∣ ≥ sinh 1 when ∣ ℑ t ∣ ≥ 1 |\Im t|\ge1 ∣ℑ t ∣ ≥ 1 ; otherwise ∣ ℜ t ∣ |\Re t| ∣ℜ t ∣ differs from ( N + 1 / 2 ) π (N+1/2)\pi ( N + 1/2 ) π by O ( N − 1 ) O(N^{-1}) O ( N − 1 ) . Rouché’s theorem gives exactly 2 N 2N 2 N zeros of G Y G_Y G Y inside each sufficiently large circle. The real zeros already counted give exactly N N N positive and N N N negative zeros there: for even N N N the circle ends at the midpoint of a positive sine interval, after its first zero, and for odd N N N it ends in the subsequent negative interval. They exhaust the count, proving the assertion. Consequently the product formula (10) holds. This is again a genus-zero product, since (8) has order 1 / 2 1/2 1/2 .
Proof of Lemma 2.2
The coefficients of log ( a / H Y ) \log(a/H_Y) log ( a / H Y ) satisfy a system of differential equations in Y Y Y whose right-hand sides have nonnegative coefficients. Since these coefficients are positive at Y = 0 Y=0 Y = 0 , where the zeros of H 0 H_0 H 0 are positive real, induction on the order of differentiation shows that all their Taylor coefficients in Y Y Y are nonnegative. The same argument applies to log ( a / K Y ) \log(a/K_Y) log ( a / K Y ) .
We first prove (11) . By (8) ,
∂ Y H Y = 2 ∂ z H Y + 3 ( H Y − a ( Y ) ) / z . \partial_Y H_Y=2\partial_z H_Y+3(H_Y-a(Y))/z. ∂ Y H Y = 2 ∂ z H Y + 3 ( H Y − a ( Y )) / z .
Writing T = log ( a / H Y ) \mathcal{T}=\log(a/H_Y) T = log ( a / H Y ) and γ = − a ′ / a \gamma=-a'/a γ = − a ′ / a , this becomes
∂ Y T = − γ + 2 ∂ z T + 3 ( e T − 1 ) / z . \partial_Y\mathcal{T}=-\gamma+2\partial_z\mathcal{T}+3(e^{\mathcal{T}}-1)/z. ∂ Y T = − γ + 2 ∂ z T + 3 ( e T − 1 ) / z .
The constant identity is γ = 5 t 1 \gamma=5t_1 γ = 5 t 1 ; for m ≥ 1 m\ge1 m ≥ 1 ,
t m ′ = ( 2 m + 5 ) t m + 1 + 3 ∑ ℓ = 2 m + 1 1 ℓ ! ∑ i 1 + ⋯ + i ℓ = m + 1 i j ≥ 1 t i 1 ⋯ t i ℓ . (281) t'_m=(2m+5)t_{m+1}+3\sum_{\ell=2}^{m+1}\frac{1}{\ell!}\sum_{\substack{i_1+\cdots+i_\ell=m+1\\ i_j\ge1}}t_{i_1}\cdots t_{i_\ell}. \tag*{(281)} t m ′ = ( 2 m + 5 ) t m + 1 + 3 ℓ = 2 ∑ m + 1 ℓ ! 1 i 1 + ⋯ + i ℓ = m + 1 i j ≥ 1 ∑ t i 1 ⋯ t i ℓ . (281)
At Y = 0 Y=0 Y = 0 ,
H 0 ( z ) = 3 sin 2 z − 2 z cos 2 z ( 2 z ) 3 . H_0(z)=3\frac{\sin\sqrt{2z}-\sqrt{2z}\cos\sqrt{2z}}{(\sqrt{2z})^3}. H 0 ( z ) = 3 ( 2 z ) 3 sin 2 z − 2 z cos 2 z .
Its zeros ρ j \rho_j ρ j are positive real. For a direct proof, a nonzero root t t t of sin t − t cos t \sin t-t\cos t sin t − t cos t gives the regular function u ( s ) = j 1 ( t s ) u(s)=j_1(ts) u ( s ) = j 1 ( t s ) , satisfying ( s 2 u ′ ) ′ + ( t 2 s 2 − 2 ) u = 0 (s^2u')'+(t^2s^2-2)u=0 ( s 2 u ′ ) ′ + ( t 2 s 2 − 2 ) u = 0 and u ( 1 ) = 0 u(1)=0 u ( 1 ) = 0 . Integration against u ˉ \bar{u} u ˉ shows
t 2 ∫ 0 1 s 2 ∣ u ∣ 2 d s = ∫ 0 1 ( s 2 ∣ u ′ ∣ 2 + 2 ∣ u ∣ 2 ) d s > 0. t^2\int_0^1s^2|u|^2\,ds=\int_0^1(s^2|u'|^2+2|u|^2)\,ds>0. t 2 ∫ 0 1 s 2 ∣ u ∣ 2 d s = ∫ 0 1 ( s 2 ∣ u ′ ∣ 2 + 2∣ u ∣ 2 ) d s > 0.
Here j 1 ( w ) = sin w / w 2 − cos w / w j_1(w)=\sin w/w^2-\cos w/w j 1 ( w ) = sin w / w 2 − cos w / w is regular at zero. The root is simple because ( sin t − t cos t ) ′ = t sin t (\sin t-t\cos t)'=t\sin t ( sin t − t cos t ) ′ = t sin t does not vanish at a nonzero root. The order-1 / 2 1/2 1/2 product therefore gives t m ( 0 ) = m − 1 ∑ j ρ j − m > 0 t_m(0)=m^{-1}\sum_j\rho_j^{-m}>0 t m ( 0 ) = m − 1 ∑ j ρ j − m > 0 .
Induct on the derivative order q q q , simultaneously for all indices m m m , in (B.3). Every differentiated term is a product of derivatives of order at most $q with nonnegative coefficients, so the induction is valid; each fixed equation has only finitely many terms. For each m m m , t m ( Y ) t_m(Y) t m ( Y ) is a polynomial in the ratios X ( j ) ( Y ) / a ( Y ) X^{(j)}(Y)/a(Y) X ( j ) ( Y ) / a ( Y ) , so it is analytic on ∣ Y ∣ < y 0 |Y|<y_0 ∣ Y ∣ < y 0 , and its Taylor series converges to t m ( Y ) t_m(Y) t m ( Y ) there.
This proves (11) , and gives strictly positive coefficients of H Y − h H_Y^{-h} H Y − h for every real h > 0 h>0 h > 0 , and nonnegative coefficients of − H Y ′ / H Y -H'_Y/H_Y − H Y ′ / H Y .
The same argument applied to S = log ( a / K Y ) = ∑ m ≥ 1 s m ( Y ) z m S=\log(a/K_Y)=\sum_{m\geq1}s_m(Y)z^m S = log ( a / K Y ) = ∑ m ≥ 1 s m ( Y ) z m gives
s m ′ = ( 2 m + 3 ) s m + 1 + ∑ ℓ = 2 m + 1 1 ℓ ! ∑ i 1 + ⋯ + i ℓ = m + 1 i j ≥ 1 s i 1 ⋯ s i ℓ . (282) s'_m=(2m+3)s_{m+1}+\sum_{\ell=2}^{m+1}\frac{1}{\ell!}\sum_{\substack{i_1+\cdots+i_\ell=m+1\\i_j\geq1}}s_{i_1}\cdots s_{i_\ell}. \tag*{(282)} s m ′ = ( 2 m + 3 ) s m + 1 + ℓ = 2 ∑ m + 1 ℓ ! 1 i 1 + ⋯ + i ℓ = m + 1 i j ≥ 1 ∑ s i 1 ⋯ s i ℓ . (282)
The initial sinc product and the same analytic argument imply that every Taylor coefficient in Y Y Y of
ϑ ℓ ( Y ) : = a ( Y ) [ z ℓ ] K Y ( z ) − 1 \vartheta_\ell(Y):=a(Y)[z^\ell]K_Y(z)^{-1} ϑ ℓ ( Y ) := a ( Y ) [ z ℓ ] K Y ( z ) − 1
is nonnegative. The product for a a a implies the same for a − 1 a^{-1} a − 1 and for ϑ ℓ = [ z ℓ ] K Y − 1 \vartheta_\ell=[z^\ell]K_Y^{-1} ϑ ℓ = [ z ℓ ] K Y − 1 . This completes the proof.
Proof of Lemma 3.1
We rewrite the equation for the stationary root in a normalized form that is strictly monotone in T T T ; the implicit function theorem then gives analyticity. The remaining assertions follow from integral formulas for I I I , J J J and β \beta β and from expansions at the two endpoints of the parameter interval.
We use from Lemma 2.1 that X > 0 X>0 X > 0 , X ′ ′ < 0 X''<0 X ′′ < 0 and X ′ ′ ′ > 0 X'''>0 X ′′′ > 0 on the relevant interval, and that 0 < y 0 < 3 / 2 0<y_0<3/2 0 < y 0 < 3/2 . The four parts of the lemma are proved in order.
Part (i): the root T T T . Normalize the left side of (2) by 2 T − Y 2\sqrt{T-Y} 2 T − Y . Its root equation becomes
Q ( Y , T ) = ∫ 0 1 X ′ ( T ( 1 − s 2 ) + Y s 2 ) d s = 0. (283) Q(Y,T)=\int_0^1 X'\bigl(T(1-s^2)+Ys^2\bigr)\,ds=0. \tag*{(283)} Q ( Y , T ) = ∫ 0 1 X ′ ( T ( 1 − s 2 ) + Y s 2 ) d s = 0. (283)
It is strictly decreasing in T T T . At T = y 0 T=y_0 T = y 0 it is positive when Y < y 0 Y<y_0 Y < y 0 . The series and the beta integral give the Sonine identity
∫ 0 T X ′ ( v ) T − v d v = sin ( 2 T ) . (284) \int_0^T\frac{X'(v)}{\sqrt{T-v}}\,dv=\sin(2\sqrt{T}). \tag*{(284)} ∫ 0 T T − v X ′ ( v ) d v = sin ( 2 T ) . (284)
At T 0 = π 2 / 4 T_0=\pi^2/4 T 0 = π 2 /4 , the full integral is zero, and removing [ 0 , Y ] [0,Y] [ 0 , Y ] makes it negative when Y > 0 Y>0 Y > 0 . Since Q > 0 Q>0 Q > 0 for T ≤ y 0 T\leq y_0 T ≤ y 0 and Q Q Q is strictly decreasing in T T T , this proves existence and uniqueness of the root of (2) in ( y 0 , π 2 / 4 ] (y_0,\pi^2/4] ( y 0 , π 2 /4 ] , including T ( 0 ) = T 0 T(0)=T_0 T ( 0 ) = T 0 . The derivative Q T Q_T Q T never vanishes there, so the implicit function theorem gives analyticity. Also T ′ < 0 T'<0 T ′ < 0 , and T → y 0 T\to y_0 T → y 0 as Y → y 0 Y\to y_0 Y → y 0 .
Twice differentiating (B.5), with v = T ( 1 − s 2 ) + Y s 2 v=T(1-s^2)+Ys^2 v = T ( 1 − s 2 ) + Y s 2 , gives
T ′ ′ ( Y ) = − ∫ 0 1 X ′ ′ ′ ( v ) [ T ′ ( 1 − s 2 ) + s 2 ] 2 d s ∫ 0 1 X ′ ′ ( v ) ( 1 − s 2 ) d s > 0. T''(Y)=-\frac{\int_0^1X'''(v)[T'(1-s^2)+s^2]^2\,ds}{\int_0^1X''(v)(1-s^2)\,ds}>0. T ′′ ( Y ) = − ∫ 0 1 X ′′ ( v ) ( 1 − s 2 ) d s ∫ 0 1 X ′′′ ( v ) [ T ′ ( 1 − s 2 ) + s 2 ] 2 d s > 0.
At ( Y , T ) = ( 0 , T 0 ) (Y,T)=(0,T_0) ( Y , T ) = ( 0 , T 0 ) , (B.6) gives Q T = − 1 / ( 2 T 0 ) Q_T=-1/(2T_0) Q T = − 1/ ( 2 T 0 ) , and differentiation of the lower limit Y Y Y in
∫ Y T X ′ ( v ) ( T − v ) − 1 / 2 d v = 2 T − Y Q \int_Y^T X'(v)(T-v)^{-1/2}\,dv=2\sqrt{T-Y}\,Q ∫ Y T X ′ ( v ) ( T − v ) − 1/2 d v = 2 T − Y Q
gives Q Y = − X ′ ( 0 ) / ( 2 T 0 ) = − 1 / ( 2 T 0 ) Q_Y=-X'(0)/(2T_0)=-1/(2T_0) Q Y = − X ′ ( 0 ) / ( 2 T 0 ) = − 1/ ( 2 T 0 ) ; hence T ′ ( 0 ) = − 1 T'(0)=-1 T ′ ( 0 ) = − 1 . Consequently − 1 ≤ T ′ < 0 -1\leq T'<0 − 1 ≤ T ′ < 0 .
The integral I I I is positive. Indeed, subtracting ( T − y 0 ) (T-y_0) ( T − y 0 ) times (2) from its definition gives
I ( Y ) = ∫ Y T X ′ ( v ) ( y 0 − v ) T − v d v > 0. (285) I(Y)=\int_Y^T\frac{X'(v)(y_0-v)}{\sqrt{T-v}}\,dv>0. \tag*{(285)} I ( Y ) = ∫ Y T T − v X ′ ( v ) ( y 0 − v ) d v > 0. (285)
The integrand is nonnegative and not identically zero, because X ′ X' X ′ is decreasing on [ 0 , π 2 / 4 ] [0,\pi^2/4] [ 0 , π 2 /4 ] and vanishes at y 0 y_0 y 0 .
Part (ii): the density λ \lambda λ . Introduce J J J and β \beta β as in (3.1). The equality for J J J is integration by parts, and the last equality for β \beta β follows from the definition of λ \lambda λ . If Q 2 = ∫ 0 1 s 2 X ′ ( T − Δ s 2 ) d s Q_2=\int_0^1s^2X'(T-\Delta s^2)\,ds Q 2 = ∫ 0 1 s 2 X ′ ( T − Δ s 2 ) d s , then Q 2 > 0 Q_2>0 Q 2 > 0 : the function X ′ ( T − Δ s 2 ) X'(T-\Delta s^2) X ′ ( T − Δ s 2 ) is strictly increasing in s s s and has mean zero by (B.5). Moreover Q 2 ≤ X ′ ( Y ) / 3 Q_2\le X'(Y)/3 Q 2 ≤ X ′ ( Y ) /3 . Differentiating J J J gives
J ′ = ( 1 − T ′ ) Q 2 ≤ 2 3 X ′ ( Y ) , β ′ = X ′ J − X J ′ J 2 ≥ X ′ 3 J > 0. (286) J'=(1-T')Q_2\le\frac{2}{3}X'(Y),\qquad\beta'=\frac{X'J-XJ'}{J^2}\ge\frac{X'}{3J}>0. \tag*{(286)} J ′ = ( 1 − T ′ ) Q 2 ≤ 3 2 X ′ ( Y ) , β ′ = J 2 X ′ J − X J ′ ≥ 3 J X ′ > 0. (286)
At zero, integrating (B.6) with respect to T T T gives
∫ 0 T X ′ ( v ) T − v d v = 1 4 ( sin 2 T − 2 T cos 2 T ) , \int_0^T X'(v)\sqrt{T-v}\,dv=\frac{1}{4}\left(\sin2\sqrt{T}-2\sqrt{T}\cos2\sqrt{T}\right), ∫ 0 T X ′ ( v ) T − v d v = 4 1 ( sin 2 T − 2 T cos 2 T ) ,
whose value at T 0 = π 2 / 4 T_0=\pi^2/4 T 0 = π 2 /4 is I ( 0 ) = π / 4 I(0)=\pi/4 I ( 0 ) = π /4 ; then J ( 0 ) = 0 + ( π / 4 ) / ( π / 2 ) = 1 / 2 J(0)=0+(\pi/4)/(\pi/2)=1/2 J ( 0 ) = 0 + ( π /4 ) / ( π /2 ) = 1/2 , and λ ( Y ) = 4 Y + O ( Y 2 ) \lambda(Y)=4Y+O(Y^2) λ ( Y ) = 4 Y + O ( Y 2 ) . At the other endpoint J → x c J\to x_c J → x c and X ( Y ) → x c X(Y)\to x_c X ( Y ) → x c , so β → 1 \beta\to1 β → 1 . This proves the asserted analytic bijection and the values in (7) at zero.
Part (iii): expansions at the critical endpoint. The endpoint constants are explicit. Put δ = y 0 − Y \delta=y_0-Y δ = y 0 − Y and b = − X ′ ′ ( y 0 ) = x c / y 0 > 0 b=-X''(y_0)=x_c/y_0>0 b = − X ′′ ( y 0 ) = x c / y 0 > 0 . The implicit equation at ( y 0 , y 0 ) (y_0,y_0) ( y 0 , y 0 ) has T ′ = − 1 / 2 T'=-1/2 T ′ = − 1/2 ; this gives the first formula in (17) . In particular T − y 0 = δ / 2 + O ( δ 2 ) = Δ / 3 + O ( Δ 2 ) T-y_0=\delta/2+O(\delta^2)=\Delta/3+O(\Delta^2) T − y 0 = δ /2 + O ( δ 2 ) = Δ/3 + O ( Δ 2 ) , so that v − y 0 = Δ / 3 − Δ s 2 + O ( Δ 2 ) v-y_0=\Delta/3-\Delta s^2+O(\Delta^2) v − y 0 = Δ/3 − Δ s 2 + O ( Δ 2 ) for v = T − Δ s 2 v=T-\Delta s^2 v = T − Δ s 2 . For the second formula, substitute X ′ ( T − Δ s 2 ) = − b ( v − y 0 ) + O ( Δ 2 ) = b Δ ( s 2 − 1 / 3 ) + O ( Δ 2 ) X'(T-\Delta s^2)=-b(v-y_0)+O(\Delta^2)=b\Delta(s^2-1/3)+O(\Delta^2) X ′ ( T − Δ s 2 ) = − b ( v − y 0 ) + O ( Δ 2 ) = b Δ ( s 2 − 1/3 ) + O ( Δ 2 ) into I = 2 Δ 3 / 2 ∫ 0 1 s 2 X ′ ( T − Δ s 2 ) d s I=2\Delta^{3/2}\int_0^1s^2X'(T-\Delta s^2)\,ds I = 2 Δ 3/2 ∫ 0 1 s 2 X ′ ( T − Δ s 2 ) d s . Together with X ( Y ) → x c X(Y)\to x_c X ( Y ) → x c , this proves (18) , and (19) follows from (7) . Since 1 − β = 2 / ( 2 + λ ) 1-\beta=2/(2+\lambda) 1 − β = 2/ ( 2 + λ ) , (18) also gives Δ ≍ 1 − β \Delta\asymp\sqrt{1-\beta} Δ ≍ 1 − β .
Part (iv): the function Ψ \Psi Ψ . By part (ii), Ψ \Psi Ψ is analytic and strictly increasing for 0 < x < x c 0<x<x_c 0 < x < x c . The first line of (20) follows from λ ( Y ) = 4 Y + O ( Y 2 ) \lambda(Y)=4Y+O(Y^2) λ ( Y ) = 4 Y + O ( Y 2 ) and x = X ( Y ) = Y + O ( Y 2 ) x=X(Y)=Y+O(Y^2) x = X ( Y ) = Y + O ( Y 2 ) . For the second line, quadratic inversion gives x c − x = ( b / 2 ) δ 2 ( 1 + O ( δ ) ) x_c-x=(b/2)\delta^2(1+O(\delta)) x c − x = ( b /2 ) δ 2 ( 1 + O ( δ )) . Combining this with (17) and (18) gives the constant 5 / 4 5/4 5/4 . All expansions here arise from convergent analytic expansions in x x x or ε \sqrt{\varepsilon} ε , so their asserted derivatives follow by termwise differentiation. In particular Ψ 2 / ( D Ψ ) \Psi^2/(\mathcal D\Psi) Ψ 2 / ( D Ψ ) has finite endpoint limits 0 0 0 and 5 / 4 5/4 5/4 , and is bounded on the full interval.
Uniform bounds for positive intersection numbers
This appendix proves Lemma C.1 , the two-sided bound 1 / 6 ≤ C g ( e ) ≤ 1 1/6\le C_g(\mathbf e)\le1 1/6 ≤ C g ( e ) ≤ 1 for the normalized pure-ψ \psi ψ intersection numbers of (23) . The bound is the only intersection estimate that the harmonic comparison of § 3.7 imports: it enters the terminal bounds of Lemma 5.2, and through them Proposition 5.1. The lemma is a consequence of the stronger bounds in [83, Lemmas 2–6]; the proof below uses their separating-term estimate, which we reproduce, and replaces their two-variable majorant by a scalar one, so that a single induction on the complexity h h h gives both bounds. The only external inputs are the classical one-point formula ⟨ τ 3 g − 2 ⟩ g = 1 / ( 24 g g ! ) \langle\tau_{3g-2}\rangle_g=1/(24^g g!) ⟨ τ 3 g − 2 ⟩ g = 1/ ( 2 4 g g !) and the two-point bound of Delecroix, Goujard, Zograf and Zorich, proved in [52] and quoted as [51, equation (6)]. After the statement we record the normalized string, dilaton and DVV identities, outline the induction, and then give the proof in five marked parts.
Lemma C.1. For g ≥ 1 g\ge1 g ≥ 1 and positive integers d 1 , … , d n d_1,\ldots,d_n d 1 , … , d n with ∑ d i = 3 g − 3 + n \sum d_i=3g-3+n ∑ d i = 3 g − 3 + n , the normalization in (23) satisfies
1 6 ≤ C g ( d 1 , … , d n ) ≤ 1. \frac{1}{6}\le C_g(d_1,\ldots,d_n)\le1. 6 1 ≤ C g ( d 1 , … , d n ) ≤ 1.
Write C ( d ) = C g ( d ) C(\mathbf d)=C_g(\mathbf d) C ( d ) = C g ( d ) and h = 2 g − 2 + n h=2g-2+n h = 2 g − 2 + n when the genus is determined by the degrees. The normalized string and dilaton identities are
C ( 0 , d ) = ∑ d j > 0 2 d j + 1 3 ( Z − 1 ) C ( d − e j ) , C ( 1 , d ) = C ( d ) , (287) C(0,\mathbf d)=\sum_{d_j>0}\frac{2d_j+1}{3(Z-1)}C(\mathbf d-\mathbf e_j),\qquad C(1,\mathbf d)=C(\mathbf d), \tag*{(287)} C ( 0 , d ) = d j > 0 ∑ 3 ( Z − 1 ) 2 d j + 1 C ( d − e j ) , C ( 1 , d ) = C ( d ) , (287)
where Z Z Z is the complexity of ( 0 , d ) (0,\mathbf{d}) ( 0 , d ) ; the terminal value C 1 ( 1 ) = 1 / 6 C_1(1)=1/6 C 1 ( 1 ) = 1/6 is kept separately. For a pivot d 1 ≥ 2 d_1 \ge2 d 1 ≥ 2 , DVV gives
C ( d ) = ∑ j = 2 n 2 d j + 1 3 ( h − 1 ) C ( d 2 , … , d j + d 1 − 1 , … , d n ) + 2 3 ( h − 1 ) ∑ a + b = d 1 − 2 C g − 1 ( a , b , d 2 , … , d n ) + ∑ a + b = d 1 − 2 I ⊔ J = { 2 , … , n } Γ ( h 1 ) Γ ( h 2 ) 6 Γ ( h ) C ( a , d I ) C ( b , d J ) . (288) \begin{aligned}
C(\mathbf{d})={}&\sum_{j=2}^{n}\frac{2d_j+1}{3(h-1)}C(d_2,\ldots,d_j+d_1-1,\ldots,d_n)\\
&+\frac{2}{3(h-1)}\sum_{a+b=d_1-2}C_{g-1}(a,b,d_2,\ldots,d_n)\\
&+\sum_{\substack{a+b=d_1-2\\I\sqcup J=\{2,\ldots,n\}}}
\frac{\Gamma(h_1)\Gamma(h_2)}{6\Gamma(h)}C(a,d_I)C(b,d_J). \tag*{(288)}
\end{aligned} C ( d ) = j = 2 ∑ n 3 ( h − 1 ) 2 d j + 1 C ( d 2 , … , d j + d 1 − 1 , … , d n ) + 3 ( h − 1 ) 2 a + b = d 1 − 2 ∑ C g − 1 ( a , b , d 2 , … , d n ) + a + b = d 1 − 2 I ⊔ J = { 2 , … , n } ∑ 6Γ ( h ) Γ ( h 1 ) Γ ( h 2 ) C ( a , d I ) C ( b , d J ) . (288)
Here h 1 , h 2 h_1,h_2 h 1 , h 2 are the complexities of the two factors, and only stable, dimension-compatible terms are included. The weights in the first two lines sum to one.
Both bounds are proved by induction on complexity. The lower bound uses the DVV recursion with its nonnegative separating terms discarded. For the upper bound we prove the stronger estimate C ( d ) ≤ F ( h ) C(\mathbf{d})\le F(h) C ( d ) ≤ F ( h ) for tuples with positive entries, where F ≤ 1 F\le1 F ≤ 1 is an explicit nondecreasing function. In the induction step the separating terms contribute at most 2 / ( ( h − 1 ) ( h − 2 ) ) 2/((h-1)(h-2)) 2/ (( h − 1 ) ( h − 2 )) ; the cases h ≤ 7 h\le7 h ≤ 7 are treated directly.
Proof of Lemma C.1. The lower bound. We prove the same lower bound for positive-genus tuples with at most two zero entries, by induction on complexity. For a tuple with z ∈ { 1 , 2 } z\in\{1,2\} z ∈ { 1 , 2 } zeros, the total string weight is ( 3 h − z ) / ( 3 h − 3 ) ≥ 1 (3h-z)/(3h-3)\ge1 ( 3 h − z ) / ( 3 h − 3 ) ≥ 1 . String lowers complexity without increasing the number of zeros. For a positive tuple, remove the ones by dilaton. Unless the tuple reduces to the genus-one base, its genus is at least two. Its merger children are positive and its handle children have positive genus and at most two zeros. Dropping the nonnegative separating terms in (C.2) therefore proves the lower bound.
The upper majorant and zero entries. Define the nondecreasing function
F ( h ) = { 1 / 3 , 1 ≤ h ≤ 7 , 1 − 4 / ( h − 1 ) , h ≥ 7. F(h)=
\begin{cases}
1/3, & 1\le h\le7,\\
1-4/(h-1), & h\ge7.
\end{cases} F ( h ) = { 1/3 , 1 − 4/ ( h − 1 ) , 1 ≤ h ≤ 7 , h ≥ 7.
We will prove C ( d ) ≤ F ( h ) C(\mathbf{d})\le F(h) C ( d ) ≤ F ( h ) for positive tuples. Its increment for h ≥ 8 h\ge8 h ≥ 8 is 4 / ( ( h − 1 ) ( h − 2 ) ) 4/((h-1)(h-2)) 4/ (( h − 1 ) ( h − 2 )) . The corresponding bound for a tuple with one zero is C ≤ 1 C\le1 C ≤ 1 : remove all ones, and let Z Z Z be the remaining complexity. String bounds the result by
( 1 + 2 3 ( Z − 1 ) ) F ( Z − 1 ) . \left(1+\frac{2}{3(Z-1)}\right)F(Z-1). ( 1 + 3 ( Z − 1 ) 2 ) F ( Z − 1 ) .
For Z ≤ 8 Z\le8 Z ≤ 8 this is at most 5 / 9 5/9 5/9 ; for Z ≥ 9 Z\ge9 Z ≥ 9 , substitution of F ( Z − 1 ) = 1 − 4 / ( Z − 2 ) F(Z-1)=1-4/(Z-2) F ( Z − 1 ) = 1 − 4/ ( Z − 2 ) gives a bound below one. This observation uses only positive tuples of smaller complexity and so is available in the induction.
The separating contribution. Suppose all d i ≥ 2 d_i\ge2 d i ≥ 2 . Both separating genera are positive by dimension. For a fixed subset I I I of size m m m and genus g 1 g_1 g 1 , the degree a a a is determined. The total separating weight is thus at most
∑ m = 0 n − 1 ( n − 1 m ) ∑ g 1 + g 2 = g g 1 , g 2 ≥ 1 ( 2 g 1 + m − 2 ) ! ( 2 g 2 + n − m − 3 ) ! 6 ( h − 1 ) ! . \sum_{m=0}^{n-1}\binom{n-1}{m}
\sum_{\substack{g_1+g_2=g\\g_1,g_2\ge1}}
\frac{(2g_1+m-2)!(2g_2+n-m-3)!}{6(h-1)!}. m = 0 ∑ n − 1 ( m n − 1 ) g 1 + g 2 = g g 1 , g 2 ≥ 1 ∑ 6 ( h − 1 )! ( 2 g 1 + m − 2 )! ( 2 g 2 + n − m − 3 )! .
As in [83, Lemma 3], the terms with at least one genus equal to one, allowing double counting, are bounded by
1 3 ( h − 1 ) ( h − 2 ) ∑ m = 0 n − 1 ( n − 1 m ) ∏ j = 1 m n − j h − 2 − j ≤ 3 2 ( h − 1 ) ( h − 2 ) . \frac{1}{3(h-1)(h-2)}
\sum_{m=0}^{n-1}\binom{n-1}{m}
\prod_{j=1}^{m}\frac{n-j}{h-2-j}
\le\frac{3}{2(h-1)(h-2)}. 3 ( h − 1 ) ( h − 2 ) 1 m = 0 ∑ n − 1 ( m n − 1 ) j = 1 ∏ m h − 2 − j n − j ≤ 2 ( h − 1 ) ( h − 2 ) 3 .
Indeed n ≤ 3 h / 5 n \le3h/5 n ≤ 3 h /5 and n ≤ h − 2 n \le h - 2 n ≤ h − 2 . The terms with m ≤ 2 m \le2 m ≤ 2 are at most one, and for m ≥ 3 m \ge3 m ≥ 3 each factor with j ≥ 3 j \ge3 j ≥ 3 is at most 3 / 5 3/5 3/5 , so the sum is at most 1 + 1 + ∑ m ≥ 2 ( 3 / 5 ) m − 2 = 9 / 2 1 + 1 + \sum_{m \ge2}(3/5)^{m-2} = 9/2 1 + 1 + ∑ m ≥ 2 ( 3/5 ) m − 2 = 9/2 . For g 1 , g 2 ≥ 2 g_1,g_2 \ge2 g 1 , g 2 ≥ 2 , the inequality ( a b ) ( c d ) ≤ ( a + c b + d ) \binom{a}{b}\binom{c}{d} \le\binom{a+c}{b+d} ( b a ) ( d c ) ≤ ( b + d a + c ) gives the bound
n ( g − 3 ) 6 ( h − 1 ) ( h − 2 ) ( 2 g − 4 2 ) ≤ 3 8 ( h − 1 ) ( h − 2 ) . \frac{n(g-3)}{6(h-1)(h-2)\binom{2g-4}{2}} \le\frac{3}{8(h-1)(h-2)}. 6 ( h − 1 ) ( h − 2 ) ( 2 2 g − 4 ) n ( g − 3 ) ≤ 8 ( h − 1 ) ( h − 2 ) 3 .
This part vanishes for g ≤ 3 g \le3 g ≤ 3 ; otherwise use n ≤ 3 g − 3 n \le3g - 3 n ≤ 3 g − 3 . The total weight is therefore at most 2 / ( ( h − 1 ) ( h − 2 ) ) 2/((h-1)(h-2)) 2/ (( h − 1 ) ( h − 2 )) . Each factor is positive or has one zero, so the same bound applies to the full separating contribution once the smaller-complexity bounds hold.
Initial cases. Put a g = C g ( 3 g − 2 ) a_g=C_g(3g-2) a g = C g ( 3 g − 2 ) and b g = C g ( 0 , 3 g − 1 ) b_g=C_g(0,3g-1) b g = C g ( 0 , 3 g − 1 ) . The one-point formula and the two-point bound of [52], quoted in [51, equation (6)], give
C g ( d 1 , d 2 ) ≤ b g = 6 g − 1 6 g − 3 a g < 1 π < 8 25 . C_g(d_1,d_2) \le b_g = \frac{6g-1}{6g-3}a_g < \frac{1}{\pi} < \frac{8}{25}. C g ( d 1 , d 2 ) ≤ b g = 6 g − 3 6 g − 1 a g < π 1 < 25 8 .
For completeness, b g + 1 / b g = ( 6 g + 5 ) ( 6 g + 1 ) ( 6 g − 3 ) / ( 108 g ( g + 1 ) ( 2 g − 1 ) ) > 1 b_{g+1}/b_g=(6g+5)(6g+1)(6g-3)/(108g(g+1)(2g-1))>1 b g + 1 / b g = ( 6 g + 5 ) ( 6 g + 1 ) ( 6 g − 3 ) / ( 108 g ( g + 1 ) ( 2 g − 1 )) > 1 , and Stirling’s formula gives b g → 1 / π b_g \to1/\pi b g → 1/ π . All positive genus-two tuples reduce by dilaton to ( 4 ) (4) ( 4 ) , ( 2 , 3 ) (2,3) ( 2 , 3 ) or ( 2 , 2 , 2 ) (2,2,2) ( 2 , 2 , 2 ) . String gives
C 1 ( 0 , 2 ) = 5 18 , C 1 ( 0 , 0 , 3 ) = 35 108 , C 1 ( 0 , 0 , 2 , 2 ) = 25 81 . C_1(0,2)=\frac{5}{18},\qquad C_1(0,0,3)=\frac{35}{108},\qquad C_1(0,0,2,2)=\frac{25}{81}. C 1 ( 0 , 2 ) = 18 5 , C 1 ( 0 , 0 , 3 ) = 108 35 , C 1 ( 0 , 0 , 2 , 2 ) = 81 25 .
Using these values in (C.2) yields
C 2 ( 4 ) = 35 144 , C 2 ( 2 , 3 ) = 1015 3888 , C 2 ( 2 , 2 , 2 ) = 175 648 < 3 11 . C_2(4)=\frac{35}{144},\qquad C_2(2,3)=\frac{1015}{3888},\qquad C_2(2,2,2)=\frac{175}{648}<\frac{3}{11}. C 2 ( 4 ) = 144 35 , C 2 ( 2 , 3 ) = 3888 1015 , C 2 ( 2 , 2 , 2 ) = 648 175 < 11 3 .
The other two values are also below 3 / 11 3/11 3/11 . Hence a genus-two tuple with one zero is at most ( 11 / 9 ) ( 3 / 11 ) = 1 / 3 (11/9)(3/11)=1/3 ( 11/9 ) ( 3/11 ) = 1/3 after removing ones; its remaining complexity is at least four. The genus-one one-zero case reduces to ( 0 , 2 ) (0,2) ( 0 , 2 ) and has the same bound.
For h ≤ 7 h \le7 h ≤ 7 , the only additional positive cores with at least three markings are the genus-three tuples ( 2 , 2 , 5 ) (2,2,5) ( 2 , 2 , 5 ) , ( 2 , 3 , 4 ) (2,3,4) ( 2 , 3 , 4 ) and ( 3 , 3 , 3 ) (3,3,3) ( 3 , 3 , 3 ) . For the first two, two string steps bound the handle term by ( 16 / 15 ) ( 11 / 9 ) ( 3 / 11 ) = 16 / 45 (16/15)(11/9)(3/11)=16/45 ( 16/15 ) ( 11/9 ) ( 3/11 ) = 16/45 . Each separating factor has genus one or two and is at most 1 / 3 1/3 1/3 ; the separating contribution is therefore at most the general separating weight bound 2 / ( ( h − 1 ) ( h − 2 ) ) = 2 / 30 2/((h-1)(h-2))=2/30 2/ (( h − 1 ) ( h − 2 )) = 2/30 at h = 7 h=7 h = 7 , multiplied by the bound 1 / 3 1/3 1/3 for each of the two factors, that is, by 1 / 9 1/9 1/9 . Thus
C ≤ 8 9 8 25 + 8 9 16 45 + 1 9 2 30 = 671 2025 < 1 3 . C \le\frac{8}{9}\frac{8}{25}+\frac{8}{9}\frac{16}{45}+\frac{1}{9}\frac{2}{30}=\frac{671}{2025}<\frac{1}{3}. C ≤ 9 8 25 8 + 9 8 45 16 + 9 1 30 2 = 2025 671 < 3 1 .
For ( 3 , 3 , 3 ) (3,3,3) ( 3 , 3 , 3 ) , both handle terms reduce to ( 0 , 3 , 3 ) (0,3,3) ( 0 , 3 , 3 ) in genus two, so
C ≤ 7 9 8 25 + 8 9 2 3 + 1 9 2 30 = 223 675 < 1 3 . C \le\frac{7}{9}\frac{8}{25}+\frac{8}{9}\frac{2}{3}+\frac{1}{9}\frac{2}{30}=\frac{223}{675}<\frac{1}{3}. C ≤ 9 7 25 8 + 9 8 3 2 + 9 1 30 2 = 675 223 < 3 1 .
Together with dilaton, the one- and two-point bounds and the genus-one base, this proves every initial case h ≤ 7 h \le7 h ≤ 7 .
Induction step. Let d \mathbf{d} d have no entry equal to one (ones having been removed by dilaton) and complexity h ≥ 8 h \ge8 h ≥ 8 ; assume n ≥ 3 n \ge3 n ≥ 3 and choose a minimal degree d = d 1 ≥ 2 d=d_1 \ge2 d = d 1 ≥ 2 . If d ≥ 3 d \ge3 d ≥ 3 , all handle children except the two endpoint terms have positive degrees. Each endpoint has one zero and at most one degree-one entry. Removing that entry if necessary and applying string bounds it by ( 1 + 2 / ( 3 ( h − 3 ) ) ) F ( h − 1 ) (1+2/(3(h-3)))F(h-1) ( 1 + 2/ ( 3 ( h − 3 ))) F ( h − 1 ) . The total endpoint weight is 4 / ( 3 ( h − 1 ) ) 4/(3(h-1)) 4/ ( 3 ( h − 1 )) . Since F ≤ 1 F \le1 F ≤ 1 , the excess over F ( h − 1 ) F(h-1) F ( h − 1 ) is at most
8 9 ( h − 1 ) ( h − 3 ) + 2 ( h − 1 ) ( h − 2 ) ≤ 46 15 ( h − 1 ) ( h − 2 ) < F ( h ) − F ( h − 1 ) . \frac{8}{9(h-1)(h-3)}+\frac{2}{(h-1)(h-2)}\le\frac{46}{15(h-1)(h-2)}<F(h)-F(h-1). 9 ( h − 1 ) ( h − 3 ) 8 + ( h − 1 ) ( h − 2 ) 2 ≤ 15 ( h − 1 ) ( h − 2 ) 46 < F ( h ) − F ( h − 1 ) .
If d = 2 d=2 d = 2 , the handle has two zeros. Two string steps, allowing an intervening dilaton step, give the factor
R h = ( 3 h − 5 ) ( 3 h − 10 ) ( 3 h − 6 ) ( 3 h − 12 ) . R_h=\frac{(3h-5)(3h-10)}{(3h-6)(3h-12)}. R h = ( 3 h − 6 ) ( 3 h − 12 ) ( 3 h − 5 ) ( 3 h − 10 ) .
The resulting positive tuples have complexity at most h − 3 h-3 h − 3 . The handle weight is 2 / ( 3 ( h − 1 ) ) 2/(3(h-1)) 2/ ( 3 ( h − 1 )) , and
2 ( R h − 1 ) 3 ( h − 1 ) = 2 ( 9 h − 22 ) 27 ( h − 4 ) ( h − 1 ) ( h − 2 ) ≤ 25 27 ( h − 1 ) ( h − 2 ) . \frac{2(R_h-1)}{3(h-1)}=\frac{2(9h-22)}{27(h-4)(h-1)(h-2)}\le\frac{25}{27(h-1)(h-2)}. 3 ( h − 1 ) 2 ( R h − 1 ) = 27 ( h − 4 ) ( h − 1 ) ( h − 2 ) 2 ( 9 h − 22 ) ≤ 27 ( h − 1 ) ( h − 2 ) 25 .
Adding separation gives 79 / ( 27 ( h − 1 ) ( h − 2 ) ) 79/(27(h-1)(h-2)) 79/ ( 27 ( h − 1 ) ( h − 2 )) , again below the increment of F F F . This closes the induction and proves the upper bound. □
Appendix D. The marked expansion beyond the leading term
This appendix proves Theorems 8.1 8.1 8.1 and 8.2 8.2 8.2 , the higher-order refinements of Proposition 7.1 7.1 7.1 stated in § 3.22 . The final marked asymptotic is stated for fixed p p p ; the intermediate recursive estimates also control larger marking ranges. The strategy is the one of § 3.19 carried to higher order: the adjacent-degree subtraction in the recursion (28) expands the normalized profiles c g , p c_{g,p} c g , p in powers of 1 / h 1/h 1/ h ; the dilaton and KdV equations convert these into expansions of the cusp and genus ratios and hence of the logarithmic derivative in Y Y Y ; and the integration constants are fixed at the pure-ψ \psi ψ endpoint, where Painlevé I supplies the asymptotics of ⟨ τ 2 3 g − 3 ⟩ g \langle\tau_2^{3g-3}\rangle_g ⟨ τ 2 3 g − 3 ⟩ g . The inputs are the marked ratios of Proposition 6.2 6.2 6.2 together with the estimates (6.6)–(6.19) from its proof, the leading term of Proposition 7.1 7.1 7.1 , the Itzykson–Zuber identity [106], and the Painlevé I asymptotics in the form of [70] and [48]. Subsections D.1–D.3 compute one further term; the coefficient to be integrated vanishes at the endpoint, so its primitive stays bounded, and this gives both the O p ( g − 1 ) O_p(g^{-1}) O p ( g − 1 ) relative error for the leading term and its first explicit correction (Theorem 8.1 8.1 8.1 ). Subsections D.4–D.8 construct the full formal series and prove polynomial bounds on its coefficients and remainders (Proposition D .3 D.3 D .3 ); Subsections D.9–D.10 sharpen these to single-factorial bounds (Proposition D .9 D.9 D .9 ); and Subsection D.11 determines the endpoint constants to all orders and completes the proof of Theorem 8.2 8.2 8.2 .
D.1. The second degree coefficient. We first extend the degree expansion (90) of the normalized profiles by one order (Lemma D .1 D.1 D .1 ). The argument is the adjacent-degree subtraction used for (90) in § 3.19 , now keeping the h − 2 h^{-2} h − 2 terms of the prefactors and the two smallest separating configurations.
We retain δ p \delta_p δ p and β p \beta_p β p from (90) and (92) . The kernel coefficients θ ℓ \theta_\ell θ ℓ are those of (21) . Set
α p = A { ( p − 3 ) Θ + 2 M 1 + 2 p − 3 } , T j = ∑ ℓ ≥ max ( 0 , j ) w ℓ . (289) \alpha_p=A\{(p-3)\Theta+2M_1+2p-3\},\qquad\mathcal{T}_j=\sum_{\ell\ge\max(0,j)}w_\ell. \tag*{(289)} α p = A {( p − 3 ) Θ + 2 M 1 + 2 p − 3 } , T j = ℓ ≥ m a x ( 0 , j ) ∑ w ℓ . (289)
For a pivot i i i , put t = d i t=d_i t = d i and let z z z count zero entries among the other p − 1 p-1 p − 1 markings. Let d i j , ℓ \mathbf d^{ij,\ell} d ij , ℓ be the merger child, with new degree t + d j + ℓ − 1 t+d_j+\ell-1 t + d j + ℓ − 1 , and let d i ; u , v \mathbf d^{i;u,v} d i ; u , v be the nonseparating child with the pivot replaced by u , v u,v u , v . Define
B p , i ( d ) = A ∑ j ≠ i ( 2 d j + 1 ) ∑ ℓ ≥ max ( 0 , 1 − t − d j ) w ℓ [ 1 − β p − 1 − δ p − 1 ( d i j , ℓ ) ] + 2 A ∑ ℓ ≥ max ( 0 , 2 − t ) w ℓ ∑ u + v = t + ℓ − 2 u , v ≥ 0 [ 1 − β p + 1 − α p − δ p + 1 ( d i ; u , v ) ] + A 2 { [ ( z 2 ) + θ 1 2 ] T 2 − t + 1 2 T 3 − t } . (290) \begin{aligned}
B_{p,i}(\mathbf d)={}&A\sum_{j\ne i}(2d_j+1)\sum_{\ell\ge\max(0,1-t-d_j)}w_\ell[1-\beta_{p-1}-\delta_{p-1}(\mathbf d^{ij,\ell})]\\
&+2A\sum_{\ell\ge\max(0,2-t)}w_\ell\sum_{\substack{u+v=t+\ell-2\\u,v\ge0}}[1-\beta_{p+1}-\alpha_p-\delta_{p+1}(\mathbf d^{i;u,v})]\\
&+A^2\left\{\left[\binom{z}{2}+\frac{\theta_1}{2}\right]\mathcal{T}_{2-t}+\frac{1}{2}\mathcal{T}_{3-t}\right\}. \tag*{(290)}
\end{aligned} B p , i ( d ) = A j = i ∑ ( 2 d j + 1 ) ℓ ≥ m a x ( 0 , 1 − t − d j ) ∑ w ℓ [ 1 − β p − 1 − δ p − 1 ( d ij , ℓ )] + 2 A ℓ ≥ m a x ( 0 , 2 − t ) ∑ w ℓ u + v = t + ℓ − 2 u , v ≥ 0 ∑ [ 1 − β p + 1 − α p − δ p + 1 ( d i ; u , v )] + A 2 { [ ( 2 z ) + 2 θ 1 ] T 2 − t + 2 1 T 3 − t } . (290)
The first line is absent at p = 1 p=1 p = 1 . Starting at zero along a fixed coordinate-increment path, define e p e_p e p by
e p ( 0 ) = 0 , e p ( d + e i ) − e p ( d ) = − B p , i ( d ) . (291) e_p(\mathbf0)=0,\qquad e_p(\mathbf d+\mathbf e_i)-e_p(\mathbf d)=-B_{p,i}(\mathbf d). \tag*{(291)} e p ( 0 ) = 0 , e p ( d + e i ) − e p ( d ) = − B p , i ( d ) . (291)
The formulas converge absolutely by (65) ; they give B p , i = O p ( ( D + 1 ) 3 ) B_{p,i}=O_p((D+1)^3) B p , i = O p (( D + 1 ) 3 ) and e p = O p ( ( D + 1 ) 4 ) e_p=O_p((D+1)^4) e p = O p (( D + 1 ) 4 ) .
Lemma D.1 (Second profile correction). For fixed p ≥ 1 p \ge1 p ≥ 1 , uniformly in Y Y Y and D ≤ h 1 / 8 D \le h^{1/8} D ≤ h 1/8 ,
c g , p ( d ) = 1 − δ p ( d ) h + e p ( d ) h 2 + O p ( ( D + 1 ) 6 h 3 ) . (292) c_{g,p}(\mathbf{d}) = 1 - \frac{\delta_p(\mathbf{d})}{h} + \frac{e_p(\mathbf{d})}{h^2} + O_p\left(\frac{(D+1)^6}{h^3}\right). \tag*{(292)} c g , p ( d ) = 1 − h δ p ( d ) + h 2 e p ( d ) + O p ( h 3 ( D + 1 ) 6 ) . (292) The function e p e_p e p is independent of the chosen path.
Proof. We first need the linear term in the genus exchange. For p ≥ 2 p \ge2 p ≥ 2 , the two genus-zero three-marking contributions to (73) give T g , p = 2 ( p − 2 ) A / h + O p ( h − 2 ) T_{g,p}=2(p-2)A/h+O_p(h^{-2}) T g , p = 2 ( p − 2 ) A / h + O p ( h − 2 ) . Every other genus-zero term loses at least two cusp factors, and the positive-genus portion is O p ( h − 2 ) O_p(h^{-2}) O p ( h − 2 ) by (74) . Thus (69) and (90) imply
W g − 1 , p + 2 W g , p = 1 − α p h + O p ( h − 2 ) . (293) \frac{W_{g-1,p+2}}{W_{g,p}} = 1 - \frac{\alpha_p}{h} + O_p(h^{-2}). \tag*{(293)} W g , p W g − 1 , p + 2 = 1 − h α p + O p ( h − 2 ) . (293) Indeed the coefficient is δ p ( e 1 ) + 3 ( p − 2 ) A = α p \delta_p(\mathbf{e}_1)+3(p-2)A=\alpha_p δ p ( e 1 ) + 3 ( p − 2 ) A = α p . For p = 0 , 1 p=0,1 p = 0 , 1 , descend from p + 1 p+1 p + 1 using (92) ; the cusp ratio changes its first coefficient by β p + 2 − β p = − A ( Θ + 2 ) \beta_{p+2}-\beta_p=-A(\Theta+2) β p + 2 − β p = − A ( Θ + 2 ) , while α p + 1 − α p = A ( Θ + 2 ) \alpha_{p+1}-\alpha_p=A(\Theta+2) α p + 1 − α p = A ( Θ + 2 ) . Consequently the two adjacent-difference prefactors are
W g , p − 1 W g , p = A h ( 1 + 1 − β p − 1 h ) + O p ( h − 3 ) , \frac{W_{g,p-1}}{W_{g,p}}=\frac{A}{h}\left(1+\frac{1-\beta_{p-1}}{h}\right)+O_p(h^{-3}), W g , p W g , p − 1 = h A ( 1 + h 1 − β p − 1 ) + O p ( h − 3 ) , 2 W g − 1 , p + 1 W g , p = 2 A h ( 1 + 1 − β p + 1 − α p h ) + O p ( h − 3 ) . (294) 2\frac{W_{g-1,p+1}}{W_{g,p}}=\frac{2A}{h}\left(1+\frac{1-\beta_{p+1}-\alpha_p}{h}\right)+O_p(h^{-3}). \tag*{(294)} 2 W g , p W g − 1 , p + 1 = h 2 A ( 1 + h 1 − β p + 1 − α p ) + O p ( h − 3 ) . (294) The first line is used only at p ≥ 2 p \ge2 p ≥ 2 . The merger and nonseparating children both have complexity h − 1 h-1 h − 1 , which accounts for the two constants + 1 +1 + 1 .
We next retain the leading separating terms of the adjacent difference. If the smaller component has genus j j j and contains s s s of the other external labels, the complementary component has complexity h − ( 2 j + s ) h-(2j+s) h − ( 2 j + s ) . Stability forces 2 j + s ≥ 2 2j+s \ge2 2 j + s ≥ 2 . Equality occurs only for ( j , s ) = ( 0 , 2 ) (j,s)=(0,2) ( j , s ) = ( 0 , 2 ) and ( 1 , 0 ) (1,0) ( 1 , 0 ) . The first component is a three-marked sphere, so its three degrees must be zero; it gives ( z 2 ) T 2 − t \binom{z}{2}T_{2-t} ( 2 z ) T 2 − t . The second is a one-marked torus, whose W W W amplitudes are θ 1 / 2 \theta_1/2 θ 1 /2 and 1 / 2 1/2 1/2 at degrees zero and one. It gives ( θ 1 / 2 ) T 2 − t + ( 1 / 2 ) T 3 − t (\theta_1/2)T_{2-t}+(1/2)T_{3-t} ( θ 1 /2 ) T 2 − t + ( 1/2 ) T 3 − t . The two orientations cancel the factor 1 / 2 1/2 1/2 in the recursion. In both cases the large all-zero component divided by the parent is A 2 / h 2 + O p ( h − 3 ) A^2/h^2+O_p(h^{-3}) A 2 / h 2 + O p ( h − 3 ) , by fixed iterations of (92) and (D.5). Its degrees can be replaced by zero at cost O p ( ( D + ℓ + 1 ) 2 / h ) O_p((D+\ell+1)^2/h) O p (( D + ℓ + 1 ) 2 / h ) .
After removing these two types the all-zero separating mass is O p ( h − 3 ) O_p(h^{-3}) O p ( h − 3 ) . To verify this sharper bound, repeat the proof of (75) : the remaining genus-zero terms have s ≥ 3 s \ge3 s ≥ 3 , the remaining genus-one terms have s ≥ 1 s \ge1 s ≥ 1 , and j ≥ 2 j \ge2 j ≥ 2 starts at h − 4 h^{-4} h − 4 . The same factorial-geometric series sums these bounds, with any fixed polynomial in j j j ; (76) handles the macroscopic splits. At p = 1 p=1 p = 1 , insert 2 j + 1 2j+1 2 j + 1 cusps in the larger component before raising its genus to marking count two, then use W g , 2 / W g , 1 = O ( g ) W_{g,2}/W_{g,1}=O(g) W g , 2 / W g , 1 = O ( g ) . This bounds its ratio by O ( ( C / g ) 2 j ) O((C/g)^{2j}) O (( C / g ) 2 j ) . After the two genus-one orientations are removed, the factorial series starts at j = 2 j=2 j = 2 and is O ( h − 4 ) O(h^{-4}) O ( h − 4 ) . This argument uses the marking-independent constant in (71) . Pair counts and the moments of w ℓ w_\ell w ℓ give an adjacent-difference error O p ( ( D + 1 ) 3 / h 3 ) O_p((D+1)^3/h^3) O p (( D + 1 ) 3 / h 3 ) for the separating part.
Now substitute (90) and (D.6) into the exact adjacent subtraction of (28) . The two retained separating types give the last line of (D.2). The result is
c ( d ) − c ( d + e i ) = δ p ( d + e i ) − δ p ( d ) h + B p , i ( d ) h 2 + O p ( ( D + 1 ) 5 / h 3 ) . c(\mathbf{d})-c(\mathbf{d}+\mathbf{e}_i)=\frac{\delta_p(\mathbf{d}+\mathbf{e}_i)-\delta_p(\mathbf{d})}{h}+\frac{B_{p,i}(\mathbf{d})}{h^2}+O_p((D+1)^5/h^3). c ( d ) − c ( d + e i ) = h δ p ( d + e i ) − δ p ( d ) + h 2 B p , i ( d ) + O p (( D + 1 ) 5 / h 3 ) . To justify the substitution, cut the kernel index ℓ \ell ℓ at h 1 / 8 h^{1/8} h 1/8 . Below this cutoff, child degrees are at most 2 h 1 / 8 + O p ( 1 ) 2h^{1/8}+O_p(1) 2 h 1/8 + O p ( 1 ) , inside the range of (90) for all sufficiently large h h h . Its remainder, multiplied by the linear prefactor and degree count, is O p ( ( D + ℓ + 1 ) 5 / h 3 ) O_p((D+\ell+1)^5/h^3) O p (( D + ℓ + 1 ) 5 / h 3 ) . Above the cutoff use 0 ≤ c ≤ 1 0 \le c \le1 0 ≤ c ≤ 1 and polynomial growth of the displayed coefficients. Since the uniform moments (65) of w ℓ w_\ell w ℓ are finite to every order, these tails are O p ( h − 3 ) O_p(h^{-3}) O p ( h − 3 ) , even after summing at most D D D increments. The dimension cutoff is farther away and is covered by the same argument. Summing increments proves (D.4). For
a fixed tuple, expansions along two paths must have the same h − 2 h^{-2} h − 2 coefficient as g → ∞ g \to\infty g → ∞ , proving path independence. □ \square □
Comparison with the fixed-cusp expansion. At Y = 0 Y = 0 Y = 0 one has A = 1 / π 2 A = 1/\pi^{2} A = 1/ π 2 , Θ = 2 \Theta= 2 Θ = 2 and M 1 = 3 / 2 M_{1} = 3/2 M 1 = 3/2 , and the normalization (64) gives W g , p / W g , p + 1 = 4 V g , p / V g , p + 1 W_{g,p}/W_{g,p+1} = 4V_{g,p}/V_{g,p+1} W g , p / W g , p + 1 = 4 V g , p / V g , p + 1 and W g − 1 , p + 2 / W g , p = V g − 1 , p + 2 / V g , p W_{g-1,p+2}/W_{g,p} = V_{g-1,p+2}/V_{g,p} W g − 1 , p + 2 / W g , p = V g − 1 , p + 2 / V g , p . With h = 2 g − 2 + p h = 2g - 2 + p h = 2 g − 2 + p , equations (92) and (D.5) become
V g , p + 1 8 π 2 g V g , p = 1 + ( 1 2 − π 2 ) p − 5 4 + 2 π − 2 g + O p ( g − 2 ) , V g − 1 , p + 2 V g , p = 1 + 3 − 2 p π 2 g + O p ( g − 2 ) . \frac{V_{g,p+1}}{8\pi^{2}gV_{g,p}} = 1 + \frac{\left(\frac{1}{2} - \pi^{2}\right)p - \frac{5}{4} + 2\pi^{-2}}{g} + O_{p}(g^{-2}), \qquad\frac{V_{g-1,p+2}}{V_{g,p}} = 1 + \frac{3 - 2p}{\pi^{2}g} + O_{p}(g^{-2}). 8 π 2 g V g , p V g , p + 1 = 1 + g ( 2 1 − π 2 ) p − 4 5 + 2 π − 2 + O p ( g − 2 ) , V g , p V g − 1 , p + 2 = 1 + π 2 g 3 − 2 p + O p ( g − 2 ) .
and (90) gives, for fixed k ≥ 1 k \ge1 k ≥ 1 ,
[ τ k τ 0 p − 1 ] g , p V g , p = 1 − k 2 + ( p − 5 2 ) k − p 2 + 3 2 π 2 g + O p ( g − 2 ) . \frac{\left[\tau_{k}\tau_{0}^{p-1}\right]_{g,p}}{V_{g,p}} = 1 - \frac{k^{2} + \left(p - \frac{5}{2}\right)k - \frac{p}{2} + \frac{3}{2}}{\pi^{2}g} + O_{p}(g^{-2}). V g , p [ τ k τ 0 p − 1 ] g , p = 1 − π 2 g k 2 + ( p − 2 5 ) k − 2 p + 2 3 + O p ( g − 2 ) .
These are the first-order formulas of Mirzakhani and Zograf [147, Theorems 1.4 and 3.2(i)], recovered here as the Y = 0 Y = 0 Y = 0 case of the parameter-dependent coefficients.
The coefficient in the differential equation
Inserting Lemma D.1 into the dilaton identity gives the next coefficient γ p \gamma_{p} γ p of the differential equation (93) . We then evaluate it with finitely many kernel moments and check that it vanishes at the pure-ψ \psi ψ endpoint.
Apply (D.4) to the dilaton child of complexity h + 1 h + 1 h + 1 and set
E p = ∑ d ≥ 1 b d e p + 1 ( d , 0 p ) , γ p = β p 2 − β p − E p / A . (295) E_{p} = \sum_{d \ge1} b_{d}e_{p+1}(d,0^{p}), \qquad\gamma_{p} = \beta_{p}^{2} - \beta_{p} - E_{p}/A. \tag*{(295)} E p = d ≥ 1 ∑ b d e p + 1 ( d , 0 p ) , γ p = β p 2 − β p − E p / A . (295)
The moments of b d b_{d} b d justify this sum and yield
d d Y log Φ g ( p ) ( X ( Y ) ) = h + β p + γ p / h + O p ( h − 2 ) A r . (296) \frac{\mathrm{d}}{\mathrm{d}Y}\log\Phi_{g}^{(p)}(X(Y)) = \frac{h + \beta_{p} + \gamma_{p}/h + O_{p}(h^{-2})}{Ar}. \tag*{(296)} d Y d log Φ g ( p ) ( X ( Y )) = A r h + β p + γ p / h + O p ( h − 2 ) . (296)
Indeed, replacing h + 1 h + 1 h + 1 by h h h in the child profile gives 1 − δ / h + ( δ + e ) / h 2 1 - \delta/h + (\delta+ e)/h^{2} 1 − δ / h + ( δ + e ) / h 2 . Thus the dilaton ratio is A { 1 − β p / h + ( β p + E p / A ) / h 2 + O p ( h − 3 ) } A\{1 - \beta_{p}/h + (\beta_{p} + E_{p}/A)/h^{2} + O_{p}(h^{-3})\} A { 1 − β p / h + ( β p + E p / A ) / h 2 + O p ( h − 3 )} , and inversion proves (D.8). To justify the sum, truncate the dilaton degree at h 1 / 8 h^{1/8} h 1/8 ; the tail is bounded using 1 + d 2 / h + d 4 / h 2 1 + d^{2}/h + d^{4}/h^{2} 1 + d 2 / h + d 4 / h 2 and a sufficiently high absolute moment of b d b_{d} b d from (68) .
The coefficient in (D.7) can be evaluated using finitely many kernel moments, as follows. Put P = p + 1 P = p + 1 P = p + 1 , and write B P ( t ) = B P , 1 ( t , 0 P − 1 ) B_{P}(t) = B_{P,1}(t,0^{P-1}) B P ( t ) = B P , 1 ( t , 0 P − 1 ) . With k = t + ℓ − 1 k = t + \ell- 1 k = t + ℓ − 1 and s = t + ℓ − 2 s = t + \ell- 2 s = t + ℓ − 2 , (D.2) becomes
R P ( k ) = 1 − β P − 1 − A { Θ ( k 2 + ( P − 5 ) k ) + 2 M 1 k + ( 4 − P ) 1 k ≥ 1 } , R_{P}(k) = 1 - \beta_{P-1} - A\{\Theta(k^{2} + (P - 5)k) + 2M_{1}k + (4 - P)1_{k \ge1}\}, R P ( k ) = 1 − β P − 1 − A { Θ ( k 2 + ( P − 5 ) k ) + 2 M 1 k + ( 4 − P ) 1 k ≥ 1 } ,
S P ( s ) = ( s + 1 ) ( 1 − β P + 1 − α P ) − A { ( s + 1 ) ( Θ ( s 2 + ( P − 3 ) s ) + 2 M 1 s ) + ( 5 − 2 P ) s − 1 + 1 s = 0 } . S_{P}(s) = (s + 1)(1 - \beta_{P+1} - \alpha_{P}) - A\{(s + 1)(\Theta(s^{2} + (P - 3)s) + 2M_{1}s) + (5 - 2P)s - 1 + 1_{s = 0}\}. S P ( s ) = ( s + 1 ) ( 1 − β P + 1 − α P ) − A {( s + 1 ) ( Θ ( s 2 + ( P − 3 ) s ) + 2 M 1 s ) + ( 5 − 2 P ) s − 1 + 1 s = 0 } .
Then
B P ( t ) = A ( P − 1 ) ∑ ℓ ≥ max ( 0 , 1 − t ) w ℓ R P ( t + ℓ − 1 ) + 2 A ∑ ℓ ≥ max ( 0 , 2 − t ) w ℓ S P ( t + ℓ − 2 ) + A 2 { [ ( P − 1 2 ) + θ 1 2 ] T 2 − t + 1 2 T 3 − t } . (297) \begin{aligned}
B_{P}(t) ={}& A(P - 1)\sum_{\ell\ge\max(0,1-t)}w_{\ell}R_{P}(t + \ell- 1) \\
&+ 2A\sum_{\ell\ge\max(0,2-t)}w_{\ell}S_{P}(t + \ell- 2) \\
&+ A^{2}\left\{\left[\binom{P - 1}{2} + \frac{\theta_{1}}{2}\right]T_{2-t} + \frac{1}{2}T_{3-t}\right\}. \tag*{(297)}
\end{aligned} B P ( t ) = A ( P − 1 ) ℓ ≥ m a x ( 0 , 1 − t ) ∑ w ℓ R P ( t + ℓ − 1 ) + 2 A ℓ ≥ m a x ( 0 , 2 − t ) ∑ w ℓ S P ( t + ℓ − 2 ) + A 2 { [ ( 2 P − 1 ) + 2 θ 1 ] T 2 − t + 2 1 T 3 − t } . (297)
The zero-degree correction in S P S_{P} S P follows by summing the q q q term of δ P + 1 \delta_{P+1} δ P + 1 : for s ≥ 1 s \ge1 s ≥ 1 the two endpoint pairs have q = 1 q = 1 q = 1 and the s − 1 s - 1 s − 1 interior pairs have q = 2 q = 2 q = 2 .
For t ≥ 3 t \ge3 t ≥ 3 , remove both lower cutoffs, replace 1 k ≥ 1 \mathbf{1}_{k\ge1} 1 k ≥ 1 by one, 1 s = 0 \mathbf{1}_{s=0} 1 s = 0 by zero, and both tails by Θ \Theta Θ . This gives a cubic polynomial B ∞ ( t ) = a 3 t 3 + a 2 t 2 + a 1 t + a 0 B_{\infty}(t)=a_{3}t^{3}+a_{2}t^{2}+a_{1}t+a_{0} B ∞ ( t ) = a 3 t 3 + a 2 t 2 + a 1 t + a 0 . Its coefficients use only M j = ∑ ℓ j w ℓ M_j=\sum\ell^{j}w_{\ell} M j = ∑ ℓ j w ℓ , 0 ≤ j ≤ 3 0\le j\le3 0 ≤ j ≤ 3 . Put Δ j = B P ( j ) − B ∞ ( j ) \Delta_j=B_P(j)-B_{\infty}(j) Δ j = B P ( j ) − B ∞ ( j ) for j = 0 , 1 , 2 j=0,1,2 j = 0 , 1 , 2 , and define
Σ ( d ) = a 3 4 d 2 ( d − 1 ) 2 + a 2 6 d ( d − 1 ) ( 2 d − 1 ) + a 1 2 d ( d − 1 ) + a 0 d , \Sigma(d)=\frac{a_3}{4}d^2(d-1)^2+\frac{a_2}{6}d(d-1)(2d-1)+\frac{a_1}{2}d(d-1)+a_0d, Σ ( d ) = 4 a 3 d 2 ( d − 1 ) 2 + 6 a 2 d ( d − 1 ) ( 2 d − 1 ) + 2 a 1 d ( d − 1 ) + a 0 d ,
e P ( d , 0 P − 1 ) = − Σ ( d ) − ∑ j = 0 2 Δ j 1 d ≥ j + 1 . (298) e_P(d,0^{P-1})=-\Sigma(d)-\sum_{j=0}^{2}\Delta_j\mathbf{1}_{d\ge j+1}. \tag*{(298)} e P ( d , 0 P − 1 ) = − Σ ( d ) − j = 0 ∑ 2 Δ j 1 d ≥ j + 1 . (298)
The three exceptional values use the same moments and w 0 , w 1 , w 2 w_0,w_1,w_2 w 0 , w 1 , w 2 ; there is no remaining infinite recursion. Set
L j = 1 3 [ ( z ∂ z + 1 ) j H ‾ Y ( z ) ] z = 1 ( 1 ≤ j ≤ 4 ) , b 1 = 1 3 , b 2 = − θ 1 5 , L_j=\frac{1}{3}\left[(z\partial_z+1)^j\overline{H}_Y(z)\right]_{z=1}\quad(1\le j\le4),\qquad b_1=\frac{1}{3},\qquad b_2=-\frac{\theta_1}{5}, L j = 3 1 [ ( z ∂ z + 1 ) j H Y ( z ) ] z = 1 ( 1 ≤ j ≤ 4 ) , b 1 = 3 1 , b 2 = − 5 θ 1 ,
and
Σ b = a 3 4 ( L 4 − 2 L 3 + L 2 ) + a 2 6 ( 2 L 3 − 3 L 2 + L 1 ) + a 1 2 ( L 2 − L 1 ) + a 0 L 1 . \Sigma_b=\frac{a_3}{4}(L_4-2L_3+L_2)+\frac{a_2}{6}(2L_3-3L_2+L_1)+\frac{a_1}{2}(L_2-L_1)+a_0L_1. Σ b = 4 a 3 ( L 4 − 2 L 3 + L 2 ) + 6 a 2 ( 2 L 3 − 3 L 2 + L 1 ) + 2 a 1 ( L 2 − L 1 ) + a 0 L 1 .
Then the finite formula for (D.7) is
γ p = β p 2 − β p + Σ b + Δ 0 A + Δ 1 ( A − b 1 ) + Δ 2 ( A − b 1 − b 2 ) A . (299) \gamma_p=\beta_p^2-\beta_p+\frac{\Sigma_b+\Delta_0A+\Delta_1(A-b_1)+\Delta_2(A-b_1-b_2)}{A}. \tag*{(299)} γ p = β p 2 − β p + A Σ b + Δ 0 A + Δ 1 ( A − b 1 ) + Δ 2 ( A − b 1 − b 2 ) . (299)
All ingredients extend analytically to Y = y 0 Y=y_0 Y = y 0 , by the removable normalized-kernel extension used in (100) .
At that endpoint A = 2 / 15 A=2/15 A = 2/15 , Θ = 1 \Theta=1 Θ = 1 , M j = 0 M_j=0 M j = 0 for j ≥ 1 j\ge1 j ≥ 1 , w 0 = θ 1 = 1 w_0=\theta_1=1 w 0 = θ 1 = 1 , and β p = − p / 5 \beta_p=-p/5 β p = − p /5 . Equation (D.9) gives B P ( 0 ) = 0 B_P(0)=0 B P ( 0 ) = 0 and B P ( 1 ) = A ( P − 1 ) ( P + 4 ) / 5 B_P(1)=A(P-1)(P+4)/5 B P ( 1 ) = A ( P − 1 ) ( P + 4 ) /5 . Only b 1 = 1 / 3 b_1=1/3 b 1 = 1/3 and b 2 = − 1 / 5 b_2=-1/5 b 2 = − 1/5 remain, so E p / A = p ( p + 5 ) / 25 = β p 2 − β p E_p/A=p(p+5)/25=\beta_p^2-\beta_p E p / A = p ( p + 5 ) /25 = β p 2 − β p . In particular,
γ p ( y 0 ) = 0 , γ p ( Y ) = O p ( y 0 − Y ) , γ p A r = O p ( 1 ) . (300) \gamma_p(y_0)=0,\qquad\gamma_p(Y)=O_p(y_0-Y),\qquad\frac{\gamma_p}{Ar}=O_p(1). \tag*{(300)} γ p ( y 0 ) = 0 , γ p ( Y ) = O p ( y 0 − Y ) , A r γ p = O p ( 1 ) . (300)
This agrees with an exact computation: string and dilaton at the pure-ψ \psi ψ endpoint give W g , p + 1 / W g , p = 3 ( h + m ) = ( 15 h − 3 p ) / 2 W_{g,p+1}/W_{g,p}=3(h+m)=(15h-3p)/2 W g , p + 1 / W g , p = 3 ( h + m ) = ( 15 h − 3 p ) /2 , so that A W g , p + 1 / ( h W g , p ) = 1 − p / ( 5 h ) AW_{g,p+1}/(hW_{g,p})=1-p/(5h) A W g , p + 1 / ( h W g , p ) = 1 − p / ( 5 h ) exactly, with A = 2 / 15 A=2/15 A = 2/15 ; there is no term of order h − 2 h^{-2} h − 2 , and γ p = 0 \gamma_p=0 γ p = 0 .
Critical matching to the next order
This subsection proves Theorem 8.1 . Recall t p t_p t p and L p L_p L p from (105) . By (D.9)–(D.12), L p L_p L p is an explicit bounded function, given by finitely many kernel derivatives and a convergent one-dimensional integral. The proof integrates the next-order transport equation (D.8) down from the pure-ψ \psi ψ endpoint, using the positive comparison (101) –(102) near the endpoint as in Section 7. There the first correction is fixed by two inputs: a sharper asymptotic formula for the critical intersection numbers, whose derivation in Subsection D.11 is independent of the transport argument, and the first correction to the critical harmonic coefficient of (99) .
Proof of Theorem 8.1 . The sharper critical input is C g ( 2 3 g − 3 ) = π − 1 ( 1 − 2 / ( 9 g ) + O ( g − 2 ) ) C_g(2^{3g-3})=\pi^{-1}(1-2/(9g)+O(g^{-2})) C g ( 2 3 g − 3 ) = π − 1 ( 1 − 2/ ( 9 g ) + O ( g − 2 )) . It is derived in Subsection D.11, as the coefficient d 1 = − 2 / 9 d_1=-2/9 d 1 = − 2/9 in (D.48), from the Itzykson–Zuber identity, the large-order relation of Garoufalidis–Its–Kapaev–Mariño and the first one-instanton coefficient v 1 = − 1 / 8 v_1=-1/8 v 1 = − 1/8 ; that derivation is independent of the transport argument of Subsections D.1–D.3, and the value agrees with [83], equation (25). A leading estimate alone would not identify this correction. Expanding the finite product (98) gives
π C g ( 0 p , 2 m ) = 1 + p ( 5 − p ) / 15 − 4 / 9 h + O p ( h − 2 ) . \pi C_g(0^p,2^m)=1+\frac{p(5-p)/15-4/9}{h}+O_p(h^{-2}). π C g ( 0 p , 2 m ) = 1 + h p ( 5 − p ) /15 − 4/9 + O p ( h − 2 ) .
The critical harmonic coefficient also has a first correction. Its summand law is geometric with mean 3 / 2 3/2 3/2 , variance ν = 15 / 4 \nu=15/4 ν = 15/4 , third cumulant 15 15 15 , and fourth cumulant 705 / 8 705/8 705/8 . At displacement d = − p / 2 d=-p/2 d = − p /2 from 3 h / 2 3h/2 3 h /2 , expansion of its exact negative-binomial coefficient by Stirling’s formula gives
M g , p ( 0 p ; y 0 ) Γ ( h ) 4 − g A ∗ − h / 2 π h ν = 1 + − p 2 / 30 + 4 p / 15 − 19 / 180 h + O p ( h − 2 ) , A ∗ = 2 15 . \frac{M_{g,p}(0^p;y_0)}{\Gamma(h)^{4-g}A_*^{-h}/\sqrt{2\pi h\nu}}
=1+\frac{-p^2/30+4p/15-19/180}{h}+O_p(h^{-2}),\qquad A_*=\frac{2}{15}. Γ ( h ) 4 − g A ∗ − h / 2 π h ν M g , p ( 0 p ; y 0 ) = 1 + h − p 2 /30 + 4 p /15 − 19/180 + O p ( h − 2 ) , A ∗ = 15 2 .
Explicitly, the coefficient before substituting d = − p / 2 d=-p/2 d = − p /2 is − d 2 / ( 2 v ) − 15 d / ( 2 v 2 ) + ( 705 / 8 ) / ( 8 v 2 ) − 5 ( 15 ) 2 / ( 24 v 3 ) -d^2/(2v)-15d/(2v^2)+(705/8)/(8v^2)-5(15)^2/(24v^3) − d 2 / ( 2 v ) − 15 d / ( 2 v 2 ) + ( 705/8 ) / ( 8 v 2 ) − 5 ( 15 ) 2 / ( 24 v 3 ) . Combining these two expansions with (99) proves
lim Y ↑ y 0 π Φ g ( p ) ( X ( Y ) ) J g , p ( Y ) = 1 + t p h + O p ( h − 2 ) . \lim_{Y \uparrow y_0}\frac{\pi\Phi_g^{(p)}(X(Y))}{\mathcal{J}_{g,p}(Y)}=1+\frac{t_p}{h}+O_p(h^{-2}). Y ↑ y 0 lim J g , p ( Y ) π Φ g ( p ) ( X ( Y )) = 1 + h t p + O p ( h − 2 ) .
Take Y ∗ = y 0 − h − 3 Y_* = y_0-h^{-3} Y ∗ = y 0 − h − 3 . The positive homogeneous-polynomial comparison (101) and the analytic normalized factor in (102) change this ratio by O p ( h ( y 0 − Y ) ) = O p ( h − 2 ) O_p(h(y_0-Y))=O_p(h^{-2}) O p ( h ( y 0 − Y )) = O p ( h − 2 ) for Y ≥ Y ∗ Y\geq Y_* Y ≥ Y ∗ . Moreover L p ( Y ) = t p + O p ( y 0 − Y ) \mathcal{L}_p(Y)=t_p+O_p(y_0-Y) L p ( Y ) = t p + O p ( y 0 − Y ) there. For Y ≤ Y ∗ Y\leq Y_* Y ≤ Y ∗ , subtract the exact logarithmic derivative of J g , p \mathcal{J}_{g,p} J g , p in (96) from (D.8), and integrate:
log π Φ g ( p ) ( X ( Y ) ) J g , p ( Y ) = L p ( Y ) h + O p ( h − 2 + 1 h 2 log Z ( Y ) Z ( Y ∗ ) ) . \log\frac{\pi\Phi_g^{(p)}(X(Y))}{\mathcal{J}_{g,p}(Y)}=\frac{\mathcal{L}_p(Y)}{h}+O_p\left(h^{-2}+\frac{1}{h^2}\log\frac{\mathcal{Z}(Y)}{\mathcal{Z}(Y_*)}\right). log J g , p ( Y ) π Φ g ( p ) ( X ( Y )) = h L p ( Y ) + O p ( h − 2 + h 2 1 log Z ( Y ∗ ) Z ( Y ) ) .
Here (D.12) controls the terminal integral, and Z ( Y ∗ ) ≍ h − 15 / 2 \mathcal{Z}(Y_*)\asymp h^{-15/2} Z ( Y ∗ ) ≍ h − 15/2 . The remainder is O p ( log h / h 2 ) O_p(\log h/h^2) O p ( log h / h 2 ) uniformly. Boundedness of L p \mathcal{L}_p L p permits exponentiation, proving (106) in both regions. □
The formal hierarchy
We now construct the full series and control its remainder as the truncation order grows. These are still estimates for generating functions: the passage to a prescribed cusp count is made in Subsection 10.1 and Appendix E. The recursion is triangular, in the sense described in Subsection 8.2. This subsection writes down the formal rules (D.13)–(D.18) and states Proposition D.3, which bounds the coefficients and the actual remainders polynomially in the order; its inputs are the ratio estimates (69) –(82) of Section 6 and the kernel bounds of Section 2. The proof divides each order-j j j coefficient by S j 2 S_j^2 S j 2 , where S j = D + p + M j S_j=D+p+Mj S j = D + p + M j ; after this division each rule becomes a coefficientwise inequality between series with nonnegative coefficients, an analytic majorant (Lemma D.2) bounds the normalized coefficients by C B j CB^j C B j , and the true remainders are controlled by a second, linear majorant system with the same normalization. Matching at the pure-ψ \psi ψ endpoint in Subsection D.11 then completes the proof of Theorem 8.2.
Use a formal variable t t t for inverse complexity, and put
P p ( d ; t ) = ∑ j ≥ 0 e p , j ( d ) t j , C p ( h ) = h A W g , p W g , p + 1 , G p ( h ) = W g − 1 , p + 2 W g , p . (301) P_p(d;t)=\sum_{j\geq0}e_{p,j}(d)t^j,\qquad C_p(h)=\frac{h}{A}\frac{W_{g,p}}{W_{g,p+1}},\qquad G_p(h)=\frac{W_{g-1,p+2}}{W_{g,p}}. \tag*{(301)} P p ( d ; t ) = j ≥ 0 ∑ e p , j ( d ) t j , C p ( h ) = A h W g , p + 1 W g , p , G p ( h ) = W g , p W g − 1 , p + 2 . (301)
The formal series for the last two quantities will also be denoted by C p ( t ) , G p ( t ) C_p(t),G_p(t) C p ( t ) , G p ( t ) . Their constant terms, and e p , 0 e_{p,0} e p , 0 , are one. Here e p , 1 = − δ p e_{p,1}=-\delta_p e p , 1 = − δ p and e p , 2 = e p e_{p,2}=e_p e p , 2 = e p from (90) and (D.3). All series operations below are truncated at a fixed finite order; convergence of a formal series is never asserted. A complexity change from h h h to h − c h-c h − c means replacing t t t by t / ( 1 − c t ) t/(1-ct) t / ( 1 − c t ) .
For a bounded component of genus j j j containing s s s other external labels, set r ∗ = 2 j + s r_* = 2j+s r ∗ = 2 j + s . Thus r ∗ r_* r ∗ is the decrease in complexity from the parent to the complementary component. Its all-zero ratio has the exact factorization
W g − j , p − s W g , p = ∏ a = 0 r ∗ − 1 A C p − s + a ( h − r ∗ + a ) h − r ∗ + a ∏ b = 0 j − 1 G p + 2 b ( h ) . (302) \frac{W_{g-j,p-s}}{W_{g,p}}=\prod_{a=0}^{r_*-1}\frac{A C_{p-s+a}(h-r_*+a)}{h-r_*+a}\prod_{b=0}^{j-1}G_{p+2b}(h). \tag*{(302)} W g , p W g − j , p − s = a = 0 ∏ r ∗ − 1 h − r ∗ + a A C p − s + a ( h − r ∗ + a ) b = 0 ∏ j − 1 G p + 2 b ( h ) . (302)
R p , j , s ( t ) = ( A t ) r ∗ ∏ a = 0 r ∗ − 1 C p − s + a ( t / ( 1 − ( r ∗ − a ) t ) ) 1 − ( r ∗ − a ) t ∏ b = 0 j − 1 G p + 2 b ( t ) . R_{p,j,s}(t)=(At)^{r_*}\prod_{a=0}^{r_*-1}\frac{C_{p-s+a}\left(t/(1-(r_*-a)t)\right)}{1-(r_*-a)t}\prod_{b=0}^{j-1}G_{p+2b}(t). R p , j , s ( t ) = ( A t ) r ∗ a = 0 ∏ r ∗ − 1 1 − ( r ∗ − a ) t C p − s + a ( t / ( 1 − ( r ∗ − a ) t ) ) b = 0 ∏ j − 1 G p + 2 b ( t ) .
The genus steps have unchanged complexity; the cusp steps remove r ∗ r_* r ∗ markings at genus g − j g-j g − j . All their marking counts are positive in the applications below. To obtain R R R through order t K t^K t K requires scalar coefficients only through order K − r ∗ K-r_* K − r ∗ .
The coefficients are computed as follows. Use the child tuples of (D.2), and set W j , q ( d ) = 4 j U ‾ j , q ( d ) W_{j,q}(\mathbf{d})=4^j\overline{U}_{j,q}(\mathbf{d}) W j , q ( d ) = 4 j U j , q ( d ) for the exact bounded-complexity amplitudes. At each order K ≥ 1 K\ge1 K ≥ 1 , form
A p , i = A t 1 − t C p − 1 ( t 1 − t ) ∑ j ≠ i ( 2 d j + 1 ) ∑ ℓ ≥ max ( 0 , 1 − d i − d j ) w ℓ P p − 1 ( d i j ; ℓ ; t 1 − t ) . (303) A_{p,i}=\frac{At}{1-t}C_{p-1}\left(\frac{t}{1-t}\right)\sum_{j\ne i}(2d_j+1)\sum_{\ell\ge\max(0,1-d_i-d_j)}w_\ell P_{p-1}\left(\mathbf{d}^{ij;\ell};\frac{t}{1-t}\right). \tag*{(303)} A p , i = 1 − t A t C p − 1 ( 1 − t t ) j = i ∑ ( 2 d j + 1 ) ℓ ≥ m a x ( 0 , 1 − d i − d j ) ∑ w ℓ P p − 1 ( d ij ; ℓ ; 1 − t t ) . (303)
N p , i = 2 A t 1 − t C p + 1 ( t 1 − t ) G p ( t ) ∑ ℓ ≥ max ( 0 , 2 − d i ) w ℓ ∑ u + v = d i + ℓ − 2 P p + 1 ( d i ; u , v ; t 1 − t ) . N_{p,i}=\frac{2At}{1-t}C_{p+1}\left(\frac{t}{1-t}\right)G_p(t)\sum_{\ell\ge\max(0,2-d_i)}w_\ell\sum_{u+v=d_i+\ell-2}P_{p+1}\left(\mathbf{d}^{i;u,v};\frac{t}{1-t}\right). N p , i = 1 − t 2 A t C p + 1 ( 1 − t t ) G p ( t ) ℓ ≥ m a x ( 0 , 2 − d i ) ∑ w ℓ u + v = d i + ℓ − 2 ∑ P p + 1 ( d i ; u , v ; 1 − t t ) .
Nonnegative node degrees are understood, and A A A is absent at p = 1 p=1 p = 1 . The retained separating contribution is
S p , i = ∑ j ≥ 0 , I ⊂ { 1 , … , p } ∖ { i } s = ∣ I ∣ , ( j , s + 1 ) s t a b l e 2 ≤ 2 j + s ≤ K R p ; j , s ( t ) ∑ ℓ ≥ 0 w ℓ ∑ u + v = d i + ℓ − 2 W j , s + 1 ( u , d I ) P p − s ( v , d J ; t 1 − ( 2 j + s ) t ) . (304) S_{p,i}=\sum_{\substack{j\ge0,\ I\subset\{1,\ldots,p\}\setminus\{i\}\\s=|I|,\ (j,s+1)\ \mathrm{stable}\\2\le2j+s\le K}}R_{p;j,s}(t)\sum_{\ell\ge0}w_\ell\sum_{u+v=d_i+\ell-2}W_{j,s+1}(u,\mathbf{d}_I)P_{p-s}\left(v,\mathbf{d}_J;\frac{t}{1-(2j+s)t}\right). \tag*{(304)} S p , i = j ≥ 0 , I ⊂ { 1 , … , p } ∖ { i } s = ∣ I ∣ , ( j , s + 1 ) stable 2 ≤ 2 j + s ≤ K ∑ R p ; j , s ( t ) ℓ ≥ 0 ∑ w ℓ u + v = d i + ℓ − 2 ∑ W j , s + 1 ( u , d I ) P p − s ( v , d J ; 1 − ( 2 j + s ) t t ) . (304)
where J J J is the remaining set of labels. Only bounded j j j occur; for large g g g these orient the smaller component uniquely. The two orientations cancel the separating factor 1 / 2 1/2 1/2 . Define
e p , K ( 0 ) = 0 , e_{p,K}(\mathbf{0})=0, e p , K ( 0 ) = 0 ,
e p , K ( d + e i ) − e p , K ( d ) = − [ t K ] ( A p , i + N p , i + S p , i ) , e_{p,K}(\mathbf{d}+\mathbf{e}_i)-e_{p,K}(\mathbf{d})=-[t^K](A_{p,i}+N_{p,i}+S_{p,i}), e p , K ( d + e i ) − e p , K ( d ) = − [ t K ] ( A p , i + N p , i + S p , i ) ,
C p ( t ) = 1 A ∑ d ≥ 1 b d P p + 1 ( d , 0 p ; t 1 + t ) . (305) C_p(t)=\frac{1}{A}\sum_{d\ge1}b_dP_{p+1}\left(\mathbf{d},\mathbf{0}^p;\frac{t}{1+t}\right). \tag*{(305)} C p ( t ) = A 1 d ≥ 1 ∑ b d P p + 1 ( d , 0 p ; 1 + t t ) . (305)
Finally, for p ≥ 2 p\ge2 p ≥ 2 , KdV gives
G p ( t ) = P p ( e 1 ; t ) − 3 ∑ 0 ≤ s ≤ p − 2 , j ≥ 0 ( j , s + 2 ) s t a b l e 1 ≤ 2 j + s ≤ K ( p − 2 s ) W j , s + 2 ( 0 ) R p ; j , s ( t ) , G_p(t)=P_p(\mathbf{e}_1;t)-3\sum_{\substack{0\le s\le p-2,\ j\ge0\\(j,s+2)\ \mathrm{stable}\\1\le2j+s\le K}}\binom{p-2}{s}W_{j,s+2}(\mathbf{0})R_{p;j,s}(t), G p ( t ) = P p ( e 1 ; t ) − 3 0 ≤ s ≤ p − 2 , j ≥ 0 ( j , s + 2 ) stable 1 ≤ 2 j + s ≤ K ∑ ( s p − 2 ) W j , s + 2 ( 0 ) R p ; j , s ( t ) ,
G p ( h ) = G p + 1 ( h + 1 ) C p + 2 ( h ) / C p ( h ) . (306) G_p(h)=G_{p+1}(h+1)C_{p+2}(h)/C_p(h). \tag*{(306)} G p ( h ) = G p + 1 ( h + 1 ) C p + 2 ( h ) / C p ( h ) . (306)
The last line descends from p = 2 p=2 p = 2 to p = 1 p=1 p = 1 , and optionally to p = 0 p=0 p = 0 . It is an exact identity; its shift h + 1 h+1 h + 1 must be retained.
The coefficient and remainder bounds below are obtained by comparing coefficientwise inequalities with an analytic system. We isolate the elementary principle used for this comparison; it is applied in the proof of Lemma D.4 and again, with the exponential degree norm, in the proof of Proposition D.9 in Subsection D.9.
Lemma D.2 (Analytic majorants). Let Φ = ( Φ 1 , … , Φ m ) \Phi=(\Phi_1,\ldots,\Phi_m) Φ = ( Φ 1 , … , Φ m ) be power series in ( t , v 1 , … , v m ) (t,v_1,\ldots,v_m) ( t , v 1 , … , v m ) with nonnegative coefficients, convergent near the origin, such that Φ ( 0 , 0 ) = 0 \Phi(0,0)=0 Φ ( 0 , 0 ) = 0 and the matrix ∂ v Φ ( 0 , 0 ) \partial_v\Phi(0,0) ∂ v Φ ( 0 , 0 ) is strictly lower triangular. Then U = Φ ( t , U ) U=\Phi(t,U) U = Φ ( t , U ) has a unique formal solution with U ( 0 ) = 0 U(0)=0 U ( 0 ) = 0 . Its coefficients are nonnegative, and it converges near t = 0 t=0 t = 0 . Let e i , j ∈ [ 0 , ∞ ] e_{i,j}\in[0,\infty] e i , j ∈ [ 0 , ∞ ] , for 1 ≤ i ≤ m 1\le i\le m 1 ≤ i ≤ m and j ≥ 1 j\ge1 j ≥ 1 , and put E i = ∑ j ≥ 1 e i , j t j E_i=\sum_{j\ge1}e_{i,j}t^j E i = ∑ j ≥ 1 e i , j t j . If e i , j ≤ [ t j ] Φ i ( t , E ) e_{i,j}\le[t^j]\Phi_i(t,E) e i , j ≤ [ t j ] Φ i ( t , E ) for all i , j i,j i , j , then e i , j ≤ [ t j ] U i e_{i,j}\le[t^j]U_i e i , j ≤ [ t j ] U i . In particular every e i , j e_{i,j} e i , j is finite, and e i , j ≤ C ρ − j e_{i,j}\le C\rho^{-j} e i , j ≤ C ρ − j with C , ρ > 0 C,\rho>0 C , ρ > 0 depending only on Φ \Phi Φ .
Proof. Order the pairs ( j , i ) (j,i) ( j , i ) lexicographically. Since E E E has no constant term, [ t j ] Φ i ( t , E ) [t^j]\Phi_i(t,E) [ t j ] Φ i ( t , E ) is a polynomial with nonnegative coefficients in the numbers e k , l e_{k,l} e k , l . A monomial t a v β t^av^\beta t a v β with a ≥ 1 a\ge1 a ≥ 1 or ∣ β ∣ ≥ 2 |\beta|\ge2 ∣ β ∣ ≥ 2 contributes only coefficients with l < j l<j l < j , since each factor has order at least one. The remaining terms are ∂ v k Φ i ( 0 , 0 ) e k , j \partial_{v_k}\Phi_i(0,0)e_{k,j} ∂ v k Φ i ( 0 , 0 ) e k , j , and they occur only for k < i k<i k < i . Thus [ t j ] Φ i ( t , E ) [t^j]\Phi_i(t,E) [ t j ] Φ i ( t , E ) involves only pairs preceding ( j , i ) (j,i) ( j , i ) . The equations U = Φ ( t , U ) U=\Phi(t,U) U = Φ ( t , U ) therefore determine the coefficients of U U U recursively and uniquely, and these are nonnegative. The same induction and the monotonicity of polynomials with nonnegative coefficients give e i , j ≤ [ t j ] U i e_{i,j}\le[t^j]U_i e i , j ≤ [ t j ] U i . The derivative of ( t , v ) ↦ v − Φ ( t , v ) (t,v)\mapsto v-\Phi(t,v) ( t , v ) ↦ v − Φ ( t , v ) with respect to v v v at the origin is I − ∂ v Φ ( 0 , 0 ) I-\partial_v\Phi(0,0) I − ∂ v Φ ( 0 , 0 ) , which is invertible. The analytic implicit-function theorem gives
a solution holomorphic near t = 0 t=0 t = 0 and vanishing at t = 0 t=0 t = 0 ; its Taylor series is the formal solution. Cauchy’s estimate on a small circle gives the final bound. □
Proposition D.3 (Uniform recursive expansions). There are C , B , M > 0 C,B,M>0 C , B , M > 0 and one complex neighborhood Ω \Omega Ω of [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] such that the preceding rules determine path-independent holomorphic coefficients with
∥ e p , j ( d ) ∥ Ω ≤ C B j ( D + p + M j ) 2 j , ∥ C p , j ∥ Ω + ∥ G p , j ∥ Ω ≤ C B j ( p + M j ) 2 j . \lVert e_{p,j}(\mathbf{d})\rVert_{\Omega}\leq CB^j(D+p+Mj)^{2j},\qquad\lVert C_{p,j}\rVert_{\Omega}+\lVert G_{p,j}\rVert_{\Omega}\leq CB^j(p+Mj)^{2j}. ∥ e p , j ( d ) ∥ Ω ≤ C B j ( D + p + M j ) 2 j , ∥ C p , j ∥ Ω + ∥ G p , j ∥ Ω ≤ C B j ( p + M j ) 2 j . For every stable state p ≥ 1 p\geq1 p ≥ 1 , h = 2 g − 2 + p ≥ 1 h=2g-2+p\geq1 h = 2 g − 2 + p ≥ 1 , every L ≥ 1 L\geq1 L ≥ 1 , and every nonnegative degree tuple, uniformly for 0 ≤ Y ≤ y 0 0\leq Y\leq y_0 0 ≤ Y ≤ y 0 ,
∣ C g , p ( d ) − ∑ j < L e p , j ( d ) h − j ∣ ≤ C B L ( D + p + M L ) 2 L h − L , \left|C_{g,p}(\mathbf{d})-\sum_{j<L}e_{p,j}(\mathbf{d})h^{-j}\right|\leq CB^L(D+p+ML)^{2L}h^{-L}, C g , p ( d ) − j < L ∑ e p , j ( d ) h − j ≤ C B L ( D + p + M L ) 2 L h − L , ∣ C p ( h ) − ∑ j < L C p , j h − j ∣ + ∣ G p ( h ) − ∑ j < L G p , j h − j ∣ ≤ C B L ( p + M L ) 2 L h − L . (307) \left|C_p(h)-\sum_{j<L}C_{p,j}h^{-j}\right|+\left|G_p(h)-\sum_{j<L}G_{p,j}h^{-j}\right|\leq CB^L(p+ML)^{2L}h^{-L}. \tag*{(307)} C p ( h ) − j < L ∑ C p , j h − j + G p ( h ) − j < L ∑ G p , j h − j ≤ C B L ( p + M L ) 2 L h − L . (307)
Here D = ∑ d i D=\sum d_i D = ∑ d i , C p , j = [ t j ] C p ( t ) C_{p,j}=[t^j]C_p(t) C p , j = [ t j ] C p ( t ) and G p , j = [ t j ] G p ( t ) G_{p,j}=[t^j]G_p(t) G p , j = [ t j ] G p ( t ) ; invalid-degree profiles and invalid genus numerators are zero. For fixed p , j p,j p , j , the rules use finitely many kernel moments, low kernel coefficients and exact bounded-complexity amplitudes.
The proof of Proposition D.3 occupies Subsections D.5–D.8. At each order the construction treats the profile, the cusp ratio, the genus ratio at p ≥ 2 p\geq2 p ≥ 2 , and the descent to p = 1 p=1 p = 1 , in this order; every profile on the right of (D.15)–(D.16) has smaller order, and the dilaton and KdV rules use the current profile only after it has been constructed. Subsection D.5 fixes the analytic neighborhood and the weights S j S_j S j and records four elementary bounds. Lemma D.4 bounds the formal coefficients, uniformly in the order, by a system of analytic majorants; this is the first assertion of the proposition. Lemmas D.5 and D.6 control the omitted topological types and give first-order estimates for the true quantities. Finally, Subsection D.8 proves the remainder bounds (D.19) by induction on the order, using a linear majorant system for the normalized errors.
Weights and elementary bounds
We fix a common analytic neighborhood and the normalizing weights, and record four elementary bounds used in both the coefficient and the remainder estimates.
Shrink one complex neighborhood Ω \Omega Ω of [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] so that the normalized kernels K ‾ Y , H ‾ Y \overline{K}_Y,\overline{H}_Y K Y , H Y and their nonvanishing quotients from Section 2 are holomorphic and bounded on a common disk ∣ w ∣ ≤ 1 + 2 η |w|\leq1+2\eta ∣ w ∣ ≤ 1 + 2 η , η > 0 \eta>0 η > 0 , for every Y ∈ Ω Y\in\Omega Y ∈ Ω , and so that A ( Y ) A(Y) A ( Y ) is bounded above and away from zero on Ω \Omega Ω . Since ∑ ℓ w ℓ w ℓ = ( 1 − w ) / K ‾ Y ( w ) \sum_{\ell}w_\ell w^\ell=(1-w)/\overline{K}_Y(w) ∑ ℓ w ℓ w ℓ = ( 1 − w ) / K Y ( w ) and b d = 1 3 [ w d − 1 ] H ‾ Y ( w ) b_d=\frac{1}{3}[w^{d-1}]\overline{H}_Y(w) b d = 3 1 [ w d − 1 ] H Y ( w ) , Cauchy’s estimate on the circle ∣ w ∣ = 1 + η |w|=1+\eta ∣ w ∣ = 1 + η gives constants C 0 C_0 C 0 and ρ = 1 + η > 1 \rho=1+\eta>1 ρ = 1 + η > 1 with
∣ w ℓ ( Y ) ∣ ≤ C 0 ρ − ℓ , ∣ b d ( Y ) ∣ ≤ C 0 ρ − d ( Y ∈ Ω , ℓ ≥ 0 , d ≥ 1 ) . (308) |w_\ell(Y)|\leq C_0\rho^{-\ell},\qquad|b_d(Y)|\leq C_0\rho^{-d}\qquad(Y\in\Omega,\ \ell\geq0,\ d\geq1). \tag*{(308)} ∣ w ℓ ( Y ) ∣ ≤ C 0 ρ − ℓ , ∣ b d ( Y ) ∣ ≤ C 0 ρ − d ( Y ∈ Ω , ℓ ≥ 0 , d ≥ 1 ) . (308)
The absolute-series argument in the proof of Lemma 5.2(iii) also works on Ω \Omega Ω : the background majorant B Y B_Y B Y of Subsection D.5.1 has a uniformly bounded germ, so one small radius z ∗ z_* z ∗ makes B Y ( z ∗ ) ≤ 1 / 2 B_Y(z_*)\leq1/2 B Y ( z ∗ ) ≤ 1/2 for all Y ∈ Ω Y\in\Omega Y ∈ Ω . A stable component of complexity r r r has at most r + 2 r+2 r + 2 markings and deficit m ≤ 3 r / 2 m\leq3r/2 m ≤ 3 r /2 , so the primitive bound of Lemma 5.2(i) and this majorant give
∣ W j , q ( d ) ∣ ≤ C 0 r Γ ( r ) ( r = 2 j − 2 + q ≥ 1 ) , (309) |W_{j,q}(\mathbf{d})|\leq C_0^r\Gamma(r)\qquad(r=2j-2+q\geq1), \tag*{(309)} ∣ W j , q ( d ) ∣ ≤ C 0 r Γ ( r ) ( r = 2 j − 2 + q ≥ 1 ) , (309)
uniformly in the markings, the degrees and Y ∈ Ω Y\in\Omega Y ∈ Ω ; enlarging C 0 C_0 C 0 covers the components of complexity one.
Fix an integer M ≥ 2 M\geq2 M ≥ 2 with e 2 / M < ρ e^{2/M}<\rho e 2/ M < ρ . For a profile at a state with p p p markings and degree sum D D D , and for a scalar with p p p markings, the order-j j j weights are
S j = D + p + M j and S j = p + M j (310) S_j=D+p+Mj\qquad\text{and}\qquad S_j=p+Mj \tag*{(310)} S j = D + p + M j and S j = p + M j (310)
respectively, so that S j ≥ M j ≥ j S_j \ge Mj \ge j S j ≥ M j ≥ j and S j ≥ p S_j \ge p S j ≥ p . The following four bounds are used repeatedly.
(a) Child weights. Let k ≤ j k \le j k ≤ j . The merger child d i j ′ , ℓ \mathrm{d}^{ij',\ell} d i j ′ , ℓ has p − 1 p-1 p − 1 markings and degree sum D + ℓ − 1 D+\ell-1 D + ℓ − 1 , the nonseparating child d i ; u , v \mathrm{d}^{i;u,v} d i ; u , v has p + 1 p+1 p + 1 markings and degree sum D + ℓ − 2 D+\ell-2 D + ℓ − 2 , and the large child ( v , d J ) (v,\mathrm{d}_J) ( v , d J ) of a separating term has p − s p-s p − s markings and degree sum at most D + ℓ − 2 D+\ell-2 D + ℓ − 2 . In all three cases the order-k k k weight of the child is at most S j + ℓ S_j+\ell S j + ℓ . A scalar factor occurring at loss r r r has at most p + r p+r p + r markings; if it enters at order k k k with k + r ≤ j k+r \le j k + r ≤ j , its weight is at most p + r + M k ≤ p + M j ≤ S j p+r+Mk \le p+Mj \le S_j p + r + M k ≤ p + M j ≤ S j , because r ≤ j − k ≤ M ( j − k ) r \le j-k \le M(j-k) r ≤ j − k ≤ M ( j − k ) . The dilaton child ( d , 0 p ) (d,0^p) ( d , 0 p ) of a scalar with p p p markings has weight at most S j + d + 1 S_j+d+1 S j + d + 1 , and the children in the KdV and descent rules have weights at most S j + 2 S_j+2 S j + 2 .
