Introduction
Non-vanishing of central L L L -values is a basic problem in analytic number theory. A standard approach is to approximate the reciprocal by a short Dirichlet polynomial and apply Cauchy–Schwarz to the resulting mollified moments. This method originates in Selberg’s work [58 ] and was developed further by Levinson and Conrey [18 , 38 ] .
For Dirichlet L L L -functions, positive-proportion non-vanishing has been established in families of primitive characters of large modulus and in quadratic families [1 , 28 , 59 ] . Further developments of mollification and improvements of the non-vanishing proportions can be found in [15 , 32 , 46 , 54 ] .
For automorphic L L L -functions on G L 2 \mathrm{GL}_2 GL 2 , we have non-vanishing results in the spectral aspect for Maass forms. See [5 , 40 , 41 , 63 ] . For the weight aspect for holomorphic forms, we can see [3 , 45 ] . In the level aspect for newforms, we refer to the work [2 , 21 , 29 , 35 , 51 , 56 , 57 ] . Results for twists of a fixed form are given in [12 , 22 , 50 , 55 ] , with further moment estimates for quadratic twists in [39 , 60 ] . Related results for Rankin–Selberg convolutions and for non-vanishing at spectrally varying special points are given in [36 , 42 , 44 , 52 , 53 , 64 ] and the references therein.
For G L 3 \mathrm{GL}_3 GL 3 , earlier work treats character twists of fixed forms [24 , 43 ] and symmetric square families [4 , 6 , 31 ] . The Kuznetsov formulas of Blomer [7 ] and Buttcane [16 ] provide tools for general G L 3 