(b) Moment bound. For 0 ≤ k ≤ 2 j 0 \le k \le2j 0 ≤ k ≤ 2 j and a ≥ 0 a \ge0 a ≥ 0 , ( S j + a ) k ≤ S j k ( 1 + a / S j ) 2 j ≤ S j k e 2 a / M (S_j+a)^k \le S_j^k(1+a/S_j)^{2j} \le S_j^k e^{2a/M} ( S j + a ) k ≤ S j k ( 1 + a / S j ) 2 j ≤ S j k e 2 a / M , since S j ≥ M j S_j \ge Mj S j ≥ M j . Hence, with C 1 = C 0 ∑ n ≥ 0 ρ − n e 2 ( n + 1 ) / M < ∞ C_1=C_0\sum_{n\ge0}\rho^{-n}e^{2(n+1)/M}<\infty C 1 = C 0 ∑ n ≥ 0 ρ − n e 2 ( n + 1 ) / M < ∞ ,
∑ ℓ ≥ 0 ∣ w ℓ ∣ ( S j + ℓ + 1 ) k ≤ C 1 S j k , ∑ d ≥ 1 ∣ b d ∣ ( S j + d + 1 ) k ≤ C 1 S j k ( 0 ≤ k ≤ 2 j ) . (311) \sum_{\ell\ge0}|w_\ell|(S_j+\ell+1)^k \le C_1S_j^k,\qquad\sum_{d\ge1}|b_d|(S_j+d+1)^k \le C_1S_j^k \qquad(0\le k\le2j). \tag*{(311)} ℓ ≥ 0 ∑ ∣ w ℓ ∣ ( S j + ℓ + 1 ) k ≤ C 1 S j k , d ≥ 1 ∑ ∣ b d ∣ ( S j + d + 1 ) k ≤ C 1 S j k ( 0 ≤ k ≤ 2 j ) . (311)
(c) Shifts. For integers a ≥ 1 a\ge1 a ≥ 1 , q ≥ 0 q\ge0 q ≥ 0 and 0 ≤ c ≤ c 0 0\le c\le c_0 0 ≤ c ≤ c 0 ,
[ t q ] ( 1 − c t ) − a = a ( a + 1 ) ⋯ ( a + q − 1 ) c q q ! ≤ ( ( a + q ) c 0 ) q q ! , [t^q](1-ct)^{-a}=\frac{a(a+1)\cdots(a+q-1)c^q}{q!}\le\frac{((a+q)c_0)^q}{q!}, [ t q ] ( 1 − c t ) − a = q ! a ( a + 1 ) ⋯ ( a + q − 1 ) c q ≤ q ! (( a + q ) c 0 ) q ,
with the empty product equal to one. For a product ∏ k ( 1 − c k t ) − a k \prod_k(1-c_kt)^{-a_k} ∏ k ( 1 − c k t ) − a k , taking absolute values if necessary, the coefficient is bounded by that of ( 1 − c 0 t ) − a (1-c_0t)^{-a} ( 1 − c 0 t ) − a , where a = ∑ k a k a=\sum_k a_k a = ∑ k a k and c 0 ≥ max k ∣ c k ∣ c_0\ge\max_k|c_k| c 0 ≥ max k ∣ c k ∣ . In an order-j j j term, a prefactor of order r r r and coefficients of total order i i i leave shift degree q = j − r − i q=j-r-i q = j − r − i . The total denominator exponent is at most r + i r+i r + i , so a + q ≤ j a+q\le j a + q ≤ j ; also c 0 ≤ max ( 1 , r ) ≤ j c_0\le\max(1,r)\le j c 0 ≤ max ( 1 , r ) ≤ j . Thus the coefficient is at most j 2 q / q ! ≤ S j 2 q / q ! j^{2q}/q!\le S_j^{2q}/q! j 2 q / q ! ≤ S j 2 q / q ! . The case a = 0 a=0 a = 0 contributes only q = 0 q=0 q = 0 . After normalization all shifts are therefore majorized by e t e^t e t .
(d) Counts. At a profile state, ∑ j ′ ≠ i ( 2 d j ′ + 1 ) ≤ 2 D + p ≤ 2 S j \sum_{j'\ne i}(2d_{j'}+1)\le2D+p\le2S_j ∑ j ′ = i ( 2 d j ′ + 1 ) ≤ 2 D + p ≤ 2 S j , the number of pairs ( u , v ) (u,v) ( u , v ) with u + v = d i + ℓ − 2 u+v=d_i+\ell-2 u + v = d i + ℓ − 2 is at most D + ℓ ≤ S j + ℓ D+\ell\le S_j+\ell D + ℓ ≤ S j + ℓ , and a coordinate path from 0 0 0 to d d d has D ≤ S j D\le S_j D ≤ S j steps. In a separating term of loss r = 2 j 0 + s r=2j_0+s r = 2 j 0 + s the number of label subsets is ( p − 1 s ) ≤ p s / s ! \binom{p-1}{s}\le p^s/s! ( s p − 1 ) ≤ p s / s ! , there are at most r + 1 ≤ 2 r r+1\le2^r r + 1 ≤ 2 r pairs ( j 0 , s ) (j_0,s) ( j 0 , s ) , and the small component has complexity r − 1 r-1 r − 1 in the profile rule and r r r in the KdV rule, so (D.21) gives
( p − 1 s ) ∣ W ∣ ≤ C 0 r p s Γ ( r ) s ! ≤ C 0 r ( p + r ) r ≤ C 0 r S j r . (312) \binom{p-1}{s}|W|\le C_0^r\frac{p^s\Gamma(r)}{s!}\le C_0^r(p+r)^r\le C_0^rS_j^r. \tag*{(312)} ( s p − 1 ) ∣ W ∣ ≤ C 0 r s ! p s Γ ( r ) ≤ C 0 r ( p + r ) r ≤ C 0 r S j r . (312)
Indeed Γ ( r ) / s ! ≤ r r − 1 − s \Gamma(r)/s!\le r^{r-1-s} Γ ( r ) / s ! ≤ r r − 1 − s when s ≤ r − 1 s\le r-1 s ≤ r − 1 , and Γ ( r ) / r ! ≤ 1 \Gamma(r)/r!\le1 Γ ( r ) / r ! ≤ 1 when s = r s=r s = r ; and p + r ≤ p + M j ≤ S j p+r\le p+Mj\le S_j p + r ≤ p + M j ≤ S j because r ≤ j r\le j r ≤ j .
We bound the formal coefficients uniformly in the order. After division by the weights, each coefficient rule becomes a coefficientwise inequality between power series with nonnegative coefficients, and the normalized coefficient sequences are dominated by the solution of a convergent analytic system.
Lemma D.4 (Coefficient bounds). The rules (D.15)–(D.18) determine coefficients e p , j ( d ) e_{p,j}(\mathbf{d}) e p , j ( d ) , C p , j C_{p,j} C p , j and G p , j G_{p,j} G p , j that are holomorphic on Ω \Omega Ω , and there are C , B 0 C,B_0 C , B 0 such that, on Ω \Omega Ω , for all p ≥ 1 p\ge1 p ≥ 1 , all degree tuples and all j ≥ 0 j\ge0 j ≥ 0 ,
∣ e p , j ( d ) ∣ ≤ C B 0 j S j 2 j , ∣ C p , j ∣ + ∣ G p , j ∣ ≤ C B 0 j S j 2 j , |e_{p,j}(\mathbf{d})|\le CB_0^jS_j^{2j},\qquad|C_{p,j}|+|G_{p,j}|\le CB_0^jS_j^{2j}, ∣ e p , j ( d ) ∣ ≤ C B 0 j S j 2 j , ∣ C p , j ∣ + ∣ G p , j ∣ ≤ C B 0 j S j 2 j , with the weights (D.22).
Proof. Define the normalized coefficient sequences
ε j = sup ∣ e p , j ( d ) ∣ S j 2 j , γ j = sup ∣ C p , j ∣ S j 2 j , η j = sup p ≥ 2 ∣ G p , j ∣ S j 2 j , ζ j = sup ∣ G 1 , j ∣ S j 2 j , \varepsilon_j=\sup\frac{|e_{p,j}(\mathbf{d})|}{S_j^{2j}},\qquad\gamma_j=\sup\frac{|C_{p,j}|}{S_j^{2j}},\qquad\eta_j=\sup_{p\ge2}\frac{|G_{p,j}|}{S_j^{2j}},\qquad\zeta_j=\sup\frac{|G_{1,j}|}{S_j^{2j}}, ε j = sup S j 2 j ∣ e p , j ( d ) ∣ , γ j = sup S j 2 j ∣ C p , j ∣ , η j = p ≥ 2 sup S j 2 j ∣ G p , j ∣ , ζ j = sup S j 2 j ∣ G 1 , j ∣ ,
e n d s m i d = 1 {ends_mid=1} e n d s m i d = 1
the suprema being taken over Y ∈ Ω Y \in\Omega Y ∈ Ω , all marking counts p ≥ 1 p \ge1 p ≥ 1 and all degree tuples; ε 0 = γ 0 = η 0 = ζ 0 = 1 \varepsilon_0=\gamma_0=\eta_0=\zeta_0=1 ε 0 = γ 0 = η 0 = ζ 0 = 1 . Write ε ( t ) = ∑ j ε j t j \varepsilon(t)=\sum_j\varepsilon_jt^j ε ( t ) = ∑ j ε j t j , and similarly γ , η , ζ \gamma,\eta,\zeta γ , η , ζ , and put η ˉ = max ( η , ζ ) \bar{\eta}=\max(\eta,\zeta) η ˉ = max ( η , ζ ) coefficientwise, which bounds every genus factor. Each coefficient of order j j j is a finite combination of coefficients of lower order, summed against w ℓ w_\ell w ℓ or b d b_d b d ; by (D.20) and the polynomial growth in ℓ \ell ℓ or d d d of the lower-order weights these sums converge absolutely and uniformly on Ω \Omega Ω , so each coefficient is holomorphic on Ω \Omega Ω and each supremum is finite, by induction on j j j . We bound the order-j j j coefficients term by term.
Profile. Fix a state ( d , p ) (\mathbf{d},p) ( d , p ) and a direction i i i . By (D.15), [ t j ] A p , i [t^j]A_{p,i} [ t j ] A p , i is the sum over j ′ ≠ i j'\ne i j ′ = i , ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 and i 1 + i 2 + q = j − 1 i_1+i_2+q=j-1 i 1 + i 2 + q = j − 1 of
A ( 2 d j ′ + 1 ) w ℓ C p − 1 , i 1 e p − 1 , i 2 ( d i j ′ , ℓ ) [ t q ] ( 1 − t ) − ( 1 + i 1 + i 2 ) . A(2d_{j'}+1)w_\ell C_{p-1,i_1}e_{p-1,i_2}(d^{ij'},\ell)[t^q](1-t)^{-(1+i_1+i_2)}. A ( 2 d j ′ + 1 ) w ℓ C p − 1 , i 1 e p − 1 , i 2 ( d i j ′ , ℓ ) [ t q ] ( 1 − t ) − ( 1 + i 1 + i 2 ) .
By (a) of Subsection D.5 the two coefficients are at most γ i 1 S j 2 i 1 \gamma_{i_1}S_j^{2i_1} γ i 1 S j 2 i 1 and ε i 2 ( S j + ℓ ) i 2 \varepsilon_{i_2}(S_j+\ell)^{i_2} ε i 2 ( S j + ℓ ) i 2 ; by (b) the sum over ℓ \ell ℓ is at most C 1 S j 2 ( i 1 + i 2 ) C_1S_j^{2(i_1+i_2)} C 1 S j 2 ( i 1 + i 2 ) ; by (c) the shift coefficient is at most S j 2 q / q ! S_j^{2q}/q! S j 2 q / q ! ; and by (d) the merger count is at most 2 S j 2S_j 2 S j . Hence
∣ [ t j ] A p , i ∣ ≤ 2 ∣ A ∣ C 1 S j 2 j − 1 [ t j − 1 ] ( e t γ ( t ) ε ( t ) ) . \left|[t^j]A_{p,i}\right|\le2|A|C_1S_j^{2j-1}[t^{j-1}](e^t\gamma(t)\varepsilon(t)). [ t j ] A p , i ≤ 2∣ A ∣ C 1 S j 2 j − 1 [ t j − 1 ] ( e t γ ( t ) ε ( t )) .
The term N p , i N_{p,i} N p , i has in addition one genus coefficient and the pair count S j + ℓ S_j+\ell S j + ℓ , which (D.23) absorbs with k = 2 ( i 1 + i 2 + i 3 ) + 1 ≤ 2 j k=2(i_1+i_2+i_3)+1\le2j k = 2 ( i 1 + i 2 + i 3 ) + 1 ≤ 2 j ; it is at most 2 ∣ A ∣ C 1 S j 2 j − 1 [ t j − 1 ] ( e t γ η ε ) 2|A|C_1S_j^{2j-1}[t^{j-1}](e^t\gamma\eta\varepsilon) 2∣ A ∣ C 1 S j 2 j − 1 [ t j − 1 ] ( e t γ η ε ) . A separating term of loss r ≥ 2 r\ge2 r ≥ 2 in (D.16) carries the prefactor ( A t ) r (At)^r ( A t ) r , r r r cusp coefficients, j 0 ≤ r / 2 j_0\le r/2 j 0 ≤ r /2 genus coefficients, one profile coefficient, the count (D.24), the pair count and the shifts. Since γ \gamma γ and η ˉ \bar{\eta} η ˉ have constant term one and nonnegative coefficients, γ r η ˉ j 0 ⪯ γ 2 r η ˉ 2 r \gamma^r\bar{\eta}^{j_0}\preceq\gamma^{2r}\bar{\eta}^{2r} γ r η ˉ j 0 ⪯ γ 2 r η ˉ 2 r , and the term is at most ( 2 ∣ A ∣ C 0 ) r C 1 S j 2 ( j − r ) S j r + 1 [ t j − r ] ( e t γ 2 r η ˉ 2 r ε ) (2|A|C_0)^rC_1S_j^{2(j-r)}S_j^{r+1}[t^{j-r}](e^t\gamma^{2r}\bar{\eta}^{2r}\varepsilon) ( 2∣ A ∣ C 0 ) r C 1 S j 2 ( j − r ) S j r + 1 [ t j − r ] ( e t γ 2 r η ˉ 2 r ε ) , where S j r + 1 ≤ S j 2 r − 1 S_j^{r+1}\le S_j^{2r-1} S j r + 1 ≤ S j 2 r − 1 because r ≥ 2 r\ge2 r ≥ 2 . Summing the increments along a coordinate path of at most S j S_j S j steps supplies one more factor S j S_j S j . After division by S j 2 j S_j^{2j} S j 2 j we obtain, for a constant b 1 b_1 b 1 depending only on sup Ω ∣ A ∣ \sup_\Omega|A| sup Ω ∣ A ∣ , C 0 C_0 C 0 and C 1 C_1 C 1 ,
ε j ≤ [ b 1 e t ε ( t ) ( t γ η ˉ + ∑ r ≥ 2 ( b 1 t ) r γ 2 r η ˉ 2 r ) ] j ( j ≥ 1 ) . (313) \varepsilon_j\le\left[b_1e^t\varepsilon(t)\left(t\gamma\bar{\eta}+\sum_{r\ge2}(b_1t)^r\gamma^{2r}\bar{\eta}^{2r}\right)\right]_j \qquad(j\ge1). \tag*{(313)} ε j ≤ [ b 1 e t ε ( t ) ( t γ η ˉ + r ≥ 2 ∑ ( b 1 t ) r γ 2 r η ˉ 2 r ) ] j ( j ≥ 1 ) . (313)
Cusp ratio. By (D.17), [ t j ] C p [t^j]C_p [ t j ] C p is the sum over d ≥ 1 d\ge1 d ≥ 1 and i + q = j i+q=j i + q = j of A − 1 b d e p + 1 , i ( d , 0 p ) [ t q ] ( 1 + t ) − i A^{-1}b_de_{p+1,i}(d,0^p)[t^q](1+t)^{-i} A − 1 b d e p + 1 , i ( d , 0 p ) [ t q ] ( 1 + t ) − i . The child weight is at most S j + d + 1 S_j+d+1 S j + d + 1 by (a), so (D.23) and (c) give
γ j ≤ [ b 2 e t ε ( t ) ] j ( j ≥ 1 ) , b 2 = C 1 inf Ω ∣ A ∣ . (314) \gamma_j\le\left[b_2e^t\varepsilon(t)\right]_j\qquad(j\ge1),\qquad b_2=\frac{C_1}{\inf_\Omega|A|}. \tag*{(314)} γ j ≤ [ b 2 e t ε ( t ) ] j ( j ≥ 1 ) , b 2 = inf Ω ∣ A ∣ C 1 . (314)
Genus ratio at p > 2 p>2 p > 2 . In (D.18) the term P p ( e 1 ; t ) P_p(e_1;t) P p ( e 1 ; t ) has weight at most S j + 1 S_j+1 S j + 1 and contributes at most e 2 / M ε j e^{2/M}\varepsilon_j e 2/ M ε j . A separating term of loss r ≥ 1 r\ge1 r ≥ 1 carries 3 ( p − 2 ) ∣ W j 0 , s + 2 ( 0 ) ∣ ≤ 3 C 0 r S j r ≤ 3 C 0 r S j 2 r 3^{(p-2)}|\mathcal{W}_{j_0,s+2}(0)|\le3C_0^rS_j^r\le3C_0^rS_j^{2r} 3 ( p − 2 ) ∣ W j 0 , s + 2 ( 0 ) ∣ ≤ 3 C 0 r S j r ≤ 3 C 0 r S j 2 r by (D.24), the prefactor ( A t ) r (At)^r ( A t ) r , at most 2 r 2r 2 r scalar coefficients and the shifts. Hence
η j ≤ [ b 3 ε ( t ) + b 3 e t ∑ r ≥ 1 ( b 3 t ) r γ 2 r η ˉ 2 r ] j ( j ≥ 1 ) , (315) \eta_j\le\left[b_3\varepsilon(t)+b_3e^t\sum_{r\ge1}(b_3t)^r\gamma^{2r}\bar{\eta}^{2r}\right]_j\qquad(j\ge1), \tag*{(315)} η j ≤ [ b 3 ε ( t ) + b 3 e t r ≥ 1 ∑ ( b 3 t ) r γ 2 r η ˉ 2 r ] j ( j ≥ 1 ) , (315)
for a constant b 3 b_3 b 3 depending only on M M M , sup Ω ∣ A ∣ \sup_\Omega|A| sup Ω ∣ A ∣ and C 0 C_0 C 0 .
Descent. The identity G 1 ( h ) = G 2 ( h + 1 ) C 3 ( h ) / C 1 ( h ) G_1(h)=G_2(h+1)C_3(h)/C_1(h) G 1 ( h ) = G 2 ( h + 1 ) C 3 ( h ) / C 1 ( h ) is a product of a genus series with the shift t / ( 1 + t ) t/(1+t) t / ( 1 + t ) , a cusp series and the formal inverse of a cusp series; the three weights are at most S j + 2 S_j+2 S j + 2 . If ∣ C 1 , i ∣ ≤ γ i S j 2 i |C_{1,i}|\le\gamma_iS_j^{2i} ∣ C 1 , i ∣ ≤ γ i S j 2 i for i ≤ j i\le j i ≤ j , then the coefficients of 1 / C 1 1/C_1 1/ C 1 through order j j j are bounded by S j 2 i S_j^{2i} S j 2 i times those of 1 / ( 2 − γ ( t ) ) = ∑ k ≥ 0 ( γ − 1 ) k 1/(2-\gamma(t))=\sum_{k\ge0}(\gamma-1)^k 1/ ( 2 − γ ( t )) = ∑ k ≥ 0 ( γ − 1 ) k . Therefore
ζ j ≤ [ b 4 e t t η ( t ) γ ( t ) 2 − γ ( t ) ] j ( j ≥ 1 ) , b 4 = e 12 / M . (316) \zeta_j\le\left[b_4e^t\frac{t\eta(t)\gamma(t)}{2-\gamma(t)}\right]_j\qquad(j\ge1),\qquad b_4=e^{12/M}. \tag*{(316)} ζ j ≤ [ b 4 e t 2 − γ ( t ) t η ( t ) γ ( t ) ] j ( j ≥ 1 ) , b 4 = e 12/ M . (316)
The majorant system. Put b = max ( 1 , b 1 , b 2 , b 3 , b 4 ) b=\max(1,b_1,b_2,b_3,b_4) b = max ( 1 , b 1 , b 2 , b 3 , b 4 ) and let E ∗ , C ∗ , H ∗ , G ∗ E_*,C_*,H_*,G_* E ∗ , C ∗ , H ∗ , G ∗ be the formal solution, with constant terms one, of
E ∗ = 1 + b e t E ∗ ( t C ∗ G ∗ + b t C ∗ 2 G ∗ 2 1 − b t C ∗ 2 G ∗ 2 ) , C ∗ = 1 + b ( e t E ∗ − 1 ) , H ∗ = 1 + b ( E ∗ − 1 ) + b e t b t C ∗ 2 G ∗ 2 1 − b t C ∗ 2 G ∗ 2 , G ∗ = 1 + b ( e t H ∗ C ∗ 2 − C ∗ − 1 ) . (317) \begin{aligned}
E_*&=1+b e^tE_*\left(tC_*G_*+\frac{btC_*^2G_*^2}{1-btC_*^2G_*^2}\right),\qquad C_*=1+b(e^tE_*-1),\\
H_*&=1+b(E_*-1)+b e^t\frac{btC_*^2G_*^2}{1-btC_*^2G_*^2},\qquad G_*=1+b\left(\frac{e^tH_*C_*}{2-C_*}-1\right). \tag*{(317)}
\end{aligned} E ∗ H ∗ = 1 + b e t E ∗ ( t C ∗ G ∗ + 1 − b t C ∗ 2 G ∗ 2 b t C ∗ 2 G ∗ 2 ) , C ∗ = 1 + b ( e t E ∗ − 1 ) , = 1 + b ( E ∗ − 1 ) + b e t 1 − b t C ∗ 2 G ∗ 2 b t C ∗ 2 G ∗ 2 , G ∗ = 1 + b ( 2 − C ∗ e t H ∗ C ∗ − 1 ) . (317)
Here E ∗ E_* E ∗ bounds the profile coefficients after division by their weights, C ∗ C_* C ∗ the cusp coefficients, H ∗ H_* H ∗ the genus coefficients at p ≥ 2 p\ge2 p ≥ 2 and G ∗ G_* G ∗ the descent. Since b ≥ 1 b\ge1 b ≥ 1 , C ∗ / ( 2 − C ∗ ) ⪰ 1 C_* /(2-C_*)\succeq1 C ∗ / ( 2 − C ∗ ) ⪰ 1 and e t ⪰ 1 e^t\succeq1 e t ⪰ 1 , the last equation gives G ∗ ⪰ H ∗ G_*\succeq H_* G ∗ ⪰ H ∗ ; so G ∗ G_* G ∗ majorizes η ˉ \bar{\eta} η ˉ as soon as H ∗ , G ∗ H_*,G_* H ∗ , G ∗ majorize η , ζ \eta,\zeta η , ζ . Writing E ∗ = 1 + v 1 E_*=1+v_1 E ∗ = 1 + v 1 , C ∗ = 1 + v 2 C_*=1+v_2 C ∗ = 1 + v 2 , H ∗ = 1 + v 3 H_*=1+v_3 H ∗ = 1 + v 3 , G ∗ = 1 + v 4 G_*=1+v_4 G ∗ = 1 + v 4 , the right-hand sides of (D.29) are power series in ( t , v ) (t,v) ( t , v ) with nonnegative coefficients, convergent near the origin and vanishing there. At t = 0 t=0 t = 0 the first equation has no linear term, the second and third are linear in v 1 v_1 v 1 only, and the fourth depends only on v 2 , v 3 v_2,v_3 v 2 , v 3 ; so the Jacobian at the origin is strictly lower triangular, and Lemma D.2 shows that the formal solution of (D.29) converges near t = 0 t=0 t = 0 . The inequalities (D.25)–(D.28) now give ϵ j ≤ [ E ∗ ] j \epsilon_j\le[E_*]_j ϵ j ≤ [ E ∗ ] j , γ j ≤ [ C ∗ ] j \gamma_j\le[C_*]_j γ j ≤ [ C ∗ ] j , η j ≤ [ H ∗ ] j \eta_j\le[H_*]_j η j ≤ [ H ∗ ] j and ζ j ≤ [ G ∗ ] j \zeta_j\le[G_*]_j ζ j ≤ [ G ∗ ] j for all j j j , by induction on j j j in the stage order: at order j j j the right-hand sides involve η ˉ \bar{\eta} η ˉ only through orders below j j j , or through η j \eta_j η j in the descent, and max ( η i , ζ i ) ≤ [ G ∗ ] i \max(\eta_i,\zeta_i)\le[G_*]_i max ( η i , ζ i ) ≤ [ G ∗ ] i as soon as η i ≤ [ H ∗ ] i \eta_i\le[H_*]_i η i ≤ [ H ∗ ] i and ζ i ≤ [ G ∗ ] i \zeta_i\le[G_*]_i ζ i ≤ [ G ∗ ] i . Cauchy’s estimate for the convergent solution gives ϵ j , γ j , η j , ζ j ≤ C B 0 j \epsilon_j,\gamma_j,\eta_j,\zeta_j\le CB_0^j ϵ j , γ j , η j , ζ j ≤ C B 0 j for all j j j , with C , B 0 C,B_0 C , B 0 depending only on the system. □
Tails over topological types and first-order estimates
We next bound the tails of the sums over topological types, before truncating any expansion, and deduce first-order estimates for the true quantities. In the sums S g , p S_{g,p} S g , p and T g , p T_{g,p} T g , p of (73) , orient the smaller-complexity component; its loss is its complexity plus one in S g , p S_{g,p} S g , p and its complexity in T g , p T_{g,p} T g , p . Write S g , p [ L ] S_{g,p}^{[L]} S g , p [ L ] and T g , p [ L ] T_{g,p}^{[L]} T g , p [ L ] for the parts of these sums in which the loss is at least L L L ; for p = 1 p=1 p = 1 , S g , 1 S_{g,1} S g , 1 is the one-marking separating sum of (81) .
Lemma D.5 (Tails over topological types). Fix M 0 > 0 M_0>0 M 0 > 0 . There are C M 0 , B M 0 C_{M_0},B_{M_0} C M 0 , B M 0 such that, for all sufficiently large h h h and all 1 ≤ p ≤ M 0 h 1\le p\le M_0\sqrt{h} 1 ≤ p ≤ M 0 h ,
S g , p [ L ] ≤ C M 0 ( B M 0 ( p + L ) h ) L , 2 ≤ L ≤ h , S_{g,p}^{[L]}\le C_{M_0}\left(\frac{B_{M_0}(p+L)}{h}\right)^L,\qquad2\le L\le\sqrt{h}, S g , p [ L ] ≤ C M 0 ( h B M 0 ( p + L ) ) L , 2 ≤ L ≤ h , T g , p [ L ] ≤ C M 0 ( B M 0 ( p + L ) h ) L , 1 ≤ L ≤ h . T_{g,p}^{[L]}\le C_{M_0}\left(\frac{B_{M_0}(p+L)}{h}\right)^L,\qquad1\le L\le\sqrt{h}. T g , p [ L ] ≤ C M 0 ( h B M 0 ( p + L ) ) L , 1 ≤ L ≤ h . the second bound being used only for p ≥ 2 p\ge2 p ≥ 2 . The ranges of L L L start at the stability minima two and one respectively.
Proof. Note first that, before specializing to fixed markings, Lemmas 5.3 and 5.6, with zero count and marked count both p p p , give, using the common local limit (36) ,
log W g , p = log Γ ( h ) − h log A + O M 0 ( h log ( h + 1 ) + log 2 ( h + 1 ) ) . \log W_{g,p}=\log\Gamma(h)-h\log A+O_{M_0}\left(\sqrt{h}\log(h+1)+\log^2(h+1)\right). log W g , p = log Γ ( h ) − h log A + O M 0 ( h log ( h + 1 ) + log 2 ( h + 1 ) ) .
Indeed the harmonic probability has displacement − p / 2 = O ( h ) -p/2=O(\sqrt{h}) − p /2 = O ( h ) and lies between c M 0 / h c_{M_0}/\sqrt{h} c M 0 / h and one; the terminal factors cost O ( p log ( p + 1 ) + ( p + log ( h + 1 ) ) log ( h + 1 ) ) O(p\log(p+1)+(p+\log(h+1))\log(h+1)) O ( p log ( p + 1 ) + ( p + log ( h + 1 )) log ( h + 1 )) . For losses between ϵ h \epsilon h ϵ h and half the parent complexity, Stirling entropy dominates this error. Both children have comparable complexity and O ( h ) O(\sqrt{h}) O ( h ) markings. The 2 p 2^p 2 p subsets cost only e O ( h ) e^{O(\sqrt{h})} e O ( h ) , so their total is O ( e − c ϵ h ) O(e^{-c_\epsilon h}) O ( e − c ϵ h ) . For r ≤ ϵ h r\le\epsilon h r ≤ ϵ h , (71) –(72) give the complementary ratio at most ( C / h ) r (C/h)^r ( C / h ) r . If its marking count is one, insert 2 j 0 + 1 2j_0+1 2 j 0 + 1 cusps before raising genus j 0 j_0 j 0 times, then the remaining s − 1 s-1 s − 1 cusps; this uses p ≥ 2 p\ge2 p ≥ 2 and avoids a one-marking genus comparison. The same subset/Gamma bound gives a r = [ C ( p + r ) / h ] r a_r=[C(p+r)/h]^r a r = [ C ( p + r ) / h ] r , after absorbing the number of types into C r C^r C r . Their consecutive ratio satisfies
a r + 1 a r ≤ e C ( p + r + 1 ) h ≤ e C ( M 0 h + ϵ + 1 h ) . (318) \frac{a_{r+1}}{a_r}\le\frac{eC(p+r+1)}{h}\le eC\left(\frac{M_0}{\sqrt{h}}+\epsilon+\frac{1}{h}\right). \tag*{(318)} a r a r + 1 ≤ h e C ( p + r + 1 ) ≤ e C ( h M 0 + ϵ + h 1 ) . (318)
Choose ε \varepsilon ε first, and then a lower bound on h h h depending on M 0 M_0 M 0 , so this is at most 1 / 2 1/2 1/2 . Thus the sum over L ≤ r ≤ ε h L \le r \le\varepsilon h L ≤ r ≤ ε h is at most 2 a L 2a_L 2 a L . The macroscopic tail also fits (D.30): for 1 ≤ L ≤ h 1 \le L \le\sqrt{h} 1 ≤ L ≤ h ,
( p + L h ) L ≥ exp { − h log h } , \left(\frac{p+L}{h}\right)^L \ge\exp\{-\sqrt{h}\log h\}, ( h p + L ) L ≥ exp { − h log h } ,
which eventually exceeds e − c ε h e^{-c_\varepsilon h} e − c ε h . Enlarging the constants covers the bounded remaining complexities and proves the asserted bound. At p = 1 p=1 p = 1 , compare to T g , 2 T_{g,2} T g , 2 : two uses of (71) and the proved fixed-one-marking cusp ratio bound multiply its summand by at most C / j 0 C/j_0 C / j 0 for j 0 ≤ g / 2 j_0 \le g/2 j 0 ≤ g /2 . The loss is 2 j 0 2j_0 2 j 0 in both sums. This proves (D.30), including p = 1 p=1 p = 1 . □
Lemma D.6 (First-order estimates). Fix M 0 > 0 M_0>0 M 0 > 0 . For all sufficiently large h h h , all 1 ≤ p ≤ M 0 h 1 \le p \le M_0\sqrt{h} 1 ≤ p ≤ M 0 h and all degree tuples,
0 ≤ 1 − c g , p ( d ) ≤ min { 1 , C M 0 D ( D + p + 1 ) / h } , ∣ C p ( h ) − 1 ∣ + ∣ G p ( h ) − 1 ∣ ≤ C M 0 ( p + 1 ) / h . 0 \le1-c_{g,p}(\mathbf{d}) \le\min\{1,C_{M_0}D(D+p+1)/h\},\qquad|C_p(h)-1|+|G_p(h)-1| \le C_{M_0}(p+1)/h. 0 ≤ 1 − c g , p ( d ) ≤ min { 1 , C M 0 D ( D + p + 1 ) / h } , ∣ C p ( h ) − 1∣ + ∣ G p ( h ) − 1∣ ≤ C M 0 ( p + 1 ) / h . Proof. Subtract the recursion at adjacent degrees, as in the proof of (78) , and bound its separating part by Lemma D.5 at the stability minima. For the profile, each increment costs C [ ( D + p + 1 ) / h + ( D + 1 ) ( p + 2 ) 2 / h 2 ] C[(D+p+1)/h+(D+1)(p+2)^2/h^2] C [( D + p + 1 ) / h + ( D + 1 ) ( p + 2 ) 2 / h 2 ] ; sum the path. Dilaton sums the child bound against absolute b d b_d b d moments, KdV uses degree one and T [ 1 ] T^{[1]} T [ 1 ] , and the exact descent gives p = 1 p=1 p = 1 . These estimates concern the true quantities, in a range of marking counts that grows with h h h . □
Remainder bounds: proof of Proposition D.3
We now compare the true quantities with their truncated formal series. With the same normalization as for the coefficients, the remainders satisfy a linear majorant system whose kernels have no constant term, and induction on the order bounds them geometrically.
Before the proof we record that the coefficient rules, which contain infinite sums over the kernel index ℓ \ell ℓ , over node degrees and over dilaton degrees, reduce at each order to finitely many kernel data, and that the coefficients grow polynomially in the degrees. For an integer b ≥ 0 b \ge0 b ≥ 0 , a stratum of threshold b b b in N p \mathbb{N}^p N p is the set of tuples obtained by prescribing, for each coordinate, either a value in { 0 , … , b − 1 } \{0,\ldots,b-1\} { 0 , … , b − 1 } or membership in [ b , ∞ ) [b,\infty) [ b , ∞ ) ; the coordinates of the second kind are called free. A function on N p \mathbb{N}^p N p is polynomial on the strata of threshold b b b if its restriction to each stratum is a polynomial in the free coordinates. The path in (D.17) is taken to be the one that raises d 1 d_1 d 1 from 0 0 0 to its value, then d 2 d_2 d 2 , and so on; path independence is proved in Proposition D.3 below.
Lemma D.7 (Polynomial stratification). Define b 0 = 0 b_0=0 b 0 = 0 and b K = 2 b K − 1 + 2 K + 4 b_K=2b_{K-1}+2K+4 b K = 2 b K − 1 + 2 K + 4 for K ≥ 1 K \ge1 K ≥ 1 . For every K ≥ 0 K \ge0 K ≥ 0 and p ≥ 1 p \ge1 p ≥ 1 :
(i) the coefficient d ↦ e p , K ( d ) \mathbf{d}\mapsto e_{p,K}(\mathbf{d}) d ↦ e p , K ( d ) is polynomial on the strata of threshold b K b_K b K , of total degree at most 2 K 2K 2 K in the free coordinates;
(ii) for K ≥ 1 K \ge1 K ≥ 1 and each direction i i i , the adjacent coefficient d ↦ [ t K ] ( A p , i + N p , i + S p , i ) ( d ) \mathbf{d}\mapsto[t^K](A_{p,i}+N_{p,i}+S_{p,i})(\mathbf{d}) d ↦ [ t K ] ( A p , i + N p , i + S p , i ) ( d ) is polynomial on the strata of threshold b K b_K b K , of total degree at most 2 K − 1 2K-1 2 K − 1 ;
(iii) the coefficients of the polynomials in (i) and (ii), and the scalar coefficients C p , K C_{p,K} C p , K , G p , K G_{p,K} G p , K , are finite combinations of the kernel moments ∑ ℓ ≥ ℓ 0 w ℓ ℓ m \sum_{\ell\ge\ell_0} w_\ell\ell^m ∑ ℓ ≥ ℓ 0 w ℓ ℓ m and ∑ d ≥ d 0 b d d m \sum_{d\ge d_0} b_d d^m ∑ d ≥ d 0 b d d m with ℓ 0 , d 0 , m ≤ C ( K , p ) \ell_0,d_0,m \le C(K,p) ℓ 0 , d 0 , m ≤ C ( K , p ) , of finitely many w ℓ w_\ell w ℓ and b d b_d b d , of exact amplitudes of complexity at most K K K , of A A A , and of scalar coefficients of lower than K K K for the descent rule in (D.18), also of the current-order scalar coefficients C 1 , K C_{1,K} C 1 , K , C 3 , K C_{3,K} C 3 , K , G 2 , K G_{2,K} G 2 , K , which precede G 1 , K G_{1,K} G 1 , K in the stage order.
In particular, for fixed p p p and K K K the rules (D.15)–(D.18) use finitely many kernel data, and ∣ e p , K ( d ) ∣ ≤ C p , K ( D + 1 ) 2 K |e_{p,K}(\mathbf{d})| \le C_{p,K}(D+1)^{2K} ∣ e p , K ( d ) ∣ ≤ C p , K ( D + 1 ) 2 K .
Proof. Induction on K K K . For K = 0 K=0 K = 0 all coefficients equal one. Let K ≥ 1 K \ge1 K ≥ 1 and assume (i)–(iii) for all orders below K K K . Write b = b K b=b_K b = b K and b ′ = b K − 1 b'=b_{K-1} b ′ = b K − 1 . The recursion gives
b ≥ 2 b ′ + 1 , b ≥ b ′ + 3 2 K , b ≥ 3 2 K − 1 , b ≥ b k ( k < K ) , (319) b \ge2b'+1,\qquad b \ge b'+\frac{3}{2}K,\qquad b \ge\frac{3}{2}K-1,\qquad b \ge b_k\quad(k<K), \tag*{(319)} b ≥ 2 b ′ + 1 , b ≥ b ′ + 2 3 K , b ≥ 2 3 K − 1 , b ≥ b k ( k < K ) , (319)
the first three because b − 2 b ′ = 2 K + 4 b-2b'=2K+4 b − 2 b ′ = 2 K + 4 and b ′ ≥ 0 b'\ge0 b ′ ≥ 0 , the last because ( b k ) (b_k) ( b k ) is increasing. Fix a stratum Σ \Sigma Σ of threshold b b b and a direction i i i ; a coordinate is low if it is prescribed on Σ \Sigma Σ and free otherwise. Every free coordinate is at least b ≥ b k b\ge b_k b ≥ b k for all k < K k<K k < K , and every low coordinate has a fixed value. Hence, for a child tuple obtained from d ∈ Σ \mathbf d\in\Sigma d ∈ Σ by one of the rules, the stratum of threshold b k b_k b k in which it lies is determined by Σ \Sigma Σ as soon as the modified coordinates are controlled; we check these coordinates rule by rule.
Merger. In A p , i \mathrm{A}_{p,i} A p , i the child d i j , ℓ \mathbf d^{ij,\ell} d ij , ℓ replaces d i , d j d_i,d_j d i , d j by d i + d j + ℓ − 1 d_i+d_j+\ell-1 d i + d j + ℓ − 1 and keeps the other coordinates, and the child coefficient has order k ≤ K − 1 k\le K-1 k ≤ K − 1 . If d i d_i d i or d j d_j d j is free, then d i + d j + ℓ − 1 ≥ b − 1 ≥ b k d_i+d_j+\ell-1\ge b-1\ge b_k d i + d j + ℓ − 1 ≥ b − 1 ≥ b k for every ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 by the first inequality of (D.31), the lower summation limit max ( 0 , 1 − d i − d j ) \max(0,1-d_i-d_j) max ( 0 , 1 − d i − d j ) is zero, and by induction e p − 1 , k ( d i j , ℓ ) e_{p-1,k}(\mathbf d^{ij,\ell}) e p − 1 , k ( d ij , ℓ ) is a polynomial of degree at most 2 k 2k 2 k in d i + d j + ℓ − 1 d_i+d_j+\ell-1 d i + d j + ℓ − 1 and in the other free coordinates; the sum over ℓ \ell ℓ against w ℓ w_\ell w ℓ converges absolutely by (D.20) and yields a polynomial in d i , d j d_i,d_j d i , d j and the free coordinates whose coefficients are the moments ∑ ℓ ≥ 0 w ℓ ℓ m \sum_{\ell\ge0}w_\ell\ell^m ∑ ℓ ≥ 0 w ℓ ℓ m , m ≤ 2 k m\le2k m ≤ 2 k . The factor 2 d j + 1 2d_j+1 2 d j + 1 raises the degree to at most 2 k + 1 ≤ 2 K − 1 2k+1\le2K-1 2 k + 1 ≤ 2 K − 1 . If d i d_i d i and d j d_j d j are both low, d i + d j d_i+d_j d i + d j is a fixed value; the finitely many ℓ \ell ℓ with d i + d j + ℓ − 1 < b k d_i+d_j+\ell-1<b_k d i + d j + ℓ − 1 < b k give fixed child tuples, and the remaining ℓ \ell ℓ give the polynomial form with the tail moments ∑ ℓ ≥ ℓ 0 w ℓ ℓ m \sum_{\ell\ge\ell_0}w_\ell\ell^m ∑ ℓ ≥ ℓ 0 w ℓ ℓ m .
Nonseparating pairs. In N p , i \mathrm{N}_{p,i} N p , i the child d i ; u , v \mathbf d^{i;u,v} d i ; u , v replaces d i d_i d i by the pair ( u , v ) (u,v) ( u , v ) with u + v = d i + ℓ − 2 u+v=d_i+\ell-2 u + v = d i + ℓ − 2 . Suppose d i d_i d i is free. Then d i ≥ b ≥ 2 b k + 1 d_i\ge b\ge2b_k+1 d i ≥ b ≥ 2 b k + 1 , so for every ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 the pairs fall into three ranges: u < b k u<b_k u < b k , in which case v = d i + ℓ − 2 − u ≥ ( 2 b k + 1 ) − 2 − ( b k − 1 ) = b k v=d_i+\ell-2-u\ge(2b_k+1)-2-(b_k-1)=b_k v = d i + ℓ − 2 − u ≥ ( 2 b k + 1 ) − 2 − ( b k − 1 ) = b k ; v < b k v<b_k v < b k , in which case u ≥ b k u\ge b_k u ≥ b k ; and u , v ≥ b k u,v\ge b_k u , v ≥ b k . The first two ranges consist of b k b_k b k fixed low values each and do not overlap, since u , v ≤ b k − 1 u,v\le b_k-1 u , v ≤ b k − 1 would force u + v ≤ 2 b k − 2 < d i + ℓ − 2 u+v\le2b_k-2<d_i+\ell-2 u + v ≤ 2 b k − 2 < d i + ℓ − 2 . On each of them the child coefficient is, by induction, a polynomial of degree at most 2 k 2k 2 k in the free coordinate v v v (respectively u u u ) and the other free coordinates. On the third range it is a polynomial P P P of degree at most 2 k 2k 2 k in ( u , v ) (u,v) ( u , v ) , and
∑ u = b k d i + ℓ − 2 − b k P ( u , d i + ℓ − 2 − u ) \sum_{u=b_k}^{d_i+\ell-2-b_k}P(u,d_i+\ell-2-u) u = b k ∑ d i + ℓ − 2 − b k P ( u , d i + ℓ − 2 − u )
is, by Faulhaber’s formula, a polynomial of degree at most 2 k + 1 2k+1 2 k + 1 in d i + ℓ − 2 d_i+\ell-2 d i + ℓ − 2 ; the summation range is consistent with this polynomial form because d i + ℓ − 2 ≥ 2 b k − 1 d_i+\ell-2\ge2b_k-1 d i + ℓ − 2 ≥ 2 b k − 1 . Summation against w ℓ w_\ell w ℓ over ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 gives a polynomial of degree at most 2 k + 1 ≤ 2 K − 1 2k+1\le2K-1 2 k + 1 ≤ 2 K − 1 in d i d_i d i and the other free coordinates, with moment coefficients. If d i d_i d i is low, the finitely many ℓ \ell ℓ with d i + ℓ − 2 < 2 b k + 1 d_i+\ell-2<2b_k+1 d i + ℓ − 2 < 2 b k + 1 give finite sums of polynomials of degree at most 2 K − 1 2K-1 2 K − 1 in the other free coordinates, and the remaining ℓ \ell ℓ are treated as before with tail moments.
Retained splits. In S p , i \mathrm{S}_{p,i} S p , i the bounded component has genus j j j , label set I I I with ∣ I ∣ = s |I|=s ∣ I ∣ = s , 2 ≤ 2 j + s ≤ K 2\le2j+s\le K 2 ≤ 2 j + s ≤ K , node degree u u u and amplitude W j , s + 1 ( u , d I ) W_{j,s+1}(u,\mathbf d_I) W j , s + 1 ( u , d I ) , which vanishes unless u + ∣ d ∣ ≤ 3 j + s − 2 ≤ 3 2 K − 2 u+|\mathbf d|\le3j+s-2\le\frac{3}{2}K-2 u + ∣ d ∣ ≤ 3 j + s − 2 ≤ 2 3 K − 2 . By the third inequality of (D.31) a free coordinate cannot belong to I I I , so the admissible sets I I I and the values d I \mathbf d_I d I are fixed on Σ \Sigma Σ , and u u u ranges over at most 3 2 K − 1 \frac{3}{2}K-1 2 3 K − 1 values. The large child ( v , d J ) (v,\mathbf d_J) ( v , d J ) has v = d i + ℓ − 2 − u v=d_i+\ell-2-u v = d i + ℓ − 2 − u . If d i d_i d i is free, the second inequality of (D.31) gives v ≥ b − 2 − ( 3 2 K − 2 ) ≥ b k v\ge b-2-\left(\frac{3}{2}K-2\right)\ge b_k v ≥ b − 2 − ( 2 3 K − 2 ) ≥ b k for all ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 , so the child coefficient is a polynomial of degree at most 2 k 2k 2 k in v v v and in the free coordinates of d J \mathbf d_J d J , and summation against w ℓ w_\ell w ℓ gives a polynomial of degree at most 2 k ≤ 2 K − 2 2k\le2K-2 2 k ≤ 2 K − 2 in d i d_i d i . If d i d_i d i is low, finitely many ℓ \ell ℓ are exceptional and the rest are polynomial with tail moments. The prefactor R p ; j , s R_{p;j,s} R p ; j , s involves A A A and scalar coefficients of order at most K − 2 j − s < K K-2j-s<K K − 2 j − s < K only.
Scalars and shifts. The scalar and shift coefficients in (D.15)–(D.16) do not depend on the degrees. This proves (ii), with degree at most 2 K − 1 2K-1 2 K − 1 .
Path sum. By (D.17), e p , K ( d ) e_{p,K}(\mathbf d) e p , K ( d ) is minus the sum of the adjacent coefficients along the chosen path. While d i d_i d i is raised, the states visited are ( d 1 , … , d i − 1 , m , 0 , … , 0 ) (d_1,\ldots,d_{i-1},m,0,\ldots,0) ( d 1 , … , d i − 1 , m , 0 , … , 0 ) , 0 ≤ m < d i 0\le m<d_i 0 ≤ m < d i , and their strata of threshold b b b are determined by Σ \Sigma Σ and by whether m < b m<b m < b . If d i d_i d i is free, the sum over m m m is, for m < b m<b m < b , a finite sum of polynomials of degree at most 2 K − 1 2K-1 2 K − 1 in the free coordinates among d 1 , … , d i − 1 d_1,\ldots,d_{i-1} d 1 , … , d i − 1 , plus, by (ii) and Faulhaber’s formula, a polynomial of degree at most 2 K 2K 2 K in d i d_i d i and in the free coordinates among d 1 , … , d i − 1 d_1,\ldots,d_{i-1} d 1 , … , d i − 1 for b ≤ m < d i b\le m<d_i b ≤ m < d i ; if d i d_i d i is low it is a finite sum. Hence e p , K e_{p,K} e p , K is polynomial on the stratum of threshold b b b , of total degree at most 2 K 2K 2 K , which is (i).
Dilaton, KdV and descent. In (D.17), C p , K C_{p,K} C p , K is a finite combination of the values e p + 1 , k ( d , 0 p ) e_{p+1,k}(d,0^p) e p + 1 , k ( d , 0 p ) , k ≤ K k \le K k ≤ K , against b d b_d b d and shift coefficients. By (i) at orders k ≤ K k \le K k ≤ K , d ↦ e p + 1 , k ( d , 0 p ) d \mapsto e_{p+1,k}(d,0^p) d ↦ e p + 1 , k ( d , 0 p ) is a polynomial of degree at most 2 k 2k 2 k for d ≥ b k d \ge b_k d ≥ b k and takes finitely many values for d < b k d < b_k d < b k , so the sum over d d d uses the signed moments ∑ d ≥ b k b d d m \sum_{d\ge b_k}b_d d^m ∑ d ≥ b k b d d m , m ≤ 2 k m \le2k m ≤ 2 k , and finitely many b d b_d b d . In (D.18), P p ( e 1 ; t ) P_p(e_1;t) P p ( e 1 ; t ) is evaluated at a single tuple, the amplitudes W j , s + 2 ( 0 ) \mathcal{W}_{j,s+2}(0) W j , s + 2 ( 0 ) have complexity 2 j + s ≤ K 2j+s \le K 2 j + s ≤ K , and the descent is a finite product of scalar series. This proves (iii). The final assertion follows from (i), since there are finitely many strata and each free coordinate is at most D D D . □
Proof of Proposition D.3. The coefficient bounds are Lemma D.4, with S j 2 j ≤ ( D + p + M j ) 2 j S_j^{2j} \le(D+p+Mj)^{2j} S j 2 j ≤ ( D + p + M j ) 2 j . It remains to prove the remainder bounds (D.19), the path independence and the finiteness of the coefficient rules.
For L ≥ 1 L \ge1 L ≥ 1 define the normalized remainders
E L = sup ∣ c g , p ( d ) − ∑ j < L e p , j ( d ) h − j ∣ S L 2 L h − L , C L = sup ∣ C p ( h ) − ∑ j < L C p , j h − j ∣ S L 2 L h − L , \mathcal{E}_L=\sup\frac{\left|c_{g,p}(\mathbf{d})-\sum_{j<L}e_{p,j}(\mathbf{d})h^{-j}\right|}{S_L^{2L}h^{-L}},\qquad\mathcal{C}_L=\sup\frac{\left|C_p(h)-\sum_{j<L}C_{p,j}h^{-j}\right|}{S_L^{2L}h^{-L}}, E L = sup S L 2 L h − L c g , p ( d ) − ∑ j < L e p , j ( d ) h − j , C L = sup S L 2 L h − L C p ( h ) − ∑ j < L C p , j h − j ,
and H L , G L \mathcal{H}_L,\mathcal{G}_L H L , G L in the same way for G p G_p G p with p ≥ 2 p \ge2 p ≥ 2 and with p = 1 p=1 p = 1 ; the suprema are over real Y ∈ [ 0 , y 0 ] Y \in[0,y_0] Y ∈ [ 0 , y 0 ] , all stable states with h ≥ 1 h \ge1 h ≥ 1 , p ≥ 1 p \ge1 p ≥ 1 , and all degree tuples, invalid tuples having profile zero. We prove by induction on L L L , in the stage order profile, cusp, genus, descent, that these numbers are finite and at most C B L CB^L C B L ; this is (D.19).
Large weights. Put δ = S L 2 / h \delta=S_L^2/h δ = S L 2 / h and fix δ 0 ∈ ( 0 , 1 ] \delta_0 \in(0,1] δ 0 ∈ ( 0 , 1 ] with B 0 δ 0 < 1 / 2 B_0\delta_0<1/2 B 0 δ 0 < 1/2 . The exact quantities are bounded: 0 ≤ c ≤ 1 0 \le c \le1 0 ≤ c ≤ 1 , ∣ C p ( h ) ∣ ≤ C \lvert C_p(h)\rvert\le C ∣ C p ( h )∣ ≤ C by (6.8), G p ( h ) ≤ 1 G_p(h)\le1 G p ( h ) ≤ 1 for p ≥ 2 p\ge2 p ≥ 2 by (6.9), and G 1 ( h ) ≤ C G_1(h)\le C G 1 ( h ) ≤ C by the case p = 1 p=1 p = 1 of (6.19). If δ ≥ δ 0 \delta\ge\delta_0 δ ≥ δ 0 , then Lemma D.4 and S j 2 j h − j ≤ δ j S_j^{2j}h^{-j}\le\delta^j S j 2 j h − j ≤ δ j (j ≤ L j\le L j ≤ L ) bound the normalized remainder by
δ − L ( C + ∑ j < L C B 0 j δ j ) ≤ C δ 0 − L ( 1 + ∑ j < L ( B 0 δ 0 ) j ) ≤ 3 C δ 0 − L , \delta^{-L}\left(C+\sum_{j<L}CB_0^j\delta^j\right)\le C\delta_0^{-L}\left(1+\sum_{j<L}(B_0\delta_0)^j\right)\le3C\delta_0^{-L}, δ − L C + j < L ∑ C B 0 j δ j ≤ C δ 0 − L 1 + j < L ∑ ( B 0 δ 0 ) j ≤ 3 C δ 0 − L ,
since δ j − L ≤ δ 0 j − L \delta^{j-L}\le\delta_0^{j-L} δ j − L ≤ δ 0 j − L for j < L j<L j < L . Invalid-degree parents satisfy D ≥ h D\ge h D ≥ h , hence δ ≥ h ≥ 1 \delta\ge h\ge1 δ ≥ h ≥ 1 , and lie in this region. In the remaining region δ < δ 0 \delta<\delta_0 δ < δ 0 one has p + M L ≤ S L < h p+ML\le S_L<\sqrt h p + M L ≤ S L < h , so every marking count that occurs, including those of children, is at most p + L + 2 ≤ 4 h p+L+2\le4\sqrt h p + L + 2 ≤ 4 h : Lemmas D.5 and D.6 apply with M 0 = 4 M_0=4 M 0 = 4 . Also h > M 2 L 2 / δ 0 h>M^2L^2/\delta_0 h > M 2 L 2 / δ 0 , and shrinking δ 0 \delta_0 δ 0 once makes Lemma D.6 give C p ( h ) ≥ 1 / 2 C_p(h)\ge1/2 C p ( h ) ≥ 1/2 wherever a cusp ratio is inverted. From now on δ < δ 0 \delta<\delta_0 δ < δ 0 .
Products of exact factors. Every term of the exact adjacent difference, of the exact dilaton identity, of the exact KdV identity and of the exact descent is a product of exact factors: a prefactor A r ∏ ( h − c k ) − 1 A^r\prod(h-c_k)^{-1} A r ∏ ( h − c k ) − 1 , cusp and genus ratios at shifted complexities, and profiles of children. Each exact factor F F F has, for every q ≥ 1 q\ge1 q ≥ 1 , a decomposition F = T < q F f o r m + R F ( q ) F=T_{<q}F^{\mathrm{form}}+R_F(q) F = T < q F form + R F ( q ) into its truncated formal series and a remainder, and the formal rule at order L L L is the truncation of the product of the formal series. The product of exact factors and the truncated formal product differ by the finite identity
∏ k F k − T < q ∏ k F k f o r m = R F 1 ( q ) ∏ k > 1 F k + ∑ j < q f 1 , j h − j ( ∏ k > 1 F k − T < q − j ∏ k > 1 F k f o r m ) . (320) \prod_k F_k-T_{<q}\prod_k F_k^{\mathrm{form}}=R_{F_1}(q)\prod_{k>1}F_k+\sum_{j<q}f_{1,j}h^{-j}\left(\prod_{k>1}F_k-T_{<q-j}\prod_{k>1}F_k^{\mathrm{form}}\right). \tag*{(320)} k ∏ F k − T < q k ∏ F k form = R F 1 ( q ) k > 1 ∏ F k + j < q ∑ f 1 , j h − j ( k > 1 ∏ F k − T < q − j k > 1 ∏ F k form ) . (320)
applied again to the bracket on the right, and so on. Each resulting term contains exactly one remainder R F k ( q k ) R_{F_k}(q_k) R F k ( q k ) , multiplied by known coefficients of the factors before it and by exact factors after it. For an inverse, with known formal inverse coefficients d j d_j d j ,
C T < q ( 1 / C ) − 1 = ∑ j < q d j h − j ( C − T < q − j C ) . (321) C T_{<q}(1/C)-1=\sum_{j<q}d_jh^{-j}(C-T_{<q-j}C). \tag*{(321)} C T < q ( 1/ C ) − 1 = j < q ∑ d j h − j ( C − T < q − j C ) . (321)
so the remainder of $1/Ci s is i s 1/Ct i m e s a s u m o f k n o w n c o e f f i c i e n t s a g a i n s t r e m a i n d e r s o f times a sum of known coefficients against remainders of t im es a s u m o f k n o w n coe f f i c i e n t s a g ain s t r e main d er so f C.
Bounding one term. Normalize by S L 2 L h − L S_L^{2L}h^{-L} S L 2 L h − L . A known coefficient of order j j j of a factor with weight S ′ S' S ′ is at most C B 0 j ( S ′ ) 2 j CB_0^j(S')^{2j} C B 0 j ( S ′ ) 2 j by Lemma D.4, where S ′ ≤ S L + ℓ S'\le S_L+\ell S ′ ≤ S L + ℓ , S L + d + 1 S_L+d+1 S L + d + 1 , S L + 2 S_L+2 S L + 2 or S L S_L S L by (a) of Subsection D.5, applied with j ′ ≤ L j'\le L j ′ ≤ L . A remainder R F k ( q k ) R_{F_k}(q_k) R F k ( q k ) of a factor at complexity h − c h-c h − c , ∣ c ∣ ≤ L \lvert c\rvert\le L ∣ c ∣ ≤ L , is at most X q k ( S ′ ) 2 q k ( h − c ) − q k \mathcal{X}_{q_k}(S')^{2q_k}(h-c)^{-q_k} X q k ( S ′ ) 2 q k ( h − c ) − q k , where X \mathcal{X} X is the error sequence of its type, and ( h − c ) − q k ≤ h − q k e 2 ∣ c ∣ q k / h ≤ e 2 δ 0 / M 2 h − q k (h-c)^{-q_k} \le h^{-q_k}e^{2|c|q_k/h} \le e^{2\delta_0/M^2}h^{-q_k} ( h − c ) − q k ≤ h − q k e 2∣ c ∣ q k / h ≤ e 2 δ 0 / M 2 h − q k because q k ≤ L q_k \le L q k ≤ L and L 2 / h ≤ δ 0 / M 2 L^2/h \le\delta_0/M^2 L 2 / h ≤ δ 0 / M 2 . The exact factors after the remainder are bounded by the constants above. The known shift tails are bounded as follows. If the orders of the known coefficients and of the prefactor sum to a = L − q a=L-q a = L − q and the first omitted shift degree is q q q , the elementary inequality
∑ n ≥ q ( a ) n n ! x n ≤ ( a ) q q ! x q ( 1 − x ) − a − q ( a ≥ 1 , 0 ≤ x < 1 ) , \sum_{n\ge q}\frac{(a)_n}{n!}x^n \le\frac{(a)_q}{q!}x^q(1-x)^{-a-q}\qquad(a\ge1,\ 0\le x<1), n ≥ q ∑ n ! ( a ) n x n ≤ q ! ( a ) q x q ( 1 − x ) − a − q ( a ≥ 1 , 0 ≤ x < 1 ) ,
which follows from ( a ) q + k = ( a ) q ( a + q ) k (a)_{q+k}=(a)_q(a+q)_k ( a ) q + k = ( a ) q ( a + q ) k and ( q + k ) ! ≥ q ! k ! (q+k)!\ge q!k! ( q + k )! ≥ q ! k ! , bounds the tail by ( L − 1 q ) ( c / h ) q ( 1 − c / h ) − L \binom{L-1}{q}(c/h)^q(1-c/h)^{-L} ( q L − 1 ) ( c / h ) q ( 1 − c / h ) − L with c ≤ L c\le L c ≤ L ; after division by S L 2 q h − q S_L^{2q}h^{-q} S L 2 q h − q this is at most e 2 δ 0 / M 2 / q ! e^{2\delta_0/M^2}/q! e 2 δ 0 / M 2 / q ! , since c ( L − 1 ) ≤ L 2 ≤ S L 2 c(L-1)\le L^2\le S_L^2 c ( L − 1 ) ≤ L 2 ≤ S L 2 and ( 1 − c / h ) − L ≤ e 2 c L / h (1-c/h)^{-L}\le e^{2cL/h} ( 1 − c / h ) − L ≤ e 2 c L / h in the small-weight region. The counts, kernel moments and path sums are handled exactly as in the proof of Lemma D.4 , with the same powers of S L S_L S L : a linear term supplies S L 2 S_L^2 S L 2 against the prefactor order one, and a separating term of loss r r r supplies S L r + 2 ≤ S L 2 r S_L^{r+2}\le S_L^{2r} S L r + 2 ≤ S L 2 r (r ≥ 2 r\ge2 r ≥ 2 ) in the profile rule and S L r ≤ S L 2 r S_L^r\le S_L^{2r} S L r ≤ S L 2 r (r ≥ 1 r\ge1 r ≥ 1 ) in the KdV rule.
Forcing terms. Three contributions do not involve the error sequences. The exact topological types of loss r ≥ L r\ge L r ≥ L , which the formal rule truncated below order L L L omits, have total mass bounded by (D.30) at L L L ; after the kernel sum and the coordinate path their profile contribution is at most CD(D+1)(B(p+L)/h)^L\le CBL^S_L^{L+2}h^{-L}\le CBL^SS_L^{2L}h^{-L} for L ≥ 2 L\ge2 L ≥ 2 , and the KdV contribution has no path or pair factor. The difference between the exact kernel sum over ℓ ≤ m \ell\le m ℓ ≤ m and the formal sum over all ℓ \ell ℓ , and the macroscopic cutoffs in the proof of Lemma D.5, are O ( e − c h ) O(e^{-ch}) O ( e − c h ) , and e − c h ≤ C L L ! h − L ≤ C L S L 2 L h − L e^{-ch}\le C^LL!h^{-L}\le C^LS_L^{2L}h^{-L} e − c h ≤ C L L ! h − L ≤ C L S L 2 L h − L . At L = 1 L=1 L = 1 the normalized remainders are bounded by Lemma D.6.
The linear majorant system. Let F 0 = C / ( 1 − B 0 t ) F_0=C/(1-B_0t) F 0 = C / ( 1 − B 0 t ) be the majorant of the normalized known coefficients from Lemma D.4 , and V = C e x + F 0 V=C_{\mathrm{ex}}+F_0 V = C ex + F 0 , where C e x ≥ 1 C_{\mathrm{ex}}\ge1 C ex ≥ 1 bounds all exact factors and 1 / C 1/C 1/ C in the region δ < δ 0 \delta<\delta_0 δ < δ 0 . Collecting the bounds above, the error sequences E , C , H , G \mathcal E,\mathcal C,\mathcal H,\mathcal G E , C , H , G , regarded as formal series in t t t , satisfy coefficientwise
E ⪯ F + K E R , C ⪯ F + b E , H ⪯ F + b E + K H R , G ⪯ F + b ( H + C ) + K G R , R = E + C + H + G . (322) \begin{aligned}
\mathcal E &\preceq F+K_E\mathcal R,\qquad\mathcal C\preceq F+b\mathcal E,\\
\mathcal H &\preceq F+b\mathcal E+K_H\mathcal R,\qquad\mathcal G\preceq F+b(\mathcal H+\mathcal C)+K_G\mathcal R,\qquad\mathcal R=\mathcal E+\mathcal C+\mathcal H+\mathcal G. \tag*{(322)}
\end{aligned} E H ⪯ F + K E R , C ⪯ F + b E , ⪯ F + b E + K H R , G ⪯ F + b ( H + C ) + K G R , R = E + C + H + G . (322)
with the zero-constant kernels
K E = b e b t [ t V 2 + ∑ r ≥ 2 ( 2 r + 1 ) ( b t ) r V 2 r ] , K H = b e b t ∑ r ≥ 1 2 r ( b t ) r V 2 r , K G = b [ e b t V 3 − V ( 0 ) 3 ] , K_E=b e^{bt}\left[tV^2+\sum_{r\ge2}(2r+1)(bt)^rV^{2r}\right],\qquad K_H=b e^{bt}\sum_{r\ge1}2r(bt)^rV^{2r},\qquad K_G=b\left[e^{bt}V^3-V(0)^3\right], K E = b e b t [ t V 2 + r ≥ 2 ∑ ( 2 r + 1 ) ( b t ) r V 2 r ] , K H = b e b t r ≥ 1 ∑ 2 r ( b t ) r V 2 r , K G = b [ e b t V 3 − V ( 0 ) 3 ] ,
and the forcing
F = b V 3 e b t 1 − b t V 2 + b 1 − δ 0 − 1 t , F=\frac{bV^3e^{bt}}{1-btV^2}+\frac{b}{1-\delta_0^{-1}t}, F = 1 − b t V 2 b V 3 e b t + 1 − δ 0 − 1 t b ,
for a constant b b b depending only on the constants already fixed. The factors 2 r + 1 2r+1 2 r + 1 and 2 r 2r 2 r count the possible positions of the single remainder among the factors of a product with r r r cusp, at most r r r genus and at most one profile factor; the term b E b\mathcal E b E in the second and third lines is the current-order remainder of the profile, which enters the dilaton and KdV rules without a prefactor t t t , and b ( H + C ) b(\mathcal H+\mathcal C) b ( H + C ) plays the same role in the descent. The first summand of F F F dominates the shift tails, the omitted topology and the cutoff terms, each of which is a product of at most 2 r + 1 2r+1 2 r + 1 known or exact factors times C L q q ! C^q_Lq! C L q q ! or C L L ! / S L 2 L C^LL!/S_L^{2L} C L L ! / S L 2 L ; the second summand dominates the region δ ≥ δ 0 \delta\ge\delta_0 δ ≥ δ 0 .
Induction. Substituting the first line of (D.34) into the others gives R ⪯ A 0 F + A 0 ( K E + K H + K G ) R \mathcal R\preceq A_0F+A_0(K_E+K_H+K_G)\mathcal R R ⪯ A 0 F + A 0 ( K E + K H + K G ) R for a constant A 0 A_0 A 0 . Since K E + K H + K G K_E+K_H+K_G K E + K H + K G has no constant term, the coefficient of t L t^L t L on the right involves R j \mathcal R_j R j only for j < L j<L j < L . Let R ∗ = A 0 F / ( 1 − A 0 ( K E + K H + K G ) ) \mathcal R_* = A_0F/(1-A_0(K_E+K_H+K_G)) R ∗ = A 0 F / ( 1 − A 0 ( K E + K H + K G )) , a series with nonnegative coefficients, analytic near t = 0 t=0 t = 0 . If R j ≤ [ R ∗ ] j \mathcal R_j\le[\mathcal R_*]_j R j ≤ [ R ∗ ] j for all j < L j<L j < L , then R L ≤ [ A 0 F + A 0 ( K E + K H + K G ) R ∗ ] L = [ R ∗ ] L \mathcal R_L\le[A_0F+A_0(K_E+K_H+K_G)\mathcal R_*]_L=[\mathcal R_*]_L R L ≤ [ A 0 F + A 0 ( K E + K H + K G ) R ∗ ] L = [ R ∗ ] L by monotonicity of series with nonnegative coefficients; the case L = 1 L=1 L = 1 is the base estimate. Cauchy’s estimate for R ∗ \mathcal R_* R ∗ gives R L ≤ C B L \mathcal R_L\le C B^L R L ≤ C B L and proves (D.19); no infinite inverse-complexity series is evaluated. For each fixed degree tuple, two coordinate paths give asymptotic expansions of the same ratio, and uniqueness of asymptotic coefficients as g → ∞ g \to\infty g → ∞ proves the asserted path independence.
Finiteness of the coefficient rules. The finite coefficient rule and the polynomial degree bounds used above are the content of Lemma D.7: at each order the degree lattice admits a finite stratification by the thresholds b K b_K b K (an admissible, not optimal, choice), the coefficients are polynomial on every stratum, the adjacent coefficient has degree at most 2 K − 1 2K-1 2 K − 1 and its path sum degree at most 2 K 2K 2 K , and every displayed infinite sum in the coefficient algorithm reduces to finitely many kernel moments, low kernel coefficients and exact small amplitudes, the latter being finite polynomials in the θ ℓ \theta_\ell θ ℓ . The common analytic disks and removable critical normalization give analyticity of that data throughout [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] . □
Exponential degree norms and single-factorial bounds
This subsection and the next prove Proposition D.9 , the single-factorial form of Proposition D.3 ; it is the version used in the proof of Theorem 8.2 and, together with that theorem, in the coefficient bounds of § 3.31 . The polynomial bounds of Proposition D.3 retain useful information at every degree. For summing the dilaton equation, however, an exponential degree norm gives a sharper dependence on the expansion order: in this norm the order-j j j bounds grow like ( p + M j ) j (p+Mj)^j ( p + M j ) j rather than ( p + M j ) 2 j (p+Mj)^{2j} ( p + M j ) 2 j , with a factor e σ D e^{\sigma D} e σ D for profiles. The argument has three parts. Lemma D.8 enlarges the range of the topological-tail estimate of Lemma D.5 to linearly many markings and a linear order range, using the all-marking envelope (D.35). Lemma D.10 isolates the mechanism that saves one factorial per order. The coefficient bounds (D.36) then follow from the majorant system (D.45) and Lemma D.2 ; the actual remainders (D.37) are proved in Subsection D.10. All constants in this subsection are uniform in the real parameter Y ∈ [ 0 , y 0 ] Y \in[0,y_0] Y ∈ [ 0 , y 0 ] unless a complex neighborhood is specified.
Dependence on earlier results. The chain of dependencies is linear. The marked ratio estimates of § 3.14 feed Proposition 7.1 in § 3.19 , whose endpoint constant is fixed by (97) , that is, by the Itzykson–Zuber identity and the large-order asymptotics of Painlevé I in the form recorded by [70, Section 2.3]. Of Subsection D.11 it uses only the leading case of Lemma D.11 (iii), a Stirling computation that rests on these two inputs alone. The leading asymptotic of Proposition 7.1 gives the envelope (D.35) used in Lemma D.8 ; Proposition D.9 uses Lemma D.8 and the formal rules (D.15)–(D.18); and Theorem 8.2 follows from Proposition D.9 together with Subsection D.11, which has two independent parts. The endpoint Gevrey expansion (Lemma D.11 ) uses only the Painlevé I input and none of the results of this subsection or the next. The vanishing (D.47) of the higher cusp coefficients at y 0 y_0 y 0 uses (D.37) and the exact string-dilaton identity at the endpoint, but not the Painlevé input. Neither part is used in the proof of Proposition D.9 . Thus § 3.14 → \to → § 3.19 → \to → Lemma D.8 → \to → Proposition D.9 → \to → Theorem 8.2 contains no cycle.
Lemma D.8 (Linear marking and order ranges). There are constants η , c ∈ ( 0 , 1 ) \eta,c \in(0,1) η , c ∈ ( 0 , 1 ) and C , B > 0 C,B>0 C , B > 0 with the following properties, uniformly in Y ∈ [ 0 , y 0 ] Y \in[0,y_0] Y ∈ [ 0 , y 0 ] . For every stable ( g , p ) (g,p) ( g , p ) with p ≥ 1 p \ge1 p ≥ 1 and h = 2 g − 2 + p h=2g-2+p h = 2 g − 2 + p ,
W g , p ≤ C Γ ( h ) A − h / h . (323) W_{g,p} \le C\Gamma(h)A^{-h}/\sqrt{h}. \tag*{(323)} W g , p ≤ C Γ ( h ) A − h / h . (323) If moreover p ≤ η h p \le\eta h p ≤ η h , the bounds (D.30) hold, with the constants C , B C,B C , B , for all L ≤ c h L \le ch L ≤ c h above the same stability minima (the bound for T g , p [ L ] T^{[L]}_{g,p} T g , p [ L ] when p ≥ 2 p \ge2 p ≥ 2 ), and, for all degree tuples d \mathbf{d} d with D = ∣ d ∣ D=|\mathbf{d}| D = ∣ d ∣ ,
∣ C p ( h ) − 1 ∣ + ∣ G p ( h ) − 1 ∣ ≤ C ( p + 1 ) / h , 0 ≤ 1 − c g , p ( d ) ≤ min { 1 , C D ( D + p + 1 ) / h } . |C_p(h)-1|+|G_p(h)-1| \le C(p+1)/h,\qquad0 \le1-c_{g,p}(\mathbf{d}) \le\min\{1,CD(D+p+1)/h\}. ∣ C p ( h ) − 1∣ + ∣ G p ( h ) − 1∣ ≤ C ( p + 1 ) / h , 0 ≤ 1 − c g , p ( d ) ≤ min { 1 , C D ( D + p + 1 ) / h } . The envelope (D.35) uses only Proposition 7.1 at p ∈ { 1 , 2 } p \in\{1,2\} p ∈ { 1 , 2 } , the exact inequalities (71) –(72) and (82) at p = 1 p=1 p = 1 ; no expansion in a growing number of markings enters.
Proof. For p = 1 , 2 p=1,2 p = 1 , 2 , the leading asymptotic of Proposition 7.1, after the normalization (64) , gives
W g , p ∼ Γ ( h ) A − h / h . W_{g,p} \sim\Gamma(h)A^{-h}/\sqrt{h}. W g , p ∼ Γ ( h ) A − h / h . Indeed its leading factor is Γ ( h ) A − h A Θ / ( π h ) \Gamma(h)A^{-h}\sqrt{A\Theta}/(\pi\sqrt{h}) Γ ( h ) A − h A Θ / ( π h ) ; all powers of r r r cancel. For each fixed ( g , p ) (g,p) ( g , p ) the normalized amplitude W g , p ( Y ) W_{g,p}(Y) W g , p ( Y ) and A ( Y ) A(Y) A ( Y ) are continuous and positive on the closed interval [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] ,
the former being a polynomial with nonnegative coefficients in the kernel coefficients θ ℓ ( Y ) \theta_\ell(Y) θ ℓ ( Y ) ; hence the two-sided bound, which Proposition 7.1 gives uniformly on [ 0 , y 0 ) [0,y_0) [ 0 , y 0 ) for large g g g , extends to Y = y 0 Y=y_0 Y = y 0 , and the finitely many remaining stable pairs are covered by compactness. Genus exchange (72) gives W g , p ≤ W g + 1 , p − 2 W_{g,p} \le W_{g+1,p-2} W g , p ≤ W g + 1 , p − 2 for p ≥ 4 p \ge4 p ≥ 4 . Iterating it reduces even p p p to two and odd p p p to three. In the odd case use the bounded one-marking genus ratio (82) once to reach p = 1 p=1 p = 1 . Each other exchange has constant one. This proves (D.35). The complexity-one bases are W 0 , 3 = 1 W_{0,3}=1 W 0 , 3 = 1 and W 1 , 1 = θ 1 / 2 W_{1,1}=\theta_1/2 W 1 , 1 = θ 1 /2 . Conversely, for g ≥ 2 g \ge2 g ≥ 2 , iteration of (71) from marking count one gives
W g , p ≥ C 0 − ( p − 1 ) Γ ( h ) Γ ( 2 g − 1 ) W g , 1 ≥ c 0 e − C 1 p Γ ( h ) A − h / h . W_{g,p} \ge C_0^{-(p-1)}\frac{\Gamma(h)}{\Gamma(2g-1)}W_{g,1} \ge c_0 e^{-C_1p}\Gamma(h)A^{-h}/\sqrt{h}. W g , p ≥ C 0 − ( p − 1 ) Γ ( 2 g − 1 ) Γ ( h ) W g , 1 ≥ c 0 e − C 1 p Γ ( h ) A − h / h .
The residual power A p − 1 A^{p-1} A p − 1 is absorbed into e − C 1 p e^{-C_1p} e − C 1 p . These estimates use only the leading fixed-marking result.
Orient the smaller component as in Lemma D.5. For loss r ≤ ϵ h r \le\epsilon h r ≤ ϵ h , its complementary ratio is bounded by ( C / h ) r (C/h)^r ( C / h ) r using (71) –(72) ; the one-marking complement is handled by the same cusp-insertion order as there. The small-component and subset bound (D.24) gives a r = [ C ( p + r ) / h ] r a_r=[C(p+r)/h]^r a r = [ C ( p + r ) / h ] r , with a r + 1 / a r ≤ ϵ C ( p + r + 1 ) / h a_{r+1}/a_r \le\epsilon C(p+r+1)/h a r + 1 / a r ≤ ϵ C ( p + r + 1 ) / h . Choose ϵ \epsilon ϵ , then η \eta η , so that this is at most 1 / 2 1/2 1/2 .
For ϵ h ≤ r ≤ h / 2 + O ( 1 ) \epsilon h \le r \le h/2+O(1) ϵ h ≤ r ≤ h /2 + O ( 1 ) both child complexities are proportional to h h h . The upper and lower envelopes above and Stirling’s formula give a bound h C 2 exp { − h H ( ϵ / 2 ) + C 3 p } h^{C_2}\exp\{-hH(\epsilon/2)+C_3p\} h C 2 exp { − h H ( ϵ /2 ) + C 3 p } , where H ( t ) = − t log t − ( 1 − t ) log ( 1 − t ) H(t)=-t\log t-(1-t)\log(1-t) H ( t ) = − t log t − ( 1 − t ) log ( 1 − t ) . The label subsets cost at most 2 p 2^p 2 p , and the number of remaining types is polynomial in h h h . Shrinking η \eta η therefore bounds this whole sector by C e − a h Ce^{-ah} C e − ah for a fixed a > 0 a>0 a > 0 . Choose c < min ( ϵ , e − 1 ) c<\min(\epsilon,e^{-1}) c < min ( ϵ , e − 1 ) with c log ( 1 / c ) < a / 2 c\log(1/c)<a/2 c log ( 1/ c ) < a /2 . For L ≤ c h L\le ch L ≤ c h ,
( ( p + L ) / h ) L ≥ ( L / h ) L ≥ e − c log ( 1 / c ) h , ((p+L)/h)^L \ge(L/h)^L \ge e^{-c\log(1/c)h}, (( p + L ) / h ) L ≥ ( L / h ) L ≥ e − c l o g ( 1/ c ) h ,
so the macroscopic tail fits the same bound as ∑ r ≥ L a r \sum_{r\ge L}a_r ∑ r ≥ L a r . The comparison with T g , 2 T_{g,2} T g , 2 in Lemma D.5 also treats S g , 1 S_{g,1} S g , 1 . Enlarging constants includes bounded complexities.
The adjacent-degree argument of Lemma D.6 now costs C [ ( D + p + 1 ) / h + ( D + 1 ) ( p + 2 ) 2 / h 2 ] C[(D+p+1)/h+(D+1)(p+2)^2/h^2] C [( D + p + 1 ) / h + ( D + 1 ) ( p + 2 ) 2 / h 2 ] per increment. Put x = D ( D + p + 1 ) / h x=D(D+p+1)/h x = D ( D + p + 1 ) / h . When x ≥ 1 x\ge1 x ≥ 1 , the profile bound follows from 0 ≤ c g , p ≤ 1 0\le c_{g,p}\le1 0 ≤ c g , p ≤ 1 . When x < 1 x<1 x < 1 and D ≥ 1 D\ge1 D ≥ 1 , one has D ( p + 2 ) / h ≤ 2 D(p+2)/h\le2 D ( p + 2 ) / h ≤ 2 . Consequently the second path contribution is bounded by the first:
D ( D + 1 ) ( p + 2 ) 2 h 2 ≤ C D ( D + p + 1 ) h . \frac{D(D+1)(p+2)^2}{h^2}\le C\frac{D(D+p+1)}{h}. h 2 D ( D + 1 ) ( p + 2 ) 2 ≤ C h D ( D + p + 1 ) .
This proves the profile estimate in the linear marking range; substituting p ≤ η h p\le\eta h p ≤ η h in each increment would only give O ( D 2 p / h ) O(D^2p/h) O ( D 2 p / h ) . Dilaton sums this bound against its geometric kernel, KdV uses degree one and the loss-one tail, and descent gives the remaining scalar case. □
Proposition D.9 (Single-factorial recursive bounds). Let ρ > 1 \rho>1 ρ > 1 be as in (D.20), and fix 0 < σ < log ρ 0<\sigma<\log\rho 0 < σ < log ρ . There are C , B , M > 0 C,B,M>0 C , B , M > 0 and a complex neighborhood Ω \Omega Ω of [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] such that
∥ e p , j ( d ) ∥ Ω ≤ C B j ( p + M j ) j e σ D , (324) \lVert e_{p,j}(\mathrm{d})\rVert_\Omega\le CB^j(p+Mj)^{j}e^{\sigma D}, \tag*{(324)} ∥ e p , j ( d ) ∥ Ω ≤ C B j ( p + M j ) j e σ D , (324) ∥ C p , j ∥ Ω + ∥ G p , j ∥ Ω ≤ C B j ( p + M j ) j . \lVert C_{p,j}\rVert_\Omega+\lVert G_{p,j}\rVert_\Omega\le CB^j(p+Mj)^j. ∥ C p , j ∥ Ω + ∥ G p , j ∥ Ω ≤ C B j ( p + M j ) j . For every stable state with p ≥ 1 p\ge1 p ≥ 1 , h ≥ 1 h\ge1 h ≥ 1 , and every L ≥ 1 L\ge1 L ≥ 1 , uniformly for real Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] ,
∣ c g , p ( d ) − ∑ j < L e p , j ( d ) h − j ∣ ≤ C B L ( p + M L ) L h − L e σ D , (325) \left|c_{g,p}(\mathrm{d})-\sum_{j<L}e_{p,j}(\mathrm{d})h^{-j}\right|
\le CB^L(p+ML)^Lh^{-L}e^{\sigma D}, \tag*{(325)} c g , p ( d ) − j < L ∑ e p , j ( d ) h − j ≤ C B L ( p + M L ) L h − L e σ D , (325) ∣ C p ( h ) − ∑ j < L C p , j h − j ∣ + ∣ G p ( h ) − ∑ j < L G p , j h − j ∣ ≤ C B L ( p + M L ) L h − L . \left|C_p(h)-\sum_{j<L}C_{p,j}h^{-j}\right|
+\left|G_p(h)-\sum_{j<L}G_{p,j}h^{-j}\right|
\le CB^L(p+ML)^Lh^{-L}. C p ( h ) − j < L ∑ C p , j h − j + G p ( h ) − j < L ∑ G p , j h − j ≤ C B L ( p + M L ) L h − L . Invalid-degree profiles and invalid genus numerators are zero, as in Proposition D.3 .
We prove the coefficient estimates here and the actual remainders in Subsection D.10. The latter are needed to pass from formal factorial bounds to an exponentially accurate approximation.
Weights and normalized suprema. Fix 0 < σ 0 < σ 1 < log ρ 0 < \sigma_0 < \sigma_1 < \log\rho 0 < σ 0 < σ 1 < log ρ and an integer M ≥ max { 2 , 2 ( σ 1 − σ 0 ) } M \ge\max\{2,2(\sigma_1-\sigma_0)\} M ≥ max { 2 , 2 ( σ 1 − σ 0 )} . For 0 < a ≤ σ 1 − σ 0 0 < a \le\sigma_1-\sigma_0 0 < a ≤ σ 1 − σ 0 , use the degree exponent σ 0 + a \sigma_0+a σ 0 + a and the weights
Z j = Z j ( p , a ) = p + M j a , B j = Z j j ( j ≥ 1 ) , B 0 = 1 , (326) Z_j=Z_j(p,a)=p+\frac{Mj}{a},\qquad B_j=Z_j^j\ (j\ge1),\qquad B_0=1, \tag*{(326)} Z j = Z j ( p , a ) = p + a M j , B j = Z j j ( j ≥ 1 ) , B 0 = 1 , (326)
so that Z j ≥ M j / a ≥ 2 j Z_j \ge Mj/a \ge2j Z j ≥ M j / a ≥ 2 j , Z j ≥ p Z_j \ge p Z j ≥ p and j / Z j ≤ a / M ≤ 1 / 2 j/Z_j \le a/M \le1/2 j / Z j ≤ a / M ≤ 1/2 . An order-j j j profile coefficient at a state with p p p markings and degree sum D D D is normalized by B j e ( σ 0 + a ) D B_j e^{(\sigma_0+a)D} B j e ( σ 0 + a ) D , a scalar coefficient by B j B_j B j . Let ε j \varepsilon_j ε j , γ j \gamma_j γ j , η j \eta_j η j and ζ j \zeta_j ζ j be the suprema of the normalized profile, cusp, genus (p ≥ 2 p \ge2 p ≥ 2 ) and descent coefficients of order j j j , taken over all p p p , all admissible a a a , all degree tuples and all Y ∈ Ω Y \in\Omega Y ∈ Ω ; put ε 0 = γ 0 = η 0 = ζ 0 = 1 \varepsilon_0=\gamma_0=\eta_0=\zeta_0=1 ε 0 = γ 0 = η 0 = ζ 0 = 1 and η ˉ j = max ( η j , ζ j ) \bar{\eta}_j=\max(\eta_j,\zeta_j) η ˉ j = max ( η j , ζ j ) . At each fixed p , a p,a p , a , the polynomial estimates of Lemma D.4 ensure convergence of the kernel sums and finiteness of each coefficient. Uniform finiteness in p p p will follow by induction; it is not a consequence of the polynomial p 2 j p^{2j} p 2 j bound. The single-factorial mechanism is isolated in the following lemma: the exponential degree norm absorbs the sum along a coordinate path at the cost of one power of Z j Z_j Z j , and in separating terms the dimension of the bounded component removes the path count altogether. Concretely, a child of positive order 0 < k < j 0<k<j 0 < k < j is measured at the smaller exponent a ′ = a k / j a'=ak/j a ′ = ak / j ; since M k / a ′ = M j / a Mk/a'=Mj/a M k / a ′ = M j / a , its weight is still comparable to Z j Z_j Z j , while the exponential gap e − ( a − a ′ ) D e^{-(a-a')D} e − ( a − a ′ ) D pays for a path of length D D D at the price of one factor Z j Z_j Z j . Over j j j orders this costs Z j j ≈ ( M j / a ) j Z_j^j\approx(Mj/a)^j Z j j ≈ ( M j / a ) j , a single factorial.
Lemma D.10 (Path-sum estimate). Let j ≥ 1 j\ge1 j ≥ 1 , let C 0 , ρ C_0,\rho C 0 , ρ be as in (D.20), and let C , C 1 C,C_1 C , C 1 denote constants depending only on σ 0 , σ 1 , ρ , C 0 \sigma_0,\sigma_1,\rho,C_0 σ 0 , σ 1 , ρ , C 0 and sup Ω ∣ A ∣ \sup_\Omega|A| sup Ω ∣ A ∣ .
(i) (Linear terms.) Consider a term of [ t j ] ( A p , i + N p , i ) [t^j](A_{p,i}+N_{p,i}) [ t j ] ( A p , i + N p , i ) in (D.15) whose child profile coefficient has order k k k , whose scalar coefficients have total order i i i (one cusp coefficient in A A A ; one cusp and one genus coefficient in N N N ) and whose shift has degree q q q , so that i + k + q = j − 1 i+k+q=j-1 i + k + q = j − 1 . Bound the child profile coefficient in the norm with exponent a ′ = a k / j a'=ak/j a ′ = ak / j , so that M k / a ′ = M j / a Mk/a'=Mj/a M k / a ′ = M j / a (for k = 0 k=0 k = 0 the child profile is the constant one and no norm is needed), and write Γ i \Gamma_i Γ i for the product of the normalized scalar suprema involved (γ i \gamma_i γ i , or γ i 1 η ˉ i 2 \gamma_{i_1}\bar{\eta}_{i_2} γ i 1 η ˉ i 2 with i 1 + i 2 = i i_1+i_2=i i 1 + i 2 = i ). At a parent state with p p p markings and degree sum m m m the term is bounded by
C ( 2 m + p + 1 ) ε k Z j a / M e a / M Γ i Z j i ⋅ ( a / M ) q Z j q q ! e ( σ 0 + a ′ ) m . (327) C(2m+p+1)\,\varepsilon_k Z_j^{a/M}e^{a/M}\,\Gamma_i Z_j^i\cdot\frac{(a/M)^q Z_j^q}{q!}\,e^{(\sigma_0+a')m}. \tag*{(327)} C ( 2 m + p + 1 ) ε k Z j a / M e a / M Γ i Z j i ⋅ q ! ( a / M ) q Z j q e ( σ 0 + a ′ ) m . (327) Summing (D.39) along a coordinate path from 0 \mathbf{0} 0 to d \mathbf{d} d , with D = ∣ d ∣ D=|\mathbf{d}| D = ∣ d ∣ , gives at most
C ( D + p + 1 ) e ( σ 0 + a ′ ) D ε k Γ i e a / M ( a / M ) q q ! Z j j − 1 . (328) C(D+p+1)e^{(\sigma_0+a')D}\,\varepsilon_k\Gamma_i e^{a/M}\frac{(a/M)^q}{q!}Z_j^{j-1}. \tag*{(328)} C ( D + p + 1 ) e ( σ 0 + a ′ ) D ε k Γ i e a / M q ! ( a / M ) q Z j j − 1 . (328) Division by the parent normalization Z j j e ( σ 0 + a ) D Z_j^j e^{(\sigma_0+a)D} Z j j e ( σ 0 + a ) D leaves the factor Z j − 1 ( D + p + 1 ) e − ( a − a ′ ) D Z_j^{-1}(D+p+1)e^{-(a-a')D} Z j − 1 ( D + p + 1 ) e − ( a − a ′ ) D , and, since a − a ′ = a ( j − k ) / j a-a'=a(j-k)/j a − a ′ = a ( j − k ) / j ,
sup D ≥ 0 ( D + p + 1 ) e − a ( j − k ) D / j ≤ p + 1 + j e a ( j − k ) ≤ 3 Z j . (329) \sup_{D\ge0}(D+p+1)e^{-a(j-k)D/j}\le p+1+\frac{j}{ea(j-k)}\le3Z_j. \tag*{(329)} D ≥ 0 sup ( D + p + 1 ) e − a ( j − k ) D / j ≤ p + 1 + e a ( j − k ) j ≤ 3 Z j . (329) Hence the normalized contribution of all linear terms of order j j j is at most
C e a / M ∑ i + k + q = j − 1 ε k Γ i ( a / M ) q q ! , Ce^{a/M}\sum_{i+k+q=j-1}\varepsilon_k\Gamma_i\frac{(a/M)^q}{q!}, C e a / M i + k + q = j − 1 ∑ ε k Γ i q ! ( a / M ) q , the power count being 1 + i + k + q = j 1+i+k+q=j 1 + i + k + q = j .
(ii) (Separating terms.) In a term of (D.16) of loss r = 2 u + s ≥ 2 r=2u+s\ge2 r = 2 u + s ≥ 2 , the bounded component has complexity r − 1 r-1 r − 1 and dimension
3 u + s − 2 = u + r − 2 ≤ 3 2 r − 2. (330) 3u+s-2=u+r-2\le\frac{3}{2}r-2. \tag*{(330)} 3 u + s − 2 = u + r − 2 ≤ 2 3 r − 2. (330)
so its node degree takes at most 2 r + 1 2r+1 2 r + 1 values, independently of D D D and ℓ \ell ℓ . Its amplitude, the label subsets and the node choice contribute at most
C 1 r ( p − 1 s ) ( 2 r + 1 ) Γ ( r − 1 ) ≤ C 1 r ( p + r ) r ≤ C 1 r Z j r . (331) C_1^r\binom{p-1}{s}(2r+1)\Gamma(r-1)\le C_1^r(p+r)^r\le C_1^rZ_j^r. \tag*{(331)} C 1 r ( s p − 1 ) ( 2 r + 1 ) Γ ( r − 1 ) ≤ C 1 r ( p + r ) r ≤ C 1 r Z j r . (331)
The degree path contributes the geometric sum ∑ m < D e ( σ 0 + a ) m ≤ C e ( σ 0 + a ) D \sum_{m<D}e^{(\sigma_0+a)m}\le Ce^{(\sigma_0+a)D} ∑ m < D e ( σ 0 + a ) m ≤ C e ( σ 0 + a ) D , with no factor D D D , so the large child may be bounded at the parent exponent a a a . A shift of degree q q q whose total denominator exponent is at most r + i + k r+i+k r + i + k , each shift being at most r r r , costs at most
( j r ) q q ! ≤ ( a M ) q Z j q e r q ! . (332) \frac{(jr)^q}{q!}\le\left(\frac{a}{M}\right)^q\frac{Z_j^q e^r}{q!}. \tag*{(332)} q ! ( j r ) q ≤ ( M a ) q q ! Z j q e r . (332)
Every scalar factor of order v v v in such a term has marking count at most p + r p+r p + r and r + v ≤ j r+v\le j r + v ≤ j , hence weight at most Z j v Z_j^v Z j v .
Proof. (i) The term consists of the prefactor A A A , which carries the power t t t , the count ∑ j ′ ≠ i ( 2 d j ′ + 1 ) ≤ 2 m + p \sum_{j'\ne i}(2d_{j'}+1)\le2m+p ∑ j ′ = i ( 2 d j ′ + 1 ) ≤ 2 m + p for a merger or the number # { ( u , v ) : u + v = d i + ℓ − 2 } ≤ m + ℓ + 1 \#\{(u,v):u+v=d_i+\ell-2\}\le m+\ell+1 # {( u , v ) : u + v = d i + ℓ − 2 } ≤ m + ℓ + 1 of nonseparating pairs, the kernel coefficient w ℓ w_\ell w ℓ , the scalar coefficients, the child profile coefficient and the shift coefficient. The child has p ∓ 1 p\mp1 p ∓ 1 markings and degree sum at most m + ℓ m+\ell m + ℓ . In the norm with exponent a ′ = a k / j a'=ak/j a ′ = ak / j its weight is Z k ( p ∓ 1 , a ′ ) = p ∓ 1 + M j / a ≤ Z j + 1 Z_k(p\mp1,a')=p\mp1+Mj/a\le Z_j+1 Z k ( p ∓ 1 , a ′ ) = p ∓ 1 + M j / a ≤ Z j + 1 , and
( Z j + 1 ) k ≤ Z j k ( 1 + 1 Z j ) j ≤ Z j k e k j / Z j ≤ Z j k e a / M ; (Z_j+1)^k\le Z_j^k\left(1+\frac{1}{Z_j}\right)^j\le Z_j^k e^{k j/Z_j}\le Z_j^k e^{a/M}; ( Z j + 1 ) k ≤ Z j k ( 1 + Z j 1 ) j ≤ Z j k e k j / Z j ≤ Z j k e a / M ;
so the child profile coefficient is at most ε k Z j k e a / M e ( σ 0 + a ′ ) ( m + ℓ ) \varepsilon_kZ_j^k e^{a/M}e^{(\sigma_0+a')(m+\ell)} ε k Z j k e a / M e ( σ 0 + a ′ ) ( m + ℓ ) . A scalar coefficient of order i ′ i' i ′ at a marking count p ′ ≤ p + 1 p'\le p+1 p ′ ≤ p + 1 is at most its normalized supremum times Z j i ′ ( p ′ , a ) ≤ ( Z j + 1 ) i ′ Z_j^{i'}(p',a)\le(Z_j+1)^{i'} Z j i ′ ( p ′ , a ) ≤ ( Z j + 1 ) i ′ , and the same inequality bounds the product of the scalar weights by e a / M Z j i e^{a/M}Z_j^i e a / M Z j i ; the constant C C C absorbs this second factor e a / M ≤ e 1 / 2 e^{a/M}\le e^{1/2} e a / M ≤ e 1/2 . Every factor of (D.15) other than G p ( t ) G_p(t) G p ( t ) carries the shift t / ( 1 − t ) t/(1-t) t / ( 1 − t ) , so the total denominator exponent is at most 1 + i + k ≤ j 1+i+k\le j 1 + i + k ≤ j ; by (c) of Subsection D.5 with c 0 = 1 c_0=1 c 0 = 1 , and since [ t q ] ( 1 − t ) − a [t^q](1-t)^{-a} [ t q ] ( 1 − t ) − a is nondecreasing in the exponent, the shift coefficient is at most ( 1 + i + k + q ) q / q ! ≤ j q / q ! ≤ ( a / M ) q Z j q / q ! (1+i+k+q)^q/q!\le j^q/q!\le(a/M)^qZ_j^q/q! ( 1 + i + k + q ) q / q ! ≤ j q / q ! ≤ ( a / M ) q Z j q / q ! , because j ≤ ( a / M ) Z j j\le(a/M)Z_j j ≤ ( a / M ) Z j . Finally, since σ 0 + a ′ ≤ σ 1 < log ρ \sigma_0+a'\le\sigma_1<\log\rho σ 0 + a ′ ≤ σ 1 < log ρ , (D.20) gives ∑ ℓ ≥ 0 ∣ w ℓ ∣ ( m + ℓ + 1 ) e ( σ 0 + a ′ ) ℓ ≤ C 0 ∑ ℓ ≥ 0 ( m + ℓ + 1 ) ρ − ℓ e σ 1 ℓ ≤ C ( m + 1 ) \sum_{\ell\ge0}|w_\ell|(m+\ell+1)e^{(\sigma_0+a')\ell}\le C_0\sum_{\ell\ge0}(m+\ell+1)\rho^{-\ell}e^{\sigma_1\ell}\le C(m+1) ∑ ℓ ≥ 0 ∣ w ℓ ∣ ( m + ℓ + 1 ) e ( σ 0 + a ′ ) ℓ ≤ C 0 ∑ ℓ ≥ 0 ( m + ℓ + 1 ) ρ − ℓ e σ 1 ℓ ≤ C ( m + 1 ) . Collecting these bounds proves (D.39).
A coordinate path from 0 \mathbf{0} 0 to d \mathbf{d} d visits each total degree m = 0 , … , D − 1 m=0,\ldots,D-1 m = 0 , … , D − 1 exactly once, and
∑ m = 0 D − 1 ( 2 m + p + 1 ) e ( σ 0 + a ′ ) m ≤ ( 2 D + p + 1 ) e ( σ 0 + a ′ ) D e σ 0 + a ′ − 1 ≤ 2 ( D + p + 1 ) e σ 0 − 1 e ( σ 0 + a ′ ) D , \sum_{m=0}^{D-1}(2m+p+1)e^{(\sigma_0+a')m}\le(2D+p+1)\frac{e^{(\sigma_0+a')D}}{e^{\sigma_0+a'}-1}\le\frac{2(D+p+1)}{e^{\sigma_0}-1}e^{(\sigma_0+a')D}, m = 0 ∑ D − 1 ( 2 m + p + 1 ) e ( σ 0 + a ′ ) m ≤ ( 2 D + p + 1 ) e σ 0 + a ′ − 1 e ( σ 0 + a ′ ) D ≤ e σ 0 − 1 2 ( D + p + 1 ) e ( σ 0 + a ′ ) D ,
which gives (D.40). Division by Z j j e ( σ 0 + a ) D Z_j^je^{(\sigma_0+a)D} Z j j e ( σ 0 + a ) D leaves Z j − 1 ( D + p + 1 ) e − ( a − a ′ ) D Z_j^{-1}(D+p+1)e^{-(a-a')D} Z j − 1 ( D + p + 1 ) e − ( a − a ′ ) D with a − a ′ = a ( j − k ) / j > 0 a-a'=a(j-k)/j>0 a − a ′ = a ( j − k ) / j > 0 . Since sup D ≥ 0 D e − α D = 1 / ( e α ) \sup_{D\ge0}De^{-\alpha D}=1/(e\alpha) sup D ≥ 0 D e − α D = 1/ ( e α ) for α > 0 \alpha>0 α > 0 , the first inequality of (D.41) follows; for the second, p + 1 ≤ Z j + 1 ≤ 2 Z j p+1\le Z_j+1\le2Z_j p + 1 ≤ Z j + 1 ≤ 2 Z j and j / ( a ( j − k ) ) ≤ j / a ≤ Z j / M ≤ Z j / 2 j/(a(j-k))\le j/a\le Z_j/M\le Z_j/2 j / ( a ( j − k )) ≤ j / a ≤ Z j / M ≤ Z j /2 .
(ii) The bounded component has genus u u u and s + 1 s+1 s + 1 markings, hence complexity 2 u − 2 + s + 1 = r − 1 2u-2+s+1=r-1 2 u − 2 + s + 1 = r − 1 and dimension 3 u − 3 + s + 1 = 3 u + s − 2 3u-3+s+1=3u+s-2 3 u − 3 + s + 1 = 3 u + s − 2 ; since 2 u + s = r 2u+s=r 2 u + s = r and u ≤ r / 2 u\le r/2 u ≤ r /2 , this gives (D.42). Its amplitude vanishes when the sum of its degrees exceeds its dimension, so its node degree is at most 3 2 r − 2 \frac{3}{2}r-2 2 3 r − 2 and takes at most 3 2 r − 1 ≤ 2 r + 1 \frac{3}{2}r-1\le2r+1 2 3 r − 1 ≤ 2 r + 1 values; the other node degree v v v is then fixed by u 0 + v = d i + ℓ − 2 u_0+v=d_i+\ell-2 u 0 + v = d i + ℓ − 2 . By (D.21) the amplitude is at most C 0 r − 1 Γ ( r − 1 ) C_0^{r-1}\Gamma(r-1) C 0 r − 1 Γ ( r − 1 ) , the number of label subsets is ( p − 1 s ) ≤ p s / s ! \binom{p-1}{s}\le p^s/s! ( s p − 1 ) ≤ p s / s ! , and Γ ( r − 1 ) / s ! ≤ r r − 2 − s \Gamma(r-1)/s!\le r^{r-2-s} Γ ( r − 1 ) / s ! ≤ r r − 2 − s for s ≤ r − 2 s\le r-2 s ≤ r − 2 , while s = r s=r s = r gives Γ ( r − 1 ) / r ! ≤ 1 \Gamma(r-1)/r!\le1 Γ ( r − 1 ) / r ! ≤ 1 ; the case s = r − 1 s=r-1 s = r − 1 does not occur since r − s = 2 u r-s=2u r − s = 2 u is even. Hence ( p − 1 s ) ( 2 r + 1 ) Γ ( r − 1 ) ≤ 3 r ( p + r ) r \binom{p-1}{s}(2r+1)\Gamma(r-1)\le3^r(p+r)^r ( s p − 1 ) ( 2 r + 1 ) Γ ( r − 1 ) ≤ 3 r ( p + r ) r , which gives (D.43) with C 1 = 3 C C_1=3C C 1 = 3 C , and p + r ≤ p + 2 j ≤ Z j p+r\le p+2j\le Z_j p + r ≤ p + 2 j ≤ Z j because r ≤ j r\le j r ≤ j and M / a ≥ 2 M/a\ge2 M / a ≥ 2 . None of these factors, nor the kernel sum ∑ ℓ ∣ w ℓ ∣ e ( σ 0 + a ) ℓ ≤ C \sum_\ell|w_\ell|e^{(\sigma_0+a)\ell}\le C ∑ ℓ ∣ w ℓ ∣ e ( σ 0 + a ) ℓ ≤ C , grows with the increments of a separating term sum along the path to at most a constant times ∑ m < D e ( σ 0 + a ) m ≤ e ( σ 0 + a ) D / ( e σ 0 − 1 ) \sum_{m<D}e^{(\sigma_0+a)m}\le e^{(\sigma_0+a)D}/(e^{\sigma_0}-1) ∑ m < D e ( σ 0 + a ) m ≤ e ( σ 0 + a ) D / ( e σ 0 − 1 ) , and after normalization by e ( σ 0 + a ) D e^{(\sigma_0+a)D} e ( σ 0 + a ) D nothing depending on D D D remains; in particular the large child, of degree sum at most m + ℓ m+\ell m + ℓ , is bounded at the parent exponent a a a with weight Z k ( p − s , a ) ≤ Z j Z_k(p-s,a) \le Z_j Z k ( p − s , a ) ≤ Z j . For the shift, (c) of Subsection D.5 with total exponent at most r + i + k r+i+k r + i + k and c 0 ≤ r c_0 \le r c 0 ≤ r gives the coefficient bound ( ( r + i + k + q ) r ) q / q ! ≤ ( j r ) q / q ! ((r+i+k+q)r)^q/q! \le(jr)^q/q! (( r + i + k + q ) r ) q / q ! ≤ ( j r ) q / q ! , because r + i + k + q ≤ j r+i+k+q \le j r + i + k + q ≤ j ; then j q ≤ ( a / M ) q Z j q j^q \le(a/M)^q Z_j^q j q ≤ ( a / M ) q Z j q and r q / q ! ≤ e r r^q/q! \le e^r r q / q ! ≤ e r prove (D.44). Finally a scalar factor of order v v v at marking count at most p + r p+r p + r has weight p + r + M v / a ≤ p + M ( r + v ) / a ≤ p + M j / a = Z j p+r+Mv/a \le p+M(r+v)/a \le p+Mj/a=Z_j p + r + M v / a ≤ p + M ( r + v ) / a ≤ p + M j / a = Z j , since M / a ≥ 2 M/a \ge2 M / a ≥ 2 and r + v ≤ j r+v \le j r + v ≤ j .
Proof of (D.36). Use the weights (D.38) and the normalized suprema defined before Lemma D.10 . By Lemma D.10 (i), the linear terms of order j j j contribute to ε j \varepsilon_j ε j at most
C e a / M ∑ i + k + q = j − 1 ε k Γ i ( a / M ) q q ! ; C e^{a/M}\sum_{i+k+q=j-1}\varepsilon_k\Gamma_i\frac{(a/M)^q}{q!}; C e a / M i + k + q = j − 1 ∑ ε k Γ i q ! ( a / M ) q ;
one factor Z j Z_j Z j is spent on the prefactor t t t together with the path sum, by (D.41), and the remaining Z j i + k + q Z_j^{i+k+q} Z j i + k + q on the coefficients and the shift. The exponent a ′ = a k / j a'=ak/j a ′ = ak / j is used only in the child profile, and a ′ = 0 a'=0 a ′ = 0 when k = 0 k=0 k = 0 .
By Lemma D.10 (ii), a separating term of loss r ≥ 2 r\ge2 r ≥ 2 contributes at most C 1 C_1 C 1 times the normalized product of its r r r cusp coefficients, of its at most r / 2 r/2 r /2 genus coefficients and of one profile coefficient, times ( a / M ) q e r (a/M)^{qe^r} ( a / M ) q e r for its shift; the degree path is a geometric sum without an extra factor D D D , and the power Z j r Z_j^r Z j r against the prefactor ( A t ) r (At)^r ( A t ) r is supplied by (D.43). The KdV split starts at r = 1 r=1 r = 1 , with bounded-component complexity r r r ; the bound ( p − 2 ) Γ ( r ) ≤ C r ( p + r ) r (p^{-2})\Gamma(r)\le C^r(p+r)^r ( p − 2 ) Γ ( r ) ≤ C r ( p + r ) r applies.
Dilaton is a bounded functional because ∑ d ∣ b d ∣ e σ 1 d < ∞ \sum_d |b_d|e^{\sigma1d}<\infty ∑ d ∣ b d ∣ e σ 1 d < ∞ . The current profile in KdV is evaluated at degree one. Descent changes the marking count by at most two and costs at most e 2 a / M e^{2a/M} e 2 a / M in the weights.
Here is an explicit majorant. Let Q ( t ) = ( 1 − b t ) − 1 Q(t)=(1-bt)^{-1} Q ( t ) = ( 1 − b t ) − 1 , with b ≥ 1 b\ge1 b ≥ 1 large enough for all the preceding constants, and give E , C , H , G E,C,H,G E , C , H , G constant terms one. Set
E = 1 + b Q E ( t C G + ( b t C 2 G 2 ) 2 1 − b t C 2 G 2 ) , C = 1 + b ( Q E − 1 ) , H = 1 + b ( E − 1 ) + b Q b t C 2 G 2 1 − b t C 2 G 2 , G = 1 + b ( Q H C 2 − C − 1 ) . (333) \begin{aligned}
E &= 1+bQE\left(tCG+\frac{(btC^2G^2)^2}{1-btC^2G^2}\right),\\
C &= 1+b(QE-1),\\
H &= 1+b(E-1)+bQ\frac{btC^2G^2}{1-btC^2G^2},\\
G &= 1+b\left(Q\frac{HC}{2-C}-1\right). \tag*{(333)}
\end{aligned} E C H G = 1 + b QE ( tC G + 1 − b t C 2 G 2 ( b t C 2 G 2 ) 2 ) , = 1 + b ( QE − 1 ) , = 1 + b ( E − 1 ) + b Q 1 − b t C 2 G 2 b t C 2 G 2 , = 1 + b ( Q 2 − C H C − 1 ) . (333)
They bound respectively profiles, cusp ratios, genus ratios at p ≥ 2 p\ge2 p ≥ 2 , and descent; the last equation also gives G ⪰ H G \succeq H G ⪰ H . A loss-r r r term has r r r cusp factors and at most r / 2 r/2 r /2 genus factors, so the powers in (D.45) dominate their products. Q Q Q bounds the normalized shifts. With the constant terms subtracted, the Jacobian at the origin is
( 0 0 0 0 b 0 0 0 b 0 0 0 0 2 b b 0 ) . \begin{pmatrix}
0 & 0 & 0 & 0\\
b & 0 & 0 & 0\\
b & 0 & 0 & 0\\
0 & 2b & b & 0
\end{pmatrix}. 0 b b 0 0 0 0 2 b 0 0 0 b 0 0 0 0 .
It is strictly lower triangular. Lemma D.2 gives an analytic solution with nonnegative coefficients bounded by C B j CB^j C B j . Induction in order and then in the stage order E , C , H , G E,C,H,G E , C , H , G proves the same bound for the normalized suprema. Taking σ 1 = σ \sigma_1=\sigma σ 1 = σ and one fixed positive a a a , and absorbing 1 / a 1/a 1/ a into M M M , proves (D.36).
Actual remainders in the exponential norm
This subsection proves the remainder bounds (D.37) and thereby completes the proof of Proposition D.9 ; the coefficient bounds (D.36) were proved in Subsection D.9. The argument has the same four stages as the remainder proof for polynomial weights in Subsection D.8, and the proof below is marked accordingly: a region of large weights, where the exact bounds and the coefficient bounds suffice; retained exact products, decomposed by the finite identities (D.32)–(D.33) into terms containing one unknown remainder each, with Lemma D.10 supplying the weight count; forcing terms from the omitted topological types, bounded by Lemma D.8; and a convergent linear majorant system for the normalized errors, solved by induction on the order. The difference from Subsection D.8 is that the marking count and the order may now be linear in h h h , so no smallness of L 2 / h L^2/h L 2 / h is available, and only the ratios r / h r/h r / h and L / h L/h L / h are used.
Proof of (D.37). Choose 0 < ς 0 < ς 1 < ς 2 < log ρ 0 < \varsigma_0 < \varsigma_1 < \varsigma_2 < \log\rho 0 < ς 0 < ς 1 < ς 2 < log ρ . Use the preceding formal bounds with base exponent ς 0 \varsigma_0 ς 0 ; for actual remainders use ς 1 + a \varsigma_1 + a ς 1 + a , with 0 < a ≤ ς 2 − ς 1 0 < a \le\varsigma_2 - \varsigma_1 0 < a ≤ ς 2 − ς 1 . Put Z L = p + M L / a Z_L = p + ML/a Z L = p + M L / a , with M ≥ max { 2 , 2 ( ς 2 − ς 1 ) } M \ge\max\{2, 2(\varsigma_2 - \varsigma_1)\} M ≥ max { 2 , 2 ( ς 2 − ς 1 )} as in (D.38), and δ = Z L / h \delta= Z_L/h δ = Z L / h . We prove geometric bounds for the remainder divided by δ L \delta^L δ L , with the additional factor e ( ς 1 + a ) D e^{(\varsigma_1+a)D} e ( ς 1 + a ) D for a profile. Precisely, for L ≥ 1 L \ge1 L ≥ 1 put
R E ( L ) = sup ∣ c g , p ( d ) − ∑ j < L e p , j ( d ) h − j ∣ δ L e ( ς 1 + a ) D , R C ( L ) = sup ∣ C p ( h ) − ∑ j < L C p , j h − j ∣ δ L , R_E(L)=\sup\frac{\left|c_{g,p}(\mathbf{d})-\sum_{j<L}e_{p,j}(\mathbf{d})h^{-j}\right|}{\delta^L e^{(\varsigma_1+a)D}}, \qquad R_C(L)=\sup\frac{\left|C_p(h)-\sum_{j<L}C_{p,j}h^{-j}\right|}{\delta^L}, R E ( L ) = sup δ L e ( ς 1 + a ) D c g , p ( d ) − ∑ j < L e p , j ( d ) h − j , R C ( L ) = sup δ L C p ( h ) − ∑ j < L C p , j h − j ,
and R H ( L ) R_H(L) R H ( L ) , R G ( L ) R_G(L) R G ( L ) in the same way for G p G_p G p with p ≥ 2 p \ge2 p ≥ 2 and with p = 1 p=1 p = 1 . The suprema are over real Y ∈ [ 0 , y 0 ] Y \in[0,y_0] Y ∈ [ 0 , y 0 ] , all stable states with p ≥ 1 p \ge1 p ≥ 1 , all degree tuples (invalid tuples having profile zero) and all a ∈ ( 0 , ς 2 − ς 1 ] a \in(0,\varsigma_2-\varsigma_1] a ∈ ( 0 , ς 2 − ς 1 ] ; the supremum over a a a is what allows a child remainder to be estimated below at the smaller exponent a ′ = a k / L a' = ak/L a ′ = ak / L . The constants are independent of the allowed a a a , p p p , L L L , h h h .
Large weights and degrees. The exact profiles and scalar ratios are bounded by (70) –(72) and (82) . The formal coefficient bounds are at most C B 0 j Z L j CB_0^jZ_L^j C B 0 j Z L j for j < L j<L j < L , with the profile exponential included: if M 0 M_0 M 0 is the constant of (D.36) at the exponent ς 0 \varsigma_0 ς 0 , then ( p + M 0 j ) j ≤ C j ( p + M L / a ) j (p+M_0j)^j \le C^j(p+ML/a)^j ( p + M 0 j ) j ≤ C j ( p + M L / a ) j for j < L j<L j < L with C = max { 1 , M 0 ( ς 2 − ς 1 ) / M } C=\max\{1,M_0(\varsigma_2-\varsigma_1)/M\} C = max { 1 , M 0 ( ς 2 − ς 1 ) / M } , because a ≤ ς 2 − ς 1 a\le\varsigma_2-\varsigma_1 a ≤ ς 2 − ς 1 gives M L / a ≥ M L / ( ς 2 − ς 1 ) ML/a\ge ML/(\varsigma_2-\varsigma_1) M L / a ≥ M L / ( ς 2 − ς 1 ) ; this comparison uses that a a a is bounded above, and would fail for unbounded a a a . Choose δ 0 \delta_0 δ 0 so that B 0 δ 0 < 1 / 2 B_0\delta_0<1/2 B 0 δ 0 < 1/2 . When δ ≥ δ 0 \delta\ge\delta_0 δ ≥ δ 0 , the normalized error is at most C δ 0 − L C\delta_0^{-L} C δ 0 − L by the large-weight calculation in the proof of Proposition D.3.
Suppose δ < δ 0 \delta<\delta_0 δ < δ 0 . Then p ≤ δ 0 h p\le\delta_0h p ≤ δ 0 h and L / h < δ 0 ( ς 2 − ς 1 ) / M L/h<\delta_0(\varsigma_2-\varsigma_1)/M L / h < δ 0 ( ς 2 − ς 1 ) / M . Shrink δ 0 \delta_0 δ 0 to fit Lemma D.8 and to bound the cusp ratios inverted below away from zero. Every state occurring below has marking count at most p + L + 2 p+L+2 p + L + 2 and complexity at least h − L h-L h − L , so for δ 0 \delta_0 δ 0 small the lemma applies at all of them. To treat D ≥ h / 4 D\ge h/4 D ≥ h /4 , note that B 0 δ < 1 / 2 B_0\delta<1/2 B 0 δ < 1/2 and the formal bounds at the exponent ς 0 \varsigma_0 ς 0 give ∑ j < L ∣ e p , j ( d ) ∣ h − j ≤ 2 C e ς 0 D \sum_{j<L}|e_{p,j}(\mathbf{d})|h^{-j}\le2Ce^{\varsigma_0D} ∑ j < L ∣ e p , j ( d ) ∣ h − j ≤ 2 C e ς 0 D , while 0 ≤ c g , p ≤ 1 0\le c_{g,p}\le1 0 ≤ c g , p ≤ 1 . After division by e ( ς 1 + a ) D e^{(\varsigma_1+a)D} e ( ς 1 + a ) D the error is therefore at most ( 1 + 2 C ) e − b 0 h (1+2C)e^{-b_0h} ( 1 + 2 C ) e − b 0 h with b 0 = ( ς 1 − ς 0 ) / 4 b_0=(\varsigma_1-\varsigma_0)/4 b 0 = ( ς 1 − ς 0 ) /4 . Since
e − b 0 h h L ≤ ( L b 0 e ) L , Z L ≥ M L ς 2 − ς 1 , e^{-b_0h}h^L\le\left(\frac{L}{b_0e}\right)^L,\qquad Z_L\ge\frac{ML}{\varsigma_2-\varsigma_1}, e − b 0 h h L ≤ ( b 0 e L ) L , Z L ≥ ς 2 − ς 1 M L ,
this contributes only a geometric forcing sequence. Invalid parents are included: their dimension is ( 3 h − p ) / 2 > h / 4 (3h-p)/2>h/4 ( 3 h − p ) /2 > h /4 in this region. We may use the exact adjacent recursion at D < h / 4 D<h/4 D < h /4 .
Retained products and shifts. Retain losses r < L r<L r < L . The factorization (D.14) is an exact finite product of shifted cusp and genus ratios. Apply (D.32) with the large profile first. Each term contains one unknown remainder, either the profile remainder or a scalar remainder preceded by known profile and scalar coefficients. All exact factors after it are scalar and bounded. There are O ( r ) O(r) O ( r ) such factors, costing at most C r C^r C r . For descent use (D.33); its exact inverse is bounded in the small-weight region.
The exact kernel sums can be extended by zero to all ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 . Indeed, with parent dimension N = 3 g − 3 + p N=3g-3+p N = 3 g − 3 + p and deficit m = N − D m=N-D m = N − D , a merger has child dimension N − 1 N-1 N − 1 and degree D + ℓ − 1 D+\ell-1 D + ℓ − 1 ; a nonseparating child has dimension N − 2 N-2 N − 2 and degree D + ℓ − 2 D+\ell-2 D + ℓ − 2 . Both deficits are m − ℓ m-\ell m − ℓ . In a split the sum of child deficits is also m − ℓ m-\ell m − ℓ , so a valid bounded child and ℓ > m \ell>m ℓ > m force the large child to be invalid. The expanded sums converge absolutely by ς 2 < log ρ \varsigma_2<\log\rho ς 2 < log ρ .
In each resulting term, the loss r r r , total known coefficient order i i i , unique remainder order q ≥ 1 q\ge1 q ≥ 1 , and shift order v v v satisfy r + i + q + v = L r+i+q+v=L r + i + q + v = L . Thus the weights are allocated to one total order L L L . Lemma D.10 applies verbatim with j j j replaced by L L L and ( ς 0 , ς 1 ) (\varsigma_0,\varsigma_1) ( ς 0 , ς 1 ) by ( ς 1 , ς 2 ) (\varsigma_1,\varsigma_2) ( ς 1 , ς 2 ) , the normalized supremum of the one unknown factor being its error sequence and those of the known factors being the normalized formal coefficients. For a linear branch let k ≤ L − 1 k \le L - 1 k ≤ L − 1 be the order of the profile factor, whether it is the remainder or a known coefficient. If it is the remainder and k > 0 k > 0 k > 0 , bound it in the norm with exponent ς 1 + a ′ \varsigma_1 + a' ς 1 + a ′ , a ′ = a k / L a' = ak/L a ′ = ak / L ; then (D.41) gives the degree-path cost
sup D ≥ 0 ( D + p + 1 ) e − a ( L − k ) D / L ≤ p + 1 + L e a ( L − k ) ≤ 3 Z L . \sup_{D \ge0}(D+p+1)e^{-a(L-k)D/L} \le p+1+\frac{L}{e\,a(L-k)} \le3Z_L. D ≥ 0 sup ( D + p + 1 ) e − a ( L − k ) D / L ≤ p + 1 + e a ( L − k ) L ≤ 3 Z L .
If it is a known coefficient, its formal bound carries the exponential e ς 0 D ′ e^{\varsigma_0D'} e ς 0 D ′ at the child degree sum D ′ ≤ D + ℓ D' \le D+\ell D ′ ≤ D + ℓ , and ς 0 < ς 1 ≤ ς 1 + a ′ \varsigma_0 < \varsigma_1 \le\varsigma_1+a' ς 0 < ς 1 ≤ ς 1 + a ′ ; hence
e ς 0 D ′ ≤ e ς 0 ℓ e ( ς 1 + a ′ ) D , e^{\varsigma_0D'} \le e^{\varsigma_0\ell}e^{(\varsigma_1+a')D}, e ς 0 D ′ ≤ e ς 0 ℓ e ( ς 1 + a ′ ) D ,
so the known coefficient enters the path sum exactly as a remainder of the same order would, with the margin ς 1 − ς 0 \varsigma_1-\varsigma_0 ς 1 − ς 0 to spare. This is how the formal-coefficient bound at the exponent ς 0 \varsigma_0 ς 0 is absorbed into the norm e ( ς 1 + a ) D e^{(\varsigma_1+a)D} e ( ς 1 + a ) D of the actual remainders. At k = 0 k=0 k = 0 the profile is one. All other weights have the same product bounds as in the formal proof. For a split, (D.42)–(D.43) remove the extra degree count, giving C r Z L r C^rZ_L^r C r Z L r against its prefactor of order r r r .
A remainder of order q ≤ L q \le L q ≤ L at complexity h − c h-c h − c , ∣ c ∣ ≤ r |c| \le r ∣ c ∣ ≤ r , costs the additional factor
( h h − c ) q ≤ e 2 r q / h ≤ e 2 r L / h ≤ C r , \left(\frac{h}{h-c}\right)^q \le e^{2rq/h} \le e^{2rL/h} \le C^r, ( h − c h ) q ≤ e 2 r q / h ≤ e 2 r L / h ≤ C r ,
because ∣ c ∣ / h ≤ r / h ≤ 1 / 2 |c|/h \le r/h \le1/2 ∣ c ∣/ h ≤ r / h ≤ 1/2 in the small-weight region, where − log ( 1 − x ) ≤ 2 x -\log(1-x) \le2x − log ( 1 − x ) ≤ 2 x for 0 ≤ x ≤ 1 / 2 0 \le x \le1/2 0 ≤ x ≤ 1/2 , and L / h ≤ δ 0 ( ς 2 − ς 1 ) / M L/h \le\delta_0(\varsigma_2-\varsigma_1)/M L / h ≤ δ 0 ( ς 2 − ς 1 ) / M is bounded. Only r / h r/h r / h and L / h L/h L / h enter here; no smallness of L 2 / h L^2/h L 2 / h is needed, in contrast to the proof of Proposition D.3 . Unit shifts in dilaton and descent cost a fixed constant. For a known shift tail whose first omitted degree is q q q , the finite negative-binomial bound used in Proposition D.3 gives
( L r ) q q ! h − q ( 1 − r / h ) − L . \frac{(Lr)^q}{q!}h^{-q}(1-r/h)^{-L}. q ! ( L r ) q h − q ( 1 − r / h ) − L .
After division by Z L q h − q Z_L^q h^{-q} Z L q h − q , the inequalities L / Z L ≤ a / M L/Z_L \le a/M L / Z L ≤ a / M and r q / q ! ≤ e r r^q/q! \le e^r r q / q ! ≤ e r leave a geometric q q q bound times C r C^r C r ; again no condition on L 2 / h L^2/h L 2 / h is needed.
Tail forcing and the base case. For L ≥ 2 L \ge2 L ≥ 2 , Lemma D.8 bounds the omitted exact topology by C [ B ( p + L ) / h ] L C[B(p+L)/h]^L C [ B ( p + L ) / h ] L . In a profile it is multiplied by at most a kernel-weighted quadratic polynomial in D D D . The norm e − ( ς 1 + a ) D e^{-(\varsigma_1+a)D} e − ( ς 1 + a ) D absorbs that polynomial uniformly. This gives C B L δ L CBL\delta^L C B L δ L . At L = 1 L=1 L = 1 the first-order estimates in the same lemma give the bound directly. The large-weight and large-degree regions have already supplied geometric forcing.
A convergent error majorant. Regard the four normalized error sequences R E , R C , R H , R G R_E,R_C,R_H,R_G R E , R C , R H , R G defined above as series in z z z , with zero constant terms and positive orders starting at L = 1 L=1 L = 1 . Known formal coefficients, inverse-cusp coefficients and exact scalar factors are bounded by one series V ( z ) = v 0 / ( 1 − v 1 z ) V(z)=v_0/(1-v_1z) V ( z ) = v 0 / ( 1 − v 1 z ) , with v 0 , v 1 ≥ 1 v_0,v_1 \ge1 v 0 , v 1 ≥ 1 (the symbol v 1 v_1 v 1 is unrelated to the instanton coefficient v 1 = − 1 / 8 v_1=-1/8 v 1 = − 1/8 of Subsection D.11). Choose a fixed b ≥ 1 b \ge1 b ≥ 1 large enough for the retained-product constants. Put x ( z ) = b z V ( z ) 8 x(z)=bzV(z)^8 x ( z ) = b z V ( z ) 8 , and set
k ( z ) = b V ( z ) 8 1 − b z ( z + 1 + x ( z ) ( 1 − x ( z ) ) 3 − 1 ) . k(z)=\frac{bV(z)^8}{1-bz}\left(z+\frac{1+x(z)}{(1-x(z))^3}-1\right). k ( z ) = 1 − b z bV ( z ) 8 ( z + ( 1 − x ( z ) ) 3 1 + x ( z ) − 1 ) .
This analytic series has nonnegative coefficients and k ( 0 ) = 0 k(0)=0 k ( 0 ) = 0 . Its second term is the sum, over r ≥ 1 r \ge1 r ≥ 1 , of b ( r + 1 ) 2 ( b z ) r V ( z ) 8 r + 8 / ( 1 − b z ) b(r+1)^2(bz)^rV(z)^{8r+8}/(1-bz) b ( r + 1 ) 2 ( b z ) r V ( z ) 8 r + 8 / ( 1 − b z ) . A separating loss-r r r product, with bounded-component genus u u u , has at most r + u + 1 ≤ 3 r / 2 + 1 r+u+1 \le3r/2+1 r + u + 1 ≤ 3 r /2 + 1 expendable factors, hence O ( r ) O(r) O ( r ) possible remainder positions. Together with the O ( r ) O(r) O ( r ) node choices, these give the displayed quadratic factor; known-order allocations are already counted by the powers of V V V . Exponential costs in r r r are absorbed in b r b^r b r . The first term covers linear branches and positive-order unit shifts. Terms whose only error is a truncated known shifted power contain no unknown remainder. Their same analytic loss sum is included in the forcing.
Let J 4 J_4 J 4 be the four-by-four matrix with all entries one, and define
N 0 = ( 0 0 0 0 b 0 0 0 b 0 0 0 0 b b 0 ) , F ( z ) = f 0 z 1 − f 1 z ( 1 , 1 , 1 , 1 ) T . N_0=\begin{pmatrix}0&0&0&0\\ b&0&0&0\\ b&0&0&0\\ 0&b&b&0\end{pmatrix},\qquad F(z)=\frac{f_0z}{1-f_1z}(1,1,1,1)^{\mathsf{T}}. N 0 = 0 b b 0 0 0 0 b 0 0 0 b 0 0 0 0 , F ( z ) = 1 − f 1 z f 0 z ( 1 , 1 , 1 , 1 ) T .
Choose f 0 f_0 f 0 , f 1 f_1 f 1 after b b b , υ 0 \upsilon_0 υ 0 , υ 1 \upsilon_1 υ 1 to bound all forcing sequences. The only current-order dependencies are profile to cusp, profile to genus at p ≥ 2 p\ge2 p ≥ 2 , and cusp/genus to descent. They are covered by N 0 N_0 N 0 ; all remaining errors have lower order. Thus, componentwise,
R ⪯ F + N 0 R + k J 4 R . R\preceq F+N_0R+kJ_4R. R ⪯ F + N 0 R + k J 4 R .
Since N 0 4 = 0 N_0^4=0 N 0 4 = 0 , ( I − N 0 ) − 1 (I-N_0)^{-1} ( I − N 0 ) − 1 has nonnegative entries. The series
U ( z ) = ( I − N 0 − k ( z ) J 4 ) − 1 F ( z ) U(z)=(I-N_0-k(z)J_4)^{-1}F(z) U ( z ) = ( I − N 0 − k ( z ) J 4 ) − 1 F ( z )
is analytic near zero, with nonnegative coefficients bounded by C B L CB^L C B L . Induction in L L L , then in stage order, proves R ⪯ U R\preceq U R ⪯ U and proves finiteness of the suprema at the same time. Finally take ς 1 + a = σ \varsigma_1+a=\sigma ς 1 + a = σ and absorb the fixed 1 / a 1/a 1/ a into M M M . This proves (D.37). □
For fixed p p p , ( p + M L ) L ≤ C p L L ! (p+ML)^L\le C_p^L L! ( p + M L ) L ≤ C p L L ! . Thus the scalar ratios have Gevrey-1 remainders. The same holds for ∑ d λ d c g , p ( d ) \sum_d\lambda_d c_{g,p}(\mathbf{d}) ∑ d λ d c g , p ( d ) whenever ∑ d ∣ λ d ∣ e σ ∣ d ∣ < ∞ \sum_d|\lambda_d|e^{\sigma|\mathbf{d}|}<\infty ∑ d ∣ λ d ∣ e σ ∣ d ∣ < ∞ for the exponent σ \sigma σ of (D.37); the dilaton functional f ↦ ∑ d ≥ 1 b d f ( d , 0 p ) f\mapsto\sum_{d\ge1}b_df(d,0^p) f ↦ ∑ d ≥ 1 b d f ( d , 0 p ) is of this kind.
Determining the constants at the pure-ψ \psi ψ endpoint
To integrate the transport equation to all orders, we need the value of every coefficient at the pure-ψ \psi ψ endpoint, with constants controlled as the order grows. This subsection supplies these values and then proves Theorem 8.2. It proceeds in four stages. First, the exact endpoint identity following (D.12) shows that the higher cusp-ratio coefficients vanish at y 0 y_0 y 0 , see (D.47). Second, the large-order asymptotics of the Painlevé I zero-instanton coefficients are made quantitative: the Borel representation of Costin, Costin and Huang [48] gives a Gevrey-1 remainder for the jump between adjacent tritronquée solutions, whose first coefficient v 1 = − 1 / 8 v_1=-1/8 v 1 = − 1/8 is derived directly, and the contour formula of Garoufalidis, Its, Kapaev and Mariño [70] transfers this remainder to the coefficients themselves. Third, the Itzykson–Zuber identity [106] converts these into the expansion (D.48) of π C g ( 2 3 g − 3 ) \pi C_g(2^{3g-3}) π C g ( 2 3 g − 3 ) and, through string and dilaton, into the exact normalized endpoint values (D.49); Lemma D.11 records the resulting Gevrey-1 bounds. Finally, the coefficients a p , j a_{p,j} a p , j and L p , j L_{p,j} L p , j of Theorem 8.2 are defined in (D.50), and the theorem is proved by integrating the logarithmic derivative from a matching point near y 0 y_0 y 0 , using Proposition D.9, the inverse identity (D.33) and the positive comparison (7.16)–(7.17).
Invert the cusp expansion and write, at each fixed order,
1 C p ( h ) = 1 + β p h + ∑ j = 1 K γ p , j h j + 1 + O p , K ( h − K − 2 ) . (334) \frac{1}{C_p(h)}=1+\frac{\beta_p}{h}+\sum_{j=1}^{K}\frac{\gamma_{p,j}}{h^{j+1}}+O_{p,K}(h^{-K-2}). \tag*{(334)} C p ( h ) 1 = 1 + h β p + j = 1 ∑ K h j + 1 γ p , j + O p , K ( h − K − 2 ) . (334)
Here γ p , 1 = γ p \gamma_{p,1}=\gamma_p γ p , 1 = γ p . Three facts determine the endpoint values of these coefficients. First, the remainder estimate (D.37) holds uniformly for real Y ∈ [ 0 , y 0 ] Y\in[0,y_0] Y ∈ [ 0 , y 0 ] , in particular at Y = y 0 Y=y_0 Y = y 0 , and the finite inverse identity (D.33) transfers it to 1 / C p ( h ; y 0 ) 1/C_p(h;y_0) 1/ C p ( h ; y 0 ) ; thus (D.46) is a valid asymptotic expansion at Y = y 0 Y=y_0 Y = y 0 . Second, the formal coefficients γ p , j \gamma_{p,j} γ p , j are holomorphic on the neighborhood Ω \Omega Ω of [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] , hence continuous at y 0 y_0 y 0 , so the coefficients of that expansion are the values β p ( y 0 ) \beta_p(y_0) β p ( y 0 ) and γ p , j ( y 0 ) \gamma_{p,j}(y_0) γ p , j ( y 0 ) . Third, the exact endpoint identity following (D.12) gives 1 / C p ( h ; y 0 ) = 1 − p / ( 5 h ) 1/C_p(h;y_0)=1-p/(5h) 1/ C p ( h ; y 0 ) = 1 − p / ( 5 h ) , a finite expansion whose only term is β p ( y 0 ) / h = − p / ( 5 h ) \beta_p(y_0)/h=-p/(5h) β p ( y 0 ) / h = − p / ( 5 h ) . Uniqueness of asymptotic expansions in h − 1 h^{-1} h − 1 therefore gives
γ p , j ( y 0 ) = 0 ( j ≥ 1 ) , γ p , j / ( A r ) is analytic at y 0 . (335) \gamma_{p,j}(y_0)=0\quad(j\ge1),\qquad\gamma_{p,j}/(Ar)\text{ is analytic at }y_0. \tag*{(335)} γ p , j ( y 0 ) = 0 ( j ≥ 1 ) , γ p , j / ( A r ) is analytic at y 0 . (335)
Analyticity follows from the finite coefficient rules, not from remainder estimates. Division by the simple zero of A r Ar A r is a bounded operation on a fixed smaller complex neighborhood: write a vanishing function as ( Y − y 0 ) ∫ 0 1 f ′ ( y 0 + t ( Y − y 0 ) ) d t (Y-y_0)\int_0^1 f'(y_0+t(Y-y_0))\,dt ( Y − y 0 ) ∫ 0 1 f ′ ( y 0 + t ( Y − y 0 )) d t and apply Cauchy’s bound once on the larger neighborhood. No successive shrinkage in order occurs.
We need a quantitative version of the pure-ψ \psi ψ endpoint input. Use the conventions of Garoufalidis–Its–Kapaev–Mariño [70]: − u ′ ′ / 6 + u 2 = Z -u''/6+u^2=Z − u ′′ /6 + u 2 = Z , zero-instanton coefficients u g , 0 u_{g,0} u g , 0 , a = 8 3 / 5 a=8\sqrt{3}/5 a = 8 3 /5 , ν g = 2 g − 1 / 2 \nu_g=2g-1/2 ν g = 2 g − 1/2 and S 1 = − i 3 1 / 4 / ( 2 π ) S_1=-i3^{1/4}/(2\sqrt{\pi}) S 1 = − i 3 1/4 / ( 2 π ) . Put
Z g = π i a ν g u g , 0 S 1 Γ ( ν g ) . Z_g=\frac{\pi i\,a^{\nu_g}u_{g,0}}{S_1\Gamma(\nu_g)}. Z g = S 1 Γ ( ν g ) π i a ν g u g , 0 .
The Borel representation of Costin–Costin–Huang [48, Theorem 2] makes the coefficient formula in [70, Section 2.3] quantitative, as follows. That formula gives the expansion of u g , 0 u_{g,0} u g , 0 to each fixed order; what is needed here, and proved below, is a remainder whose factorial growth in the order is uniform. Set ϱ = 6 − 1 / 5 e i π / 5 \varrho=6^{-1/5}e^{i\pi/5} ϱ = 6 − 1/5 e iπ /5 ; y ( Z c ) = ϱ 2 u ( ϱ Z c ) y(Z_c)=\varrho^2u(\varrho Z_c) y ( Z c ) = ϱ 2 u ( ϱ Z c ) satisfies y ′ ′ = 6 y 2 + Z c y''=6y^2+Z_c y ′′ = 6 y 2 + Z c . Their Boutroux coordinate is ξ = a Z 5 / 4 \xi=aZ^{5/4} ξ = a Z 5/4 , and the normalized remainder is u ( Z ) / Z − 1 + 4 / ( 25 ξ 2 ) u(Z)/\sqrt{Z}-1+4/(25\xi^2) u ( Z ) / Z − 1 + 4/ ( 25 ξ 2 ) . The CCH variable x x x is our ξ \xi ξ , and their unknown h ( x ) h(x) h ( x ) is this normalized remainder. The adjacent GIKM solutions u 1 , u 0 u_1,u_0 u 1 , u 0 have sectors ( − 2 π / 5 , 6 π / 5 ) (-2\pi/5,6\pi/5) ( − 2 π /5 , 6 π /5 ) and ( − 6 π / 5 , 2 π / 5 ) (-6\pi/5,2\pi/5) ( − 6 π /5 , 2 π /5 ) respectively. Their normalized remainders are O ( ξ − 4 ) O(\xi^{-4}) O ( ξ − 4 ) on + i ∞ +i\infty + i ∞ for u 1 u_1 u 1 and on − i ∞ -i\infty − i ∞ for u 0 u_0 u 0 , each ray lying inside the corresponding sector. Thus the respective lateral constants in the exact Borel representation vanish: multiplying by e ξ ξ 1 / 2 e^{\xi}\xi^{1/2} e ξ ξ 1/2 gives zero on those rays. This is the characterization of the tritronquée solutions as the tronquée solutions whose constant beyond all orders vanishes, [48, Theorem 7 and Note 8(ii)]. It identifies the lateral Borel sums with the solutions themselves, not merely with their formal expansions.
The Borel transform at its first positive singularity has the form S ( 1 − s ) − 1 / 2 α ( s ) + β ( s ) S(1-s)^{-1/2}\alpha(s)+\beta(s) S ( 1 − s ) − 1/2 α ( s ) + β ( s ) , where α , β \alpha,\beta α , β are analytic near one and α ( 1 ) = 1 \alpha(1)=1 α ( 1 ) = 1 [48, Proposition 3(i)]. We keep the constant S S S symbolic: only the structure of this singularity is used. The jump between the two solutions is normalized by the Stokes multiplier S 1 S_1 S 1 of Kapaev [116], in the form recorded in [70, Section 2.1]; with this value of S 1 S_1 S 1 ,
J ( ξ ) = u 1 ( Z ) − u 0 ( Z ) − S 1 e − ξ ξ − 1 / 8 , J ( ξ ) ⟶ 1. \mathfrak{J}(\xi)=\frac{u_1(Z)-u_0(Z)}{-S_1e^{-\xi}\xi^{-1/8}},\qquad\mathfrak{J}(\xi)\longrightarrow1. J ( ξ ) = − S 1 e − ξ ξ − 1/8 u 1 ( Z ) − u 0 ( Z ) , J ( ξ ) ⟶ 1.
Choose a fixed small r > 0 r>0 r > 0 and angles ± ϕ \pm\phi ± ϕ so ( 1 + r ) cos ϕ > 1 + r / 2 (1+r)\cos\phi>1+r/2 ( 1 + r ) cos ϕ > 1 + r /2 . Deform only the finite initial parts of the two Laplace rays into a local Hankel contour about [ 1 , 1 + r ] [1,1+r] [ 1 , 1 + r ] and compact joining arcs on ∣ s ∣ = 1 + r |s|=1+r ∣ s ∣ = 1 + r . The infinite ray tails are kept unchanged. The analytic continuation and weighted L 1 L^1 L 1 estimates of [48, Lemma 17(c),(f)] bound these outer parts by C e − ( 1 + ϵ ) ξ Ce^{-(1+\epsilon)\xi} C e − ( 1 + ϵ ) ξ , with constants independent of Taylor order. The local analytic term has no jump. Cauchy’s estimate on ∣ s − 1 ∣ ≤ 2 r |s-1|\leq2r ∣ s − 1∣ ≤ 2 r bounds the degree-L − 1 L-1 L − 1 remainder of α ( 1 + t ) \alpha(1+t) α ( 1 + t ) by C ( 2 r ) − L t L C(2r)^{-L}t^L C ( 2 r ) − L t L . Its normalized Laplace integral is bounded by C B L Γ ( L + 1 / 2 ) ξ − L CB^L\Gamma(L+1/2)\xi^{-L} C B L Γ ( L + 1/2 ) ξ − L . Extending the retained polynomial integrals from r r r to infinity obeys the same bound; the outer term O ( ξ e − ξ ) O(\sqrt{\xi}e^{-\xi}) O ( ξ e − ξ ) is at most C B L L ! ξ − L CB^LL!\xi^{-L} C B L L ! ξ − L . Consequently, for all L ≥ 1 L\geq1 L ≥ 1 and real ξ ≥ ξ 0 \xi\geq\xi_0 ξ ≥ ξ 0 ,
∣ J ( ξ ) − ∑ j < L v j ξ − j ∣ ≤ C B L L ! ξ − L , v j = a j u j , 1 . \left|\mathfrak{J}(\xi)-\sum_{j<L}v_j\xi^{-j}\right|\leq C B^L L!\xi^{-L},\qquad v_j=a^j u_{j,1}. J ( ξ ) − j < L ∑ v j ξ − j ≤ C B L L ! ξ − L , v j = a j u j , 1 .
The value of the Borel constant S S S of [48] is not used; the only numerical Stokes datum is Kapaev’s S 1 S_1 S 1 .
The first coefficient is v 1 = − 1 / 8 v_1=-1/8 v 1 = − 1/8 , and we derive it here because it determines the first correction at the endpoint. The difference u 1 − u 0 u_1-u_0 u 1 − u 0 is exponentially small and satisfies the linearization ( u 1 − u 0 ) ′ ′ = 12 u 0 ( u 1 − u 0 ) (u_1-u_0)''=12u_0(u_1-u_0) ( u 1 − u 0 ) ′′ = 12 u 0 ( u 1 − u 0 ) of − u ′ ′ / 6 + u 2 = Z -u''/6+u^2=Z − u ′′ /6 + u 2 = Z up to a term quadratic in u 1 − u 0 u_1-u_0 u 1 − u 0 , which is O ( e − 2 ξ ) O(e^{-2\xi}) O ( e − 2 ξ ) and does not affect the coefficients v j v_j v j . Here 12 u 0 = 12 Z 1 / 2 − Z − 2 / 4 + O ( Z − 9 / 2 ) 12u_0=12Z^{1/2}-Z^{-2/4}+O(Z^{-9/2}) 12 u 0 = 12 Z 1/2 − Z − 2/4 + O ( Z − 9/2 ) , because u 1 , 0 = − 1 / 48 u_{1,0}=-1/48 u 1 , 0 = − 1/48 . Put f = Z − 1 / 8 e − ξ f=Z^{-1/8}e^{-\xi} f = Z − 1/8 e − ξ . Since ξ ′ = 2 3 Z 1 / 4 \xi'=2\sqrt{3}Z^{1/4} ξ ′ = 2 3 Z 1/4 and ξ ′ ′ = ( 3 / 2 ) Z − 3 / 4 \xi''=(\sqrt{3}/2)Z^{-3/4} ξ ′′ = ( 3 /2 ) Z − 3/4 ,
f ′ ′ f = ( 1 8 Z + ξ ′ ) 2 + 1 8 Z 2 − ξ ′ ′ = 12 Z 1 / 2 + 9 64 Z − 2 . \frac{f''}{f}=\left(\frac{1}{8Z}+\xi'\right)^2+\frac{1}{8Z^2}-\xi''=12Z^{1/2}+\frac{9}{64}Z^{-2}. f f ′′ = ( 8 Z 1 + ξ ′ ) 2 + 8 Z 2 1 − ξ ′′ = 12 Z 1/2 + 64 9 Z − 2 .
the terms of order Z − 3 / 4 Z^{-3/4} Z − 3/4 cancel, which is why the prefactor is Z − 1 / 8 Z^{-1/8} Z − 1/8 . Writing u 1 − u 0 = − S 1 f ȷ ~ u_1-u_0=-S_1f\tilde{\jmath} u 1 − u 0 = − S 1 f ~ and dividing the linearized equation by − S 1 f -S_1f − S 1 f gives ( f ′ ′ / f ) ȷ ~ + 2 ( f ′ / f ) ȷ ~ ′ + ȷ ~ ′ ′ = 12 u 0 ȷ ~ (f^{\prime\prime}/f)\tilde{\jmath}+2(f^\prime/f)\tilde{\jmath}^\prime+\tilde{\jmath}^{\prime\prime}=12u_0\tilde{\jmath} ( f ′′ / f ) ~ + 2 ( f ′ / f ) ~ ′ + ~ ′′ = 12 u 0 ~ . The local Laplace representation above gives the expansion of ȷ ~ \tilde{\jmath} ~ in a sector about the positive ray; Cauchy’s estimate therefore permits differentiation of its remainder on a smaller sector. With ȷ ~ = 1 + v 1 ξ − 1 + O ( ξ − 2 ) \tilde{\jmath}=1+v_1\xi^{-1}+O(\xi^{-2}) ~ = 1 + v 1 ξ − 1 + O ( ξ − 2 ) one has ȷ ~ ′ = − v 1 ξ ′ ξ − 2 + O ( Z − 7 / 2 ) = − ( 2 3 v 1 / a 2 ) Z − 9 / 4 + O ( Z − 7 / 2 ) \tilde{\jmath}^\prime=-v_1\xi^\prime\xi^{-2}+O(Z^{-7/2})=-(2\sqrt{3}v_1/a^2)Z^{-9/4}+O(Z^{-7/2}) ~ ′ = − v 1 ξ ′ ξ − 2 + O ( Z − 7/2 ) = − ( 2 3 v 1 / a 2 ) Z − 9/4 + O ( Z − 7/2 ) , so 2 ( f ′ / f ) ȷ ~ ′ = − 2 ξ ′ ȷ ~ ′ + O ( Z − 13 / 4 ) = ( 24 v 1 / a 2 ) Z − 2 + O ( Z − 13 / 4 ) 2(f^\prime/f)\tilde{\jmath}^\prime=-2\xi^\prime\tilde{\jmath}^\prime+O(Z^{-13/4})=(24v_1/a^2)Z^{-2}+O(Z^{-13/4}) 2 ( f ′ / f ) ~ ′ = − 2 ξ ′ ~ ′ + O ( Z − 13/4 ) = ( 24 v 1 / a 2 ) Z − 2 + O ( Z − 13/4 ) , while ȷ ~ ′ ′ = O ( Z − 13 / 4 ) \tilde{\jmath}^{\prime\prime}=O(Z^{-13/4}) ~ ′′ = O ( Z − 13/4 ) . Comparing the coefficients of Z − 2 Z^{-2} Z − 2 gives 9 / 64 + 24 v 1 / a 2 = − 1 / 4 9/64+24v_1/a^2=-1/4 9/64 + 24 v 1 / a 2 = − 1/4 , that is v 1 = − 25 a 2 / 1536 = − 1 / 8 v_1=-25a^2/1536=-1/8 v 1 = − 25 a 2 /1536 = − 1/8 , since a 2 = 192 / 25 a^2=192/25 a 2 = 192/25 . Equivalently u 1 , 1 = − 5 / ( 64 3 ) u_{1,1}=-5/(64\sqrt{3}) u 1 , 1 = − 5/ ( 64 3 ) .
The exact coefficient contour in [70, Section 2.3] expresses u g , 0 u_{g,0} u g , 0 as a circle integral and deforms it onto the five rays arg t = 2 π k / 5 \arg t=2\pi k/5 arg t = 2 π k /5 , k = 0 , ± 1 , ± 2 k=0,\pm1,\pm2 k = 0 , ± 1 , ± 2 ; the symmetry of the tritronquée solutions reduces their sum to five times the positive ray [70, equations (2.25) and (2.29)]$:
u g , 0 = − 5 2 π i ∫ R 0 ∞ t 5 g − 2 ( u 1 ( t 2 ) − u 0 ( t 2 ) ) d t + r g ( R 0 ) , ∣ r g ( R 0 ) ∣ ≤ C R 0 5 g . u_{g,0}=-\frac{5}{2\pi i}\int_{R_0}^{\infty}t^{5g-2}\left(u_1(t^2)-u_0(t^2)\right)\,\mathrm{d}t+r_g(R_0),\qquad\lvert r_g(R_0)\rvert\le CR_0^{5g}. u g , 0 = − 2 π i 5 ∫ R 0 ∞ t 5 g − 2 ( u 1 ( t 2 ) − u 0 ( t 2 ) ) d t + r g ( R 0 ) , ∣ r g ( R 0 )∣ ≤ C R 0 5 g .
for a fixed radius R 0 R_0 R 0 , large enough for [70] and such that ξ 0 = a R 0 5 / 2 \xi_0=aR_0^{5/2} ξ 0 = a R 0 5/2 is at least the threshold in the jump expansion above. Put ξ = a t 5 / 2 \xi=at^{5/2} ξ = a t 5/2 , so that t 5 g − 9 / 4 d t = 2 5 a − ν g ξ ν g − 1 d ξ t^{5g-9/4}\,\mathrm{d}t=\frac{2}{5}a^{-\nu_g}\xi^{\nu_g-1}\,\mathrm{d}\xi t 5 g − 9/4 d t = 5 2 a − ν g ξ ν g − 1 d ξ , and write u 1 − u 0 = − S 1 e − ξ ξ − 1 / 4 ȷ ~ ( ξ ) u_1-u_0=-S_1e^{-\xi}\xi^{-1/4}\tilde{\jmath}(\xi) u 1 − u 0 = − S 1 e − ξ ξ − 1/4 ~ ( ξ ) . The factor 5 ⋅ 2 5 5\cdot\frac{2}{5} 5 ⋅ 5 2 cancels the 2 2 2 in 2 π i 2\pi i 2 π i , and the definition of Z g Z_g Z g gives exactly
Z g = 1 Γ ( ν g ) ∫ ξ 0 ∞ e − ξ ξ ν g − 1 ȷ ~ ( ξ ) d ξ + O ( D 1 g Γ ( ν g ) ) , D 1 = a 2 R 0 5 . Z_g=\frac{1}{\Gamma(\nu_g)}\int_{\xi_0}^{\infty}e^{-\xi}\xi^{\nu_g-1}\tilde{\jmath}(\xi)\,\mathrm{d}\xi+O\left(\frac{D_1^g}{\Gamma(\nu_g)}\right),\qquad D_1=a^2R_0^5. Z g = Γ ( ν g ) 1 ∫ ξ 0 ∞ e − ξ ξ ν g − 1 ~ ( ξ ) d ξ + O ( Γ ( ν g ) D 1 g ) , D 1 = a 2 R 0 5 .
the error term is ( π i / S 1 ) a ν g r g ( R 0 ) / Γ ( ν g ) (\pi i/S_1)a^{\nu_g}r_g(R_0)/\Gamma(\nu_g) ( π i / S 1 ) a ν g r g ( R 0 ) /Γ ( ν g ) , of modulus at most C ∣ π / S 1 ∣ a − 1 / 2 D 1 g / Γ ( ν g ) C\lvert\pi/S_1\rvert a^{-1/2}D_1^g/\Gamma(\nu_g) C ∣ π / S 1 ∣ a − 1/2 D 1 g /Γ ( ν g ) since ν g = 2 g − 1 2 \nu_g=2g-\frac12 ν g = 2 g − 2 1 . Applying the jump remainder with L = K + 1 ≤ g / 2 L=K+1\le g/2 L = K + 1 ≤ g /2 gives C B L L ! Γ ( ν g − L ) / Γ ( ν g ) ≤ C B L L ! g − L CB^LL!\Gamma(\nu_g-L)/\Gamma(\nu_g)\le CB^LL!g^{-L} C B L L ! Γ ( ν g − L ) /Γ ( ν g ) ≤ C B L L ! g − L . The omitted low integrals and inner circle are O ( e − c g log g ) O(e^{-c_g\log g}) O ( e − c g l o g g ) uniformly for K ≤ c 0 g K\le c_0g K ≤ c 0 g , with c 0 c_0 c 0 small. For the low integrals, bound the finite sum by C ξ 0 ν g ∑ j < L B j ξ 0 − j / Γ ( ν g ) C\xi_0^{\nu_g}\sum_{j<L}B^j\xi_0^{-j}/\Gamma(\nu_g) C ξ 0 ν g ∑ j < L B j ξ 0 − j /Γ ( ν g ) and use Stirling’s lower bound. Thus the finite falling-factorial expansion of Z g Z_g Z g has a Gevrey-1 remainder. Re-expansion in g − 1 g^{-1} g − 1 preserves it:
Γ ( ν g − j ) Γ ( ν g ) = ( 2 g ) − j ∏ a = 1 j ( 1 − a + 1 / 2 2 g ) − 1 . \frac{\Gamma(\nu_g-j)}{\Gamma(\nu_g)}=(2g)^{-j}\prod_{a=1}^{j}\left(1-\frac{a+1/2}{2g}\right)^{-1}. Γ ( ν g ) Γ ( ν g − j ) = ( 2 g ) − j a = 1 ∏ j ( 1 − 2 g a + 1/2 ) − 1 .
Cauchy on radius 1 / ( D 0 ( K + 1 ) ) 1/(D_0(K+1)) 1/ ( D 0 ( K + 1 )) for the product, with a fixed D 0 ≥ 4 D_0\ge4 D 0 ≥ 4 , on which each factor has modulus at most two and which contains 1 / ( 2 g ) 1/(2g) 1/ ( 2 g ) in its half-radius disk once the linear-order constant is small, followed by j ! ( K + 1 ) K + 1 − j ≤ e K + 1 ( K + 1 ) ! j!(K+1)^{K+1-j}\le e^{K+1}(K+1)! j ! ( K + 1 ) K + 1 − j ≤ e K + 1 ( K + 1 )! , bounds the total omitted inverse-genus powers by C B K + 1 ( K + 1 ) ! g − K − 1 CB^{K+1}(K+1)!g^{-K-1} C B K + 1 ( K + 1 )! g − K − 1 .
The exact identity of Itzykson–Zuber [106, equation (6.10)] is
⟨ τ 2 3 g − 3 ⟩ g = 2 g ( 3 g − 3 ) ! ( 5 g − 5 ) ( 5 g − 3 ) ϕ g , g ≥ 2. \left\langle\tau_2^{3g-3}\right\rangle_g=\frac{2^{g}(3g-3)!}{(5g-5)(5g-3)}\phi_g,\qquad g\ge2. ⟨ τ 2 3 g − 3 ⟩ g = ( 5 g − 5 ) ( 5 g − 3 ) 2 g ( 3 g − 3 )! ϕ g , g ≥ 2.
where ϕ ( t ) = ∑ g ≥ 0 ϕ g t ( 1 − 5 g ) / 2 \phi(t)=\sum_{g\ge0}\phi_gt^{(1-5g)/2} ϕ ( t ) = ∑ g ≥ 0 ϕ g t ( 1 − 5 g ) /2 , ϕ 0 = − 1 \phi_0=-1 ϕ 0 = − 1 , satisfies ϕ ′ ′ / 3 + ϕ 2 = t \phi^{\prime\prime}/3+\phi^2=t ϕ ′′ /3 + ϕ 2 = t . The substitution ϕ ( t ) = − 2 1 / 5 u ( 2 − 2 / 5 t ) \phi(t)=-2^{1/5}u(2^{-2/5}t) ϕ ( t ) = − 2 1/5 u ( 2 − 2/5 t ) gives exactly ϕ g = − 2 g u g , 0 \phi_g=-2^gu_{g,0} ϕ g = − 2 g u g , 0 . Combined with the expansion of Z g Z_g Z g above, this gives, in the normalization of (23) and for every fixed K K K ,
π C g ( 2 3 g − 3 ) = 1 + ∑ j = 1 K d j g − j + O K ( g − K − 1 ) . (336) \pi C_g(2^{3g-3})=1+\sum_{j=1}^{K}d_jg^{-j}+O_K(g^{-K-1}). \tag*{(336)} π C g ( 2 3 g − 3 ) = 1 + j = 1 ∑ K d j g − j + O K ( g − K − 1 ) . (336)
Lemma D.11(iii) below gives the Gevrey-1 form of this remainder, uniform for K ≤ c 0 g K\le c_0g K ≤ c 0 g . The leading constant is also independently given by [83, Theorem 1]. To include the markings, put h 0 = 2 g − 2 h_0=2g-2 h 0 = 2 g − 2 , h = h 0 + p h=h_0+p h = h 0 + p , a ∗ = 2 / 15 a_\ast=2/15 a ∗ = 2/15 , and define the exact normalized endpoint
E g , p = π U ‾ g , p ( 0 p ; y 0 ) Γ ( h ) 4 g − a ∗ − h a ∗ / ( π h ) . E_{g,p}=\frac{\pi\overline{U}_{g,p}(0^p;y_0)}{\Gamma(h)^{4g-a_\ast-h}\sqrt{a_\ast/(\pi h)}}. E g , p = Γ ( h ) 4 g − a ∗ − h a ∗ / ( π h ) π U g , p ( 0 p ; y 0 ) .
For p ≥ 1 p \ge1 p ≥ 1 this is the limit of π Φ g ( p ) / J g , p \pi\Phi_g^{(p)}/\mathcal{J}_{g,p} π Φ g ( p ) / J g , p ; at p = 0 p=0 p = 0 it is defined by the same exact pure-ψ \psi ψ expression. String and dilaton give E g , p + 1 / E g , p = ( h 0 + 4 p / 5 ) ( h 0 + p ) − 1 ( h 0 + p + 1 ) / ( h 0 + p ) E_{g,p+1}/E_{g,p}=(h_0+4p/5)(h_0+p)^{-1}\sqrt{(h_0+p+1)/(h_0+p)} E g , p + 1 / E g , p = ( h 0 + 4 p /5 ) ( h 0 + p ) − 1 ( h 0 + p + 1 ) / ( h 0 + p ) . The base conversion gives exactly U ‾ g , 0 ( y 0 ) = − 108 9 u g , 0 / [ 27 ( 5 g − 5 ) ( 5 g − 3 ) ] \overline{U}_{g,0}(y_0)=-108^9u_{g,0}/[27(5g-5)(5g-3)] U g , 0 ( y 0 ) = − 10 8 9 u g , 0 / [ 27 ( 5 g − 5 ) ( 5 g − 3 )] . Substitution therefore proves
E g , p = Z g Γ ( h 0 + 3 / 2 ) Γ ( h 0 ) h 0 3 / 2 R p ( h 0 ) 1 + 4 / ( 5 h 0 ) , R p ( h 0 ) = 1 + p / h 0 ∏ i = 0 p − 1 1 + 4 i / ( 5 h 0 ) 1 + i / h 0 . (337) \begin{aligned}
E_{g,p}&=Z_g\frac{\Gamma(h_0+3/2)}{\Gamma(h_0)h_0^{3/2}}\frac{R_p(h_0)}{1+4/(5h_0)},\\
R_p(h_0)&=\sqrt{1+p/h_0}\prod_{i=0}^{p-1}\frac{1+4i/(5h_0)}{1+i/h_0}. \tag*{(337)}
\end{aligned} E g , p R p ( h 0 ) = Z g Γ ( h 0 ) h 0 3/2 Γ ( h 0 + 3/2 ) 1 + 4/ ( 5 h 0 ) R p ( h 0 ) , = 1 + p / h 0 i = 0 ∏ p − 1 1 + i / h 0 1 + 4 i / ( 5 h 0 ) . (337)
This identity is exact for g ≥ 2 g\ge2 g ≥ 2 , p ≥ 0 p\ge0 p ≥ 0 , with the empty product one.
The quantitative endpoint input is the following lemma. It is used at the matching point in the proof of Theorem 8.2 , and it gives the Gevrey-1 form of (D.48).
Lemma D.11 (Gevrey-1 endpoint expansion). Let h 0 = 2 g − 2 h_0=2g-2 h 0 = 2 g − 2 and h = h 0 + p h=h_0+p h = h 0 + p . There are constants c 0 , C , B > 0 c_0,C,B>0 c 0 , C , B > 0 and, for each p ≥ 0 p\ge0 p ≥ 0 , constants C p , B p > 0 C_p,B_p>0 C p , B p > 0 such that the following hold for all g ≥ 2 g\ge2 g ≥ 2 and 0 ≤ K ≤ c 0 g 0\le K\le c_0g 0 ≤ K ≤ c 0 g .
(i) With v 0 = 1 v_0=1 v 0 = 1 and v j = a j u j , 1 v_j=a^j u_{j,1} v j = a j u j , 1 ,
∣ Z g − ∑ j = 0 K v j Γ ( ν g − j ) Γ ( ν g ) ∣ ≤ C B K + 1 ( K + 1 ) ! g − K − 1 , \left|Z_g-\sum_{j=0}^{K}v_j\frac{\Gamma(\nu_g-j)}{\Gamma(\nu_g)}\right|\le CB^{K+1}(K+1)!g^{-K-1}, Z g − j = 0 ∑ K v j Γ ( ν g ) Γ ( ν g − j ) ≤ C B K + 1 ( K + 1 )! g − K − 1 , and the same bound holds after re-expanding the sum in powers of g − 1 g^{-1} g − 1 or of h 0 − 1 h_0^{-1} h 0 − 1 and truncating at order K K K . For each fixed p p p , re-expansion in powers of h − 1 h^{-1} h − 1 has remainder at most C p B p K + 1 ( K + 1 ) ! h − K − 1 C_pB_p^{K+1}(K+1)!h^{-K-1} C p B p K + 1 ( K + 1 )! h − K − 1 .
(ii) The function log E g , p \log E_{g,p} log E g , p has an asymptotic expansion ∑ j ≥ 1 a p , j ∗ h − j \sum_{j\ge1}a^*_{p,j}h^{-j} ∑ j ≥ 1 a p , j ∗ h − j as g → ∞ g\to\infty g → ∞ , whose coefficients are obtained by re-expanding (D.49) and (i) in powers of h − 1 h^{-1} h − 1 , and
∣ log E g , p − ∑ j = 1 K a p , j ∗ h − j ∣ + ∣ E g , p − ∑ j = 0 K ε p , j h − j ∣ ≤ C p B p K + 1 ( K + 1 ) ! h − K − 1 , \left|\log E_{g,p}-\sum_{j=1}^{K}a^*_{p,j}h^{-j}\right|+\left|E_{g,p}-\sum_{j=0}^{K}\varepsilon_{p,j}h^{-j}\right|\le C_pB_p^{K+1}(K+1)!h^{-K-1}, log E g , p − j = 1 ∑ K a p , j ∗ h − j + E g , p − j = 0 ∑ K ε p , j h − j ≤ C p B p K + 1 ( K + 1 )! h − K − 1 , where ε p , j = [ t j ] exp ( ∑ i = 1 j a p , i ∗ t i ) \varepsilon_{p,j}=[t^j]\exp\left(\sum_{i=1}^{j}a^*_{p,i}t^i\right) ε p , j = [ t j ] exp ( ∑ i = 1 j a p , i ∗ t i ) .
(iii) For the coefficients d j d_j d j of (D.48), with d 0 = 1 d_0=1 d 0 = 1 ,
∣ π C g ( 2 3 g − 3 ) − ∑ j = 0 K d j g − j ∣ ≤ C B K + 1 ( K + 1 ) ! g − K − 1 . \left|\pi C_g(2^{3g-3})-\sum_{j=0}^{K}d_jg^{-j}\right|\le CB^{K+1}(K+1)!g^{-K-1}. π C g ( 2 3 g − 3 ) − j = 0 ∑ K d j g − j ≤ C B K + 1 ( K + 1 )! g − K − 1 .
Proof. (i) This is the Gevrey-1 remainder for Z g Z_g Z g established above: the jump remainder in the integral representation of Z g Z_g Z g gives the bound in falling factorials for L = K + 1 ≤ g / 2 L=K+1\le g/2 L = K + 1 ≤ g /2 , and Cauchy’s estimate on the radius 1 / ( D 0 ( K + 1 ) ) 1/(D_0(K+1)) 1/ ( D 0 ( K + 1 )) for the product ∏ a ≤ j ( 1 − ( a + 1 / 2 ) / ( 2 g ) ) − 1 \prod_{a\le j}(1-(a+1/2)/(2g))^{-1} ∏ a ≤ j ( 1 − ( a + 1/2 ) / ( 2 g ) ) − 1 , followed by j ! ( K + 1 ) K + 1 − j ≤ e K + 1 ( K + 1 ) ! j!(K+1)^{K+1-j}\le e^{K+1}(K+1)! j ! ( K + 1 ) K + 1 − j ≤ e K + 1 ( K + 1 )! , converts it into a bound in powers of g − 1 g^{-1} g − 1 . Since g − 1 = 2 h − 1 ( 1 + ( 2 − p ) h − 1 ) − 1 g^{-1}=2h^{-1}(1+(2-p)h^{-1})^{-1} g − 1 = 2 h − 1 ( 1 + ( 2 − p ) h − 1 ) − 1 is a rational function of h − 1 h^{-1} h − 1 , analytic on a disk about zero depending on the fixed value of p p p , the same Cauchy estimate applied to the degree-K K K polynomial on the radius 1 / ( D p ( K + 1 ) ) 1/(D_p(K+1)) 1/ ( D p ( K + 1 )) gives the re-expansion in h − 1 h^{-1} h − 1 with constants depending on p p p ; taking p = 0 p=0 p = 0 gives the re-expansion in h 0 − 1 h_0^{-1} h 0 − 1 .
(ii) By (D.49),
log E g , p = log Z g + log Γ ( h 0 + 3 / 2 ) Γ ( h 0 ) h 0 3 / 2 + log R p ( h 0 ) − log ( 1 + 4 5 h 0 ) . \log E_{g,p}=\log Z_g+\log\frac{\Gamma(h_0+3/2)}{\Gamma(h_0)h_0^{3/2}}+\log R_p(h_0)-\log\left(1+\frac{4}{5h_0}\right). log E g , p = log Z g + log Γ ( h 0 ) h 0 3/2 Γ ( h 0 + 3/2 ) + log R p ( h 0 ) − log ( 1 + 5 h 0 4 ) .
The first term is covered by (i), since Z g = 1 + O ( g − 1 ) Z_g=1+O(g^{-1}) Z g = 1 + O ( g − 1 ) stays in a fixed disk on which the logarithm is analytic. For the Gamma factor, for fixed α > 0 \alpha>0 α > 0 ,
log Γ ( h 0 + α ) Γ ( h 0 ) h 0 α = ∫ 0 ∞ e − h 0 t [ α − 1 − e − α t 1 − e − t ] d t t . \log\frac{\Gamma(h_0+\alpha)}{\Gamma(h_0)h_0^\alpha}=\int_0^\infty e^{-h_0t}\left[\alpha-\frac{1-e^{-\alpha t}}{1-e^{-t}}\right]\frac{\mathrm{d}t}{t}. log Γ ( h 0 ) h 0 α Γ ( h 0 + α ) = ∫ 0 ∞ e − h 0 t [ α − 1 − e − t 1 − e − α t ] t d t .
The kernel is analytic at zero on a fixed disk and bounded outside a small positive interval. Its global Taylor remainder is bounded by C B L t L − 1 CB^Lt^{L-1} C B L t L − 1 , giving C B L ( L − 1 ) ! h 0 − L CB^L(L-1)!h_0^{-L} C B L ( L − 1 )! h 0 − L after integration. The finite insertion factors R p ( h 0 ) R_p(h_0) R p ( h 0 ) and ( 1 + 4 / ( 5 h 0 ) ) − 1 (1+4/(5h_0))^{-1} ( 1 + 4/ ( 5 h 0 ) ) − 1 have convergent expansions at h 0 − 1 = 0 h_0^{-1}=0 h 0 − 1 = 0 . Finite shifts from h 0 h_0 h 0 to h = h 0 + p h=h_0+p h = h 0 + p preserve the factorial error by the same Cauchy estimate for rational products used in (i). For logarithms, exponentials and inverses, take the degree-K K K polynomial on radius 1 / ( D 0 ( K + 1 ) ) 1/(D_0(K+1)) 1/ ( D 0 ( K + 1 )) ; its unit or zero-constant part stays bounded away from the relevant singularity. Cauchy’s tail and the bounded real derivative of the composed function preserve the remainder bound. This proves (ii); uniqueness of asymptotic expansions identifies a p , j ∗ a_{p,j}^* a p , j ∗ with the coefficients of the formal re-expansion.
(iii) By the Itzykson–Zuber identity, the conversion ϕ g = − 2 g u g , 0 \phi_g=-2^g u_{g,0} ϕ g = − 2 g u g , 0 and the definition of Z g Z_g Z g , in the normalization (23) ,
π C g ( 2 3 g − 3 ) = Z g Ξ ( g ) , Ξ ( g ) = 3 1 / 4 2 π 4 g 15 3 g − 3 2 2 g Γ ( 3 g − 2 ) Γ ( ν g ) 3 5 g − 5 Γ ( 5 g − 5 ) ( 5 g − 5 ) ( 5 g − 3 ) α ν g . \pi C_g(2^{3g-3})=Z_g\Xi(g),\qquad\Xi(g)=\frac{3^{1/4}}{2\sqrt{\pi}}\frac{4^g15^{3g-3}2^{2g}\Gamma(3g-2)\Gamma(\nu_g)}{3^{5g-5}\Gamma(5g-5)(5g-5)(5g-3)\alpha^{\nu_g}}. π C g ( 2 3 g − 3 ) = Z g Ξ ( g ) , Ξ ( g ) = 2 π 3 1/4 3 5 g − 5 Γ ( 5 g − 5 ) ( 5 g − 5 ) ( 5 g − 3 ) α ν g 4 g 1 5 3 g − 3 2 2 g Γ ( 3 g − 2 ) Γ ( ν g ) .
where 3 1 / 4 / ( 2 π ) = − S 1 / i 3^{1/4}/(2\sqrt{\pi})=-S_1/i 3 1/4 / ( 2 π ) = − S 1 / i ; this is an explicit product of powers c g c^g c g , of Gamma functions at arguments α g + β \alpha g+\beta α g + β and of a rational function of g g g . Stirling’s formula Γ ( α g + β ) ∼ 2 π ( α g ) α g + β − 1 / 2 e − α g \Gamma(\alpha g+\beta)\sim\sqrt{2\pi}(\alpha g)^{\alpha g+\beta-1/2}e^{-\alpha g} Γ ( α g + β ) ∼ 2 π ( α g ) α g + β − 1/2 e − α g gives Ξ ( g ) = 1 + O ( g − 1 ) \Xi(g)=1+O(g^{-1}) Ξ ( g ) = 1 + O ( g − 1 ) directly: the exponentials and the powers of g g g and of 5 5 5 cancel, the powers of 2 2 2 and 3 3 3 reduce to 2 1 / 2 3 − 1 / 4 2^{1/2}3^{-1/4} 2 1/2 3 − 1/4 , and 2 π 2 1 / 2 3 − 1 / 4 3 1 / 4 / ( 2 π ) = 1 \sqrt{2\pi}\,2^{1/2}3^{-1/4}3^{1/4}/(2\sqrt{\pi})=1 2 π 2 1/2 3 − 1/4 3 1/4 / ( 2 π ) = 1 . Together with Z g = 1 + O ( g − 1 ) Z_g=1+O(g^{-1}) Z g = 1 + O ( g − 1 ) , which is the leading case of [70, Section 2.3], this proves (97) . Binet’s formula log Γ ( x ) = ( x − 1 2 ) log x − x + 1 2 log 2 π + μ ( x ) \log\Gamma(x)=(x-\frac{1}{2})\log x-x+\frac{1}{2}\log2\pi+\mu(x) log Γ ( x ) = ( x − 2 1 ) log x − x + 2 1 log 2 π + μ ( x ) , with μ ( x ) = ∫ 0 ∞ ( 1 e t − 1 − 1 t + 1 2 ) e − x t t d t \mu(x)=\int_0^\infty\left(\frac{1}{e^t-1}-\frac{1}{t}+\frac{1}{2}\right)\frac{e^{-xt}}{t}\,\mathrm{d}t μ ( x ) = ∫ 0 ∞ ( e t − 1 1 − t 1 + 2 1 ) t e − x t d t , expresses log Ξ ( g ) \log\Xi(g) log Ξ ( g ) as a sum of elementary terms and of integrals ± μ ( α g + β ) \pm\mu(\alpha g+\beta) ± μ ( α g + β ) . The elementary terms are convergent series in g − 1 g^{-1} g − 1 once their contributions of orders g log g g\log g g log g , g g g , log g \log g log g and g 0 g^0 g 0 are collected, and these must cancel because log Ξ ( g ) → 0 \log\Xi(g)\to0 log Ξ ( g ) → 0 . The kernel of μ \mu μ is analytic on ∣ t ∣ < 2 π |t|<2\pi ∣ t ∣ < 2 π and bounded on the positive axis, so the Taylor-remainder argument just used for the Gamma factor gives for each μ ( α g + β ) \mu(\alpha g+\beta) μ ( α g + β ) an expansion in ( α g + β ) − 1 (\alpha g+\beta)^{-1} ( α g + β ) − 1 with remainder C B L ( L − 1 ) ! ( α g + β ) − L CB^L(L-1)!(\alpha g+\beta)^{-L} C B L ( L − 1 )! ( α g + β ) − L , and the rational re-expansion in g − 1 g^{-1} g − 1 is handled as in (i). Hence log Ξ ( g ) \log\Xi(g) log Ξ ( g ) , and by exponentiation Ξ ( g ) \Xi(g) Ξ ( g ) , have Gevrey-1 expansions in g − 1 g^{-1} g − 1 for K ≤ c 0 g K\le c_0g K ≤ c 0 g . Multiplying by (i), as in (ii), proves (iii). □ \square □
At first order, (D.49), the value v 1 = − 1 / 8 v_1=-1/8 v 1 = − 1/8 derived above, and ν g − 1 = h 0 + 1 / 2 \nu_g-1=h_0+1/2 ν g − 1 = h 0 + 1/2 give
a p , 1 ∗ = − 1 8 + 3 8 − 4 5 + p 2 − p ( p − 1 ) 10 = − 2 p 2 + 12 p − 11 20 = t p , a_{p,1}^*=-\frac{1}{8}+\frac{3}{8}-\frac{4}{5}+\frac{p}{2}-\frac{p(p-1)}{10}=\frac{-2p^2+12p-11}{20}=t_p, a p , 1 ∗ = − 8 1 + 8 3 − 5 4 + 2 p − 10 p ( p − 1 ) = 20 − 2 p 2 + 12 p − 11 = t p ,
the constant in (105) ; replacing h 0 h_0 h 0 by h = h 0 + p h=h_0+p h = h 0 + p does not change the first-order coefficient. In particular the first coefficient in (D.48) is d 1 = − 2 / 9 d_1=-2/9 d 1 = − 2/9 . Indeed, at p = 0 p=0 p = 0 the definition of E g , 0 E_{g,0} E g , 0 and (99) give
π C g ( 2 3 g − 3 ) = E g , 0 N h 0 , N h 0 = M g , 0 ( y 0 ) Γ ( h 0 ) 4 g − a ∗ h 0 ν g 2 π h 0 ν g = 2 π h 0 ν g Pr ( S h 0 = 3 h 0 / 2 ) , ν = 15 4 , \pi C_g(2^{3g-3})=\frac{E_{g,0}}{N_{h_0}},\qquad N_{h_0}=\frac{\mathcal{M}_{g,0}(y_0)}{\Gamma(h_0)4^{g-a_*}h_0^{\nu_g}}\sqrt{2\pi h_0\nu_g}=\sqrt{2\pi h_0\nu_g}\Pr(S_{h_0}=3h_0/2),\qquad\nu=\frac{15}{4}, π C g ( 2 3 g − 3 ) = N h 0 E g , 0 , N h 0 = Γ ( h 0 ) 4 g − a ∗ h 0 ν g M g , 0 ( y 0 ) 2 π h 0 ν g = 2 π h 0 ν g Pr ( S h 0 = 3 h 0 /2 ) , ν = 4 15 ,
where S h 0 S_{h_0} S h 0 is the sum of h 0 h_0 h 0 independent copies of the geometric law with generating function ( 2 / 5 ) / ( 1 − 3 w / 5 ) (2/5)/(1-3w/5) ( 2/5 ) / ( 1 − 3 w /5 ) ; this is the normalized negative-binomial coefficient of Theorem 8.1 at displacement zero, so N h 0 = 1 − 19 / ( 180 h 0 ) + O ( h 0 − 2 ) N_{h_0}=1-19/(180h_0)+O(h_0^{-2}) N h 0 = 1 − 19/ ( 180 h 0 ) + O ( h 0 − 2 ) . On the other side, Z g = 1 + v 1 Γ ( ν g − 1 ) / Γ ( ν g ) + O ( g − 2 ) = 1 − 1 / ( 8 h 0 ) + O ( h 0 − 2 ) Z_g=1+v_1\Gamma(\nu_g-1)/\Gamma(\nu_g)+O(g^{-2})=1-1/(8h_0)+O(h_0^{-2}) Z g = 1 + v 1 Γ ( ν g − 1 ) /Γ ( ν g ) + O ( g − 2 ) = 1 − 1/ ( 8 h 0 ) + O ( h 0 − 2 ) , because ν g − 1 = h 0 + 1 / 2 \nu_g-1=h_0+1/2 ν g − 1 = h 0 + 1/2 and v 1 = − 1 / 8 v_1=-1/8 v 1 = − 1/8 ; Γ ( h 0 + 3 / 2 ) / ( Γ ( h 0 ) h 0 3 / 2 ) = 1 + 3 / ( 8 h 0 ) + O ( h 0 − 2 ) \Gamma(h_0+3/2)/(\Gamma(h_0)h_0^{3/2})=1+3/(8h_0)+O(h_0^{-2}) Γ ( h 0 + 3/2 ) / ( Γ ( h 0 ) h 0 3/2 ) = 1 + 3/ ( 8 h 0 ) + O ( h 0 − 2 ) ; and ( 1 + 4 / ( 5 h 0 ) ) − 1 = 1 − 4 / ( 5 h 0 ) + O ( h 0 − 2 ) (1+4/(5h_0))^{-1}=1-4/(5h_0)+O(h_0^{-2}) ( 1 + 4/ ( 5 h 0 ) ) − 1 = 1 − 4/ ( 5 h 0 ) + O ( h 0 − 2 ) . Hence, by (D.49),
π C g ( 2 3 g − 3 ) = 1 + 1 h 0 ( − 1 8 + 3 8 − 4 5 + 19 180 ) + O ( h 0 − 2 ) = 1 − 4 9 h 0 + O ( h 0 − 2 ) = 1 − 2 9 g + O ( g − 2 ) , \pi C_g(2^{3g-3})=1+\frac{1}{h_0}\left(-\frac{1}{8}+\frac{3}{8}-\frac{4}{5}+\frac{19}{180}\right)+O(h_0^{-2})=1-\frac{4}{9h_0}+O(h_0^{-2})=1-\frac{2}{9g}+O(g^{-2}), π C g ( 2 3 g − 3 ) = 1 + h 0 1 ( − 8 1 + 8 3 − 5 4 + 180 19 ) + O ( h 0 − 2 ) = 1 − 9 h 0 4 + O ( h 0 − 2 ) = 1 − 9 g 2 + O ( g − 2 ) ,
since h 0 = 2 g − 2 h_0 = 2g - 2 h 0 = 2 g − 2 . This value agrees with the expansion recorded by Guo, Yang and Zagier [83, equation (25)], which rests on the same Painlevé I input; it is used in the proof of Theorem 8.1 .
Define
a p , j ( Y ) = a p , j ∗ − ∫ Y y 0 γ p , j ( v ) A ( v ) r ( v ) d v , L p , j ( Y ) = [ t j ] exp ( ∑ i = 1 j a p , i ( Y ) t i ) . (338) a_{p,j}(Y) = a^*_{p,j} - \int_Y^{y_0} \frac{\gamma_{p,j}(v)}{A(v)r(v)}\,\mathrm{d}v,\qquad\mathcal{L}_{p,j}(Y) = [t^j]\exp\left(\sum_{i=1}^{j} a_{p,i}(Y)t^i\right). \tag*{(338)} a p , j ( Y ) = a p , j ∗ − ∫ Y y 0 A ( v ) r ( v ) γ p , j ( v ) d v , L p , j ( Y ) = [ t j ] exp ( i = 1 ∑ j a p , i ( Y ) t i ) . (338)
Proposition D.9 , inversion and the bounded division and integration above give common-domain bounds ∥ a p , j ∥ + ∥ L p , j ∥ ≤ C p B p j j ! \lVert a_{p,j}\rVert+ \lVert\mathcal{L}_{p,j}\rVert\le C_p \mathcal{B}_p^j j! ∥ a p , j ∥ + ∥ L p , j ∥ ≤ C p B p j j ! . For formal exponentiation this follows also by Cauchy on the finite log polynomial at radius 1 / ( D 0 j ) 1/(D_0j) 1/ ( D 0 j ) . The first coefficient is L p , 1 = L p \mathcal{L}_{p,1} = \mathcal{L}_p L p , 1 = L p . These are the coefficients in Theorem 8.2 .
Proof of Theorem 8.2. We first expand the logarithmic derivative, then fix its integration constant near the endpoint. After integration, exponentiation gives the claimed expansion. The extra orders retained below compensate for the error accumulated near the zero of A r A_r A r .
Put F g ( Y ) = log ( π Φ g ( p ) / J g , p ) F_g(Y) = \log(\pi\Phi_g^{(p)}/J_{g,p}) F g ( Y ) = log ( π Φ g ( p ) / J g , p ) . For fixed p p p the inverse cusp ratio is bounded, and the finite inverse identity (D.33), with Proposition D.9 , transfers the remainder to 1 / C p 1/C_p 1/ C p with error C B N N ! h − N CB^N N!h^{-N} C B N N ! h − N , for N ≤ c h N \le ch N ≤ c h . The exact dilaton and (96) give
F g ′ = h A r ( 1 / C p − 1 − β p / h ) . F'_g = \frac{h}{A_r}\left(1/C_p - 1 - \beta_p/h\right). F g ′ = A r h ( 1/ C p − 1 − β p / h ) .
For L = K + 1 L = K + 1 L = K + 1 , use N = L + 2 = K + 3 N = L + 2 = K + 3 N = L + 2 = K + 3 , retaining γ p , j \gamma_{p,j} γ p , j through j = L j = L j = L . Its unknown derivative error is C B L + 2 ( L + 2 ) ! h − L − 1 / ( A r ) CB^{L+2}(L + 2)!h^{-L-1}/(A_r) C B L + 2 ( L + 2 )! h − L − 1 / ( A r ) . Choose δ ∗ = h − L − 1 \delta_* = h^{-L-1} δ ∗ = h − L − 1 . The positive comparison (101) and the exact first equality of (102) change F g F_g F g from its endpoint value by at most C h δ ∗ = C h − L Ch\delta_* = Ch^{-L} C h δ ∗ = C h − L on this strip. The leading harmonic approximation in the second equality of (102) is not used. Lemma D.11 (ii) supplies the value at the matching point through order L L L , with remainder C p B p L + 1 ( L + 1 ) ! h − L − 1 C_pB_p^{L+1}(L + 1)!h^{-L-1} C p B p L + 1 ( L + 1 )! h − L − 1 ; relative to L ! h − L L!h^{-L} L ! h − L this is C p B p L + 1 ( L + 1 ) / h ≤ C p ( c + h − 1 ) B p L + 1 ≤ 2 C p B p L + 1 C_pB_p^{L+1}(L + 1)/h \le C_p(c + h^{-1})B_p^{L+1} \le2C_pB_p^{L+1} C p B p L + 1 ( L + 1 ) / h ≤ C p ( c + h − 1 ) B p L + 1 ≤ 2 C p B p L + 1 , a fixed exponential in L L L . The bounded coefficient derivatives from (D.47) show that the whole retained coefficient sum changes by O ( δ ∗ / h ) O(\delta_*/h) O ( δ ∗ / h ) there, after choosing the linear-order constant small.
Integrate backwards from y 0 − δ ∗ y_0 - \delta_* y 0 − δ ∗ . Since A r A_r A r has a simple zero, the unknown error costs at most
C B L + 2 ( L + 2 ) ! h − L − 1 [ 1 + ( L + 1 ) log h ] . CB^{L+2}(L + 2)!h^{-L-1}\left[1 + (L + 1)\log h\right]. C B L + 2 ( L + 2 )! h − L − 1 [ 1 + ( L + 1 ) log h ] .
Relative to L ! h − L L!h^{-L} L ! h − L , its extra factor is a fixed exponential base times ( L + 1 ) ( L + 2 ) [ 1 + ( L + 1 ) log h ] / h (L + 1)(L + 2)[1 + (L + 1)\log h]/h ( L + 1 ) ( L + 2 ) [ 1 + ( L + 1 ) log h ] / h , bounded by another fixed exponential in L L L . Indeed L ≤ c h L \le ch L ≤ c h implies h ≥ L / c h \ge L/c h ≥ L / c ; the decrease of log h / h \log h/h log h / h for h ≥ e h \ge e h ≥ e bounds this factor by C ( L + 2 ) 2 log ( L + 2 ) ≤ C D 2 L C(L + 2)^2\log(L + 2) \le CD_2^L C ( L + 2 ) 2 log ( L + 2 ) ≤ C D 2 L for a fixed D 2 > 1 D_2 > 1 D 2 > 1 . The bounded exceptional values of h h h , L L L are absorbed into C C C . The known coefficients integrate to (D.50). Discarding the last term a p , L h − L a_{p,L}h^{-L} a p , L h − L proves the asserted log error, on both the strip and its complement. Choose c p c_p c p smaller and h p h_p h p larger once so that the inverse-cusp remainder estimate remains valid through order K + 3 K + 3 K + 3 . The endpoint range is linear in h h h .
Finally the finite degree-K K K log polynomial is uniformly bounded on ∣ t ∣ ≤ 1 / ( D 0 ( K + 1 ) ) |t| \le1/(D_0(K + 1)) ∣ t ∣ ≤ 1/ ( D 0 ( K + 1 )) . Cauchy’s exponential tail at t = 1 / h t = 1/h t = 1/ h is C [ D 0 ( K + 1 ) / h ] K + 1 ≤ C ( e D 0 ) K + 1 ( K + 1 ) ! h − K − 1 C[D_0(K + 1)/h]^{K+1} \le C(eD_0)^{K+1}(K + 1)!h^{-K-1} C [ D 0 ( K + 1 ) / h ] K + 1 ≤ C ( e D 0 ) K + 1 ( K + 1 )! h − K − 1 . At the real evaluation point the polynomial is O ( h − 1 ) O(h^{-1}) O ( h − 1 ) and the function F g F_g F g itself is bounded by Proposition 7.1. Thus the exponential is Lipschitz on the required bounded interval and transfers the error bound for the logarithm to (107) . Choosing K = [ c p ′ h ] K = [c'_ph] K = [ c p ′ h ] after the final error base, with B p c p ′ < 1 \mathcal{B}_pc'_p < 1 B p c p ′ < 1 , gives the exponential upper error. □
Appendix E. Coefficient extraction with inverse-genus errors
This appendix proves Theorems 10.1 and 10.2 of § 3.30 , which extract a single coefficient of Φ g ( p ) \Phi_g^{(p)} Φ g ( p ) with an error of order 1 / h 1/h 1/ h , and to all orders, uniformly in the coefficient index. The strategy is the third idea of Subsection 1.5: the coefficient is a contour integral around the real saddle point, the true generating function is replaced there by a finite marked expansion transferred to a small complex neighborhood, and the resulting model integral is evaluated against an exact Poisson factor rather than a Gaussian. The inputs are the marked expansion of Theorem 8.2 , used only through its statement; the positivity of Lemma 9.1, which bounds the modulus of Φ g ( p ) \Phi_g^{(p)} Φ g ( p ) on a circle by its value on the real axis and supplies the Fourier bounds of Subsection 9.3; the endpoint expansions {(9.3)} of Lemma 3.1 ; and, near zero cusp weight, Temme’s change of variables [178]. Subsection E.1 makes the local limit theorem of § 3.25 quantitative; its complex comparison and Fourier estimates are reused afterwards. Subsections E.2–E.4 prove Theorem 10.1 , and Subsections E.5–E.7 prove Theorem 10.2 and compute its first coefficient.
A quantitative local limit
We first make the local limit theorem (114) quantitative. The argument transfers a real-axis error to a complex neighborhood and applies Fourier inversion; the positive convolution factor controls the remaining part of the integral. This gives an error bound by a negative power of min ( h , μ ) \min(h,\mu) min ( h , μ ) . The complex comparison and Fourier estimates established in the proof will also be used in Subsection E.2, where retaining a Poisson factor gives an inverse-genus error uniformly over all coefficient indices.
Proposition E.1. Fix p ≥ 1 p \ge1 p ≥ 1 , and use h , μ , σ h,\mu,\sigma h , μ , σ from (111) . There are constants C p , κ p > 0 C_p,\kappa_p > 0 C p , κ p > 0 such that, with M = min ( h , μ ) M=\min(h,\mu) M = min ( h , μ ) ,
sup k ∈ Z ∣ σ P x ( N = k ) − e − ( k − μ ) 2 / ( 2 σ 2 ) 2 π ∣ ≤ C p M − κ p , (339) \sup_{k\in\mathbb{Z}}\left|\sigma\mathbb{P}_x(N=k)-\frac{e^{-(k-\mu)^2/(2\sigma^2)}}{\sqrt{2\pi}}\right|\le C_pM^{-\kappa_p}, \tag*{(339)} k ∈ Z sup σ P x ( N = k ) − 2 π e − ( k − μ ) 2 / ( 2 σ 2 ) ≤ C p M − κ p , (339) whenever M M M is sufficiently large, uniformly for 0 < x < x c 0<x<x_c 0 < x < x c .
Proof. Write Z = ( N − μ ) / σ Z=(N-\mu)/\sigma Z = ( N − μ ) / σ and L ( z ) = E x e z Z \mathcal{L}(z)=\mathbb{E}_x e^{zZ} L ( z ) = E x e z Z . The integral identity (112) and the convergent endpoint expansions (110) give constants c , C p > 0 c,C_p>0 c , C p > 0 for which
∣ log L ( u ) − u 2 2 ∣ ≤ C p M − 1 / 2 ( ∣ u ∣ + ∣ u ∣ 3 ) , ∣ u ∣ ≤ c M . (340) \left|\log\mathcal{L}(u)-\frac{u^2}{2}\right|\le C_pM^{-1/2}(|u|+|u|^3),\qquad|u|\le c\sqrt{M}. \tag*{(340)} log L ( u ) − 2 u 2 ≤ C p M − 1/2 ( ∣ u ∣ + ∣ u ∣ 3 ) , ∣ u ∣ ≤ c M . (340)
To check the uniformity, note that near zero Ψ \Psi Ψ , D Ψ \mathcal{D}\Psi D Ψ and D 2 Ψ \mathcal{D}^2\Psi D 2 Ψ are comparable to x x x , so the model Taylor remainder is O ( ∣ u ∣ 3 / μ ) O(|u|^3/\sqrt{\mu}) O ( ∣ u ∣ 3 / μ ) and the integrated error is O ( ∣ u ∣ x / μ ) O(|u|x/\sqrt{\mu}) O ( ∣ u ∣ x / μ ) . On an interior compact interval both are O p ( h − 1 / 2 ( ∣ u ∣ + ∣ u ∣ 3 ) ) O_p(h^{-1/2}(|u|+|u|^3)) O p ( h − 1/2 ( ∣ u ∣ + ∣ u ∣ 3 )) . Near criticality, σ ≍ h / ε \sigma\asymp\sqrt{h}/\varepsilon σ ≍ h / ε ; for ∣ u ∣ ≤ c h |u|\le c\sqrt{h} ∣ u ∣ ≤ c h the tilted distance to the critical point stays comparable to ε \varepsilon ε . The bounds Ψ = O ( ε − 1 ) \Psi=O(\varepsilon^{-1}) Ψ = O ( ε − 1 ) and D 2 Ψ = O ( ε − 3 ) \mathcal{D}^2\Psi=O(\varepsilon^{-3}) D 2 Ψ = O ( ε − 3 ) give the same estimate. Shrinking c c c ensures the real moment generating function is defined on the whole interval. As before, the error in (83) is never differentiated.
Put T = log M / 80 T=\sqrt{\log M}/80 T = log M /80 and f ( z ) = L ( z ) − e z 2 / 2 f(z)=\mathcal{L}(z)-e^{z^2/2} f ( z ) = L ( z ) − e z 2 /2 . The function is analytic on ∣ z ∣ ≤ 4 T |z|\le4T ∣ z ∣ ≤ 4 T , because 4 T ≤ c M 4T\le c\sqrt{M} 4 T ≤ c M for all sufficiently large M M M and the moment generating function is defined for ∣ ℜ z ∣ ≤ c M |\Re z|\le c\sqrt{M} ∣ℜ z ∣ ≤ c M , the range of (E.2). Positivity gives ∣ L ( z ) ∣ ≤ L ( ℜ z ) |\mathcal{L}(z)|\le\mathcal{L}(\Re z) ∣ L ( z ) ∣ ≤ L ( ℜ z ) , whence (E.2) implies
sup ∣ z ∣ ≤ 4 T ∣ f ( z ) ∣ ≤ 2 e 9 T 2 = B , sup u ∈ [ − T , T ] ∣ f ( u ) ∣ ≤ C p M − 1 / 2 ( 1 + T ) 3 e T 2 = b . (341) \sup_{|z|\le4T}|f(z)|\le2e^{9T^2}=B,\qquad
\sup_{u\in[-T,T]}|f(u)|\le C_pM^{-1/2}(1+T)^3e^{T^2}=b. \tag*{(341)} ∣ z ∣ ≤ 4 T sup ∣ f ( z ) ∣ ≤ 2 e 9 T 2 = B , u ∈ [ − T , T ] sup ∣ f ( u ) ∣ ≤ C p M − 1/2 ( 1 + T ) 3 e T 2 = b . (341)
For sufficiently large M M M , b ≤ B b\le B b ≤ B . In the disk ∣ z ∣ < 4 T |z|<4T ∣ z ∣ < 4 T slit along [ − T , T ] [-T,T] [ − T , T ] , the harmonic measure of the slit is at least 1 / 2 1/2 1/2 for ∣ z ∣ ≤ T |z|\le T ∣ z ∣ ≤ T . To check this, rescale T T T to one and use G ( w ) = log ∣ w + w 2 − 1 ∣ G(w)=\log|w+\sqrt{w^2-1}| G ( w ) = log ∣ w + w 2 − 1 ∣ , with the slit as its infinity. It vanishes on the slit, is at most log ( 1 + 2 ) \log(1+\sqrt{2}) log ( 1 + 2 ) on ∣ w ∣ ≤ 1 |w|\le1 ∣ w ∣ ≤ 1 , and is at least log ( 4 + 15 ) \log(4+\sqrt{15}) log ( 4 + 15 ) on ∣ w ∣ = 4 |w|=4 ∣ w ∣ = 4 . The maximum principle compares harmonic measure to 1 − G / log ( 4 + 15 ) 1-G/\log(4+\sqrt{15}) 1 − G / log ( 4 + 15 ) , which is at least 1 / 2 1/2 1/2 on the inner disk. Applying the two-constants principle to log ∣ f ∣ \log|f| log ∣ f ∣ therefore yields
∣ f ( i t ) ∣ ≤ b B ≤ C p M − 1 / 4 ( 1 + T ) 3 / 2 e 5 T 2 / 2 , ∣ t ∣ ≤ T . (342) |f(it)|\le\sqrt{bB}\le C_pM^{-1/4}(1+T)^{3/2}e^{5T^2/2},\qquad|t|\le T. \tag*{(342)} ∣ f ( i t ) ∣ ≤ b B ≤ C p M − 1/4 ( 1 + T ) 3/2 e 5 T 2 /2 , ∣ t ∣ ≤ T . (342)
Its integral over this interval is O p ( M − 1 / 8 ) O_p(M^{-1/8}) O p ( M − 1/8 ) . The Fourier estimates of Subsection 9.3 give constants c 0 , c 1 , C > 0 c_0,c_1,C>0 c 0 , c 1 , C > 0 such that
∫ T < ∣ t ∣ ≤ π σ ∣ E x e i t Z ∣ d t ≤ C e − c 0 T 2 + C e − c 0 h , 1 ≤ T ≤ c 1 M . (343) \int_{T<|t|\le\pi\sigma} \left|\mathbb{E}_x e^{itZ}\right|\,dt \le C e^{-c_0T^2}+C e^{-c_0h},\qquad1\le T\le c_1\sqrt{M}. \tag*{(343)} ∫ T < ∣ t ∣ ≤ π σ E x e i tZ d t ≤ C e − c 0 T 2 + C e − c 0 h , 1 ≤ T ≤ c 1 M . (343)
Away from a fixed critical neighborhood this follows from (115) : the product of the zero and one probabilities of V ( x z ) / V ( x ) V(xz)/V(x) V ( x z ) / V ( x ) , divided by D Ψ ( x ) \mathcal{D}\Psi(x) D Ψ ( x ) , is bounded below. Near zero both numerator and denominator are comparable to x x x . In the critical neighborhood (118) becomes ( 1 + c t 2 / h ) − h / 8 (1+ct^2/h)^{-h/8} ( 1 + c t 2 / h ) − h /8 . For ∣ t ∣ ≤ h |t|\le\sqrt{h} ∣ t ∣ ≤ h this is bounded by e − c ′ t 2 e^{-c't^2} e − c ′ t 2 . Beyond h \sqrt{h} h , split the exponent in half, bound one half at h \sqrt{h} h , and integrate the other against a fixed integrable power. The integral is O ( e − c ′ h ) O(e^{-c'h}) O ( e − c ′ h ) . The outer arc is exponentially small uniformly in ε \varepsilon ε by (121) . These observations prove (E.5).
Fourier inversion, (E.4), (E.5), and the Gaussian Fourier tail prove (E.1). The three contributions are O p ( M − 1 / 8 ) O_p(M^{-1/8}) O p ( M − 1/8 ) , O ( M − c 0 / 80 ) + O ( e − c 0 h ) O(M^{-c_0/80})+O(e^{-c_0h}) O ( M − c 0 /80 ) + O ( e − c 0 h ) , and O ( e − T 2 / 2 ) = O ( M − 1 / 160 ) O(e^{-T^2/2})=O(M^{-1/160}) O ( e − T 2 /2 ) = O ( M − 1/160 ) , so one may take κ p = min ( c 0 , 1 / 2 ) / 80 \kappa_p=\min(c_0,1/2)/80 κ p = min ( c 0 , 1/2 ) /80 . This exponent is not intended to be sharp. □
As a first application, and only as an illustration of Proposition E.1, we make the volume-ratio limits quantitative; the rates actually used in Section 11 are the sharper ones that follow from Theorem 10.1, and (E.6) is not used elsewhere. Put κ = min ( κ 1 , κ 2 , κ 3 ) \kappa=\min(\kappa_1,\kappa_2,\kappa_3) κ = min ( κ 1 , κ 2 , κ 3 ) , so that the exponent serves for each of the marking counts p = 1 , 2 , 3 p=1,2,3 p = 1 , 2 , 3 . At the common saddle (122) , put M = min ( 2 g − 1 , n − 1 ) M=\min(2g-1,n-1) M = min ( 2 g − 1 , n − 1 ) . Repeating the coefficient comparisons in (123) –(128) now gives
B g , n = π 2 2 r ( Y ) ( 1 + O ( M − κ ) ) , C g , n = π 2 ( n − 1 ) g x ( 1 + O ( M − κ ) ) (344) B_{g,n}=\frac{\pi^2}{2r(Y)}\left(1+O(M^{-\kappa})\right),\qquad C_{g,n}=\frac{\pi^2(n-1)}{gx}\left(1+O(M^{-\kappa})\right) \tag*{(344)} B g , n = 2 r ( Y ) π 2 ( 1 + O ( M − κ ) ) , C g , n = g x π 2 ( n − 1 ) ( 1 + O ( M − κ ) ) (344)
when g , n → ∞ g,n\to\infty g , n → ∞ . The center shift in (126) contributes only O ( g − 1 ) O(g^{-1}) O ( g − 1 ) , since its Gaussian argument is O ( g − 1 / 2 ) O(g^{-1/2}) O ( g − 1/2 ) .
The next three subsections prove Theorem 10.1. We first compare the true and the model generating functions on small complex neighborhoods of the real saddle point, and then treat large and small coefficient indices separately.
Local complex comparison
To extract a coefficient by a contour integral we need the marked expansion on complex neighborhoods of the real saddle, not only on the real axis. We transfer a finite-order expansion to such neighborhoods and compare the true and model Fourier integrals.
Choose a fixed finite order K K K , to be enlarged below, and define
F h , [ K ] ( z ) = J g , p ( z ) π ( 1 + ∑ j = 1 K L p , j ( z ) h j ) . (345) F_{h,[K]}(z)=\frac{\mathcal{J}_{g,p}(z)}{\pi}\left(1+\sum_{j=1}^{K}\frac{\mathcal{L}_{p,j}(z)}{h^j}\right). \tag*{(345)} F h , [ K ] ( z ) = π J g , p ( z ) ( 1 + j = 1 ∑ K h j L p , j ( z ) ) . (345)
This is initially a local analytic function near the positive real interval. We do not assume that it is analytic on a whole cusp-weight disk. Theorem 8.2 gives its real relative error O p , K ( h − K − 1 ) O_{p,K}(h^{-K-1}) O p , K ( h − K − 1 ) .
The convergent endpoint expansions and compactness give a fixed c > 0 c>0 c > 0 such that all functions in (E.7) are analytic at z = x exp ( ϖ / σ ) z=x\exp(\varpi/\sigma) z = x exp ( ϖ / σ ) for ∣ w ∣ ≤ c M |w|\le c\sqrt{M} ∣ w ∣ ≤ c M . Near zero this is an ordinary inverse-function neighborhood. Near criticality the radius in z z z is proportional to x c − x x_c-x x c − x , and σ ( x c − x ) / x c ∼ h \sigma(x_c-x)/x_c\sim\sqrt{h} σ ( x c − x ) / x c ∼ h ; the square-root inverse and the analytic δ \delta δ coefficients give the required branch. Each fixed collection of L p , j \mathcal{L}_{p,j} L p , j is bounded on a smaller such neighborhood. The following lemma makes the analytic radius and the resulting derivative bounds explicit, with constants that are geometric in the order of differentiation. It is used for (E.13), for the disk bound following (E.13), for (E.14), for the remainder in (E.17), and for the disk bound in the proof of Theorem 10.2.
Lemma E.2 (Normalized analytic radius). Fix p ≥ 1 p\ge1 p ≥ 1 . There are constants c 0 , c , C > 0 c_0,c,C>0 c 0 , c , C > 0 such that for every 0 < x < x c 0<x<x_c 0 < x < x c , with ε = 1 − x / x c \varepsilon=1-x/x_c ε = 1 − x / x c and
ρ x = c 0 min ( 1 , ε ) , \rho_x=c_0\min(1,\varepsilon), ρ x = c 0 min ( 1 , ε ) ,
the functions
ϕ x ( s ) = S ( x e s ) − S ( x ) − Ψ ( x ) s , ψ x ( s ) = log B p ( x e s ) − log B p ( x ) \phi_x(s)=S(xe^s)-S(x)-\Psi(x)s,\qquad\psi_x(s)=\log B_p(xe^s)-\log B_p(x) ϕ x ( s ) = S ( x e s ) − S ( x ) − Ψ ( x ) s , ψ x ( s ) = log B p ( x e s ) − log B p ( x )
are analytic on ∣ s ∣ ≤ ρ x |s|\le\rho_x ∣ s ∣ ≤ ρ x and satisfy
sup ∣ s ∣ ≤ ρ x ∣ ϕ x ( s ) ∣ + sup ∣ s ∣ ≤ ρ x ∣ ψ x ( s ) ∣ ≤ C x . (346) \sup_{|s|\le\rho_x}|\phi_x(s)|+\sup_{|s|\le\rho_x}|\psi_x(s)|\le Cx. \tag*{(346)} ∣ s ∣ ≤ ρ x sup ∣ ϕ x ( s ) ∣ + ∣ s ∣ ≤ ρ x sup ∣ ψ x ( s ) ∣ ≤ C x . (346)
Consequently, by Cauchy’s estimate,
∣ D m S ( x ) ∣ ≤ C m ! x ρ x − m ( m ≥ 2 ) , ∣ D m log B p ( x ) ∣ ≤ C m ! x ρ x − m ( m ≥ 1 ) . (347) |\mathcal{D}^mS(x)|\le Cm!x\rho_x^{-m}\qquad(m\ge2),\qquad|\mathcal{D}^m\log B_p(x)|\le Cm!x\rho_x^{-m}\qquad(m\ge1). \tag*{(347)} ∣ D m S ( x ) ∣ ≤ C m ! x ρ x − m ( m ≥ 2 ) , ∣ D m log B p ( x ) ∣ ≤ C m ! x ρ x − m ( m ≥ 1 ) . (347)
Moreover, with k = h Ψ ( x ) k=h\Psi(x) k = h Ψ ( x ) , M = min ( h , k ) M=\min(h,k) M = min ( h , k ) and σ 2 = h D Ψ ( x ) \sigma^2=h\mathcal{D}\Psi(x) σ 2 = h D Ψ ( x ) ,
σ ρ x ≥ c M , c σ 2 ≤ h x ρ x − 2 ≤ C σ 2 , (348) \sigma\rho_x\ge c\sqrt{M},\qquad c\sigma^2\le hx\rho_x^{-2}\le C\sigma^2, \tag*{(348)} σ ρ x ≥ c M , c σ 2 ≤ h x ρ x − 2 ≤ C σ 2 , (348)
so that for ∣ w ∣ ≤ σ ρ x |w|\le\sigma\rho_x ∣ w ∣ ≤ σ ρ x
h ∣ ϕ x ( w / σ ) ∣ ≤ C ∣ w ∣ 2 , ∣ ψ x ( w / σ ) ∣ ≤ C ∣ w ∣ , (349) h|\phi_x(w/\sigma)|\le C|w|^2,\qquad|\psi_x(w/\sigma)|\le C|w|, \tag*{(349)} h ∣ ϕ x ( w / σ ) ∣ ≤ C ∣ w ∣ 2 , ∣ ψ x ( w / σ ) ∣ ≤ C ∣ w ∣ , (349)
and, for m ≥ 3 m\ge3 m ≥ 3 and m ≥ 1 m\ge1 m ≥ 1 respectively,
h ∣ D m S ( x ) ∣ σ m ≤ C m ! ( σ ρ x ) 2 − m ≤ C m ! c 2 − m M 1 − m / 2 , \frac{h|\mathcal{D}^mS(x)|}{\sigma^m}\le Cm!(\sigma\rho_x)^{2-m}\le Cm!c^{2-m}M^{1-m/2}, σ m h ∣ D m S ( x ) ∣ ≤ C m ! ( σ ρ x ) 2 − m ≤ C m ! c 2 − m M 1 − m /2 ,
∣ D m log B p ( x ) ∣ σ m ≤ C m ! ( σ ρ x ) − m ≤ C m ! c − m M − m / 2 . (350) \frac{|\mathcal{D}^m\log B_p(x)|}{\sigma^m}\le Cm!(\sigma\rho_x)^{-m}\le Cm!c^{-m}M^{-m/2}. \tag*{(350)} σ m ∣ D m log B p ( x ) ∣ ≤ C m ! ( σ ρ x ) − m ≤ C m ! c − m M − m /2 . (350)
Proof. Near zero, S S S and log B p \log B_p log B p are analytic on a fixed disk ∣ z ∣ ≤ z 0 |z|\le z_0 ∣ z ∣ ≤ z 0 and vanish at z = 0 z=0 z = 0 , so ∣ S ( z ) ∣ + ∣ log B p ( z ) ∣ ≤ C ∣ z ∣ |S(z)|+|\log B_p(z)|\le C|z| ∣ S ( z ) ∣ + ∣ log B p ( z ) ∣ ≤ C ∣ z ∣ there; for x ≤ z 0 / e x\le z_0/e x ≤ z 0 / e and ∣ s ∣ ≤ 1 |s|\le1 ∣ s ∣ ≤ 1 one has ∣ x e s ∣ ≤ e x ≤ z 0 |xe^s|\le ex\le z_0 ∣ x e s ∣ ≤ e x ≤ z 0 , and Ψ ( x ) ≤ C x \Psi(x)\le Cx Ψ ( x ) ≤ C x , which gives (E.8) with ρ x = 1 \rho_x=1 ρ x = 1 . On a compact subinterval of ( 0 , x c ) (0,x_c) ( 0 , x c ) the same holds with a fixed radius, by compactness. Near x c x_c x c , write δ = y 0 − Y \delta=y_0-Y δ = y 0 − Y and use the convergent local factors
x c − X ( Y ) = δ 2 α ( δ ) , Z ( Y ) = δ 5 / 2 ζ ( δ ) , B p ( X ( Y ) ) = δ p / 2 χ p ( δ ) , x_c-X(Y)=\delta^2\alpha(\delta),\qquad Z(Y)=\delta^{5/2}\zeta(\delta),\qquad B_p(X(Y))=\delta^{p/2}\chi_p(\delta), x c − X ( Y ) = δ 2 α ( δ ) , Z ( Y ) = δ 5/2 ζ ( δ ) , B p ( X ( Y )) = δ p /2 χ p ( δ ) ,
with α , ζ , χ p \alpha,\zeta,\chi_p α , ζ , χ p analytic and nonzero at zero; they follow from the analytic endpoint parametrization constructed in the proof of Lemma 3.1(iii) in Appendix B.§ 3.3 , Z = 2 I Z=\sqrt{2I} Z = 2 I and (136) , and are recorded again in Subsection E.7. For z = x e s z=xe^s z = x e s the parameter δ z = δ q \delta_z=\delta_q δ z = δ q solves q 2 α ( δ q ) = α ( δ ) − x ( e s − 1 ) δ − 2 q^2\alpha(\delta_q)=\alpha(\delta)-x(e^s-1)\delta^{-2} q 2 α ( δ q ) = α ( δ ) − x ( e s − 1 ) δ − 2 , since ∂ q [ q 2 α ( δ q ) ] = 2 α ( 0 ) ≠ 0 \partial_q[q^2\alpha(\delta_q)]=2\alpha(0)\ne0 ∂ q [ q 2 α ( δ q )] = 2 α ( 0 ) = 0 at ( q , δ ) = ( 1 , 0 ) (q,\delta)=(1,0) ( q , δ ) = ( 1 , 0 ) , the implicit function theorem gives q q q analytic in δ \delta δ and in τ = x ( e s − 1 ) δ − 2 \tau=x(e^s-1)\delta^{-2} τ = x ( e s − 1 ) δ − 2 for ∣ τ ∣ ≤ c 1 |\tau|\le c_1 ∣ τ ∣ ≤ c 1 , hence for ∣ s ∣ ≤ c δ 2 / x |s|\le c\delta^2/x ∣ s ∣ ≤ c δ 2 / x ; as ε = δ 2 α ( δ ) / x c \varepsilon=\delta^2\alpha(\delta)/x_c ε = δ 2 α ( δ ) / x c , this is a disk ∣ s ∣ ≤ c 0 ε |s|\le c_0\varepsilon ∣ s ∣ ≤ c 0 ε on which q q q stays in a fixed neighborhood of one. There
S ( z ) − S ( x ) = − 5 2 log q − log ζ ( δ q ) ζ ( δ ) , log B p ( z ) − log B p ( x ) = p 2 log q + log χ p ( δ q ) χ p ( δ ) S(z)-S(x)=-\frac{5}{2}\log q-\log\frac{\zeta(\delta q)}{\zeta(\delta)},\qquad
\log B_p(z)-\log B_p(x)=\frac{p}{2}\log q+\log\frac{\chi_p(\delta q)}{\chi_p(\delta)} S ( z ) − S ( x ) = − 2 5 log q − log ζ ( δ ) ζ ( δ q ) , log B p ( z ) − log B p ( x ) = 2 p log q + log χ p ( δ ) χ p ( δ q )
are bounded, and Ψ ( x ) ∣ s ∣ ≤ C ε − 1 ⋅ c 0 ε ≤ C \Psi(x)|s|\le C\varepsilon^{-1}\cdot c_0\varepsilon\le C Ψ ( x ) ∣ s ∣ ≤ C ε − 1 ⋅ c 0 ε ≤ C by (20) . This proves (E.8) near x c x_c x c , where C x ≍ C C_x\asymp C C x ≍ C .
Since d m d s m S ( x e s ) ∣ s = 0 = D m S ( x ) \left.\frac{d^m}{ds^m}S(xe^s)\right|_{s=0}=\mathcal{D}^mS(x) d s m d m S ( x e s ) s = 0 = D m S ( x ) and D S = Ψ \mathcal{D}S=\Psi D S = Ψ , the Taylor coefficients of ϕ x \phi_x ϕ x at s = 0 s=0 s = 0 are D m S ( x ) / m ! \mathcal{D}^mS(x)/m! D m S ( x ) / m ! for m ≥ 2 m\ge2 m ≥ 2 , and those of ψ x \psi_x ψ x are D m log B p ( x ) / m ! \mathcal{D}^m\log B_p(x)/m! D m log B p ( x ) / m ! for m ≥ 1 m\ge1 m ≥ 1 ; Cauchy’s estimate on ∣ s ∣ = ρ x |s|=\rho_x ∣ s ∣ = ρ x gives (E.9). For (E.10), σ 2 = h D Ψ \sigma^2=h\mathcal{D}\Psi σ 2 = h D Ψ with D Ψ ≍ x \mathcal{D}\Psi\asymp x D Ψ ≍ x near zero, ≍ 1 \asymp1 ≍ 1 on compact subintervals and ≍ ε − 2 \asymp\varepsilon^{-2} ≍ ε − 2 near x c x_c x c , by (20) ; hence h x ρ x − 2 ≍ h x ≍ σ 2 hx\rho_x^{-2}\asymp hx\asymp\sigma^2 h x ρ x − 2 ≍ h x ≍ σ 2 near zero and h x ρ x − 2 ≍ h ε − 2 ≍ σ 2 hx\rho_x^{-2}\asymp h\varepsilon^{-2}\asymp\sigma^2 h x ρ x − 2 ≍ h ε − 2 ≍ σ 2 near x c x_c x c , while σ ρ x ≍ h x ≍ h Ψ ( x ) = k ≥ M \sigma\rho_x\asymp\sqrt{hx}\asymp\sqrt{h\Psi(x)}=\sqrt{k}\ge\sqrt{M} σ ρ x ≍ h x ≍ h Ψ ( x ) = k ≥ M near zero and σ ρ x ≍ h ≥ M \sigma\rho_x\asymp\sqrt{h}\ge\sqrt{M} σ ρ x ≍ h ≥ M otherwise. Since ϕ x \phi_x ϕ x vanishes to second order and ψ x \psi_x ψ x to first order at s = 0 s=0 s = 0 , Schwarz’s lemma on ∣ s ∣ ≤ ρ x |s|\le\rho_x ∣ s ∣ ≤ ρ x gives ∣ ϕ x ( s ) ∣ ≤ C x ∣ s ∣ 2 ρ x − 2 |\phi_x(s)|\le Cx|s|^2\rho_x^{-2} ∣ ϕ x ( s ) ∣ ≤ C x ∣ s ∣ 2 ρ x − 2 and ∣ ψ x ( s ) ∣ ≤ C x ∣ s ∣ ρ x − 1 |\psi_x(s)|\le Cx|s|\rho_x^{-1} ∣ ψ x ( s ) ∣ ≤ C x ∣ s ∣ ρ x − 1 ; at s = w / σ s=w/\sigma s = w / σ this is (E.11), by (E.10) and x / ( σ ρ x ) ≤ C x / h ≤ C x/(\sigma\rho_x)\le C\sqrt{x/h}\le C x / ( σ ρ x ) ≤ C x / h ≤ C for h ≥ 1 h\ge1 h ≥ 1 . Finally (E.12) follows from (E.9) and (E.10), since h x ρ x − m σ − m = ( h x ρ x − 2 ) ( σ ρ x ) 2 − m hx\rho_x^{-m}\sigma^{-m}=(hx\rho_x^{-2})(\sigma\rho_x)^{2-m} h x ρ x − m σ − m = ( h x ρ x − 2 ) ( σ ρ x ) 2 − m and x ρ x − m σ − m ≤ x c ( σ ρ x ) − m x\rho_x^{-m}\sigma^{-m}\le x_c(\sigma\rho_x)^{-m} x ρ x − m σ − m ≤ x c ( σ ρ x ) − m .
The model derivatives obey
h D m S σ m = O m ( M 1 − m / 2 ) ( m ≥ 3 ) , D m log B p σ m = O p , m ( M − m / 2 ) ( m ≥ 1 ) . (351) \frac{h\mathcal{D}^mS}{\sigma^m}=O_m(M^{1-m/2})\qquad(m\ge3),\qquad
\frac{\mathcal{D}^m\log B_p}{\sigma^m}=O_{p,m}(M^{-m/2})\qquad(m\ge1). \tag*{(351)} σ m h D m S = O m ( M 1 − m /2 ) ( m ≥ 3 ) , σ m D m log B p = O p , m ( M − m /2 ) ( m ≥ 1 ) . (351)
with constants of the form C m ! c − m C\,m!c^{-m} C m ! c − m , by (E.12); after division by the factor m ! m! m ! of the Taylor coefficients this dependence on m m m is geometric, which is what the disk bounds below require. At zero use D m S = 2 x + O m ( x 2 ) \mathcal{D}^m S=2x+O_m(x^2) D m S = 2 x + O m ( x 2 ) and D m log B p = O p , m ( x ) \mathcal{D}^m\log B_p=O_{p,m}(x) D m log B p = O p , m ( x ) . At criticality, with ε = 1 − x / x c \varepsilon=1-x/x_c ε = 1 − x / x c , both numerator derivatives before their h h h factor are O ( ε − m ) O(\varepsilon^{-m}) O ( ε − m ) , whereas σ ≍ h / ε \sigma\asymp\sqrt{h/\varepsilon} σ ≍ h / ε . These bounds also hold locally in the complex neighborhoods by the same convergent expansions.
Put
L h ( w ) = e − k w / σ F h ( x e w / σ ) F h ( x ) , L h , [ K ] ( w ) = e − k w / σ F h , [ K ] ( x e w / σ ) F h , [ K ] ( x ) . \mathcal{L}_h(w)=e^{-kw/\sigma}\frac{F_h(xe^{w/\sigma})}{F_h(x)},\qquad\mathcal{L}_{h,[K]}(w)=e^{-kw/\sigma}\frac{F_{h,[K]}(xe^{w/\sigma})}{F_{h,[K]}(x)}. L h ( w ) = e − k w / σ F h ( x ) F h ( x e w / σ ) , L h , [ K ] ( w ) = e − k w / σ F h , [ K ] ( x ) F h , [ K ] ( x e w / σ ) .
Since k = h Ψ ( x ) k=h\Psi(x) k = h Ψ ( x ) , the logarithm of the finite analytic model is
log L h , [ K ] ( w ) = h ϕ x ( w / σ ) + ψ x ( w / σ ) + log P K ( x e w / σ ) P K ( x ) , P K = 1 + ∑ j = 1 K L p , j h − j , \log\mathcal{L}_{h,[K]}(w)=h\phi_x(w/\sigma)+\psi_x(w/\sigma)+\log\frac{P_K(xe^{w/\sigma})}{P_K(x)},\qquad P_K=1+\sum_{j=1}^{K}\mathcal{L}_{p,j}h^{-j}, log L h , [ K ] ( w ) = h ϕ x ( w / σ ) + ψ x ( w / σ ) + log P K ( x ) P K ( x e w / σ ) , P K = 1 + j = 1 ∑ K L p , j h − j ,
in the notation of Lemma E.2 ; the last ratio lies between 1 / 2 1/2 1/2 and 2 2 2 in modulus on the chosen neighborhoods for large h h h . Hence (E.11) gives the disk bound C p , K e C ∣ w ∣ 2 + C ∣ w ∣ C_{p,K}e^{C|w|^2+C|w|} C p , K e C ∣ w ∣ 2 + C ∣ w ∣ , for ∣ w ∣ ≤ c M |w|\leq c\sqrt{M} ∣ w ∣ ≤ c M , for the centered, normalized finite analytic model, with C C C independent of K K K ; the centered, normalized true generating function obeys the same bound, because positivity bounds its modulus by its value at R e w \mathop{\rm Re}w Re w , where the real expansion applies. No error term is differentiated. The explicit model also satisfies
∣ J g , p ( x e i t / σ ) J g , p ( x ) ∣ ≤ C p e − c ′ t 2 , ∣ t ∣ ≤ c ∗ M . (352) \left|\frac{\mathcal{J}_{g,p}(xe^{it/\sigma})}{\mathcal{J}_{g,p}(x)}\right|\leq C_pe^{-c't^2},\qquad|t|\leq c_*\sqrt{M}. \tag*{(352)} J g , p ( x ) J g , p ( x e i t / σ ) ≤ C p e − c ′ t 2 , ∣ t ∣ ≤ c ∗ M . (352)
for a suitable c ∗ > 0 c_*>0 c ∗ > 0 and all sufficiently large M M M . Indeed the real action is
R e h ϕ x ( i t / σ ) = − t 2 2 + R e ∑ m ≥ 4 h D m S ( i t ) m m ! σ m . \mathop{\rm Re}h\phi_x(it/\sigma)=-\frac{t^2}{2}+\mathop{\rm Re}\sum_{m\geq4}\frac{h\mathcal{D}^mS(it)^m}{m!\sigma^m}. Re h ϕ x ( i t / σ ) = − 2 t 2 + Re m ≥ 4 ∑ m ! σ m h D m S ( i t ) m .
the odd terms being purely imaginary, and by (E.12) the sum is at most C M ∑ m ≥ 4 ( ∣ t ∣ / ( c M ) ) m ≤ C 1 t 4 / M CM\sum_{m\geq4}(|t|/(c\sqrt{M}))^m\leq C_1t^4/M C M ∑ m ≥ 4 ( ∣ t ∣/ ( c M ) ) m ≤ C 1 t 4 / M for ∣ t ∣ ≤ c M / 2 |t|\leq c\sqrt{M}/2 ∣ t ∣ ≤ c M /2 , with C 1 = 2 C / c 4 C_1=2C/c^4 C 1 = 2 C / c 4 ; the real amplitude change R e ψ x ( i t / σ ) \mathop{\rm Re}\psi_x(it/\sigma) Re ψ x ( i t / σ ) is at most C ∑ m ≥ 2 ( ∣ t ∣ / ( c M ) ) m ≤ C 1 t 2 / M C\sum_{m\geq2}(|t|/(c\sqrt{M}))^m\leq C_1t^2/M C ∑ m ≥ 2 ( ∣ t ∣/ ( c M ) ) m ≤ C 1 t 2 / M on the same range. For ∣ t ∣ ≤ c ∗ M |t|\leq c_*\sqrt{M} ∣ t ∣ ≤ c ∗ M , with c ∗ ≤ c / 2 c_*\leq c/2 c ∗ ≤ c /2 and C 1 c ∗ 2 ≤ 1 / 8 C_1c_*^2\leq1/8 C 1 c ∗ 2 ≤ 1/8 , the former is at most t 2 / 8 t^2/8 t 2 /8 , and for M ≥ 8 C 1 M\geq8C_1 M ≥ 8 C 1 so is the latter; this gives (E.14) with c ′ = 1 / 4 c'=1/4 c ′ = 1/4 .
Take T 2 = A 0 log h T^2=A_0\log h T 2 = A 0 log h , with A 0 A_0 A 0 large enough that the Fourier tail (E.5) of the true distribution, the polynomial-Gaussian tails below, and the sparse-model Fourier tail below are O ( h − 3 ) O(h^{-3}) O ( h − 3 ) . These tail constants depend on the fixed action S S S and amplitude B p B_p B p , not on the finite truncation order. The central truncation order will first be fixed below; then choose C 0 C_0 C 0 so that M ≥ C 0 log h M\geq C_0\log h M ≥ C 0 log h implies both 4 T ≤ c M 4T\leq c\sqrt{M} 4 T ≤ c M for its analytic neighborhood and T ≤ min ( c 1 , c ∗ ) M T\leq\min(c_1,c_*)\sqrt{M} T ≤ min ( c 1 , c ∗ ) M , the ranges of (E.5) and (E.14).
Lemma E.3 (Local comparison). There are a fixed order K c K_c K c and a constant C 0 C_0 C 0 such that, uniformly for M ≥ C 0 log h M\geq C_0\log h M ≥ C 0 log h ,
σ P x ( N = k ) = 1 2 π ∫ − T T L h , [ K c ] ( i t ) d t + O ( h − 2 ) , T 2 = A 0 log h . \sigma\mathbb{P}_x(N=k)=\frac{1}{2\pi}\int_{-T}^{T}\mathcal{L}_{h,[K_c]}(it)\,dt+O(h^{-2}),\qquad T^2=A_0\log h. σ P x ( N = k ) = 2 π 1 ∫ − T T L h , [ K c ] ( i t ) d t + O ( h − 2 ) , T 2 = A 0 log h .
Proof. Here k ≥ 1 k\geq1 k ≥ 1 is an integer and x x x is its saddle, h Ψ ( x ) = k h\Psi(x)=k h Ψ ( x ) = k , so L h ( i t ) \mathcal{L}_h(it) L h ( i t ) is the characteristic function of ( N − k ) / σ (N-k)/\sigma ( N − k ) / σ . Fourier inversion gives P x ( N = k ) = ( 2 π ) − 1 ∫ − π σ π σ L h ( i t ) d t \mathbb{P}_x(N=k)=(2\pi)^{-1}\int_{-\pi\sigma}^{\pi\sigma}\mathcal{L}_h(it)\,dt P x ( N = k ) = ( 2 π ) − 1 ∫ − π σ π σ L h ( i t ) d t , and by (E.5) and the choice of A 0 A_0 A 0 the part T < ∣ t ∣ ≤ π σ T<|t|\leq\pi\sigma T < ∣ t ∣ ≤ π σ is O ( h − 3 ) O(h^{-3}) O ( h − 3 ) . It remains to compare the two integrands on [ − T , T ] [-T,T] [ − T , T ] . The difference between the two centered, normalized functions has real-part bound C h − K − 1 e C T 2 Ch^{-K-1}e^{CT^2} C h − K − 1 e C T 2 on [ − T , T ] [-T,T] [ − T , T ] and disk bound C e C T 2 Ce^{CT^2} C e C T 2 . The harmonic-measure argument of (E.4) then gives
sup ∣ t ∣ ≤ T ∣ L h ( i t ) − L h , [ K ] ( i t ) ∣ ≤ C p , K h − ( K + 1 ) / 2 + C ′ ′ A 0 . (353) \sup_{|t|\leq T}\left|\mathcal{L}_h(it)-\mathcal{L}_{h,[K]}(it)\right|\leq C_{p,K}h^{-(K+1)/2+C''A_0}. \tag*{(353)} ∣ t ∣ ≤ T sup L h ( i t ) − L h , [ K ] ( i t ) ≤ C p , K h − ( K + 1 ) /2 + C ′′ A 0 . (353)
The exponential constants here are independent of K K K : for each fixed K K K , the modulus of its finite correction factor lies between 1 / 2 1/2 1/2 and 2 on the chosen neighborhoods for all sufficiently large h h h . Only the threshold and multiplicative constants depend on K K K . Fix the central order K = K c K=K_c K = K c sufficiently large that the integral of (E.15) is O ( h − 2 ) O(h^{-2}) O ( h − 2 ) , and then fix C 0 C_0 C 0 . Thus σ P x ( N = k ) \sigma P_x(N=k) σ P x ( N = k ) equals the centered finite-model integral on [ − T , T ] [-T,T] [ − T , T ] , divided by 2 π 2\pi 2 π , up to O ( h − 2 ) O(h^{-2}) O ( h − 2 ) . The normalization F h , [ K ] ( x ) / F h ( x ) = 1 + O ( h − K − 1 ) F_{h,[K]}(x)/F_h(x)=1+O(h^{-K-1}) F h , [ K ] ( x ) / F h ( x ) = 1 + O ( h − K − 1 ) is retained when recovering coefficients. All orders chosen here are fixed independently of g , x , k g,x,k g , x , k . □
Cancellation of the Poisson singularity
For indices k ≥ h 1 / 4 k\ge h^{1/4} k ≥ h 1/4 , so that M = min ( h , k ) ≥ h 1 / 4 M=\min(h,k)\ge h^{1/4} M = min ( h , k ) ≥ h 1/4 , we compute the model coefficient by the saddle-point method. At order h − j h^{-j} h − j its Gaussian coefficient has a leading pole s j Ψ − j s_j\Psi^{-j} s j Ψ − j as the cusp weight tends to zero, matching the corresponding Poisson coefficient. Division by the Poisson factor removes these leading poles; the weaker residual bounds in (E.19) suffice for the error in (137) .
Lemma E.4 (Large indices). The estimate (137) holds uniformly for integers k ≥ h 1 / 4 k\ge h^{1/4} k ≥ h 1/4 .
Proof. Here M = min ( h , k ) ≥ h 1 / 4 M=\min(h,k)\ge h^{1/4} M = min ( h , k ) ≥ h 1/4 . Put b = log B p b=\log B_p b = log B p . For a formal variable u u u define the model polynomials through the exponential
exp { ∑ m ≥ 3 u m − 2 D m − 1 Ψ m ! ν m / 2 w m + ∑ m ≥ 1 u m D m b m ! ν m / 2 w m } (354) \exp\left\{\sum_{m\ge3}\frac{u^{m-2}\mathcal{D}^{m-1}\Psi}{m!\nu^{m/2}}w^m+\sum_{m\ge1}\frac{u^m\mathcal{D}^m b}{m!\nu^{m/2}}w^m\right\} \tag*{(354)} exp { m ≥ 3 ∑ m ! ν m /2 u m − 2 D m − 1 Ψ w m + m ≥ 1 ∑ m ! ν m /2 u m D m b w m } (354)
Let E j ( x ) E_j(x) E j ( x ) be the Gaussian Fourier integral of its u 2 j u^{2j} u 2 j coefficient: replace w 2 a w^{2a} w 2 a by ( − 1 ) a ( 2 a − 1 ) ! ! (-1)^a(2a-1)!! ( − 1 ) a ( 2 a − 1 )!! and odd powers by zero. Each coefficient uses only finitely many derivatives. By Lemma E.3 it suffices to evaluate the central model integral. Expansion under the Gaussian bound (E.14) gives
1 2 π ∫ − T T e − i k t / σ J g , p ( x e i t / σ ) J g , p ( x ) d t = 1 2 π ( 1 + ∑ j = 1 3 E j ( x ) h j + O p ( M − 4 + h − 2 ) ) . (355) \frac{1}{2\pi}\int_{-T}^{T}e^{-ikt/\sigma}\frac{\mathcal{J}_{g,p}(xe^{it/\sigma})}{\mathcal{J}_{g,p}(x)}\,dt
=
\frac{1}{\sqrt{2\pi}}\left(1+\sum_{j=1}^{3}\frac{E_j(x)}{h^j}+O_p(M^{-4}+h^{-2})\right). \tag*{(355)} 2 π 1 ∫ − T T e − ik t / σ J g , p ( x ) J g , p ( x e i t / σ ) d t = 2 π 1 ( 1 + j = 1 ∑ 3 h j E j ( x ) + O p ( M − 4 + h − 2 ) ) . (355)
To justify the stated remainder, retain weights through seven in (E.16); a term of weight q q q has parity q q q in w w w , so odd weights integrate to zero. By (E.12), a factor of weight m − 2 m-2 m − 2 in the action sum of (E.16) is at most C c 2 − m M 1 − m / 2 ∣ t ∣ m Cc^{2-m}M^{1-m/2}|t|^m C c 2 − m M 1 − m /2 ∣ t ∣ m and a factor of weight m m m in the amplitude sum is at most C c − m M − m / 2 ∣ t ∣ m Cc^{-m}M^{-m/2}|t|^m C c − m M − m /2 ∣ t ∣ m , geometrically in m m m ; so the terms of total weight q q q are at most C q M − q / 2 ( 1 + ∣ t ∣ ) 3 q C^qM^{-q/2}(1+|t|)^{3q} C q M − q /2 ( 1 + ∣ t ∣ ) 3 q , and their sum over q ≥ 8 q\ge8 q ≥ 8 is at most 2 ( C ( 1 + ∣ t ∣ ) 3 / M ) 8 2(C(1+|t|)^3/\sqrt{M})^8 2 ( C ( 1 + ∣ t ∣ ) 3 / M ) 8 as long as C ( 1 + ∣ t ∣ ) 3 ≤ M / 2 C(1+|t|)^3\le\sqrt{M}/2 C ( 1 + ∣ t ∣ ) 3 ≤ M /2 . Thus the remaining terms are bounded by M − 4 M^{-4} M − 4 times a fixed polynomial in ∣ t ∣ |t| ∣ t ∣ times e − c ′ t 2 e^{-c't^2} e − c ′ t 2 . Their integral is O ( M − 4 ) O(M^{-4}) O ( M − 4 ) . The cutoff is O ( log h ) O(\sqrt{\log h}) O ( log h ) , which is smaller than any fixed positive power of M M M in this range; the Taylor remainder and exponential remainder bounds therefore apply uniformly. The removed polynomial-Gaussian tails are O ( h − 2 ) O(h^{-2}) O ( h − 2 ) by the initial choice of A 0 A_0 A 0 . Replacing the finite model by J \mathcal{J} J in its central integral costs O p ( h − 1 ) O_p(h^{-1}) O p ( h − 1 ) , by boundedness of its corrections and (E.14).
The coefficients E j E_j E j are bounded at criticality by (E.13). Near zero, v = Ψ + O ( Ψ 2 ) v=\Psi+O(\Psi^2) v = Ψ + O ( Ψ 2 ) and D Ψ = Ψ + O m ( Ψ 2 ) \mathcal{D}\Psi=\Psi+O_m(\Psi^2) D Ψ = Ψ + O m ( Ψ 2 ) . In a monomial of weight 2 j 2j 2 j with a a a action factors and b ∗ b_* b ∗ amplitude factors, the total degree in w w w is 2 j + 2 a 2j+2a 2 j + 2 a . Its leading power in Ψ \Psi Ψ is therefore Ψ b ∗ − j \Psi^{b_*-j} Ψ b ∗ − j . Every amplitude factor, or higher term in an action coefficient, improves the pole by at least one. Consequently
E j ( x ) = s j Ψ ( x ) − j + O p , j ( Ψ ( x ) 1 − j ) ( x ↓ 0 ) , (356) E_j(x)=s_j\Psi(x)^{-j}+O_{p,j}(\Psi(x)^{1-j})\qquad(x\downarrow0), \tag*{(356)} E j ( x ) = s j Ψ ( x ) − j + O p , j ( Ψ ( x ) 1 − j ) ( x ↓ 0 ) , (356)
where s j s_j s j is the coefficient of k − j k^{-j} k − j in 2 π k p k \sqrt{2\pi k}\,p_k 2 π k p k . Indeed the leading action terms in (E.16) are exactly those of a centered Poisson law: for the exact Poisson generating function e h z e^{hz} e h z one has Ψ ( z ) = z \Psi(z)=z Ψ ( z ) = z , v = Ψ v=\Psi v = Ψ and b = 0 b=0 b = 0 , the coefficient is h k / k ! h^k/k! h k / k ! , each E j E_j E j equals e j Ψ − j e_j\Psi^{-j} e j Ψ − j by homogeneity, and (E.17), carried to any fixed order, reads 2 π k p k = 1 + ∑ j ≤ J e j k − j + O ( k − J − 1 ) \sqrt{2\pi k}\,p_k=1+\sum_{j\le J}e_jk^{-j}+O(k^{-J-1}) 2 π k p k = 1 + ∑ j ≤ J e j k − j + O ( k − J − 1 ) , so that e j = s j e_j=s_j e j = s j . Parity makes the remaining Laurent powers integral. For example, direct Gaussian integration gives
E 1 = − D 2 b + ( D b ) 2 2 v + ( D b ) D 2 Ψ 2 v 2 + D 3 Ψ 8 v 2 − 5 ( D 2 Ψ ) 2 24 v 3 . E_1=-\frac{\mathcal{D}^2b+(\mathcal{D}b)^2}{2v}+\frac{(\mathcal{D}b)\mathcal{D}^2\Psi}{2v^2}+\frac{\mathcal{D}^3\Psi}{8v^2}-\frac{5(\mathcal{D}^2\Psi)^2}{24v^3}. E 1 = − 2 v D 2 b + ( D b ) 2 + 2 v 2 ( D b ) D 2 Ψ + 8 v 2 D 3 Ψ − 24 v 3 5 ( D 2 Ψ ) 2 .
The amplitude terms are needed because k = h Ψ k=h\Psi k = h Ψ is a model center, not an assertion about the exact finite-genus mean.
Divide (E.17) by the central Poisson expansion 2 π k p k = 1 + ∑ j = 1 3 s j / ( h Ψ ) j + O ( k − 4 ) \sqrt{2\pi k}\,p_k=1+\sum_{j=1}^{3}s_j/(h\Psi)^j+O(k^{-4}) 2 π k p k = 1 + ∑ j = 1 3 s j / ( h Ψ ) j + O ( k − 4 ) . The principal pole cancels at each order. The quotient coefficients R j R_j R j satisfy
∣ R j ( x ) ∣ ≤ C p , j ( 1 + Ψ ( x ) 1 − j ) , j = 1 , 2 , 3. (357) \lvert R_j(x)\rvert\le C_{p,j}(1+\Psi(x)^{1-j}),\qquad j=1,2,3. \tag*{(357)} ∣ R j ( x )∣ ≤ C p , j ( 1 + Ψ ( x ) 1 − j ) , j = 1 , 2 , 3. (357)
This follows inductively: a product of a Poisson pole of order i i i and a previous quotient pole of order at most j − i − 1 j-i-1 j − i − 1 has order at most j − 1 j-1 j − 1 . The quotient is bounded at the critical end as well. Since k = h Ψ ≥ 1 k=h\Psi\ge1 k = h Ψ ≥ 1 , every h − j R j h^{-j}R_j h − j R j is O p ( h − 1 ) O_p(h^{-1}) O p ( h − 1 ) . Also M − 4 ≤ h − 1 M^{-4}\le h^{-1} M − 4 ≤ h − 1 . Combining this with (E.15) and the bounded finite amplitude corrections proves (137) in this range. □
Small coefficient indices
For k ≤ h 1 / 4 k\le h^{1/4} k ≤ h 1/4 the preceding finite-order Gaussian estimate does not give the required uniform error O ( h − 1 ) O(h^{-1}) O ( h − 1 ) . We expand the model coefficient directly around the Poisson coefficient of e s 1 h z e^{s_1hz} e s 1 h z .
Lemma E.5 (Small indices). The estimate (137) holds uniformly for integers 0 ≤ k ≤ h 1 / 4 0\le k\le h^{1/4} 0 ≤ k ≤ h 1/4 .
Proof. Near zero write S ( z ) = s 1 z + s 2 z 2 + O ( z 3 ) S(z)=s_1z+s_2z^2+O(z^3) S ( z ) = s 1 z + s 2 z 2 + O ( z 3 ) , s 1 = 2 s_1=2 s 1 = 2 , and B p ( z ) = 1 + d 1 z + O ( z 2 ) B_p(z)=1+d_1z+O(z^2) B p ( z ) = 1 + d 1 z + O ( z 2 ) . The full model (E.7) is analytic on a fixed neighborhood of zero. For k ≤ h 1 / 4 k\le h^{1/4} k ≤ h 1/4 , coefficient majorants give
k ! [ z k ] F h , [ K ] ( z ) ( J g , p ( 0 ) / π ) ( s 1 h ) k = 1 + ( s 2 / s 1 2 ) k ( k − 1 ) + ( d 1 / s 1 ) k + L p , 1 ( 0 ) h + O p , K ( ( k + 1 ) 4 h 2 ) . (358) \frac{k![z^k]F_{h,[K]}(z)}{(\mathcal{J}_{g,p}(0)/\pi)(s_1h)^k}
=1+\frac{(s_2/s_1^2)k(k-1)+(d_1/s_1)k+\mathcal{L}_{p,1}(0)}{h}+O_{p,K}\left(\frac{(k+1)^4}{h^2}\right). \tag*{(358)} ( J g , p ( 0 ) / π ) ( s 1 h ) k k ! [ z k ] F h , [ K ] ( z ) = 1 + h ( s 2 / s 1 2 ) k ( k − 1 ) + ( d 1 / s 1 ) k + L p , 1 ( 0 ) + O p , K ( h 2 ( k + 1 ) 4 ) . (358)
Factor e s 1 h z e^{s_1hz} e s 1 h z from the model. Its retained multiplier is 1 + h s 2 z 2 + d 1 z + L p , 1 ( 0 ) / h 1+hs_2z^2+d_1z+\mathcal{L}_{p,1}(0)/h 1 + h s 2 z 2 + d 1 z + L p , 1 ( 0 ) / h . Analytic absolute-coefficient majorants bound the omitted terms by convergent products starting with h z 3 hz^3 h z 3 , h 2 z 4 h^2z^4 h 2 z 4 , z 2 z^2 z 2 , z / h z/h z / h , h − 2 h^{-2} h − 2 . In convolution with the exponential, a degree-a a a term contributes the factor ( k ) a / ( s 1 h ) a (k)_a/(s_1h)^a ( k ) a / ( s 1 h ) a , bounded by ( k / ( s 1 h ) ) a (k/(s_1h))^a ( k / ( s 1 h ) ) a . Evaluation of the majorants at that last argument gives the remainder in (E.20); their exponent is O ( k 2 / h ) = o ( 1 ) O(k^2/h)=o(1) O ( k 2 / h ) = o ( 1 ) . This avoids a spurious k \sqrt{k} k loss from a supremum Cauchy estimate.
At the saddle, direct expansion gives the same polynomial without L p , 1 ( 0 ) \mathcal{L}_{p,1}(0) L p , 1 ( 0 ) for the right-hand side of (137) , relative to this Poisson baseline. Indeed, h S − k + k log ( k / ( s 1 h x ) ) hS-k+k\log(k/(s_1hx)) h S − k + k log ( k / ( s 1 h x )) contributes s 2 k 2 / ( s 1 2 h ) s_2k^2/(s_1^2h) s 2 k 2 / ( s 1 2 h ) , log B p \log B_p log B p contributes d 1 k / ( s 1 h ) d_1k/(s_1h) d 1 k / ( s 1 h ) , and 1 2 log ( k / ( h v ) ) \frac{1}{2}\log(k/(hv)) 2 1 log ( k / ( h v )) contributes − s 2 k / ( s 1 2 h ) -s_2k/(s_1^2h) − s 2 k / ( s 1 2 h ) . Their error is O p ( ( k + 1 ) 4 / h 2 ) O_p((k+1)^4/h^2) O p (( k + 1 ) 4 / h 2 ) . Thus the model coefficient divided by the proposed saddle expression is 1 + O p ( h − 1 ) 1+O_p(h^{-1}) 1 + O p ( h − 1 ) for k ≤ h 1 / 4 k\le h^{1/4} k ≤ h 1/4 .
For C 0 log h ≤ k ≤ h 1 / 4 C_0\log h\le k\le h^{1/4} C 0 log h ≤ k ≤ h 1/4 , comparison (E.15) applies. Here the entire circle ∣ z ∣ = x \lvert z\rvert=x ∣ z ∣ = x is inside the fixed zero-neighborhood. The convergent series for S S S and ∣ 1 − cos ( m θ ) ∣ ≤ m 2 ( 1 − cos θ ) \lvert1-\cos(m\theta)\rvert\le m^2(1-\cos\theta) ∣ 1 − cos ( m θ )∣ ≤ m 2 ( 1 − cos θ ) give
ℜ ( S ( x e i θ ) − S ( x ) ) ≤ − c x ( 1 − cos θ ) , ∣ θ ∣ ≤ π . \Re(S(xe^{i\theta})-S(x))\le-cx(1-\cos\theta),\qquad\lvert\theta\rvert\le\pi. ℜ ( S ( x e i θ ) − S ( x )) ≤ − c x ( 1 − cos θ ) , ∣ θ ∣ ≤ π .
The model amplitude is bounded on this circle, and h x ≍ k hx\asymp k h x ≍ k , so its standardized Fourier tail beyond T T T is O ( h − 2 ) O(h^{-2}) O ( h − 2 ) by the initial choice of A 0 A_0 A 0 . Complete the central model integral to its full coefficient integral and apply (E.20). This proves (137) in the middle sparse range. The scaled model central mass is bounded below by a positive constant here, by (E.20) and the saddle comparison; therefore the additive central-integral errors remain relative errors.
For 0 ≤ k ≤ C 0 log h 0\le k\le C_0\log h 0 ≤ k ≤ C 0 log h use a different contour and, if needed, a different fixed truncation order K = K ℓ K=K_\ell K = K ℓ , chosen after C 0 C_0 C 0 . Put r k = ( k + 1 ) / ( 2 h ) r_k=(k+1)/(2h) r k = ( k + 1 ) / ( 2 h ) and consider ( F h ( r k z ) − F h , [ K ] ( r k z ) ) / ( J g , p ( 0 ) / π ) (F_h(r_kz)-F_{h,[K]}(r_kz))/(\mathcal{J}_{g,p}(0)/\pi) ( F h ( r k z ) − F h , [ K ] ( r k z )) / ( J g , p ( 0 ) / π ) in the disk ∣ z − 1 ∣ < 16 \lvert z-1\rvert<16 ∣ z − 1 ∣ < 16 slit along [ 0 , 2 ] [0,2] [ 0 , 2 ] . Both functions are analytic there for all large h h h ; only a fixed neighborhood of zero is used. Positivity, the real expansion, and the explicit model give disk bound C e C ′ ( k + 1 ) Ce^{C'(k+1)} C e C ′ ( k + 1 ) and slit bound C p , K h − K − 1 e C ′ ( k + 1 ) C_{p,K}h^{-K-1}e^{C'(k+1)} C p , K h − K − 1 e C ′ ( k + 1 ) . On ∣ z ∣ = 1 \lvert z\rvert=1 ∣ z ∣ = 1 the slit harmonic measure is greater than 1/2: translate w = z − 1 w = z - 1 w = z − 1 , so ∣ w ∣ ≤ 2 \lvert w\rvert\le2 ∣ w ∣ ≤ 2 , and compare the Green function in the proof of (E.4) on radii two and sixteen. Its lower bound is
1 − log ( 2 + 5 ) log ( 16 + 255 ) > 1 2 . 1 - \frac{\log(2+\sqrt{5})}{\log(16+\sqrt{255})} > \frac{1}{2}. 1 − log ( 16 + 255 ) log ( 2 + 5 ) > 2 1 .
Thus the unit-circle error is at most C p , K h − ( K + 1 ) / 2 e C ′ ′ ( k + 1 ) = O ( h − 2 ) C_{p,K}h^{-(K+1)/2}e^{C''(k+1)} = O(h^{-2}) C p , K h − ( K + 1 ) /2 e C ′′ ( k + 1 ) = O ( h − 2 ) upon fixing K K K sufficiently large. Cauchy extraction changes the coefficient, relative to the baseline of (E.20), by at most
C h − 2 k ! ( 2 h r k ) k = C h − 2 k ! ( k + 1 ) k ≤ C h − 2 . Ch^{-2}\frac{k!}{(2hr_k)^k}=Ch^{-2}\frac{k!}{(k+1)^k}\le Ch^{-2}. C h − 2 ( 2 h r k ) k k ! = C h − 2 ( k + 1 ) k k ! ≤ C h − 2 .
Equation (E.20) and its saddle comparison finish this range, including k = 0 k=0 k = 0 . The two finite orders K c , K ℓ K_c,K_\ell K c , K ℓ are fixed before h h h tends to infinity, and each is used only in its own range. Enlarging the low-mean order does not change the already fixed central analytic neighborhood or C 0 C_0 C 0 . □
Proof of Theorem 10.1 . Lemmas E.4 and E.5 cover all k ≥ 0 k\ge0 k ≥ 0 ; taking common constants and a common threshold h p h_p h p proves the uniform assertion. □
The analytic phase change at zero
To obtain an expansion that is uniform down to k = 0 k=0 k = 0 , we change variables so that the phase becomes exactly that of a Poisson coefficient, following [178]. This subsection and the next supply the two lemmas used in the small-density part of the proof of Theorem 10.2 in Subsection E.7: Lemma E.6 constructs the change of variables and records the resulting coefficient identity (E.22), and Lemma E.7 expands the transformed amplitude with a remainder that is uniform in the index k k k and factorial in the order. Only the local analyticity of the action S S S and the amplitude near zero is used.
Lemma E.6 (Analytic phase change). Let S S S be analytic near zero with S ( 0 ) = 0 S(0)=0 S ( 0 ) = 0 and S ′ ( 0 ) = s > 0 S'(0)=s>0 S ′ ( 0 ) = s > 0 , and let x = x ( ν ) x=x(\nu) x = x ( ν ) be the solution of x S ′ ( x ) = ν xS'(x)=\nu x S ′ ( x ) = ν near zero, so that x = ν / s + O ( ν 2 ) x=\nu/s+O(\nu^2) x = ν / s + O ( ν 2 ) . There is a jointly analytic map z = Z ( u , ν ) z=Z(u,\nu) z = Z ( u , ν ) on a fixed polydisk, with Z ( 0 , ν ) = 0 Z(0,\nu)=0 Z ( 0 , ν ) = 0 , such that
S ( z ) − ν log z = u − ν log u + H ( ν ) , H ( ν ) = S ( x ) − ν + ν log ( ν / x ) , Z ( ν , ν ) = x . (359) S(z)-\nu\log z=u-\nu\log u+H(\nu),\qquad H(\nu)=S(x)-\nu+\nu\log(\nu/x),\qquad Z(\nu,\nu)=x. \tag*{(359)} S ( z ) − ν log z = u − ν log u + H ( ν ) , H ( ν ) = S ( x ) − ν + ν log ( ν / x ) , Z ( ν , ν ) = x . (359) If H H H is analytic near zero and f ( u , ν ) = H ( Z ( u , ν ) ) u Z u ( u , ν ) / Z ( u , ν ) f(u,\nu)=H(Z(u,\nu))uZ_u(u,\nu)/Z(u,\nu) f ( u , ν ) = H ( Z ( u , ν )) u Z u ( u , ν ) / Z ( u , ν ) , then f f f is jointly analytic, including at u = 0 u=0 u = 0 , and for all sufficiently small ν = k / h \nu=k/h ν = k / h ,
[ z k ] e h S ( z ) H ( z ) = e h H ( ν ) [ u k ] e h u f ( u , ν ) . (360) [z^k]e^{hS(z)}H(z)=e^{hH(\nu)}[u^k]e^{hu}f(u,\nu). \tag*{(360)} [ z k ] e h S ( z ) H ( z ) = e h H ( ν ) [ u k ] e h u f ( u , ν ) . (360)
Proof. The logarithmic terms in (E.21) are to be read as − ν log z + ν log u = − ν log ( Z / u ) -\nu\log z+\nu\log u=-\nu\log(Z/u) − ν log z + ν log u = − ν log ( Z / u ) : the quotient Z ( u , ν ) / u Z(u,\nu)/u Z ( u , ν ) / u constructed below is analytic and nonvanishing on the polydisk, so log ( Z / u ) \log(Z/u) log ( Z / u ) is a single-valued analytic function there, and no branch of log z \log z log z or of log u \log u log u separately is needed. We construct Z Z Z directly. Put c = S ′ ( x ) c=S'(x) c = S ′ ( x ) at the saddle x = x ( ν ) x=x(\nu) x = x ( ν ) , x S ′ ( x ) = ν xS'(x)=\nu x S ′ ( x ) = ν ; thus c = ν / x c=\nu/x c = ν / x for ν ≠ 0 \nu\ne0 ν = 0 , the function c c c is analytic near ν = 0 \nu=0 ν = 0 with c ( 0 ) = S ′ ( 0 ) c(0)=S'(0) c ( 0 ) = S ′ ( 0 ) , and H ( ν ) = S ( x ) − ν + ν log c H(\nu)=S(x)-\nu+\nu\log c H ( ν ) = S ( x ) − ν + ν log c is analytic with H ( 0 ) = 0 H(0)=0 H ( 0 ) = 0 . Put d = z − x d=z-x d = z − x and ϕ = S ( z ) − c z + ν log c − H \phi=S(z)-cz+\nu\log c-H ϕ = S ( z ) − cz + ν log c − H . Then ϕ ( x , ν ) = ϕ z ( x , ν ) = 0 \phi(x,\nu)=\phi_z(x,\nu)=0 ϕ ( x , ν ) = ϕ z ( x , ν ) = 0 , and its double-zero quotient is ∫ 0 1 ( 1 − t ) S ′ ′ ( x + t d ) d t \int_0^1(1-t)S''(x+td)\,dt ∫ 0 1 ( 1 − t ) S ′′ ( x + t d ) d t , jointly analytic. Seek u = z ( c + d η ) u=z(c+d\eta) u = z ( c + d η ) . The phase equation becomes
ϕ d 2 − η + ν ∑ m ≥ 2 ( − 1 ) m − 1 m c − m d m − 2 η m = 0. (361) \frac{\phi}{d^2}-\eta+\nu\sum_{m\ge2}\frac{(-1)^{m-1}}{m}c^{-m}d^{m-2}\eta^m=0. \tag*{(361)} d 2 ϕ − η + ν m ≥ 2 ∑ m ( − 1 ) m − 1 c − m d m − 2 η m = 0. (361)
At z = ν = 0 z=\nu=0 z = ν = 0 it is S ′ ′ ( 0 ) / 2 − η = 0 S''(0)/2-\eta=0 S ′′ ( 0 ) /2 − η = 0 , with derivative − 1 -1 − 1 in η \eta η . The analytic implicit-function theorem gives η ( z , ν ) \eta(z,\nu) η ( z , ν ) . Since u z ( 0 , 0 ) = s u_z(0,0)=s u z ( 0 , 0 ) = s , the inverse Z Z Z is jointly analytic on a fixed polydisk. The construction gives u ( x , ν ) = ν u(x,\nu)=\nu u ( x , ν ) = ν and u ( z , 0 ) = S ( z ) u(z,0)=S(z) u ( z , 0 ) = S ( z ) . Only local analyticity of S S S is required.
The function f f f is jointly analytic because u / Z = c + d η u/Z=c+d\eta u / Z = c + d η is analytic and nonvanishing, its value at z = ν = 0 z=\nu=0 z = ν = 0 being s s s . For (E.22), write the coefficient as a residue and change variables z = Z ( u , ν ) z=Z(u,\nu) z = Z ( u , ν ) . For ν > 0 \nu>0 ν > 0 sufficiently small, the circle ∣ u ∣ = ν \lvert u\rvert=\nu ∣ u ∣ = ν lies in the fixed analytic neighborhood, and its image surrounds zero once. At k = ν = 0 k=\nu=0 k = ν = 0 , both sides equal H ( 0 ) H(0) H ( 0 ) , since S ( 0 ) = H ( 0 ) = 0 S(0)=\mathcal{H}(0)=0 S ( 0 ) = H ( 0 ) = 0 and f ( 0 , 0 ) = H ( 0 ) f(0,0)=H(0) f ( 0 , 0 ) = H ( 0 ) .
At the saddle, differentiating (E.21) twice in u u u gives ( S ′ ′ ( x ) + ν / x 2 ) Z u ( ν , ν ) 2 = 1 / ν (S''(x)+\nu/x^2)Z_u(\nu,\nu)^2=1/\nu ( S ′′ ( x ) + ν / x 2 ) Z u ( ν , ν ) 2 = 1/ ν . With v = D 2 S ( x ) = ν + x 2 S ′ ′ ( x ) v=\mathcal{D}^2S(x)=\nu+x^2S''(x) v = D 2 S ( x ) = ν + x 2 S ′′ ( x ) this is ( ν Z u ( ν , ν ) / x ) 2 = ν / v (\nu Z_u(\nu,\nu)/x)^2=\nu/v ( ν Z u ( ν , ν ) / x ) 2 = ν / v , and Z u > 0 Z_u>0 Z u > 0 near the origin, so f ( ν , ν ) = H ( x ) ν / v f(\nu,\nu)=H(x)\sqrt{\nu/v} f ( ν , ν ) = H ( x ) ν / v , with the positive branch. For H = B p H=B_p H = B p the normalization on the right is exactly e h ν h k B p ( x ) ν / v / k ! e^{h\nu}h^kB_p(x)\sqrt{\nu/v}/k! e h ν h k B p ( x ) ν / v / k ! , which equals the x x x -dependent part of (138) .
Poisson central polynomials with a uniform remainder
In the new variable the coefficient is a Poisson coefficient of an analytic amplitude. Expanding the amplitude about u = k / h u=k/h u = k / h expresses it through the central polynomials T m ( k ) T_m(k) T m ( k ) below; we bound the remainder uniformly in k k k .
Define the normalized coefficient operator by
P h , k f = k ! h − k [ u k ] e h u f ( u , ν ) , ν = k / h . \mathcal{P}_{h,k}f=k!h^{-k}[u^k]e^{hu}f(u,\nu),\qquad\nu=k/h. P h , k f = k ! h − k [ u k ] e h u f ( u , ν ) , ν = k / h .
Its central polynomial moments are
P h , k ( u − ν ) m = h − m T m ( k ) , ∑ m ≥ 0 T m ( k ) t m m ! = e − k t ( 1 + t ) k . (362) \mathcal{P}_{h,k}(u-\nu)^m=h^{-m}T_m(k),\qquad\sum_{m\ge0}T_m(k)\frac{t^m}{m!}=e^{-kt}(1+t)^k. \tag*{(362)} P h , k ( u − ν ) m = h − m T m ( k ) , m ≥ 0 ∑ T m ( k ) m ! t m = e − k t ( 1 + t ) k . (362)
Thus T 0 = 1 T_0=1 T 0 = 1 , T 1 = 0 T_1=0 T 1 = 0 ; for m ≥ 2 m\ge2 m ≥ 2 , T m T_m T m has degree at most ⌊ m / 2 ⌋ \lfloor m/2\rfloor ⌊ m /2 ⌋ and zero constant term. Write T m ( k ) = ∑ j = ⌈ m / 2 ⌉ m − 1 t m , j k m − j T_m(k)=\sum_{j=\lceil m/2\rceil}^{m-1}t_{m,j}k^{m-j} T m ( k ) = ∑ j = ⌈ m /2 ⌉ m − 1 t m , j k m − j .
Lemma E.7 (Poisson central expansion). Let f ( u , ν ) f(u,\nu) f ( u , ν ) be jointly analytic and bounded by one on a fixed polydisk. There are ν 0 , C > 0 \nu_0,C>0 ν 0 , C > 0 such that, uniformly for 0 ≤ ν = k / h ≤ ν 0 0\le\nu=k/h\le\nu_0 0 ≤ ν = k / h ≤ ν 0 and 0 ≤ K ≤ h / 2 0\le K\le h/2 0 ≤ K ≤ h /2 ,
P h , k f = ∑ j = 0 K h − j T j [ f ] ( ν ) + O ( C K + 1 ( K + 1 ) ! h − K − 1 ) , (363) \mathcal{P}_{h,k}f=\sum_{j=0}^{K}h^{-j}\mathcal{T}_j[f](\nu)+O\left(C^{K+1}(K+1)!h^{-K-1}\right), \tag*{(363)} P h , k f = j = 0 ∑ K h − j T j [ f ] ( ν ) + O ( C K + 1 ( K + 1 )! h − K − 1 ) , (363) T 0 [ f ] = f ( ν , ν ) , T j [ f ] = ∑ m = j + 1 2 j t m , j ν m − j ∂ u m f ( ν , ν ) m ! ( j ≥ 1 ) . \mathcal{T}_0[f]=f(\nu,\nu),\qquad\mathcal{T}_j[f]=\sum_{m=j+1}^{2j}t_{m,j}\nu^{m-j}\frac{\partial_u^m f(\nu,\nu)}{m!}\qquad(j\ge1). T 0 [ f ] = f ( ν , ν ) , T j [ f ] = m = j + 1 ∑ 2 j t m , j ν m − j m ! ∂ u m f ( ν , ν ) ( j ≥ 1 ) . Moreover ∥ T j [ f ] ∥ ≤ C j + 1 j ! \|\mathcal{T}_j[f]\|\le C^{j+1}j! ∥ T j [ f ] ∥ ≤ C j + 1 j ! on a smaller complex disk in ν \nu ν , each T j [ f ] \mathcal{T}_j[f] T j [ f ] with j ≥ 1 j\ge1 j ≥ 1 vanishes at ν = 0 \nu=0 ν = 0 , and ∣ P h , k f ∣ ≤ C |\mathcal{P}_{h,k}f|\le C ∣ P h , k f ∣ ≤ C .
Proof. Choose ν 0 > 0 \nu_0>0 ν 0 > 0 small and a fixed Cauchy radius r 0 r_0 r 0 about every u = ν u=\nu u = ν , with 2 ν 0 < r 0 / 2 2\nu_0<r_0/2 2 ν 0 < r 0 /2 . The Taylor remainder through degree 2 K + 1 2K+1 2 K + 1 on ∣ u ∣ = ν |u|=\nu ∣ u ∣ = ν is at most 2 r 0 − 2 K − 2 ∣ u − ν ∣ 2 K + 2 2r_0^{-2K-2}|u-\nu|^{2K+2} 2 r 0 − 2 K − 2 ∣ u − ν ∣ 2 K + 2 . For k ≥ 1 k\ge1 k ≥ 1 , normalized coefficient extraction therefore bounds it by
C r 0 − 2 K − 2 p k − 1 ν 2 K + 2 ∫ − π π ∣ θ ∣ 2 K + 2 e − c k θ 2 d θ ≤ C K + 1 ( K + 1 ) ! ν K + 1 h − K − 1 ≤ C 1 K + 1 ( K + 1 ) ! h − K − 1 . (364) Cr_0^{-2K-2}p_k^{-1}\nu^{2K+2}\int_{-\pi}^{\pi}|\theta|^{2K+2}e^{-ck\theta^2}\,d\theta
\le C^{K+1}(K+1)!\nu^{K+1}h^{-K-1}
\le C_1^{K+1}(K+1)!h^{-K-1}. \tag*{(364)} C r 0 − 2 K − 2 p k − 1 ν 2 K + 2 ∫ − π π ∣ θ ∣ 2 K + 2 e − c k θ 2 d θ ≤ C K + 1 ( K + 1 )! ν K + 1 h − K − 1 ≤ C 1 K + 1 ( K + 1 )! h − K − 1 . (364)
Here p k − 1 ≤ C k + 1 p_k^{-1}\le C\sqrt{k+1} p k − 1 ≤ C k + 1 and 1 − cos θ ≥ c θ 2 1-\cos\theta\ge c\theta^2 1 − cos θ ≥ c θ 2 on [ − π , π ] [-\pi,\pi] [ − π , π ] . The Gaussian integral is an upper bound at every order, even when K K K exceeds k k k . At k = 0 k=0 k = 0 the operator is evaluation at zero and this remainder vanishes.
The retained moment terms above inverse order K K K must also be bounded. The generating function in (E.24) gives
∣ [ k a ] T m ( k ) ∣ ≤ m ! a ! [ t m ] ( ∑ r ≥ 2 t r r ) a ≤ 2 m m ! a ! , 1 ≤ a ≤ m / 2. |[k^a]T_m(k)|\le\frac{m!}{a!}[t^m]\left(\sum_{r\ge2}\frac{t^r}{r}\right)^a\le2^m\frac{m!}{a!},\qquad1\le a\le m/2. ∣ [ k a ] T m ( k ) ∣ ≤ a ! m ! [ t m ] ( r ≥ 2 ∑ r t r ) a ≤ 2 m a ! m ! , 1 ≤ a ≤ m /2.
For ℓ = m − a = K + 1 + d > K \ell=m-a=K+1+d>K ℓ = m − a = K + 1 + d > K , the retained range has 0 ≤ d ≤ K − 1 0\le d\le K-1 0 ≤ d ≤ K − 1 , m ≤ 2 K + 1 m\le2K+1 m ≤ 2 K + 1 , and
m ! ( m − ℓ ) ! ≤ ( 2 K + 1 ) ! ( K − d ) ! ≤ C K ( K + 1 ) ! K d . \frac{m!}{(m-\ell)!}\le\frac{(2K+1)!}{(K-d)!}\le C^K(K+1)!K^d. ( m − ℓ )! m ! ≤ ( K − d )! ( 2 K + 1 )! ≤ C K ( K + 1 )! K d .
The sum of ( K / h ) d (K/h)^d ( K / h ) d is bounded when K ≤ h / 2 K \le h/2 K ≤ h /2 ; the polynomial number of terms is absorbed into C K C^K C K . At K = 0 K=0 K = 0 this sum is empty. This proves (E.25). At order j j j , m ! / ( m − j ) ! ≤ ( 2 j ) ! / j ! ≤ 4 j j ! m!/(m-j)! \le(2j)!/j! \le4^j j! m ! / ( m − j )! ≤ ( 2 j )! / j ! ≤ 4 j j ! . Cauchy’s bound thus gives ∥ T j [ f ] ∥ ≤ C j + 1 j ! \|\mathcal{T}_j[f]\| \le C^{j+1}j! ∥ T j [ f ] ∥ ≤ C j + 1 j ! on one smaller complex ν \nu ν disk. Each positive-order coefficient vanishes at zero. The same contour estimate with no vanishing factor gives ∣ P h , k f ∣ ≤ C ∥ f ∥ |\mathcal{P}_{h,k}f| \le C\|f\| ∣ P h , k f ∣ ≤ C ∥ f ∥ .
Apply Lemmas E.6 and E.7 to f s ( u , ν ) = B p ( Z ( u , ν ) ) L p , s ( Z ( u , ν ) ) u Z u / Z f_s(u,\nu)=B_p(Z(u,\nu))\mathcal{L}_{p,s}(Z(u,\nu))uZ_u/Z f s ( u , ν ) = B p ( Z ( u , ν )) L p , s ( Z ( u , ν )) u Z u / Z . Put f 0 ∗ ( ν ) = B p ( x ) ν / v f_0^*(\nu)=B_p(x)\sqrt{\nu/v} f 0 ∗ ( ν ) = B p ( x ) ν / v , which is bounded away from zero on a fixed small complex disk, with f 0 ∗ ( 0 ) = 1 f_0^*(0)=1 f 0 ∗ ( 0 ) = 1 . The extracted coefficients there are
A p , j ( Y ) = 1 f 0 ∗ ( ν ) ∑ s = 0 j T j − s [ f s ] ( ν ) , ν = Ψ ( X ( Y ) ) . (365) \mathcal{A}_{p,j}(Y)=\frac{1}{f_0^*(\nu)}\sum_{s=0}^{j}\mathcal{T}_{j-s}[f_s](\nu),\qquad\nu=\Psi(X(Y)). \tag*{(365)} A p , j ( Y ) = f 0 ∗ ( ν ) 1 s = 0 ∑ j T j − s [ f s ] ( ν ) , ν = Ψ ( X ( Y )) . (365)
This proves their zero-end analyticity and endpoint values. The common norms of Theorem 8.2 give ∥ f s ∥ ≤ C p ( B p F E ) s s ! \|f_s\|\le C_p(\mathcal{B}_p^{\mathrm{FE}})^s s! ∥ f s ∥ ≤ C p ( B p FE ) s s ! on one polydisk. Applying (E.25) at order K − s K-s K − s uses s ! ( K − s + 1 ) ! ≤ ( K + 1 ) ! s!(K-s+1)!\le(K+1)! s ! ( K − s + 1 )! ≤ ( K + 1 )! . For a finite marked model of order J = ⌊ κ h ⌋ > K J=\lfloor\kappa h\rfloor>K J = ⌊ κh ⌋ > K , the bounded operator controls all unused amplitudes by their first geometric tail term, since B p F E ( s + 1 ) / h < 1 / 2 \mathcal{B}_p^{\mathrm{FE}}(s+1)/h<1/2 B p FE ( s + 1 ) / h < 1/2 through J J J . Thus both model errors and coefficient norms have the required Gevrey bound. The functions (E.27) agree with (140) where both are defined; this is proved at the end of the proof of Theorem 10.2 , from the two expansions on a common range of densities.
This subsection proves Theorem 10.2 and then computes the first extracted coefficient A p , 1 \mathcal{A}_{p,1} A p , 1 , see (E.29), together with its value (E.30) at the critical endpoint. The inputs are Theorem 8.2 over its whole linear order range, which provides one finite marked model with exponentially small relative error on the real axis; the Fourier bounds (E.5) and the two-constants argument of (E.4), which transfer this error to the contour; Lemma E.2 for the disk bounds; and Lemmas E.6 and E.7 for the Poisson range. The proof is marked in stages: the order in which the constants are fixed, the comparison of the true coefficients with those of the model, the Poisson range of small densities, uniform Gaussian coordinates through the critical point, the Gaussian expansion, the exact Poisson normalization, the identification of the two coefficient families on their overlap, and the absorption of the residual exponential error.
Proof of Theorem 10.2 . We first replace the true generating function by one finite marked expansion, with exponentially small error after coefficient extraction. For small cusp density the Poisson calculation above applies. For the remaining densities we use uniform Gaussian coordinates and divide by the exact Poisson normalization. The last step absorbs the exponential comparison error into the stated remainder at every allowed order.
Order of the constants. The constants below are chosen in the order
κ → ν 0 → ν ∗ → d → η → τ → a → c p → c p ′ , \kappa\to\nu_0 \to\nu_* \to d \to\eta\to\tau\to a \to c_p \to c_p', κ → ν 0 → ν ∗ → d → η → τ → a → c p → c p ′ ,
each depending only on p p p and on those before it. The proportion κ \kappa κ fixes the model order J = ⌊ κ h ⌋ J=\lfloor\kappa h\rfloor J = ⌊ κh ⌋ and satisfies κ B p F E < 1 / 2 \kappa\mathcal{B}_p^{\mathrm{FE}}<1/2 κ B p FE < 1/2 , where we write B p F E \mathcal{B}_p^{\mathrm{FE}} B p FE for the Gevrey base of Theorem 8.2 throughout this proof, to distinguish it from the error base B p \mathcal{B}_p B p of the present theorem; we also take κ \kappa κ at most the linear-order constant of that theorem, so that it applies at order J J J . The proportion κ \kappa κ determines the exponential rate a 0 a_0 a 0 of the model. The radius ν 0 \nu_0 ν 0 is that of Lemma E.7 for the amplitudes f s f_s f s of (E.27), and 0 < ν ∗ < ν 0 0<\nu_*<\nu_0 0 < ν ∗ < ν 0 is the switching density. The small-disk radius d d d is chosen so that the two-constants bound on ∣ z ∣ ≤ d |z|\le d ∣ z ∣ ≤ d is exponentially small, then η \eta η so that the saddle for k < η h k<\eta h k < η h lies in ∣ z ∣ ≤ d / 4 |z|\le d/4 ∣ z ∣ ≤ d /4 and
η ≤ ν ∗ / 2 , \eta\le\nu_*/2, η ≤ ν ∗ /2 ,
so that the Gaussian comparison (E.28), proved for k ≥ η h k\ge\eta h k ≥ η h , covers every density Ψ ≥ ν ∗ / 2 \Psi\ge\nu_*/2 Ψ ≥ ν ∗ /2 . The Fourier cutoff T = τ h T=\tau\sqrt{h} T = τ h is fixed after η \eta η and a a a , and a > 0 a>0 a > 0 denotes the smallest of the exponential rates obtained. Two further smallness conditions are imposed where they arise in the Gaussian range: τ \tau τ is small compared with the radius of the common disk of the Gaussian coordinates, and c p c_p c p is small compared with that radius and with τ \tau τ ; neither depends on a constant chosen later. Finally c p c_p c p satisfies
c p < κ / 2 , c p ≤ c R ν ∗ / 2 , 2 c p < 1 , 2 c p log e 2 c p < a . c_p < \kappa/2,\qquad c_p \le c_R\nu_*/2,\qquad2c_p < 1,\qquad2c_p\log\frac{e}{2c_p} < a. c p < κ /2 , c p ≤ c R ν ∗ /2 , 2 c p < 1 , 2 c p log 2 c p e < a .
where c R c_R c R is the constant in the Gevrey remainder of R ( k ) R(k) R ( k ) below, which requires K ≤ c R k K \le c_R k K ≤ c R k while k ≥ ( ν ∗ / 2 ) h k \ge(\nu_*/2)h k ≥ ( ν ∗ /2 ) h on the Gaussian range; and c p ′ < c p c'_p < c_p c p ′ < c p satisfies B p c p ′ < 1 \mathcal{B}_p c'_p < 1 B p c p ′ < 1 for the error base B p \mathcal{B}_p B p of the present theorem.
The finite model. Write F h = Φ g ( p ) F_h=\Phi_g^{(p)} F h = Φ g ( p ) and J h = J g , p / π J_h=\mathcal{J}_{g,p}/\pi J h = J g , p / π . Choose a sufficiently small fixed κ > 0 \kappa>0 κ > 0 , set J = ⌊ κ h ⌋ J=\lfloor\kappa h\rfloor J = ⌊ κh ⌋ , and use the single finite model
M h ( z ) = J h ( z ) P J ( z ) , P J = ∑ s = 0 J L p , s h − s . M_h(z)=J_h(z)P_J(z),\qquad P_J=\sum_{s=0}^{J}\mathcal{L}_{p,s}h^{-s}. M h ( z ) = J h ( z ) P J ( z ) , P J = s = 0 ∑ J L p , s h − s .
Theorem 8.2 gives ∣ F h / M h − 1 ∣ ≤ C e − a 0 h |F_h/M_h-1|\le Ce^{-a_0h} ∣ F h / M h − 1∣ ≤ C e − a 0 h on the entire real interval. Its common coefficient norm also gives ∥ P J − 1 ∥ ≤ C / h \|P_J-1\|\le C/h ∥ P J − 1∥ ≤ C / h on one complex parameter neighborhood, since the consecutive majorant ratio is B p F E ( s + 1 ) / h < 1 / 2 \mathcal{B}^{\mathrm{FE}}_p(s+1)/h<1/2 B p FE ( s + 1 ) / h < 1/2 . All choices below use this same model. The constants in the two coefficient ranges can be chosen in the order d d d , then η \eta η , then τ \tau τ ; the small-disk argument below specifies the first two choices.
Comparison with the model: large indices. We first compare the true coefficients with those of this model. Put m ∗ = min ( h , k ) m_*=\min(h,k) m ∗ = min ( h , k ) and, for k ≥ 1 k\ge1 k ≥ 1 ,
F x ( w ) = e − k w / σ F h ( x e w / σ ) F h ( x ) , M x ( w ) = e − k w / σ M h ( x e w / σ ) M h ( x ) . \mathcal{F}_x(w)=e^{-kw/\sigma}\frac{F_h(xe^{w/\sigma})}{F_h(x)},\qquad
\mathcal{M}_x(w)=e^{-kw/\sigma}\frac{M_h(xe^{w/\sigma})}{M_h(x)}. F x ( w ) = e − k w / σ F h ( x ) F h ( x e w / σ ) , M x ( w ) = e − k w / σ M h ( x ) M h ( x e w / σ ) .
There is a fixed c > 0 c>0 c > 0 on whose disks ∣ w ∣ ≤ c m ∗ |w|\le c\sqrt{m_*} ∣ w ∣ ≤ c m ∗ the complex inverse parameter is defined and stays in the common coefficient neighborhood. At criticality, ε ( x e w / σ ) / ε ( x ) = 1 + O ( w / h ) \varepsilon(xe^{w/\sigma})/\varepsilon(x)=1+O(w/\sqrt{h}) ε ( x e w / σ ) / ε ( x ) = 1 + O ( w / h ) ; choosing c c c small fixes its square-root branch and keeps the tilted real radius below x c x_c x c . Near zero use ordinary analytic inversion, and use compactness between the endpoints. The centered model action has zero first derivative and bounded second derivative: its second derivative is v ( x e w / σ ) / v ( x ) v(xe^{w/\sigma})/v(x) v ( x e w / σ ) / v ( x ) . In the notation of Lemma E.2, with k = h Ψ ( x ) k=h\Psi(x) k = h Ψ ( x ) ,
log M x ( w ) = h ϕ x ( w / σ ) + ψ x ( w / σ ) + log P J ( x e w / σ ) P J ( x ) , \log\mathcal{M}_x(w)=h\phi_x(w/\sigma)+\psi_x(w/\sigma)+\log\frac{P_J(xe^{w/\sigma})}{P_J(x)}, log M x ( w ) = h ϕ x ( w / σ ) + ψ x ( w / σ ) + log P J ( x ) P J ( x e w / σ ) ,
and ∥ P J − 1 ∥ ≤ C / h \|P_J-1\|\le C/h ∥ P J − 1∥ ≤ C / h on the common neighborhood. Consequently, by (E.11), for ∣ w ∣ ≤ c m ∗ |w|\le c\sqrt{m_*} ∣ w ∣ ≤ c m ∗ ,
∣ M x ( w ) ∣ ≤ C e C ∣ w ∣ 2 + C ∣ w ∣ . |\mathcal{M}_x(w)|\le Ce^{C|w|^2+C|w|}. ∣ M x ( w ) ∣ ≤ C e C ∣ w ∣ 2 + C ∣ w ∣ .
Positivity bounds ∣ F x ( w ) ∣ |\mathcal{F}_x(w)| ∣ F x ( w ) ∣ by F x ( Re w ) \mathcal{F}_x(\operatorname{Re}w) F x ( Re w ) . On the real diameter the exact identity
F x ( u ) = M x ( u ) 1 + e h ( x e u / σ ) 1 + e h ( x ) , e h = F h / M h − 1 , \mathcal{F}_x(u)=\mathcal{M}_x(u)\frac{1+e_h(xe^{u/\sigma})}{1+e_h(x)},\qquad e_h=F_h/M_h-1, F x ( u ) = M x ( u ) 1 + e h ( x ) 1 + e h ( x e u / σ ) , e h = F h / M h − 1 ,
therefore gives the same disk bound and an additional factor e − a 0 h e^{-a_0h} e − a 0 h for their difference on that diameter.
For a fixed small η > 0 \eta>0 η > 0 , take k ≥ η h k\ge\eta h k ≥ η h and T = τ h T=\tau\sqrt{h} T = τ h . Choose τ > 0 \tau>0 τ > 0 after η , a 0 \eta,a_0 η , a 0 , so that 4 T ≤ c m ∗ 4T\le c\sqrt{m_*} 4 T ≤ c m ∗ , T ≤ c 1 m ∗ T\le c_1\sqrt{m_*} T ≤ c 1 m ∗ in (E.5), T ≤ c ∗ m ∗ T\le c_*\sqrt{m_*} T ≤ c ∗ m ∗ in (E.14), and T < π σ T<\pi\sigma T < π σ . The scale-invariant harmonic measure used in (E.4) gives
sup ∣ t ∣ ≤ T ∣ F x ( i t ) − M x ( i t ) ∣ ≤ C exp { − a 0 h / 2 + C T 2 + C T } ≤ C e − a 1 h . \sup_{|t|\le T}|\mathcal{F}_x(it)-\mathcal{M}_x(it)|\le C\exp\{-a_0h/2+CT^2+CT\}\le Ce^{-a_1h}. ∣ t ∣ ≤ T sup ∣ F x ( i t ) − M x ( i t ) ∣ ≤ C exp { − a 0 h /2 + C T 2 + C T } ≤ C e − a 1 h .
Its integral has the same bound after decreasing a 1 a_1 a 1 . The integrated tail (E.5) of the true distribution is also exponentially small in h h h ; the factor σ ≍ h / ε \sigma\asymp\sqrt{h/\varepsilon} σ ≍ h / ε from the outer-arc estimate is absorbed into this exponential bound. The model central integral is bounded by (E.14) and the bounded P J P_J P J ratio. Restoring the real normalization thus shows
∣ [ z k ] F h − x − k 2 π σ ∫ − T T e − i k t / σ M h ( x e i t / σ ) d t ∣ ≤ C Q g , p , k e − a 2 h . (366) \left|[z^k]F_h-\frac{x^{-k}}{2\pi\sigma}\int_{-T}^{T}e^{-ikt/\sigma}M_h(xe^{it/\sigma})\,dt\right|\le CQ_{g,p,k}e^{-a_2h}. \tag*{(366)} [ z k ] F h − 2 π σ x − k ∫ − T T e − ik t / σ M h ( x e i t / σ ) d t ≤ C Q g , p , k e − a 2 h . (366)
Indeed the error before division is at most C F h ( x ) x − k σ − 1 e − a 2 h CF_h(x)x^{-k}\sigma^{-1}e^{-a_2h} C F h ( x ) x − k σ − 1 e − a 2 h , and p k k p_k\sqrt{k} p k k is bounded above and below for integer k ≥ 1 k\ge1 k ≥ 1 . The real matching factor F h ( x ) / M h ( x ) = 1 + O ( e − a 2 h ) F_h(x)/M_h(x)=1+O(e^{-a_2h}) F h ( x ) / M h ( x ) = 1 + O ( e − a 2 h ) has been absorbed into the error. Here k ≥ η h k\ge\eta h k ≥ η h with η ≤ ν ∗ / 2 \eta\le\nu_*/2 η ≤ ν ∗ /2 , so that (E.28) covers every density Ψ ( x ) ≥ ν ∗ / 2 \Psi(x)\ge\nu_*/2 Ψ ( x ) ≥ ν ∗ /2 used in the Gaussian range below.
Comparison with the model: small indices. For the remaining k k k , put ϱ = d \varrho=d ϱ = d . On ∣ z ∣ ≤ 4 ϱ |z|\le4\varrho ∣ z ∣ ≤ 4 ϱ , positivity and the near-zero model give ∣ ( F h − M h ) / J h ( 0 ) ∣ ≤ C e C h ϱ |(F_h-M_h)/J_h(0)|\le Ce^{Ch\varrho} ∣ ( F h − M h ) / J h ( 0 ) ∣ ≤ C e C h ϱ ; on [ 0 , ϱ ] [0,\varrho] [ 0 , ϱ ] this improves to C e − a 0 h + C h ϱ Ce^{-a_0h+Ch\varrho} C e − a 0 h + C h ϱ . In the disk ∣ z − ϱ / 2 ∣ < 2 ϱ |z-\varrho/2|<2\varrho ∣ z − ϱ /2∣ < 2 ϱ slit along [ 0 , ϱ ] [0,\varrho] [ 0 , ϱ ] , the slit harmonic measure is greater than 1 / 4 1/4 1/4 on ∣ z ∣ ≤ ϱ / 4 |z|\le\varrho/4 ∣ z ∣ ≤ ϱ /4 . To see this, rescale to ∣ w ∣ < 4 |w|<4 ∣ w ∣ < 4 with slit [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] . On the target ∣ w ∣ ≤ 3 / 2 |w|\le3/2 ∣ w ∣ ≤ 3/2 its Green function is less than log 4 \log4 log 4 , and on ∣ w ∣ = 4 |w|=4 ∣ w ∣ = 4 it is greater than log 7 \log7 log 7 ; comparison gives 1 − log 4 / log 7 > 1 / 4 1-\log4/\log7>1/4 1 − log 4/ log 7 > 1/4 . Choose d d d small, then η \eta η small, so the two-constants bound is C e − a 3 h Ce^{-a_3h} C e − a 3 h and the saddle for 1 ≤ k < η h 1\le k<\eta h 1 ≤ k < η h lies in ∣ z ∣ ≤ ϱ / 4 |z|\le\varrho/4 ∣ z ∣ ≤ ϱ /4 . Cauchy’s bound on ∣ z ∣ = x |z|=x ∣ z ∣ = x is on the Q Q Q scale with a cost at most C k + 1 C\sqrt{k+1} C k + 1 : v / Ψ v/\Psi v /Ψ is bounded and J h ( x ) / J h ( 0 ) ≥ c J_h(x)/J_h(0)\ge c J h ( x ) / J h ( 0 ) ≥ c near zero. This polynomial cost is absorbed into a smaller exponential. At k = 0 k=0 k = 0 use the marked estimate at zero directly.
The Poisson range. Choose a switching density ν ∗ > 0 \nu_*>0 ν ∗ > 0 strictly inside the fixed Poisson neighborhood. For Ψ ( x ) ≤ ν ∗ \Psi(x)\le\nu_* Ψ ( x ) ≤ ν ∗ , the full model circle is analytic and
ℜ ( S ( x e i θ ) − S ( x ) ) ≤ − c x ( 1 − cos θ ) . \Re(S(xe^{i\theta})-S(x))\le-cx(1-\cos\theta). ℜ ( S ( x e i θ ) − S ( x )) ≤ − c x ( 1 − cos θ ) .
This follows from S ( z ) = 2 z + ∑ m ≥ 2 s m z m S(z)=2z+\sum_{m\ge2}s_mz^m S ( z ) = 2 z + ∑ m ≥ 2 s m z m and ∑ m ≥ 2 m 2 ∣ s m ∣ x m = O ( x 2 ) \sum_{m\ge2}m^2|s_m|x^m=O(x^2) ∑ m ≥ 2 m 2 ∣ s m ∣ x m = O ( x 2 ) , using 1 − cos ( m θ ) ≤ m 2 ( 1 − cos θ ) 1-\cos(m\theta)\le m^2(1-\cos\theta) 1 − cos ( m θ ) ≤ m 2 ( 1 − cos θ ) . The amplitudes are bounded, and h ν ≍ k h\nu\asymp k h ν ≍ k , so the standardized model tail is Gaussian on the whole circle. For k ≥ η h k\ge\eta h k ≥ η h it completes the integral in (E.28) to [ z k ] M h [z^k]M_h [ z k ] M h at exponential precision. Together with the small disk, this compares the true coefficients with those of the full model for every Ψ ( x ) ≤ ν ∗ \Psi(x)\le\nu_* Ψ ( x ) ≤ ν ∗ . The preceding Poisson argument now gives (10.6), with one additional fixed error C e − a h Ce^{-ah} C e − ah , throughout this range.
Uniform coordinates near criticality. For the other range we justify the growing-order Gaussian expansion uniformly through the critical point. In δ = y 0 − Y \delta=y_0-Y δ = y 0 − Y , write the convergent local factors as
x c − X ( Y ) = δ 2 α ( δ ) , Z ( Y ) = δ 5 / 2 ζ ( δ ) , B p ( X ( Y ) ) = δ p / 2 χ p ( δ ) , v ( X ( Y ) ) = δ − 4 v 0 ( δ ) , x_c-X(Y)=\delta^2\alpha(\delta),\qquad\mathcal{Z}(Y)=\delta^{5/2}\zeta(\delta),\qquad B_p(X(Y))=\delta^{p/2}\chi_p(\delta),\qquad v(X(Y))=\delta^{-4}v_0(\delta), x c − X ( Y ) = δ 2 α ( δ ) , Z ( Y ) = δ 5/2 ζ ( δ ) , B p ( X ( Y )) = δ p /2 χ p ( δ ) , v ( X ( Y )) = δ − 4 v 0 ( δ ) ,
where all four factors are analytic and nonzero at zero. For z = X ( Y ) e i u / v z=X(Y)e^{iu/\sqrt{v}} z = X ( Y ) e i u / v the inverse has δ z = δ q ( δ , u ) \delta_z=\delta q(\delta,u) δ z = δ q ( δ , u ) , jointly analytic and close to one: divide its equation x c − X ( y 0 − δ q ) = x c − z x_c-X(y_0-\delta q)=x_c-z x c − X ( y 0 − δ q ) = x c − z by δ 2 \delta^2 δ 2 and apply the implicit-function theorem at q = 1 , u = δ = 0 q=1,u=\delta=0 q = 1 , u = δ = 0 . Its q q q -derivative is 2 α ( 0 ) ≠ 0 2\alpha(0)\ne0 2 α ( 0 ) = 0 . Thus S ( z ) − S ( x ) S(z)-S(x) S ( z ) − S ( x ) and B p ( z ) / B p ( x ) B_p(z)/B_p(x) B p ( z ) / B p ( x ) are analytic joint families, expressed through log q \log q log q , ( δ q ) / ζ ( δ ) (\delta q)/\zeta(\delta) ( δ q ) / ζ ( δ ) and q p / 2 χ p ( δ q ) / χ p ( δ ) q^{p/2}\chi_p(\delta q)/\chi_p(\delta) q p /2 χ p ( δ q ) / χ p ( δ ) . Also Ψ / v \Psi/\sqrt v Ψ/ v extends analytically. Compactness gives common neighborhoods for Ψ ≥ ν ∗ / 2 \Psi\ge\nu_*/2 Ψ ≥ ν ∗ /2 .
Set
ψ Y ( u ) = S ( x e i u / v ) − S ( x ) − i Ψ u / v , W Y ( u ) = u − 2 ψ Y ( u ) / u 2 . \psi_Y(u)=S(xe^{iu/\sqrt v})-S(x)-i\Psi u/\sqrt v,\qquad W_Y(u)=u\sqrt{-2\psi_Y(u)/u^2}. ψ Y ( u ) = S ( x e i u / v ) − S ( x ) − i Ψ u / v , W Y ( u ) = u − 2 ψ Y ( u ) / u 2 .
The removable quotient is one at zero; W Y ′ ( 0 ) = 1 W_Y'(0)=1 W Y ′ ( 0 ) = 1 . On one fixed disk W Y W_Y W Y has a jointly analytic inverse U Y U_Y U Y , with ψ Y ( U Y ( w ) ) = − w 2 / 2 \psi_Y(U_Y(w))=-w^2/2 ψ Y ( U Y ( w )) = − w 2 /2 . Put u 0 = T / h = τ u_0=T/\sqrt h=\tau u 0 = T / h = τ . The image endpoints are ± u 0 + O ( u 0 2 ) \pm u_0+O(u_0^2) ± u 0 + O ( u 0 2 ) . Straight connectors to ± u 0 \pm u_0 ± u 0 have length O ( u 0 2 ) O(u_0^2) O ( u 0 2 ) and ℜ ( w 2 ) ≥ u 0 2 / 2 \Re(w^2)\ge u_0^2/2 ℜ ( w 2 ) ≥ u 0 2 /2 . A finite contour deformation therefore costs at most C h u 0 2 e − h u 0 2 / 4 C\sqrt h\,u_0^2e^{-hu_0^2/4} C h u 0 2 e − h u 0 2 /4 . After factoring J h ( x ) x − k / σ J_h(x)x^{-k/\sigma} J h ( x ) x − k / σ , the central integral becomes
h 2 π ∫ − u 0 u 0 e − h w 2 / 2 ∑ s = 0 J f ~ s ( w , Y ) h − s d w + O ( e − a h ) , \frac{\sqrt h}{2\pi}\int_{-u_0}^{u_0}e^{-hw^2/2}\sum_{s=0}^{J}\widetilde f_s(w,Y)h^{-s}\,dw+O(e^{-ah}), 2 π h ∫ − u 0 u 0 e − h w 2 /2 s = 0 ∑ J f s ( w , Y ) h − s d w + O ( e − ah ) ,
where f ~ s = [ B p ( z ( U Y ( w ) ) ) / B p ( x ) ] L p , s ( z ( U Y ( w ) ) ) U Y ′ ( w ) \widetilde{f}_s=[B_p(z(U_Y(w)))/B_p(x)]\mathcal{L}_{p,s}(z(U_Y(w)))U'_Y(w) f s = [ B p ( z ( U Y ( w ))) / B p ( x )] L p , s ( z ( U Y ( w ))) U Y ′ ( w ) has common norm C p ( B p F E ) s s ! C_p(\mathcal{B}^{\mathrm{FE}}_p)^s s! C p ( B p FE ) s s ! and f ~ 0 ( 0 , Y ) = 1 \widetilde{f}_0(0,Y)=1 f 0 ( 0 , Y ) = 1 .
Gaussian expansion. For s ≤ K s\le K s ≤ K , Taylor expand through degree 2 ( K − s ) + 1 2(K-s)+1 2 ( K − s ) + 1 on that fixed disk. The absolute Gaussian moment bounds the remainder by C K − s + 1 ( K − s + 1 ) ! h − K + s − 1 ∥ f ~ s ∥ C^{K-s+1}(K-s+1)!h^{-K+s-1}\|\widetilde{f}_s\| C K − s + 1 ( K − s + 1 )! h − K + s − 1 ∥ f s ∥ . Odd terms vanish. To extend the polynomial moments to the whole line, use, for ∣ w ∣ ≥ u 0 |w|\ge u_0 ∣ w ∣ ≥ u 0 ,
e − h w 2 / 2 ≤ e − h u 0 2 / 4 e − h w 2 / 4 . e^{-hw^2/2}\le e^{-hu_0^2/4}e^{-hw^2/4}. e − h w 2 /2 ≤ e − h u 0 2 /4 e − h w 2 /4 .
The remaining moment majorants C j j h − j Cj^j h^{-j} C j j h − j have bounded sum for j ≤ K ≤ c p h j\le K\le c_p h j ≤ K ≤ c p h , after choosing c p c_p c p sufficiently small relative to the fixed disk and u 0 u_0 u 0 . The sum over amplitude orders is bounded by the same choice. All polynomial tails are therefore O ( e − a h ) O(e^{-ah}) O ( e − ah ) , even at growing Taylor degree. The unused s > K s>K s > K amplitudes again form a geometric tail. The factorial convolution used above yields a numerator remainder C B K + 1 ( K + 1 ) ! h − K − 1 + C e − a h CB^{K+1}(K+1)!h^{-K-1}+Ce^{-ah} C B K + 1 ( K + 1 )! h − K − 1 + C e − ah . After factoring 1 / 2 π 1/\sqrt{2\pi} 1/ 2 π , its coefficients have the same common-domain bounds and agree with (139) . Indeed, for fixed Y Y Y and each fixed s s s , the model integral with the single amplitude L p , s \mathcal{L}_{p,s} L p , s is defined for every real h h h , and both families are the coefficients of its asymptotic expansion as h → ∞ h\to\infty h → ∞ through real values: (139) before and the present ones after the analytic change of variable u = U Y ( w ) u=U_Y(w) u = U Y ( w ) . Such expansion coefficients are unique. The argument is applied separately to each s s s , and the coefficients of the full model are then assembled by the formal product in (139) ; no uniqueness statement is needed for the model with the h h h -dependent order J J J .
The exact Poisson normalization. The divisor in (138) is exact. Put R ( k ) = 2 π k p k R(k)=\sqrt{2\pi k}\,p_k R ( k ) = 2 π k p k . Its logarithm satisfies
log R ( k ) = − ∫ 0 ∞ e − k t ( 1 e t − 1 − 1 t + 1 2 ) d t t . \log R(k)=-\int_0^\infty e^{-kt}\left(\frac{1}{e^t-1}-\frac{1}{t}+\frac{1}{2}\right)\frac{dt}{t}. log R ( k ) = − ∫ 0 ∞ e − k t ( e t − 1 1 − t 1 + 2 1 ) t d t .
The kernel has a removable origin, analytic radius 2 π 2\pi 2 π , and is bounded on the positive axis. Its Taylor remainder through degree L − 2 L-2 L − 2 is bounded globally there by C D L t L − 1 CD^L t^{L-1} C D L t L − 1 : use Cauchy near zero and bound each retained power outside a fixed small interval. Integration gives C D L ( L − 1 ) ! k − L CD^L(L-1)!k^{-L} C D L ( L − 1 )! k − L . For exponentiation and inversion, the degree-K K K log polynomial is bounded on ∣ t ∣ ≤ 1 / ( D 0 ( K + 1 ) ) |t|\le1/(D_0(K+1)) ∣ t ∣ ≤ 1/ ( D 0 ( K + 1 )) for a fixed D 0 > 0 D_0>0 D 0 > 0 , as in the proof of Lemma D.11. Cauchy’s tail at t = 1 / k t=1/k t = 1/ k , for K ≤ c R k K\le c_Rk K ≤ c R k , is at most C [ D 0 ( K + 1 ) / k ] K + 1 ≤ C ( e D 0 ) K + 1 ( K + 1 ) ! k − K − 1 C[D_0(K+1)/k]^{K+1}\le C(eD_0)^{K+1}(K+1)!k^{-K-1} C [ D 0 ( K + 1 ) / k ] K + 1 ≤ C ( e D 0 ) K + 1 ( K + 1 )! k − K − 1 . Since the exponential is bounded on the relevant real interval, the remainder for log R \log R log R transfers to R R R and to R − 1 R^{-1} R − 1 , which therefore have Gevrey-1 remainders. Here k ≥ ( ν ∗ / 2 ) h k\ge(\nu_*/2)h k ≥ ( ν ∗ /2 ) h and Ψ − 1 \Psi^{-1} Ψ − 1 is bounded analytic through criticality. By (138) , Q g , p , k / ( J g , p ( x ) x − k / ( π σ ) ) = R ( k ) / 2 π Q_{g,p,k}/(\mathcal{J}_{g,p}(x)x^{-k}/(\pi\sigma))=R(k)/\sqrt{2\pi} Q g , p , k / ( J g , p ( x ) x − k / ( π σ )) = R ( k ) / 2 π . The two factors 1 / 2 π 1/\sqrt{2\pi} 1/ 2 π cancel. Dividing the normalized numerator by R ( k ) R(k) R ( k ) consequently preserves the Gevrey-1 error and gives exactly (140) .
Identification on the overlap. For ν ∗ / 2 ≤ Ψ ≤ ν ∗ \nu_*/2\le\Psi\le\nu_* ν ∗ /2 ≤ Ψ ≤ ν ∗ both expansions hold; write A p , j P \mathcal{A}^P_{p,j} A p , j P for the Poisson coefficients (E.27) and A j G \mathcal{A}^G_j A j G for the Gaussian coefficients (140) . Fixed-Y Y Y uniqueness is not available directly, since k = h Ψ ( X ( Y ) ) k=h\Psi(X(Y)) k = h Ψ ( X ( Y )) must be an integer, so we argue by density. Suppose the two families agree below order j j j . Subtracting the expansions at order K = j K=j K = j gives A p , j P ( Y h , k ) − A j G ( Y h , k ) = O ( h − 1 ) \mathcal{A}^P_{p,j}(Y_{h,k})-\mathcal{A}^G_j(Y_{h,k})=O(h^{-1}) A p , j P ( Y h , k ) − A j G ( Y h , k ) = O ( h − 1 ) , uniformly over all admissible ( h , k ) (h,k) ( h , k ) with Y h , k Y_{h,k} Y h , k in the overlap, where h Ψ ( X ( Y h , k ) ) = k h\Psi(X(Y_{h,k}))=k h Ψ ( X ( Y h , k )) = k . Given Y Y Y in the open overlap, take k = ⌊ h Ψ ( X ( Y ) ) ⌋ k=\lfloor h\Psi(X(Y))\rfloor k = ⌊ h Ψ ( X ( Y ))⌋ ; since Ψ ∘ X \Psi\circ X Ψ ∘ X is continuous and strictly increasing, Y h , k → Y Y_{h,k}\to Y Y h , k → Y as h → ∞ h\to\infty h → ∞ . Both coefficients are continuous, so they agree at Y Y Y . By induction the two families coincide on the open overlap. Both are analytic on 0 ≤ Ψ ( X ( Y ) ) < ν ∗ 0\le\Psi(X(Y))<\nu_* 0 ≤ Ψ ( X ( Y )) < ν ∗ , so the identity theorem extends the equality to this whole interval. Hence (E.27) is the analytic continuation of (140) to Y = 0 Y=0 Y = 0 , and the coefficients are analytic on one complex neighborhood of [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] .
Absorbing the residual error. It remains to absorb the fixed exponential error at every allowed order. After fixing the model, neighborhoods and cutoff, choose c p < κ / 2 c_p<\kappa/2 c p < κ /2 so small that 2 c p < 1 2c_p<1 2 c p < 1 and 2 c p log ( e / ( 2 c p ) ) < a 2c_p\log(e/(2c_p))<a 2 c p log ( e / ( 2 c p )) < a . For L = K + 1 ≤ 2 c p h L=K+1\le2c_ph L = K + 1 ≤ 2 c p h and B p ≥ 1 \mathcal{B}_p\ge1 B p ≥ 1 ,
B p L L ! h − L ≥ e − L log ( e h / L ) ≥ e − 2 c p log ( e / ( 2 c p ) ) h . \mathcal{B}_p^L L!h^{-L}\ge e^{-L\log(eh/L)}\ge e^{-2c_p\log(e/(2c_p))h}. B p L L ! h − L ≥ e − L l o g ( e h / L ) ≥ e − 2 c p l o g ( e / ( 2 c p )) h .
This bounds the fixed residual by the stated remainder simultaneously for every K ≤ c p h K \le c_p h K ≤ c p h , including zero, once h h h is large enough. Finally choose c p ′ < c p c'_p < c_p c p ′ < c p after the final B p B_p B p , with B p c p ′ < 1 B_p c'_p < 1 B p c p ′ < 1 . The upper bound B p L L ! h − L ≤ ( B p L / h ) L \mathfrak{B}_p^L L!h^{-L} \le(B_pL/h)^L B p L L ! h − L ≤ ( B p L / h ) L proves the exponential approximation at K = ⌊ c p ′ h ⌋ K=\lfloor c'_p h\rfloor K = ⌊ c p ′ h ⌋ . □ \square □
The first coefficient after extraction follows from (139) –(140) , in which s 1 = − 1 / 12 s_1=-1/12 s 1 = − 1/12 : with E 1 E_1 E 1 the Gaussian coefficient displayed in the proof of Lemma E.4 , one has A p , 1 = L p , 1 + E 1 + 1 / ( 12 Ψ ) \mathcal{A}_{p,1}=\mathcal{L}_{p,1}+E_1+1/(12\Psi) A p , 1 = L p , 1 + E 1 + 1/ ( 12Ψ ) , that is,
A p , 1 = L p , 1 − D 2 b + ( D b ) 2 2 v + ( D b ) D 2 Ψ 2 v 2 + D 3 Ψ 8 v 2 − 5 ( D 2 Ψ ) 2 24 v 3 + 1 12 Ψ , b = log B p . (367) \mathcal{A}_{p,1}=\mathcal{L}_{p,1}-\frac{\mathcal{D}^2b+(\mathcal{D}b)^2}{2v}+\frac{(\mathcal{D}b)\mathcal{D}^2\Psi}{2v^2}+\frac{\mathcal{D}^3\Psi}{8v^2}-\frac{5(\mathcal{D}^2\Psi)^2}{24v^3}+\frac{1}{12\Psi},\qquad b=\log B_p. \tag*{(367)} A p , 1 = L p , 1 − 2 v D 2 b + ( D b ) 2 + 2 v 2 ( D b ) D 2 Ψ + 8 v 2 D 3 Ψ − 24 v 3 5 ( D 2 Ψ ) 2 + 12Ψ 1 , b = log B p . (367)
Its apparent pole at zero is removable by (E.27). By Theorem 10.2 with K = 1 K=1 K = 1 , retaining this term improves the relative remainder to O p ( h − 2 ) O_p(h^{-2}) O p ( h − 2 ) over every coefficient index.
At the critical endpoint its value is
A p , 1 ( y 0 ) = − p 2 8 + p 2 − 37 60 . (368) \mathcal{A}_{p,1}(y_0)=-\frac{p^2}{8}+\frac{p}{2}-\frac{37}{60}. \tag*{(368)} A p , 1 ( y 0 ) = − 8 p 2 + 2 p − 60 37 . (368)
Indeed, writing ε = 1 − x / x c \varepsilon=1-x/x_c ε = 1 − x / x c , the convergent factorizations recorded above in the proof of Theorem 10.2 , together with the exact relation ε = δ 2 α ( δ ) / x c \varepsilon=\delta^2\alpha(\delta)/x_c ε = δ 2 α ( δ ) / x c , give S = − ( 5 / 4 ) log ε + α S ( δ ) S=-(5/4)\log\varepsilon+\alpha_S(\delta) S = − ( 5/4 ) log ε + α S ( δ ) and log B p = ( p / 4 ) log ε + α B ( δ ) \log B_p=(p/4)\log\varepsilon+\alpha_B(\delta) log B p = ( p /4 ) log ε + α B ( δ ) with α S , α B \alpha_S,\alpha_B α S , α B analytic at δ = 0 \delta=0 δ = 0 . Since D ε = − ( 1 − ε ) \mathcal{D}\varepsilon=-(1-\varepsilon) D ε = − ( 1 − ε ) , one has D m log ε = − ( m − 1 ) ! ε − m ( 1 + O ( ε ) ) \mathcal{D}^m\log\varepsilon=-(m-1)!\varepsilon^{-m}(1+O(\varepsilon)) D m log ε = − ( m − 1 )! ε − m ( 1 + O ( ε )) , while D m \mathcal{D}^m D m applied to an analytic function of δ \delta δ is O ( δ 1 − 2 m ) = O ( ε 1 / 2 − m ) O(\delta^{1-2m})=O(\varepsilon^{1/2-m}) O ( δ 1 − 2 m ) = O ( ε 1/2 − m ) , since D = − ( x / a ( Y ) ) d / d δ \mathcal{D}=-(x/a(Y))\,d/d\delta D = − ( x / a ( Y )) d / d δ and a ( Y ) ≍ δ a(Y)\asymp\delta a ( Y ) ≍ δ ; relative to ε − m \varepsilon^{-m} ε − m this is O ( ε ) O(\sqrt{\varepsilon}) O ( ε ) . With v = D 2 S = 5 4 ε − 2 ( 1 + O ( ε ) ) v=\mathcal{D}^2S=\frac{5}{4}\varepsilon^{-2}(1+O(\sqrt{\varepsilon})) v = D 2 S = 4 5 ε − 2 ( 1 + O ( ε )) , the normalized derivatives have the endpoint limits
D m S v m / 2 → ( 5 4 ) 1 − m / 2 ( m − 1 ) ! ( m ≥ 2 ) , D m b v m / 2 → − p 4 ( m − 1 ) ! ( 5 4 ) − m / 2 ( m ≥ 1 ) . \frac{\mathcal{D}^mS}{v^{m/2}}\to\left(\frac{5}{4}\right)^{1-m/2}(m-1)!\quad(m\ge2),\qquad\frac{\mathcal{D}^m b}{v^{m/2}}\to-\frac{p}{4}(m-1)!\left(\frac{5}{4}\right)^{-m/2}\quad(m\ge1). v m /2 D m S → ( 4 5 ) 1 − m /2 ( m − 1 )! ( m ≥ 2 ) , v m /2 D m b → − 4 p ( m − 1 )! ( 4 5 ) − m /2 ( m ≥ 1 ) .
and the four terms of E 1 E_1 E 1 in (E.29) tend respectively to
− D 2 b + ( D b ) 2 2 v → − p 2 40 + p 10 , ( D b ) D 2 Ψ 2 v 2 → − p 5 , D 3 Ψ 8 v 2 → 3 5 , − 5 ( D 2 Ψ ) 2 24 v 3 → − 2 3 , -\frac{\mathcal{D}^2b+(\mathcal{D}b)^2}{2v}\to-\frac{p^2}{40}+\frac{p}{10},\qquad\frac{(\mathcal{D}b)\mathcal{D}^2\Psi}{2v^2}\to-\frac{p}{5},\qquad\frac{\mathcal{D}^3\Psi}{8v^2}\to\frac{3}{5},\qquad-\frac{5(\mathcal{D}^2\Psi)^2}{24v^3}\to-\frac{2}{3}, − 2 v D 2 b + ( D b ) 2 → − 40 p 2 + 10 p , 2 v 2 ( D b ) D 2 Ψ → − 5 p , 8 v 2 D 3 Ψ → 5 3 , − 24 v 3 5 ( D 2 Ψ ) 2 → − 3 2 ,
whose sum is E 1 ( y 0 ) = − ( p 2 + 4 p ) / 40 − 1 / 15 E_1(y_0)=-(p^2+4p)/40-1/15 E 1 ( y 0 ) = − ( p 2 + 4 p ) /40 − 1/15 . Combine this with L p , 1 ( y 0 ) = t p = ( − 2 p 2 + 12 p − 11 ) / 20 \mathcal{L}_{p,1}(y_0)=t_p=(-2p^2+12p-11)/20 L p , 1 ( y 0 ) = t p = ( − 2 p 2 + 12 p − 11 ) /20 from (105) and Ψ − 1 → 0 \Psi^{-1}\to0 Ψ − 1 → 0 .
Appendix F. The rate for adjacent coefficient ratios
This appendix proves Theorem F.1 , which contains the rate in Theorem 1.4 . We quantify the three degree ranges treated in § 3.35 . The logarithmic comparison and concavity give the rate away from the lowest degrees, including bounded positive codimension at the top. Bracket difference estimates supply the remaining range. Balancing the bounds gives O ( g − 1 / 8 ) O(g^{-1/8}) O ( g − 1/8 ) uniformly; we do not claim that this exponent is optimal. The coefficients and the exact density are those of Theorem 1.4 .
Theorem F.1 (Quantitative adjacent ratios). There is a constant C C C such that, for all sufficiently large g g g ,
sup 0 ≤ k < 3 g − 2 ∣ a g , k / a g , k + 1 4 π 2 / t g , k 2 − 1 ∣ ≤ C g − 1 / 8 . (369) \sup_{0\le k<3g-2}\left|\frac{a_{g,k}/a_{g,k+1}}{4\pi^2/t_{g,k}^2}-1\right|\le Cg^{-1/8}. \tag*{(369)} 0 ≤ k < 3 g − 2 sup 4 π 2 / t g , k 2 a g , k / a g , k + 1 − 1 ≤ C g − 1/8 . (369) For every fixed ε ∈ ( 0 , 3 ) \varepsilon\in(0,3) ε ∈ ( 0 , 3 ) there is C ε C_\varepsilon C ε such that
sup ε g ≤ k < 3 g − 2 ∣ a g , k / a g , k + 1 4 π 2 / t g , k 2 − 1 ∣ ≤ C ε g − 1 / 4 . (370) \sup_{\varepsilon g\le k<3g-2}\left|\frac{a_{g,k}/a_{g,k+1}}{4\pi^2/t_{g,k}^2}-1\right|\le C_\varepsilon g^{-1/4}. \tag*{(370)} ε g ≤ k < 3 g − 2 sup 4 π 2 / t g , k 2 a g , k / a g , k + 1 − 1 ≤ C ε g − 1/4 . (370) In the second supremum k k k is an integer. Both bounds include bounded positive codimension from the top coefficient.
The logarithmic model up to its endpoints
We first make the logarithmic comparison of § 3.35 quantitative, up to both ends of the degree range. The result is Lemma F.4: the logarithm of the normalized coefficient sequence agrees with g f ( k / g ) gf(k/g) g f ( k / g ) to within O ( g ) O(\sqrt{g}) O ( g ) for every degree 0 ≤ k ≤ N 0 \le k \le N 0 ≤ k ≤ N . Its inputs are the exact model (159) , the comparison (160) and the saddle-point description (165) –(170) of the function f f f , all from § 3.35 . The new point is the degenerate end m / h → 0 m/h \to0 m / h → 0 , where the tilted summands depend on h h h ; Lemma F.2 supplies the local limit theorem for such a triangular array, and Lemma F.3 uses it to bound the model coefficients uniformly in the codimension m m m .
Use H H H , A ~ m \widetilde{A}_m A m , μ \mu μ from (159) and (165) . Let ζ ( u ) \zeta(u) ζ ( u ) be the inverse of μ ( z ) \mu(z) μ ( z ) on 0 ≤ u ≤ 3 / 2 0 \le u \le3/2 0 ≤ u ≤ 3/2 , with ζ ( 0 ) = 0 \zeta(0) = 0 ζ ( 0 ) = 0 . Since μ ( z ) = z / 5 + O ( z 2 ) \mu(z) = z/5 + O(z^2) μ ( z ) = z /5 + O ( z 2 ) at zero and μ ′ ( z ) > 0 \mu'(z) > 0 μ ′ ( z ) > 0 throughout [ 0 , π 2 / 2 ] [0,\pi^2/2] [ 0 , π 2 /2 ] , the function
χ ( u ) = u ζ ( u ) , χ ( 0 ) = 1 5 , \chi(u) = \frac{u}{\zeta(u)}, \qquad\chi(0) = \frac{1}{5}, χ ( u ) = ζ ( u ) u , χ ( 0 ) = 5 1 ,
is positive and analytic in a neighborhood of the closed real interval. Write f f f for the function in (169) , extended continuously to [ 0 , 3 ] [0,3] [ 0 , 3 ] , and set
q ( α ) = 4 α χ ( ( 3 − α ) / 2 ) , f ′ ( α ) = − log q ( α ) , ∣ f ′ ′ ( α ) ∣ ≤ C ( 1 + α − 1 ) ( 0 < α ≤ 3 ) . (371) q(\alpha) = 4\alpha\chi((3-\alpha)/2), \qquad f'(\alpha) = -\log q(\alpha), \qquad|f''(\alpha)| \le C(1+\alpha^{-1}) \quad(0 < \alpha\le3). \tag*{(371)} q ( α ) = 4 α χ (( 3 − α ) /2 ) , f ′ ( α ) = − log q ( α ) , ∣ f ′′ ( α ) ∣ ≤ C ( 1 + α − 1 ) ( 0 < α ≤ 3 ) . (371)
These follow from (170) ; in particular q ( 3 ) = 12 / 5 q(3) = 12/5 q ( 3 ) = 12/5 and f ( 3 ) = 0 f(3) = 0 f ( 3 ) = 0 .
We need the endpoint-uniform version of the logarithmic saddle estimate (167) . At the degenerate end m / h → 0 m/h \to0 m / h → 0 the tilted summands depend on h h h , so the local limit theorem required is one for a triangular array; we isolate it.
Lemma F.2 (Triangular local limit). For each h h h let ( p j ) j ≥ 0 = ( p j ( h ) ) j ≥ 0 (p_j)_{j\ge0} = (p_j^{(h)})_{j\ge0} ( p j ) j ≥ 0 = ( p j ( h ) ) j ≥ 0 be a probability law on N \mathbb{N} N with characteristic function φ 1 ( θ ) = ∑ j p j e i j θ \varphi_1(\theta) = \sum_j p_j e^{ij\theta} φ 1 ( θ ) = ∑ j p j e ij θ , mean μ 1 \mu_1 μ 1 and variance Var 1 \operatorname{Var}_1 Var 1 , and let S h S_h S h be the sum of h h h independent copies. Suppose that there are z = z h > 0 z = z_h > 0 z = z h > 0 and constants c , C > 0 c,C > 0 c , C > 0 such that φ 1 \varphi_1 φ 1 does not vanish on [ − π , π ] [-\pi,\pi] [ − π , π ] and
p 0 p 1 ≥ c z , Var 1 ≤ C z , sup ∣ θ ∣ ≤ π ∣ ∂ θ 3 log φ 1 ( θ ) ∣ ≤ C z . (372) p_0p_1 \ge cz, \qquad\operatorname{Var}_1 \le Cz, \qquad\sup_{|\theta|\le\pi} |\partial_\theta^3 \log\varphi_1(\theta)| \le Cz. \tag*{(372)} p 0 p 1 ≥ cz , Var 1 ≤ C z , ∣ θ ∣ ≤ π sup ∣ ∂ θ 3 log φ 1 ( θ ) ∣ ≤ C z . (372) Then Var 1 ≥ p 0 p 1 ≥ c z \operatorname{Var}_1 \ge p_0p_1 \ge cz Var 1 ≥ p 0 p 1 ≥ cz , and for every integer m m m with ∣ m − h μ 1 ∣ ≤ 1 |m-h\mu_1| \le1 ∣ m − h μ 1 ∣ ≤ 1 , as h z → ∞ hz \to\infty h z → ∞ ,
Pr ( S h = m ) = ( 2 π h Var 1 ) − 1 / 2 ( 1 + o ( 1 ) ) , (373) \Pr(S_h=m) = (2\pi h\operatorname{Var}_1)^{-1/2}(1+o(1)), \tag*{(373)} Pr ( S h = m ) = ( 2 π h Var 1 ) − 1/2 ( 1 + o ( 1 )) , (373) uniformly over all arrays satisfying (F.4) with the same c , C c,C c , C .
Proof. First, for every θ \theta θ ,
∣ φ 1 ( θ ) ∣ 2 = ∑ j , k p j p k cos ( ( j − k ) θ ) = 1 − ∑ j , k p j p k ( 1 − cos ( ( j − k ) θ ) ) ≤ 1 − 2 p 0 p 1 ( 1 − cos θ ) , (374) |\varphi_1(\theta)|^2 = \sum_{j,k}p_jp_k\cos((j-k)\theta) = 1-\sum_{j,k}p_jp_k(1-\cos((j-k)\theta)) \le1-2p_0p_1(1-\cos\theta), \tag*{(374)} ∣ φ 1 ( θ ) ∣ 2 = j , k ∑ p j p k cos (( j − k ) θ ) = 1 − j , k ∑ p j p k ( 1 − cos (( j − k ) θ )) ≤ 1 − 2 p 0 p 1 ( 1 − cos θ ) , (374)
since ∑ j p j = 1 \sum_j p_j=1 ∑ j p j = 1 and every term of the middle sum is nonnegative; this is the bound (115) , reproved here. Also Var 1 = 1 2 ∑ j , k p j p k ( j − k ) 2 ≥ p 0 p 1 \operatorname{Var}_1 = \frac{1}{2}\sum_{j,k}p_jp_k(j-k)^2 \ge p_0p_1 Var 1 = 2 1 ∑ j , k p j p k ( j − k ) 2 ≥ p 0 p 1 . Write log φ 1 ( θ ) = i μ 1 θ − 1 2 Var 1 θ 2 + R ( θ ) \log\varphi_1(\theta)=i\mu_1\theta-\frac{1}{2}\operatorname{Var}_1\theta^2+R(\theta) log φ 1 ( θ ) = i μ 1 θ − 2 1 Var 1 θ 2 + R ( θ ) ; by Taylor’s formula and the third hypothesis, ∣ R ( θ ) ∣ ≤ C z ∣ θ ∣ 3 / 6 |R(\theta)|\le Cz|\theta|^3/6 ∣ R ( θ ) ∣ ≤ C z ∣ θ ∣ 3 /6 on [ − π , π ] [-\pi,\pi] [ − π , π ] . Fourier inversion gives
h Var 1 Pr ( S h = m ) = 1 2 π ∫ − π h Var 1 π h Var 1 exp { − ϑ 2 2 + h R ( ϑ h Var 1 ) − i ( m − h μ 1 ) ϑ h Var 1 } d ϑ , \sqrt{h\operatorname{Var}_1}\Pr(S_h=m) = \frac{1}{2\pi}\int_{-\pi\sqrt{h\operatorname{Var}_1}}^{\pi\sqrt{h\operatorname{Var}_1}} \exp\left\{-\frac{\vartheta^2}{2}+hR\left(\frac{\vartheta}{\sqrt{h\operatorname{Var}_1}}\right)-i(m-h\mu_1)\frac{\vartheta}{\sqrt{h\operatorname{Var}_1}}\right\}\,d\vartheta, h Var 1 Pr ( S h = m ) = 2 π 1 ∫ − π h Var 1 π h Var 1 exp { − 2 ϑ 2 + h R ( h Var 1 ϑ ) − i ( m − h μ 1 ) h Var 1 ϑ } d ϑ ,
after the substitution θ = ϑ / h Var 1 \theta=\vartheta/\sqrt{h\operatorname{Var}_1} θ = ϑ / h Var 1 . For fixed ϑ \vartheta ϑ the exponent tends to − ϑ 2 / 2 -\vartheta^2/2 − ϑ 2 /2 , because h ∣ R ∣ ≤ C h z ∣ ϑ ∣ 3 ( h Var 1 ) − 3 / 2 ≤ C ′ ∣ ϑ ∣ 3 ( h z ) − 1 / 2 → 0 h|R|\le Chz|\vartheta|^3(h\operatorname{Var}_1)^{-3/2}\le C'|\vartheta|^3(hz)^{-1/2}\to0 h ∣ R ∣ ≤ C h z ∣ ϑ ∣ 3 ( h Var 1 ) − 3/2 ≤ C ′ ∣ ϑ ∣ 3 ( h z ) − 1/2 → 0 by Var 1 ≥ c z \operatorname{Var}_1\ge cz Var 1 ≥ cz , and ∣ m − h μ 1 ∣ ∣ ϑ ∣ ( h Var 1 ) − 1 / 2 → 0 |m-h\mu_1||\vartheta|(h\operatorname{Var}_1)^{-1/2}\to0 ∣ m − h μ 1 ∣∣ ϑ ∣ ( h Var 1 ) − 1/2 → 0 . For domination, (F.6) and 1 − cos θ ≥ 2 θ 2 / π 2 1-\cos\theta\ge2\theta^2/\pi^2 1 − cos θ ≥ 2 θ 2 / π 2 on [ − π , π ] [-\pi,\pi] [ − π , π ] give ∣ φ 1 ( θ ) ∣ h ≤ exp { − h p 0 p 1 ( 1 − cos θ ) } ≤ exp { − 2 c h z θ 2 / π 2 } ≤ exp { − 2 c ϑ 2 / ( π 2 C ) } |\varphi_1(\theta)|^h\le\exp\{-hp_0p_1(1-\cos\theta)\}\le\exp\{-2chz\theta^2/\pi^2\}\le\exp\{-2c\vartheta^2/(\pi^2C)\} ∣ φ 1 ( θ ) ∣ h ≤ exp { − h p 0 p 1 ( 1 − cos θ )} ≤ exp { − 2 c h z θ 2 / π 2 } ≤ exp { − 2 c ϑ 2 / ( π 2 C )} , using Var 1 ≤ C z \operatorname{Var}_1\le Cz Var 1 ≤ C z . Dominated convergence gives the limit ( 2 π ) − 1 / 2 (2\pi)^{-1/2} ( 2 π ) − 1/2 of the right side, which is (F.5). The convergence is uniform over the class: on ∣ ϑ ∣ ≤ ( h z ) 1 / 8 |\vartheta|\le(hz)^{1/8} ∣ ϑ ∣ ≤ ( h z ) 1/8 the exponent differs from − ϑ 2 / 2 -\vartheta^2/2 − ϑ 2 /2 by at most C ′ ′ ( h z ) − 1 / 8 C''(hz)^{-1/8} C ′′ ( h z ) − 1/8 by the two displayed bounds, and the complement is controlled by the Gaussian majorant; every constant depends only on c , C c,C c , C . □
Lemma F.3 (Endpoint-uniform saddle estimate). For a positive integer h h h and integers 0 ≤ m ≤ 3 h / 2 0 \le m \le3h/2 0 ≤ m ≤ 3 h /2 , let z = ζ ( m / h ) z = \zeta(m/h) z = ζ ( m / h ) , interpreting the m = 0 m = 0 m = 0 expressions by continuity. Then, uniformly in m m m ,
log A ~ m ( h ) = − h log H ( z ) − m log z + O ( log ( h + 1 ) ) . (375) \log\widetilde{A}_m(h) = -h \log H(z) - m \log z + O(\log(h+1)). \tag*{(375)} log A m ( h ) = − h log H ( z ) − m log z + O ( log ( h + 1 )) . (375)
Proof. Tilt H − 1 H^{-1} H − 1 at z z z and let S h S_h S h be the sum of h h h independent variables with probability generating function H ( z ) / H ( z w ) H(z)/H(zw) H ( z ) / H ( z w ) . Exactly,
A ~ m ( h ) = H ( z ) − h z − m Pr ( S h = m ) . \widetilde{A}_m(h) = H(z)^{-h}z^{-m}\Pr(S_h=m). A m ( h ) = H ( z ) − h z − m Pr ( S h = m ) .
There is c > 0 c>0 c > 0 such that Pr ( S h = m ) ≥ c / m + 1 \Pr(S_h=m) \ge c/\sqrt{m+1} Pr ( S h = m ) ≥ c / m + 1 uniformly. We give the details at the degenerate endpoint m / h → 0 m/h \to0 m / h → 0 , where Lemma F.2 applies with the parameter z = ζ ( m / h ) z=\zeta(m/h) z = ζ ( m / h ) . Indeed, the summand law has p j = H ( z ) z j [ w j ] H ( w ) − 1 = H ( z ) h j z j p_j=H(z)z^j[w^j]H(w)^{-1}=H(z)h_jz^j p j = H ( z ) z j [ w j ] H ( w ) − 1 = H ( z ) h j z j with the positive coefficients h j = [ z j ] H − 1 h_j=[z^j]H^{-1} h j = [ z j ] H − 1 (h 0 = 1 h_0=1 h 0 = 1 , h 1 = 1 / 5 h_1=1/5 h 1 = 1/5 ), so p 0 = H ( z ) p_0=H(z) p 0 = H ( z ) and p 1 = H ( z ) z / 5 p_1=H(z)z/5 p 1 = H ( z ) z /5 , and p 0 p 1 = H ( z ) 2 z / 5 ≥ c z p_0p_1=H(z)^2z/5 \ge cz p 0 p 1 = H ( z ) 2 z /5 ≥ cz on 0 < z ≤ π 2 / 2 0<z\le\pi^2/2 0 < z ≤ π 2 /2 . Its characteristic function is φ 1 ( θ ) = H ( z ) / H ( z e i θ ) \varphi_1(\theta)=H(z)/H(ze^{i\theta}) φ 1 ( θ ) = H ( z ) / H ( z e i θ ) , which does not vanish for z ≤ π 2 / 2 z\le\pi^2/2 z ≤ π 2 /2 , and
log φ 1 ( θ ) = log H ( z ) + ∑ j ≥ 1 c j z j e i j θ , − log H ( z ) = ∑ j ≥ 1 c j z j , \log\varphi_1(\theta)=\log H(z)+\sum_{j\ge1}c_jz^je^{ij\theta},\qquad-\log H(z)=\sum_{j\ge1}c_jz^j, log φ 1 ( θ ) = log H ( z ) + j ≥ 1 ∑ c j z j e ij θ , − log H ( z ) = j ≥ 1 ∑ c j z j ,
where the series converges for ∣ z ∣ |z| ∣ z ∣ below the first zero of H H H , which exceeds π 2 / 2 \pi^2/2 π 2 /2 . Hence ∂ θ 3 log φ 1 ( θ ) = ∑ j ( i j ) 3 c j z j e i j θ \partial_\theta^3\log\varphi_1(\theta)=\sum_j(ij)^3c_jz^je^{ij\theta} ∂ θ 3 log φ 1 ( θ ) = ∑ j ( ij ) 3 c j z j e ij θ has modulus at most ∑ j j 3 ∣ c j ∣ z j ≤ C z \sum_jj^3|c_j|z^j\le Cz ∑ j j 3 ∣ c j ∣ z j ≤ C z , and Var 1 = ∑ j j 2 c j z j ≤ C z \operatorname{Var}_1=\sum_jj^2c_jz^j\le Cz Var 1 = ∑ j j 2 c j z j ≤ C z , uniformly for 0 < z ≤ π 2 / 2 0<z\le\pi^2/2 0 < z ≤ π 2 /2 . When m → ∞ m\to\infty m → ∞ and m / h → 0 m/h\to0 m / h → 0 , one has z ≍ m / h z\asymp m/h z ≍ m / h , so h z ≍ m → ∞ hz\asymp m\to\infty h z ≍ m → ∞ , and the mean of S h S_h S h is exactly h μ ( z ) = m h\mu(z)=m h μ ( z ) = m . Lemma F.2 therefore gives Pr ( S h = m ) ∼ ( 2 π h Var 1 ) − 1 / 2 ≥ c / h z ≥ c ′ / m + 1 \Pr(S_h=m)\sim(2\pi h\operatorname{Var}_1)^{-1/2}\ge c/\sqrt{hz}\ge c'/\sqrt{m+1} Pr ( S h = m ) ∼ ( 2 π h Var 1 ) − 1/2 ≥ c / h z ≥ c ′ / m + 1 . When m / h m/h m / h stays away from zero, the summand law lies in a compact tilted family with positive zero and one probabilities and variance bounded above and below; hence the sum has variance comparable to h h h . If m m m stays bounded, pass to a fixed integer m m m ; then z = 5 m / h + O m ( h − 2 ) z=5m/h+O_m(h^{-2}) z = 5 m / h + O m ( h − 2 ) , and A ~ m ( h ) = ( h / 5 ) m / m ! ( 1 + O m ( h − 1 ) ) \widetilde{A}_m(h)=(h/5)^m/m!(1+O_m(h^{-1})) A m ( h ) = ( h /5 ) m / m ! ( 1 + O m ( h − 1 )) , so the central mass tends to e − m m m / m ! > 0 e^{-m}m^m/m!>0 e − m m m / m ! > 0 . At m = 0 m=0 m = 0 it is one. These cases and a subsequence contradiction prove the asserted uniform lower bound. The upper bound is one, proving (F.7). □
Put
G g ( k ) = log J k J N g 2 ( N − k ) , N = 3 g − 2. G_g(k)=\log\frac{J_k}{J_Ng^{2(N-k)}},\qquad N=3g-2. G g ( k ) = log J N g 2 ( N − k ) J k , N = 3 g − 2.
Lemma F.4 (Logarithmic comparison up to the endpoints). There is a constant C C C such that
max 0 ≤ k ≤ N ∣ G g ( k ) − g f ( k / g ) ∣ ≤ C g . (376) \max_{0\le k\le N}|G_g(k)-gf(k/g)|\le C\sqrt{g}. \tag*{(376)} 0 ≤ k ≤ N max ∣ G g ( k ) − g f ( k / g ) ∣ ≤ C g . (376)
Proof. This follows from the model (159) , the comparison (160) and Lemma F.3. The passage to f f f is uniform, even at m = 0 m=0 m = 0 . To see this, use the exact model for the intersection number, J k m o d = 3 g Γ ( h ) 4 − g m ! A ~ m ( h ) / ( 2 k + 1 ) ! ! J_k^{\mathrm{mod}}=3^g\Gamma(h)4^{-g}m!\widetilde{A}_m(h)/(2k+1)!! J k mod = 3 g Γ ( h ) 4 − g m ! A m ( h ) / ( 2 k + 1 )!! from § 3.35 , with h = 2 g − 1 h=2g-1 h = 2 g − 1 and m = N − k m=N-k m = N − k . Its quotient by the top model value is m ! A ~ m ( h ) ( 2 N + 1 ) ! ! / ( 2 k + 1 ) ! ! m!\widetilde{A}_m(h)(2N+1)!!/(2k+1)!! m ! A m ( h ) ( 2 N + 1 )!! / ( 2 k + 1 )!! , so by Lemma F.3
G g m o d ( k ) : = log J k m o d J N m o d g 2 ( N − k ) = log m ! + Φ ( h , m ) + ∑ j = k + 1 N log ( 2 j + 1 ) − 2 m log g + O ( log ( g + 1 ) ) , G_g^{\mathrm{mod}}(k):=\log\frac{J_k^{\mathrm{mod}}}{J_N^{\mathrm{mod}}g^{2(N-k)}}=\log m!+\Phi(h,m)+\sum_{j=k+1}^{N}\log(2j+1)-2m\log g+O(\log(g+1)), G g mod ( k ) := log J N mod g 2 ( N − k ) J k mod = log m ! + Φ ( h , m ) + j = k + 1 ∑ N log ( 2 j + 1 ) − 2 m log g + O ( log ( g + 1 )) ,
where the saddle action is
Φ ( h , m ) = − h log H ( ζ ( m / h ) ) − m log ζ ( m / h ) . \Phi(h,m)=-h\log H(\zeta(m/h))-m\log\zeta(m/h). Φ ( h , m ) = − h log H ( ζ ( m / h )) − m log ζ ( m / h ) .
Since ζ ( m / h ) \zeta(m/h) ζ ( m / h ) is the stationary point of z ↦ − h log H ( z ) − m log z z\mapsto-h\log H(z)-m\log z z ↦ − h log H ( z ) − m log z , whose derivative is ( h μ ( z ) − m ) / z (h\mu(z)-m)/z ( h μ ( z ) − m ) / z by (165) , the envelope identity gives
∂ h Φ ( h , m ) = − log H ( ζ ( m / h ) ) , ∂ m Φ ( h , m ) = − log ζ ( m / h ) . \partial_h\Phi(h,m)=-\log H(\zeta(m/h)),\qquad\partial_m\Phi(h,m)=-\log\zeta(m/h). ∂ h Φ ( h , m ) = − log H ( ζ ( m / h )) , ∂ m Φ ( h , m ) = − log ζ ( m / h ) .
The first is bounded, since H H H decreases from $1$ to 3 / π 2 3/\pi^2 3/ π 2 on [ 0 , π 2 / 2 ] [0,\pi^2/2] [ 0 , π 2 /2 ] ; the second has only the integrable singularity log ζ ( u ) = log u + O ( 1 ) \log\zeta(u)=\log u+O(1) log ζ ( u ) = log u + O ( 1 ) at u = 0 u=0 u = 0 , since ζ ( u ) ≍ u \zeta(u)\asymp u ζ ( u ) ≍ u . Hence replacing ( h , m ) = ( 2 g − 1 , 3 g − 2 − k ) (h,m)=(2g-1,3g-2-k) ( h , m ) = ( 2 g − 1 , 3 g − 2 − k ) by ( 2 g , 3 g − k ) (2g,3g-k) ( 2 g , 3 g − k ) along a straight path, on which m ′ / h ′ m'/h' m ′ / h ′ stays in [ 0 , 3 / 2 ] [0,3/2] [ 0 , 3/2 ] , changes $\Phi by O ( 1 ) + ∫ ∣ log ζ ( m ′ / h ′ ) ∣ d m ′ = O ( log ( g + 1 ) ) O(1)+\int|\log\zeta(m'/h')|\,dm'=O(\log(g+1)) O ( 1 ) + ∫ ∣ log ζ ( m ′ / h ′ ) ∣ d m ′ = O ( log ( g + 1 )) , and Φ ( 2 g , 3 g − k ) = g [ − 2 log H ( z ) − ( 3 − α ) log z ] \Phi(2g,3g-k)=g[-2\log H(z)-(3-\alpha)\log z] Φ ( 2 g , 3 g − k ) = g [ − 2 log H ( z ) − ( 3 − α ) log z ] with α = k / g \alpha=k/g α = k / g and z = ζ ( ( 3 − α ) / 2 ) z=\zeta((3-\alpha)/2) z = ζ (( 3 − α ) /2 ) , the first two terms of g f ( α ) gf(\alpha) g f ( α ) in (169) . For the remaining terms, Stirling’s bound with logarithmic error and the convention 0 log 0 = 0 0\log0=0 0 log 0 = 0 give, uniformly in 0 ≤ k ≤ N 0\le k\le N 0 ≤ k ≤ N ,
log m ! = g ( 3 − α ) log g + g ( 3 − α ) log ( 3 − α ) − g ( 3 − α ) + O ( log ( g + 1 ) ) , \log m!=g(3-\alpha)\log g+g(3-\alpha)\log(3-\alpha)-g(3-\alpha)+O(\log(g+1)), log m ! = g ( 3 − α ) log g + g ( 3 − α ) log ( 3 − α ) − g ( 3 − α ) + O ( log ( g + 1 )) ,
∑ j = k + 1 N log ( 2 j + 1 ) = [ 3 g log g + 3 g log 6 − 3 g ] − [ g α log g + g α log ( 2 α ) − g α ] + O ( log ( g + 1 ) ) , \sum_{j=k+1}^{N}\log(2j+1)=[3g\log g+3g\log6-3g]-[g\alpha\log g+g\alpha\log(2\alpha)-g\alpha]+O(\log(g+1)), j = k + 1 ∑ N log ( 2 j + 1 ) = [ 3 g log g + 3 g log 6 − 3 g ] − [ g α log g + g α log ( 2 α ) − g α ] + O ( log ( g + 1 )) ,
− 2 m log g = − 2 g ( 3 − α ) log g + O ( log ( g + 1 ) ) , -2m\log g=-2g(3-\alpha)\log g+O(\log(g+1)), − 2 m log g = − 2 g ( 3 − α ) log g + O ( log ( g + 1 )) ,
the second line using log ( 2 n + 1 ) ! ! = n log ( 2 n ) − n + O ( log ( n + 1 ) ) \log(2n+1)!!=n\log(2n)-n+O(\log(n+1)) log ( 2 n + 1 )!! = n log ( 2 n ) − n + O ( log ( n + 1 )) at n = N n=N n = N and n = k n=k n = k . The coefficients of g log g g\log g g log g add up to ( 3 − α ) + ( 3 − α ) − 2 ( 3 − α ) = 0 (3-\alpha)+(3-\alpha)-2(3-\alpha)=0 ( 3 − α ) + ( 3 − α ) − 2 ( 3 − α ) = 0 , and the remaining terms are g [ ( 3 − α ) log ( 3 − α ) + 3 log 6 − α log ( 2 α ) − 2 ( 3 − α ) ] g[(3-\alpha)\log(3-\alpha)+3\log6-\alpha\log(2\alpha)-2(3-\alpha)] g [( 3 − α ) log ( 3 − α ) + 3 log 6 − α log ( 2 α ) − 2 ( 3 − α )] , which completes g f ( α ) gf(\alpha) g f ( α ) . Thus G g mod ( k ) = g f ( k / g ) + O ( log ( g + 1 ) ) G_g^{\operatorname{mod}}(k)=gf(k/g)+O(\log(g+1)) G g mod ( k ) = g f ( k / g ) + O ( log ( g + 1 )) uniformly. The two logarithmic errors between the true and the model values are at most C g C\sqrt{g} C g by (160) , including the top coefficient, which proves (F.8). □
Quantitative secants and the top endpoint
Concavity converts the logarithmic comparison into two-sided secant bounds for the adjacent ratios. Together with the exact top anchor, these give the rate for all k ≥ g 3 / 4 k\ge g^{3/4} k ≥ g 3/4 ; the remaining range k ≤ g 3 / 4 k\le g^{3/4} k ≤ g 3/4 is treated in the final part of the same proof.
Proof of Theorem F.1 . Secant bounds. The sequence G g G_g G g is concave by (162) . Write q k = J k / J k + 1 q_k=J_k/J_{k+1} q k = J k / J k + 1 . For an integer ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 with ℓ ≤ k / 2 \ell\le k/2 ℓ ≤ k /2 and k + ℓ ≤ N k+\ell\le N k + ℓ ≤ N , its secants imply
G g ( k ) − G g ( k − ℓ ) ℓ ≤ log ( q k / g 2 ) ≤ − G g ( k + ℓ ) − G g ( k ) ℓ . \frac{G_g(k)-G_g(k-\ell)}{\ell}\le\log(q_k/g^2)\le-\frac{G_g(k+\ell)-G_g(k)}{\ell}. ℓ G g ( k ) − G g ( k − ℓ ) ≤ log ( q k / g 2 ) ≤ − ℓ G g ( k + ℓ ) − G g ( k ) .
Equations (F.3) and (F.8) yield
∣ log q k g 2 q ( k / g ) ∣ ≤ C ( g ℓ + ℓ k ) . (377) \left|\log\frac{q_k}{g^2q(k/g)}\right|\le C\left(\frac{\sqrt{g}}{\ell}+\frac{\ell}{k}\right). \tag*{(377)} log g 2 q ( k / g ) q k ≤ C ( ℓ g + k ℓ ) . (377)
where ℓ ≤ k / 2 \ell\le k/2 ℓ ≤ k /2 ensures α − 1 ≤ 2 g / k \alpha^{-1}\le2g/k α − 1 ≤ 2 g / k on both secant intervals [ ( k − ℓ ) / g , k / g ] [(k-\ell)/g,k/g] [( k − ℓ ) / g , k / g ] and [ k / g , ( k + ℓ ) / g ] [k/g,(k+\ell)/g] [ k / g , ( k + ℓ ) / g ] , so that the second-derivative bound in (F.3) gives the secant error O ( ℓ / k ) O(\ell/k) O ( ℓ / k ) . If k + ℓ > N k+\ell>N k + ℓ > N , retain the lower secant and use q k ≤ q N − 1 q_k\le q_{N-1} q k ≤ q N − 1 for the upper bound. The exact top anchor (173) gives q N − 1 / g 2 = 12 / 5 + O ( g − 1 ) q_{N-1}/g^2=12/5+O(g^{-1}) q N − 1 / g 2 = 12/5 + O ( g − 1 ) . Since log q \log q log q is Lipschitz near three and 3 − k / g ≤ ( ℓ + 2 ) / g 3-k/g\le(\ell+2)/g 3 − k / g ≤ ( ℓ + 2 ) / g , this gives the same bound (F.9), with a changed constant.
Take ℓ = ⌊ g 5 / 8 ⌋ \ell=\lfloor g^{5/8}\rfloor ℓ = ⌊ g 5/8 ⌋ . For k ≥ g 3 / 4 k\ge g^{3/4} k ≥ g 3/4 it is at most k / 2 k/2 k /2 for large g g g , and (F.9) is O ( g − 1 / 8 ) O(g^{-1/8}) O ( g − 1/8 ) . For k ≤ ϵ g k\le\epsilon g k ≤ ϵ g , instead take ℓ = ⌊ g 3 / 4 ⌋ \ell=\lfloor g^{3/4}\rfloor ℓ = ⌊ g 3/4 ⌋ to obtain O ϵ ( g − 1 / 4 ) O_\epsilon(g^{-1/4}) O ϵ ( g − 1/4 ) .
From intersection numbers to coefficient ratios. These estimates for the intersection numbers translate into the statement of the theorem without losing the bounded-codimension range. Indeed, put m = N − k m=N-k m = N − k , h = 2 g − 1 h=2g-1 h = 2 g − 1 , and z g , k = t g , k 2 / 2 z_{g,k}=t_{g,k}^2/2 z g , k = t g , k 2 /2 . By (166) , μ ( z g , k ) = m / h \mu(z_{g,k})=m/h μ ( z g , k ) = m / h . Therefore the value of q k = J k / J k + 1 q_k=J_k/J_{k+1} q k = J k / J k + 1 that corresponds, through (158) , to the proposed equivalent 4 π 2 / t g , k 2 4\pi^2/t_{g,k}^2 4 π 2 / t g , k 2 of a g , k / a g , k + 1 a_{g,k}/a_{g,k+1} a g , k / a g , k + 1 is
( 2 k + 3 ) m z g , k = ( 2 k + 3 ) h χ ( m / h ) = g 2 q ( k / g ) ( 1 + O ( k − 1 + g − 1 ) ) . (378) \frac{(2k+3)m}{z_{g,k}}=(2k+3)h\chi(m/h)=g^2q(k/g)\left(1+O(k^{-1}+g^{-1})\right). \tag*{(378)} z g , k ( 2 k + 3 ) m = ( 2 k + 3 ) h χ ( m / h ) = g 2 q ( k / g ) ( 1 + O ( k − 1 + g − 1 ) ) . (378)
The last error is uniform for k ≥ 1 k\ge1 k ≥ 1 : the two arguments of χ \chi χ differ by O ( g − 1 ) O(g^{-1}) O ( g − 1 ) , while χ \chi χ is positive with bounded logarithmic derivative on the closed interval. This calculation retains m m m exactly, including m = 1 m=1 m = 1 . By (158) , the two secant estimates prove (F.1) for k ≥ g 3 / 4 k\ge g^{3/4} k ≥ g 3/4 and all of (F.2).
The small-degree range. It remains to consider k ≤ g 3 / 4 k\le g^{3/4} k ≤ g 3/4 . Put ρ k = b g , k / b g , k + 1 ≥ 1 \rho_k=b_{g,k}/b_{g,k+1}\ge1 ρ k = b g , k / b g , k + 1 ≥ 1 and ω k = ( 2 k + 3 ) ( N − k ) \omega_k=(2k+3)(N-k) ω k = ( 2 k + 3 ) ( N − k ) as before. Equation (180) gives ρ k = 1 + O ( g − 1 / 2 ) \rho_k=1+O(g^{-1/2}) ρ k = 1 + O ( g − 1/2 ) for k ≤ δ 0 g k\le\delta_0\sqrt{g} k ≤ δ 0 g with a fixed small δ 0 > 0 \delta_0 > 0 δ 0 > 0 . For the remaining indices we claim
0 ≤ ρ k − 1 ≤ C ( log ( g + 1 ) k + k g + 1 k ) (379) 0 \le\rho_k - 1 \le C\left(\frac{\log(g + 1)}{\sqrt{k}} + \sqrt{\frac{k}{g} + \frac{1}{k}}\right) \tag*{(379)} 0 ≤ ρ k − 1 ≤ C ( k log ( g + 1 ) + g k + k 1 ) (379)
First ρ k < 2 \rho_k < 2 ρ k < 2 uniformly in this range for large g g g . Otherwise monotonicity of q j q_j q j and ω j / ω k ≤ 1 + δ \omega_j/\omega_k \le1 + \delta ω j / ω k ≤ 1 + δ propagate a fixed ratio gap over a block of length comparable to k k k : with δ = 1 / 3 \delta= 1/3 δ = 1/3 , for k ≤ j ≤ 4 k / 3 k \le j \le4k/3 k ≤ j ≤ 4 k /3 one has ω j / ω k ≤ ( 2 j + 3 ) / ( 2 k + 3 ) ≤ 4 / 3 \omega_j/\omega_k \le(2j + 3)/(2k + 3) \le4/3 ω j / ω k ≤ ( 2 j + 3 ) / ( 2 k + 3 ) ≤ 4/3 , hence ρ j ≥ ρ k ω k / ω j ≥ 3 / 2 \rho_j \ge\rho_k\omega_k/\omega_j \ge3/2 ρ j ≥ ρ k ω k / ω j ≥ 3/2 on a block of at least k / 3 k/3 k /3 indices. Its product would force log ( V g , 1 / b g , J + 1 ) ≥ k 3 log 3 2 \log(V_{g,1}/b_{g,J+1}) \ge\frac{k}{3}\log\frac{3}{2} log ( V g , 1 / b g , J + 1 ) ≥ 3 k log 2 3 at J = ⌊ 4 k / 3 ⌋ J = \lfloor4k/3\rfloor J = ⌊ 4 k /3 ⌋ , contradicting (178) at J + 1 ≤ 2 k J + 1 \le2k J + 1 ≤ 2 k , whose upper bound C log 2 ( g + 1 ) + C k 2 / g C\log^2(g + 1) + Ck^2/g C log 2 ( g + 1 ) + C k 2 / g is o ( k ) o(k) o ( k ) uniformly on this range.
Now put η = ρ k − 1 ∈ [ 0 , 1 ] \eta= \rho_k - 1 \in[0,1] η = ρ k − 1 ∈ [ 0 , 1 ] and δ = η / 6 \delta= \eta/6 δ = η /6 . For k ≤ j ≤ ⌊ ( 1 + δ ) k ⌋ k \le j \le\lfloor(1 + \delta)k\rfloor k ≤ j ≤ ⌊( 1 + δ ) k ⌋ , the same monotonicity and the exact factor ω j \omega_j ω j give ρ j ≥ ( 1 + η ) / ( 1 + δ ) ≥ 1 + η / 2 \rho_j \ge(1 + \eta)/(1 + \delta) \ge1 + \eta/2 ρ j ≥ ( 1 + η ) / ( 1 + δ ) ≥ 1 + η /2 . This block has ⌊ η k / 6 ⌋ + 1 ≥ η k / 6 \lfloor\eta k/6\rfloor+ 1 \ge\eta k/6 ⌊ η k /6 ⌋ + 1 ≥ η k /6 indices. Summing logarithms over the block and using (178) at its last index plus one gives
c η 2 k ≤ C log 2 ( g + 1 ) + C k 2 / g . c\eta^2 k \le C\log^2(g + 1) + Ck^2/g. c η 2 k ≤ C log 2 ( g + 1 ) + C k 2 / g .
All these indices lie below the fixed c 0 g c_0g c 0 g range of (178) for large g g g . This proves (F.11). Its right side is O ( g − 1 / 8 ) O(g^{-1/8}) O ( g − 1/8 ) on the displayed range. Thus ρ k = 1 + O ( g − 1 / 8 ) \rho_k = 1 + O(g^{-1/8}) ρ k = 1 + O ( g − 1/8 ) uniformly for 0 ≤ k ≤ g 3 / 4 0 \le k \le g^{3/4} 0 ≤ k ≤ g 3/4 .
Finally the saddle equation at t = π t = \pi t = π gives π 2 / t g , k 2 = 1 + O ( ( k + 1 ) / g ) \pi^2/t_{g,k}^2 = 1 + O((k+1)/g) π 2 / t g , k 2 = 1 + O (( k + 1 ) / g ) uniformly on this small-degree range: by (165) –(166) , t = t g , k t = t_{g,k} t = t g , k solves R ( t ) = α g , k = ( 2 k + 1 ) / ( 2 g − 1 ) R(t) = \alpha_{g,k} = (2k + 1)/(2g - 1) R ( t ) = α g , k = ( 2 k + 1 ) / ( 2 g − 1 ) with R ( t ) = t 2 sin t / ( sin t − t cos t ) R(t) = t^2\sin t/(\sin t - t\cos t) R ( t ) = t 2 sin t / ( sin t − t cos t ) , and R ( π ) = 0 R(\pi) = 0 R ( π ) = 0 , R ′ ( π ) = − π R'(\pi) = -\pi R ′ ( π ) = − π , so π − t g , k = α g , k / π + O ( α g , k 2 ) \pi- t_{g,k} = \alpha_{g,k}/\pi+ O(\alpha_{g,k}^2) π − t g , k = α g , k / π + O ( α g , k 2 ) and π 2 / t g , k 2 = 1 + 2 α g , k / π 2 + O ( α g , k 2 ) \pi^2/t_{g,k}^2 = 1 + 2\alpha_{g,k}/\pi^2 + O(\alpha_{g,k}^2) π 2 / t g , k 2 = 1 + 2 α g , k / π 2 + O ( α g , k 2 ) . Since a g , k / a g , k + 1 = 4 ρ k a_{g,k}/a_{g,k+1} = 4\rho_k a g , k / a g , k + 1 = 4 ρ k , the desired relative error is O ( g − 1 / 8 ) O(g^{-1/8}) O ( g − 1/8 ) there as well. Together with the secant range this proves both assertions. □
The argument is a quantitative version of the one used for the leading equivalent, based on the same logarithmic comparison and concavity. Liu–Xu’s fixed-codimension expansion has finer errors at each fixed codimension [129]; the bounds here hold with a single constant across all valid degrees.
Appendices to Part 2: spectral laws
Proof of the short-curve theorem
This appendix proves Theorem 14.2 by the method of moments, following the outline at the end of Subsection 14.2. The proof is a single argument in three steps. Step 1 derives a coefficient-ratio bound, (G.1), from the positive convolution factor (108) and Lemma 9.3 ; it shows that curves cutting off many cusps form a uniformly summable tail, and it is reused in the proofs of Lemmas L.1 and L.2 (Appendix § 10.20 ). Step 2 computes the mixed factorial moments for finitely many fixed cusp types: the disjoint configurations give the Poisson moments, while nested configurations are negligible. Step 3 shows that, on the same length scale, all short curves other than planar separating ones are rare, which identifies the full primitive process and the systole. The volume inputs are Mirzakhani’s integration formula (187) , the boundary estimate (188) , the volume ratio (189) obtained from Theorem 1.3 , and the coefficient identity (124) ; we use the notation of Subsection 14.2, in particular the intensity q g q_g q g in (182) and the weights (184) .
Proof of Theorem 14.2 . Write P = 2 π 2 P = 2\pi^2 P = 2 π 2 , h = 2 g − 1 h = 2g - 1 h = 2 g − 1 , m = n − 1 m = n - 1 m = n − 1 , and F h = Φ g ′ F_h = \Phi'_g F h = Φ g ′ . Choose x ∈ ( 0 , x c ) x \in(0,x_c) x ∈ ( 0 , x c ) by h Ψ ( x ) = m h\Psi(x) = m h Ψ ( x ) = m and put q g ∗ = n x / ( 4 P ) q_g^* = nx/(4P) q g ∗ = n x / ( 4 P ) . By (7) , C ( λ ) = π 2 λ / X ( Y λ ) \mathcal{C}(\lambda) = \pi^2\lambda/X(Y_\lambda) C ( λ ) = π 2 λ / X ( Y λ ) , so q g = ( n − 1 ) X ( Y n / α ) / ( 8 π 2 ) q_g = (n - 1)X(Y_{n/\alpha})/(8\pi^2) q g = ( n − 1 ) X ( Y n / α ) / ( 8 π 2 ) and q g ∗ = n X ( Y λ e f f ) / ( 8 π 2 ) q_g^* = nX(Y_{\lambda_{\mathrm{eff}}})/(8\pi^2) q g ∗ = n X ( Y λ eff ) / ( 8 π 2 ) with the effective density λ e f f = 2 ( n − 1 ) / ( 2 g − 1 ) \lambda_{\mathrm{eff}} = 2(n - 1)/(2g - 1) λ eff = 2 ( n − 1 ) / ( 2 g − 1 ) . Since λ e f f g / n → 1 \lambda_{\mathrm{eff}}g/n \to1 λ eff g / n → 1 and λ ↦ X ( Y λ ) \lambda\mapsto X(Y_\lambda) λ ↦ X ( Y λ ) is continuous and positive on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , with X ( Y λ ) ∼ λ / 4 X(Y_\lambda) \sim\lambda/4 X ( Y λ ) ∼ λ /4 as λ → 0 \lambda\to0 λ → 0 and X ( Y λ ) → x c X(Y_\lambda) \to x_c X ( Y λ ) → x c as λ → ∞ \lambda\to\infty λ → ∞ , a subsequence argument gives q g ∗ / q g → 1 q_g^*/q_g \to1 q g ∗ / q g → 1 ; and q g ≍ n 2 / ( g + n ) → ∞ q_g \asymp n^2/(g + n) \to\infty q g ≍ n 2 / ( g + n ) → ∞ by Subsection 14.2.
Step 1: a uniform tail bound over cusp types. We first show that curves cutting off a large number of cusps contribute negligibly, uniformly along the sequence. The input is the coefficient-ratio bound (G.1), which compares volumes whose cusp counts differ by j j j . By the positive convolution (108) and Lemma 9.3, every point mass is at most C / σ C/\sigma C / σ , where σ 2 = h H Ψ ( x ) \sigma^2=h\mathcal{H}\Psi(x) σ 2 = h H Ψ ( x ) . The local limit (114) gives a lower bound c / σ c/\sigma c / σ at m m m . Consequently, for all sufficiently large h , m h,m h , m ,
[ z m − j ] F h ( z ) [ z m ] F h ( z ) = x j P x ( N = m − j ) P x ( N = m ) ≤ C x j ( 0 ≤ j ≤ m ) . (380) \frac{[z^{m-j}]F_h(z)}{[z^m]F_h(z)}=x^j\frac{\mathbb{P}_x(N=m-j)}{\mathbb{P}_x(N=m)}\leq Cx^j\qquad(0\leq j\leq m). \tag*{(380)} [ z m ] F h ( z ) [ z m − j ] F h ( z ) = x j P x ( N = m ) P x ( N = m − j ) ≤ C x j ( 0 ≤ j ≤ m ) . (380)
Only the denominator uses the local limit; the numerator may be at any index, including zero.
Let N k , L N_{k,L} N k , L count curves of type k k k with length at most L L L , and put a k = V 0 , k + 1 / ( k ! P k − 2 ) a_k=V_{0,k+1}/(k!P^{k-2}) a k = V 0 , k + 1 / ( k ! P k − 2 ) . Mirzakhani integration gives
E N k , L = ( n k ) V g , n ∫ 0 L ℓ V 0 , k + 1 ( ℓ , 0 k ) V g , n − k + 1 ( ℓ , 0 n − k ) d ℓ , \mathbb{E}N_{k,L}=\frac{\binom{n}{k}}{V_{g,n}}\int_0^L\ell V_{0,k+1}(\ell,0^k)V_{g,n-k+1}(\ell,0^{n-k})\,\mathrm{d}\ell, E N k , L = V g , n ( k n ) ∫ 0 L ℓ V 0 , k + 1 ( ℓ , 0 k ) V g , n − k + 1 ( ℓ , 0 n − k ) d ℓ ,
( n k ) V 0 , k + 1 V g , n − k + 1 V g , n = n P a k [ z n − k ] F h [ z n − 1 ] F h . (381) \binom{n}{k}\frac{V_{0,k+1}V_{g,n-k+1}}{V_{g,n}}=\frac{n}{P}a_k\frac{[z^{n-k}]F_h}{[z^{n-1}]F_h}. \tag*{(381)} ( k n ) V g , n V 0 , k + 1 V g , n − k + 1 = P n a k [ z n − 1 ] F h [ z n − k ] F h . (381)
The two sides have different genera, so there is no side-exchange factor. The coefficient identity follows from (124) ; it holds up to k = n k=n k = n . Bound (188) changes the integrand’s two volumes by a factor between one and e L e^L e L , independently of k k k . Using (G.1) with j = k − 1 j=k-1 j = k − 1 therefore gives, for each fixed T T T ,
E ∑ k ≥ K N k , T q g ∗ ≤ C T 2 ∑ k ≥ K a k x c k − 2 ⟶ 0 ( K → ∞ ) , (382) \mathbb{E}\sum_{k\geq K}\frac{N_{k,T}}{\sqrt{q_g^*}}\leq CT^2\sum_{k\geq K}a_kx_c^{k-2}\longrightarrow0\qquad(K\to\infty), \tag*{(382)} E k ≥ K ∑ q g ∗ N k , T ≤ C T 2 k ≥ K ∑ a k x c k − 2 ⟶ 0 ( K → ∞ ) , (382)
uniformly along the sequence. To justify summability directly, apply the string-shift identity (27) at ( g , n ) = ( 0 , 3 ) (g,n)=(0,3) ( g , n ) = ( 0 , 3 ) : by (25) with h = 1 h=1 h = 1 and m = 0 m=0 m = 0 , Φ 0 ′ ′ ( x ) = U 0 , 3 Y ( 0 , 0 , 0 ) = ⟨ τ 0 3 ⟩ 0 / a ( Y ) = 1 / a ( Y ( x ) ) = Y ′ ( x ) \Phi_0''(x)=U^Y_{0,3}(0,0,0)=\langle\tau_0^3\rangle_0/a(Y)=1/a(Y(x))=Y'(x) Φ 0 ′′ ( x ) = U 0 , 3 Y ( 0 , 0 , 0 ) = ⟨ τ 0 3 ⟩ 0 / a ( Y ) = 1/ a ( Y ( x )) = Y ′ ( x ) . Since Φ 0 ( x ) = x 3 / 6 + O ( x 4 ) \Phi_0(x)=x^3/6+O(x^4) Φ 0 ( x ) = x 3 /6 + O ( x 4 ) has no unstable terms, Φ 0 ′ ′ ( 0 ) = Φ 0 ′ ( 0 ) = 0 \Phi_0''(0)=\Phi_0'(0)=0 Φ 0 ′′ ( 0 ) = Φ 0 ′ ( 0 ) = 0 ; with Y ( 0 ) = 0 Y(0)=0 Y ( 0 ) = 0 this gives Φ 0 ′ ′ = Y \Phi_0''=Y Φ 0 ′′ = Y and Φ 0 ′ ( x ) = ∫ 0 x Y ( u ) d u \Phi_0'(x)=\int_0^xY(u)\,\mathrm{d}u Φ 0 ′ ( x ) = ∫ 0 x Y ( u ) d u . This is finite at x c x_c x c ; positivity gives convergence of the coefficient series there. Its initial term x 2 / 2 x^2/2 x 2 /2 also gives W ( 0 ) = 1 W(0)=1 W ( 0 ) = 1 .
Step 2: fixed types and nested configurations. Next we compute the mixed factorial moments for finitely many fixed cusp types. The main contribution comes from configurations whose planar sides are pairwise disjoint, which we call stars ; nested configurations are shown to be negligible. Fix r r r ordered distinct curves of prescribed types k 1 , … , k r k_1,\ldots,k_r k 1 , … , k r , and write K = ∑ k i K=\sum k_i K = ∑ k i . For large g g g all rescaled windows lie below 2 arcsinh 1 2\operatorname{arcsinh}1 2 arcsinh 1 , so the collar lemma makes their boundaries disjoint. For a star, the ordered cusp-label multiplicity is ( n ) K / ∏ i k i ! (n)_K/\prod_i k_i! ( n ) K / ∏ i k i ! , and the complement has type ( g , n − K + r ) (g,n-K+r) ( g , n − K + r ) . There is no r ! r! r ! divisor. The boundary error is 1 + O r ( L ) 1+O_r(L) 1 + O r ( L ) , as each cut occurs twice in (188) . For fixed d d d , (189) and q g ∗ / q g → 1 q_g^*/q_g\to1 q g ∗ / q g → 1 give V g , n − d / V g , n = ( x / ( P n ) ) d ( 1 + o ( 1 ) ) V_{g,n-d}/V_{g,n}=(x/(Pn))^d(1+o(1)) V g , n − d / V g , n = ( x / ( P n ) ) d ( 1 + o ( 1 )) . Thus for windows [ a i , b i ] / q g ∗ [a_i,b_i]/\sqrt{q_g^*} [ a i , b i ] / q g ∗ the star contribution tends to
∏ i = 1 r w k i ( x ) ( b i 2 − a i 2 ) . (383) \prod_{i=1}^r w_{k_i}(x)(b_i^2-a_i^2). \tag*{(383)} i = 1 ∏ r w k i ( x ) ( b i 2 − a i 2 ) . (383)
Before the limit is taken, the star contribution equals the right side at x = x ( g , n ) x=x(g,n) x = x ( g , n ) up to a factor 1 + o ( 1 ) 1+o(1) 1 + o ( 1 ) ; the error is relative, so the statement keeps its meaning when x → 0 x\to0 x → 0 and w k ( x ) → 0 w_k(x)\to0 w k ( x ) → 0 for k > 2 k>2 k > 2 .
Filling cusps turns each planar side into a disk. Since each curve separates and g ≥ 1 g\geq1 g ≥ 1 , disjoint boundaries therefore have nested or disjoint disks; their cut graph is a tree with one genus-g g g root and r r r planar vertices. Let d d d be the root degree, a v a_v a v the number of cusps at a planar vertex and d v d_v d v its degree; put a = ∑ v a v a=\sum_v a_v a = ∑ v a v . Prescribed types bound a a a by K K K . There are finitely many such trees with ordered edges and cusp allocations, depending only on the fixed types. Classification of surfaces gives the corresponding mapping-class orbits with bounded symmetry factors; assigning cusp labels costs at most C n a Cn^a C n a . The root volume ratio is O ( ( x / ( P n ) ) a − d ) O((x/(Pn))^{a-d}) O (( x / ( P n ) ) a − d ) , the planar volumes are fixed, and the r r r length integrals cost O ( ( n x ) − r ) O((nx)^{-r}) O (( n x ) − r ) . Each cut type contributes at most
C n d − r x a − d − r = O ( n − 1 ) if it is not a star . (384) Cn^{d-r}x^{a-d-r}=O(n^{-1})\quad\text{if it is not a star}. \tag*{(384)} C n d − r x a − d − r = O ( n − 1 ) if it is not a star . (384)
Indeed stability gives a v + d v ≥ 3 a_v+d_v \ge3 a v + d v ≥ 3 and ∑ v d v = 2 r − d \sum_v d_v=2r-d ∑ v d v = 2 r − d , hence a ≥ r + d a \ge r+d a ≥ r + d . The power of x x x is nonnegative, so no loss occurs as x → 0 x \to0 x → 0 . A nonstar has d ≤ r − 1 d \le r-1 d ≤ r − 1 ; if d = r d=r d = r , every planar vertex is a leaf and the configuration is a star. This accounts for every fixed-type mixed factorial moment.
The limits (G.4) are the mixed factorial moments of independent Poisson counts. Higher fixed moments give uniform integrability, and the product Poisson law is moment determinate. First moments give tightness on compact sets in the marked space. This proves (185) . Its projection onto the length coordinate requires more than vague convergence: truncate the types, pass to the finite-type limit, then let the cutoff grow using (G.3) and the convergent limiting series. This proves the stated unmarked process. Finally replace q g b q_g^b q g b by q g q_g q g . The type-two restriction gives (183) without assuming a limit for n / g n/g n / g , by subsequences and compactness of [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] . Its zero-count probabilities give the pants minimum law.
Step 3: the remaining curves and the systole. It remains to show that, on the scale q g − 1 / 2 q_g^{-1/2} q g − 1/2 , there are no short curves other than planar separating ones: nonseparating curves, and separating curves with positive genus on both sides, are rare. Nonseparating simple curves of length at most L ≤ L 0 L \le L_0 L ≤ L 0 have expected count at most C e L L 2 V g − 1 , n + 2 / V g , n ≤ C L 2 Ce^L L^2V_{g-1,n+2}/V_{g,n} \le CL^2 C e L L 2 V g − 1 , n + 2 / V g , n ≤ C L 2 , by integration, (188) , and the uniform lower bound for B g , n B_{g,n} B g , n . To sum every positive-genus separating cut, put S g ( z ) = ∑ i = 1 g − 1 Φ i ( z ) Φ g − i ′ ( z ) S_g(z)=\sum_{i=1}^{g-1}\Phi_i(z)\Phi'_{g-i}(z) S g ( z ) = ∑ i = 1 g − 1 Φ i ( z ) Φ g − i ′ ( z ) . Exact positivity (108) includes genus one, so the child powers add to h − 1 h-1 h − 1 :
S g ( z ) = V ( z ) h − 1 Q g ( z ) , Q g ≥ 0 coefficientwise , S g ( x ) F h ( x ) ≤ C a ( Y ) r ( Y ) 2 g 2 . (385) S_g(z)=V(z)^{h-1}Q_g(z),\qquad Q_g\ge0\ \text{coefficientwise},\qquad\frac{S_g(x)}{F_h(x)}\le\frac{C a(Y)r(Y)^2}{g^2}. \tag*{(385)} S g ( z ) = V ( z ) h − 1 Q g ( z ) , Q g ≥ 0 coefficientwise , F h ( x ) S g ( x ) ≤ g 2 C a ( Y ) r ( Y ) 2 . (385)
The last bound is (81) : undoing (64) leaves exactly a r 2 ar^2 a r 2 , as the child complexities sum to 2 g − 2 2g-2 2 g − 2 and their deficits to 3 g − 4 3g-4 3 g − 4 . In particular bounded child genera are retained. Lemma 9.3, applied to the positive factor V h − 1 V^{h-1} V h − 1 , gives sup j P X S ( N = j ) ≤ C / σ \sup_j\mathbb{P}^{S}_X(N=j)\le C/\sigma sup j P X S ( N = j ) ≤ C / σ . Against the parent lower bound P X F ( N = n − 1 ) ≥ c / σ \mathbb{P}^{F}_X(N=n-1)\ge c/\sigma P X F ( N = n − 1 ) ≥ c / σ , the standard deviations cancel. The coefficient identity (124) therefore yields
A g , n : = ∑ i = 1 g − 1 ∑ k = 0 n ( n k ) V i , k + 1 V g − i , n − k + 1 V g , n = [ z n ] S g ( z ) [ z n ] Φ g ( z ) = n P X F h ( x ) S g ( x ) F h ( x ) P X S ( N = n ) P X F ( N = n − 1 ) ≤ C n a r 2 x g 2 ≤ C g . (386) \begin{aligned}
A_{g,n}:=\sum_{i=1}^{g-1}\sum_{k=0}^{n}\binom{n}{k}\frac{V_{i,k+1}V_{g-i,n-k+1}}{V_{g,n}}=\frac{[z^n]S_g(z)}{[z^n]\Phi_g(z)} \\
&=\frac{n}{P_XF_h(x)}\frac{S_g(x)}{F_h(x)}\frac{\mathbb{P}^{S}_X(N=n)}{\mathbb{P}^{F}_X(N=n-1)}
\le\frac{Cnar^2}{xg^2}\le\frac{C}{g}. \tag*{(386)}
\end{aligned} A g , n := i = 1 ∑ g − 1 k = 0 ∑ n ( k n ) V g , n V i , k + 1 V g − i , n − k + 1 = [ z n ] Φ g ( z ) [ z n ] S g ( z ) = P X F h ( x ) n F h ( x ) S g ( x ) P X F ( N = n − 1 ) P X S ( N = n ) ≤ x g 2 C na r 2 ≤ g C . (386)
Indeed ( n − 1 ) / x = h / ( A a r ) (n-1)/x=h/(Aar) ( n − 1 ) / x = h / ( A a r ) , while A A A is bounded below and r r r above. Thus no small-density factor is lost. Each separating orbit with two positive genera occurs in this ordered sum. Its integration symmetry and one-handle factors are at most one; applying (188) to both components bounds its total expected count by C e L L 2 A g , n ≤ C L 2 / g Ce^L L^2A_{g,n}\le CL^2/g C e L L 2 A g , n ≤ C L 2 / g . This includes all cusp partitions.
For fixed T T T , take L = T / q g L=T/\sqrt{q_g} L = T / q g . Markov’s inequality shows that with probability 1 − O ( T 2 / q g ) 1-O(T^2/q_g) 1 − O ( T 2 / q g ) every simple geodesic in this window is planar. Primitive geodesics shorter than 2 arcsinh 1 2\operatorname{arcsinh}1 2 arcsinh 1 are simple, by the standard collar theorem used in [99, Section 1.1]. Since q g → ∞ q_g\to\infty q g → ∞ , the full primitive process and planar process agree on each bounded rescaled window with probability tending to one. This transfers the process limit and its zero-count law to the systole.
Finally restrict the first point to [ n , T ] [n,T] [ n , T ] and finitely many cusp types. The first-point map is continuous when its first length is unique and avoids the endpoints, as happens almost surely in the limiting Poisson process. Let the type cutoff grow using (G.3), then η ↓ 0 \eta\downarrow0 η ↓ 0 and T → ∞ T\to\infty T → ∞ . The omitted length probabilities tend to zero by the zero-count law; W ≥ 1 W\ge1 W ≥ 1 . Nonplanar competitors are excluded on each fixed window. This also makes any tie rule immaterial and proves planarity of every systolic curve with probability tending to one. The first-arrival density of the limiting marked process is (186) .
Proofs for the tight length spectrum
This appendix proves Theorem 14.8 and Corollary 14.9 , following the outline in Subsection 14.5. The scale α \alpha α and the intensity ν \nu ν are those of (200) . The tight integration formula of Budd and Lions [32, Proposition 3.2] expresses factorial moments of tight curve counts through tight volumes of the cut surfaces; the inputs from Part 1 are the shifted amplitudes of § 3.4 , the coefficient extraction of § 3.25 (Lemma 9.3 and the estimates (6.16)–(6.21)), and the prefactor of Proposition 7.1. Subsection H.1 expresses the generating functions of tight volumes through the shifted amplitudes (Lemma H.1 ); Subsection H.2 extracts the coefficient at the prescribed cusp count for connected complements (Lemma H.2); Subsection H.3 shows that disconnected complements are negligible (Lemma H.3); Subsection H.4 identifies the Poisson limit, and Subsection H.5 averages over the random cusp count to obtain the corollary.
Tight volume generating functions
This subsection proves Lemma H.1 , which identifies the tight generating functions with sums of the shifted amplitudes U g , p Y U^Y_{g,p} U g , p Y of (25) , compares them with the ordinary marked generating functions, and records their positive convolution structure. The identity comes from the tight recursion of Budd and Zonneveld [34]; the comparison and positivity come from the factorial series (70) , (78) and the positive factor (108) . Both are used in the next two subsections: the comparison transfers local limits from the marked functions to the tight ones (proof of Lemma H.2), and positivity supplies the point-mass bounds (proof of Lemma H.3). Put P = 2 π 2 P = 2\pi^2 P = 2 π 2 . Let T g , p , k ( L ) T_{g,p,k}(\mathbf{L}) T g , p , k ( L ) be the tight volume with p p p protected geodesic boundaries and k k k cusps that may be filled, in the conventions of [30, 34, 32]. For stable ( g , p ) (g,p) ( g , p ) define
T g , p ( x ; L ) = ∑ k ≥ 0 T g , p , k ( L ) P 3 g − 3 + p + k x k k ! , F g , p = Φ g ( p ) . \mathcal{T}_{g,p}(x;\mathbf{L})=\sum_{k\geq0}\frac{T_{g,p,k}(\mathbf{L})}{P^{3g-3+p+k}}\frac{x^k}{k!},\qquad F_{g,p}=\Phi_g^{(p)}. T g , p ( x ; L ) = k ≥ 0 ∑ P 3 g − 3 + p + k T g , p , k ( L ) k ! x k , F g , p = Φ g ( p ) .
Unmarked functions are used only for g ≥ 2 g\geq2 g ≥ 2 . Set ϱ ( Y ) = r ( Y ) / P \varrho(Y)=\sqrt{r(Y)/P} ϱ ( Y ) = r ( Y ) / P and S ( t ) = sinh ( t ) / t S(t)=\sinh(t)/t S ( t ) = sinh ( t ) / t , with S ( 0 ) = 1 S(0)=1 S ( 0 ) = 1 .
Lemma H.1 (Tight generating functions). Let ( g , p ) (g,p) ( g , p ) be stable and 0 ≤ Y < y 0 0\leq Y<y_0 0 ≤ Y < y 0 . The tight generating function is expressed through the shifted amplitudes by the exact identity
T g , p ( X ( Y ) ; L ) = ∑ d U g , p Y ( d ) ∏ i = 1 p L i 2 d i P d i ( 2 d i + 1 ) ! , (387) \mathcal{T}_{g,p}(X(Y);\mathbf{L})=\sum_{\mathbf d}U^Y_{g,p}(\mathbf d)\prod_{i=1}^{p}\frac{L_i^{2d_i}}{P^{d_i}(2d_i+1)!}, \tag*{(387)} T g , p ( X ( Y ) ; L ) = d ∑ U g , p Y ( d ) i = 1 ∏ p P d i ( 2 d i + 1 )! L i 2 d i , (387) which fixes both the normalization and the meaning of the protected boundaries. Its ratio to the ordinary marked generating function satisfies
1 ≤ T g , p ( X ( Y ) ; L ) F g , p ( X ( Y ) ) ≤ ∏ i = 1 p S ( ϱ ( Y ) L i ) , 1\leq\frac{\mathcal{T}_{g,p}(X(Y);\mathbf{L})}{F_{g,p}(X(Y))}\leq\prod_{i=1}^{p}S(\varrho(Y)L_i), 1 ≤ F g , p ( X ( Y )) T g , p ( X ( Y ) ; L ) ≤ i = 1 ∏ p S ( ϱ ( Y ) L i ) , T g , p ( X ( Y ) ; L ) F g , p ( X ( Y ) ) = ∏ i = 1 p S ( ϱ ( Y ) L i ) ( 1 + O p , B ( g − 1 ) ) if max i ϱ ( Y ) L i ≤ B , (388) \frac{\mathcal{T}_{g,p}(X(Y);\mathbf{L})}{F_{g,p}(X(Y))}
=\prod_{i=1}^{p}S(\varrho(Y)L_i)\left(1+O_{p,B}(g^{-1})\right)
\quad\text{if }\max_i\varrho(Y)L_i\leq B, \tag*{(388)} F g , p ( X ( Y )) T g , p ( X ( Y ) ; L ) = i = 1 ∏ p S ( ϱ ( Y ) L i ) ( 1 + O p , B ( g − 1 ) ) if i max ϱ ( Y ) L i ≤ B , (388) where the inequality holds at every stable genus. Finally, for each fixed physical length vector,
T g , p ( x ; L ) = V ( x ) 2 g − 2 + p R g , p , L ( x ) , [ x j ] R g , p , L ( x ) ≥ 0 , (389) \mathcal{T}_{g,p}(x;\mathbf{L})=V(x)^{2g-2+p}R_{g,p,\mathbf{L}}(x),\qquad[x^j]R_{g,p,\mathbf{L}}(x)\geq0, \tag*{(389)} T g , p ( x ; L ) = V ( x ) 2 g − 2 + p R g , p , L ( x ) , [ x j ] R g , p , L ( x ) ≥ 0 , (389) and the series converge for 0 ≤ x < x c 0\leq x<x_c 0 ≤ x < x c .
Proof. The identity (H.1) follows from [34, Proposition 12]. Indeed, their root and moments are R = Y / P R=Y/P R = Y / P and M j = P j X ( j + 1 ) ( Y ) M_j=P^jX^{(j+1)}(Y) M j = P j X ( j + 1 ) ( Y ) . With q = P / 2 q=P/2 q = P /2 , their shifted times are q k − 1 t k ′ q^{k-1}t'_k q k − 1 t k ′ , where the t k ′ t'_k t k ′ are the times in (27) . The dimension constraint gives the factor q 3 g − 3 + p − ∣ d ∣ q^{3g-3+p-|\mathbf d|} q 3 g − 3 + p − ∣ d ∣ . Combining their external insertion 2 L 2 d / ( 4 d d ! ) 2L^{2d}/(4^d d!) 2 L 2 d / ( 4 d d !) with the free-energy factor 2 3 g − 3 2^{3g-3} 2 3 g − 3 and the double factorial in (25) gives (H.1). This also covers genus zero with p ≥ 3 p\geq3 p ≥ 3 . Three special cases check the normalization in (H.1). The dependence on L L L is tested at Y = 0 Y=0 Y = 0 . There x = 0 x=0 x = 0 , no cusp is present, tightness is vacuous and T g , p , 0 ( L ) = V g , p ( L ) T_{g,p,0}(\mathbf{L})=V_{g,p}(\mathbf{L}) T g , p , 0 ( L ) = V g , p ( L ) . By (24) , U g , p Y ( d ) = [ τ d 1 ⋯ τ d p ] g , p / ( 4 ∣ d ∣ P m ) U^Y_{g,p}(\mathbf{d})=[\tau_{d_1}\cdots\tau_{d_p}]_{g,p}/(4^{|\mathbf{d}|}P^m) U g , p Y ( d ) = [ τ d 1 ⋯ τ d p ] g , p / ( 4 ∣ d ∣ P m ) with m = 3 g − 3 + p − ∣ d ∣ m=3g-3+p-|\mathbf{d}| m = 3 g − 3 + p − ∣ d ∣ , so the right side of (H.1) is
P − ( 3 g − 3 + p ) ∑ d [ τ d 1 ⋯ τ d p ] g , p ∏ i = 1 p ( L i / 2 ) 2 d i ( 2 d i + 1 ) ! = V g , p ( L ) P 3 g − 3 + p . P^{-(3g-3+p)}\sum_{\mathbf{d}}[\tau_{d_1}\cdots\tau_{d_p}]_{g,p}\prod_{i=1}^{p}\frac{(L_i/2)^{2d_i}}{(2d_i+1)!}=\frac{V_{g,p}(\mathbf{L})}{P^{3g-3+p}}. P − ( 3 g − 3 + p ) d ∑ [ τ d 1 ⋯ τ d p ] g , p i = 1 ∏ p ( 2 d i + 1 )! ( L i /2 ) 2 d i = P 3 g − 3 + p V g , p ( L ) .
by the expansion of V g , p ( 2 L ) V_{g,p}(2\mathbf{L}) V g , p ( 2 L ) in § 3.4 . Next let L = 0 \mathbf{L}=\mathbf{0} L = 0 . Only d = 0 \mathbf{d}=\mathbf{0} d = 0 survives on the right, and (27) gives
T g , p ( X ( Y ) ; 0 ) = U g , p Y ( 0 ) = Φ g ( p ) ( X ( Y ) ) = ∑ k ≥ 0 V g , p + k P 3 g − 3 + p + k X ( Y ) k k ! . \mathcal{T}_{g,p}(X(Y);0)=U^Y_{g,p}(0)=\Phi_g^{(p)}(X(Y))=\sum_{k\geq0}\frac{V_{g,p+k}}{P^{3g-3+p+k}}\frac{X(Y)^k}{k!}. T g , p ( X ( Y ) ; 0 ) = U g , p Y ( 0 ) = Φ g ( p ) ( X ( Y )) = k ≥ 0 ∑ P 3 g − 3 + p + k V g , p + k k ! X ( Y ) k .
that is, T g , p , k ( 0 ) = V g , p + k T_{g,p,k}(0)=V_{g,p+k} T g , p , k ( 0 ) = V g , p + k for every k k k . This is the correct value: a protected boundary of length zero is a cusp, every other closed geodesic in its extended free homotopy class has positive length, so the tightness condition is vacuous and the tight moduli space is the full moduli space; Budd and Lions record T g , n , 0 = T g , 0 , n = V g , n T_{g,n,0}=T_{g,0,n}=V_{g,n} T g , n , 0 = T g , 0 , n = V g , n and M g , n , p t i g h t ( 0 , 0 ) = M g , n + p ( 0 , 0 ) \mathcal{M}^{\mathrm{tight}}_{g,n,p}(0,0)=\mathcal{M}_{g,n+p}(0,0) M g , n , p tight ( 0 , 0 ) = M g , n + p ( 0 , 0 ) in [32, Section 2]. Finally let g = 0 g=0 g = 0 and p = 3 p=3 p = 3 . Then m = 3 g − 3 + p − ∣ d ∣ = − ∣ d ∣ m=3g-3+p-|\mathbf{d}|=-|\mathbf{d}| m = 3 g − 3 + p − ∣ d ∣ = − ∣ d ∣ in (25) , so again only d = 0 \mathbf{d}=\mathbf{0} d = 0 contributes, with h = 2 g − 2 + p = 1 h=2g-2+p=1 h = 2 g − 2 + p = 1 , no τ k \tau_k τ k insertions and ⟨ τ ∅ ⟩ 0 = 1 \langle\tau_{\varnothing}\rangle_0=1 ⟨ τ ∅ ⟩ 0 = 1 :
∑ d U 0 , 3 Y ( d ) ∏ i = 1 3 L i 2 d i P d i ( 2 d i + 1 ) ! = U 0 , 3 Y ( 0 ) = 1 a ( Y ) = 1 X ′ ( Y ) . \sum_{\mathbf{d}}U^Y_{0,3}(\mathbf{d})\prod_{i=1}^{3}\frac{L_i^{2d_i}}{P^{d_i}(2d_i+1)!}=U^Y_{0,3}(0)=\frac{1}{a(Y)}=\frac{1}{X'(Y)}. d ∑ U 0 , 3 Y ( d ) i = 1 ∏ 3 P d i ( 2 d i + 1 )! L i 2 d i = U 0 , 3 Y ( 0 ) = a ( Y ) 1 = X ′ ( Y ) 1 .
independently of L \mathbf{L} L ; here a = X ′ a=X' a = X ′ as in § 3.4 . On the other side, in the normalization of [34, 32] with cusp weight μ = x / P \mu=x/P μ = x / P ,
T 0 , 3 ( x ; L ) = ∑ k ≥ 0 μ k k ! T 0 , 3 , k ( L , 0 k ) = T 0 , 3 ( L , μ ) = 1 M 0 ( μ ) . \mathcal{T}_{0,3}(x;\mathbf{L})=\sum_{k\geq0}\frac{\mu^k}{k!}T_{0,3,k}(\mathbf{L},0^k)=T_{0,3}(\mathbf{L},\mu)=\frac{1}{M_0(\mu)}. T 0 , 3 ( x ; L ) = k ≥ 0 ∑ k ! μ k T 0 , 3 , k ( L , 0 k ) = T 0 , 3 ( L , μ ) = M 0 ( μ ) 1 .
the last identity being the genus-zero initial case P 0 , 3 ≡ 1 P_{0,3}\equiv1 P 0 , 3 ≡ 1 of the tight recursion [34, Theorem 4], as recalled in [32, Theorem 2.2]. Their root R ( μ ) R(\mu) R ( μ ) solves μ = R 2 π J 1 ( 2 π 2 R ) \mu=\frac{\sqrt{R}}{\sqrt{2\pi}}J_1(2\pi\sqrt{2R}) μ = 2 π R J 1 ( 2 π 2 R ) and M 0 ( μ ) = d d r [ r 2 π J 1 ( 2 π 2 r ) ] ∣ r = R ( μ ) M_0(\mu)=\left.\frac{\mathrm{d}}{\mathrm{d}r}\left[\frac{\sqrt{r}}{\sqrt{2\pi}}J_1(2\pi\sqrt{2r})\right]\right|_{r=R(\mu)} M 0 ( μ ) = d r d [ 2 π r J 1 ( 2 π 2 r ) ] r = R ( μ ) . Since X ( y ) = y J 1 ( 2 y ) X(y)=\sqrt{y}J_1(2\sqrt{y}) X ( y ) = y J 1 ( 2 y ) and P = 2 π 2 P=2\pi^2 P = 2 π 2 , the function differentiated is X ( P r ) / P X(Pr)/P X ( P r ) / P ; hence R ( μ ) = Y / P R(\mu)=Y/P R ( μ ) = Y / P when μ = X ( Y ) / P \mu=X(Y)/P μ = X ( Y ) / P , and M 0 ( μ ) = X ′ ( Y ) = a ( Y ) M_0(\mu)=X'(Y)=a(Y) M 0 ( μ ) = X ′ ( Y ) = a ( Y ) , in agreement with the root and moments quoted above. Both sides of (H.1) therefore equal 1 / a ( Y ) 1/a(Y) 1/ a ( Y ) for g = 0 g=0 g = 0 , p = 3 p=3 p = 3 . In particular T 0 , 3 , k ( L ) = V 0 , 3 + k T_{0,3,k}(\mathbf{L})=V_{0,3+k} T 0 , 3 , k ( L ) = V 0 , 3 + k for all L \mathbf{L} L and k k k : the tight volumes with three protected boundaries in genus zero do not depend on the boundary lengths, which is the content of P 0 , 3 ≡ 1 P_{0,3}\equiv1 P 0 , 3 ≡ 1 .
Equations (70) and (78) give (H.2). For the asymptotic estimate, the factorial series with the extra weight ∣ d ∣ ( ∣ d ∣ + 1 ) |\mathbf{d}|(|\mathbf{d}|+1) ∣ d ∣ ( ∣ d ∣ + 1 ) is uniformly summable on the indicated box. Thus the lengths may grow as r ( Y ) − 1 / 2 r(Y)^{-1/2} r ( Y ) − 1/2 .
For each fixed physical length vector, the proof of (108) applies to every amplitude in (H.1), which gives (H.3). The finite sum in (H.1) and (H.2) justify convergence throughout 0 ≤ x < x c 0\leq x<x_c 0 ≤ x < x c . □
Prescribing the cusp count
This subsection proves Lemma H.2, the volume ratio for a cut surface with connected complement, as it occurs in the tight integration formula. We extract the coefficient at the prescribed cusp count by a local limit theorem at a suitably chosen cusp weight, or by a Poisson limit when the cusp count stays bounded; the inputs are the real and Fourier estimates (79) –(84) and (115) –(121) of § 3.25 , transferred from the marked generating functions to the tight ones by Lemma H.1. Fix a number k k k of cuts and put h 0 = 2 g − 2 h_0=2g-2 h 0 = 2 g − 2 . When n → ∞ n\to\infty n → ∞ , choose x x x by
( h 0 + 1 ) Ψ ( x ) = n − 1 , σ 2 = ( h 0 + 1 ) V ( x ) , x = X ( Y ) , (h_0+1)\Psi(x)=n-1,\qquad\sigma^2=(h_0+1)\mathscr{V}(x),\qquad x=X(Y), ( h 0 + 1 ) Ψ ( x ) = n − 1 , σ 2 = ( h 0 + 1 ) V ( x ) , x = X ( Y ) ,
and write α x = 2 ρ ( Y ) \alpha_x=2\rho(Y) α x = 2 ρ ( Y ) . If n n n stays bounded, we pass to a subsequence on which n n n is fixed and take instead x = c / h 0 x=c/h_0 x = c / h 0 with c > 0 c>0 c > 0 .
Lemma H.2 (Connected cut configurations). Fix k k k and a compact box of scaled lengths t ∈ [ 0 , ∞ ) 2 k \mathbf{t} \in[0,\infty)^{2k} t ∈ [ 0 , ∞ ) 2 k . With x x x chosen as above, uniformly for t \mathbf{t} t in the box,
[ z n ] T g − k , 2 k ( z ; t / α x ) [ z n ] Φ g ( z ) = ( 4 r ( Y ) ) k ∏ i = 1 2 k S ( t i / 2 ) ( 1 + o ( 1 ) ) . (390) \frac{[z^n]\mathcal{T}_{g-k,2k}(z;\mathbf{t}/\alpha_x)}{[z^n]\Phi_g(z)}=(4r(Y))^k\prod_{i=1}^{2k}S(t_i/2)(1+o(1)). \tag*{(390)} [ z n ] Φ g ( z ) [ z n ] T g − k , 2 k ( z ; t / α x ) = ( 4 r ( Y ) ) k i = 1 ∏ 2 k S ( t i /2 ) ( 1 + o ( 1 )) . (390) Moreover α x / α ( n / g ) → 1 \alpha_x/\alpha(n/g)\to1 α x / α ( n / g ) → 1 .
Proof. Suppose first that n → ∞ n\to\infty n → ∞ . The denominator F g , 1 = Φ g ′ F_{g,1}=\Phi'_g F g , 1 = Φ g ′ has its local limit at n − 1 n-1 n − 1 . For the connected-complement numerator G ( x ) = T g − k , 2 k ( x ; t / α x ) G(x)=\mathcal{T}_{g-k,2k}(x;\mathbf{t}/\alpha_x) G ( x ) = T g − k , 2 k ( x ; t / α x ) , the physical lengths are held fixed when its argument is varied. Its model center is h 0 Ψ ( x ) h_0\Psi(x) h 0 Ψ ( x ) , and
n − h 0 Ψ ( x ) σ = 1 + Ψ ( x ) σ ⟶ 0. \frac{n-h_0\Psi(x)}{\sigma}=\frac{1+\Psi(x)}{\sigma}\longrightarrow0. σ n − h 0 Ψ ( x ) = σ 1 + Ψ ( x ) ⟶ 0. The endpoint expansions give
sup 0 < x < x c ∣ D log r ( Y ( x ) ) ∣ D Ψ ( x ) < ∞ . \sup_{0<x<x_c}\frac{\left|\mathcal{D}\log r(Y(x))\right|}{\sqrt{\mathcal{D}\Psi(x)}}<\infty. 0 < x < x c sup D Ψ ( x ) ∣ D log r ( Y ( x )) ∣ < ∞. Consequently a real tilt x e u / σ xe^{u/\sigma} x e u / σ changes r r r by a relative O u ( g − 1 / 2 ) O_u(g^{-1/2}) O u ( g − 1/2 ) , uniformly up to either endpoint. Equation (H.2) therefore transfers the real Gaussian limit from F g − k , 2 k F_{g-k,2k} F g − k , 2 k to G G G . The positive factor (H.3) transfers the Fourier estimates (115) –(121) . The two local probabilities at n n n and n − 1 n-1 n − 1 are asymptotic to ( 2 π σ ) − 1 (\sqrt{2\pi\sigma})^{-1} ( 2 π σ ) − 1 .
We use the derivative only in the denominator: adding a protected cusp to a tight surface would change which homotopies are allowed. For a generating function J J J with nonnegative coefficients write P x J ( N = m ) = [ z m ] J ( z ) x m / J ( x ) \mathbb{P}^{J}_{x}(N=m)=[z^m]J(z)x^m/J(x) P x J ( N = m ) = [ z m ] J ( z ) x m / J ( x ) . The exact identity is
[ z n ] G ( z ) [ z n ] Φ g ( z ) = n x G ( x ) Φ g ′ ( x ) P x G ( N = n ) P x Φ g ′ ( N = n − 1 ) . (391) \frac{[z^n]G(z)}{[z^n]\Phi_g(z)}
=\frac{n}{x}\frac{G(x)}{\Phi'_g(x)}
\frac{\mathbb{P}^{G}_{x}(N=n)}{\mathbb{P}^{\Phi'_g}_{x}(N=n-1)}. \tag*{(391)} [ z n ] Φ g ( z ) [ z n ] G ( z ) = x n Φ g ′ ( x ) G ( x ) P x Φ g ′ ( N = n − 1 ) P x G ( N = n ) . (391) Equation (79) at p = 0 p=0 p = 0 shows that n Φ g ( x ) / ( x Φ g ′ ( x ) ) → 1 n\Phi_g(x)/(x\Phi'_g(x))\to1 n Φ g ( x ) / ( x Φ g ′ ( x )) → 1 . Iterating (84) now yields (H.4).
For bounded n n n , with x = c / h 0 x=c/h_0 x = c / h 0 , equation (79) gives (83) also at p = 0 p=0 p = 0 . Integrating that estimate between x x x and x z xz x z , and using Ψ ( x ) = 2 x + O ( x 2 ) \Psi(x)=2x+O(x^2) Ψ ( x ) = 2 x + O ( x 2 ) , proves convergence of the probability generating functions to exp ( 2 c ( z − 1 ) ) \exp(2c(z-1)) exp ( 2 c ( z − 1 )) . The same holds for G G G by (H.2). The probabilities at the fixed index n n n thus converge to the same positive Poisson probability e − 2 c ( 2 c ) n / n ! e^{-2c}(2c)^n/n! e − 2 c ( 2 c ) n / n ! . In this case we use, instead of (H.5), the undifferentiated identity
[ z n ] G ( z ) [ z n ] Φ g ( z ) = G ( x ) Φ g ( x ) P x G ( N = n ) P x Φ g ( N = n ) . \frac{[z^n]G(z)}{[z^n]\Phi_g(z)}
=\frac{G(x)}{\Phi_g(x)}
\frac{\mathbb{P}^{G}_{x}(N=n)}{\mathbb{P}^{\Phi_g}_{x}(N=n)}. [ z n ] Φ g ( z ) [ z n ] G ( z ) = Φ g ( x ) G ( x ) P x Φ g ( N = n ) P x G ( N = n ) . which is meaningful also at n = 0 n=0 n = 0 . This proves (H.4) in this case too.
Finally α x / α ( n / g ) → 1 \alpha_x/\alpha(n/g)\to1 α x / α ( n / g ) → 1 . At finite density this follows by continuity; at infinite density use r ( Y λ ) ∼ ( 3 5 j 0 , 1 / 2 ) λ − 1 / 2 r(Y_\lambda)\sim(3\sqrt{5}j_{0,1/2})\lambda^{-1/2} r ( Y λ ) ∼ ( 3 5 j 0 , 1/2 ) λ − 1/2 . This relative comparison is needed because both scales tend to zero at infinite density.
H.3. Separating configurations. This subsection proves Lemma H.3: tuples whose union disconnects the filled surface contribute negligibly to the tight integration formula. The mechanism is that, compared with the parent, the product of the component generating functions loses one inverse power of the genus for each additional component. The inputs are the two-sided prefactor of Proposition 7.1, a convolution bound for the resulting Gamma-function profiles, and the point-mass bound of Lemma 9.3 applied to the positive factor of Lemma H.1 .
Lemma H.3 (Separating configurations). Fix k k k and L L L . Uniformly in n n n , with lengths scaled by α x \alpha_x α x ,
E g , n W P [ ordered separating tight k -tuples with scaled lengths at most L ] = o ( 1 ) , (392) \mathbb{E}^{\mathrm{WP}}_{g,n}\left[\text{ordered separating tight }k\text{-tuples with scaled lengths at most }L\right]=o(1), \tag*{(392)} E g , n WP [ ordered separating tight k -tuples with scaled lengths at most L ] = o ( 1 ) , (392)
where only simple, pairwise disjoint tuples are counted, and separating means that their union disconnects the filled surface.
Proof. For a stable connected cut graph with k k k edges and Q Q Q vertices, let ( g i , p i ) (g_i,p_i) ( g i , p i ) be the component types after filling the cusps. No filled component is a disk or an annulus: the first would make a cut curve contractible on the filled surface, and the second would make two distinct tight curves homotopic there, contrary to the uniqueness in the definition of tightness. Hence every component is stable, and
p i ≥ 1 , h i = 2 g i − 2 + p i ≥ 1 , ∑ i h i = h 0 , ∑ i p i = 2 k , ∑ i g i = g − k + Q − 1. p_i \ge1,\qquad h_i=2g_i-2+p_i\ge1,\qquad\sum_i h_i=h_0,\qquad\sum_i p_i=2k,\qquad\sum_i g_i=g-k+Q-1. p i ≥ 1 , h i = 2 g i − 2 + p i ≥ 1 , i ∑ h i = h 0 , i ∑ p i = 2 k , i ∑ g i = g − k + Q − 1.
For fixed positive p p p , Proposition 7.1 implies the two-sided bound
F g , p ( X ( Y ) ) ≍ p 4 − g Z ( Y ) − h r ( Y ) p / 2 f ( h ) , f ( h ) = Γ ( h ) h , h = 2 g − 2 + p . (393) F_{g,p}(X(Y))\asymp_p 4^{-g}\mathcal{Z}(Y)^{-h}r(Y)^{p/2}f(h),\qquad f(h)=\frac{\Gamma(h)}{\sqrt{h}},\qquad h=2g-2+p. \tag*{(393)} F g , p ( X ( Y )) ≍ p 4 − g Z ( Y ) − h r ( Y ) p /2 f ( h ) , f ( h ) = h Γ ( h ) , h = 2 g − 2 + p . (393)
The constants are uniform in Y Y Y . Bounded stable genera are included by positivity and continuity of the normalized finite amplitudes on [ 0 , y 0 ] [0,y_0] [ 0 , y 0 ] . For p = 0 p=0 p = 0 , g ≥ 2 g\ge2 g ≥ 2 , the same bound follows from p = 1 p=1 p = 1 and (79) .
The identities for the component types cancel all powers of Z \mathcal{Z} Z in the product-to-parent ratio. They leave
r ( Y ) − k ∏ i F g i , p i ( X ( Y ) ) Φ g ( X ( Y ) ) ≤ C k ∏ i f ( h i ) f ( h 0 ) . r(Y)^{-k}\frac{\prod_i F_{g_i,p_i}(X(Y))}{\Phi_g(X(Y))}\le C_k\frac{\prod_i f(h_i)}{f(h_0)}. r ( Y ) − k Φ g ( X ( Y )) ∏ i F g i , p i ( X ( Y )) ≤ C k f ( h 0 ) ∏ i f ( h i ) .
To sum over genera, observe that ∑ a = 1 ℓ − 1 f ( a ) f ( ℓ − a ) ≤ C f ( ℓ − 1 ) \sum_{a=1}^{\ell-1}f(a)f(\ell-a)\le Cf(\ell-1) ∑ a = 1 ℓ − 1 f ( a ) f ( ℓ − a ) ≤ C f ( ℓ − 1 ) . Indeed,
Γ ( a ) Γ ( ℓ − a ) Γ ( ℓ ) = 1 ( ℓ − 1 ) ( ℓ − 2 a − 1 ) , \frac{\Gamma(a)\Gamma(\ell-a)}{\Gamma(\ell)}=\frac{1}{(\ell-1)\binom{\ell-2}{a-1}}, Γ ( ℓ ) Γ ( a ) Γ ( ℓ − a ) = ( ℓ − 1 ) ( a − 1 ℓ − 2 ) 1 ,
the reciprocal binomial coefficients have uniformly bounded sum, and ℓ / a ( ℓ − a ) \sqrt{\ell}/\sqrt{a(\ell-a)} ℓ / a ( ℓ − a ) is bounded. Iteration gives f ∗ Q ( ℓ ) ≤ C Q f ( ℓ − Q + 1 ) f^{*Q}(\ell)\le C_Qf(\ell-Q+1) f ∗ Q ( ℓ ) ≤ C Q f ( ℓ − Q + 1 ) . Since the marking allocations and gluings have bounded multiplicity for fixed k k k , it follows that
r ( Y ) − k ∑ stable cut types Q components ∏ i F g i , p i ( X ( Y ) ) Φ g ( X ( Y ) ) ≤ C k g − ( Q − 1 ) , 2 ≤ Q ≤ k + 1. (394) r(Y)^{-k}\sum_{\substack{\text{stable cut types}\\Q\text{ components}}}\frac{\prod_i F_{g_i,p_i}(X(Y))}{\Phi_g(X(Y))}\le C_kg^{-(Q-1)},\qquad2\le Q\le k+1. \tag*{(394)} r ( Y ) − k stable cut types Q components ∑ Φ g ( X ( Y )) ∏ i F g i , p i ( X ( Y )) ≤ C k g − ( Q − 1 ) , 2 ≤ Q ≤ k + 1. (394)
For scaled lengths in a fixed compact box, (H.2) bounds each boundary factor uniformly, including on small-genus components. Every product in this positive sum has the factor V h 0 V^{h_0} V h 0 by (H.3). Its largest tilted point probability is at most C / σ C/\sigma C / σ , by Lemma 9.3. Applying (H.5), against the denominator probability at least c / σ c/\sigma c / σ , therefore transfers (H.8) to coefficients. For bounded n n n , use x = c / h 0 x=c/h_0 x = c / h 0 and the undifferentiated identity above, whose denominator probability has a positive limit; the numerator probability is at most one.
In the tight integration formula [32, Proposition 3.2], the physical normalization costs P − k P^{-k} P − k and the length Jacobian is α x 2 k = ( P / ( 4 r ) ) k \alpha_x^{2k}=(P/(4r))^k α x 2 k = ( P / ( 4 r ) ) k . These cancel the remaining r k r^k r k . The symmetry factors are at most one. This proves (H.6). □
Proof of Theorem 14.8
Lemmas H.2 and H.3 give the Poisson limits of the moments of disjoint tight tuples. The proof below relates these to the moments of the actual counts, using the deterministic multicurve constructions of Budd and Lिओns [32] to show that intersecting tight curves are excluded at the level of expectations, and then passes from moments to the point process.
Proof of Theorem 14.8. For fixed disjoint bounded intervals in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , count ordered tuples of simple, pairwise disjoint tight curves in those intervals. The connected-complement symmetry is 2 − k 2^{-k} 2 − k in the tight integration formula. Lemma H.2, with each cut length appearing on two boundary components, gives the limiting integrand
∏ i = 1 k 1 2 t i S ( t i / 2 ) 2 d t i = ∏ i = 1 k cosh t i − 1 t i d t i . \prod_{i=1}^{k} \frac{1}{2}t_i S(t_i/2)^2\,\mathrm{d}t_i = \prod_{i=1}^{k} \frac{\cosh t_i-1}{t_i}\,\mathrm{d}t_i. i = 1 ∏ k 2 1 t i S ( t i /2 ) 2 d t i = i = 1 ∏ k t i cosh t i − 1 d t i .
Lemma H.3 removes all other cut types. Thus the disjoint-tuple moments have the Poisson limits of every fixed order.
Here the disjoint-tuple moments must be related to the actual count moments: a small probability of intersections alone would not suffice. Choose L L L above all interval endpoints, and let A L A_L A L be the event that scaled lengths at most L L L include a self-intersecting tight curve or two intersecting tight curves. Write N ^ r \widehat{N}_{\mathbf r} N r for the disjoint-tuple count in the prescribed intervals, with multiplicities r = ( r 1 , … , r j ) \mathbf r=(r_1,\ldots,r_j) r = ( r 1 , … , r j ) . We use the deterministic constructions in [32, proofs of Propositions 5.2–5.3 and Theorem 1.2, Section 5.2], rather than their probabilistic conclusions in the near-critical model. These constructions use only curve surgery at intersection points, the subadditivity of length under surgery and a lower bound on the scaled tight systole; they therefore apply after rescaling by any positive length scale, with the explicit bounds on the number and lengths of the added curves given in [32, Section 5.1].
Fix ε > 0 \varepsilon>0 ε > 0 below the interval endpoints. On A L A_L A L , outside the event of a tight systole shorter than ε \varepsilon ε , the construction extends a counted tuple to a separating tight multicurve. Its number of components and scaled lengths are bounded in terms of ε \varepsilon ε , L L L , r \mathbf r r , independently of g , n g,n g , n . If there is a shorter tight systole, either it adds a disjoint curve in [ 0 , ε ] [0,\varepsilon] [ 0 , ε ] , or the tuple again extends to a separating multicurve. The original ordered tuple is retained in these extensions. Thus the separating terms vanish by (H.6), and the mixed disjoint-tuple limit gives
lim sup g → ∞ E [ N ^ r 1 A L ] ≤ ν ( [ 0 , ε ] ) ∏ i = 1 j ν ( I i ) r i . \limsup_{g\to\infty}\mathbb{E}[\widehat{N}_{\mathbf r}\mathbf{1}_{A_L}]
\le\nu([0,\varepsilon])\prod_{i=1}^{j}\nu(I_i)^{r_i}. g → ∞ lim sup E [ N r 1 A L ] ≤ ν ([ 0 , ε ]) i = 1 ∏ j ν ( I i ) r i .
where I i I_i I i are the prescribed intervals. Letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 removes the right side. The same construction without a marked tuple gives P ( A L ) → 0 \mathbb{P}(A_L)\to0 P ( A L ) → 0 . Conditional on A L c A_L^c A L c , factorial counts are exactly disjoint-tuple counts. Their limiting moments therefore remain unchanged, and the moment-determinate Poisson law gives joint count convergence.
The first-moment bound for simple tight curves in [ 0 , ε ] [0,\varepsilon] [ 0 , ε ] tends to ν ( [ 0 , ε ] ) → 0 \nu([0,\varepsilon])\to0 ν ([ 0 , ε ]) → 0 . A shortest tight curve is simple: resolving a self-intersection yields two closed curves, at least one of which is noncontractible on the filled surface, and its tight representative is strictly shorter. So the same bound controls the event of any tight curve in this window. This extends the convergence to [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) and gives tightness on compact intervals. Finally the disjoint-tuple moments bound the number of curves in a fixed window in probability. For any fixed upper bound on that number, (H.6) rules out every disconnected complement. Letting the upper bound increase proves the last claim. □ \square □
Proof of Corollary 14.9
Proof of Corollary 14.9. Under the Boltzmann measure P g , μ W P \mathbb{P}^{\mathrm{WP}}_{g,\mu} P g , μ WP of Subsection 14.5 the cusp count is random. We show, using the real central limit theorem of § 3.25 , that the scale evaluated at the random density is asymptotic in probability to the deterministic scale a g a_g a g , and then average the conditional limits of Theorem 14.8 over the cusp count.
Write x = P μ g x=P\mu_g x = P μ g and h = 2 g − 2 h=2g-2 h = 2 g − 2 . The case x = 0 x=0 x = 0 has N = 0 N=0 N = 0 . Otherwise (79) at p = 0 p=0 p = 0 gives
E N = h Ψ ( x ) + O ( Ψ ( x ) ) . \mathbb{E}N=h\Psi(x)+O(\Psi(x)). E N = h Ψ ( x ) + O ( Ψ ( x )) .
If h Ψ ( x ) h\Psi(x) h Ψ ( x ) stays bounded, N N N is tight and both N / g N/g N / g and 2 Ψ ( x ) 2\Psi(x) 2Ψ ( x ) tend to zero. If h Ψ ( x ) → ∞ h\Psi(x)\to\infty h Ψ ( x ) → ∞ , integrating (83) at p = 0 p=0 p = 0 gives the same real central limit theorem as (113) , hence
N = h Ψ ( x ) + O P ( h ∂ Ψ ( x ) ) . N=h\Psi(x)+O_{\mathbb{P}}\left(\sqrt{h\partial\Psi(x)}\right). N = h Ψ ( x ) + O P ( h ∂ Ψ ( x ) ) .
The endpoint estimates (110) imply D Ψ / Ψ 2 ≤ C ( 1 + 1 / Ψ ) D\Psi/\Psi^2 \le C(1+1/\Psi) D Ψ/ Ψ 2 ≤ C ( 1 + 1/Ψ ) . It follows that N / ( h Ψ ( x ) ) → 1 N/(h\Psi(x)) \to1 N / ( h Ψ ( x )) → 1 in probability in this second case. Continuity of α \alpha α at finite density and its λ − 1 / 4 \lambda^{-1/4} λ − 1/4 asymptotic at infinity therefore give, in both cases,
α ( N / g ) α ( 2 Ψ ( x ) ) = α ( N / g ) a g ⟶ 1 in probability . \frac{\alpha(N/g)}{\alpha(2\Psi(x))}=\frac{\alpha(N/g)}{a_g}\longrightarrow1 \quad\text{in probability}. α ( 2Ψ ( x )) α ( N / g ) = a g α ( N / g ) ⟶ 1 in probability .
Subsequences cover arbitrary choices of x x x .
For each fixed nonnegative compactly supported continuous test function, the conditional Laplace-functional convergence in Theorem 14.8 is uniform in n n n : a sequence violating uniformity would contradict that theorem. Averaging over N N N gives the Poisson limit with random normalization α ( N / g ) \alpha(N/g) α ( N / g ) . The displayed scale comparison and continuity of dilation of point measures replace this by a g a_g a g . For μ g → μ < μ c r \mu_g \to\mu< \mu_{\mathrm{cr}} μ g → μ < μ cr , a change of scale gives the stated unscaled intensity. At the critical endpoint, r ∼ 3 ( y 0 − Y ) r \sim3(y_0-Y) r ∼ 3 ( y 0 − Y ) and x c − x ∼ ( − X ′ ′ ( y 0 ) ) ( y 0 − Y ) 2 / 2 x_c-x \sim(-X''(y_0))(y_0-Y)^2/2 x c − x ∼ ( − X ′′ ( y 0 )) ( y 0 − Y ) 2 /2 show that a g a_g a g is asymptotic to the scale in [32, Theorem 1.2]. □
Appendix I. Proofs for the critical length spectrum
This appendix proves Theorem 14.10 , following the outline in Subsection 14.6. Throughout, n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g} \to c \in(0,\infty) n / g → c ∈ ( 0 , ∞ ) , and P s P_s P s , K s K_s K s , μ c \mu_c μ c and ν \nu ν are as in (202) . The volume inputs are the bounds of Mirzakhani–Zograf and Shen–Wu [147, 172], the zero-density case of Theorem 1.3 and the boundary factorization of Proposition 9.4 ; the geometric inputs are the intersection and filling arguments of [146, 95]. The four subsections bound the contribution of each cut graph (Lemma I.1 ), reduce intersecting simple tuples and nonsimple curves to cut graphs (Lemmas I.2 and I.3 ), compute the joint Poisson limit of nonseparating lengths and waists of two-cusp pairs of pants (Lemma I.4 ), and finally attach to each waist the primitive length spectrum of its pair of pants. The first three lemmas are also used outside this appendix: Lemmas I.1 and I.3 isolate the pants in § 4.7 , and Lemma I.2 is used in Subsection 14.6 and in Appendices T, V and W.
I.1. Ordinary cuts and cusp allocations. Short multicurves are organized by their cut graphs: the vertices are the components of the complement, labeled by their genera, and the edges are the curves. Lemma I.1 bounds the expected number of multicurves with a given cut graph, using only Mirzakhani integration and the volume bounds of [172]. It shows that, in the limit, only nonseparating curves and boundaries of two-cusp pairs of pants contribute; the later subsections compute these contributions jointly.
Fix a connected cut graph with k ≥ 1 k \ge1 k ≥ 1 edges and Q Q Q vertices; loops and multiple edges are allowed. At vertex i i i let a i a_i a i be the genus and p i ≥ 1 p_i \ge1 p i ≥ 1 its degree. Then
∑ i a i = g − k + Q − 1 , ∑ i p i = 2 k . \sum_i a_i = g-k+Q-1,\qquad\sum_i p_i = 2k. i ∑ a i = g − k + Q − 1 , i ∑ p i = 2 k .
Partition the vertices into u u u of type ( 0 , 1 ) (0,1) ( 0 , 1 ) , v v v of type ( 0 , 2 ) (0,2) ( 0 , 2 ) , and w w w with h i = 2 a i − 2 + p i ≥ 1 h_i=2a_i-2+p_i \ge1 h i = 2 a i − 2 + p i ≥ 1 . Put h 0 = 2 g − 2 h_0=2g-2 h 0 = 2 g − 2 . For large g g g , w ≥ 1 w \ge1 w ≥ 1 , and the latter complexities sum to H = h 0 + u H=h_0+u H = h 0 + u . A component of type ( 0 , 1 ) (0,1) ( 0 , 1 ) needs at least two cusps, and one of type ( 0 , 2 ) (0,2) ( 0 , 2 ) at least one.
Lemma I.1 (Critical cut estimate). Fix k k k and a bound on the cut lengths. At n / g → c > 0 n/\sqrt{g} \to c>0 n / g → c > 0 , the expected number of cut multicurves of the specified graph and vertex partition, summed over all admissible genera and cusp allocations, is O ( g 1 − w − v / 2 ) O(g^{1-w-v/2}) O ( g 1 − w − v /2 ) . If a specified ( 0 , 1 ) (0,1) ( 0 , 1 ) component has at least three cusps, the bound gains a factor g − 1 / 2 g^{-1/2} g − 1/2 .
Proof. Write h n = 2 g − 2 + n h_n=2g-2+n h n = 2 g − 2 + n for the complexity of the cusped surface. The volume bounds in [172, Lemmas 18–19] give
V a , b ≤ C ( 4 π 2 ) d − 1 Γ ( d ) d , d = 2 a − 2 + b ≥ 1 , V g , n ≥ c 0 ( 4 π 2 ) h n − 1 Γ ( h n ) h n . (395) V_{a,b}\le C\frac{(4\pi^2)^{d-1}\Gamma(d)}{\sqrt{d}},\qquad d=2a-2+b\ge1,\qquad V_{g,n}\ge c_0\frac{(4\pi^2)^{h_n-1}\Gamma(h_n)}{\sqrt{h_n}}. \tag*{(395)} V a , b ≤ C d ( 4 π 2 ) d − 1 Γ ( d ) , d = 2 a − 2 + b ≥ 1 , V g , n ≥ c 0 h n ( 4 π 2 ) h n − 1 Γ ( h n ) . (395)
The first bound includes d = 1 d=1 d = 1 after enlarging C C C for ( 0 , 3 ) (0,3) ( 0 , 3 ) and ( 1 , 1 ) (1,1) ( 1 , 1 ) ; the second uses n = O ( g ) n=O(\sqrt{g}) n = O ( g ) . If vertex i i i receives q i q_i q i cusps, put d i = h i + q i d_i=h_i+q_i d i = h i + q i . Then ∑ i d i = h n \sum_i d_i=h_n ∑ i d i = h n and h n / ∏ i d i ≤ Q \sqrt{h_n}/\prod_i\sqrt{d_i}\le\sqrt{Q} h n / ∏ i d i ≤ Q . Mirzakhani integration, the bounded-boundary estimate [144, Lemma 3.2], and the multinomial allocation of cusp labels therefore bound the expected cut count E g , n E_{g,n} E g , n by
E g , n ≤ C k , c n ! Γ ( h n ) ∑ admissible genera q 1 + ⋯ + q Q = n ∏ i Γ ( h i + q i ) q i ! E_{g,n}\le\frac{C_{k,c}n!}{\Gamma(h_n)}\sum_{\substack{\text{admissible genera}\\q_1+\cdots+q_Q=n}}\prod_i\frac{\Gamma(h_i+q_i)}{q_i!} E g , n ≤ Γ ( h n ) C k , c n ! admissible genera q 1 + ⋯ + q Q = n ∑ i ∏ q i ! Γ ( h i + q i )
Here and below the constants may also depend on the length bound. The powers of 4 π 2 4\pi^2 4 π 2 leave only the fixed factor ( 4 π 2 ) 1 − Q (4\pi^2)^{1-Q} ( 4 π 2 ) 1 − Q .
The cusp sums at the two unstable vertex types are bounded using
ϕ 0 , 1 ( z ) = ∑ q ≥ 2 z q q ( q − 1 ) = z 2 2 + O ( z 3 ) , ϕ 0 , 2 ( z ) = ∑ q ≥ 1 z q q = z + O ( z 2 ) . (396) \phi_{0,1}(z)=\sum_{q\ge2}\frac{z^q}{q(q-1)}=\frac{z^2}{2}+O(z^3),\qquad\phi_{0,2}(z)=\sum_{q\ge1}\frac{z^q}{q}=z+O(z^2). \tag*{(396)} ϕ 0 , 1 ( z ) = q ≥ 2 ∑ q ( q − 1 ) z q = 2 z 2 + O ( z 3 ) , ϕ 0 , 2 ( z ) = q ≥ 1 ∑ q z q = z + O ( z 2 ) . (396)
At a stable vertex their generating function is Γ ( h ) ( 1 − z ) − h \Gamma(h)(1-z)^{-h} Γ ( h ) ( 1 − z ) − h . Moreover
∑ h 1 + ⋯ + h w = H h i ≥ 1 ∏ i Γ ( h i ) ≤ C w Γ ( H − w + 1 ) . \sum_{\substack{h_1+\cdots+h_w=H\\h_i\ge1}}\prod_i\Gamma(h_i)\le C_w\Gamma(H-w+1). h 1 + ⋯ + h w = H h i ≥ 1 ∑ i ∏ Γ ( h i ) ≤ C w Γ ( H − w + 1 ) .
For w = 2 w=2 w = 2 this follows by dividing by Γ ( H − 1 ) \Gamma(H-1) Γ ( H − 1 ) and summing ( H − 2 h 1 − 1 ) \binom{H-2}{h_1-1} ( h 1 − 1 H − 2 ) ; iteration gives the general bound. Dropping genus parity constraints only increases the sum. Let a m = [ z m ] ϕ 0 , 1 ( z ) u ϕ 0 , 2 ( z ) v a_m=[z^m]\phi_{0,1}(z)^u\phi_{0,2}(z)^v a m = [ z m ] ϕ 0 , 1 ( z ) u ϕ 0 , 2 ( z ) v . Expanding ( 1 − z ) − H (1-z)^{-H} ( 1 − z ) − H gives
E g , n ≤ C k , c Γ ( H − w + 1 ) Γ ( H ) ∑ m ≤ n a m ( n ) m Γ ( h n + u − m ) Γ ( h n ) . E_{g,n}\le C_{k,c}\frac{\Gamma(H-w+1)}{\Gamma(H)}\sum_{m\le n}a_m(n)_m\frac{\Gamma(h_n+u-m)}{\Gamma(h_n)}. E g , n ≤ C k , c Γ ( H ) Γ ( H − w + 1 ) m ≤ n ∑ a m ( n ) m Γ ( h n ) Γ ( h n + u − m ) .
Uniformly for m ≤ n = O ( h n ) m\le n=O(\sqrt{h_n}) m ≤ n = O ( h n ) , the finite-product formula for the Gamma ratio gives Γ ( h n + u − m ) / Γ ( h n ) ≤ C k , c h n u − m \Gamma(h_n+u-m)/\Gamma(h_n)\le C_{k,c}h_n^{u-m} Γ ( h n + u − m ) /Γ ( h n ) ≤ C k , c h n u − m : the logarithmic error is O ( ( m + 1 ) 2 / h n ) O((m+1)^2/h_n) O (( m + 1 ) 2 / h n ) . Since ( n ) m ≤ n m (n)_m\le n^m ( n ) m ≤ n m , positivity yields
E g , n ≤ C k , c h n u − w + 1 ϕ 0 , 1 ( n / h n ) u ϕ 0 , 2 ( n / h n ) v = O ( g 1 − w − v / 2 ) . (397) E_{g,n}\le C_{k,c}h_n^{u-w+1}\phi_{0,1}(n/h_n)^u\phi_{0,2}(n/h_n)^v=O(g^{1-w-v/2}). \tag*{(397)} E g , n ≤ C k , c h n u − w + 1 ϕ 0 , 1 ( n / h n ) u ϕ 0 , 2 ( n / h n ) v = O ( g 1 − w − v /2 ) . (397)
Requiring three cusps at a specified ( 0 , 1 ) (0,1) ( 0 , 1 ) vertex replaces one ϕ 0 , 1 ( z ) \phi_{0,1}(z) ϕ 0 , 1 ( z ) by ϕ 0 , 1 ( z ) − z 2 / 2 \phi_{0,1}(z)-z^2/2 ϕ 0 , 1 ( z ) − z 2 /2 and gains n / h n = O ( g − 1 / 2 ) n/h_n=O(g^{-1/2}) n / h n = O ( g − 1/2 ) . □
Only w = 1 w=1 w = 1 , v = 0 v=0 v = 0 can survive: the graph then has one large-genus vertex, loops at that vertex, and two-cusp pairs of pants attached as leaves. All other graphs, and all leaves with more than two cusps, vanish.
Excluding complex short subsurfaces
We now reduce intersecting simple tuples, and nonsimple curves that do not fill two-cusp pairs of pants, to the cut graphs just estimated (Lemma I.3 ). Since we need expectation bounds, not only a bound on the probability of a bad configuration, we give the multiplicity argument explicitly. It rests on a uniform bound for the number of short primitive geodesics on a surface of bounded topology (Lemma I.2), on the local intersection argument of Mirzakhani–Petri [146] extended to cusp neighborhoods, and on the filling construction of Hide [95].
Lemma I.2 (Bounded primitive count). For fixed K K K , L L L , a finite-area hyperbolic surface S S S with geodesic boundary and 0 < − χ ( S ) ≤ K 0<-\chi(S)\le K 0 < − χ ( S ) ≤ K has at most C ( K , L ) C(K,L) C ( K , L ) primitive unoriented geodesics of length at most L L L , uniformly in its metric.
Proof. Double S S S along its boundaries. The double has area at most 4 π K 4\pi K 4 π K . Fix a Margulis constant ε > 0 \varepsilon>0 ε > 0 . A primitive geodesic wholly in a thin component is its core; there are boundedly many such cores. Every other geodesic meets the thick part. A maximal ε / 4 \varepsilon/4 ε /4 -separated net there has bounded cardinality by area packing. At each net point, orbit points of a lift are uniformly separated, so hyperbolic ball packing bounds the number of based loops of length at most L + ε L+\varepsilon L + ε . Connecting a short geodesic to a net point proves the assertion. Primitivity is preserved upon inclusion in the double or in an ambient surface: a root would have the same geodesic image and would already be a closed curve in S S S . □ \square □
Lemma I.3 (Complex short configurations). Fix r r r and L L L . As g → ∞ g \to\infty g → ∞ with n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g} \to c \in(0,\infty) n / g → c ∈ ( 0 , ∞ ) ,
E [ intersecting ordered simple r -tuples of lengths at most L ] = O r , L , c ( g − 1 / 2 ) , (398) \mathbb{E}[\textit{intersecting ordered simple }r\textit{-tuples of lengths at most }L] = O_{r,L,c}(g^{-1/2}), \tag*{(398)} E [ intersecting ordered simple r -tuples of lengths at most L ] = O r , L , c ( g − 1/2 ) , (398) E [ nonsimple primitive curves of length at most L not filling two-cusp pants ] = O L , c ( g − 1 / 2 ) , (399) \mathbb{E}[\textit{nonsimple primitive curves of length at most }L\textit{ not filling two-cusp pants}] = O_{L,c}(g^{-1/2}), \tag*{(399)} E [ nonsimple primitive curves of length at most L not filling two-cusp pants ] = O L , c ( g − 1/2 ) , (399) and the expected number of simple separating curves of length at most L L L that are not waists of two-cusp pairs of pants tends to zero.
Proof. For fixed r r r , L L L , a tuple of r r r distinct simple geodesics of length at most L L L has bounded total intersection number, on every finite-area hyperbolic surface. The segment argument in the proof of [146, Proposition 4.5] rests on [146, Lemma 4.4]: if two geodesic segments $I$, $J$ have common endpoints, meet transversally at at least one of them, and $I$ lies on a closed geodesic of length at most $L$, then $\ell(I) + \ell(J)$ is bounded below by a positive constant depending only on $L$. On a cusped surface this holds whenever the loop $I \cup J$ is not peripheral; here $I$ and $J$ lie on two distinct geodesics of the tuple, which meet transversally. Fix $Y = 4$ and write $\{y > Y\}$ for the cusp neighborhoods in width-one cusp coordinates. The larger regions $\{y > 2\}$ are embedded and pairwise disjoint by [101, Lemma 5.2]. A peripheral loop through a point of height $y \le Y$ has length at least $2\operatorname{arcsinh}(1/(2Y))$, so the argument of {citep}[146]` bounds the number of intersection points outside the cusp neighborhoods. For those inside, note that no closed geodesic lies entirely in a cusp neighborhood, and that two consecutive excursions of a geodesic into ⋃ { y > Y } \bigcup\{y > Y\} ⋃ { y > Y } are separated by length at least 2 log ( Y / 2 ) = 2 log 2 2\log(Y/2) = 2\log2 2 log ( Y /2 ) = 2 log 2 ; a closed geodesic of length at most L L L therefore makes at most 1 + L / ( 2 log 2 ) 1 + L/(2\log2) 1 + L / ( 2 log 2 ) excursions. Each excursion lifts to an arc of a Euclidean semicircle centered on the real axis, with endpoints on { y = Y } \{y = Y\} { y = Y } and radius at most Y e L / 2 Y e^{L/2} Y e L /2 ; two distinct semicircles meet at most once, and a semicircle meets at most 4 Y e L / 2 + 1 4Y e^{L/2} + 1 4 Y e L /2 + 1 of the integer translates of another. Hence the total intersection number of the tuple is bounded by a constant depending only on r r r and L L L . Thicken each nonannular component of their union, fill complementary disks and peripheral once-punctured disks, and tighten the essential boundaries. The resulting pieces have bounded topology and total boundary length at most 2 r L 2rL 2 r L . Keep one ambient geodesic for each parallel boundary class. Empty annuli between parallel copies collapse; their two incidences are the two sides of a single edge, possibly a loop. An annulus with cusps does not collapse and is a ( 0 , 2 ) (0,2) ( 0 , 2 ) component counted among the v v v vertices. Isolated tuple curves are retained as additional cut edges, unless they are already one of the tightened boundaries. Minimal intersection ensures that these boundaries are disjoint from the original tuple.
There are finitely many resulting graph shapes and markings at fixed r r r , L L L . For a fixed cut multicurve, Lemma I.2 bounds the choices of original curves in each marked piece. Hence the tuple count is bounded pointwise by a constant times a finite sum of marked cut-multicurve counts. An intersecting simple tuple cannot fill a two-cusp pair of pants. Its marked component either has h i ≥ 1 h_i \ge1 h i ≥ 1 , forcing a second stable vertex to carry the unbounded genus, or has type ( 0 , 2 ) (0,2) ( 0 , 2 ) , or has type ( 0 , 1 ) (0,1) ( 0 , 1 ) with at least three cusps. The bounds in Lemma I.1 give (I.3).
For a single nonsimple primitive curve of length at most L L L , use the filling construction of [95, Definition 4.10 and Lemma 4.11]. The filled piece has boundary length at most 2 L 2L 2 L and area at most 4 L 4L 4 L , so its topology is bounded. The same recording retains formal boundary sides that tighten to the same ambient curve as sides of a cut edge. Lemma I.2 controls the number of curves filling each piece. The preceding classification therefore gives (I.4). A filled piece cannot be the whole ambient surface for large g g g , since its topology is bounded. Finally, the one-edge case of Lemma I.1 shows that simple separating curves of bounded length other than waists of two-cusp pairs of pants vanish in expectation. □ \square □
Joint nonseparating and pants counts
With intersecting tuples and the other cut graphs excluded, it remains to compute the mixed factorial moments of the two surviving counts (Lemma I.4 ). This is where the two-sided volume ratios enter: the cusp ratios come from [147], the genus ratios from the zero-density case of Theorem 1.3 , and the boundary factors from Proposition 9.4 . Let N g \mathcal{N}_g N g count simple nonseparating lengths and let Q X \mathcal{Q}_X Q X count the waist lengths of two-cusp pairs of pants.
Lemma I.4 (Joint Poisson limit). As g → ∞ g \to\infty g → ∞ with n / g → c ∈ ( 0 , ∞ ) n/\sqrt{g} \to c \in(0,\infty) n / g → c ∈ ( 0 , ∞ ) ,
( N g , Q X ) ⟹ ( N , Q c ) (400) (\mathcal{N}_g,\mathcal{Q}_X) \Longrightarrow(\mathcal{N},\mathcal{Q}_c) \tag*{(400)} ( N g , Q X ) ⟹ ( N , Q c ) (400) jointly in the vague topology on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , where N \mathcal{N} N and Q c \mathcal{Q}_c Q c are independent Poisson point processes of intensities ν \nu ν and μ c \mu_c μ c .
Proof. In the mixed factorial moments of ( N g , Q X ) (\mathcal{N}_g,\mathcal{Q}_X) ( N g , Q X ) , (I.3) removes intersecting tuples. Lemma I.1 removes every disjoint cut type except the graph with a single large component, k k k nonseparating loops and r r r two-cusp leaves. For prescribed bounded intervals for the k + r k+r k + r lengths, its exact contribution is
( n ) 2 r 2 r + k V g , n ∫ V g − k , n − r + 2 k ( s , t , t , 0 n − 2 r ) ∏ j = 1 r s j d s j ∏ i = 1 k t i d t i . (401) \frac{(n)_{2r}}{2^{r+k}V_{g,n}}\int V_{g-k,n-r+2k}(\mathbf{s},\mathbf{t},\mathbf{t},0^{n-2r})\prod_{j=1}^{r}s_j\,\mathrm{d}s_j\prod_{i=1}^{k}t_i\,\mathrm{d}t_i . \tag*{(401)} 2 r + k V g , n ( n ) 2 r ∫ V g − k , n − r + 2 k ( s , t , t , 0 n − 2 r ) j = 1 ∏ r s j d s j i = 1 ∏ k t i d t i . (401)
Here ( n ) 2 r (n)_{2r} ( n ) 2 r is the falling factorial. The factor 2 − r 2^{-r} 2 − r counts unordered pairs of cusp labels at the ordered pants leaves, as in [172, Lemma 28]; 2 − k 2^{-k} 2 − k exchanges the sides of each nonseparating cut, as in [146, Proposition 4.2]. There is no factor r ! r! r ! or k ! k! k ! , since both tuples are ordered.
The volume ratio is a product of r r r cusp ratios and k k k genus ratios. The cusp ratios satisfy V g , j + 1 / V g , j ∼ 8 π 2 g V_{g,j+1}/V_{g,j}\sim8\pi^2 g V g , j + 1 / V g , j ∼ 8 π 2 g for j ≤ n j\le n j ≤ n by [147, Lemma 5.1(ii)], since n = o ( g ) n=o(g) n = o ( g ) . For the genus ratios, [147, Lemma 5.1(iii)] gives only an upper bound, and the two-sided limit at n ∼ g n\sim\sqrt{g} n ∼ g is the zero-density case B ( 0 ) = 1 \mathcal{B}(0)=1 B ( 0 ) = 1 of Theorem 1.3 . Together,
V g − k , n − r + 2 k V g , n ∼ ( 8 π 2 g ) − r . \frac{V_{g-k,n-r+2k}}{V_{g,n}}\sim(8\pi^2g)^{-r}. V g , n V g − k , n − r + 2 k ∼ ( 8 π 2 g ) − r .
Proposition 9.4 , applied at zero limiting density, where b 0 ( L ) = sinh ( L / 2 ) / ( L / 2 ) b_0(L)=\sinh(L/2)/(L/2) b 0 ( L ) = sinh ( L /2 ) / ( L /2 ) , supplies, uniformly on the fixed length box, the boundary factor ∏ j S ( s j / 2 ) 2 ∏ i S ( t i / 2 ) 2 \prod_j S(s_j/2)^2\prod_i S(t_i/2)^2 ∏ j S ( s j /2 ) 2 ∏ i S ( t i /2 ) 2 , with S ( t ) = sinh ( t ) / t S(t)=\sinh(t)/t S ( t ) = sinh ( t ) / t ; compare Shen–Wu’s bounded-boundary estimate [172, Proposition 25], which they attribute essentially to [146, Proposition 3.1] and [155, Lemma 20]. Consequently the limiting factors in (I.6) are
1 2 t S ( t / 2 ) 2 d t = ν ( d t ) , c 2 16 π 2 s S ( s / 2 ) d s = μ c ( d s ) . \frac{1}{2}tS(t/2)^2\,\mathrm{d}t=\nu(\mathrm{d}t),\qquad
\frac{c^2}{16\pi^2}sS(s/2)\,\mathrm{d}s=\mu_c(\mathrm{d}s). 2 1 tS ( t /2 ) 2 d t = ν ( d t ) , 16 π 2 c 2 s S ( s /2 ) d s = μ c ( d s ) .
Hence, for disjoint bounded intervals, the mixed factorial moments of the counts of N g \mathcal{N}_g N g and Q X \mathcal{Q}_X Q X converge to those of independent Poisson variables with the corresponding ν \nu ν - and μ c \mu_c μ c -masses. A multivariate Poisson law is determined by its moments, so these counts converge jointly in distribution. The intensities ν \nu ν and μ c \mu_c μ c are diffuse and finite on bounded intervals, including neighborhoods of zero, where their masses are O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) ; convergence of the counts on finite families of disjoint bounded intervals is therefore equivalent to vague convergence in distribution of the pair of point processes [115, Chapter 4]. This proves the lemma. □
Proof of Theorem 14.10
By Lemmas I.1 and I.3 , with probability tending to one the primitive length spectrum below a fixed bound consists of the simple nonseparating lengths together with the primitive length spectra of the two-cusp pairs of pants with short waists. The proof combines this with the joint Poisson limit of Lemma I.4 and the continuity of the pants spectra in the waist length.
Completion of the proof of Theorem 14.10. Fix L < ∞ L<\infty L < ∞ . By Lemma I.3 , with probability tending to one every nonsimple primitive curve of length at most L L L fills a two-cusp pair of pants with waist at most 2 L 2L 2 L . By (I.3), the pants waists up to 2 L 2L 2 L are pairwise disjoint with probability tending to one. The pants interiors are then disjoint as well. Other simple separating curves up to L L L are absent with probability tending to one. On this event, the full primitive counting measure restricted to [ 0 , L ] [0,L] [ 0 , L ] is exactly
( N g + ∑ s ∈ Q X ∩ [ 0 , 2 L ] K s ) ∣ [ 0 , L ] . \left(\mathcal{N}_g+\sum_{s\in\mathcal{Q}_X\cap[0,2L]}K_s\right)\bigg|_{[0,L]}. N g + s ∈ Q X ∩ [ 0 , 2 L ] ∑ K s [ 0 , L ] .
Every interior primitive geodesic of P s P_s P s is nonsimple and fills it. The filling boundary bound implies s ≤ 2 ℓ s\le2\ell s ≤ 2 ℓ for such a geodesic of length ℓ \ell ℓ . Thus pants with waist greater than 2 L 2L 2 L cannot contribute, and the limiting sum is locally finite.
For ε > 0 \varepsilon>0 ε > 0 , the family P s P_s P s , ε ≤ s ≤ 2 L \varepsilon\le s\le2L ε ≤ s ≤ 2 L , is compact in its fixed-topology moduli space. Lengths of marked classes vary continuously, and only finitely many such classes have length at most L L L somewhere on this compact family. For the latter assertion, use a finite cover by uniformly bilipschitz markings, with standard neighborhoods at the two cusps, and the finiteness of each center metric’s length spectrum below a fixed bound. Hence s ↦ ⟨ f , K s ⟩ s\mapsto\langle f,K_s\rangle s ↦ ⟨ f , K s ⟩ is continuous away from zero for continuous compactly supported f f f . Write F ( s ) = ⟨ f , K s ⟩ F(s)=\langle f,K_s\rangle F ( s ) = ⟨ f , K s ⟩ and choose L L L larger than the support of f f f ; then F F F is bounded by Lemma I.2 and vanishes for s ≥ 2 L s\ge2L s ≥ 2 L . For ε > 0 \varepsilon>0 ε > 0 the map ( η , ξ ) ↦ ⟨ f , η ⟩ + ∫ [ ε , ∞ ) F d ξ (\eta,\xi)\mapsto\langle f,\eta\rangle+\int_{[\varepsilon,\infty)}F\,\mathrm{d}\xi ( η , ξ ) ↦ ⟨ f , η ⟩ + ∫ [ ε , ∞ ) F d ξ is vaguely continuous at every pair with ξ ( { ε } ) = 0 \xi(\{\varepsilon\})=0 ξ ({ ε }) = 0 , hence almost surely at ( N , Q c ) (\mathcal{N},\mathcal{Q}_c) ( N , Q c ) , because μ c \mu_c μ c has no atoms. By (I.5) and the continuous mapping theorem,
⟨ f , N g ⟩ + ∑ s ∈ Q X , s ≥ ε F ( s ) ⟹ ⟨ f , N ⟩ + ∑ s ∈ Q c , s ≥ ε F ( s ) . \langle f,\mathcal{N}_g\rangle+\sum_{s\in\mathcal{Q}_X,\ s\ge\varepsilon}F(s)\Longrightarrow\langle f,\mathcal{N}\rangle+\sum_{s\in\mathcal{Q}_c,\ s\ge\varepsilon}F(s). ⟨ f , N g ⟩ + s ∈ Q X , s ≥ ε ∑ F ( s ) ⟹ ⟨ f , N ⟩ + s ∈ Q c , s ≥ ε ∑ F ( s ) .
On the event described above, the left side differs from the pairing of f f f with the primitive length spectrum only if a waist is shorter than ε \varepsilon ε , which has probability at most E Q X ( [ 0 , ε ] ) \mathbb{E}\mathcal{Q}_X([0,\varepsilon]) E Q X ([ 0 , ε ]) . By the case k = 0 k=0 k = 0 , r = 1 r=1 r = 1 of (I.6) this expectation converges to μ c ( [ 0 , ε ] ) = O ( ε 2 ) \mu_c([0,\varepsilon])=O(\varepsilon^2) μ c ([ 0 , ε ]) = O ( ε 2 ) , and the right side converges almost surely to ⟨ f , Λ c ⟩ \langle f,\Lambda_c\rangle ⟨ f , Λ c ⟩ as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . Hence the Laplace functionals converge, which proves (203) . The independent Poisson Laplace functionals give (204) . □
Trace estimates on bounded critical envelopes
This appendix extends Huang’s orbital trace estimate [101] to cusp counts of order at most g \sqrt{g} g . Its two outputs are the polynomial trace estimate (216) and the local expected-count bound for eigenvalues below a fixed threshold, both used in the proof of Proposition 15.3 ; Appendix S restates the corresponding inputs from [101] for the sublinear range, with tail bounds proved there. Huang’s stated additive genus threshold contains C ( n + 1 ) 2 C(n+1)^2 C ( n + 1 ) 2 , so it cannot simply be invoked for an arbitrary fixed limit n / g = c n/\sqrt{g}=c n / g = c . We retain his canonical coefficients and modify only the complementary-volume tail estimates. Subsection J.1 restates, in the notation of this paper, every result of [101] that the appendix uses, and records which two of his estimates are replaced. Subsection J.2 proves the replaced tail bounds on the new envelope, Subsection J.3 checks the retained records and integrates over orbital types, and Subsection J.4 passes to polynomial tests of growing degree and to eigenvalue counts. All numbered references to [101] below refer to version 1.
Fix M ≥ 1 M\ge1 M ≥ 1 and write s = n + 1 ≤ M g s=n+1\le M\sqrt{g} s = n + 1 ≤ M g ; the letter s s s is reserved for waist lengths, as in § 4.7 . Constants may depend on M M M and on the fixed multiplier, but not on g , n g,n g , n or the expansion order. There are θ > 0 \theta>0 θ > 0 and g M g_M g M such that for g ≥ g M g\ge g_M g ≥ g M and ρ = R + ∣ T ∣ + 1 ≤ g θ \rho=R+|T|+1\le g^\theta ρ = R + ∣ T ∣ + 1 ≤ g θ , the orbital expansion obeys
∣ E O H − Ω ( H ) − ∑ j = 1 R g − j κ j , n ( H ) ∣ ≤ g − R − 1 s R + 2 ( C M ρ ) C M ρ e 3 T ∥ H ∥ ∞ (402) \left|\mathbb{E}O_H-\Omega(H)-\sum_{j=1}^{R}g^{-j}\kappa_{j,n}(H)\right|
\le g^{-R-1}s^{R+2}(C_{M\rho})^{C_{M\rho}}e^{3T}\|H\|_\infty \tag*{(402)} E O H − Ω ( H ) − j = 1 ∑ R g − j κ j , n ( H ) ≤ g − R − 1 s R + 2 ( C M ρ ) C M ρ e 3 T ∥ H ∥ ∞ (402)
for bounded even Borel H H H supported in [ − T , T ] [-T,T] [ − T , T ] . The functionals O H O_H O H , Ω \Omega Ω and κ j , n \kappa_{j,n} κ j , n are recalled in Definition J.1 ; the coefficients are precisely those in Theorem 4.8, [101]. We justify the changed parameter domain, including the retained kernels, before transferring this estimate to the positive spectral trace.
Inputs from Huang’s expansion
This subsection contains no new mathematics. It fixes notation, introduces the orbital functional and the carrier bookkeeping of [101] (Definitions J.1 and J.2), and restates as Lemmas J.3–J.12 the estimates of that paper that the later subsections use, each with the hypotheses under which it is proved there or, for Lemmas J.9 and J.10, under the hypotheses of the deterministic parts of its proofs that we use. The paragraph at the end of the subsection lists the two estimates whose genus thresholds are quadratic in the cusp count and which are therefore replaced on the envelope s ≤ M g s \le M\sqrt{g} s ≤ M g .
Throughout, g ≥ 3 g \ge3 g ≥ 3 and n ≥ 0 n \ge0 n ≥ 0 are integers, N = 2 g + n − 3 N = 2g+n-3 N = 2 g + n − 3 is the factorial index of V g , n V_{g,n} V g , n , and A g , n = 2 π ( 2 g − 2 + n ) = area ( X ) A_{g,n}=2\pi(2g-2+n)=\operatorname{area}(X) A g , n = 2 π ( 2 g − 2 + n ) = area ( X ) . The fixed multiplier is that of Proposition 15.3: h = b 0 ∗ b 0 h=b_0*b_0 h = b 0 ∗ b 0 with supp h ⊂ [ − L , L ] \operatorname{supp}h\subset[-L,L] supp h ⊂ [ − L , L ] and ∥ h ∥ 1 = 1 \lVert h\rVert_1=1 ∥ h ∥ 1 = 1 , f = h ^ f=\widehat{h} f = h with H ^ ( r ) = ∫ R H ( u ) e − i r u d u \widehat{H}(r)=\int_{\mathbb{R}}H(u)e^{-iru}\,du H ( r ) = ∫ R H ( u ) e − i r u d u , and a = f ( i / 2 ) > 1 a=f(i/2)>1 a = f ( i /2 ) > 1 ; thus 0 ≤ f ≤ 1 0\le f\le1 0 ≤ f ≤ 1 on R \mathbb{R} R and 1 ≤ f ( i τ ) ≤ a 1\le f(i\tau)\le a 1 ≤ f ( i τ ) ≤ a for 0 ≤ τ ≤ 1 / 2 0\le\tau\le1/2 0 ≤ τ ≤ 1/2 . For a real polynomial p ( x ) = p ~ ( x ) p(x)=\widetilde{p}(x) p ( x ) = p ( x ) we write N ∗ ( p ) = ∥ p ~ ∥ [ 0 , a ] N_*(p)=\lVert\widetilde{p}\rVert_{[0,a]} N ∗ ( p ) = ∥ p ∥ [ 0 , a ] and N R ( p ) = ∥ p ~ ∥ [ 0 , 1 ] N_{\mathbb{R}}(p)=\lVert\widetilde{p}\rVert_{[0,1]} N R ( p ) = ∥ p ∥ [ 0 , 1 ] . Put ϕ ( x ) = sinh ( x / 2 ) / ( x / 2 ) \phi(x)=\sinh(x/2)/(x/2) ϕ ( x ) = sinh ( x /2 ) / ( x /2 ) , ϕ ( 0 ) = 1 \phi(0)=1 ϕ ( 0 ) = 1 . Volumes V u , v ( ℓ ) V_{u,v}(\ell) V u , v ( ℓ ) are Weil–Petersson orbifold volumes with V 0 , 3 = 1 V_{0,3}=1 V 0 , 3 = 1 and V 1 , 1 ( ℓ ) = ( ℓ 2 + 4 π 2 ) / 48 V_{1,1}(\ell)=(\ell^2+4\pi^2)/48 V 1 , 1 ( ℓ ) = ( ℓ 2 + 4 π 2 ) /48 ; a zero length denotes a cusp, and a type ( u , v ) (u,v) ( u , v ) is stable when its complexity c = 2 u + v − 2 c=2u+v-2 c = 2 u + v − 2 is at least one.
Definition J.1 (Orbital functional and canonical coefficients). Let H H H be a bounded even Borel function supported in [ − T , T ] [-T,T] [ − T , T ] . The orbital functional is
O H ( X ) = ∑ γ ∈ P o r ( X ) ∑ ν ≥ 1 ℓ γ H ( ν ℓ γ ) 2 sinh ( ν ℓ γ / 2 ) . (403) \mathcal{O}_{H}(X)=\sum_{\gamma\in\mathcal{P}_{\mathrm{or}}(X)}\sum_{\nu\ge1}\frac{\ell_\gamma H(\nu\ell_\gamma)}{2\sinh(\nu\ell_\gamma/2)}. \tag*{(403)} O H ( X ) = γ ∈ P or ( X ) ∑ ν ≥ 1 ∑ 2 sinh ( ν ℓ γ /2 ) ℓ γ H ( ν ℓ γ ) . (403)
where P o r ( X ) \mathcal{P}_{\mathrm{or}}(X) P or ( X ) is the set of primitive oriented closed geodesics, both orientations are counted, and all repetitions ν ≥ 1 \nu\ge1 ν ≥ 1 occur [101, (3.1)]; it is the functional O F s ( H ) \mathcal{O}_{F_s}(H) O F s ( H ) of (217) evaluated on X X X . The leading functional is Ω ( H ) = ∫ 0 ∞ ∑ ν ≥ 1 2 sinh 2 ( ℓ / 2 ) sinh ( ν ℓ / 2 ) H ( ν ℓ ) d ℓ \Omega(H)=\int_0^\infty\sum_{\nu\ge1}\frac{2\sinh^2(\ell/2)}{\sinh(\nu\ell/2)}H(\nu\ell)\,d\ell Ω ( H ) = ∫ 0 ∞ ∑ ν ≥ 1 s i n h ( ν ℓ /2 ) 2 s i n h 2 ( ℓ /2 ) H ( ν ℓ ) d ℓ [101, (3.20)]. By [101, Theorems 3.5 and 4.8] there are linear functionals κ j , n \kappa_{j,n} κ j , n , j ≥ 1 j\ge1 j ≥ 1 , on such tests, independent of g g g , of the truncation order R R R and of the support cutoff T T T , such that E g , n O H = Ω ( H ) + ∑ j = 1 R g − j κ j , n ( H ) + ( remainder ) \mathbb{E}_{g,n}\mathcal{O}_H=\Omega(H)+\sum_{j=1}^{R}g^{-j}\kappa_{j,n}(H)+(\text{remainder}) E g , n O H = Ω ( H ) + ∑ j = 1 R g − j κ j , n ( H ) + ( remainder ) under Huang’s genus thresholds; J.1 asserts the same expansion, with the same κ j , n \kappa_{j,n} κ j , n , on the present envelope. By [101, Proposition 3.8],
κ 1 , n = Λ 0 + n Λ 1 + ( n 2 ) Λ 2 (404) \kappa_{1,n}=\Lambda_0+n\Lambda_1+\binom{n}{2}\Lambda_2 \tag*{(404)} κ 1 , n = Λ 0 + n Λ 1 + ( 2 n ) Λ 2 (404)
with linear functionals Λ i \Lambda_i Λ i independent of n n n ; Λ 2 \Lambda_2 Λ 2 is the functional (217) . For a polynomial p ( x ) = ∑ t ≥ 1 a t x t p(x)=\sum_{t\ge1}a_tx^t p ( x ) = ∑ t ≥ 1 a t x t put
H p = ∑ t ≥ 1 a t h ∗ t , O X ( p ) = O H p ( X ) , ω 1 , n ( p ) = ω 1 ( p ) = Ω ( H p ) , ω j , n ( p ) = κ j − 1 , n ( H p ) ( j ≥ 2 ) , W i ( p ) = Λ i ( H p ) ( 0 ≤ i ≤ 2 ) . (405) \begin{aligned}
H_p&=\sum_{t\ge1}a_th^{*t},&
\mathcal{O}_X(p)&=\mathcal{O}_{H_p}(X),&
\omega_{1,n}(p)&=\omega_1(p)=\Omega(H_p),\\
\omega_{j,n}(p)&=\kappa_{j-1,n}(H_p)\quad(j\ge2),&
W_i(p)&=\Lambda_i(H_p)\quad(0\le i\le2). \tag*{(405)}
\end{aligned} H p ω j , n ( p ) = t ≥ 1 ∑ a t h ∗ t , = κ j − 1 , n ( H p ) ( j ≥ 2 ) , O X ( p ) W i ( p ) = O H p ( X ) , = Λ i ( H p ) ( 0 ≤ i ≤ 2 ) . ω 1 , n ( p ) = ω 1 ( p ) = Ω ( H p ) , (405)
so that H ^ p = p ∘ f \widehat{H}_p=p\circ f H p = p ∘ f , supp H p ⊂ [ − L Q , L Q ] \operatorname{supp}H_p\subset[-LQ,LQ] supp H p ⊂ [ − L Q , L Q ] when deg p ≤ Q \deg p\le Q deg p ≤ Q , ∥ h ∗ t ∥ ∞ ≤ ∥ h ∥ ∞ \lVert h^{*t}\rVert_\infty\le\lVert h\rVert_\infty ∥ h ∗ t ∥ ∞ ≤ ∥ h ∥ ∞ , and J.3 reads
ω 2 , n = W 0 + n W 1 + ( n 2 ) W 2 . (406) \omega_{2,n}=W_0+nW_1+\binom{n}{2}W_2. \tag*{(406)} ω 2 , n = W 0 + n W 1 + ( 2 n ) W 2 . (406)
These are [101, (3.43), (3.56) and (6.55)]. The leading coefficient ω 1 \omega_1 ω 1 does not depend on n n n [101, Proposition 5.1].
Definition J.2 (Carriers, patterns, tails and records). (i) Carriers. The normalized carrier Y Y Y of a nonsimple primitive geodesic γ \gamma γ on X X X is the subsurface filled by γ \gamma γ : thicken the geodesic graph, fill disks, complete peripheral punctured disks to the inherited cusps, and tighten the remaining boundary curves to geodesics, keeping two boundary slots when two of them tighten to the same ambient geodesic [101, Section 3.3]. Its type ( h 0 , a 0 , k ) (h_0,a_0,k) ( h 0 , a 0 , k ) records the genus, the number of inherited cusps and the number k ≥ 1 k\ge1 k ≥ 1 of positive boundary slots; its complexity is m = 2 h 0 + a 0 + k − 2 ≥ 1 m=2h_0+a_0+k-2\ge1 m = 2 h 0 + a 0 + k − 2 ≥ 1 .
By [101, Lemma 3.3], Λ : = ℓ ( ∂ Y ) ≤ 2 ℓ γ \Lambda:= \ell(\partial Y) \le2\ell_\gamma Λ := ℓ ( ∂ Y ) ≤ 2 ℓ γ and ℓ γ ≥ π m / 2 \ell_\gamma\ge\pi m/2 ℓ γ ≥ π m /2 . A carrier without stable complementary component has m = 2 g + n − 2 m = 2g + n - 2 m = 2 g + n − 2 , hence ℓ γ ≥ π ( g − 1 + n / 2 ) \ell_\gamma\ge\pi(g - 1 + n/2) ℓ γ ≥ π ( g − 1 + n /2 ) [101, (3.38)].
(ii) Patterns and complementary data. A gluing pattern P P P on the k k k slots consists of a matching of 2 ς 2\varsigma 2 ς slots into pairs, glued to one another, and a partition of the remaining slots into b ≥ 1 b \ge1 b ≥ 1 nonempty blocks I 1 , … , I b I_1,\ldots,I_b I 1 , … , I b , the block I i I_i I i being attached to the i i i th complementary component; k i = ∣ I i ∣ k_i = |I_i| k i = ∣ I i ∣ . The number of patterns is at most ( 4 t ) 4 t (4t)^{4t} ( 4 t ) 4 t whenever k ≤ t + 1 k \le t+1 k ≤ t + 1 . Complementary data are integers ( g i , a i ) i ≤ b (g_i,a_i)_{i\le b} ( g i , a i ) i ≤ b with
∑ i a i = n − a 0 , ∑ i g i = g − h 0 − k + ς + b , c i = 2 g i + a i + k i − 2 ≥ 1. \sum_i a_i = n-a_0,\qquad\sum_i g_i = g-h_0-k+\varsigma+b,\qquad c_i = 2g_i+a_i+k_i-2 \ge1. i ∑ a i = n − a 0 , i ∑ g i = g − h 0 − k + ς + b , c i = 2 g i + a i + k i − 2 ≥ 1.
We write R i = c i − 1 R_i = c_i-1 R i = c i − 1 , so that ∑ i R i = N − d \sum_i R_i = N-d ∑ i R i = N − d with d = m + b − 1 d=m+b-1 d = m + b − 1 , and define the deficit
δ = N − max i R i ≥ d ≥ m , \delta= N-\max_i R_i \ge d \ge m, δ = N − i max R i ≥ d ≥ m ,
the loss of factorial index on passing to a largest complementary component [101, (2.5)–(2.6), (2.11)]. The labeled weight and the complementary tail of order R R R are
W = n ! a 0 ! ∏ i a i ! V h 0 , a 0 + k ∏ i V g i , a i + k i V g , n , T g , n h 0 , a 0 , k ( R ) = ∑ P ∑ δ ≥ R + 1 W (407) W=\frac{n!}{a_0!\prod_i a_i!}\frac{V_{h_0,a_0+k}\prod_i V_{g_i,a_i+k_i}}{V_{g,n}},\qquad
\mathcal{T}_{g,n}^{h_0,a_0,k}(R)=\sum_P\sum_{\delta\ge R+1}W \tag*{(407)} W = a 0 ! ∏ i a i ! n ! V g , n V h 0 , a 0 + k ∏ i V g i , a i + k i , T g , n h 0 , a 0 , k ( R ) = P ∑ δ ≥ R + 1 ∑ W (407)
[101, (2.7)]. For a fixed pattern the boundary version assigns a length y e ≥ 0 y_e\ge0 y e ≥ 0 to each of the e P = k − ς e_P=k-\varsigma e P = k − ς actual edges, induces slot lengths x j x_j x j (paired slots share one length), replaces the numerator volumes by boundary volumes with these slot lengths and zero cusp lengths, and multiplies by the Jacobian J P = ∏ e y e J_P=\prod_e y_e J P = ∏ e y e ; then Λ = ∑ j = 1 k x j \Lambda=\sum_{j=1}^k x_j Λ = ∑ j = 1 k x j .
(iii) Separating tail. For a simple separating cut with a 0 a_0 a 0 cusps and genus i i i on a distinguished side put c 1 = 2 i + a 0 − 1 c_1=2i+a_0-1 c 1 = 2 i + a 0 − 1 , c 2 = 2 ( g − i ) + n − a 0 − 1 c_2=2(g-i)+n-a_0-1 c 2 = 2 ( g − i ) + n − a 0 − 1 and δ = min ( c 1 , c 2 ) \delta=\min(c_1,c_2) δ = min ( c 1 , c 2 ) . Then
S g , n ≥ r ( x ) = ∑ δ ≥ r ( n a 0 ) V i , a 0 + 1 ( 0 a 0 , x ) V g − i , n − a 0 + 1 ( 0 n − a 0 , x ) V g , n (408) S_{g,n}^{\ge r}(x)=\sum_{\delta\ge r}\binom{n}{a_0}\frac{V_{i,a_0+1}(0_{a_0},x)V_{g-i,n-a_0+1}(0_{n-a_0},x)}{V_{g,n}} \tag*{(408)} S g , n ≥ r ( x ) = δ ≥ r ∑ ( a 0 n ) V g , n V i , a 0 + 1 ( 0 a 0 , x ) V g − i , n − a 0 + 1 ( 0 n − a 0 , x ) (408)
is the ordered separating kernel restricted to deficit at least r r r [101, (3.3), (3.9), (3.13)].
(iv) Records. A retained record of a pattern consists of a distinguished block ∗ \ast ∗ and stable data ( g i , a i ) i ≠ ∗ (g_i,a_i)_{i\ne\ast} ( g i , a i ) i = ∗ ; the component attached to I ∗ I_\ast I ∗ is reconstructed from the ambient genus. For a record put
δ = m + ∑ i ≠ ∗ c i , A o u t = a 0 + ∑ i ≠ ∗ a i , H 0 = h 0 + k − ς − b + ∑ i ≠ ∗ g i , \delta=m+\sum_{i\ne\ast}c_i,\qquad A_{\mathrm{out}}=a_0+\sum_{i\ne\ast}a_i,\qquad H_0=h_0+k-\varsigma-b+\sum_{i\ne\ast}g_i, δ = m + i = ∗ ∑ c i , A out = a 0 + i = ∗ ∑ a i , H 0 = h 0 + k − ς − b + i = ∗ ∑ g i ,
k ∗ = ∣ I ∗ ∣ , p = n − A o u t + k ∗ , X ∗ = ∑ j ∈ I ∗ x j (409) k_\ast=|I_\ast|,\qquad p=n-A_{\mathrm{out}}+k_\ast,\qquad X_\ast=\sum_{j\in I_\ast}x_j \tag*{(409)} k ∗ = ∣ I ∗ ∣ , p = n − A out + k ∗ , X ∗ = j ∈ I ∗ ∑ x j (409)
Thus A o u t A_{\mathrm{out}} A out is the number of cusps outside the large component, H 0 H_0 H 0 its genus shift, p p p its number of marked ends, k ∗ k_\ast k ∗ its number of gluing boundaries and X ∗ X_\ast X ∗ their total length. One has 2 H 0 + A o u t − k ∗ = δ 2H_0+A_{\mathrm{out}}-k_\ast=\delta 2 H 0 + A out − k ∗ = δ , A o u t ≤ δ + 1 A_{\mathrm{out}}\le\delta+1 A out ≤ δ + 1 , p ≥ k ∗ ≥ 1 p\ge k_\ast\ge1 p ≥ k ∗ ≥ 1 , the large component has genus g − H 0 g-H_0 g − H 0 and factorial index N − δ N-\delta N − δ , and there are at most b ( δ + 3 ) 2 ( b − 1 ) b(\delta+3)^{2(b-1)} b ( δ + 3 ) 2 ( b − 1 ) records of deficit δ \delta δ [101, (2.40)–(2.42)]. For a truncation order R ≥ δ R\ge\delta R ≥ δ , J l o c = R − δ J_{\mathrm{loc}}=R-\delta J loc = R − δ denotes the order through which the shifted volume ratio of the large component is expanded.
Lemma J.3 (Volume envelope [101, Lemma 2.1]). Put A 0 = 4 π 2 A_0=4\pi^2 A 0 = 4 π 2 . There are absolute constants U ≥ 1 U\ge1 U ≥ 1 , β ν > 0 \beta_\nu>0 β ν > 0 and κ ν > 0 \kappa_\nu>0 κ ν > 0 such that, for every stable type ( u , ν ) (u,\nu) ( u , ν ) of complexity c c c , every g ≥ 2 g\ge2 g ≥ 2 , n ≥ 0 n\ge0 n ≥ 0 , and every length vector ℓ \ell ℓ ,
V u , ν ≤ U A 0 c − 1 ( c − 1 ) ! c , V g , n ≥ β ν A 0 N N ! g exp ( − κ ν n ( n + 1 ) 2 g ) , V u , ν ( ℓ ) ≤ V u , ν exp ( 1 2 ∑ j ℓ j ) . (410) V_{u,\nu}\le U A_0^{c-1}\frac{(c-1)!}{\sqrt{c}},\qquad
V_{g,n}\ge\beta_\nu A_0^N\frac{N!}{\sqrt{g}}\exp\left(-\frac{\kappa_\nu n(n+1)}{2g}\right),\qquad
V_{u,\nu}(\ell)\le V_{u,\nu}\exp\left(\frac12\sum_j\ell_j\right). \tag*{(410)} V u , ν ≤ U A 0 c − 1 c ( c − 1 )! , V g , n ≥ β ν A 0 N g N ! exp ( − 2 g κ ν n ( n + 1 ) ) , V u , ν ( ℓ ) ≤ V u , ν exp ( 2 1 j ∑ ℓ j ) . (410)
The lower bound is the only volume estimate in which the cusp count enters exponentially; the cusp-label counts ( b + 1 ) n (b+1)^n ( b + 1 ) n and 2 n 2^n 2 n of the deep sector are the second place where it does, and they are absorbed in (J.34). Huang imposes g ≥ κ v s 2 g \ge\kappa_v s^2 g ≥ κ v s 2 , which makes the exponential at least e − 1 / 2 e^{-1/2} e − 1/2 ; on the envelope s ≤ M g s \le M\sqrt{g} s ≤ M g it is at least e − κ v M 2 / 2 e^{-\kappa_v M^2/2} e − κ v M 2 /2 , see (J.30).
Lemma J.4 (Tail skeleton [101], proofs of Theorem 2.2 and Lemma 4.6). Let ( h 0 , a 0 , k ) (h_0,a_0,k) ( h 0 , a 0 , k ) be a carrier type of complexity m ≥ 1 m \ge1 m ≥ 1 , let R ≥ 0 R \ge0 R ≥ 0 , r = R + 1 r=R+1 r = R + 1 , t = R + m + 1 ≥ 2 t=R+m+1 \ge2 t = R + m + 1 ≥ 2 , D = max ( m , r ) D=\max(m,r) D = max ( m , r ) , J = ⌊ N / ( 64 s ) ⌋ J=\lfloor N/(64s)\rfloor J = ⌊ N / ( 64 s )⌋ and L s = log ( 64 s ) L_s=\log(64s) L s = log ( 64 s ) , and let Θ ≥ 1 \Theta\ge1 Θ ≥ 1 be such that the exponential in (J.9) is at least Θ − 1 \Theta^{-1} Θ − 1 . Assume (J.10) g ≥ 128 t s , g ≤ N ≤ 3 g , N − D + 1 ≥ g , J ≥ t , N / ( 128 s ) ≤ J ≤ N / 2. g \ge128ts,\qquad g \le N \le3g,\qquad N-D+1 \ge g,\qquad J \ge t,\qquad N/(128s) \le J \le N/2. g ≥ 128 t s , g ≤ N ≤ 3 g , N − D + 1 ≥ g , J ≥ t , N / ( 128 s ) ≤ J ≤ N /2.
Fix a pattern P P P with b b b blocks, so that b ≤ k ≤ m + 2 ≤ t + 1 b \le k \le m+2 \le t+1 b ≤ k ≤ m + 2 ≤ t + 1 , and put P b = Θ B b + 2 b ( m − 1 ) ! P_b=\Theta B^{b+2}\sqrt{b}(m-1)! P b = Θ B b + 2 b ( m − 1 )! with an absolute constant B ≥ 2 B \ge2 B ≥ 2 . Then:
(i) every labeled weight satisfies
W ≤ P b ∏ i R i ! N ! n ! a 0 ! ∏ i a i ! ; W \le P_b \frac{\prod_i R_i!}{N!}\frac{n!}{a_0!\prod_i a_i!}; W ≤ P b N ! ∏ i R i ! a 0 ! ∏ i a i ! n ! ; (ii) A o u t ≤ ϑ + 1 A_{\mathrm{out}} \le\vartheta+1 A out ≤ ϑ + 1 , the multinomial in (i) is at most min { s ϑ + 1 , ( b + 1 ) n } \min\{s^{\vartheta+1},(b+1)^n\} min { s ϑ + 1 , ( b + 1 ) n } , and the total weight at deficit ϑ \vartheta ϑ is at most P b b x ϑ P_b b x_\vartheta P b b x ϑ , where
x ϑ = s ϑ + 1 ( ϑ + 3 ) 2 b ( N ϑ ) ; x_\vartheta=s^{\vartheta+1}(\vartheta+3)^{2b}\binom{N}{\vartheta}; x ϑ = s ϑ + 1 ( ϑ + 3 ) 2 b ( ϑ N ) ; (iii) (near sector) for D ≤ ϑ < J D \le\vartheta< J D ≤ ϑ < J one has x ϑ + 1 / x ϑ ≤ e 2 / 63 < 1 / 2 x_{\vartheta+1}/x_\vartheta\le e^2/63 < 1/2 x ϑ + 1 / x ϑ ≤ e 2 /63 < 1/2 , and the sum over D ≤ ϑ ≤ J D \le\vartheta\le J D ≤ ϑ ≤ J is at most
2 P b b ( D + 3 ) 2 b D ! g − D s D + 1 ; 2P_b b(D+3)^{2b}D!g^{-D}s^{D+1}; 2 P b b ( D + 3 ) 2 b D ! g − D s D + 1 ; (iv) (deep sector) for b ≥ 2 b \ge2 b ≥ 2 the sum over ϑ > J \vartheta>J ϑ > J is at most
P b ( b + 1 ) n ( N + 1 ) 2 b ( N J ) − 1 , P_b(b+1)^n(N+1)^{2b}\binom{N}{J}^{-1}, P b ( b + 1 ) n ( N + 1 ) 2 b ( J N ) − 1 , and
( N J ) ≥ ( 64 s ) J ≥ exp ( N L s / ( 128 s ) ) ; \binom{N}{J}\ge(64s)^J\ge\exp(NL_s/(128s)); ( J N ) ≥ ( 64 s ) J ≥ exp ( N L s / ( 128 s )) ; for b = 1 b=1 b = 1 the deep sector is empty;
(v)
r log g + n log ( b + 1 ) + 2 b log ( N + 1 ) ≤ F t ( g ) : = 4 t log g + s log ( 2 t ) + 3 t log 4 ; r\log g+n\log(b+1)+2b\log(N+1)\le F_t(g):=4t\log g+s\log(2t)+3t\log4; r log g + n log ( b + 1 ) + 2 b log ( N + 1 ) ≤ F t ( g ) := 4 t log g + s log ( 2 t ) + 3 t log 4 ; (vi) there are at most ( 4 t ) 4 t (4t)^{4t} ( 4 t ) 4 t patterns, and
B b + 2 b ( m − 1 ) ! ≤ ( C t ) 3 t , 2 b ( D + 3 ) 2 b D ! ≤ ( 2 t ) 5 t ; B^{b+2}\sqrt{b}(m-1)!\le(Ct)^{3t},\qquad2b(D+3)^{2b}D!\le(2t)^{5t}; B b + 2 b ( m − 1 )! ≤ ( C t ) 3 t , 2 b ( D + 3 ) 2 b D ! ≤ ( 2 t ) 5 t ; (vii) in the boundary version each of the bounds above acquires the factor ( 1 + Λ ) k e x (1+\Lambda)^k e^x ( 1 + Λ ) k e x .
For the separating kernel, under (J.10) with t = r ≥ 1 t=r\ge1 t = r ≥ 1 : each zero-length summand of (J.7) at deficit ϑ \vartheta ϑ is at most B Θ ( N a 0 ) / ( N ϑ ) B\Theta\binom{N}{a_0}/\binom{N}{\vartheta} B Θ ( a 0 N ) / ( ϑ N ) ; the sum at fixed deficit is at most 6 B Θ s ϑ + 1 / ( N ϑ ) 6B\Theta s^{\vartheta+1}/\binom{N}{\vartheta} 6 B Θ s ϑ + 1 / ( ϑ N ) , with successive ratio at most 1 / 63 1/63 1/63 for r ≤ ϑ < J r\le\vartheta<J r ≤ ϑ < J , so the near sector r ≤ ϑ ≤ J r\le\vartheta\le J r ≤ ϑ ≤ J is at most 12 B Θ r ! g − r s r + 1 12B\Theta r!g^{-r}s^{r+1} 12 B Θ r ! g − r s r + 1 ; the deep sector ϑ > J \vartheta>J ϑ > J is at most B Θ ( g + 1 ) 2 n ( N J ) − 1 B\Theta(g+1)^{2n}\binom{N}{J}^{-1} B Θ ( g + 1 ) 2 n ( J N ) − 1 , and log ( g r ( g + 1 ) 2 n ) ≤ F r ( g ) \log(g^r(g+1)^{2n})\le F_r(g) log ( g r ( g + 1 ) 2 n ) ≤ F r ( g ) ; boundary lengths cost the factor e x e^x e x .
Proof. These are the displayed steps (2.12)–(2.19) and (4.19)–(4.24) in Huang’s proofs of Theorem 2.2 and Lemma 4.6, and (3.10)–(3.12) for the separating kernel, with the following bookkeeping of hypotheses. (i) uses (J.9) for every factor of W W W : the largest complementary complexity is at least ( N + 1 − m ) / b ≥ g / b (N+1-m)/b\ge g/b ( N + 1 − m ) / b ≥ g / b because g ≥ m + 2 g\ge m+2 g ≥ m + 2 , which cancels the ambient g \sqrt{g} g up to b \sqrt{b} b ; the powers of A 0 A_0 A 0 total − b -b − b ; and the exponential of the lower bound is absorbed into Θ \Theta Θ . (ii) is combinatorial. (iii) uses ϑ + 1 ≤ J ≤ N / ( 64 s ) \vartheta+1\le J\le N/(64s) ϑ + 1 ≤ J ≤ N / ( 64 s ) , hence s ( ϑ + 1 ) / ( N − ϑ ) ≤ 1 / 63 s(\vartheta+1)/(N-\vartheta)\le1/63 s ( ϑ + 1 ) / ( N − ϑ ) ≤ 1/63 , together with ( 1 + 1 / ( ϑ + 3 ) ) 2 b ≤ e 2 (1+1/(\vartheta+3))^{2b}\le e^2 ( 1 + 1/ ( ϑ + 3 ) ) 2 b ≤ e 2 from b ≤ ϑ + 2 b\le\vartheta+2 b ≤ ϑ + 2 , and ( N D ) = ( N N − D ) / D ! ≥ g D / D ! \binom{N}{D}=\binom{N}{N-D}/D!\ge g^D/D! ( D N ) = ( N − D N ) / D ! ≥ g D / D ! from N − D + 1 ≥ g N-D+1\ge g N − D + 1 ≥ g . (iv) uses the capped factorial inequality ∏ i z i ! ≤ ( N − J ) ! J ! \prod_i z_i!\le(N-J)!J! ∏ i z i ! ≤ ( N − J )! J ! for nonnegative integers with ∑ i z i ≤ N \sum_i z_i\le N ∑ i z i ≤ N , z i ≤ N − J z_i\le N-J z i ≤ N − J , 1 ≤ J ≤ N / 2 1\le J\le N/2 1 ≤ J ≤ N /2 [101, (2.18)], applied to the R i R_i R i (all of which are smaller than N − J N-J N − J when ϑ > J \vartheta>J ϑ > J ), the count ( N + 1 ) 2 b (N+1)^{2b} ( N + 1 ) 2 b of deep allocations, and N / J ≥ 64 s N/J\ge64s N / J ≥ 64 s , J ≥ N / ( 128 s ) J\ge N/(128s) J ≥ N / ( 128 s ) . (v) uses N + 1 ≤ 4 g N+1\le4g N + 1 ≤ 4 g , r + 2 b ≤ 4 t r+2b\le4t r + 2 b ≤ 4 t , b + 1 ≤ 2 t b+1\le2t b + 1 ≤ 2 t and 2 b ≤ 3 t 2b\le3t 2 b ≤ 3 t . (vi) is elementary and (vii) is the boundary version estimate in (J.9) together with ( 1 + Λ ) k (1+\Lambda)^k ( 1 + Λ ) k [101, (2.10)]. The separating statements use (J.9) in the form [101, (3.10)], the bound ( n a 0 ) ≤ s ϑ + 1 \binom{n}{a_0}\le s^{\vartheta+1} ( a 0 n ) ≤ s ϑ + 1 for the smaller side, at most 2 ( ϑ + 2 ) 2(\vartheta+2) 2 ( ϑ + 2 ) ordered records at fixed deficit, binomial unimodality for ϑ > J \vartheta>J ϑ > J , and ∑ a 0 ( n a 0 ) = 2 n \sum_{a_0}\binom{n}{a_0}=2^n ∑ a 0 ( a 0 n ) = 2 n with at most g + 1 g+1 g + 1 genus choices.
In [101] the single hypothesis g ≥ K ( s 2 + t 2 s ) g\ge K(s^2+t^2s) g ≥ K ( s 2 + t 2 s ) , K ≥ max { 4096 , κ v } K\ge\max\{4096,\kappa_v\} K ≥ max { 4096 , κ v } , serves two purposes: it yields (J.10) with Θ = e 1 / 2 \Theta=e^{1/2} Θ = e 1/2 [101, (4.20)], and it closes the deep sector through the absorption inequality F t ( g ) < g L s / ( 128 s ) F_t(g)<gL_s/(128s) F t ( g ) < g L s / ( 128 s ) [101, (4.25)]. Only these two uses depend on a quadratic genus condition; they are replaced in Subsection J.2. □
Lemma J.5 (Retained records and orbital integration [101, Lemmas 4.1–4.2, Propositions 3.4 and 4.7, Theorem 4.8]). (a) Graded shifted-volume error. There is an absolute C ≥ 2 C\ge2 C ≥ 2 such that, for
integers J l o c , H 0 ≥ 0 J_{\mathrm{loc}}, H_{0} \ge0 J loc , H 0 ≥ 0 and δ ≥ 1 \delta\ge1 δ ≥ 1 with p = n + 2 H 0 − δ ≥ 1 p = n + 2H_{0} - \delta\ge1 p = n + 2 H 0 − δ ≥ 1 , for L s h = J l o c + H 0 + δ + 1 L_{\mathrm{sh}} = J_{\mathrm{loc}} + H_{0} + \delta+ 1 L sh = J loc + H 0 + δ + 1 , 0 ≤ ℓ ≤ p 0 \le\ell\le p 0 ≤ ℓ ≤ p , x ∈ [ 0 , ∞ ) ℓ \mathbf{x} \in[0,\infty)^{\ell} x ∈ [ 0 , ∞ ) ℓ , X = ∑ i x i X = \sum_i x_i X = ∑ i x i and g ≥ [ C L s h ] C s g \ge[CL_{\mathrm{sh}}]^{C_s} g ≥ [ C L sh ] C s ,
∣ V g − H 0 , p ( 0 p − ℓ , x ) V g , n − ( 8 π 2 ) − δ g − δ ∑ j = 0 J l o c g − j Q j ( x ) ∣ ≤ g − δ − J l o c − 1 s J l o c + 1 [ C L s h ] C ( J l o c + 1 ) ( 1 + X ) 2 J l o c + 2 e X / 2 (411) \left|\frac{V_{g-H_{0},p}(\mathbf{0}_{p-\ell},\mathbf{x})}{V_{g,n}}-(8\pi^2)^{-\delta}g^{-\delta}\sum_{j=0}^{J_{\mathrm{loc}}}g^{-j}Q_j(\mathbf{x})\right|
\le g^{-\delta-J_{\mathrm{loc}}-1}s^{J_{\mathrm{loc}}+1}[CL_{\mathrm{sh}}]^{C(J_{\mathrm{loc}}+1)}(1+X)^{2J_{\mathrm{loc}}+2}e^{X/2} \tag*{(411)} V g , n V g − H 0 , p ( 0 p − ℓ , x ) − ( 8 π 2 ) − δ g − δ j = 0 ∑ J loc g − j Q j ( x ) ≤ g − δ − J loc − 1 s J loc + 1 [ C L sh ] C ( J loc + 1 ) ( 1 + X ) 2 J loc + 2 e X /2 (411)
where the Q j = Q j n , H 0 , δ ; ℓ Q_j=Q_j^{n,H_0,\delta;\ell} Q j = Q j n , H 0 , δ ; ℓ are canonical functions independent of g g g and J l o c J_{\mathrm{loc}} J loc , with ∣ Q j ( x ) ∣ ≤ s j [ C L s h ] C j ( 1 + X ) 2 j e X / 2 |Q_j(\mathbf{x})|\le s^j[CL_{\mathrm{sh}}]^{Cj}(1+X)^{2j}e^{X/2} ∣ Q j ( x ) ∣ ≤ s j [ C L sh ] C j ( 1 + X ) 2 j e X /2 . The boundary expansion of V G , p ( 0 p − ℓ , x ) / V G , p V_{G,p}(\mathbf{0}_{p-\ell},\mathbf{x})/V_{G,p} V G , p ( 0 p − ℓ , x ) / V G , p through order J J J holds for G ≥ [ C ( J + 1 ) ] C ( p + 1 ) G\ge[C(J+1)]^{C(p+1)} G ≥ [ C ( J + 1 ) ] C ( p + 1 ) [101, Lemma 4.1], and the handle-ratio and source thresholds of [101, Proposition 2.3] used inside these lemmas are bounded by [ C t ] C s [Ct]^{C_s} [ C t ] C s whenever p ≤ n + 3 t p\le n+3t p ≤ n + 3 t and the order is at most t t t [101, proof of Proposition 4.7]. In particular every retained-record genus threshold used here from [101, Section 4.1] is of the form [ C t ] C s [Ct]^{C_s} [ C t ] C s .
(b) Small-volume prefactor. For a record e e e of deficit δ \delta δ , without any genus condition,
A e ( y ) = J P V h 0 , a 0 + k ( 0 a 0 , x ) n ! a 0 ! ( n − A o u t ) ! ∏ i ≠ ∗ a i ! ∏ i ≠ ∗ V g i , a i + k i ( 0 a i , x I i ) , (412) A_e(\mathbf{y})=J_PV_{h_0,a_0+k}(\mathbf{0}_{a_0},\mathbf{x})\frac{n!}{a_0!(n-A_{\mathrm{out}})!\prod_{i\ne*}a_i!}\prod_{i\ne*}V_{g_i,a_i+k_i}(\mathbf{0}_{a_i},\mathbf{x}_{I_i}), \tag*{(412)} A e ( y ) = J P V h 0 , a 0 + k ( 0 a 0 , x ) a 0 ! ( n − A out )! ∏ i = ∗ a i ! n ! i = ∗ ∏ V g i , a i + k i ( 0 a i , x I i ) , (412)
∣ A e ( y ) ∣ ≤ s A o u t [ C ( δ + 1 ) ] C ( δ + 1 ) ( 1 + Λ ) k e Λ − X ∗ / 2 |A_e(\mathbf{y})|\le s^{A_{\mathrm{out}}}[C(\delta+1)]^{C(\delta+1)}(1+\Lambda)^k e^{\Lambda-X_*/2} ∣ A e ( y ) ∣ ≤ s A out [ C ( δ + 1 ) ] C ( δ + 1 ) ( 1 + Λ ) k e Λ − X ∗ /2
[101, (2.43), (4.13)].
(c) Kernel errors. Let t = R + m + 1 t=R+m+1 t = R + m + 1 . For a record of deficit δ ≤ R \delta\le R δ ≤ R take J l o c = R − δ J_{\mathrm{loc}}=R-\delta J loc = R − δ ; then H 0 ≤ t H_0\le t H 0 ≤ t , L s h = R + H 0 + 1 ≤ 2 t L_{\mathrm{sh}}=R+H_0+1\le2t L sh = R + H 0 + 1 ≤ 2 t and p ≥ k ∗ ≥ 1 p\ge k_*\ge1 p ≥ k ∗ ≥ 1 . The product of (J.11) and (J.12) has cusp power s A o u t + J l o c + 1 ≤ s R + 2 s^{A_{\mathrm{out}}+J_{\mathrm{loc}}+1}\le s^{R+2} s A out + J loc + 1 ≤ s R + 2 , exponential factor e Λ e^\Lambda e Λ and length power at most k + 2 ( R − δ ) + 2 ≤ 2 t k+2(R-\delta)+2\le2t k + 2 ( R − δ ) + 2 ≤ 2 t . Reconstruction of the large component requires N > 2 R N>2R N > 2 R and g > H 0 g>H_0 g > H 0 ; unavailable cusp allocations are omitted before any volume is evaluated. The sum over records of the retained terms is Huang’s canonical coefficient F j , n h 0 , a 0 , k ; P F_{j,n}^{h_0,a_0,k;P} F j , n h 0 , a 0 , k ; P [101, (2.45)], and the full nonnegative kernel
F g , n h 0 , a 0 , k ; P ( y ) = J P V h 0 , a 0 + k ( 0 a 0 , x ) ∑ n ! a 0 ! ∏ i a i ! ∏ i V g i , a i + k i ( 0 a i , x I i ) V g , n (413) F_{g,n}^{h_0,a_0,k;P}(\mathbf{y})=J_PV_{h_0,a_0+k}(\mathbf{0}_{a_0},\mathbf{x})\sum\frac{n!}{a_0!\prod_i a_i!}\frac{\prod_iV_{g_i,a_i+k_i}(\mathbf{0}_{a_i},\mathbf{x}_{I_i})}{V_{g,n}} \tag*{(413)} F g , n h 0 , a 0 , k ; P ( y ) = J P V h 0 , a 0 + k ( 0 a 0 , x ) ∑ a 0 ! ∏ i a i ! n ! V g , n ∏ i V g i , a i + k i ( 0 a i , x I i ) (413)
differs from ∑ j = 1 R g − j F j , n h 0 , a 0 , k ; P \sum_{j=1}^{R}g^{-j}F_{j,n}^{h_0,a_0,k;P} ∑ j = 1 R g − j F j , n h 0 , a 0 , k ; P by at most the sum of the record errors plus the boundary version of T g , n h 0 , a 0 , k ( R ) T_{g,n}^{h_0,a_0,k}(R) T g , n h 0 , a 0 , k ( R ) for the pattern. The nonseparating kernel U g , n ( x ) = V g − 1 , n + 2 ( 0 n , x , x ) / V g , n U_{g,n}(x)=V_{g-1,n+2}(\mathbf{0}_n,x,x)/V_{g,n} U g , n ( x ) = V g − 1 , n + 2 ( 0 n , x , x ) / V g , n is the product of a scalar handle ratio and a boundary ratio and has no tail; its expansion through order R R R has error g − R − 1 s R + 1 [ C r ] C r ( 1 + x ) 2 R + 2 e x g^{-R-1}s^{R+1}[C_r]^{C_r}(1+x)^{2R+2}e^x g − R − 1 s R + 1 [ C r ] C r ( 1 + x ) 2 R + 2 e x under g ≥ [ C r ] C s g\ge[C_r]^{C_s} g ≥ [ C r ] C s . The separating kernel splits into retained endpoints δ ≤ R \delta\le R δ ≤ R , treated by (J.11) with H 0 = h 0 ≤ ( δ + 1 ) / 2 H_0=h_0\le(\delta+1)/2 H 0 = h 0 ≤ ( δ + 1 ) /2 and the prefactor V h 0 , a 0 + 1 ( 0 a 0 , x ) ≤ s a 0 [ C ( δ + 1 ) ] C ( δ + 1 ) e x / 2 V_{h_0,a_0+1}(\mathbf{0}_{a_0},x)\le s^{a_0}[C(\delta+1)]^{C(\delta+1)}e^{x/2} V h 0 , a 0 + 1 ( 0 a 0 , x ) ≤ s a 0 [ C ( δ + 1 ) ] C ( δ + 1 ) e x /2 , and the omitted tail S g , n ≥ R + 1 ( x ) S_{g,n}^{\ge R+1}(x) S g , n ≥ R + 1 ( x ) .
(d) Integration. With K g , n = U g , n + S g , n K_{g,n}=U_{g,n}+S_{g,n} K g , n = U g , n + S g , n , where S g , n S_{g,n} S g , n is the full ordered separating kernel,
E g , n O H s i m p = ∫ 0 ∞ ∑ ν ≥ 1 ℓ 2 H ( ν ℓ ) 2 sinh ( ν ℓ / 2 ) K g , n ( ℓ ) d ℓ , (414) \mathbb{E}_{g,n}O^{\mathrm{simp}}_H=\int_0^\infty\sum_{\nu\ge1}\frac{\ell^2H(\nu\ell)}{2\sinh(\nu\ell/2)}K_{g,n}(\ell)\,d\ell, \tag*{(414)} E g , n O H simp = ∫ 0 ∞ ν ≥ 1 ∑ 2 sinh ( ν ℓ /2 ) ℓ 2 H ( ν ℓ ) K g , n ( ℓ ) d ℓ , (414)
and a kernel bound ∣ k ( ℓ ) ∣ ≤ A ( 1 + ℓ ) d e ℓ |k(\ell)|\le A(1+\ell)^d e^\ell ∣ k ( ℓ ) ∣ ≤ A ( 1 + ℓ ) d e ℓ gives
∣ ∫ 0 ∞ ∑ ν ≥ 1 ℓ 2 k ( ℓ ) H ( ν ℓ ) 2 sinh ( ν ℓ / 2 ) d ℓ ∣ ≤ ζ ( 3 ) 2 A T 2 ( 1 + T ) d e T ∥ H ∥ ∞ \left|\int_0^\infty\sum_{\nu\ge1}\frac{\ell^2k(\ell)H(\nu\ell)}{2\sinh(\nu\ell/2)}\,d\ell\right|
\le\frac{\zeta(3)}{2}AT^2(1+T)^d e^T\|H\|_\infty ∫ 0 ∞ ν ≥ 1 ∑ 2 sinh ( ν ℓ /2 ) ℓ 2 k ( ℓ ) H ( ν ℓ ) d ℓ ≤ 2 ζ ( 3 ) A T 2 ( 1 + T ) d e T ∥ H ∥ ∞
[101, Lemma 3.2 and (3.19)]. For a carrier type with at least one stable complementary component,
E g , n O H h 0 , a 0 , k = 1 k ! ∑ P ∫ ( 0 , ∞ ) P Ψ H h 0 , a 0 , k ( x P ( y ) ) F g , n h 0 , a 0 , k ; P ( y ) d y , (415) \mathbb{E}_{g,n}O_H^{h_0,a_0,k}=\frac{1}{k!}\sum_P\int_{(0,\infty)^P}\Psi_H^{h_0,a_0,k}(\mathbf{x}_P(\mathbf{y}))F_{g,n}^{h_0,a_0,k;P}(\mathbf{y})\,d\mathbf{y}, \tag*{(415)} E g , n O H h 0 , a 0 , k = k ! 1 P ∑ ∫ ( 0 , ∞ ) P Ψ H h 0 , a 0 , k ( x P ( y )) F g , n h 0 , a 0 , k ; P ( y ) d y , (415)
where Ψ H \Psi_H Ψ H is the fiber integral over the carrier moduli space with the prescribed boundary lengths, of the sum (J.2) restricted to filling geodesics, divided by the carrier volume; it satisfies ∣ Ψ H ∣ ≤ C m ( 1 + T ) e T ∥ H ∥ ∞ |\Psi_H|\le Cm(1+T)e^T\|H\|_\infty ∣ Ψ H ∣ ≤ C m ( 1 + T ) e T ∥ H ∥ ∞ and vanishes unless Λ ≤ 2 T \Lambda\le2T Λ ≤ 2 T and m ≤ 2 T / π m\le2T/\pi m ≤ 2 T / π , and Leb { Λ P ≤ 2 T } ≤ ( 1 + 2 T ) k \operatorname{Leb}\{\Lambda_P\le2T\}\le(1+2T)^k Leb { Λ P ≤ 2 T } ≤ ( 1 + 2 T ) k [101, (3.23), (3.25), (3.29), (3.33)]. At fixed m m m there are at most ( m + 2 ) 2 (m+2)^2 ( m + 2 ) 2 carrier types and k ≤ m + 2 k\le m+2 k ≤ m + 2 . The vanishing rules [101, (3.32)] identify the sum of the retained coefficients over all types with κ j , n \kappa_{j,n} κ j , n for every R R R and T T T . Apart from the exclusion of carriers without stable complementary component, which requires π ( g − 1 + n / 2 ) > T \pi(g-1+n/2)>T π ( g − 1 + n /2 ) > T , no statement in (d) involves a genus condition.
Lemma J.6 (Truncated positive trace [101, Theorem 5.3, Propositions 5.1 and 5.4, (5.44)–(5.45)]).
Let X X X be a complete hyperbolic surface of finite area with n ≥ 1 n\ge1 n ≥ 1 cusps and no orbifold points, let A X = f ( Δ X − 1 / 4 ) A_X=f(\sqrt{\Delta_X-1/4}) A X = f ( Δ X − 1/4 ) be the bounded self-adjoint operator with Laplace multiplier f ( r ( λ ) ) f(r(\lambda)) f ( r ( λ )) , r ( λ ) = λ − 1 / 4 r(\lambda)=\sqrt{\lambda-1/4} r ( λ ) = λ − 1/4 for λ ≥ 1 / 4 \lambda\ge1/4 λ ≥ 1/4 and r ( λ ) = i 1 / 4 − λ r(\lambda)=i\sqrt{1/4-\lambda} r ( λ ) = i 1/4 − λ for 0 ≤ λ < 1 / 4 0\le\lambda<1/4 0 ≤ λ < 1/4 , so that 0 ≤ A X ≤ a 0\le A_X\le a 0 ≤ A X ≤ a , the constant eigenfunction has A X A_X A X -eigenvalue a a a , and the spectrum of Δ X \Delta_X Δ X in [ 1 / 4 , ∞ ) [1/4,\infty) [ 1/4 , ∞ ) is mapped into [ 0 , 1 ] [0,1] [ 0 , 1 ] . For Y > 1 Y>1 Y > 1 let C b ( Y ) C_b(Y) C b ( Y ) be the width-one cusp regions above height Y Y Y , which are embedded and pairwise disjoint, let X ( Y ) = X ∖ ⋃ b C b ( Y ) X(Y)=X\setminus\bigcup_b C_b(Y) X ( Y ) = X ∖ ⋃ b C b ( Y ) , χ Y = 1 X ( Y ) \chi_Y=1_{X(Y)} χ Y = 1 X ( Y ) , and fix
Y Q = 3 e L Q ( Q ≥ 1 ) (416) Y_Q=3e^{LQ}\qquad(Q\ge1) \tag*{(416)} Y Q = 3 e L Q ( Q ≥ 1 ) (416) [101, (5.23), (6.1)]. Let p p p be a real polynomial with p ( 0 ) = 0 p(0)=0 p ( 0 ) = 0 and deg p ≤ Q \deg p\le Q deg p ≤ Q .
(i) The compression χ Y Q p ( A X ) χ Y Q \chi_{Y_Q}p(A_X)\chi_{Y_Q} χ Y Q p ( A X ) χ Y Q is trace class and
τ X , Q ( p ) : = Tr ( χ Y Q p ( A X ) χ Y Q ) = ( A g , n − n Y Q ) I ( p ) + n P Y Q ( p ) + O X ( p ) , (417) \tau_{X,Q}(p):=\operatorname{Tr}\bigl(\chi_{Y_Q}p(A_X)\chi_{Y_Q}\bigr)=\left(A_{g,n}-\frac{n}{Y_Q}\right)\mathcal{I}(p)+nP_{Y_Q}(p)+\mathcal{O}_X(p), \tag*{(417)} τ X , Q ( p ) := Tr ( χ Y Q p ( A X ) χ Y Q ) = ( A g , n − Y Q n ) I ( p ) + n P Y Q ( p ) + O X ( p ) , (417) where I ( p ) = 1 4 π ∫ R r tanh ( π r ) p ( f ( r ) ) d r \mathcal{I}(p)=\frac{1}{4\pi}\int_{\mathbb{R}}r\tanh(\pi r)p(f(r))\,dr I ( p ) = 4 π 1 ∫ R r tanh ( π r ) p ( f ( r )) d r and P Y ( p ) = 2 ∫ 0 L Q K H p ( ϱ ) cosh ( ϱ / 2 ) H ⌊ 2 Y sinh ( ϱ / 2 ) ⌋ d ϱ P_Y(p)=2\int_0^{LQ}K_{H_p}(\varrho)\cosh(\varrho/2)H_{\lfloor2Y\sinh(\varrho/2)\rfloor}\,d\varrho P Y ( p ) = 2 ∫ 0 L Q K H p ( ϱ ) cosh ( ϱ /2 ) H ⌊ 2 Y s i n h ( ϱ /2 )⌋ d ϱ , with K H K_H K H the Abel transform K H ( ϱ ) = − 1 2 π ∫ ϱ ∞ H ′ ( u ) d u cosh u − cosh ϱ K_H(\varrho)=-\frac{1}{\sqrt{2\pi}}\int_\varrho^\infty\frac{H'(u)\,du}{\sqrt{\cosh u-\cosh\varrho}} K H ( ϱ ) = − 2 π 1 ∫ ϱ ∞ c o s h u − c o s h ϱ H ′ ( u ) d u and H m = ∑ k = 1 m k − 1 H_m=\sum_{k=1}^m k^{-1} H m = ∑ k = 1 m k − 1 .
(ii) If p ≥ 0 p\ge0 p ≥ 0 on the spectrum of A X A_X A X then τ X , Q ( p ) ≥ 0 \tau_{X,Q}(p)\ge0 τ X , Q ( p ) ≥ 0 ; in general ∣ τ X , Q ( p ) ∣ ≤ N ∗ ( p ) τ X , Q ( x ) |\tau_{X,Q}(p)|\le N_*(p)\tau_{X,Q}(x) ∣ τ X , Q ( p ) ∣ ≤ N ∗ ( p ) τ X , Q ( x ) . More generally, if ψ = x ξ \psi=x\xi ψ = x ξ with ξ \xi ξ bounded and Borel on [ 0 , a ] [0,a] [ 0 , a ] , then χ Y Q ψ ( A X ) χ Y Q \chi_{Y_Q}\psi(A_X)\chi_{Y_Q} χ Y Q ψ ( A X ) χ Y Q is trace class and its trace, also after insertion of 1 − P 0 1-P_0 1 − P 0 with P 0 P_0 P 0 as in (v), has absolute value at most ∥ ξ ∥ [ 0 , a ] τ X , Q ( x ) \|\xi\|_{[0,a]}\tau_{X,Q}(x) ∥ ξ ∥ [ 0 , a ] τ X , Q ( x ) [101, (5.29)].
(iii) There are constants I f I_f I f , J f J_f J f , b f b_f b f depending only on f f f with
∣ I ( p ) ∣ ≤ I f N R ( p ) , ∣ P Y Q ( p ) ∣ ≤ J f Q 2 N R ( p ) , ∣ P Y Q ( x ) ∣ ≤ b f Q . (418) |\mathcal{I}(p)|\le I_fN_{\mathbb{R}}(p),\qquad|P_{Y_Q}(p)|\le J_fQ^2N_{\mathbb{R}}(p),\qquad|P_{Y_Q}(x)|\le b_fQ. \tag*{(418)} ∣ I ( p ) ∣ ≤ I f N R ( p ) , ∣ P Y Q ( p ) ∣ ≤ J f Q 2 N R ( p ) , ∣ P Y Q ( x ) ∣ ≤ b f Q . (418) (iv) There is a finite signed measure μ 0 \mu_0 μ 0 on [ 0 , 1 ] [0,1] [ 0 , 1 ] , independent of n n n , with ∥ μ 0 ∥ T V < 5 \|\mu_0\|_{\mathrm{TV}}<5 ∥ μ 0 ∥ TV < 5 and ω 1 ( p ) = p ( a ) + ∫ [ 0 , 1 ] p d μ 0 \omega_1(p)=p(a)+\int_{[0,1]}p\,d\mu_0 ω 1 ( p ) = p ( a ) + ∫ [ 0 , 1 ] p d μ 0 ; in particular ∣ ω 1 ( x ) ∣ ≤ 6 a |\omega_1(x)|\le6a ∣ ω 1 ( x ) ∣ ≤ 6 a .
(v) Let P 0 P_0 P 0 be the orthogonal projection onto constants. The compressed trace τ X , Q ∘ ( p ) : = Tr ( χ Y Q p ( A X ) ( 1 − P 0 ) χ Y Q ) = τ X , Q ( p ) − ( 1 − n A g , n Y Q ) p ( a ) \tau^\circ_{X,Q}(p):=\operatorname{Tr}\bigl(\chi_{Y_Q}p(A_X)(1-P_0)\chi_{Y_Q}\bigr)=\tau_{X,Q}(p)-\left(1-\frac{n}{A_{g,n}Y_Q}\right)p(a) τ X , Q ∘ ( p ) := Tr ( χ Y Q p ( A X ) ( 1 − P 0 ) χ Y Q ) = τ X , Q ( p ) − ( 1 − A g , n Y Q n ) p ( a ) is nonnegative when p ≥ 0 p\ge0 p ≥ 0 on [ 0 , a ] [0,a] [ 0 , a ] , and
τ X , Q ∘ ( p ) = ( A g , n − n Y Q ) I ( p ) + n P Y Q ( p ) + μ 0 ( p ) + ( O X ( p ) − ω 1 ( p ) ) + n p ( a ) A g , n Y Q . (419) \tau^\circ_{X,Q}(p)=\left(A_{g,n}-\frac{n}{Y_Q}\right)\mathcal{I}(p)+nP_{Y_Q}(p)+\mu_0(p)+\bigl(\mathcal{O}_X(p)-\omega_1(p)\bigr)+\frac{np(a)}{A_{g,n}Y_Q}. \tag*{(419)} τ X , Q ∘ ( p ) = ( A g , n − Y Q n ) I ( p ) + n P Y Q ( p ) + μ 0 ( p ) + ( O X ( p ) − ω 1 ( p ) ) + A g , n Y Q n p ( a ) . (419) The last term is the contribution of the mass of the constant eigenfunction removed by the truncation.
Lemma J.7 (Retained mass [101, Lemma 6.5]). If ψ \psi ψ is a normalized L 2 L^2 L 2 eigenfunction of Δ X \Delta_X Δ X with eigenvalue 0 < λ ≤ 1 / 4 − ε 0<\lambda\le1/4-\varepsilon 0 < λ ≤ 1/4 − ε , including residual eigenfunctions, then ∥ χ Y Q ψ ∥ 2 2 ≥ 1 − e − 2 ε L Q \|\chi_{Y_Q}\psi\|_2^2\ge1-e^{-2\sqrt{\varepsilon}LQ} ∥ χ Y Q ψ ∥ 2 2 ≥ 1 − e − 2 ε L Q , uniformly in the surface and in its number of cusps.
Lemma J.8 (Coefficient bounds [101, Lemmas 6.1 and 6.7]). (a) If p ( x ) = ∑ t = 1 Q a t x t p(x)=\sum_{t=1}^Q a_tx^t p ( x ) = ∑ t = 1 Q a t x t , then ∑ t = 1 Q ∣ a t ∣ ≤ 7 Q N ∗ ( p ) \sum_{t=1}^Q|a_t|\le7^QN_*(p) ∑ t = 1 Q ∣ a t ∣ ≤ 7 Q N ∗ ( p ) .
(b) Let c f ≥ 2 c_f\ge2 c f ≥ 2 be the integer of [101, Proposition 3.6] and D 0 = 2 c f + 4 D_0=2c_f+4 D 0 = 2 c f + 4 . There is C 0 ≥ 2 C_0\ge2 C 0 ≥ 2 , depending only on f f f , such that for all integers j ≥ 2 j\ge2 j ≥ 2 , Q , n ≥ 1 Q,n\ge1 Q , n ≥ 1 and all real p = x ~ p p=\widetilde{x}p p = x p of degree at most Q Q Q ,
∣ ω j , n ( p ) ∣ ≤ s j [ C 0 ( Q + j ) ] D 0 j N ∗ ( p ) , (420) |\omega_{j,n}(p)|\le s^j[C_0(Q+j)]^{D_0j}N_*(p), \tag*{(420)} ∣ ω j , n ( p ) ∣ ≤ s j [ C 0 ( Q + j ) ] D 0 j N ∗ ( p ) , (420) with no condition on the genus. With A Q = ( 3 C 0 Q ) D 0 A_Q=(3C_0Q)^{D_0} A Q = ( 3 C 0 Q ) D 0 this gives ∣ ω j , n ( p ) ∣ ≤ s j A Q j N ∗ ( p ) |\omega_{j,n}(p)|\le s^jA_Q^jN_*(p) ∣ ω j , n ( p ) ∣ ≤ s j A Q j N ∗ ( p ) for 2 ≤ j ≤ 2 Q 2\le j\le2Q 2 ≤ j ≤ 2 Q .
Lemma J.9 (First moment of the truncated trace [101, proof of Lemma 6.8]). Suppose 1 ≤ n ≤ g 1 \le n \le g 1 ≤ n ≤ g and ∣ E g , n O X ( x ) − ω 1 ( x ) ∣ ≤ g |\mathbb{E}_{g,n}\mathcal{O}_{X}(x)-\omega_{1}(x)| \le g ∣ E g , n O X ( x ) − ω 1 ( x ) ∣ ≤ g . Then, for every Q ≥ 1 Q \ge1 Q ≥ 1 ,
E g , n τ X , Q ( x ) ≤ D f ( g + n Q ) , D f = 1 + 6 a + 6 π I f + b f , (421) \mathbb{E}_{g,n}\tau_{X,Q}(x) \le D_f(g+nQ), \qquad D_f=1+6a+6\pi I_f+b_f, \tag*{(421)} E g , n τ X , Q ( x ) ≤ D f ( g + n Q ) , D f = 1 + 6 a + 6 π I f + b f , (421) and every polynomial trace and orbital statistic of degree at most Q Q Q with zero constant term is absolutely integrable, with E g , n ∣ O X ( p ) ∣ ≤ N ∗ ( p ) { D f ( g + n Q ) + ( 4 π g + 2 π n ) I f + n J f Q 2 } \mathbb{E}_{g,n}|\mathcal{O}_{X}(p)| \le N_*(p)\{D_f(g+nQ)+(4\pi g+2\pi n)I_f+nJ_fQ^2\} E g , n ∣ O X ( p ) ∣ ≤ N ∗ ( p ) { D f ( g + n Q ) + ( 4 π g + 2 π n ) I f + n J f Q 2 } .
Proof. Take expectations in (J.17) at p = x p=x p = x , where H x = h H_x=h H x = h and O X ( x ) ≥ 0 \mathcal{O}_X(x)\ge0 O X ( x ) ≥ 0 . By (J.18), Lemma J.6(iv), the hypothesis and A g , n − n / Y Q ≤ 4 π g + 2 π n ≤ 6 π g \mathcal{A}_{g,n}-n/Y_Q \le4\pi g+2\pi n \le6\pi g A g , n − n / Y Q ≤ 4 π g + 2 π n ≤ 6 π g ,
E τ X , Q ( x ) ≤ ( 4 π g + 2 π n ) I f + n b f Q + 6 a + g ≤ D f ( g + n Q ) . \mathbb{E}\tau_{X,Q}(x) \le(4\pi g+2\pi n)I_f+nb_fQ+6a+g \le D_f(g+nQ). E τ X , Q ( x ) ≤ ( 4 π g + 2 π n ) I f + n b f Q + 6 a + g ≤ D f ( g + n Q ) . Lemma J.6(ii) and the exact identity give the remaining statements [101, (6.5)–(6.6)]. □
Lemma J.10 (Polynomial transfer [101, proof of Proposition 6.9]). Let Q ≥ 1 Q \ge1 Q ≥ 1 , q = Q + 2 q=Q+2 q = Q + 2 , and let p = x p ~ p=x\tilde p p = x p ~ be real of degree at most Q Q Q . Suppose that for every 1 ≤ t ≤ Q 1 \le t \le Q 1 ≤ t ≤ Q
∣ g − 1 E g , n O X ( x t ) − ∑ j = 1 q − 1 g − j ω j , n ( x t ) ∣ ≤ E , (422) \left|g^{-1}\mathbb{E}_{g,n}\mathcal{O}_X(x^t)-\sum_{j=1}^{q-1}g^{-j}\omega_{j,n}(x^t)\right|\le E, \tag*{(422)} g − 1 E g , n O X ( x t ) − j = 1 ∑ q − 1 g − j ω j , n ( x t ) ≤ E , (422) and that s A Q / g ≤ 1 / 2 sA_Q/g \le1/2 s A Q / g ≤ 1/2 . Then
∣ E g , n O X ( p ) − ω 1 ( p ) − g − 1 ω 2 , n ( p ) ∣ ≤ { 2 s 3 A Q 3 q 2 + 7 q g E } N ∗ ( p ) . (423) \left|\mathbb{E}_{g,n}\mathcal{O}_X(p)-\omega_1(p)-g^{-1}\omega_{2,n}(p)\right|
\le\left\{\frac{2s^3A_Q^3}{q^2}+7^q gE\right\}N_*(p). \tag*{(423)} E g , n O X ( p ) − ω 1 ( p ) − g − 1 ω 2 , n ( p ) ≤ { q 2 2 s 3 A Q 3 + 7 q g E } N ∗ ( p ) . (423) Proof. Linearity, Lemma J.8(a) and (J.22) give E O X ( p ) = ω 1 ( p ) + ∑ j = 2 Q + 1 g 1 − j ω j , n ( p ) + R \mathbb{E}\mathcal{O}_X(p)=\omega_1(p)+\sum_{j=2}^{Q+1}g^{1-j}\omega_{j,n}(p)+R E O X ( p ) = ω 1 ( p ) + ∑ j = 2 Q + 1 g 1 − j ω j , n ( p ) + R with ∣ R ∣ ≤ 7 q g E N ∗ ( p ) |R|\le7^qgE N_*(p) ∣ R ∣ ≤ 7 q g E N ∗ ( p ) [101, (6.51)]. For 3 ≤ j ≤ Q + 1 3\le j\le Q+1 3 ≤ j ≤ Q + 1 , Lemma J.8(b) gives
∑ j = 3 Q + 1 g 1 − j ∣ ω j , n ( p ) ∣ ≤ s 3 A Q 3 g 2 N ∗ ( p ) ∑ j ≥ 3 ( s A Q / g ) j − 3 ≤ 2 s 3 A Q 3 g 2 N ∗ ( p ) . \sum_{j=3}^{Q+1}g^{1-j}|\omega_{j,n}(p)|\le\frac{s^3A_Q^3}{g^2}N_*(p)\sum_{j\ge3}(sA_Q/g)^{j-3}\le\frac{2s^3A_Q^3}{g^2}N_*(p). j = 3 ∑ Q + 1 g 1 − j ∣ ω j , n ( p ) ∣ ≤ g 2 s 3 A Q 3 N ∗ ( p ) j ≥ 3 ∑ ( s A Q / g ) j − 3 ≤ g 2 2 s 3 A Q 3 N ∗ ( p ) . In [101] the hypothesis is supplied by Lemma 6.8 with E = g − s q [ c 3 q ] E=g^{-sq}[c3q] E = g − s q [ c 3 q ] under the additive threshold (6.46), and 7 q g E ≤ ( s 2 / g ) 2 − Q 7^qgE\le(s^2/g)2^{-Q} 7 q g E ≤ ( s 2 / g ) 2 − Q there by (6.52); here it is supplied by (J.37). □
Lemma J.11 (Weighted distributions and the exceptional measure [101, Lemma 6.10, Theorem 6.11]). Put m = 2 D 0 + 2 m=2D_0+2 m = 2 D 0 + 2 (here m m m is an order of differentiation, not the carrier complexity of Subsection J.1). For each n ≥ 1 n\ge1 n ≥ 1 the functional K n ( q ) = ω 2 , n ( x q ) K_n(q)=\omega_{2,n}(x^q) K n ( q ) = ω 2 , n ( x q ) on polynomials q q q extends uniquely to a distribution on R \mathbb{R} R supported in [ 0 , a ] [0,a] [ 0 , a ] and of order at most m m m , with ∣ K n ( φ ) ∣ ≤ K f ( n + 1 ) 2 ∥ φ ∥ C m ( [ 0 , a ] ) |\mathcal{K}_n(\varphi)|\le K_f(n+1)^2\|\varphi\|_{C^m([0,a])} ∣ K n ( φ ) ∣ ≤ K f ( n + 1 ) 2 ∥ φ ∥ C m ([ 0 , a ]) for a constant K f K_f K f depending only on f f f . The unweighted extension V n ( ψ ) = K n ( ( ψ − ψ ( 0 ) ) / x ) V_n(\psi)=K_n((\psi-\psi(0))/x) V n ( ψ ) = K n (( ψ − ψ ( 0 )) / x ) restricts on ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) to a positive Radon measure μ n \mu_n μ n supported in ( 1 , a ] (1,a] ( 1 , a ] , with μ n ( [ b , a ] ) ≤ C f , b ( n + 1 ) 2 \mu_n([b,a])\le C_{f,b}(n+1)^2 μ n ([ b , a ]) ≤ C f , b ( n + 1 ) 2 for 1 < b ≤ a 1<b\le a 1 < b ≤ a [101, (6.57)–(6.58)]. The second difference in n n n of (J.5) gives the weighted extension
T 2 = K 3 − 2 K 2 + K 1 , T 2 ( q ) = W 2 ( x q ) , ∣ T 2 ( φ ) ∣ ≤ 38 K f ∥ φ ∥ C m ( [ 0 , a ] ) . (424) \mathcal{T}_2=\mathcal{K}_3-2\mathcal{K}_2+\mathcal{K}_1,\qquad\mathcal{T}_2(q)=W_2(xq),\qquad|\mathcal{T}_2(\varphi)|\le38K_f\|\varphi\|_{C^m([0,a])}. \tag*{(424)} T 2 = K 3 − 2 K 2 + K 1 , T 2 ( q ) = W 2 ( x q ) , ∣ T 2 ( φ ) ∣ ≤ 38 K f ∥ φ ∥ C m ([ 0 , a ]) . (424) There is a positive Radon measure η 2 \eta_2 η 2 on ( 1 , a ] (1,a] ( 1 , a ] such that
T 2 ( ψ / x ) = ∫ ψ d η 2 , η 2 ( [ b , a ] ) ≤ C f , b ( 1 < b ≤ a ) , (425) \mathcal{T}_2(\psi/x)=\int\psi\,d\eta_2,\qquad\eta_2([b,a])\le C_{f,b}\quad(1<b\le a), \tag*{(425)} T 2 ( ψ / x ) = ∫ ψ d η 2 , η 2 ([ b , a ]) ≤ C f , b ( 1 < b ≤ a ) , (425) for all ψ ∈ C ε ∞ ( ( 1 , ∞ ) ) \psi\in C^\infty_\varepsilon((1,\infty)) ψ ∈ C ε ∞ (( 1 , ∞ )) ; these tests may be nonzero at a a a , and the statement allows η 2 \eta_2 η 2 to have an atom at a a a .
Lemma J.12 (Detector and count bound [101, Lemma 6.4, (6.26), Proposition 6.12]). Fix 0 < ε < 1 / 4 0<\varepsilon<1/4 0 < ε < 1/4 and put b ε = f ( i ε ) ∈ ( 1 , a ) b_\varepsilon=f(i\sqrt{\varepsilon})\in(1,a) b ε = f ( i ε ) ∈ ( 1 , a ) . There are κ ε > 0 \kappa_\varepsilon>0 κ ε > 0 , B ε < ∞ B_\varepsilon<\infty B ε < ∞ , an integer Q ε Q_\varepsilon Q ε and, for every Q ≥ 1 Q\ge1 Q ≥ 1 , a real polynomial p Q p_Q p Q of degree Q Q Q with p Q ( 0 ) = 0 p_Q(0)=0 p Q ( 0 ) = 0 , p Q ( a ) = 1 p_Q(a)=1 p Q ( a ) = 1 and 0 ≤ p Q ≤ 1 0\le p_Q\le1 0 ≤ p Q ≤ 1 on [ 0 , a ] [0,a] [ 0 , a ] , such that for Q ≥ Q ε Q\ge Q_\varepsilon Q ≥ Q ε
N ∗ ( p Q ) ≤ B ε , N R ( p Q ) ≤ B ε e − κ ε Q , inf [ b ε , a ] p Q ≥ 1 − e − 8 κ ε Q . (426) N_*(p_Q)\le B_\varepsilon,\qquad N_{\mathbb{R}}(p_Q)\le B_\varepsilon e^{-\kappa_\varepsilon Q},\qquad\inf_{[b_\varepsilon,a]}p_Q\ge1-e^{-8\kappa_\varepsilon Q}. \tag*{(426)} N ∗ ( p Q ) ≤ B ε , N R ( p Q ) ≤ B ε e − κ ε Q , [ b ε , a ] inf p Q ≥ 1 − e − 8 κ ε Q . (426)
Let N ε ( X ) N_{\varepsilon}(X) N ε ( X ) be the number of eigenvalues of Δ X \Delta_X Δ X in ( 0 , 1 / 4 − ε ] (0,1/4-\varepsilon] ( 0 , 1/4 − ε ] , counted with multiplicity and including residual eigenvalues. Moreover e − 8 κ ε Q ≤ 1 2 e^{-8\kappa\varepsilon Q}\le\frac{1}{2} e − 8 κ εQ ≤ 2 1 and e − 2 ε L Q ≤ 1 2 e^{-2\sqrt{\varepsilon}LQ}\le\frac{1}{2} e − 2 ε L Q ≤ 2 1 for Q ≥ Q ε Q\ge Q_{\varepsilon} Q ≥ Q ε [101, (6.14)]. Then, by (J.26) and Lemma J.7,
N ε ( X ) ≤ 4 τ X , Q ∘ ( p Q ) ( Q ≥ Q ε ) . (427) N_{\varepsilon}(X)\le4\tau^\circ_{X,Q}(p_Q)\qquad(Q\ge Q_{\varepsilon}). \tag*{(427)} N ε ( X ) ≤ 4 τ X , Q ∘ ( p Q ) ( Q ≥ Q ε ) . (427)
Finally, there is C f , ε C_{f,\varepsilon} C f , ε such that, for all n ≥ 1 n\ge1 n ≥ 1 and Q ≥ Q ε Q\ge Q_{\varepsilon} Q ≥ Q ε , without any genus condition,
∣ ω 2 , n ( p Q ) ∣ ≤ C f , ε ( n + 1 ) 2 . (428) |\omega_{2,n}(p_Q)|\le C_{f,\varepsilon}(n+1)^2. \tag*{(428)} ∣ ω 2 , n ( p Q ) ∣ ≤ C f , ε ( n + 1 ) 2 . (428)
The estimate (J.28) is deduced in [101] from Lemma J.11 by splitting p Q p_Q p Q at a cutoff strictly below the detector transition; it uses only the local mass bound for the measures μ n \mu_n μ n associated with K n \mathcal{K}_n K n and the exponential smallness of p Q / x p_Q/x p Q / x below the transition, and hence holds for every n n n .
Verbatim inputs and replaced estimates. Lemmas J.3–J.12 are used below as stated. Lemmas J.9 and J.10 record the deterministic parts of the proofs of [101, Lemma 6.8 and Proposition 6.9] under the hypotheses displayed there, which replace his additive genus thresholds. Two estimates of [101] are not available on the envelope s ≤ M g s\le M\sqrt{g} s ≤ M g and are replaced:
(R1) the use of the volume lower bound in (J.9) with exponential factor at least e − 1 / 2 e^{-1/2} e − 1/2 , which Huang obtains from g ≥ κ v s 2 g\ge\kappa_v s^2 g ≥ κ v s 2 ; it is replaced by the uniform loss e κ v M 2 / 2 e^{\kappa_vM^2/2} e κ v M 2 /2 in (J.30);
(R2) the deep-sector absorption [101, (4.25)], which closes Lemma J.4(iv)–(v) under g ≥ K s ( s + t 2 ) g\ge Ks(s+t^2) g ≥ K s ( s + t 2 ) ; it is replaced by (J.33)–(J.34), which use t ≤ g θ t\le g^\theta t ≤ g θ and s ≤ M g s\le M\sqrt{g} s ≤ M g instead.
Besides these, the linear thresholds [ C t ] [C t] [ C t ] C s C s C s of Lemma J.5(a) and the carrier exclusion π ( g − 1 + n / 2 ) > T \pi(g-1+n/2)>T π ( g − 1 + n /2 ) > T are verified on the envelope in Subsection J.3, and Huang’s additive threshold for polynomial transfer is replaced by inserting (J.37) into Lemma J.10 in Subsection J.4. Every inequality that is not displayed in Subsections J.2–J.4 is taken verbatim from the lemmas above.
Near and deep complementary-volume tails
This subsection proves the graded tail bound (J.35) for carrier types and the separating tail bound (J.36) on the envelope s ≤ M g s\le M\sqrt{g} s ≤ M g , t ≤ g θ t\le g^\theta t ≤ g θ ; these replace the two estimates (R1)–(R2) listed at the end of Subsection J.1. The input is the tail skeleton of Lemma J.4. The tails are split into a near sector, summed as a geometric series, and a deep sector, bounded by the capped factorial inequality; the only new step is the comparison (J.34), which closes the deep sector using the envelope in place of a quadratic genus condition. Use the notation of Lemma J.4: put N = 2 g + n − 3 N=2g+n-3 N = 2 g + n − 3 , r = R + 1 r=R+1 r = R + 1 , t = R + m + 1 t=R+m+1 t = R + m + 1 , D = max ( m , r ) D=\max(m,r) D = max ( m , r ) , and J = ⌊ N / ( 64 s ) ⌋ J=\lfloor N/(64s)\rfloor J = ⌊ N / ( 64 s )⌋ . Choose initially
0 < θ < min { 1 / 4 , 1 / ( 512 M 2 ) } . 0<\theta<\min\{1/4,1/(512M^2)\}. 0 < θ < min { 1/4 , 1/ ( 512 M 2 )} .
For 1 ≤ t ≤ g θ 1\le t\le g^\theta 1 ≤ t ≤ g θ , uniformly at sufficiently large g g g ,
g ≥ 128 t s , g ≤ N ≤ 3 g , N − D + 1 ≥ g , J ≥ t , N / ( 128 s ) ≤ J ≤ N / 2. (429) g\ge128ts,\qquad g\le N\le3g,\qquad N-D+1\ge g,\qquad J\ge t,\qquad N/(128s)\le J\le N/2. \tag*{(429)} g ≥ 128 t s , g ≤ N ≤ 3 g , N − D + 1 ≥ g , J ≥ t , N / ( 128 s ) ≤ J ≤ N /2. (429)
Indeed g ≥ 128 t s g\ge128ts g ≥ 128 t s follows from g 1 / 2 − θ ≥ 128 M g^{1/2-\theta}\ge128M g 1/2 − θ ≥ 128 M ; g ≤ N ≤ 3 g g\le N\le3g g ≤ N ≤ 3 g from g ≥ 3 g\ge3 g ≥ 3 and n ≤ g + 3 n\le g+3 n ≤ g + 3 ; N − D + 1 ≥ g N-D+1\ge g N − D + 1 ≥ g from D ≤ t ≤ g θ ≤ g − 2 D\le t\le g^\theta\le g-2 D ≤ t ≤ g θ ≤ g − 2 ; and N ≥ g ≥ 128 t s N\ge g\ge128ts N ≥ g ≥ 128 t s gives N / ( 64 s ) ≥ 2 t N/(64s)\ge2t N / ( 64 s ) ≥ 2 t , hence J ≥ N / ( 128 s ) ≥ t J\ge N/(128s)\ge t J ≥ N / ( 128 s ) ≥ t . Thus (J.10) holds. The inverse volume lower bound in (J.9) loses at most a bounded factor on the envelope: since n ( n + 1 ) ≤ s 2 ≤ M 2 g n(n+1)\le s^2\le M^2g n ( n + 1 ) ≤ s 2 ≤ M 2 g ,
V g , n ≥ β v A 0 N N ! g exp ( − κ v n ( n + 1 ) 2 g ) ≥ e − κ v M 2 / 2 β v A 0 N N ! g , (430) V_{g,n}\ge\beta_v A_0^N\frac{N!}{\sqrt{g}}\exp\left(-\frac{\kappa_v n(n+1)}{2g}\right)\ge e^{-\kappa_vM^2/2}\beta_v A_0^N\frac{N!}{\sqrt{g}}, \tag*{(430)} V g , n ≥ β v A 0 N g N ! exp ( − 2 g κ v n ( n + 1 ) ) ≥ e − κ v M 2 /2 β v A 0 N g N ! , (430)
so Lemma J.4 applies with Θ = e κ v M 2 / 2 \Theta=e^{\kappa_vM^2/2} Θ = e κ v M 2 /2 . Its factorial envelope therefore remains valid with prefactor
P b , M = e κ v M 2 / 2 B b + 2 b ( m − 1 ) ! . P_{b,M}=e^{\kappa_vM^2/2}B^{b+2}\sqrt{b}(m-1)!. P b , M = e κ v M 2 /2 B b + 2 b ( m − 1 )! .
Here b b b is the number of complementary components. The largest complementary complexity is at least g / b g/b g / b , by the same Euler identity; the terminal types ( 0 , 3 ) (0,3) ( 0 , 3 ) and ( 1 , 1 ) (1,1) ( 1 , 1 ) are included in the envelope.
Near and deep sectors. The exact cusp budget is A o u t ≤ δ + 1 A_{\mathrm{out}} \le\delta+ 1 A out ≤ δ + 1 , and the number of fixed-deficit allocations is at most b ( δ + 3 ) 2 ( b − 1 ) b(\delta+ 3)2^{(b-1)} b ( δ + 3 ) 2 ( b − 1 ) (Lemma J.4 (ii)). Neither statement changes with the genus threshold. In the near sector, the successive ratio up to J J J is at most e 2 / 63 < 1 / 2 e^2/63 < 1/2 e 2 /63 < 1/2 by (J.29) (Lemma J.4 (iii)). Summing gives
2 P b , M b ( D + 3 ) 2 b D ! g − D s D + 1 ≤ 2 P b , M b ( D + 3 ) 2 b D ! g − r s r + 1 (431) 2P_{b,M}b(D+3)^{2b}D!g^{-D}s^{D+1}\le2P_{b,M}b(D+3)^{2b}D!g^{-r}s^{r+1} \tag*{(431)} 2 P b , M b ( D + 3 ) 2 b D ! g − D s D + 1 ≤ 2 P b , M b ( D + 3 ) 2 b D ! g − r s r + 1 (431)
where the inequality uses D ≥ r D\ge r D ≥ r and s / g ≤ 1 s/g\le1 s / g ≤ 1 . For b = 1 b=1 b = 1 this includes every allocation. For b ≥ 2 b\ge2 b ≥ 2 , the capped factorial inequality and multinomial sum of Lemma J.4 (iv), before Huang’s final absorption step, give the deep-sector bound
P b , M ( b + 1 ) n ( N + 1 ) 2 b exp { − N log ( 64 s ) / ( 128 s ) } . (432) P_{b,M}(b+1)^n(N+1)^{2b}\exp\{-N\log(64s)/(128s)\}. \tag*{(432)} P b , M ( b + 1 ) n ( N + 1 ) 2 b exp { − N log ( 64 s ) / ( 128 s )} . (432)
Replacing the absorption step. Write L s = log ( 64 s ) L_s=\log(64s) L s = log ( 64 s ) and F t ( g ) = 4 t log g + s log ( 2 t ) + 3 t log 4 F_t(g)=4t\log g+s\log(2t)+3t\log4 F t ( g ) = 4 t log g + s log ( 2 t ) + 3 t log 4 . By Lemma J.4 (iv)–(v) and N ≥ g N\ge g N ≥ g ,
g r P b , M ( b + 1 ) n ( N + 1 ) 2 b ( N J ) − 1 ≤ P b , M exp { F t ( g ) − N L s 128 s } ≤ P b , M exp { F t ( g ) − g L s 128 s } . (433) g^rP_{b,M}(b+1)^n(N+1)^{2b}\binom{N}{J}^{-1}
\le P_{b,M}\exp\left\{F_t(g)-\frac{NL_s}{128s}\right\}
\le P_{b,M}\exp\left\{F_t(g)-\frac{gL_s}{128s}\right\}. \tag*{(433)} g r P b , M ( b + 1 ) n ( N + 1 ) 2 b ( J N ) − 1 ≤ P b , M exp { F t ( g ) − 128 s N L s } ≤ P b , M exp { F t ( g ) − 128 s g L s } . (433)
so the deep sector is at most P b , M g − r P_{b,M}g^{-r} P b , M g − r as soon as F t ( g ) < g L s / ( 128 s ) F_t(g)<gL_s/(128s) F t ( g ) < g L s / ( 128 s ) . Uniformly on the envelope 1 ≤ t ≤ g θ 1\le t\le g^\theta 1 ≤ t ≤ g θ , 1 ≤ s ≤ M g 1\le s\le M\sqrt{g} 1 ≤ s ≤ M g ,
s F t ( g ) g L s = s ( 4 t log g + 3 t log 4 ) g L s + s 2 log ( 2 t ) g L s , \frac{sF_t(g)}{gL_s}
=\frac{s(4t\log g+3t\log4)}{gL_s}+\frac{s^2\log(2t)}{gL_s}, g L s s F t ( g ) = g L s s ( 4 t log g + 3 t log 4 ) + g L s s 2 log ( 2 t ) ,
s ( 4 t log g + 3 t log 4 ) g L s ≤ M g g θ ( 4 log g + 3 log 4 ) g log 64 = O ( g θ − 1 / 2 log g ) = o ( 1 ) , \frac{s(4t\log g+3t\log4)}{gL_s}
\le\frac{M\sqrt{g}\,g^\theta(4\log g+3\log4)}{g\log64}
=O\left(g^{\theta-1/2}\log g\right)=o(1), g L s s ( 4 t log g + 3 t log 4 ) ≤ g log 64 M g g θ ( 4 log g + 3 log 4 ) = O ( g θ − 1/2 log g ) = o ( 1 ) ,
s 2 log ( 2 t ) g L s ≤ M 2 ( θ log g + log 2 ) log ( 64 M g ) = 2 M 2 θ + o ( 1 ) < 1 256 + o ( 1 ) . (434) \frac{s^2\log(2t)}{gL_s}
\le\frac{M^2(\theta\log g+\log2)}{\log(64M\sqrt{g})}
=2M^2\theta+o(1)<\frac{1}{256}+o(1). \tag*{(434)} g L s s 2 log ( 2 t ) ≤ log ( 64 M g ) M 2 ( θ log g + log 2 ) = 2 M 2 θ + o ( 1 ) < 256 1 + o ( 1 ) . (434)
For the last inequality, use the monotonicity of x 2 / log ( 64 x ) x^2/\log(64x) x 2 / log ( 64 x ) for x ≥ 1 x\ge1 x ≥ 1 to replace s s s by M g M\sqrt{g} M g , and log ( 2 t ) ≤ θ log g + log 2 \log(2t)\le\theta\log g+\log2 log ( 2 t ) ≤ θ log g + log 2 . Thus eventually s F t ( g ) / ( g L s ) < 1 / 128 sF_t(g)/(gL_s)<1/128 s F t ( g ) / ( g L s ) < 1/128 , that is, F t ( g ) < g L s / ( 128 s ) F_t(g)<gL_s/(128s) F t ( g ) < g L s / ( 128 s ) , and (J.32) is at most P b , M g − r P_{b,M}g^{-r} P b , M g − r . Summing the at most ( 4 t ) 4 t (4t)^{4t} ( 4 t ) 4 t patterns with (J.31), and using Lemma J.4 (vi), gives the graded tail
T g , n h 0 , a 0 , k ( R ) ≤ g − R − 1 s R + 2 ( C M t ) C M t . (435) \mathcal{T}^{h_0,a_0,k}_{g,n}(R)\le g^{-R-1}s^{R+2}(C_Mt)^{C_Mt}. \tag*{(435)} T g , n h 0 , a 0 , k ( R ) ≤ g − R − 1 s R + 2 ( C M t ) C M t . (435)
The boundary version has the same extra factor ( 1 + Λ ) k e Λ (1+\Lambda)^{k}e^\Lambda ( 1 + Λ ) k e Λ as in Huang’s lemma, since that factor comes from the numerator boundary estimate (Lemma J.4 (vii)).
The separating kernel. For the ordered separating kernel, repeat the split with t = r ≥ 1 t=r\ge1 t = r ≥ 1 , using the separating part of Lemma J.4 with Θ = e κ v M 2 / 2 \Theta=e^{\kappa_vM^2/2} Θ = e κ v M 2 /2 and B M = B e κ v M 2 / 2 B_M=Be^{\kappa_vM^2/2} B M = B e κ v M 2 /2 . The near sector is at most 12 B M r ! g − r s r + 1 12B_Mr!g^{-r}s^{r+1} 12 B M r ! g − r s r + 1 , and the deep sector before absorption is B M ( g + 1 ) 2 n exp { − N log ( 64 s ) / ( 128 s ) } B_M(g+1)^{2n}\exp\{-N\log(64s)/(128s)\} B M ( g + 1 ) 2 n exp { − N log ( 64 s ) / ( 128 s )} . The logarithm of g r ( g + 1 ) 2 n g^r(g+1)^{2n} g r ( g + 1 ) 2 n is bounded by F r ( g ) F_r(g) F r ( g ) . The same comparison (J.34), including r = 1 r=1 r = 1 , gives
S g , n ≥ r ( x ) ≤ g − r s r + 1 ( C M r ) C M r e x . (436) \mathcal{S}^{\ge r}_{g,n}(x)\le g^{-r}s^{r+1}(C_Mr)^{C_Mr}e^x. \tag*{(436)} S g , n ≥ r ( x ) ≤ g − r s r + 1 ( C M r ) C M r e x . (436)
The original allocation sums include the central tie and empty sums, so no endpoint configuration is discarded.
Retained records and orbital integration
This subsection completes the proof of the orbital estimate (J.1). We verify that the remaining inputs of Huang’s orbital expansion, namely the linear genus thresholds of Lemma J.5 (a) and the whole-carrier exclusion of Definition J.2 (i), hold on the new envelope; we then combine the record prefactors of Lemma J.5 (b)–(c) with the tail bound (J.35) into a graded carrier error, and integrate it over orbital types using Lemma J.5 (d).
Linear thresholds. The graded boundary and shifted-volume estimates of Lemma J.5(a) require only g ≥ [ C t ] C s g \ge[Ct]^{C}s g ≥ [ C t ] C s . Decrease θ \theta θ so that C θ < 1 / 2 C\theta< 1/2 C θ < 1/2 for each of their finitely many fixed constants; then
[ C t ] C s ≤ C C g C θ M g = o ( g ) ( t ≤ g θ , s ≤ M g ) , [Ct]^{C}s \le C^{C}g^{C\theta}M\sqrt{g}=o(g)\qquad(t\le g^{\theta},\ s\le M\sqrt{g}), [ C t ] C s ≤ C C g C θ M g = o ( g ) ( t ≤ g θ , s ≤ M g ) ,
so these requirements, including fixed multiples of t t t , hold uniformly on t ≤ g θ t\le g^{\theta} t ≤ g θ , s ≤ M g s\le M\sqrt{g} s ≤ M g . Retained records. For a retained carrier record of deficit δ ≤ R \delta\le R δ ≤ R , its exact identities (J.8) and Lemma J.5(c) give
H 0 ≤ t , p = n − A o u t + k ∗ ≥ k ∗ ≥ 1 , J l o c = R − δ , A o u t + J l o c + 1 ≤ R + 2. H_0\le t,\qquad p=n-A_{\mathrm{out}}+k_* \ge k_* \ge1,\qquad J_{\mathrm{loc}}=R-\delta,\qquad A_{\mathrm{out}}+J_{\mathrm{loc}}+1\le R+2. H 0 ≤ t , p = n − A out + k ∗ ≥ k ∗ ≥ 1 , J loc = R − δ , A out + J loc + 1 ≤ R + 2.
The small-volume prefactor is bounded by s A o u t [ C ( δ + 1 ) ] C ( δ + 1 ) ( 1 + Λ ) k e Λ − X ∗ / 2 s^{A_{\mathrm{out}}}[C(\delta+1)]^{C(\delta+1)}(1+\Lambda)^{k}e^{\Lambda-X_*/2} s A out [ C ( δ + 1 ) ] C ( δ + 1 ) ( 1 + Λ ) k e Λ − X ∗ /2 by (J.12), without a genus condition. Multiplication by the local shifted-volume error (J.11) therefore has cusp power at most s R + 2 s^{R+2} s R + 2 , length power at most 2 t 2t 2 t and exponential e Λ e^{\Lambda} e Λ . There are at most b ( δ + 3 ) 2 ( b − 1 ) b(\delta+3)^{2(b-1)} b ( δ + 3 ) 2 ( b − 1 ) records. Together with (J.35), this proves the same graded carrier error as [101, Proposition 4.7], on the new envelope:
∣ F g , n h 0 , a 0 , k ; P ( y ) − ∑ j = 1 R g − j F j , n h 0 , a 0 , k ; P ( y ) ∣ ≤ g − R − 1 s R + 2 ( C M t ) C M t ( 1 + Λ ) 2 t e Λ . \left|F_{g,n}^{h_0,a_0,k;P}(y)-\sum_{j=1}^{R}g^{-j}F_{j,n}^{h_0,a_0,k;P}(y)\right|
\le g^{-R-1}s^{R+2}(C_Mt)^{C_Mt}(1+\Lambda)^{2t}e^\Lambda. F g , n h 0 , a 0 , k ; P ( y ) − j = 1 ∑ R g − j F j , n h 0 , a 0 , k ; P ( y ) ≤ g − R − 1 s R + 2 ( C M t ) C M t ( 1 + Λ ) 2 t e Λ .
For a nonseparating kernel, the scalar handle-ratio expansion times the two-boundary expansion has the smaller cusp power s R + 1 s^{R+1} s R + 1 . For a separating kernel, the retained endpoints use the same shifted estimate and the omitted ones use (J.36).
The inequalities N > 2 R N>2R N > 2 R and g > H 0 g>H_0 g > H 0 , which hold since R , H 0 ≤ t ≤ g θ R,H_0\le t\le g^{\theta} R , H 0 ≤ t ≤ g θ , justify reconstruction of retained records. Unavailable cusp allocations are omitted before evaluating a volume. When R = 0 R=0 R = 0 or m > R m>R m > R , apply the tail estimate directly rather than a negative-order shifted expansion. The canonical finite sums are consequently unchanged; this argument does not interpolate in n n n .
Integration over orbital types. For a length support [ − T , T ] [-T,T] [ − T , T ] , the whole-carrier bound π ( g − 1 + n / 2 ) > T \pi(g-1+n/2)>T π ( g − 1 + n /2 ) > T of Definition J.2(i), which holds since T ≤ ρ ≤ g θ T\le\rho\le g^{\theta} T ≤ ρ ≤ g θ , excludes filled surfaces and entirely self-paired carriers. The remaining types have 1 ≤ m ≤ ⌊ 2 T / π ⌋ 1\le m\le\lfloor2T/\pi\rfloor 1 ≤ m ≤ ⌊ 2 T / π ⌋ , k ≤ m + 2 k\le m+2 k ≤ m + 2 and t = R + m + 1 ≤ R + ⌈ T ⌉ + 1 = ρ t=R+m+1\le R+\lceil T\rceil+1=\rho t = R + m + 1 ≤ R + ⌈ T ⌉ + 1 = ρ . The exact rooted integration formula (J.15) and its fiber and boundary-simplex bounds (Lemma J.5(d)) give, per pattern, an error at most
g − R − 1 s R + 2 ( C M t ) C M t C m ( 1 + T ) ( 1 + 2 T ) 2 t + k e 3 T ∥ H ∥ ∞ . g^{-R-1}s^{R+2}(C_Mt)^{C_Mt}C_m(1+T)(1+2T)^{2t+k}e^{3T}\lVert H\rVert_\infty. g − R − 1 s R + 2 ( C M t ) C M t C m ( 1 + T ) ( 1 + 2 T ) 2 t + k e 3 T ∥ H ∥ ∞ .
There are at most ( k + 1 ) 2 k (k+1)^{2k} ( k + 1 ) 2 k patterns and ( m + 2 ) 2 (m+2)^2 ( m + 2 ) 2 types at each m m m . Absorbing these factors, and adding the simple part obtained from (J.14) with the kernel errors of Lemma J.5(c), proves (J.1). Simple repetitions have the same absolutely integrable bound. The zero-support case is zero and an unavailable nonsimple type contributes an empty sum. Absolute bounds also cover signed tests. The coefficient vanishing rules [101, (3.32)], recalled in Lemma J.5(d), identify the retained finite sum with the canonical coefficients at every R , T R,T R , T .
Growing-degree polynomial tests and eigenvalue counts
Finally we apply the orbital estimate (J.1) to polynomial tests whose degree grows with the genus, and deduce the local expected-count estimate for eigenvalues. There are three stages: the monomial estimate (J.37), which also supplies the first-moment hypothesis of Lemma J.9; the transfer to polynomials of degree Q = ⌊ g σ ⌋ Q=\lfloor g^\sigma\rfloor Q = ⌊ g σ ⌋ through Lemma J.10, which proves (216) ; and the count bound, obtained from the detector polynomial and the positive trace of Lemmas J.6–J.7 and J.12.
Monomial tests. Apply (J.1) to H = h ∗ j H=h^{*j} H = h ∗ j with 1 ≤ j < q 1\le j<q 1 ≤ j < q , R = q − 2 R=q-2 R = q − 2 , and T = j L T=jL T = j L , where supp h ⊂ [ − L , L ] \operatorname{supp}h\subset[-L,L] supp h ⊂ [ − L , L ] . Its supremum is at most ∥ h ∥ ∞ \lVert h\rVert_\infty ∥ h ∥ ∞ , and ρ = q − 1 + ∣ j L ∣ ≤ ( 1 + L ) q \rho=q-1+\lvert jL\rvert\le(1+L)q ρ = q − 1 + ∣ j L ∣ ≤ ( 1 + L ) q . For q ≤ g θ / ( 2 + L ) q\le g^\theta/(2+L) q ≤ g θ / ( 2 + L ) , division by g g g gives
∣ g − 1 E O ( x j ) − ∑ i = 1 q − 1 g − i ω i , n ( x j ) ∣ ≤ g − q s q ( C M , f q ) C M , f q . (437) \left|g^{-1}\mathbb{E}O(x^j)-\sum_{i=1}^{q-1}g^{-i}\omega_{i,n}(x^j)\right|
\le g^{-q}s^q(C_{M,f}q)^{C_{M,f}q}. \tag*{(437)} g − 1 E O ( x j ) − i = 1 ∑ q − 1 g − i ω i , n ( x j ) ≤ g − q s q ( C M , f q ) C M , f q . (437)
since ( C M ρ ) C M , ρ e 3 j L ∥ h ∥ ∞ ≤ ( C M , f q ) C M , f q (C_{M}\rho)^{C_{M,\rho}}e^{3jL}\|h\|_{\infty}\leq(C_{M,f}q)^{C_{M,f}q} ( C M ρ ) C M , ρ e 3 j L ∥ h ∥ ∞ ≤ ( C M , f q ) C M , f q . The base case q = 2 q=2 q = 2 , j = 1 j=1 j = 1 gives ∣ E C X ( x ) − ω 1 ( x ) ∣ ≤ s 2 ( 2 C M , f ) 2 C M , f / g ≤ M 2 ( 2 C M , f ) 2 C M , f ≤ g |\mathbb{E}C_X(x)-\omega_1(x)|\leq s^2(2C_{M,f})^{2C_{M,f}}/g\leq M^2(2C_{M,f})^{2C_{M,f}}\leq g ∣ E C X ( x ) − ω 1 ( x ) ∣ ≤ s 2 ( 2 C M , f ) 2 C M , f / g ≤ M 2 ( 2 C M , f ) 2 C M , f ≤ g for large g g g , so Lemma J.9 bounds the first moment of the positive truncated trace by D f ( g + n Q ) D_f(g+nQ) D f ( g + n Q ) for every cutoff Q Q Q ; a large-degree genus threshold is not required for this base-moment bound.
Polynomial transfer. Take Q = ⌊ g σ ⌋ Q=\lfloor g^\sigma\rfloor Q = ⌊ g σ ⌋ with
0 < σ < θ , σ C M , f < 1 / 4 , 3 σ D 0 < 1 / 2. 0<\sigma<\theta,\qquad\sigma C_{M,f}<1/4,\qquad3\sigma D_0<1/2. 0 < σ < θ , σ C M , f < 1/4 , 3 σ D 0 < 1/2.
For p = x ~ p p=\widetilde{x}p p = x p of degree at most Q Q Q , put q = Q + 2 q=Q+2 q = Q + 2 in (J.37), so that (J.22) holds with E = q − q s q ( C M , f q ) C M , f q E=q^{-q}s^q(C_{M,f}q)^{C_{M,f}q} E = q − q s q ( C M , f q ) C M , f q . The monomial coefficient bound of Lemma J.8(a) contributes 7 Q 7^Q 7 Q . Relative to ( s 2 / g ) ∥ p ~ ∥ [ 0 , a ] (s^2/g)\|\widetilde{p}\|_{[0,a]} ( s 2 / g ) ∥ p ∥ [ 0 , a ] , the terminal error 7 Q g E N ∗ ( p ) 7^QgE N_*(p) 7 Q g E N ∗ ( p ) is bounded by
7 Q [ C M , f ( Q + 2 ) ] C M , f ( Q + 2 ) ( s / g ) Q ≤ 2 − Q 7^Q[C_{M,f}(Q+2)]^{C_{M,f}(Q+2)}(s/g)^Q\leq2^{-Q} 7 Q [ C M , f ( Q + 2 ) ] C M , f ( Q + 2 ) ( s / g ) Q ≤ 2 − Q
for sufficiently large g g g : its logarithm divided by Q log g Q\log g Q log g has limit superior at most σ C M , f − 1 / 2 < 0 \sigma C_{M,f}-1/2<0 σ C M , f − 1/2 < 0 , because log ( s / g ) ≤ − 1 2 log g + log M \log(s/g)\leq-\frac{1}{2}\log g+\log M log ( s / g ) ≤ − 2 1 log g + log M . The genus-independent canonical coefficient estimate of Lemma J.8(b), with A Q = ( 3 C 0 Q ) D 0 A_Q=(3C_0Q)^{D_0} A Q = ( 3 C 0 Q ) D 0 , gives
∑ i = 3 Q + 1 g 1 − i ∣ ω i , n ( p ) ∣ ≤ 2 s 3 A Q 3 g 2 ∥ p ~ ∥ [ 0 , a ] , s A Q / g = O M ( g σ D 0 − 1 / 2 ) = o ( 1 ) . \sum_{i=3}^{Q+1}g^{1-i}|\omega_{i,n}(p)|\leq\frac{2s^3A_Q^3}{g^2}\|\widetilde{p}\|_{[0,a]},\qquad sA_Q/g=O_M(g^{\sigma D_0-1/2})=o(1). i = 3 ∑ Q + 1 g 1 − i ∣ ω i , n ( p ) ∣ ≤ g 2 2 s 3 A Q 3 ∥ p ∥ [ 0 , a ] , s A Q / g = O M ( g σ D 0 − 1/2 ) = o ( 1 ) .
Lemma J.10 therefore proves (216) , with an o ( 1 ) o(1) o ( 1 ) right side since s 3 A Q 3 / g 2 = O M ( g 3 σ D 0 − 1 / 2 ) s^3A_Q^3/g^2=O_M(g^{3\sigma D_0-1/2}) s 3 A Q 3 / g 2 = O M ( g 3 σ D 0 − 1/2 ) . The same reasoning holds for Q = ⌈ K log g ⌉ Q=\lceil K\log g\rceil Q = ⌈ K log g ⌉ for any fixed K K K .
Eigenvalue counts. Finally apply the binomial detector, exact constant cancellation, parabolic bound and retained-mass estimate of Lemmas J.6, J.7 and J.12 at Q = ⌈ K ε log g ⌉ Q=\lceil K_\varepsilon\log g\rceil Q = ⌈ K ε log g ⌉ with K ε ≥ max { 3 / κ ε , 2 / L } K_\varepsilon\geq\max\{3/\kappa_\varepsilon,2/L\} K ε ≥ max { 3/ κ ε , 2/ L } . Taking expectations in (J.19) at p = p Q p=p_Q p = p Q , using (J.27), the bounds (J.18) and ∣ μ 0 ( p Q ) ∣ ≤ 5 N R ( p Q ) |\mu_0(p_Q)|\leq5N_R(p_Q) ∣ μ 0 ( p Q ) ∣ ≤ 5 N R ( p Q ) , the detector norms (J.26), the degree-independent bound (J.28), and the polynomial error (216) just established at this Q Q Q , we obtain, exactly as in [101, proof of Theorem 6.13, (6.74)],
E N X ( 1 / 4 − ε ) ≤ 4 E τ X , Q ( p Q ) ≤ C M , f , ε { ( g + n Q 2 ) e − κ ε Q + s 2 g + s 3 Q 3 D 0 g 2 + s 2 g 2 − Q + n area ( X ) Y Q } . \mathbb{E}N_X(1/4-\varepsilon)\leq4\mathbb{E}\tau_{X,Q}(p_Q)\leq C_{M,f,\varepsilon}\left\{(g+nQ^2)e^{-\kappa_\varepsilon Q}+\frac{s^2}{g}+\frac{s^3Q^{3D_0}}{g^2}+\frac{s^2}{g}2^{-Q}+\frac{n}{\operatorname{area}(X)Y_Q}\right\}. E N X ( 1/4 − ε ) ≤ 4 E τ X , Q ( p Q ) ≤ C M , f , ε { ( g + n Q 2 ) e − κ ε Q + g s 2 + g 2 s 3 Q 3 D 0 + g s 2 2 − Q + area ( X ) Y Q n } .
Here N X ( 1 / 4 − ε ) = N ε ( X ) N_X(1/4-\varepsilon)=N_\varepsilon(X) N X ( 1/4 − ε ) = N ε ( X ) is the count of Lemma J.12. The exponential term is O ( g − 2 log 2 g ) O(g^{-2}\log^2g) O ( g − 2 log 2 g ) by the choice of K ε K_\varepsilon K ε , since n ≤ g n\leq g n ≤ g and e − κ ε Q ≤ g − 3 e^{-\kappa_\varepsilon Q}\leq g^{-3} e − κ ε Q ≤ g − 3 ; the third term divided by s 2 / g s^2/g s 2 / g is O M ( Q 3 D 0 / g ) = o ( 1 ) O_M(Q^{3D_0}/\sqrt{g})=o(1) O M ( Q 3 D 0 / g ) = o ( 1 ) ; and the truncation height Y Q = 3 e L Q ≥ 3 g 2 Y_Q=3e^{LQ}\geq3g^2 Y Q = 3 e L Q ≥ 3 g 2 makes the last term O ( g − 2 ) O(g^{-2}) O ( g − 2 ) , because area ( X ) ≥ 2 π n \operatorname{area}(X)\geq2\pi n area ( X ) ≥ 2 π n . Thus all terms are O M , f , ε ( s 2 / g ) O_{M,f,\varepsilon}(s^2/g) O M , f , ε ( s 2 / g ) . This is the local expected-count estimate used to remove the eigenfunction-mass weights in Proposition 15.3.
Appendix K. Proof of the moving spectral law
This appendix proves the moving spectral law, Theorem 16.1; Subsections K.4 and K.5 then prove Corollaries 16.2 and 16.3. Throughout, c = n / g c=n/\sqrt{g} c = n / g , κ 0 = ( 8 π 2 ) − 1 \kappa_0=(8\pi^2)^{-1} κ 0 = ( 8 π 2 ) − 1 , and the model functions are those of (227) . For a > 0 a>0 a > 0 , let N g ( a ) = D X ( [ a , ∞ ) ) N_g(a)=D_X([a,\infty)) N g ( a ) = D X ([ a , ∞ )) and let Z g ( a ) Z_g(a) Z g ( a ) count two-cusp pairs of pants whose model deficits are at least a c 2 ac^2 a c 2 . The inputs are Huang’s trace expansion [101], used here under his original genus thresholds, the model identification of Proposition 15.5, the cut-graph bound (2) of Appendix I, and the funnel model F s F_s F s of Section 15. The proof has three parts, kept separate so that their ranges are explicit: a positive trace estimate for the mean of the eigenvalue counts (first subsection), Poisson convergence and conditional isolation of the pants counts (second subsection), and a spectral comparison that matches the two counts and identifies the eigenfunctions (third subsection).
The moving expected count
This subsection proves the limit (K.5) of the mean of the eigenvalue counts, using Huang’s trace expansion with test functions adapted to the moving scale. The main step is Lemma K.1, a trace approximation whose error terms are explicit in c c c ; the inputs (I1)–(I6) from [101] are listed before it, and the identification of the limit with the model intensity uses Proposition 15.5 . In this subsection we assume only c → 0 c \to0 c → 0 and log ( 1 / c ) = o ( log g ) \log(1/c) = o(\log g) log ( 1/ c ) = o ( log g ) . This is implied by (229) : if c log g → ∞ c \log g \to\infty c log g → ∞ then c ≥ 1 / log g c \ge1/\log g c ≥ 1/ log g eventually, so log ( 1 / c ) ≤ log log g \log(1/c) \le\log\log g log ( 1/ c ) ≤ log log g ; it also forces n = g 1 / 2 − o ( 1 ) → ∞ n = g^{1/2-o(1)} \to\infty n = g 1/2 − o ( 1 ) → ∞ . Keep the fixed multiplier f f f , its endpoint a 0 = f ( i / 2 ) a_0 = f(i/2) a 0 = f ( i /2 ) and its support constant L L L from § 4.7 . The function F ( t ) = f ( i t ) F(t) = f(i\sqrt{t}) F ( t ) = f ( i t ) is analytic on [ 0 , 1 / 4 ] [0,1/4] [ 0 , 1/4 ] and F ′ ( t ) > 0 F'(t) > 0 F ′ ( t ) > 0 there. At zero this follows from the positive second moment of h = b 0 ∗ b 0 h = b_0 * b_0 h = b 0 ∗ b 0 ; at positive t t t it follows by differentiating the even integral against cosh ( t u ) \cosh(\sqrt{t}\,u) cosh ( t u ) . Its inverse J J J is smooth on [ 1 , a 0 ] [1,a_0] [ 1 , a 0 ] .
Let ϕ \phi ϕ be smooth on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , zero below some a > 0 a > 0 a > 0 , and either compactly supported or constant beyond a finite point. Put ψ c ( x ) = ϕ ( J ( x ) / c 2 ) \psi_c(x) = \phi(J(x)/c^2) ψ c ( x ) = ϕ ( J ( x ) / c 2 ) for x ≥ 1 x \ge1 x ≥ 1 and zero for x ≤ 1 x \le1 x ≤ 1 . Extend smoothly beyond a 0 a_0 a 0 when necessary. For each fixed r r r ,
∥ ψ c / x ∥ C r ( [ 0 , a 0 ] ) ≤ C ϕ , r c − 2 r . (438) \left\|\psi_c/x\right\|_{C^r([0,a_0])} \le C_{\phi,r}c^{-2r}. \tag*{(438)} ∥ ψ c / x ∥ C r ([ 0 , a 0 ]) ≤ C ϕ , r c − 2 r . (438)
Choose Q = ⌊ g σ ⌋ Q = \lfloor g^\sigma\rfloor Q = ⌊ g σ ⌋ with 0 < σ < min { ( 2 D 3 ) − 1 , ( 6 D 0 ) − 1 , 1 / 4 } 0 < \sigma< \min\{(2D_3)^{-1},(6D_0)^{-1},1/4\} 0 < σ < min {( 2 D 3 ) − 1 , ( 6 D 0 ) − 1 , 1/4 } , using Huang’s constants [101, Proposition 6.9]. His original threshold applies: ( n + 1 ) 2 / g → 0 (n+1)^2/g \to0 ( n + 1 ) 2 / g → 0 and Q D 3 ( n + 1 ) / g → 0 Q^{D_3}(n+1)/g \to0 Q D 3 ( n + 1 ) / g → 0 . No critical-envelope extension is needed in this range.
Let r Q r_Q r Q be the shifted Chebyshev truncation of ψ c / x \psi_c/x ψ c / x of degree Q − 1 Q-1 Q − 1 and v Q = x r Q v_Q = xr_Q v Q = x r Q . For fixed K , m K,m K , m , integration by parts after x = a 0 ( 1 + cos θ ) / 2 x = a_0(1+\cos\theta)/2 x = a 0 ( 1 + cos θ ) /2 and (K.1) with r = K + 2 m + 2 r = K+2m+2 r = K + 2 m + 2 give
e m : = ∥ r Q − ψ c / x ∥ C m ( [ 0 , a 0 ] ) ≤ C K , m Q − K c − 2 ( K + 2 m + 2 ) = Q − K g o ( 1 ) . (439) e_m := \left\|r_Q-\psi_c/x\right\|_{C^m([0,a_0])} \le C_{K,m}Q^{-K}c^{-2(K+2m+2)} = Q^{-K}g^{o(1)}. \tag*{(439)} e m := ∥ r Q − ψ c / x ∥ C m ([ 0 , a 0 ]) ≤ C K , m Q − K c − 2 ( K + 2 m + 2 ) = Q − K g o ( 1 ) . (439)
Indeed the j j j th Chebyshev coefficient costs j − K − 2 m − 2 j^{-K-2m-2} j − K − 2 m − 2 and its m m m th derivative costs at most C m j 2 m C_mj^{2m} C m j 2 m .
Inputs from Huang’s trace expansion. We now record, in the notation of [101] and of Proposition 15.3, the statements of that paper that enter the trace estimate below; each is quoted with its own hypotheses, and none of them is modified here. Let A X A_X A X be the bounded self-adjoint operator with multiplier f f f : a discrete eigenvalue 0 < λ j < 1 / 4 0 < \lambda_j < 1/4 0 < λ j < 1/4 with eigenfunction u j u_j u j has A X A_X A X -eigenvalue x j = F ( 1 / 4 − λ j ) = f ( i 1 / 4 − λ j ) ∈ ( 1 , a 0 ] x_j = F(1/4-\lambda_j) = f(i\sqrt{1/4-\lambda_j}) \in(1,a_0] x j = F ( 1/4 − λ j ) = f ( i 1/4 − λ j ) ∈ ( 1 , a 0 ] , the constant eigenfunction has A X A_X A X -eigenvalue a 0 a_0 a 0 , and the part of the spectrum of Δ X \Delta_X Δ X in [ 1 / 4 , ∞ ) [1/4,\infty) [ 1/4 , ∞ ) is mapped into [ 0 , 1 ] [0,1] [ 0 , 1 ] . Let χ Q \chi_Q χ Q be the indicator of the complement of the n n n width-one cusp regions above height Y Q = 3 e L Q Y_Q = 3e^{LQ} Y Q = 3 e L Q , and let P 0 P_0 P 0 be the projection onto constants. For a real polynomial p = x p ~ p = x\widetilde p p = x p with p ( 0 ) = 0 p(0)=0 p ( 0 ) = 0 and deg p ≤ Q \deg p \le Q deg p ≤ Q , write N ∗ ( p ) = ∥ p ~ ∥ [ 0 , a 0 ] N_*(p)=\|\widetilde p\|_{[0,a_0]} N ∗ ( p ) = ∥ p ∥ [ 0 , a 0 ] , N R ( p ) = ∥ p ~ ∥ [ 0 , 1 ] N_{\mathbb R}(p)=\|\widetilde p\|_{[0,1]} N R ( p ) = ∥ p ∥ [ 0 , 1 ] , and τ X , Q ( p ) = Tr ( χ Q p ( A X ) χ Q ) \tau_{X,Q}(p)=\operatorname{Tr}(\chi_Qp(A_X)\chi_Q) τ X , Q ( p ) = Tr ( χ Q p ( A X ) χ Q ) . We write I ( p ) \mathcal I(p) I ( p ) and P Y Q ( p ) \mathcal P_{Y_Q}(p) P Y Q ( p ) for Huang’s identity and parabolic terms, O X ( p ) \mathcal O_X(p) O X ( p ) for his orbital term, and ω 1 , ω 2 , n \omega_1,\omega_{2,n} ω 1 , ω 2 , n for the first two canonical coefficients of the expansion of E O X ( p ) \mathbb E\mathcal O_X(p) E O X ( p ) in powers of 1 / g 1/g 1/ g , as in (216) . Put s = n + 1 s=n+1 s = n + 1 , as in Appendix § 10.10 .
(I1) ([101, Theorem 5.3, Proposition 5.1 and equations (5.44)–(5.45)]). For every such X , Q X,Q X , Q and p p p , the compression χ Q p ( A X ) χ Q \chi_Qp(A_X)\chi_Q χ Q p ( A X ) χ Q is trace class and
Tr ( χ Q p ( A X ) ( I − P 0 ) χ Q ) = ( area ( X ) − n Y Q ) I ( p ) + n P Y Q ( p ) + μ 0 ( p ) + n p ( a 0 ) area ( X ) Y Q ) + ( O X ( p ) − ω 1 ( p ) ) . \operatorname{Tr}\bigl(\chi_Qp(A_X)(I-P_0)\chi_Q\bigr)
=
\left(\operatorname{area}(X)-\frac{n}{Y_Q}\right)\mathcal I(p)+n\mathcal P_{Y_Q}(p)+\mu_0(p)+\frac{np(a_0)}{\operatorname{area}(X)Y_Q)}+\bigl(\mathcal O_X(p)-\omega_1(p)\bigr). Tr ( χ Q p ( A X ) ( I − P 0 ) χ Q ) = ( area ( X ) − Y Q n ) I ( p ) + n P Y Q ( p ) + μ 0 ( p ) + area ( X ) Y Q ) n p ( a 0 ) + ( O X ( p ) − ω 1 ( p ) ) .
where μ 0 \mu_0 μ 0 is a fixed signed measure supported in [ 0 , 1 ] [0,1] [ 0 , 1 ] with total variation less than 5, and ω 1 ( p ) = p ( a 0 ) + μ 0 ( p ) \omega_1(p)=p(a_0)+\mu_0(p) ω 1 ( p ) = p ( a 0 ) + μ 0 ( p ) is independent of n n n . Moreover τ X , Q ( p ) ≥ 0 \tau_{X,Q}(p)\ge0 τ X , Q ( p ) ≥ 0 when p ≥ 0 p\ge0 p ≥ 0 on the spectrum of A X A_X A X , and ∣ τ X , Q ( p ) ∣ ≤ N ∗ ( p ) τ X , Q ( x ) |\tau_{X,Q}(p)|\le N_*(p)\tau_{X,Q}(x) ∣ τ X , Q ( p ) ∣ ≤ N ∗ ( p ) τ X , Q ( x ) . More generally, if ψ = x ξ \psi=x\xi ψ = x ξ with ξ \xi ξ bounded and Borel on [ 0 , a 0 ] [0,a_0] [ 0 , a 0 ] , then χ Q ψ ( A X ) ( I − P 0 ) χ Q \chi_Q\psi(A_X)(I-P_0)\chi_Q χ Q ψ ( A X ) ( I − P 0 ) χ Q is trace class and ∣ Tr ( χ Q ψ ( A X ) ( I − P 0 ) χ Q ) ∣ ≤ ∥ ξ ∥ [ 0 , a 0 ] τ X , Q ( x ) |\operatorname{Tr}(\chi_Q\psi(A_X)(I-P_0)\chi_Q)|\le\|\xi\|_{[0,a_0]}\tau_{X,Q}(x) ∣ Tr ( χ Q ψ ( A X ) ( I − P 0 ) χ Q ) ∣ ≤ ∥ ξ ∥ [ 0 , a 0 ] τ X , Q ( x ) [101, (5.29) and (6.63)].
(I2) ([101, Proposition 5.4]). There are constants I f , J f I_f,J_f I f , J f , depending only on f f f , such that for every integer Q ≥ 1 Q\ge1 Q ≥ 1 and every such p p p , ∣ I ( p ) ∣ ≤ I f N R ( p ) |\mathcal I(p)|\le I_fN_{\mathbb R}(p) ∣ I ( p ) ∣ ≤ I f N R ( p ) and ∣ P Y Q ( p ) ∣ ≤ J f Q 2 N R ( p ) |\mathcal P_{Y_Q}(p)|\le J_fQ^2N_{\mathbb R}(p) ∣ P Y Q ( p ) ∣ ≤ J f Q 2 N R ( p ) . These bounds use only the values of p ~ \widetilde p p on the real spectral image [ 0 , 1 ] [0,1] [ 0 , 1 ] .
(I3) ([101, Lemma 6.8, equation (6.43)]). There is an integer c 3 ≥ 2 c_3\ge2 c 3 ≥ 2 , depending only on f f f , such that E τ X , Q ( x ) ≤ D f ( g + n Q ) \mathbb E\tau_{X,Q}(x)\le D_f(g+nQ) E τ X , Q ( x ) ≤ D f ( g + n Q ) for every Q ≥ 1 Q\ge1 Q ≥ 1 whenever g ≥ c 3 s 2 + [ 2 c 3 ] c 3 s g\ge c_3s^2+[2c_3]^ {c_3}s g ≥ c 3 s 2 + [ 2 c 3 ] c 3 s ; here D f D_f D f depends only on f f f . At this threshold every degree-at-most-Q Q Q polynomial trace is absolutely integrable. (I4) ([101, Proposition 6.9, equations (6.46)–(6.47)].) With D 0 D_0 D 0 , C 0 C_0 C 0 from [101, Lemma 6.7], D 3 = max { D 0 , 3 c 3 } D_3=\max\{D_0,3c_3\} D 3 = max { D 0 , 3 c 3 } and A Q = ( 3 C 0 Q ) D 0 A_Q=(3C_0Q)^{D_0} A Q = ( 3 C 0 Q ) D 0 , there is C a d d ≥ 2 C_{\mathrm{add}}\ge2 C add ≥ 2 , depending only on f f f , such that for all integers n n n , Q ≥ 1 Q\ge1 Q ≥ 1 with g ≥ C a d d s 2 + C a d d Q D 3 s g\ge C_{\mathrm{add}}s^2+C_{\mathrm{add}}Q^{D_3}s g ≥ C add s 2 + C add Q D 3 s and every such p p p ,
∣ E O X ( p ) − ω 1 ( p ) − g − 1 ω 2 , n ( p ) ∣ ≤ { 2 s 3 A Q 3 g 2 + s 2 g 2 − Q } N ∗ ( p ) . \left|\mathbb{E}O_X(p)-\omega_1(p)-g^{-1}\omega_{2,n}(p)\right|\le\left\{\frac{2s^3A_Q^3}{g^2}+\frac{s^2}{g}2^{-Q}\right\}N_*(p). E O X ( p ) − ω 1 ( p ) − g − 1 ω 2 , n ( p ) ≤ { g 2 2 s 3 A Q 3 + g s 2 2 − Q } N ∗ ( p ) .
This threshold implies the one in (I3).
(I5) ([101, Lemma 6.10 and Theorem 6.11], with the quadratic identity [101, (3.56)].) Put m = 2 D 0 + 2 m=2D_0+2 m = 2 D 0 + 2 . For each n ≥ 1 n\ge1 n ≥ 1 the functional K n ( q ) = ω 2 , n ( x q ) \mathcal{K}_n(q)=\omega_{2,n}(xq) K n ( q ) = ω 2 , n ( x q ) on polynomials extends uniquely to a distribution supported in [ 0 , a 0 ] [0,a_0] [ 0 , a 0 ] of order at most m m m , with ∣ K n ( φ ) ∣ ≤ K f ( n + 1 ) 2 ∥ φ ∥ C m ( [ 0 , a 0 ] ) |\mathcal{K}_n(\varphi)|\le K_f(n+1)^2\|\varphi\|_{C^m([0,a_0])} ∣ K n ( φ ) ∣ ≤ K f ( n + 1 ) 2 ∥ φ ∥ C m ([ 0 , a 0 ]) for a constant K f K_f K f depending only on f f f . The exact identity ω 2 , n = W 0 + n W 1 + ( n 2 ) W 2 \omega_{2,n}=W_0+nW_1+\binom{n}{2}W_2 ω 2 , n = W 0 + n W 1 + ( 2 n ) W 2 gives
K n = K ( 0 ) + n K ( 1 ) + ( n 2 ) T 2 , \mathcal{K}_n=\mathcal{K}^{(0)}+n\mathcal{K}^{(1)}+\binom{n}{2}\mathcal{T}_2, K n = K ( 0 ) + n K ( 1 ) + ( 2 n ) T 2 ,
where T 2 = K 3 − 2 K 2 + K 1 \mathcal{T}_2=\mathcal{K}_3-2\mathcal{K}_2+\mathcal{K}_1 T 2 = K 3 − 2 K 2 + K 1 , K ( 1 ) = 3 K 2 − K 3 − 2 K 1 \mathcal{K}^{(1)}=3\mathcal{K}_2-\mathcal{K}_3-2\mathcal{K}_1 K ( 1 ) = 3 K 2 − K 3 − 2 K 1 and K ( 0 ) = 3 K 1 − 3 K 2 + K 3 \mathcal{K}^{(0)}=3\mathcal{K}_1-3\mathcal{K}_2+\mathcal{K}_3 K ( 0 ) = 3 K 1 − 3 K 2 + K 3 are the weighted extensions of W 2 W_2 W 2 , W 1 W_1 W 1 and W 0 W_0 W 0 ; each has order at most m m m and norm at most C f C_f C f on C m ( [ 0 , a 0 ] ) C^m([0,a_0]) C m ([ 0 , a 0 ]) . On ( 1 , a 0 ] (1,a_0] ( 1 , a 0 ] the distribution T 2 \mathcal{T}_2 T 2 is the positive measure x η 2 x\eta_2 x η 2 of (214) , and T 2 ( ψ / x ) = ∫ ψ d η 2 \mathcal{T}_2(\psi/x)=\int\psi\,d\eta_2 T 2 ( ψ / x ) = ∫ ψ d η 2 for smooth ψ \psi ψ vanishing near [ 0 , 1 ] [0,1] [ 0 , 1 ] , including tests with ψ ( a 0 ) ≠ 0 \psi(a_0)\ne0 ψ ( a 0 ) = 0 .
(I6) ([101, Lemma 6.5].) If u u u is a normalized eigenfunction with eigenvalue 0 < λ ≤ 1 / 4 − ε 0<\lambda\le1/4-\varepsilon 0 < λ ≤ 1/4 − ε , then
∥ χ Q u ∥ 2 2 ≥ 1 − e − 2 ε L Q , \|\chi_Q u\|_2^2\ge1-e^{-2\sqrt{\varepsilon}L_Q}, ∥ χ Q u ∥ 2 2 ≥ 1 − e − 2 ε L Q ,
uniformly in the surface and in its cusp count.
Inputs (I1)–(I5) are used in the following lemma; (I6) is used afterwards to remove the truncation weights.
Lemma K.1 (Trace approximation for c c c -dependent tests). Let ψ c \psi_c ψ c be smooth on [ 0 , a 0 ] [0,a_0] [ 0 , a 0 ] , vanishing on [ 0 , 1 ] [0,1] [ 0 , 1 ] , with
∥ ψ c / x ∥ C r ( [ 0 , a 0 ] ) ≤ M r c − 2 r \|\psi_c/x\|_{C^r([0,a_0])}\le M_rc^{-2r} ∥ ψ c / x ∥ C r ([ 0 , a 0 ]) ≤ M r c − 2 r for every fixed r ≥ 0 r\ge0 r ≥ 0 , the constants M r M_r M r being independent of g g g ; by (K.1) this includes ψ c = φ ( J ( x ) / c 2 ) \psi_c=\varphi(J(x)/c^2) ψ c = φ ( J ( x ) / c 2 ) for φ \varphi φ as above. Let r Q r_Q r Q , v Q = x r Q v_Q=xr_Q v Q = x r Q and e m e_m e m be as in (K.2), whose bound holds for every such ψ c \psi_c ψ c , and let Q = ⌊ g σ ⌋ Q=\lfloor g^\sigma\rfloor Q = ⌊ g σ ⌋ with σ \sigma σ as above. Suppose ( n + 1 ) 2 / g → 0 (n+1)^2/g\to0 ( n + 1 ) 2 / g → 0 and Q D 3 ( n + 1 ) / g → 0 Q^{D_3}(n+1)/g\to0 Q D 3 ( n + 1 ) / g → 0 . For an orthonormal basis u j u_j u j of eigenfunctions with 0 < λ j < 1 / 4 0<\lambda_j<1/4 0 < λ j < 1/4 and x j = F ( 1 / 4 − λ j ) x_j=F(1/4-\lambda_j) x j = F ( 1/4 − λ j ) , put
D X , Q ( ψ c ) = Tr ( χ Q ψ c ( A X ) ( I − P 0 ) χ Q ) = ∑ 0 < λ j < 1 / 4 ψ c ( x j ) ∥ χ Q u j ∥ 2 2 . D_{X,Q}(\psi_c)=\operatorname{Tr}(\chi_Q\psi_c(A_X)(I-P_0)\chi_Q)=\sum_{0<\lambda_j<1/4}\psi_c(x_j)\|\chi_Q u_j\|_2^2. D X , Q ( ψ c ) = Tr ( χ Q ψ c ( A X ) ( I − P 0 ) χ Q ) = 0 < λ j < 1/4 ∑ ψ c ( x j ) ∥ χ Q u j ∥ 2 2 . Then, with m = 2 D 0 + 2 m=2D_0+2 m = 2 D 0 + 2 and a constant C C C depending only on f f f and the constants M r M_r M r , for all large g g g ,
∣ E D X , Q ( ψ c ) − 1 g ( n 2 ) T 2 ( ψ c / x ) ∣ ≤ C { ( g + n Q 2 ) e 0 + ( n + 1 ) 2 g e m + n + 1 g c − 2 m + ( n + 1 ) 3 Q 3 D 0 g 2 + ( n + 1 ) 2 g 2 − Q + n area ( X ) Y Q } , Y Q = 3 e L Q . \left|\mathbb{E}D_{X,Q}(\psi_c)-\frac{1}{g}\binom{n}{2}\mathcal{T}_2(\psi_c/x)\right|
\le C\left\{(g+nQ^2)e_0+\frac{(n+1)^2}{g}e_m+\frac{n+1}{g}c^{-2m}+\frac{(n+1)^3Q^{3D_0}}{g^2}+\frac{(n+1)^2}{g}2^{-Q}+\frac{n}{\operatorname{area}(X)}Y_Q\right\},
\qquad Y_Q=3e^{L_Q}. E D X , Q ( ψ c ) − g 1 ( 2 n ) T 2 ( ψ c / x ) ≤ C { ( g + n Q 2 ) e 0 + g ( n + 1 ) 2 e m + g n + 1 c − 2 m + g 2 ( n + 1 ) 3 Q 3 D 0 + g ( n + 1 ) 2 2 − Q + area ( X ) n Y Q } , Y Q = 3 e L Q .
Proof. Write ξ c = ψ c / x \xi_c=\psi_c/x ξ c = ψ c / x . Since ψ c \psi_c ψ c vanishes on [ 0 , 1 ] [0,1] [ 0 , 1 ] , the test is zero on the whole real spectral image, and no sign is attributed to the parabolic term. Hence N R ( v Q ) = ∥ r Q ∥ [ 0 , 1 ] ≤ e 0 N_{\mathbb{R}}(v_Q)=\|r_Q\|_{[0,1]}\le e_0 N R ( v Q ) = ∥ r Q ∥ [ 0 , 1 ] ≤ e 0 , N ∗ ( v Q ) ≤ ∥ ξ c ∥ ∞ + e 0 ≤ M 0 + e 0 N_*(v_Q)\le\|\xi_c\|_\infty+e_0\le M_0+e_0 N ∗ ( v Q ) ≤ ∥ ξ c ∥ ∞ + e 0 ≤ M 0 + e 0 , and ∣ v Q ( a 0 ) ∣ ≤ a 0 N ∗ ( v Q ) ≤ C |v_Q(a_0)|\le a_0N_*(v_Q)\le C ∣ v Q ( a 0 ) ∣ ≤ a 0 N ∗ ( v Q ) ≤ C . For ψ c = φ ( J ( x ) / c 2 ) \psi_c=\varphi(J(x)/c^2) ψ c = φ ( J ( x ) / c 2 ) the vanishing holds on [ 0 , F ( a c 2 ) ] ⊃ [ 0 , 1 ] [0,F(ac^2)]\supset[0,1] [ 0 , F ( a c 2 )] ⊃ [ 0 , 1 ] because φ \varphi φ vanishes below a a a . The hypotheses ( n + 1 ) 2 / g → 0 (n+1)^2/g\to0 ( n + 1 ) 2 / g → 0 and Q D 3 ( n + 1 ) / g → 0 Q^{D_3}(n+1)/g\to0 Q D 3 ( n + 1 ) / g → 0 make the threshold of (I4), and hence that of (I3), hold for all large g g g . The operator ψ c ( A X ) ( I − P 0 ) \psi_c(A_X)(I-P_0) ψ c ( A X ) ( I − P 0 ) has finite rank, so the displayed sum representation of D X , Q ( ψ c ) D_{X,Q}(\psi_c) D X , Q ( ψ c ) holds, and D X , Q ( v Q ) = Tr ( χ Q v Q ( A X ) ( I − P 0 ) χ Q ) D_{X,Q}(v_Q)=\operatorname{Tr}(\chi_Qv_Q(A_X)(I-P_0)\chi_Q) D X , Q ( v Q ) = Tr ( χ Q v Q ( A X ) ( I − P 0 ) χ Q ) is the quantity in (I1); all expectations below are finite by (I3). The estimate follows from the exact identity and the following four bounds, each attributed to the input it uses. Smooth test to polynomial , by (I1) and (I3). The spectral domination of (I1), applied to ψ c − v Q = x ( ξ c − r Q ) \psi_c-v_Q=x(\xi_c-r_Q) ψ c − v Q = x ( ξ c − r Q ) , whose quotient is bounded by e 0 e_0 e 0 on [ 0 , a 0 ] [0,a_0] [ 0 , a 0 ] , and the first moment of (I3) give
∣ E D X , Q ( ψ c ) − E D X , Q ( v Q ) ∣ ≤ e 0 E τ X , Q ( x ) ≤ D f ( g + n Q ) e 0 . \left|\mathbb{E}D_{X,Q}(\psi_c)-\mathbb{E}D_{X,Q}(v_Q)\right|\le e_0\mathbb{E}\tau_{X,Q}(x)\le D_f(g+n_Q)e_0. ∣ E D X , Q ( ψ c ) − E D X , Q ( v Q ) ∣ ≤ e 0 E τ X , Q ( x ) ≤ D f ( g + n Q ) e 0 .
Exact identity , by (I1).
D X , Q ( v Q ) = ( area ( X ) − n Y Q ) I ( v Q ) + n P Y Q ( v Q ) + μ 0 ( v Q ) + n v Q ( a 0 ) area ( X ) Y Q + O X ( v Q ) − ω 1 ( v Q ) . D_{X,Q}(v_Q)=\left(\operatorname{area}(X)-\frac{n}{Y_Q}\right)\mathcal{I}(v_Q)+n\mathcal{P}_{Y_Q}(v_Q)+\mu_0(v_Q)+\frac{nv_Q(a_0)}{\operatorname{area}(X)Y_Q}+\mathcal{O}_X(v_Q)-\omega_1(v_Q). D X , Q ( v Q ) = ( area ( X ) − Y Q n ) I ( v Q ) + n P Y Q ( v Q ) + μ 0 ( v Q ) + area ( X ) Y Q n v Q ( a 0 ) + O X ( v Q ) − ω 1 ( v Q ) .
The constant eigenfunction is removed here, before any approximation; its remaining truncation correction is the fourth term, and nothing is required of ψ c ( a 0 ) \psi_c(a_0) ψ c ( a 0 ) .
Identity, parabolic and μ 0 \mu_0 μ 0 terms , by (I2) and (I1). Since ∣ v Q ∣ ≤ N R ( v Q ) ≤ e 0 |v_Q|\le N_{\mathbb{R}}(v_Q)\le e_0 ∣ v Q ∣ ≤ N R ( v Q ) ≤ e 0 on [ 0 , 1 ] [0,1] [ 0 , 1 ] and ∣ v Q ( a 0 ) ∣ ≤ C |v_Q(a_0)|\le C ∣ v Q ( a 0 ) ∣ ≤ C ,
∣ ( area ( X ) − n Y Q ) I ( v Q ) + n P Y Q ( v Q ) + μ 0 ( v Q ) ∣ ≤ ( I f area ( X ) + J f n Q 2 + 5 ) e 0 ≤ C ( g + n Q 2 ) e 0 , \left|\left(\operatorname{area}(X)-\frac{n}{Y_Q}\right)\mathcal{I}(v_Q)+n\mathcal{P}_{Y_Q}(v_Q)+\mu_0(v_Q)\right|\le(I_f\operatorname{area}(X)+J_fnQ^2+5)e_0\le C(g+nQ^2)e_0, ( area ( X ) − Y Q n ) I ( v Q ) + n P Y Q ( v Q ) + μ 0 ( v Q ) ≤ ( I f area ( X ) + J f n Q 2 + 5 ) e 0 ≤ C ( g + n Q 2 ) e 0 ,
∣ n v Q ( a 0 ) area ( X ) Y Q ∣ ≤ C n area ( X ) Y Q . \left|\frac{nv_Q(a_0)}{\operatorname{area}(X)Y_Q}\right|\le\frac{Cn}{\operatorname{area}(X)Y_Q}. area ( X ) Y Q n v Q ( a 0 ) ≤ area ( X ) Y Q C n .
Orbital expansion , by (I4). Since N ∗ ( v Q ) ≤ C N_*(v_Q)\le C N ∗ ( v Q ) ≤ C ,
∣ E O X ( v Q ) − ω 1 ( v Q ) − 1 g ω 2 , n ( v Q ) ∣ ≤ C { ( n + 1 ) 3 Q 3 D 0 g 2 + ( n + 1 ) 2 g 2 − Q } . \left|\mathbb{E}\mathcal{O}_X(v_Q)-\omega_1(v_Q)-\frac{1}{g}\omega_{2,n}(v_Q)\right|\le C\left\{\frac{(n+1)^3Q^{3D_0}}{g^2}+\frac{(n+1)^2}{g}2^{-Q}\right\}. E O X ( v Q ) − ω 1 ( v Q ) − g 1 ω 2 , n ( v Q ) ≤ C { g 2 ( n + 1 ) 3 Q 3 D 0 + g ( n + 1 ) 2 2 − Q } .
The quadratic coefficient , by (I5). By the order-m m m bound and the hypothesis with r = m r=m r = m ,
∣ ω 2 , n ( v Q ) − K n ( ξ c ) ∣ = ∣ K n ( r Q − ξ c ) ∣ ≤ K f ( n + 1 ) 2 e m , \left|\omega_{2,n}(v_Q)-\mathcal{K}_n(\xi_c)\right|=\left|\mathcal{K}_n(r_Q-\xi_c)\right|\le K_f(n+1)^2e_m, ∣ ω 2 , n ( v Q ) − K n ( ξ c ) ∣ = ∣ K n ( r Q − ξ c ) ∣ ≤ K f ( n + 1 ) 2 e m ,
∣ K n ( ξ c ) − ( n 2 ) T 2 ( ξ c ) ∣ = ∣ K ( 0 ) ( ξ c ) + n K ( 1 ) ( ξ c ) ∣ ≤ C ( n + 1 ) ∥ ξ c ∥ C m ≤ C ( n + 1 ) c − 2 m . \left|\mathcal{K}_n(\xi_c)-\binom{n}{2}\mathcal{T}_2(\xi_c)\right|=\left|\mathcal{K}^{(0)}(\xi_c)+n\mathcal{K}^{(1)}(\xi_c)\right|\le C(n+1)\|\xi_c\|_{C^m}\le C(n+1)c^{-2m}. K n ( ξ c ) − ( 2 n ) T 2 ( ξ c ) = K ( 0 ) ( ξ c ) + n K ( 1 ) ( ξ c ) ≤ C ( n + 1 ) ∥ ξ c ∥ C m ≤ C ( n + 1 ) c − 2 m .
Taking the expectation of the exact identity, dividing the last two bounds by g g g , and adding all the bounds proves (K.3). □
The proof uses c c c only through the seminorm hypothesis. If a test ψ \psi ψ , smooth on [ 0 , a 0 ] [0,a_0] [ 0 , a 0 ] and vanishing on [ 0 , 1 ] [0,1] [ 0 , 1 ] , satisfies ∥ ψ / x ∥ C r ( [ 0 , a 0 ] ) ≤ M r \|\psi/x\|_{C^r([0,a_0])}\le M_r ∥ ψ / x ∥ C r ([ 0 , a 0 ]) ≤ M r for every fixed r r r , with M r M_r M r independent of g g g , the same argument gives (K.3) with $c^{-2m}$ replaced by $1$ and with $e_m\le C_{K,m}Q^{-K}$; this form is used for fixed physical tests in Subsection K.5`.
For large fixed K K K , every term in (K.3) tends to zero. Indeed ( g + n Q 2 ) e 0 ≤ g 1 + 2 σ − σ K + o ( 1 ) (g+nQ^2)e_0\le g^{1+2\sigma-\sigma K+o(1)} ( g + n Q 2 ) e 0 ≤ g 1 + 2 σ − σ K + o ( 1 ) by (K.2); c − 2 m = g o ( 1 ) c^{-2m}=g^{o(1)} c − 2 m = g o ( 1 ) because log ( 1 / c ) = o ( log g ) \log(1/c)=o(\log g) log ( 1/ c ) = o ( log g ) , while ( n + 1 ) / g → 0 (n+1)/g\to0 ( n + 1 ) / g → 0 ; ( n + 1 ) 3 Q 3 D 0 / g 2 ≤ [ ( n + 1 ) 2 / g ] g 3 σ D 0 − 1 / 2 (n+1)^3Q^{3D_0}/g^2\le[(n+1)^2/g]g^{3\sigma D_0-1/2} ( n + 1 ) 3 Q 3 D 0 / g 2 ≤ [( n + 1 ) 2 / g ] g 3 σ D 0 − 1/2 eventually, and 3 σ D 0 < 1 / 2 3\sigma D_0<1/2 3 σ D 0 < 1/2 ; and n / ( area ( X ) Y Q ) ≤ e − L Q n/(\operatorname{area}(X)Y_Q)\le e^{-LQ} n / ( area ( X ) Y Q ) ≤ e − L Q . For ψ c = ϕ ( J ( x ) / c 2 ) \psi_c=\phi(J(x)/c^2) ψ c = ϕ ( J ( x ) / c 2 ) , Proposition 15.5 and (I5) give
T 2 ( ψ c / x ) = ∫ ψ c d η 2 = 2 κ 0 ∫ 0 ∞ ϕ ( 1 / 4 − β ( s ) c 2 ) sinh ( s / 2 ) d s . \mathcal{T}_2(\psi_c/x)=\int\psi_c\,d\eta_2=2\kappa_0\int_0^\infty\phi\left(\frac{1/4-\beta(s)}{c^2}\right)\sinh(s/2)\,ds. T 2 ( ψ c / x ) = ∫ ψ c d η 2 = 2 κ 0 ∫ 0 ∞ ϕ ( c 2 1/4 − β ( s ) ) sinh ( s /2 ) d s .
since w ( s ) = sinh ( s / 2 ) / ( 4 π ) = 2 κ 0 sinh ( s / 2 ) w(s)=\sinh(s/2)/(4\pi)=2\kappa_0\sinh(s/2) w ( s ) = sinh ( s /2 ) / ( 4 π ) = 2 κ 0 sinh ( s /2 ) and ψ c ( x s ) = ϕ ( ( 1 / 4 − β ( s ) ) / c 2 ) \psi_c(x_s)=\phi((1/4-\beta(s))/c^2) ψ c ( x s ) = ϕ (( 1/4 − β ( s )) / c 2 ) . The third term in (K.3) is the cost of removing W 0 + n W 1 W_0+nW_1 W 0 + n W 1 . The resulting weighted spectral sum has mean
E D X , Q ( ψ c ) = n ( n − 1 ) g κ 0 ∫ 0 ∞ ϕ ( h ∗ ( y ) c 2 ) d y + o ( 1 ) . (440) \mathbb{E}D_{X,Q}(\psi_c)=\frac{n(n-1)}{g}\kappa_0\int_0^\infty\phi\left(\frac{h_*(y)}{c^2}\right)\,dy+o(1). \tag*{(440)} E D X , Q ( ψ c ) = g n ( n − 1 ) κ 0 ∫ 0 ∞ ϕ ( c 2 h ∗ ( y ) ) d y + o ( 1 ) . (440)
This is the physical form of the local model identification, since y = 2 ( cosh ( s / 2 ) − 1 ) y=2(\cosh(s/2)-1) y = 2 ( cosh ( s /2 ) − 1 ) gives d y = sinh ( s / 2 ) d s dy=\sinh(s/2)\,ds d y = sinh ( s /2 ) d s and h ∗ ( y ) = 1 / 4 − β ( s ) h_*(y)=1/4-\beta(s) h ∗ ( y ) = 1/4 − β ( s ) .
For modes on the support, the omitted mass is at most ϵ g = e − 2 c a L Q → 0 \epsilon_g=e^{-2c\sqrt{a}LQ}\to0 ϵ g = e − 2 c a L Q → 0 , by [101, Lemma 6.5]. Apply (K.4) first to a nonnegative smooth majorant of ∣ ϕ ∣ |\phi| ∣ ϕ ∣ . Its weighted expectation is bounded, by (228) and the change y = z / c 2 y = z/c^2 y = z / c 2 . The unweighted-minus-weighted expectation is at most ϵ g / ( 1 − ϵ g ) \epsilon_g/(1-\epsilon_g) ϵ g / ( 1 − ϵ g ) times that bound. Dominated convergence in the same change of variables gives
E ⟨ ϕ , D X ⟩ ⟶ ∫ 0 ∞ ϕ ( x ) κ 0 d x x 2 , E N g ( a ) ⟶ κ 0 a . (441) \mathbb{E}\langle\phi,D_X\rangle\longrightarrow\int_0^\infty\phi(x)\frac{\kappa_0\,dx}{x^2},\qquad\mathbb{E}N_g(a)\longrightarrow\frac{\kappa_0}{a}. \tag*{(441)} E ⟨ ϕ , D X ⟩ ⟶ ∫ 0 ∞ ϕ ( x ) x 2 κ 0 d x , E N g ( a ) ⟶ a κ 0 . (441)
For the second assertion, squeeze the tail indicator by smooth tests constant near infinity. The limiting measure has no atom at a a a . This argument bounds unweighted counts without assuming such a bound in advance.
Growing waists and conditional isolation
We next prove geometric Poisson convergence of the pants counts Z g ( a ) Z_g(a) Z g ( a ) and their isolation for any R = o ( log g ) R=o(\log g) R = o ( log g ) , R ≥ 1 R\ge1 R ≥ 1 , and any cutoff width ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 with ℓ = o ( log g ) \ell=o(\log g) ℓ = o ( log g ) , still assuming only log ( 1 / c ) = o ( log g ) \log(1/c)=o(\log g) log ( 1/ c ) = o ( log g ) . The argument is the factorial-moment method of Appendix I with waists growing like log ( 1 / c ) \log(1/c) log ( 1/ c ) : the boundary volumes are controlled at growing lengths by [101, Proposition 2.3], and the cut-graph bound (I.2) is applied conditionally on a marked tuple to show that almost every pair of pants is good, that is, isolated at radius R ′ R' R ′ with an embedded funnel neighborhood. The outputs used later are the conditional bound (K.11), the discard estimate (K.12) and the Poisson limits of the good counts. The isolation radius is
R ′ = R + ℓ + 1 , (442) R'=R+\ell+1, \tag*{(442)} R ′ = R + ℓ + 1 , (442)
chosen so that the cutoff functions of the next subsection, which are supported in the radius-( R + ℓ ) (R+\ell) ( R + ℓ ) funnel neighborhoods, have disjoint supports for distinct good pants; there ℓ = 1 / c \ell=1/c ℓ = 1/ c and R R R are fixed by (K.17), and R ′ = o ( log g ) R'=o(\log g) R ′ = o ( log g ) exactly when R = o ( log g ) R=o(\log g) R = o ( log g ) . For fixed a > 0 a>0 a > 0 , let S a S_a S a solve q ( 2 cosh ( S a / 4 ) ) = a c 2 q(2\cosh(S_a/4))=ac^2 q ( 2 cosh ( S a /4 )) = a c 2 . Equation (228) gives
S a = 4 log ( 1 / c ) − 2 log a + o ( 1 ) . (443) S_a=4\log(1/c)-2\log a+o(1). \tag*{(443)} S a = 4 log ( 1/ c ) − 2 log a + o ( 1 ) . (443)
Boundary volumes at growing lengths. For fixed r r r , the order-zero joint intersection estimates [101, Proposition 2.3], originating in [96, Theorem 4.2], imply, uniformly for p ≤ g p\le\sqrt{g} p ≤ g and 0 ≤ s i ≤ S a 0\le s_i\le S_a 0 ≤ s i ≤ S a ,
V g , p ( s , 0 ) V g , p = ∏ i = 1 r sinh ( s i / 2 ) s i / 2 ( 1 + O r ( p + 1 g ( 1 + r S a ) r + 2 ) ) . (444) \frac{V_{g,p}(\mathbf{s},0)}{V_{g,p}}=\prod_{i=1}^{r}\frac{\sinh(s_i/2)}{s_i/2}\left(1+O_r\left(\frac{p+1}{g}(1+rS_a)^{r+2}\right)\right). \tag*{(444)} V g , p V g , p ( s , 0 ) = i = 1 ∏ r s i /2 sinh ( s i /2 ) ( 1 + O r ( g p + 1 ( 1 + r S a ) r + 2 ) ) . (444)
For the growing-length uniformity, write [ τ d ] g , p [\tau_{\mathbf d}]_{g,p} [ τ d ] g , p , d ∈ Z ≥ 0 p \mathbf d\in\mathbb{Z}_{\ge0}^p d ∈ Z ≥ 0 p , ∣ d ∣ = ∑ i d i |\mathbf d|=\sum_i d_i ∣ d ∣ = ∑ i d i , for the normalized intersection numbers of [101, Proposition 2.3], and E 2 ( v ) = 1 + v + v 2 / 2 E_2(v)=1+v+v^2/2 E 2 ( v ) = 1 + v + v 2 /2 . The order-zero case of that proposition bounds ∣ [ τ d ] g , p / V g , p − 1 ∣ \lvert[\tau_{\mathbf d}]_{g,p}/V_{g,p}-1\rvert ∣[ τ d ] g , p / V g , p − 1 ∣ by C ( p + 1 ) g − 1 E 2 ( ∣ d ∣ ) C(p+1)g^{-1}E_2(|\mathbf d|) C ( p + 1 ) g − 1 E 2 ( ∣ d ∣ ) . Beyond the dimension cutoff the bracket is zero and ∣ d ∣ ≥ g |\mathbf d|\ge g ∣ d ∣ ≥ g , so the same bound holds there. Summation against the boundary series gives C r ( p + 1 ) g − 1 e ∑ s i / 2 ( 1 + ∑ s i ) 2 C_r(p+1)g^{-1}e^{\sum s_i/2}(1+\sum s_i)^2 C r ( p + 1 ) g − 1 e ∑ s i /2 ( 1 + ∑ s i ) 2 . Divide by the leading product using sinh ( s / 2 ) / ( s / 2 ) ≥ e s / 2 / ( 1 + s ) \sinh(s/2)/(s/2)\ge e^{s/2}/(1+s) sinh ( s /2 ) / ( s /2 ) ≥ e s /2 / ( 1 + s ) . The order-zero genus threshold is linear in p p p , and hence permits these parameters. Also V g , j / V g , j + 1 = ( 8 π 2 g ) − 1 ( 1 + O ( ( j + 1 ) / g ) ) V_{g,j}/V_{g,j+1}=(8\pi^2g)^{-1}(1+O((j+1)/g)) V g , j / V g , j + 1 = ( 8 π 2 g ) − 1 ( 1 + O (( j + 1 ) / g )) .
The Campbell integral. The exact contribution of ordered disjoint r r r -tuples of pants with waists in prescribed sets A i ⊂ [ 0 , S a ] A_i\subset[0,S_a] A i ⊂ [ 0 , S a ] is
( n 2 r ) 2 r V g , n ∫ ∏ i A i V g , n − 2 r ( s , 0 ) ∏ i s i d s i . (445) \frac{\binom{n}{2r}}{2^rV_{g,n}}\int_{\prod_i A_i}V_{g,n-2r}(\mathbf{s},0)\prod_i s_i\,ds_i. \tag*{(445)} 2 r V g , n ( 2 r n ) ∫ ∏ i A i V g , n − 2 r ( s , 0 ) i ∏ s i d s i . (445)
The factor 2 − r 2^{-r} 2 − r chooses unordered cusp pairs; the tuple is ordered. Distinct disjoint pants cannot be properly nested, since their difference would be an annulus between distinct geodesics. Their complement is connected. Thus there are no other complementary types in (K.9). By (K.8), this integral is asymptotic to the product with one-variable measure
μ g ( d s ) = κ 0 c 2 sinh ( s / 2 ) d s . \mu_g(ds)=\kappa_0c^2\sinh(s/2)\,ds. μ g ( d s ) = κ 0 c 2 sinh ( s /2 ) d s .
Under s ↦ a ( 2 cosh ( s / 4 ) ) / c 2 s\mapsto a(2\cosh(s/4))/c^2 s ↦ a ( 2 cosh ( s /4 )) / c 2 , its upper-tail mass tends to κ 0 / a \kappa_0/a κ 0 / a , since y = 2 ( cosh ( s / 2 ) − 1 ) y=2(\cosh(s/2)-1) y = 2 ( cosh ( s /2 ) − 1 ) and h ∗ ( y ) ∼ 1 / y h_*(y)\sim1/y h ∗ ( y ) ∼ 1/ y .
Conditional cut-graph bounds. Conditioning in the Campbell integral leaves a WP complement of genus g g g , with n − 2 r n-2r n − 2 r cusps, r r r labeled boundaries and uniform gluing twists. Its total number of ends is N = n − r N=n-r N = n − r . More explicitly, if s s s are the marked waist lengths, the conditional law of the complement is its WP measure divided by V g , N ( s , 0 ) V_{g,N}(s,0) V g , N ( s , 0 ) ; each gluing twist has law d t i / s i dt_i/s_i d t i / s i . Cutting a witness graph Γ \Gamma Γ therefore gives, for fixed cusp labels, integrals of the form
C Γ ∫ ∑ e ℓ e ≤ B ∏ v V g v , n v ( s v , ℓ v , 0 ) V g , N ( s , 0 ) ∏ e ℓ e d ℓ e , C_{\Gamma}\int_{\sum_e \ell_e\le B}\frac{\prod_v V_{g_v,n_v}(s_v,\ell_v,0)}{V_{g,N}(s,0)}\prod_e \ell_e\,d\ell_e, C Γ ∫ ∑ e ℓ e ≤ B V g , N ( s , 0 ) ∏ v V g v , n v ( s v , ℓ v , 0 ) e ∏ ℓ e d ℓ e ,
where C Γ C_{\Gamma} C Γ is the fixed symmetry factor and the product runs through the cut components. Summing the admissible genera and cusp allocations is exactly the zero-boundary graph sum below. Positivity gives V g , N ( s , 0 ) ≥ V g , N V_{g,N}(s,0)\ge V_{g,N} V g , N ( s , 0 ) ≥ V g , N , while the boundary upper bound charges each old length once and each new length twice. These observations justify applying the graph estimate under the conditional, rather than the original unmarked, law. For a fixed cut graph with ν ≤ 2 \nu\le2 ν ≤ 2 new edges and total new length at most B B B , positivity of the boundary polynomials and the boundary upper bound [101, Lemma 2.1] cost at most
e r S a / 2 + B ( 1 + B ) 2 ν (446) e^{rS_a/2+B}(1+B)^{2\nu} \tag*{(446)} e r S a /2 + B ( 1 + B ) 2 ν (446)
relative to the zero-boundary allocation. Applying the proof of (I.2) with h = 2 g − 2 + N h=2g-2+N h = 2 g − 2 + N , z = N / h z=N/h z = N / h gives C ν h u − w + 1 ϕ 0 , 1 ( z ) u ϕ 0 , 2 ( z ) v C_\nu h^{u-w+1}\phi_{0,1}(z)^u\phi_{0,2}(z)^v C ν h u − w + 1 ϕ 0 , 1 ( z ) u ϕ 0 , 2 ( z ) v . The constants are uniform for N ≤ g N\le\sqrt{g} N ≤ g : the ambient-volume loss is e O ( N 2 / g ) e^{O(N^2/g)} e O ( N 2 / g ) and the finite Gamma-product loss is e O ( ( m + 1 ) 2 / h ) e^{O((m+1)^2/h)} e O (( m + 1 ) 2 / h ) , m ≤ N m\le N m ≤ N .
A degree-two planar vertex, or a planar leaf with at least three ends, supplies O ( g − 1 / 2 ) O(g^{-1/2}) O ( g − 1/2 ) after the remaining bounded factors. For a planar leaf required to contain one specified old boundary, the multinomial factor changes by q / N q/N q / N when that leaf has q q q ends. Before summation this replaces ϕ 0 , 1 \phi_{0,1} ϕ 0 , 1 by z ϕ 0 , 1 ′ / N z\phi'_{0,1}/N z ϕ 0 , 1 ′ / N . In the leading one-leaf graph the factor is
h z ϕ 0 , 1 ′ ( z ) N = O ( N / g ) . h\frac{z\phi'_{0,1}(z)}{N}=O(N/g). h N z ϕ 0 , 1 ′ ( z ) = O ( N / g ) .
Two specified old ends instead give h z 2 ϕ 0 , 1 ′ ′ ( z ) / ( N ( N − 1 ) ) = O ( 1 / g ) hz^2\phi''_{0,1}(z)/(N(N-1))=O(1/g) h z 2 ϕ 0 , 1 ′′ ( z ) / ( N ( N − 1 )) = O ( 1/ g ) . Additional leaves cost h ϕ 0 , 1 ( z ) = O ( 1 ) h\phi_{0,1}(z)=O(1) h ϕ 0 , 1 ( z ) = O ( 1 ) and additional stable vertices only decrease the bound. The distinguished labels cannot be treated as freely chosen cusps.
Good pants and their witnesses. Declare a pair of pants good if its exterior radius-R ′ R' R ′ funnel embeds and its distance from every other pair of pants of waist at most S a S_a S a exceeds 2 R ′ 2R' 2 R ′ , with R ′ = R + ℓ + 1 R'=R+\ell+1 R ′ = R + ℓ + 1 as in (K.6). Every failure has a cut witness of total length at most B = 20 ( r + 1 ) ( S a + R ′ + 1 ) = 20 ( r + 1 ) ( S a + R + ℓ + 2 ) B=20(r+1)(S_a+R'+1)=20(r+1)(S_a+R+\ell+2) B = 20 ( r + 1 ) ( S a + R ′ + 1 ) = 20 ( r + 1 ) ( S a + R + ℓ + 2 ) and at most two new edges. Indeed a first normal collision supplies an essential returning or connecting arc of length at most 2 R ′ 2R' 2 R ′ ; there are no focal points or collisions from boundary-parallel arcs in the cyclic funnel cover. A candidate waist crossing a marked pair of pants gives an essential excursion of length at most S a S_a S a in the complement: a peripheral excursion would bound a geodesic bigon. Otherwise a nearby unmarked pair of pants lies in the complement and is joined to an old boundary by an arc of length at most 2 R ′ 2R' 2 R ′ , unless such an arc first meets another old boundary.
Thicken a returning arc and its old boundary annulus. The result has one old boundary and two new sides. After tightening and filling peripheral once-punctured disks or annuli leading to old boundaries, either both sides remain and give a planar degree-two vertex, or one remains and gives a leaf with the old label and another end. An empty disk would make the arc peripheral; both sides cannot disappear in large genus. Coincident tightened sides are the two incidences of one cut edge. A connecting arc gives a leaf with two specified old ends. A joining arc to an unmarked pair of pants gives a leaf with the old end and its two cusp ends. Thus every witness has one of the suppressed factors above. The graph shapes are finite in number; their genera and all end allocations have already been summed. Only existence of a witness is needed, not a bound on the number of arcs producing it. Markov’s inequality gives, uniformly in the marked lengths, labels and twists,
P ( a marked pants is not good ∣ disjoint tuple ) ≤ C r e C r ( S a + R ′ + 1 ) ( 1 + S a + R ′ ) C r g − 1 / 2 = : ϵ g , r ⟶ 0. (447) \mathbb{P}(\text{a marked pants is not good}\mid\text{disjoint tuple}) \le C_r e^{C_r(S_a+R'+1)}(1+S_a+R')C_r g^{-1/2} =:\epsilon_{g,r}\longrightarrow0. \tag*{(447)} P ( a marked pants is not good ∣ disjoint tuple ) ≤ C r e C r ( S a + R ′ + 1 ) ( 1 + S a + R ′ ) C r g − 1/2 =: ϵ g , r ⟶ 0. (447)
The limit uses S a = O ( log ( 1 / c ) ) = o ( log g ) S_a=O(\log(1/c))=o(\log g) S a = O ( log ( 1/ c )) = o ( log g ) and R ′ = R + ℓ + 1 = o ( log g ) R'=R+\ell+1=o(\log g) R ′ = R + ℓ + 1 = o ( log g ) ; since ℓ = o ( log g ) \ell=o(\log g) ℓ = o ( log g ) , the enlargement of the radius from R R R to R ′ R' R ′ leaves the condition R = o ( log g ) R=o(\log g) R = o ( log g ) unchanged.
Poisson limits of the good counts. Let Z g g o o d Z_g^{\mathrm{good}} Z g good count good pants, using the same S a , R , ℓ S_a,R,\ell S a , R , ℓ for all windows under consideration. The one-pants Campbell formula gives
E ( Z g ( a ) − Z g g o o d ( a ) ) ≤ ϵ g , 1 E Z g ( a ) = o ( 1 ) . (448) \mathbb{E}\bigl(Z_g(a)-Z_g^{\mathrm{good}}(a)\bigr)\le\epsilon_{g,1}\mathbb{E}Z_g(a)=o(1). \tag*{(448)} E ( Z g ( a ) − Z g good ( a ) ) ≤ ϵ g , 1 E Z g ( a ) = o ( 1 ) . (448)
Every good tuple is disjoint, and the conditional probability that a disjoint tuple is good is at least 1 − r ϵ g , r 1-r\epsilon_{g,r} 1 − r ϵ g , r , by the union bound over its r r r marked pants in (K.11). Consequently its mixed factorial moments have the Poisson limits given by (K.9). Higher fixed factorial moments give tightness and moment passage; the Poisson law is determinate. The good counts have bounded second moments by the disjoint-pair integral. Equation (K.12) transfers both convergence and uniform integrability to the full counts. No factorial-moment bound for all intersecting tuples is inferred from a high-probability event.
Model cutoffs, count matching and eigenfunctions
This subsection completes the proof of Theorem 16.1 under the range (229) . We cut off the ground states of the funnel models F s F_s F s at a radius chosen in (K.17) and transplant them to the good pants of the preceding subsection. Min–max then bounds the eigenvalue counts from below by the pants counts; together with the limiting mean (K.5) this matches the two counts (K.19), and the transplanted states identify the eigenfunctions. The mechanism is that the limiting mean equals the pants intensity exactly. After localization errors and threshold shifts are removed, an extra spectral count persisting with positive probability in the limit would contradict this equality; min–max supplies the lower bound from the pants. The cutoff estimates (K.15)–(K.16) are reused in Subsections K.4 and K.5.
Cutoff model states. On the model funnel d r 2 + s 2 cosh 2 r d θ 2 dr^2+s^2\cosh^2r\,d\theta^2 d r 2 + s 2 cosh 2 r d θ 2 , write the normalized eigenfunction of energy β = 1 / 4 − k 2 \beta=1/4-k^2 β = 1/4 − k 2 in Fourier modes, and put v m = s cosh r u m v_m=\sqrt{s\cosh r}\,u_m v m = s cosh r u m . Then
− v m ′ ′ + [ k 2 + ( 1 4 + 4 π 2 m 2 s 2 ) sech 2 r ] v m = 0. (449) -v_m''+\left[k^2+\left(\frac{1}{4}+\frac{4\pi^2m^2}{s^2}\right)\operatorname{sech}^2r\right]v_m=0. \tag*{(449)} − v m ′′ + [ k 2 + ( 4 1 + s 2 4 π 2 m 2 ) sech 2 r ] v m = 0. (449)
Each mode has finite transformed H 1 H^1 H 1 energy. An L 2 L^2 L 2 solution with zero boundary value vanishes by integration by parts, so a nonzero mode divided by its value at any starting point is real and positive on the remaining half-line. It tends to zero at infinity. The maximum principle applied to its difference from e − k t e^{-kt} e − k t gives ∣ v m ( r + t ) ∣ ≤ e − k t ∣ v m ( r ) ∣ |v_m(r+t)|\le e^{-kt}|v_m(r)| ∣ v m ( r + t ) ∣ ≤ e − k t ∣ v m ( r ) ∣ . Integration and Parseval imply
∫ r ≥ R ∣ u ∣ 2 ≤ e − 2 k R . (450) \int_{r\ge R}|u|^2\le e^{-2kR}. \tag*{(450)} ∫ r ≥ R ∣ u ∣ 2 ≤ e − 2 k R . (450)
For R ≥ ℓ > 0 R\ge\ell>0 R ≥ ℓ > 0 , multiply (K.13) by 2 ζ v m ‾ 2\zeta\overline{v_m} 2 ζ v m , where ζ \zeta ζ changes from zero to one on [ R − ℓ , R ] [R-\ell,R] [ R − ℓ , R ] . Cauchy’s inequality and the nonnegative potential give
∑ m ∫ R ∞ ∣ v m ′ ∣ 2 ≤ C ℓ − 2 e − 2 k ( R − ℓ ) . \sum_m\int_R^\infty|v_m'|^2\le C\ell^{-2}e^{-2k(R-\ell)}. m ∑ ∫ R ∞ ∣ v m ′ ∣ 2 ≤ C ℓ − 2 e − 2 k ( R − ℓ ) .
Cutoffs at infinity justify the identity by finite energy. Choose χ = 1 \chi=1 χ = 1 up to R R R , zero beyond R + ℓ R+\ell R + ℓ , with derivatives bounded by C ℓ − 1 C\ell^{-1} C ℓ − 1 and C ℓ − 2 C\ell^{-2} C ℓ − 2 , retaining the whole pair of pants and both complete cusps. The conjugated commutator is exactly − χ ′ ′ v − 2 χ ′ v ′ -\chi''v-2\chi'v' − χ ′′ v − 2 χ ′ v ′ . Therefore
∥ ( Δ − β ) ( χ u ) ∥ 2 ≤ C ℓ − 2 e − k ( R − ℓ ) . (451) \|(\Delta-\beta)(\chi u)\|_2\le C\ell^{-2}e^{-k(R-\ell)}. \tag*{(451)} ∥ ( Δ − β ) ( χ u ) ∥ 2 ≤ C ℓ − 2 e − k ( R − ℓ ) . (451)
E ( χ u ) ∥ χ u ∥ 2 2 − β ≤ C ℓ − 2 e − 2 k R 1 − e − 2 k R . (452) \frac{\mathcal{E}(\chi u)}{\|\chi u\|_2^2}-\beta\le\frac{C\ell^{-2}e^{-2kR}}{1-e^{-2kR}}. \tag*{(452)} ∥ χ u ∥ 2 2 E ( χ u ) − β ≤ 1 − e − 2 k R C ℓ − 2 e − 2 k R . (452)
These are operator-domain vectors: their distributional Laplacian is in L 2 L^2 L 2 , and the complete-surface Laplacian has its maximal L 2 L^2 L 2 domain. Smooth extension by zero at the outer cutoff preserves this.
Transplantation and count matching. Now assume (229) . This is where the condition c log g → ∞ c\log g\to\infty c log g → ∞ itself enters: the cutoff radius R R R must satisfy both c R → ∞ cR\to\infty c R → ∞ , for the funnel tail, and R = o ( log g ) R=o(\log g) R = o ( log g ) , for the isolation estimate. Take
ℓ = c − 1 , R = ( 2 + c log g ) / c . (453) \ell=c^{-1},\qquad R=(2+\sqrt{c\log g})/c. \tag*{(453)} ℓ = c − 1 , R = ( 2 + c log g ) / c . (453)
Then R ′ = R + ℓ + 1 = o ( log g ) R'=R+\ell+1=o(\log g) R ′ = R + ℓ + 1 = o ( log g ) , so the good event of the preceding subsection, with isolation radius (K.6), is available, and the cutoff vectors χ u \chi_u χ u of distinct good pants, supported in their radius-( R + ℓ ) (R+\ell) ( R + ℓ ) funnel neighborhoods together with the pants and cusps, have disjoint supports. Uniformly for k 2 ≥ a c 2 k^2\geq ac^2 k 2 ≥ a c 2 , the operator error divided by c 2 c^2 c 2 is at most C e − a ( 1 + c log g ) → 0 Ce^{-\sqrt{a}(1+\sqrt{c\log g})}\to0 C e − a ( 1 + c l o g g ) → 0 . The support area is at most 2 π + s sinh ( R + ℓ + 1 ) = g o ( 1 ) 2\pi+s\sinh(R+\ell+1)=g^{o(1)} 2 π + s sinh ( R + ℓ + 1 ) = g o ( 1 ) . Projecting a normalized vector off the global constant changes its norm and operator residual by at most C area ( supp u ) / area ( X ) = g − 1 / 2 + o ( 1 ) = o ( c 2 ) C\sqrt{\operatorname{area}(\operatorname{supp}u)/\operatorname{area}(X)}=g^{-1/2+o(1)}=o(c^2) C area ( supp u ) / area ( X ) = g − 1/2 + o ( 1 ) = o ( c 2 ) . For a bounded number of disjoint vectors the same estimate controls their Gram matrix. Min–max and tightness of the model counts yield
P { N g ( a ) ≥ Z g ( a + η ) } ⟶ 1 ( a , η > 0 ) . (454) \mathbb{P}\{N_g(a)\geq Z_g(a+\eta)\}\longrightarrow1\qquad(a,\eta>0). \tag*{(454)} P { N g ( a ) ≥ Z g ( a + η )} ⟶ 1 ( a , η > 0 ) . (454)
On this event ( Z g ( a ) − N g ( a ) ) + ≤ Z g ( a ) − Z g ( a + η ) (Z_g(a)-N_g(a))_+\leq Z_g(a)-Z_g(a+\eta) ( Z g ( a ) − N g ( a ) ) + ≤ Z g ( a ) − Z g ( a + η ) ; on its complement it is at most Z g ( a ) Z_g(a) Z g ( a ) . Uniform integrability makes the latter contribution vanish. Let g → ∞ g\to\infty g → ∞ , then η ↓ 0 \eta\downarrow0 η ↓ 0 , and use (K.5). This gives
E ∣ N g ( a ) − Z g ( a ) ∣ = E N g ( a ) − E Z g ( a ) + 2 E ( Z g ( a ) − N g ( a ) ) + ⟶ 0. (455) \mathbb{E}|N_g(a)-Z_g(a)|=\mathbb{E}N_g(a)-\mathbb{E}Z_g(a)+2\mathbb{E}(Z_g(a)-N_g(a))_+\longrightarrow0. \tag*{(455)} E ∣ N g ( a ) − Z g ( a ) ∣ = E N g ( a ) − E Z g ( a ) + 2 E ( Z g ( a ) − N g ( a ) ) + ⟶ 0. (455)
Integer-valued counts therefore match with probability tending to one at every finite list of thresholds. Their differences give joint interval counts; the local mean bounds give point-measure tightness. The void probability for [ x , ∞ ) [x,\infty) [ x , ∞ ) gives (231) . The transformation x = κ 0 / t x=\kappa_0/t x = κ 0 / t gives the asserted arrival-time representation. Sending a fixed lower threshold to zero shows that each fixed number of eigenvalues exists with probability tending to one.
Eigenfunctions. For eigenfunctions, first restrict model deficits to a fixed [ a , b ] [a,b] [ a , b ] . Partition it into finitely many bins, then exclude thin margins at their endpoints and bins with two model points. The limiting Poisson process is simple and atomless, so the limiting exceptional probability can be made arbitrarily small by choosing the partition and margins. Count matching at the endpoints gives one actual eigenvalue, counted with multiplicity, in each occupied bin. It is simple. A model point away from the margins has distance at least ρ c 2 \rho c^2 ρ c 2 from the complement of its bin’s spectral projection, for a fixed ρ > 0 \rho>0 ρ > 0 . The spectral theorem and (K.15) project its normalized transplanted state to this one-dimensional eigenspace with o ( 1 ) o(1) o ( 1 ) error. Refining the bins gives matching ordered energies. Send a ↓ 0 a\downarrow0 a ↓ 0 , b ↑ ∞ b\uparrow\infty b ↑ ∞ after g → ∞ g\to\infty g → ∞ . A deterministic waist cap 4 log ( 1 / c ) + log ( c log g ) 4\log(1/c)+\log(c\log g) 4 log ( 1/ c ) + log ( c log g ) eventually exceeds every fixed S a S_a S a ; together with (K.17) it gives the asserted deterministic area bound. This proves Theorem 16.1.
K.4. Residual exhaustion. We prove Corollary 16.2. The spectral subspace below the moving threshold is shown to be residual by observing it on fixed annuli in the cusps of the pants: the observation annihilates cusp forms, while outside events of small probability it is injective on that subspace. The observation argument of (226) requires a quantitative lower bound when the model waist grows; the first part of this subsection proves such a bound, (K.20), from the Hecke polygon of the funnel model [174] and the Cheng–Yau gradient estimate [46]. The second part combines it with the cutoff estimate (K.15), the count matching (K.19) and the Poisson limit of Theorem 16.1. Let u s > 0 u_s>0 u s > 0 be the normalized ground state, w = 2 cosh ( s / 4 ) w=2\cosh(s/4) w = 2 cosh ( s /4 ) and k = 1 / 4 − β ( s ) k=\sqrt{1/4-\beta(s)} k = 1/4 − β ( s ) . Only large s s s are relevant below, and we assume throughout that w ≥ 3 w\geq3 w ≥ 3 , that is s ≥ 4 arcosh ( 3 / 2 ) s\geq4\operatorname{arcosh}(3/2) s ≥ 4 arcosh ( 3/2 ) . Cut each cusp at a horocycle of length b = 1 / 4 b=1/4 b = 1/4 , and let K s K_s K s be the remaining pants core. In the first cusp write the metric as d t 2 + b 2 e − 2 t d θ 2 dt^2+b^2e^{-2t}d\theta^2 d t 2 + b 2 e − 2 t d θ 2 and set A s = { 0 ≤ t ≤ 1 } A_s=\{0\leq t\leq1\} A s = { 0 ≤ t ≤ 1 } , h s = 1 A s / area ( A s ) h_s=1_{A_s}/\sqrt{\operatorname{area}(A_s)} h s = 1 A s / area ( A s ) . We claim that for universal positive constants,
m s : = ⟨ u s , h s ⟩ ≥ C 0 k 1 + s w − C . (456) m_s := \langle u_s,h_s\rangle\ge C_0\sqrt{\frac{k}{1+s}}\,w^{-C}. \tag*{(456)} m s := ⟨ u s , h s ⟩ ≥ C 0 1 + s k w − C . (456)
The cusp observation bound. In a cusp, conjugation of a Fourier mode by b e − t / 2 \sqrt{b}e^{-t/2} b e − t /2 gives v m ′ ′ = [ k 2 + ( 2 π m / b ) 2 e 2 t ] v m v_m'' = [k^2 + (2\pi m/b)^2e^{2t}]v_m v m ′′ = [ k 2 + ( 2 π m / b ) 2 e 2 t ] v m . Together with (K.13), the same half-line comparison proves, for each of the three ends,
∥ u s ∥ L 2 ( e n d ) 2 ≤ ( 2 k ) − 1 ∫ ∂ e n d ∣ u s ∣ 2 . (457) \|u_s\|_{L^2(\mathrm{end})}^2 \le(2k)^{-1}\int_{\partial\mathrm{end}} |u_s|^2. \tag*{(457)} ∥ u s ∥ L 2 ( end ) 2 ≤ ( 2 k ) − 1 ∫ ∂ end ∣ u s ∣ 2 . (457)
The boundary circumference is included by the conjugation and Parseval.
The needed core diameter bound follows directly from the Hecke polygon [174, equations (1), (6)–(7)]. Let S ( z ) = − 1 / z S(z)=-1/z S ( z ) = − 1/ z and F w = { ∣ ℜ z ∣ ≤ w / 2 , ∣ z ∣ ≥ 1 } F_w=\{|\Re z|\le w/2,\ |z|\ge1\} F w = { ∣ℜ z ∣ ≤ w /2 , ∣ z ∣ ≥ 1 } . The index-two surface has polygon F w ∪ S F w F_w\cup SF_w F w ∪ S F w . The right funnel boundary lies on the axis
( x − w / 2 ) 2 + y 2 = w 2 / 4 − 1. (x-w/2)^2+y^2=w^2/4-1. ( x − w /2 ) 2 + y 2 = w 2 /4 − 1.
Its intersection with ∣ z ∣ = 1 |z|=1 ∣ z ∣ = 1 is at ( 2 / w , 1 − 4 / w 2 ) (2/w,\sqrt{1-4/w^2}) ( 2/ w , 1 − 4/ w 2 ) . The truncated core in F w F_w F w lies above the appropriate circle arcs, hence above 5 / 3 \sqrt{5}/3 5 /3 , and vertical motion upward remains in the core. The cusp horocycle is at Y = w / b Y=w/b Y = w / b , above those arcs; its S S S -image truncates the other cusp, disjointly. A core point can be joined vertically to height Y Y Y , horizontally to i Y iY iY , then down to i i i , with total length O ( 1 + log w ) O(1+\log w) O ( 1 + log w ) . The S S S -copy has the same bound and joins at i i i . Thus diam ( K s ) ≤ C ( 1 + log w ) \operatorname{diam}(K_s)\le C(1+\log w) diam ( K s ) ≤ C ( 1 + log w ) .
Lift u s u_s u s to H 2 \mathbb{H}^2 H 2 . The positive function U ( z , t ) = e β ( s ) t u s ( z ) U(z,t)=e^{\sqrt{\beta(s)}t}u_s(z) U ( z , t ) = e β ( s ) t u s ( z ) is harmonic on H 2 × R \mathbb{H}^2\times\mathbb{R} H 2 × R in the analytic sign convention. The Cheng–Yau gradient estimate [46] on fixed-radius balls gives ∣ ∇ log u s ∣ ≤ C |\nabla\log u_s|\le C ∣∇ log u s ∣ ≤ C , independently of injectivity radius. If p p p is on the observed horocycle and M s = sup K s u s M_s=\sup_{K_s}u_s M s = sup K s u s , the diameter bound gives M s ≤ C w C u s ( p ) M_s\le Cw^C u_s(p) M s ≤ C w C u s ( p ) . Since area ( K s ) ≤ 2 π \operatorname{area}(K_s)\le2\pi area ( K s ) ≤ 2 π , (K.21) yields
1 ≤ ( 2 π + s + 2 b 2 k ) M s 2 ≤ C ( 1 + s ) k − 1 w 2 C u s ( p ) 2 . 1\le\left(2\pi+\frac{s+2b}{2k}\right)M_s^2\le C(1+s)k^{-1}w^{2C}u_s(p)^2. 1 ≤ ( 2 π + 2 k s + 2 b ) M s 2 ≤ C ( 1 + s ) k − 1 w 2 C u s ( p ) 2 .
Every point of A s A_s A s is joined to p p p by a path of bounded length; its area is the fixed number b ( 1 − e − 1 ) b(1-e^{-1}) b ( 1 − e − 1 ) . Harnack’s inequality proves (K.20). In particular, for each fixed 0 < a < b < ∞ 0<a<b<\infty 0 < a < b < ∞ and a ≤ k 2 / c 2 ≤ b a\le k^2/c^2\le b a ≤ k 2 / c 2 ≤ b , the model asymptotic gives
m s ≥ C a , b c M . (458) m_s\ge C_{a,b}c^M. \tag*{(458)} m s ≥ C a , b c M . (458)
for some fixed M M M : here k ≍ c k\asymp c k ≍ c , w ≍ c − 1 w\asymp c^{-1} w ≍ c − 1 and s = 4 log ( 1 / c ) + O a , b ( 1 ) s=4\log(1/c)+O_{a,b}(1) s = 4 log ( 1/ c ) + O a , b ( 1 ) .
Transplanted vectors. Under (232) , put t g = log ( 1 / c ) t_g=\log(1/c) t g = log ( 1/ c ) , H g = c log g / t g → ∞ H_g=c\log g/t_g\to\infty H g = c log g / t g → ∞ , and choose R g = 3 + t g log g / c R_g=3+\sqrt{t_g\log g/c} R g = 3 + t g log g / c with transition width one. Then R g = o ( log g ) R_g=o(\log g) R g = o ( log g ) and c R g / t g → ∞ cR_g/t_g\to\infty c R g / t g → ∞ . Equation (K.15) gives an error smaller than every fixed power of c c c on a fixed rescaled deficit window. After transplanting and dividing by m s m_s m s , the vectors f s f_s f s have
⟨ f s , h s ⟩ = 1 , ∥ ( Δ X − β ( s ) ) f s ∥ 2 = o ( c 2 ) . \langle f_s,h_s\rangle=1,\qquad\|(\Delta_X-\beta(s))f_s\|_2=o(c^2). ⟨ f s , h s ⟩ = 1 , ∥ ( Δ X − β ( s )) f s ∥ 2 = o ( c 2 ) .
They retain the complete cusps and are not projected off constants. Their norms need not be bounded; the observation and residual are the quantities used below.
Injectivity on the spectral subspace. Fix a > 0 a>0 a > 0 . Exclude a model deficit above b b b , a model point in [ a , a + η ] [a,a+\eta] [ a , a + η ] , and a count Z g ( a ) > K Z_g(a)>K Z g ( a ) > K . By the Poisson limit their limiting probabilities sum to at most
κ 0 b + κ 0 ( 1 a − 1 a + η ) + P { Pois ( κ 0 / a ) > K } . \frac{\kappa_0}{b}+\kappa_0\left(\frac{1}{a}-\frac{1}{a+\eta}\right)+\mathbb{P}\{\operatorname{Pois}(\kappa_0/a)>K\}. b κ 0 + κ 0 ( a 1 − a + η 1 ) + P { Pois ( κ 0 / a ) > K } .
For fixed b , η , K b,\eta,K b , η , K , isolation and count matching fail with probability o ( 1 ) o(1) o ( 1 ) . Otherwise Z g ( a ) = N g ( a ) ≤ K Z_g(a)=N_g(a)\le K Z g ( a ) = N g ( a ) ≤ K and the selected deficits lie in [ a + η , b ] [a+\eta,b] [ a + η , b ] . Set E = 1 ( 0 , 1 / 4 − a c 2 ] ( Δ X ) L 2 ( X ) E=1_{(0,1/4-ac^2]}(\Delta_X)L^2(X) E = 1 ( 0 , 1/4 − a c 2 ] ( Δ X ) L 2 ( X ) and let P P P project onto E E E . For the transplanted vectors set F v = ∑ i v i f i Fv=\sum_i v_i f_i F v = ∑ i v i f i and T h = ( ⟨ h , h A i ⟩ ) i Th=(\langle h,h_{A_i}\rangle)_i T h = (⟨ h , h A i ⟩ ) i . Disjoint annuli give $TF=I and ∥ T ∥ ≤ 1 \lVert T\rVert\le1 ∥ T ∥ ≤ 1 . Each model energy is eventually at least 1 / 8 1/8 1/8 and has distance at least η c 2 \eta c^2 η c 2 from the spectral complement of E E E . Hence
∥ T P F − I ∥ ≤ K max i ∥ ( Δ X − β i ) f i ∥ 2 η c 2 = o ( 1 ) . \lVert TPF-I\rVert\le\sqrt{K}\max_i\frac{\lVert(\Delta_X-\beta_i)f_i\rVert_2}{\eta c^2}=o(1). ∥ T P F − I ∥ ≤ K i max η c 2 ∥( Δ X − β i ) f i ∥ 2 = o ( 1 ) .
Since dim E = N g ( a ) \dim E=N_g(a) dim E = N g ( a ) , T T T is injective on E E E . It annihilates cusp forms, whose constant Fourier coefficients vanish. The residual–cuspidal decomposition therefore makes E E E entirely residual. Let b , K → ∞ b,K\to\infty b , K → ∞ , η ↓ 0 \eta\downarrow0 η ↓ 0 after g → ∞ g\to\infty g → ∞ . The cuspidal count tends to zero in probability. Equation (K.19) and uniform integrability of Z g ( a ) Z_g(a) Z g ( a ) give uniform integrability of N g ( a ) N_g(a) N g ( a ) , so the expected cuspidal count also tends to zero. This proves all the corollary’s assertions.
Rare events and actual gap moments
This subsection proves Corollary 16.3 . Since the limiting gap law (231) has infinite mean, the moment statements require expectation-level estimates for discarded geometries rather than an exchange of limit and expectation. The trace approximation of Lemma K.1 , applied to tests whose seminorms are controlled uniformly in the threshold, gives the upper bounds; the expected counts of good pants from the second subsection of this appendix, together with the cutoff estimate (K.16), give the lower bounds for the rare-gap law and for moments of order at least one. The four parts below treat the relative trace estimate, the rare-gap law (233) , the upper moment bounds and the lower moment bounds.
The relative trace estimate. Put α g = n ( n − 1 ) / g \alpha_g=n(n-1)/g α g = n ( n − 1 ) / g . First assume only n → ∞ n\to\infty n → ∞ and n = o ( g ) n=o(\sqrt{g}) n = o ( g ) . For a fixed smooth physical-deficit test ϕ \phi ϕ vanishing near zero, Lemma K.1 , in the c c c -independent form recorded after its proof and applied to the fixed test ψ ( x ) = ϕ ( J ( x ) ) \psi(x)=\phi(J(x)) ψ ( x ) = ϕ ( J ( x )) , gives the relative estimate
E ∑ 0 < λ j < 1 / 4 ϕ ( 1 / 4 − λ j ) = α g κ 0 ∫ 0 ∞ ϕ ( h ∗ ( y ) ) d y + o ( c 2 ) . (459) \mathbb{E}\sum_{0<\lambda_j<1/4}\phi(1/4-\lambda_j)=\alpha_g\kappa_0\int_0^\infty\phi(h_*(y))\,\mathrm{d}y+o(c^2). \tag*{(459)} E 0 < λ j < 1/4 ∑ ϕ ( 1/4 − λ j ) = α g κ 0 ∫ 0 ∞ ϕ ( h ∗ ( y )) d y + o ( c 2 ) . (459)
Indeed, retain the same Q = g σ Q=g^\sigma Q = g σ . After division by c 2 c^2 c 2 , the two lower cusp powers cost O ( 1 / n ) O(1/n) O ( 1/ n ) , the cubic remainder O ( ( n + 1 ) Q 3 D 0 / g ) O((n+1)Q^3D_0/g) O (( n + 1 ) Q 3 D 0 / g ) , and the terminal remainder O ( 2 − Q ) O(2^{-Q}) O ( 2 − Q ) . The uniform approximation error is at most O ( g 2 Q − K ) O(g^{2Q-K}) O ( g 2 Q − K ) ; choose σ K > 2 \sigma K>2 σ K > 2 . The fixed-order distributional error tends to zero. The common-truncation and retained-mass errors are exponentially small in Q Q Q . A nonnegative majorant first bounds the weighted mean by O ( c 2 ) O(c^2) O ( c 2 ) , so removing weights preserves the relative estimate. Smooth squeezing gives E N X p h y s ( t ) / c 2 → κ 0 H ( t ) \mathbb{E}N_X^{\mathrm{phys}}(t)/c^2\to\kappa_0H(t) E N X phys ( t ) / c 2 → κ 0 H ( t ) .
The rare-gap law. For a fixed physical threshold t t t , the waists have a fixed bound. Let Z X p h y s ( t ) Z_X^{\mathrm{phys}}(t) Z X phys ( t ) be the number of two-cusp pairs of pants of X X X whose model deficit 1 / 4 − β ( s ) 1/4-\beta(s) 1/4 − β ( s ) is at least t t t , that is, whose waist s s s satisfies 2 ( cosh ( s / 2 ) − 1 ) ≤ H ( t ) 2(\cosh(s/2)-1)\le H(t) 2 ( cosh ( s /2 ) − 1 ) ≤ H ( t ) . The one-pants integral gives E Z X p h y s ( t ) = α g κ 0 H ( t ) ( 1 + o ( 1 ) ) \mathbb{E}Z_X^{\mathrm{phys}}(t)=\alpha_g\kappa_0H(t)(1+o(1)) E Z X phys ( t ) = α g κ 0 H ( t ) ( 1 + o ( 1 )) ; the disjoint-pair integral is O ( c 4 ) O(c^4) O ( c 4 ) . At every fixed radius, (K.11) is uniform for N ≤ g N\le\sqrt{g} N ≤ g , so expected discarded counts are O ( c 2 ) O(c^2) O ( c 2 ) and good-pair counts are O ( c 4 ) O(c^4) O ( c 4 ) . Fix η > 0 \eta>0 η > 0 with t + η < 1 / 4 t+\eta<1/4 t + η < 1/4 and choose the radius so large that (K.16) is less than η / 2 \eta/2 η /2 for every model of deficit at least t + η t+\eta t + η . The constant-projection error is O ( 1 / g ) O(1/g) O ( 1/ g ) . Every good such model therefore implies D ( X ) ≥ t D(X)\ge t D ( X ) ≥ t for large g g g . For its good count Z Z Z ,
P ( D ≥ t ) ≥ P ( Z ≥ 1 ) ≥ E Z − 1 2 E Z ( Z − 1 ) . \mathbb{P}(D\ge t)\ge\mathbb{P}(Z\ge1)\ge\mathbb{E}Z-\frac{1}{2}\mathbb{E}Z(Z-1). P ( D ≥ t ) ≥ P ( Z ≥ 1 ) ≥ E Z − 2 1 E Z ( Z − 1 ) .
Divide by c 2 c^2 c 2 , take the lower limit and let η ↓ 0 \eta\downarrow0 η ↓ 0 . Together with P ( D ≥ t ) ≤ E N X p h y s ( t ) \mathbb{P}(D\ge t)\le\mathbb{E}N_X^{\mathrm{phys}}(t) P ( D ≥ t ) ≤ E N X phys ( t ) , this proves (233) . The identity E ( N − 1 ) + = E N − P ( N ≥ 1 ) = o ( c 2 ) \mathbb{E}(N-1)_+=\mathbb{E}N-\mathbb{P}(N\ge1)=o(c^2) E ( N − 1 ) + = E N − P ( N ≥ 1 ) = o ( c 2 ) proves the conditional single-mode assertion.
Upper bounds for the moments. For the moments, assume (229) . Fix a smooth 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 , zero below 1 / 2 1/2 1/2 and one above 1 1 1 . For fixed p ≥ 0 p\ge0 p ≥ 0 and c 2 ≤ b ≤ 1 / 4 c^2\le b\le1/4 c 2 ≤ b ≤ 1/4 , the tests d p χ ( d / b ) \mathrm{d}^p\chi(\mathrm{d}/b) d p χ ( d / b ) , extended by zero near d = 0 \mathrm{d}=0 d = 0 , have r r r th derivative norm at most C p , r c − r C_{p,r}c^{-r} C p , r c − r . This includes noninteger p p p . Lemma K.1 , that is, (K.3), gives, uniformly in b b b ,
E ∑ 0 < λ j < 1 / 4 ( 1 / 4 − λ j ) p χ ( ( 1 / 4 − λ j ) / b ) = α g κ 0 ∫ h ∗ ( y ) p χ ( h ∗ ( y ) / b ) d y + O ( r g ) . (460) \mathbb{E}\sum_{0<\lambda_j<1/4}(1/4-\lambda_j)^p\chi((1/4-\lambda_j)/b)
=\alpha_g\kappa_0\int h_*(y)^p\chi(h_*(y)/b)\,\mathrm{d}y+O(r_g). \tag*{(460)} E 0 < λ j < 1/4 ∑ ( 1/4 − λ j ) p χ (( 1/4 − λ j ) / b ) = α g κ 0 ∫ h ∗ ( y ) p χ ( h ∗ ( y ) / b ) d y + O ( r g ) . (460)
where r g = g − ρ + o ( 1 ) r_g=g^{-\rho+o(1)} r g = g − ρ + o ( 1 ) for some ρ > 0 \rho>0 ρ > 0 depending on the fixed approximation choices. The negative polynomial exponents include − 1 / 2 + 3 D 0 σ -1/2+3D_0\sigma − 1/2 + 3 D 0 σ and − 1 / 2 -1/2 − 1/2 ; all test seminorms are subpowers of g g g . Weight removal costs at most an exponential in − c Q -cQ − c Q times a bounded weighted mean. Since c − 1 = o ( log g ) c^{-1}=o(\log g) c − 1 = o ( log g ) , r g = o ( c A ) r_g=o(c^A) r g = o ( c A ) for every fixed A > 0 A>0 A > 0 . In particular the case p = 0 p=0 p = 0 yields
P ( D ≥ b ) ≤ E N X p h y s ( b ) ≤ C c 2 / b + O ( r g ) , c 2 ≤ b ≤ 1 / 4. (461) \mathbb{P}(D\ge b)\le\mathbb{E}N_X^{\mathrm{phys}}(b)\le Cc^2/b+O(r_g),\qquad c^2\le b\le1/4. \tag*{(461)} P ( D ≥ b ) ≤ E N X phys ( b ) ≤ C c 2 / b + O ( r g ) , c 2 ≤ b ≤ 1/4. (461)
For 0 < p < 1 0<p<1 0 < p < 1 , (K.25) and layer cake imply
lim sup g E [ ( D / c 2 ) p 1 { D / c 2 > K } ] ≤ C p K p − 1 . \limsup_{g}\mathbb{E}\left[(D/c^2)^p\mathbf{1}_{\{D/c^2>K\}}\right]\le C_pK^{p-1}. g lim sup E [ ( D / c 2 ) p 1 { D / c 2 > K } ] ≤ C p K p − 1 .
The additive error contributes at most C p r g c − 2 p → 0 C_pr_gc^{-2p}\to0 C p r g c − 2 p → 0 . Uniform integrability and (231) now give (234) , since Z = κ 0 / E Z=\kappa_0/E Z = κ 0 / E for a rate-one exponential E E E .
For p ≥ 1 p\ge1 p ≥ 1 , the pointwise inequality
D p ≤ c 2 p + ∑ 0 < λ j < 1 / 4 ( 1 / 4 − λ j ) p χ ( ( 1 / 4 − λ j ) / c 2 ) D^p\le c^{2p}+\sum_{0<\lambda_j<1/4}(1/4-\lambda_j)^p\chi\left((1/4-\lambda_j)/c^2\right) D p ≤ c 2 p + 0 < λ j < 1/4 ∑ ( 1/4 − λ j ) p χ ( ( 1/4 − λ j ) / c 2 )
gives upper bounds. At p = 1 p=1 p = 1 , integrability of h ∗ − ( 1 + y ) − 1 h_*- (1+y)^{-1} h ∗ − ( 1 + y ) − 1 and H ( b ) ∼ 1 / b H(b)\sim1/b H ( b ) ∼ 1/ b show
∫ h ∗ ( y ) χ ( h ∗ ( y ) / b ) d y = log ( 1 / b ) + O ( 1 ) . (462) \int h_*(y)\chi(h_*(y)/b)\,dy=\log(1/b)+O(1). \tag*{(462)} ∫ h ∗ ( y ) χ ( h ∗ ( y ) / b ) d y = log ( 1/ b ) + O ( 1 ) . (462)
The transition region lies between H ( b ) H(b) H ( b ) and H ( b / 2 ) H(b/2) H ( b /2 ) and costs O ( 1 ) O(1) O ( 1 ) . Replacing α g \alpha_g α g by c 2 c^2 c 2 costs O ( c 2 log ( 1 / c ) / n ) = o ( c 2 ) O(c^2\log(1/c)/n)=o(c^2) O ( c 2 log ( 1/ c ) / n ) = o ( c 2 ) . This proves the upper mean bound. For p > 1 p>1 p > 1 , h ∗ p h_*^p h ∗ p is integrable and dominated convergence in (K.24) proves the required upper limit.
Lower bounds for the moments. The lower bounds require expected discarded counts, rather than an exceptional event of unspecified probability. Put b = c 2 b=c^2 b = c 2 and use good pants relative to all waists of model deficit at least b b b , with the isolation radius R ′ = R + ℓ + 1 R'=R+\ell+1 R ′ = R + ℓ + 1 of (K.6) built from the common radius (K.17). For every t ≥ b t\ge b t ≥ b let Z g o o d ( t ) Z^{\mathrm{good}}(t) Z good ( t ) count these good pants of deficit at least t t t , and let D g g o o d D_g^{\mathrm{good}} D g good be their largest deficit, or zero if there are none. Uniformly for b ≤ t < 1 / 4 b\le t<1/4 b ≤ t < 1/4 ,
E Z X p h y s ( t ) = α g κ 0 H ( t ) ( 1 + O ( v g ) ) , \mathbb{E}Z_X^{\mathrm{phys}}(t)=\alpha_g\kappa_0H(t)(1+O(v_g)), E Z X phys ( t ) = α g κ 0 H ( t ) ( 1 + O ( v g )) ,
E ( Z X p h y s ( t ) − Z g o o d ( t ) ) ≤ ϵ g , 1 E Z X p h y s ( t ) , \mathbb{E}(Z_X^{\mathrm{phys}}(t)-Z^{\mathrm{good}}(t))\le\epsilon_{g,1}\mathbb{E}Z_X^{\mathrm{phys}}(t), E ( Z X phys ( t ) − Z good ( t )) ≤ ϵ g , 1 E Z X phys ( t ) ,
E [ Z g o o d ( t ) ( Z g o o d ( t ) − 1 ) ] ≤ C c 4 / t 2 . \mathbb{E}[Z^{\mathrm{good}}(t)(Z^{\mathrm{good}}(t)-1)]\le Cc^4/t^2. E [ Z good ( t ) ( Z good ( t ) − 1 )] ≤ C c 4 / t 2 .
with v g , ϵ g , 1 ≤ g − 1 / 2 + o ( 1 ) v_g,\epsilon_{g,1}\le g^{-1/2+o(1)} v g , ϵ g , 1 ≤ g − 1/2 + o ( 1 ) . These are respectively (K.8), (K.11), and the disjoint-pair integral; S b = O ( log log g ) S_b=O(\log\log g) S b = O ( log log g ) and the radius is o ( log g ) o(\log g) o ( log g ) . Bonferroni and layer cake give
E ( D g g o o d ) p ≥ ( 1 − O ( v g + ϵ g , 1 ) ) α g κ 0 ∫ b 1 / 4 p t p − 1 H ( t ) d t − C p c 4 ∫ b 1 / 4 t p − 3 d t . (463) \mathbb{E}(D_g^{\mathrm{good}})^p\ge(1-O(v_g+\epsilon_{g,1}))\alpha_g\kappa_0\int_b^{1/4}pt^{p-1}H(t)\,dt-C_pc^4\int_b^{1/4}t^{p-3}\,dt. \tag*{(463)} E ( D g good ) p ≥ ( 1 − O ( v g + ϵ g , 1 )) α g κ 0 ∫ b 1/4 p t p − 1 H ( t ) d t − C p c 4 ∫ b 1/4 t p − 3 d t . (463)
The main integral equals ∫ h ∗ ≥ b ( h ∗ p − b p ) d y \int_{h_*\ge b}(h_*^p-b^p)\,dy ∫ h ∗ ≥ b ( h ∗ p − b p ) d y . For p = 1 p=1 p = 1 it is log ( 1 / b ) + O ( 1 ) \log(1/b)+O(1) log ( 1/ b ) + O ( 1 ) , the pair error is O ( c 2 ) O(c^2) O ( c 2 ) , and the relative geometric error contributes o ( c 2 ) o(c^2) o ( c 2 ) . For p > 1 p>1 p > 1 the main integral converges to ∫ h ∗ p \int h_*^p ∫ h ∗ p ; the pair errors are respectively O ( c 2 p ) O(c^{2p}) O ( c 2 p ) , O ( c 4 log ( 1 / c ) ) O(c^4\log(1/c)) O ( c 4 log ( 1/ c )) , and O ( c 4 ) O(c^4) O ( c 4 ) for p < 2 p<2 p < 2 , p = 2 p=2 p = 2 , and p > 2 p>2 p > 2 , all o ( c 2 ) o(c^2) o ( c 2 ) .
Finally (K.16) and the area projection bound give a deterministic e g = o ( c 2 ) e_g=o(c^2) e g = o ( c 2 ) such that, on every surface, D ( X ) ≥ ( D g g o o d − e g ) + D(X)\ge(D_g^{\mathrm{good}}-e_g)_+ D ( X ) ≥ ( D g good − e g ) + . If a good pair of pants exists, apply its trial vector; otherwise the assertion is trivial. Since x p x^p x p is Lipschitz on [ 0 , 1 / 4 ] [0,1/4] [ 0 , 1/4 ] for p ≥ 1 p\ge1 p ≥ 1 , this transfers the lower moment bounds to without a probability-to-expectation substitution. Equations (235) –(236) follow. The variance uses ( E D ) 2 = o ( c 2 ) (\mathbb{E}D)^2=o(c^2) ( E D ) 2 = o ( c 2 ) ; tightness of D / c 2 D/c^2 D / c 2 and (235) give D / E D → 0 D/\mathbb{E}D\to0 D / E D → 0 in probability, completing Corollary 16.3 .
Geometric exclusion in the supercritical regime
This appendix proves the two geometric lemmas used in the proof of Theorem 17.1 in § 4.13 . The first, Lemma L.1, shows that, apart from single planar separating curves, cuts into two connected pieces with small ratio of length to smaller area are rare; its bound B i , j , r ( v ) B_{i,j,r}(v) B i , j , r ( v ) for two positive-genus sides is reused in Appendix Q. The second, Lemma L.2, combines this with first-moment bounds for short curves: with probability tending to one, the short curves are planar separating, their cut tree has branches of controlled depth and area, and every cut piece has an explicit lower bound on its Neumann gap. The volume inputs are Mirzakhani integration with the boundary estimate (188) , the coefficient-ratio bound (G.1) and identity (G.2) from Appendix § 10.1 , the point-mass and local-limit estimates of § 3.25 , and the all-marking envelope (D.35); the Neumann gaps come from Lemma M.1 of Appendix M. Throughout, n / g → ∞ n/\sqrt{g}\to\infty n / g → ∞ , n ≤ M g n\le Mg n ≤ M g , P = 2 π 2 P=2\pi^{2} P = 2 π 2 , ( 2 g − 1 ) Ψ ( x ) = n − 1 (2g-1)\Psi(x)=n-1 ( 2 g − 1 ) Ψ ( x ) = n − 1 and Q = n x / ( 4 P ) Q=nx/(4P) Q = n x / ( 4 P ) . Thus x ≤ x ˉ ( M ) < x c x\le\bar{x}(M)<x_c x ≤ x ˉ ( M ) < x c and n / x ≤ C M g n/x\le C_Mg n / x ≤ C M g . Constants may depend on M M M .
Lemma L.1 (Competing connected cuts). For each M > 0 M>0 M > 0 there are u 0 = u 0 ( M ) > 0 u_0=u_0(M)>0 u 0 = u 0 ( M ) > 0 , C M C_M C M and integers g 0 = g 0 ( M ) g_0=g_0(M) g 0 = g 0 ( M ) , n 0 = n 0 ( M ) ≥ 2 n_0=n_0(M)\ge2 n 0 = n 0 ( M ) ≥ 2 such that, for g ≥ g 0 g\ge g_0 g ≥ g 0 , n 0 ≤ n ≤ M g n_0\le n\le Mg n 0 ≤ n ≤ M g and 0 < u ≤ u 0 0<u\le u_0 0 < u ≤ u 0 , the P g , n W P \mathbb{P}^{\mathrm{WP}}_{g,n} P g , n WP -probability that X X X carries a geodesic multicurve which divides X X X into two connected components, has length at most u u u times the smaller of the two areas, and is not a single planar separating curve, is at most C M ( u 4 + u 2 / g ) C_M(u^4+u^2/g) C M ( u 4 + u 2 / g ) . The number of components of the multicurve and the distribution of the cusps between the two sides are arbitrary.
Proof. The lower thresholds on g g g and n n n ensure the central coefficient bound used below. Indeed, Proposition 9.2 applies along every sequence with h → ∞ h\to\infty h → ∞ and μ → ∞ \mu\to\infty μ → ∞ ; at the integer saddle h Ψ ( x ) = m h\Psi(x)=m h Ψ ( x ) = m , it therefore gives the central lower bound uniformly once both h h h and m m m exceed fixed thresholds. Otherwise a sequence violating that bound with h , m → ∞ h,m\to\infty h , m → ∞ would contradict the proposition. In the planar terms below with r ≤ g / 2 r\le g/2 r ≤ g /2 , we have G ≥ g / 2 G\ge g/2 G ≥ g /2 and N = n + 2 r − 2 ≥ n N=n+2r-2\ge n N = n + 2 r − 2 ≥ n , so the same thresholds apply simultaneously to h = 2 G − 1 h=2G-1 h = 2 G − 1 and m = N − 1 m=N-1 m = N − 1 .
We use Mirzakhani integration with the boundary bound (188) . For r r r unlabeled cut curves, the unordered count equals 1 / r ! 1/r! 1/ r ! times the labeled-boundary integral, in which the r r r curves are ordered, and ∫ ∑ ℓ i ≤ L ∏ ℓ i d ℓ i = L 2 r / ( 2 r ) ! \int_{\sum\ell_i\le L}\prod\ell_i\,d\ell_i=L^{2r}/(2r)! ∫ ∑ ℓ i ≤ L ∏ ℓ i d ℓ i = L 2 r / ( 2 r )! ; see also [172, equations (30)–(31)].
A planar side with r ≥ 2 r\ge2 r ≥ 2 boundaries. If this side has a a a cusps, put χ = a + r − 2 > 0 \chi=a+r-2>0 χ = a + r − 2 > 0 , minus its Euler characteristic, and G = g − r + 1 G=g-r+1 G = g − r + 1 . The planar side has area 2 π χ 2\pi\chi 2 π χ , so a cut of ratio at most u u u has total boundary length at most u u u times the smaller area, hence at most 2 π u χ 2\pi u\chi 2 π u χ , whether or not the planar side is the smaller side. Its expected count with total boundary length at most 2 π u χ 2\pi u\chi 2 π u χ is bounded by
( n a ) V 0 , a + r V G , n − a + r V g , n e 2 π u χ ( 2 π u χ ) 2 r r ! ( 2 r ) ! . (464) \binom{n}{a}\frac{V_{0,a+r}V_{G,n-a+r}}{V_{g,n}}e^{2\pi u\chi}\frac{(2\pi u\chi)^{2r}}{r!(2r)!}. \tag*{(464)} ( a n ) V g , n V 0 , a + r V G , n − a + r e 2 π u χ r ! ( 2 r )! ( 2 π u χ ) 2 r . (464)
For r ≤ g / 2 r\le g/2 r ≤ g /2 , set N = n + 2 r − 2 N=n+2r-2 N = n + 2 r − 2 and choose ( 2 G − 1 ) Ψ ( x r ) = N − 1 (2G-1)\Psi(x_r)=N-1 ( 2 G − 1 ) Ψ ( x r ) = N − 1 . Then x r ≤ x ∗ ( M ) < x c x_r\le x_*(M)<x_c x r ≤ x ∗ ( M ) < x c . Successive genus exchanges and Theorem 1.3 give V G , N / V g , n ≤ C r − 1 V_{G,N}/V_{g,n}\le C^{r-1} V G , N / V g , n ≤ C r − 1 . By (124) and (G.1),
( n a ) V G , n − a + r V G , N ≤ C a ! ( x r P ) a + r − 2 ( n − a + 1 ) r − 1 ( n + 1 ) 2 r − 3 . \binom{n}{a}\frac{V_{G,n-a+r}}{V_{G,N}}\le\frac{C}{a!}\left(\frac{x_r}{P}\right)^{a+r-2}\frac{(n-a+1)^{r-1}}{(n+1)^{2r-3}}. ( a n ) V G , N V G , n − a + r ≤ a ! C ( P x r ) a + r − 2 ( n + 1 ) 2 r − 3 ( n − a + 1 ) r − 1 .
where the bars denote rising factorials. Pairing numerator factors with the first r − 1 r-1 r − 1 denominator factors bounds their quotient by ( n + r ) − ( r − 2 ) (n+r)^{-(r-2)} ( n + r ) − ( r − 2 ) . Let
F r ( z ) = Φ 0 ( r ) ( z ) = ∑ a ≥ max ( 0 , 3 − r ) V 0 , a + r z a a ! P a + r − 3 . F_r(z)=\Phi_0^{(r)}(z)=\sum_{a\ge\max(0,3-r)}\frac{V_{0,a+r}z^a}{a!P^{a+r-3}}. F r ( z ) = Φ 0 ( r ) ( z ) = a ≥ m a x ( 0 , 3 − r ) ∑ a ! P a + r − 3 V 0 , a + r z a .
The volume prefactor in (L.1) is therefore bounded by [ z a ] F r ( z ) x r a [z^a]F_r(z)x_r^a [ z a ] F r ( z ) x r a times C r P − 1 ( x r / ( n + r ) ) r − 2 C^rP^{-1}(x_r/(n+r))^{r-2} C r P − 1 ( x r / ( n + r ) ) r − 2 . Choose fixed s , u 0 > 0 s,u_0>0 s , u 0 > 0 and z 1 < x c z_1<x_c z 1 < x c with x ∗ e 2 π u 0 + s < z 1 x_*e^{2\pi u_0+s}<z_1 x ∗ e 2 π u 0 + s < z 1 . The bound χ 2 r ≤ ( 2 r ) ! s − 2 r e s χ \chi^{2r}\le(2r)!s^{-2r}e^{s\chi} χ 2 r ≤ ( 2 r )! s − 2 r e sχ sums the cusp factor into e ( 2 π u + s ) ( r − 2 ) F r ( x r e 2 π u + s ) e^{(2\pi u+s)(r-2)}F_r(x_re^{2\pi u+s}) e ( 2 π u + s ) ( r − 2 ) F r ( x r e 2 π u + s ) . Cauchy’s estimate on a fixed neighborhood of [ 0 , z 1 ] [0,z_1] [ 0 , z 1 ] gives F r ≤ C r ! d − r F_r \le Cr!d^{-r} F r ≤ C r ! d − r . The factorials cancel those in the integral, leaving
E { such cuts with r boundaries } ≤ C ( C M u 2 ) r ( n + r ) r − 2 . \mathbb{E}\{\text{such cuts with }r\text{ boundaries}\}\le C\frac{(C_Mu^2)^r}{(n+r)^{r-2}}. E { such cuts with r boundaries } ≤ C ( n + r ) r − 2 ( C M u 2 ) r .
Summing 2 ≤ r ≤ g / 2 2\le r\le g/2 2 ≤ r ≤ g /2 gives O M ( u 4 ) O_M(u^4) O M ( u 4 ) . For r > g / 2 r>g/2 r > g /2 , the total boundary length is at most u u u times half the area, π u ( 2 g − 2 + n ) < π ( M + 2 ) u g \pi u(2g-2+n)<\pi(M+2)ug π u ( 2 g − 2 + n ) < π ( M + 2 ) ug , so the average boundary length is at most 2 π ( M + 2 ) u 2\pi(M+2)u 2 π ( M + 2 ) u . At least r / 2 > g / 4 r/2>g/4 r /2 > g /4 boundaries have length at most 4 π ( M + 2 ) u 4\pi(M+2)u 4 π ( M + 2 ) u . Each is nonseparating in X X X , since the other r − 1 r-1 r − 1 curves still join the two sides. Integration and the bounded genus ratio give E N n o n s e p ( L ) ≤ C e L L 2 \mathbb{E}N_{\mathrm{nonsep}}(L)\le C e^L L^2 E N nonsep ( L ) ≤ C e L L 2 . Markov’s inequality bounds this event by C M u 2 / g C_Mu^2/g C M u 2 / g .
Two positive-genus sides. Write their genera as i , j ≥ 1 i,j\ge1 i , j ≥ 1 , with i + j = g − r + 1 i+j=g-r+1 i + j = g − r + 1 , and set h 1 = 2 i − 2 + r ≤ h 2 = 2 j − 2 + r h_1=2i-2+r\le h_2=2j-2+r h 1 = 2 i − 2 + r ≤ h 2 = 2 j − 2 + r , H = h 1 + h 2 = 2 g − 2 H=h_1+h_2=2g-2 H = h 1 + h 2 = 2 g − 2 . Even if the first side is not smaller in area, a cut of ratio at most u u u has length at most 2 π u ( h 1 + a ) 2\pi u(h_1+a) 2 π u ( h 1 + a ) , where a a a is its cusp count. The same integral and the exponential moment bound reduce its expected count to
( 2 π u / s ) 2 r r ! e v h 1 B i , j , r ( v ) , B i , j , r ( v ) = ∑ a ( n a ) V i , a + r V j , n − a + r V g , n e v a , v = 2 π u + s . \frac{(2\pi u/s)^{2r}}{r!}e^{vh_1}B_{i,j,r}(v),\qquad
B_{i,j,r}(v)=\sum_a\binom{n}{a}\frac{V_{i,a+r}V_{j,n-a+r}}{V_{g,n}}e^{va},\qquad v=2\pi u+s. r ! ( 2 π u / s ) 2 r e v h 1 B i , j , r ( v ) , B i , j , r ( v ) = a ∑ ( a n ) V g , n V i , a + r V j , n − a + r e v a , v = 2 π u + s .
Put F i = Φ i ( r ) F_i=\Phi_i^{(r)} F i = Φ i ( r ) , F j = Φ j ( r ) F_j=\Phi_j^{(r)} F j = Φ j ( r ) and F g = Φ g ′ F_g=\Phi'_g F g = Φ g ′ . The exact coefficient dictionary gives
B i , j , r ( v ) = n P r [ z n ] F i ( e v z ) F j ( z ) [ z n − 1 ] F g ( z ) ≤ C n P r x F i ( x e v ) F j ( x ) F g ( x ) . B_{i,j,r}(v)=\frac{n}{P^r}\frac{[z^n]F_i(e^vz)F_j(z)}{[z^{n-1}]F_g(z)}
\le\frac{C n}{Pr x}\frac{F_i(xe^v)F_j(x)}{F_g(x)}. B i , j , r ( v ) = P r n [ z n − 1 ] F g ( z ) [ z n ] F i ( e v z ) F j ( z ) ≤ P r x C n F g ( x ) F i ( x e v ) F j ( x ) .
For the inequality use the positive convolution factor of F j F_j F j , whose power h 2 ≥ g − 1 h_2\ge g-1 h 2 ≥ g − 1 gives a maximal point mass C / σ C/\sigma C / σ by Lemma 9.3 ; convolution with the tilted first component cannot increase it. The parent’s point mass at n − 1 n-1 n − 1 is at least c / σ c/\sigma c / σ by Proposition 9.2. Only the unshifted large component uses the point-mass lemma, so the smaller radius of F i ( e v z ) F_i(e^vz) F i ( e v z ) causes no difficulty.
Write ρ ( z ) = r ( Y ( z ) ) \rho(z)=r(Y(z)) ρ ( z ) = r ( Y ( z )) and S ( z ) = a ( Y ( z ) ) ρ ( z ) 3 / 2 A ( Y ( z ) ) S(z)=a(Y(z))\rho(z)^{3/2}A(Y(z)) S ( z ) = a ( Y ( z )) ρ ( z ) 3/2 A ( Y ( z )) . The actual all-marking envelope (D.35), with (64) , gives for every k , p ≥ 1 k,p\ge1 k , p ≥ 1 , h = 2 k − 2 + p h=2k-2+p h = 2 k − 2 + p ,
Φ k ( p ) ( z ) ≤ C 4 − k Γ ( h ) h S ( z ) − h ρ ( z ) p / 2 . \Phi_k^{(p)}(z)\le C4^{-k}\frac{\Gamma(h)}{\sqrt{h}}S(z)^{-h}\rho(z)^{p/2}. Φ k ( p ) ( z ) ≤ C 4 − k h Γ ( h ) S ( z ) − h ρ ( z ) p /2 .
The parent has a matching lower bound at p = 1 p=1 p = 1 by Proposition 7.1. Bounded logarithmic derivatives on a compact interval containing x x x and x e v xe^v x e v imply
e v h 1 F i ( x e v ) F j ( x ) F g ( x ) ≤ C M r e C M v h 1 Γ ( h 1 ) Γ ( h 2 ) Γ ( H + 1 ) H + 1 h 1 h 2 . e^{vh_1}\frac{F_i(xe^v)F_j(x)}{F_g(x)}
\le C_M^r e^{C_Mvh_1}\frac{\Gamma(h_1)\Gamma(h_2)}{\Gamma(H+1)}
\sqrt{\frac{H+1}{h_1h_2}}. e v h 1 F g ( x ) F i ( x e v ) F j ( x ) ≤ C M r e C M v h 1 Γ ( H + 1 ) Γ ( h 1 ) Γ ( h 2 ) h 1 h 2 H + 1 .
The powers of 4 4 4 leave 4 r − 1 4^{r-1} 4 r − 1 and those of S S S leave S ( x ) ( S ( x ) / S ( x e v ) ) h 1 S(x)(S(x)/S(xe^v))^{h_1} S ( x ) ( S ( x ) / S ( x e v ) ) h 1 : the tilt costs h 1 h_1 h 1 , not H H H . For h 1 ≤ H / 2 h_1\le H/2 h 1 ≤ H /2 ,
Γ ( h 1 ) Γ ( h 2 ) Γ ( H ) = 1 ( H − 1 ) ( H − 2 h 1 − 1 ) ≤ 2 − ( h 1 − 1 ) H − 1 . \frac{\Gamma(h_1)\Gamma(h_2)}{\Gamma(H)}
=\frac{1}{(H-1)\binom{H-2}{h_1-1}}
\le\frac{2^{-(h_1-1)}}{H-1}. Γ ( H ) Γ ( h 1 ) Γ ( h 2 ) = ( H − 1 ) ( h 1 − 1 H − 2 ) 1 ≤ H − 1 2 − ( h 1 − 1 ) .
Choose s , u 0 s,u_0 s , u 0 so that C M ( s + 2 π u 0 ) < ( log 2 ) / 2 C_M(s+2\pi u_0)<(\log2)/2 C M ( s + 2 π u 0 ) < ( log 2 ) /2 . Using n / x ≤ C M g n/x\le C_Mg n / x ≤ C M g yields e v h 1 B i , j , r ( v ) ≤ C M r g − 1 e − c h 1 / h 1 e^{vh_1}B_{i,j,r}(v)\le C_M^r g^{-1}e^{-ch_1}/\sqrt{h_1} e v h 1 B i , j , r ( v ) ≤ C M r g − 1 e − c h 1 / h 1 . Summing first the admissible genera and then all r ≥ 1 r\ge1 r ≥ 1 gives
C g − 1 ∑ r ≥ 1 ( C M u 2 ) r r ! = O M ( u 2 / g ) . C g^{-1}\sum_{r\ge1}\frac{(C_Mu^2)^r}{r!}=O_M(u^2/g). C g − 1 r ≥ 1 ∑ r ! ( C M u 2 ) r = O M ( u 2 / g ) .
These two cases exhaust the non-single-planar connected cuts. □
Lemma L.2 (Short-cap geometry). With the cutoffs (240) , with probability tending to one all δ \delta δ -short curves are planar separating; their cut tree has root-branch depth at most K K K and branch area at most 2 π K 2\pi K 2 π K . Every cut piece has Neumann gap at least (241) . For any growing R R R with R δ → 0 R\delta\to0 R δ → 0 , every δ \delta δ -short curve satisfying Q ℓ / ( k − 1 ) ≤ R \sqrt{Q\ell}/(k-1)\le R Q ℓ / ( k − 1 ) ≤ R is an isolated root leaf with probability tending to
one. Moreover the geodesic Cheeger minimum, on bounded Q − 1 / 2 Q^{-1/2} Q − 1/2 windows, is attained by single planar waists of bounded cusp type up to a tail tending to zero.
Proof. The proof has four stages: planarity of the short curves together with the depth and area bounds; isolation of the retained leaves; the Neumann gaps of the cut pieces; and the Cheeger minimum.
Planarity, depth and area. The nonplanar first-moment bounds in the proof of Theorem 14.2 give exceptional probability O ( δ 2 ) O(\delta^2) O ( δ 2 ) . The collar lemma makes the remaining curves disjoint. Their planar sides are disjoint or nested, so the dual graph is a tree. By (G.1)–(G.2) and (188) , for every k k k , L L L ,
E N k , L ≤ C Q L 2 w k ( x ) e L (465) \mathbb{E}N_{k,L} \le C Q L^2 w_k(x)e^L \tag*{(465)} E N k , L ≤ C Q L 2 w k ( x ) e L (465)
On a compact cusp-weight interval, the w k w_k w k have a uniform exponential tail. Hence the probability of a δ \delta δ -short cap with more than K = C M log ( Q + 2 ) K=C_M\log(Q+2) K = C M log ( Q + 2 ) cusps is at most C Q δ 2 e − c M K = o ( 1 ) C Q \delta^2 e^{-c_M K}=o(1) C Q δ 2 e − c M K = o ( 1 ) . A maximal cap with k k k cusps has total Euler complexity k − 1 k-1 k − 1 ; each planar vertex has positive integer complexity. This bounds its number of vertices, its depth and its area as asserted.
Isolation. Consider a nested pair with a ≥ 1 a\ge1 a ≥ 1 cusps between the curves and b ≥ 2 b\ge2 b ≥ 2 inside. The coefficient dictionary bounds its zero-boundary prefactor by
C n P 2 x V 0 , a + 2 x a a ! P a − 1 V 0 , b + 1 x b b ! P b − 2 . \frac{C n}{P^2 x}\frac{V_{0,a+2}x^a}{a!P^{a-1}}\frac{V_{0,b+1}x^b}{b!P^{b-2}}. P 2 x C n a ! P a − 1 V 0 , a + 2 x a b ! P b − 2 V 0 , b + 1 x b .
The root coefficient deficit is a + b − 1 a+b-1 a + b − 1 and the remaining factorial factor is exactly n n n , including a + b = n a+b=n a + b = n . If either curve has score at most R R R , its squared length integral is bounded by R 2 ( b − 1 ) 2 / ( 2 Q ) R^2(b-1)^2/(2Q) R 2 ( b − 1 ) 2 / ( 2 Q ) or R 2 ( a + b − 1 ) 2 / ( 2 Q ) R^2(a+b-1)^2/(2Q) R 2 ( a + b − 1 ) 2 / ( 2 Q ) ; the other integral is at most δ 2 / 2 \delta^2/2 δ 2 /2 . Actual boundary lengths remain bounded by δ \delta δ , so their volume error is at most e 2 δ e^{2\delta} e 2 δ . Writing F = Φ 0 ′ F=\Phi_0' F = Φ 0 ′ , Y = Φ 0 ′ ′ Y=\Phi_0'' Y = Φ 0 ′′ and D = x ∂ x D=x\partial_x D = x ∂ x , the total is
C n R 2 δ 2 P 2 x Q { Y ( D − 1 ) 2 F + ( D − 1 ) 2 ( Y F ) } ≤ C M Q n R 2 δ 2 = o ( 1 ) . \frac{C n R^2\delta^2}{P^2 x Q}\left\{Y(D-1)^2F+(D-1)^2(YF)\right\}\le C_M\frac{Q}{n}R^2\delta^2=o(1). P 2 x Q C n R 2 δ 2 { Y ( D − 1 ) 2 F + ( D − 1 ) 2 ( Y F ) } ≤ C M n Q R 2 δ 2 = o ( 1 ) .
The bracket is analytic and O ( x 3 ) O(x^3) O ( x 3 ) on the compact interval, and Q / n Q/n Q / n is bounded. Absence of such pairs makes each relevant edge a root leaf, with no short curve inside its cap.
Neumann gaps. Let Y 0 Y_0 Y 0 be the root obtained by deleting maximal caps. An internal geodesic cut of Y 0 Y_0 Y 0 fills to a connected two-sided cut of X X X with the same length and with both areas increased. Lemma L.1 excludes all competing topologies at ratio u u u with error C M ( u 4 + u 2 / g ) C_M(u^4+u^2/g) C M ( u 4 + u 2 / g ) . A remaining single planar waist longer than δ \delta δ and of ratio at most u u u must have k − 1 > δ / ( 2 π u ) k-1>\delta/(2\pi u) k − 1 > δ / ( 2 π u ) . Summing (L.2) with L = 2 π u ( k − 1 ) L=2\pi u(k-1) L = 2 π u ( k − 1 ) gives error C M Q u 2 e − c M δ / u C_M Q u^2 e^{-c_M\delta/u} C M Q u 2 e − c M δ / u , after absorbing the boundary exponential in w k ( x e 2 π u ) w_k(xe^{2\pi u}) w k ( x e 2 π u ) . Choose u = δ / [ D M log ( Q + 2 ) ] u=\delta/[D_M\log(Q+2)] u = δ / [ D M log ( Q + 2 )] with D M D_M D M large. Outside events of probability tending to zero, the root has internal geodesic Cheeger constant at least u u u : otherwise the filled cut would have to be a δ \delta δ -short waist internal to the root. Lemma M.1 gives its Neumann gap at least u 2 / 144 u^2/144 u 2 /144 . Each other piece has area at most 2 π K 2\pi K 2 π K and no internal geodesic of length at most δ \delta δ , so its internal geodesic Cheeger constant is at least δ / ( 2 π K ) \delta/(2\pi K) δ / ( 2 π K ) . The same lemma gives the required gap.
The Cheeger minimum. Finally use u = t / Q u=t/\sqrt{Q} u = t / Q . Lemma L.1 excludes all non-single-planar cuts with this ratio. That lemma concerns cuts with two connected sides; this is no restriction, since by the graph-averaging reduction in the proof of Lemma M.1 every separating geodesic multicurve of ratio at most u u u contains a subfamily cutting X X X into two connected sides with ratio at most u u u . For a truncation level K 0 K_0 K 0 , unrelated to the cutoff K K K , the tail over k > K 0 k>K_0 k > K 0 is at most C t ∑ k > K 0 ( k − 1 ) 2 w k ( x e 2 π t / Q ) → 0 C_t\sum_{k>K_0}(k-1)^2w_k(xe^{2\pi t/\sqrt{Q}})\to0 C t ∑ k > K 0 ( k − 1 ) 2 w k ( x e 2 π t / Q ) → 0 uniformly as K 0 → ∞ K_0\to\infty K 0 → ∞ . For bounded types the planar side is the smaller side, of area 2 π ( k − 1 ) 2\pi(k-1) 2 π ( k − 1 ) . This proves the last assertion.
This appendix proves the two deterministic comparisons used in the proof of Theorem 17.1; both are also used in Appendix Q. The first, Lemma M.1 , bounds the relative Cheeger constant of a surface with short geodesic boundary from below in terms of its internal geodesic cuts, and hence gives a Neumann gap for each cut piece; it is applied in the proof of Lemma L.2, and its graph-averaging step is invoked in § 4.13 . The second, Lemma M.2, compares the low eigenvalue counts of a surface, cut along short separating geodesics into a tree of pieces, from above and below with those of a weighted star. No probabilistic input is used: the tools are the relative Cheeger inequality of He–Wu [90], the collar lemma [40], isotopy of multicurves [68], and Poincaré, trace and collar-mass estimates with min–max.
Lemma M.1 (Relative geodesic comparison). Let Y Y Y be a connected finite-area hyperbolic surface whose geodesic boundary circles have lengths at most δ \delta δ ; cusps are allowed. Let h N h_N h N count only the interior perimeter of a relative cut, and let H i n t H_{\mathrm{int}} H int be the infimum of length divided by smaller area over internal geodesic multicurves with two connected complementary components, excluding boundary-parallel classes. An empty infimum is infinity. Set w = arcsinh ( 1 / sinh ( δ / 2 ) ) w = \operatorname{arcsinh}(1/\sinh(\delta/2)) w = arcsinh ( 1/ sinh ( δ /2 )) and a δ = 1 + δ / ( 2 w ) a_\delta= 1 + \delta/(2w) a δ = 1 + δ / ( 2 w ) . Then
h N ≥ H i n t a δ + ( a δ + 2 ) H i n t . (466) h_N \ge\frac{H_{\mathrm{int}}}{a_\delta+ (a_\delta+ 2)H_{\mathrm{int}}}. \tag*{(466)} h N ≥ a δ + ( a δ + 2 ) H int H int . (466) For small δ \delta δ , if H i n t ≥ u H_{\mathrm{int}} \ge u H int ≥ u and 0 < u ≤ 1 0 < u \le1 0 < u ≤ 1 , the Neumann gap is at least u 2 / 144 u^2/144 u 2 /144 , independently of the topology.
Proof. The proof has five steps: the relative Cheeger inequality in the presence of cusps; surgery of an interface into a closed internal one; straightening it to geodesics with controlled area change; the isotopy justification of that step; and the reduction to cuts with connected sides, which gives (M.1).
Step 1: the relative Cheeger inequality with cusps. The relative Cheeger inequality is gap N ≥ h N 2 / 4 \operatorname{gap}_N \ge h_N^2/4 gap N ≥ h N 2 /4 ; see [90, Theorem 9]. That theorem is stated for compact hyperbolic surfaces with geodesic boundary. For finite-area Y Y Y with cusps the same coarea argument applies verbatim to every f ∈ H 1 ( Y ) f \in H^1(Y) f ∈ H 1 ( Y ) that vanishes above some horocycle in each cusp: the level sets of such f f f are compact relative cuts of Y Y Y , to which the defining inequality of h N h_N h N applies. These f f f are dense in H 1 ( Y ) H^1(Y) H 1 ( Y ) , because truncating at the horocycle of length ϵ \epsilon ϵ with a cutoff of gradient at most one changes ∥ f ∥ H 1 \|f\|_{H^1} ∥ f ∥ H 1 by o ( 1 ) o(1) o ( 1 ) as ϵ → 0 \epsilon\to0 ϵ → 0 , the cusp tail having area ϵ \epsilon ϵ ; since the Rayleigh quotient is continuous on H 1 ( Y ) H^1(Y) H 1 ( Y ) , the gap bound extends. We prove (M.1).
Step 2: from a relative interface to a closed internal one. Start with a smooth finite interface of interior length L L L , separating areas with minimum m m m . Such interfaces suffice by smooth approximation and truncation of cusp tails, whose added horocycle lengths and areas tend to zero. Remove contractible loops and boundary-parallel proper arcs by filling or emptying innermost disks. For a boundary-parallel arc the corresponding geodesic boundary subarc is no longer than the arc; the disk has perimeter at most twice the arc length and area at most that perimeter. Each removed original component is charged once. Thus the total area change is at most 2 L 2L 2 L and the remaining interface length is at most L L L . The isoperimetric inequality used here is [90, Lemma 10], and the factor two for boundary-parallel arcs is the one in the proof of [90, Lemma 12].
Two standard facts are used in the rest of this step and in the next one. First, since each boundary circle has length at most δ \delta δ , the half-collars of width w = arcsinh ( 1 / sinh ( δ / 2 ) ) w = \operatorname{arcsinh}(1/\sinh(\delta/2)) w = arcsinh ( 1/ sinh ( δ /2 )) about the boundary circles are embedded and pairwise disjoint: this is the collar lemma [40, Chapter 4], applied to the double of Y Y Y across its geodesic boundary, whose collars about the doubled circles are embedded and disjoint and restrict to the half-collars of Y Y Y . Second, the closed interface constructed below may contain components homotopic to a boundary circle; such components are not internal geodesic classes, and in the straightening step they are pushed into the collar and disappear from the interior interface, with a change of the side areas bounded by their length, by the same Fermi-coordinate primitive as for the other classes. Every remaining proper arc exits the embedded half-collars at both ends, even if both ends lie on the same circle. It has length at least 2 w 2w 2 w . Hence at most L / ( 2 w ) L/(2w) L / ( 2 w ) arcs, and at most L / w L/w L / w boundary circles, are involved. On each such circle assign a sufficiently thin strip wholly to one side, choosing the shorter of the two traces. The added interface length tends to at most half the circle length, and its area tends to zero. After smoothing, this gives a closed internal interface with length L ′ ≤ a δ L + o ( 1 ) L' \le a_\delta L + o(1) L ′ ≤ a δ L + o ( 1 ) and smaller area m ′ ≥ m − 2 L − o ( 1 ) m' \ge m - 2L - o(1) m ′ ≥ m − 2 L − o ( 1 ) .
Step 3: straightening. Straightening this closed interface gives a geodesic partition with
L g e o ≤ L ′ , m g e o ≥ m ′ − L ′ . (467) L_{\mathrm{geo}} \le L', \qquad m_{\mathrm{geo}} \ge m' - L'. \tag*{(467)} L geo ≤ L ′ , m geo ≥ m ′ − L ′ . (467)
Here is the area justification, including curves crossing their own geodesic representatives. Double across the boundary if necessary. In the cyclic cover of an essential curve, the area form in Fermi coordinates is ℓ cosh r d r d θ \ell\cosh r\,\mathrm{d}r\,\mathrm{d}\theta ℓ cosh r d r d θ , with primitive ℓ sinh r d θ \ell\sinh r\,\mathrm{d}\theta ℓ sinh r d θ of norm at most one, vanishing on the axis. The trace annulus of a homotopy lifts to this cover. Stokes bounds its signed area by the original curve length plus the absolute primitive integral on the final curve. On a parallel curve approaching the axis, the latter tends to zero. The same construction treats classes parallel to the original boundary: the parallel curve is pushed onto the boundary circle, where it ceases to be part of the interior interface, and the swept area is at most the curve length, which is already counted in L ′ L' L ′ . For cusp classes use d x / y \mathrm{d}x/y d x / y in a parabolic cover, whose norm is one and whose integral on a shrinking horocycle tends to zero. For contractible classes use ( cosh r − 1 ) d θ (\cosh r - 1)\mathrm{d}\theta ( cosh r − 1 ) d θ in polar coordinates on the hyperbolic plane; its norm is tanh ( r / 2 ) ≤ 1 \tanh(r/2) \le1 tanh ( r /2 ) ≤ 1 .
Step 4: the isotopy justification. The finite disjoint family can be simultaneously isotoped to disjoint nearby representatives, keeping parallel copies in their original order, pushing peripheral curves into collars or cusp tails, and shrinking nested contractible families in disks. The standard topological inputs are minimal position, homotopy versus isotopy and ambient extension [68, Corollary 1.9 and Propositions 1.10–1.11]. At each finite stage the ambient isotopy has compact support and fixes the boundary of Y Y Y . If D ′ D' D ′ is one side and D e ′ D'_e D e ′ its image, area transport expresses area ( D e ′ ) − area ( D ′ ) \operatorname{area}(D'_e)-\operatorname{area}(D') area ( D e ′ ) − area ( D ′ ) as the sum of the oriented trace-annulus integrals. Its absolute value is at most L ′ + o ( 1 ) L' + o(1) L ′ + o ( 1 ) by the primitives just described. As the neighborhoods shrink, the side converges in area to a union of geodesic complementary components. Parallel interfaces cancel in pairs because their side labels alternate; an odd family leaves one geodesic. Each retained geodesic minimizes length in its class. Since a ↦ min ( a , area ( Y ) − a ) a \mapsto\min(a,\operatorname{area}(Y)-a) a ↦ min ( a , area ( Y ) − a ) is 1-Lipschitz, this proves (M.2). No estimate for the entire swept region or for the symmetric difference of the sides has been used.
Step 5: reduction to connected sides. The resulting sides need not be connected. Regard the complementary pieces as vertices of a finite graph weighted by area, with interface lengths as edge weights. A connected component of the smaller side has no greater boundary-to-mass ratio. If its complement is disconnected and one complementary component has more than half the total mass, take the complement of that component: it is connected, has greater mass and no larger boundary. Otherwise averaging over complementary components selects one of ratio no greater than the previous ratio; its complement is connected through the original chosen component. Thus some cut with both sides connected has no larger ratio, using only the existing geodesics. It follows that L g e o ≥ H 0 m g e o L_{\mathrm{geo}} \ge H_0m_{\mathrm{geo}} L geo ≥ H 0 m geo for every finite H 0 ≤ H i n t H_0 \le H_{\mathrm{int}} H 0 ≤ H int , also when the interface is empty. Consequently L ′ ≥ H 0 ( m − 2 L − L ′ − o ( 1 ) ) L' \ge H_0(m - 2L - L' - o(1)) L ′ ≥ H 0 ( m − 2 L − L ′ − o ( 1 )) . Let the strip widths tend to zero and then let H 0 ↑ H i n t H_0 \uparrow H_{\mathrm{int}} H 0 ↑ H int to obtain (M.1). For a δ ≤ 2 a_\delta\le2 a δ ≤ 2 , h N ≥ u / ( 2 + 4 u ) ≥ u / 6 h_N \ge u/(2+4u) \ge u/6 h N ≥ u / ( 2 + 4 u ) ≥ u /6 , proving the gap bound. □
The second lemma is the surface–graph comparison. Its hypotheses, the Neumann gaps of the pieces and the depth and area bounds on the branches, are supplied in the proof of Theorem 17.1 by Lemma L.2; the proof below is a min–max argument with explicit trial maps in both directions, and its intermediate estimates (M.5)–(M.6) are the ones reused in Appendix Q.
Lemma M.2 (Comparison with the retained star). Cut a finite-area surface along separating geodesics of length at most δ \delta δ . Suppose their dual graph is a tree with a positive-genus root and planar other vertices. Every geodesic piece has Neumann gap at least 0 < γ ≤ 1 0 < \gamma\le1 0 < γ ≤ 1 , each root branch has depth at most K K K , and its area is at most C K CK C K . Choose δ < b \delta< b δ < b with b b b small. Trim each half-collar at
circumference b b b , let m v m_v m v be the remaining areas and W e W_e W e their total below edge e e e . Put
r e = arcosh ( b / ℓ e ) , I e = ∫ − r e r e sech r d r , κ e = ℓ e / I e . r_e=\operatorname{arcosh}(b/\ell_e),\qquad I_e=\int_{-r_e}^{r_e}\operatorname{sech}r\,dr,\qquad\kappa_e=\ell_e/I_e. r e = arcosh ( b / ℓ e ) , I e = ∫ − r e r e sech r d r , κ e = ℓ e / I e .
Retain edges with κ e / W e ≤ τ \kappa_e/W_e\leq\tau κ e / W e ≤ τ , and suppose each is an isolated root leaf. Contract the other vertices into the root, adding their masses, to form a weighted star. There are absolute constants C C C and T ≤ C K / ( b γ ) T\leq CK/(b\gamma) T ≤ C K / ( bγ ) such that, for Λ < 1 / 4 \Lambda<1/4 Λ < 1/4 ,
N ⋆ + ( Λ ) ≤ N X + ( Λ ) ≤ N ⋆ + ( B Λ / a ) , B = ( 1 + 2 τ T ) 2 , (468) N_\star^+(\Lambda)\leq N_X^+(\Lambda)\leq N_\star^+(B\Lambda/a),\qquad B=(1+\sqrt{2\tau T})^2, \tag*{(468)} N ⋆ + ( Λ ) ≤ N X + ( Λ ) ≤ N ⋆ + ( B Λ/ a ) , B = ( 1 + 2 τ T ) 2 , (468)
where
a = [ 1 − C Λ / γ 1 + C b − ( 2 K / τ + 4 K T ) Λ ] 2 , (469) a=\left[\sqrt{\frac{1-C\Lambda/\gamma}{1+Cb}}-\sqrt{(2K/\tau+4KT)\Lambda}\right]^2, \tag*{(469)} a = [ 1 + C b 1 − C Λ/ γ − ( 2 K / τ + 4 K T ) Λ ] 2 , (469)
provided the first square root and the bracket are positive. Moreover m v = ( 1 + O ( b ) ) area ( P v ) m_v=(1+O(b))\operatorname{area}(P_v) m v = ( 1 + O ( b )) area ( P v ) and κ e = ( ℓ e / π ) ( 1 + O ( δ / b ) ) \kappa_e=(\ell_e/\pi)(1+O(\delta/b)) κ e = ( ℓ e / π ) ( 1 + O ( δ / b )) .
Proof. Write B v B_v B v for a trimmed piece, a v a_v a v for the mean of a real H 1 H^1 H 1 function f f f on it, and p e , q e p_e,q_e p e , q e for its endpoint angular averages. The standard disjoint half-collars extend at least one unit beyond r e r_e r e when b b b is small. Their outer unit strips are therefore disjoint. Each removed half-collar has area b 2 − ℓ e 2 ≤ b \sqrt{b^2-\ell_e^2}\leq b b 2 − ℓ e 2 ≤ b . If χ v = − χ ( P v ) > 0 \chi_v=-\chi(P_v)>0 χ v = − χ ( P v ) > 0 and d v d_v d v is its boundary count, then d v ≤ 3 χ v d_v\leq3\chi_v d v ≤ 3 χ v . Hence m v = ( 1 + O ( b ) ) 2 π χ v m_v=(1+O(b))2\pi\chi_v m v = ( 1 + O ( b )) 2 π χ v and d v ≤ C m v d_v\leq Cm_v d v ≤ C m v .
Poincaré and trace estimates. Extend a function on B v B_v B v harmonically inward on every removed half-collar, with a free endpoint at the waist. In conformal coordinates, Fourier mode k ≠ 0 k\neq0 k = 0 has extension energy 2 π ∣ k ∣ tanh ( 2 π ∣ k ∣ T i n ) ∣ f k ∣ 2 2\pi|k|\tanh(2\pi|k|T_{\mathrm{in}})|f_k|^2 2 π ∣ k ∣ tanh ( 2 π ∣ k ∣ T in ) ∣ f k ∣ 2 . The outer strip has conformal length T o u t ≥ c / b T_{\mathrm{out}}\geq c/b T out ≥ c / b , and its energy for this trace is at least 2 π ∣ k ∣ tanh ( 2 π ∣ k ∣ T o u t ) ∣ f k ∣ 2 2\pi|k|\tanh(2\pi|k|T_{\mathrm{out}})|f_k|^2 2 π ∣ k ∣ tanh ( 2 π ∣ k ∣ T out ) ∣ f k ∣ 2 , at least half the extension bound. The zero mode extends constantly. Summing disjoint strips gives extension energy at most three times the original energy. Variance on B v B_v B v is no larger than variance of its extension, so
∫ B v ∣ f − a v ∣ 2 ≤ 3 γ − 1 ∫ B v ∣ ∇ f ∣ 2 . \int_{B_v}|f-a_v|^2\leq3\gamma^{-1}\int_{B_v}|\nabla f|^2. ∫ B v ∣ f − a v ∣ 2 ≤ 3 γ − 1 ∫ B v ∣∇ f ∣ 2 .
The one-dimensional trace inequality and angular averaging on each outer strip give ∣ p e − a v ∣ 2 ≤ C b − 1 ∫ s t r i p ( ∣ f − a v ∣ 2 + ∣ ∇ f ∣ 2 ) |p_e-a_v|^2\leq Cb^{-1}\int_{\mathrm{strip}}(|f-a_v|^2+|\nabla f|^2) ∣ p e − a v ∣ 2 ≤ C b − 1 ∫ strip ( ∣ f − a v ∣ 2 + ∣∇ f ∣ 2 ) . Sum over disjoint strips before applying Poincaré. With E = ∫ X ∣ ∇ f ∣ 2 E=\int_X|\nabla f|^2 E = ∫ X ∣∇ f ∣ 2 , this yields
∑ e n d p o i n t s ∣ p e − a v ∣ 2 ≤ C b γ E , ∑ e W e ( ( p e − a p a r e n t ) 2 + ( q e − a c h i l d ) 2 ) ≤ T E . (470) \sum_{\mathrm{endpoints}}|p_e-a_v|^2\leq\frac{C}{b\gamma}E,\qquad\sum_eW_e\left((p_e-a_{\mathrm{parent}})^2+(q_e-a_{\mathrm{child}})^2\right)\leq TE. \tag*{(470)} endpoints ∑ ∣ p e − a v ∣ 2 ≤ bγ C E , e ∑ W e ( ( p e − a parent ) 2 + ( q e − a child ) 2 ) ≤ T E . (470)
Here W e ≤ C K W_e\leq CK W e ≤ C K ; no boundary-count factor is introduced.
Mass in the collars. On a half-collar, for z ( r e ) = 0 z(r_e)=0 z ( r e ) = 0 ,
∫ 0 r e z 2 ℓ cosh r d r ≤ 4 ∫ 0 r e ( z ′ ) 2 ℓ cosh r d r . \int_0^{r_e}z^2\ell\cosh r\,dr\leq4\int_0^{r_e}(z')^2\ell\cosh r\,dr. ∫ 0 r e z 2 ℓ cosh r d r ≤ 4 ∫ 0 r e ( z ′ ) 2 ℓ cosh r d r .
Indeed the ground state ϕ = ( cosh r ) − 1 / 2 \phi=(\cosh r)^{-1/2} ϕ = ( cosh r ) − 1/2 has potential − ( cosh r ϕ ′ ) ′ / ( cosh r ϕ ) = 1 / 4 + 1 / ( 4 cosh 2 r ) -(\cosh r\,\phi')'/(\cosh r\,\phi)=1/4+1/(4\cosh^2r) − ( cosh r ϕ ′ ) ′ / ( cosh r ϕ ) = 1/4 + 1/ ( 4 cosh 2 r ) ; the boundary term vanishes at 0 0 0 since ϕ ′ ( 0 ) = 0 \phi'(0)=0 ϕ ′ ( 0 ) = 0 and at r e r_e r e since z ( r e ) = 0 z(r_e)=0 z ( r e ) = 0 . Subtract the outer average from the angular zero mode. Its mass is at most 2 b p e 2 2bp_e^2 2 b p e 2 plus eight times its radial energy. Nonzero modes have mass at most ( b / ( 2 π ) ) 2 (b/(2\pi))^2 ( b / ( 2 π ) ) 2 times their angular energy. Thus a full collar C e C_e C e has mass at most 2 b ( p e 2 + q e 2 ) + C ∫ C e ∣ ∇ f ∣ 2 2b(p_e^2+q_e^2)+C\int_{C_e}|\nabla f|^2 2 b ( p e 2 + q e 2 ) + C ∫ C e ∣∇ f ∣ 2 . Combining this with (M.5) and d v ≤ C m v d_v\leq Cm_v d v ≤ C m v , and writing M ( f ) = ∑ v m v a v 2 M(f)=\sum_vm_va_v^2 M ( f ) = ∑ v m v a v 2 , gives
∥ f ∥ 2 ≤ ( 1 + C b ) M ( f ) + C γ − 1 E , ∑ e κ e ( p e − q e ) 2 ≤ E . (471) \|f\|^2\leq(1+Cb)M(f)+C\gamma^{-1}E,\qquad\sum_e\kappa_e(p_e-q_e)^2\leq E. \tag*{(471)} ∥ f ∥ 2 ≤ ( 1 + C b ) M ( f ) + C γ − 1 E , e ∑ κ e ( p e − q e ) 2 ≤ E . (471)
The second inequality follows by radial Cauchy–Schwarz after angular averaging. Also I e = π + O ( δ / b ) I_e=\pi+O(\delta/b) I e = π + O ( δ / b ) , proving the conductance estimate. All extension arguments hold for H 1 H^1 H 1 traces by approximation.
Contracting branches. A nonretained edge has W e < κ e / τ W_e<\kappa_e/\tau W e < κ e / τ . Sum Cauchy–Schwarz along paths of length at most K K K , then over vertex masses. Since retained edges are root leaves, every other path consists only of nonretained edges. Equations (M.5)–(M.6) give
∑ v c o n t r a c t e d m v ∣ a v − a 0 ∣ 2 ≤ K ∑ e n o t r e t a i n e d W e ∣ a c h i l d − a p a r e n t ∣ 2 ≤ ( 2 K / τ + 4 K T ) E . \sum_{v\ \mathrm{contracted}}m_v|a_v-a_0|^2\le K\sum_{e\ \mathrm{not\ retained}}W_e|a_{\mathrm{child}}-a_{\mathrm{parent}}|^2\le(2K/\tau+4KT)E. v contracted ∑ m v ∣ a v − a 0 ∣ 2 ≤ K e not retained ∑ W e ∣ a child − a parent ∣ 2 ≤ ( 2 K / τ + 4 K T ) E .
The left side is the squared distance in the full core-mass norm between the original means and the vector replacing all contracted means by a 0 a_0 a 0 . On a subspace with E ≤ Λ ∥ f ∥ 2 E\le\Lambda\|f\|^2 E ≤ Λ∥ f ∥ 2 , the triangle inequality and (M.6) give M ⋆ ( J f ) ≥ a ∥ f ∥ 2 M_\star(Jf)\ge a\|f\|^2 M ⋆ ( J f ) ≥ a ∥ f ∥ 2 , with a a a in (M.4), where J J J keeps the root mean and the retained leaf means. For retained edges W e = m e W_e=m_e W e = m e and κ e ≤ τ m e \kappa_e\le\tau m_e κ e ≤ τ m e . Minkowski’s inequality in the edge-energy norm gives Q ⋆ ( J f ) ≤ ( 1 + 2 τ T ) 2 E Q_\star(Jf)\le(1+\sqrt{2\tau T})^2E Q ⋆ ( J f ) ≤ ( 1 + 2 τ T ) 2 E . The positivity of the bracket in (M.4) ensures that J J J is injective on the spectral subspace. Min–max proves the upper count bound in (M.3).
For the other direction, assign each star value constantly to its core, the root value to all contracted cores, and use radial harmonic interpolation on retained collars and constants on the other collars. This injective trial map has energy exactly Q ⋆ Q_\star Q ⋆ and mass at least M ⋆ M_\star M ⋆ . Min–max gives the lower count bound. Both comparisons include the constant zero mode before subtracting one; neither map needs to preserve orthogonality to constants. □
Uniform cutoffs at a cusp of a large planar surface
This appendix constructs low-energy cutoff functions around a typical cusp of a large WP sphere, with constants independent of the number of cusps; this is Lemma N.1 below, the first of the two inputs to the proof of Theorem 18.1 in Appendix P. The same construction identifies the conformal type of the area-rooted infinite WP sphere, which is the recurrent endpoint of Theorem 20.2 (§ 5.1 ). The strategy is to regularize the metric conformally to bounded geometry, and then to apply a finite planar resistance estimate on a triangulation of the regularized sphere; the resulting discrete potential is interpolated and smoothed into the cutoff. The inputs are the genus-zero volume bounds of [99], the local convergence of [31, Theorem 34], the conformal limit theorem of [153, Theorem 1.4], the triangulation of [27] and the resistance estimate of [85, Lemma 3.6].
We write E X ( f ) = ∫ X ∣ ∇ f ∣ 2 E_X(f)=\int_X|\nabla f|^2 E X ( f ) = ∫ X ∣∇ f ∣ 2 and let U b ( c ) U_b(c) U b ( c ) be the complete tail of cusp c c c bounded by a horocycle of length b b b . All horocycles below are chosen inside disjoint Margulis regions.
Lemma N.1 (Uniform cusp cutoffs). There are b > 0 b>0 b > 0 and M < ∞ M<\infty M < ∞ such that, for every e , p > 0 e,p>0 e , p > 0 , there are finite A ( e , p ) , N 0 ( e , p ) A(e,p),N_0(e,p) A ( e , p ) , N 0 ( e , p ) with the following property. For every N ≥ N 0 ( e , p ) N\ge N_0(e,p) N ≥ N 0 ( e , p ) , a WP sphere X N ∈ M 0 , N X_N\in\mathcal{M}_{0,N} X N ∈ M 0 , N with a fixed labeled cusp c c c admits, with probability at least 1 − p 1-p 1 − p , a function ϕ ∈ H 1 ( X N ) \phi\in H^1(X_N) ϕ ∈ H 1 ( X N ) satisfying
0 ≤ ϕ ≤ 1 , ϕ ∣ U b ( c ) = 1 , Lip h y p ϕ ≤ M , E X N ( ϕ ) ≤ e , area { ϕ ≠ 0 } ≤ A ( e , p ) . (472) 0\le\phi\le1,\qquad\phi|_{U_b(c)}=1,\qquad\operatorname{Lip}_{\mathrm{hyp}}\phi\le M,\qquad E_{X_N}(\phi)\le e,\qquad\operatorname{area}\{\phi\ne0\}\le A(e,p). \tag*{(472)} 0 ≤ ϕ ≤ 1 , ϕ ∣ U b ( c ) = 1 , Lip hyp ϕ ≤ M , E X N ( ϕ ) ≤ e , area { ϕ = 0 } ≤ A ( e , p ) . (472) The support need not be compact in the hyperbolic metric. More precisely, there are constants C , u 0 C,u_0 C , u 0 such that, for every N ≥ 3 N\ge3 N ≥ 3 and u ≥ u 0 u\ge u_0 u ≥ u 0 , with probability at least 1 − C u − 1 / 3 log u 1-Cu^{-1/3}\log u 1 − C u − 1/3 log u , the first three conditions in (N.1) hold together with
E X N ( ϕ ) ≤ C log u , area { ϕ ≠ 0 } ≤ C u . (473) E_{X_N}(\phi)\le\frac{C}{\log u},\qquad\operatorname{area}\{\phi\ne0\}\le Cu. \tag*{(473)} E X N ( ϕ ) ≤ log u C , area { ϕ = 0 } ≤ C u . (473) All constants are independent of N N N and u u u .
The proof also establishes the parabolic conformal type conjectured in [31, Section 7.3] for the puncture-completed, area-rooted local limit. We use the strong local convergence proved there, and derive the conformal assertion by changing the metric and controlling the resulting change of root law; the change of metric is needed because the original cusped metrics have no uniform injectivity-radius bound. The first subsection constructs the regularized metric and records its geometry in (N.4), the second checks the locality of the construction and identifies the conformal type of the local limit, and the third proves the lemma. The regularized metric is also used, for surfaces of arbitrary genus, in the proof of Proposition U.1 in Appendix U.
We replace the hyperbolic metric by a conformal metric on the filled sphere with bounded curvature and injectivity radius bounded below. The construction fills every cusp and inflates every short collar by fixed models, so that all constants are independent of N N N and of the surface; its output is the geometry summary (N.4) together with the explicit local form of the modified metric (the fixed cusp-disk model of diameter D D D and central area v v v , the collar inflation F / f F/f F / f , and the domination of the old metric outside the cusp disks); these are the only properties of the construction that the later subsections use. The only probabilistic input is the first moment (N.3) of the inverse collar lengths, which controls the area added by the inflated collars. Fix 0 < δ ≪ b ≪ 1 0<\delta\ll b\ll1 0 < δ ≪ b ≪ 1 . Let Z δ ( X N ) = ∑ ℓ γ ≤ δ ℓ γ − 1 Z_\delta(X_N)=\sum_{\ell_\gamma\leq\delta}\ell_\gamma^{-1} Z δ ( X N ) = ∑ ℓ γ ≤ δ ℓ γ − 1 , where the sum is over simple closed geodesics. Mirzakhani integration and the two-sided genus-zero volume bounds give
E Z δ ( X N ) ≤ C δ N . (474) \mathbb{E}Z_\delta(X_N)\leq C\delta N. \tag*{(474)} E Z δ ( X N ) ≤ C δ N . (474)
Indeed the zero-boundary factor for k k k cusps on one side satisfies
( N k ) V 0 , k + 1 V 0 , N − k + 1 V 0 , N ≤ C N 7 / 2 k 5 / 2 ( N − k ) 5 / 2 , 2 ≤ k ≤ N − 2. \binom{N}{k}\frac{V_{0,k+1}V_{0,N-k+1}}{V_{0,N}}\leq\frac{CN^{7/2}}{k^{5/2}(N-k)^{5/2}},\qquad2\leq k\leq N-2. ( k N ) V 0 , N V 0 , k + 1 V 0 , N − k + 1 ≤ k 5/2 ( N − k ) 5/2 C N 7/2 , 2 ≤ k ≤ N − 2.
These bounds include the stable small indices; see [99, Theorems 2.1–2.2]. The twist factor ℓ \ell ℓ cancels 1 / ℓ 1/\ell 1/ ℓ , the boundary factors are bounded for ℓ ≤ δ \ell\leq\delta ℓ ≤ δ , and summing C N k − 5 / 2 CNk^{-5/2} C N k − 5/2 for k ≤ N / 2 k\leq N/2 k ≤ N /2 proves the claim. Counting the two sides separately only enlarges the upper bound.
Fill every cusp conformally and inflate every collar of central length ℓ < δ \ell<\delta ℓ < δ . Denote the resulting closed sphere by M ( X N ) M(X_N) M ( X N ) . Below, M ( X N ) M(X_N) M ( X N ) , abbreviated M M M as in K M K_M K M , q M q_M q M and area ( M ) \operatorname{area}(M) area ( M ) , always denotes this regularized sphere. It is unrelated to the Lipschitz constant M M M of Lemma N.1, which reappears only as M ( t , α ) M(t,\alpha) M ( t , α ) at the end of the proof. Here is a construction with constants independent of N N N and X N X_N X N . On a geodesic collar use θ ∈ R / Z \theta\in\mathbb{R}/\mathbb{Z} θ ∈ R / Z and d s = d r / ( ℓ cosh r ) ds=dr/(\ell\cosh r) d s = d r / ( ℓ cosh r ) . Its metric is f ( s ) ( d s 2 + d θ 2 ) f(s)(ds^2+d\theta^2) f ( s ) ( d s 2 + d θ 2 ) , where f ( s ) = ℓ sec ( ℓ s ) f(s)=\ell\sec(\ell s) f ( s ) = ℓ sec ( ℓ s ) . Choose a smooth positive function H H H with H ( u ) = 2 H(u)=2 H ( u ) = 2 for u ≤ 1 u\leq1 u ≤ 1 , H ( u ) = u H(u)=u H ( u ) = u for u ≥ 3 u\geq3 u ≥ 3 , and H ( u ) ≥ max ( 2 , u ) H(u)\geq\max(2,u) H ( u ) ≥ max ( 2 , u ) , and replace f f f by F = b H ( f / b ) F=bH(f/b) F = b H ( f / b ) . Such an H H H is obtained by symmetric smoothing of max ( 2 , u ) \max(2,u) max ( 2 , u ) on its transition interval. The change is confined to f < 3 b f<3b f < 3 b , with a buffer out to f = 6 b f=6b f = 6 b . The central cylinder is flat, has circumference 2 b 2b 2 b , and adds area at most C / ℓ C/\ell C / ℓ because its conformal length is at most π / ℓ \pi/\ell π / ℓ .
On the transition band f f f is between fixed positive constants and
( f ′ ) 2 = f 2 ( f 2 − ℓ 2 ) , f ′ ′ = 2 f 3 − ℓ 2 f . (f')^2=f^2(f^2-\ell^2),\qquad f''=2f^3-\ell^2f. ( f ′ ) 2 = f 2 ( f 2 − ℓ 2 ) , f ′′ = 2 f 3 − ℓ 2 f .
These identities and their derivatives bound the curvature − F − 2 ( log F ) ′ ′ -F^{-2}(\log F)'' − F − 2 ( log F ) ′′ and all its covariant derivatives uniformly. For a cusp use u ≥ 0 u\geq0 u ≥ 0 , f ( u ) = b / ( 1 + b u ) f(u)=b/(1+bu) f ( u ) = b / ( 1 + b u ) and interpolate log f \log f log f smoothly, on a fixed bounded interval, to log c 0 − 2 π u \log c_0-2\pi u log c 0 − 2 π u . Thereafter use c 0 e − 2 π u c_0e^{-2\pi u} c 0 e − 2 π u . In the coordinate z = e − 2 π ( u + i θ ) z=e^{-2\pi(u+i\theta)} z = e − 2 π ( u + i θ ) this is a smooth flat metric at z = 0 z=0 z = 0 . Keep the old metric near u = 0 u=0 u = 0 and out to the horocycle of length 2 b 2b 2 b . All filled disks are the same fixed smooth model, of diameter at most D D D , each containing an embedded central disk of area v > 0 v>0 v > 0 . The metric outside these cusp disks dominates the old metric.
Choose b b b so that the buffered regions lie in disjoint Margulis components. In each modified collar the new arclength coordinate t t t gives d t 2 + F ( t ) 2 d θ 2 dt^2+F(t)^2d\theta^2 d t 2 + F ( t ) 2 d θ 2 , with 2 b ≤ F ≤ 6 b 2b\leq F\leq6b 2 b ≤ F ≤ 6 b and bounded ∣ ∂ t log F ∣ |\partial_t\log F| ∣ ∂ t log F ∣ . Fixed small rectangles give embedded uniformly bilipschitz disk charts, including across the buffered interfaces. The cusp models have the same property. Away from these regions the old injectivity radius is bounded below in terms of δ , b \delta,b δ , b . Thus every sufficiently small new metric ball contains a subchart of uniformly positive area. The curvature and noncollapse estimate of Cheeger–Gromov–Taylor [42, Theorem 4.7(i), (4.23)] gives a uniform positive injectivity radius. We have proved
∣ K M ∣ ≤ C , inj ( M ) ≥ c > 0 , c N ≤ area ( M ) ≤ C N + C Z δ ( X N ) , E area ( M ) ≤ C N . (475) |K_M|\leq C,\qquad\operatorname{inj}(M)\geq c>0,\qquad cN\leq\operatorname{area}(M)\leq CN+CZ_\delta(X_N),\qquad\mathbb{E}\operatorname{area}(M)\leq CN. \tag*{(475)} ∣ K M ∣ ≤ C , inj ( M ) ≥ c > 0 , c N ≤ area ( M ) ≤ C N + C Z δ ( X N ) , E area ( M ) ≤ C N . (475)
For the lower area bound, at most N b Nb N b of old cusp area is removed from 2 π ( N − 2 ) 2\pi(N-2) 2 π ( N − 2 ) . All curvature derivatives are uniformly bounded, so pointed smooth compactness is available.
Locality and the change of root measure
This subsection proves the conformal assertion for the area-rooted infinite WP sphere, the recurrent endpoint of Theorem 20.2 ; the proof of Lemma N.1 in the next subsection works on the finite surfaces and uses (N.4) together with the local description of the regularized metric from the first subsection, but not the recurrence argument of this subsection. The same regularization is defined on the infinite local limit of [31, Theorem 34]. We first check its compatibility with that theorem’s total-variation convergence of each fixed-radius metric ball, then apply the conformal limit theorem of [153, Theorem 1.4] to the regularized spheres rooted by their new area, and finally compare the new root law with the old one through the domination (N.5). In the fixed filled disk, any two boundary points can be joined along its boundary with length at most C 0 C_0 C 0 times their intrinsic distance in the disk. This follows near the diagonal from a smooth boundary collar and away from it by compactness. Replacing disk excursions by these arcs, a new path of length R R R between old-core points gives an old-core path of length at most C 0 R C_0R C 0 R . If the root has old cusp depth τ \tau τ , first move it to its cusp boundary. A new R R R -ball is consequently determined by an old ball of radius τ + C 1 ( R + D + 1 ) \tau+C_1(R+D+1) τ + C 1 ( R + D + 1 ) , together with the attached fixed disks.
The data identifying the modifications are local as well. A short primitive based loop in a Margulis region has developing holonomy which distinguishes a parabolic component from a hyperbolic one. In the latter case its translation length and distance to its axis determine ℓ \ell ℓ and f f f ; the central geodesic need not lie in the observed ball. In the parabolic case the holonomy determines the circumference coordinate. The loop and its development fit in a fixed additional neighborhood. Exact local isometries therefore preserve these data, including the threshold ℓ < δ \ell<\delta ℓ < δ . The filled disk is rotationally invariant, so no gluing phase is needed. There is no assertion that the operation is continuous at this sharp threshold for arbitrary weak metric convergence.
The modified infinite surface is complete. A bounded-length escaping path would, by the path comparison, remain in a compact part of the complete old truncated core. Only finitely many cusp disks can meet such a compact set, because their disjoint fixed-width buffers have positive area. Each such disk is compact. This excludes escape. The old area fraction at cusp depth greater than τ \tau τ is at most C e − τ Ce^{-\tau} C e − τ . Truncate this event and apply the exact local couplings of [31, Theorem 34]. It follows that M ( X N ) M(X_N) M ( X N ) rooted by old hyperbolic area converges locally to the modified area-rooted limit.
To identify its conformal type, let q M q_M q M be the law of M M M rooted uniformly in its new area. Bias the law of X N X_N X N by area ( M ) / area ( M ) \operatorname{area}(M)/\operatorname{area}(M) area ( M ) / area ( M ) , and let μ N \mu_N μ N be the barycenter of the resulting laws q M q_M q M . Namazi–Pankka–Souto [153, Theorem 1.4], after a fixed metric rescaling, applies to any deterministic sequence of these closed spheres whose areas diverge: their distributional limits are supported on surfaces conformally C \mathbb{C} C or C ∗ \mathbb{C}^* C ∗ . It also applies to our mixtures. To see this, let L \mathcal{L} L be the compact set of limit laws q M q_M q M along such deterministic sequences. Pointed bounded-geometry compactness implies that uniformly over spheres of area at least c N cN c N , the distance from q M q_M q M to L \mathcal{L} L tends to zero; otherwise a violating sequence would have a limit in L \mathcal{L} L . Every limiting distribution of the random law q M q_M q M is therefore supported on L \mathcal{L} L . Taking barycenters proves that every limit μ \mu μ of μ N \mu_N μ N is parabolic.
Let ν N \nu_N ν N instead root M ( X N ) M(X_N) M ( X N ) by old hyperbolic area, whose total mass is the deterministic 2 π ( N − 2 ) 2\pi(N-2) 2 π ( N − 2 ) . Outside old cusp depth τ \tau τ , the density d A o l d / d A n e w dA_{\mathrm{old}}/dA_{\mathrm{new}} d A old / d A new is at most C τ C_\tau C τ . Equation (N.4) consequently gives, for every bounded measurable 0 ≤ F ≤ 1 0\le F\le1 0 ≤ F ≤ 1 on pointed surfaces,
∫ F d ν N ≤ C τ ′ ∫ F d μ N + C e − τ . (476) \int F\,d\nu_N \le C'_\tau\int F\,d\mu_N+Ce^{-\tau}. \tag*{(476)} ∫ F d ν N ≤ C τ ′ ∫ F d μ N + C e − τ . (476)
Pass to simultaneous limits first for continuous tests, then for Borel sets by measure regularity. Sending τ → ∞ \tau\to\infty τ → ∞ shows ν ≪ μ \nu\ll\mu ν ≪ μ . Thus the modified Budd–Curien limit is parabolic. Its underlying surface before filling is homeomorphic to R 2 ∖ Z 2 \mathbb{R}^2\setminus\mathbb{Z}^2 R 2 ∖ Z 2 by the proof of [31, Lemma 33]; after filling it is simply connected and hence conformally C \mathbb{C} C . The unfilled surface is therefore C \mathbb{C} C minus a closed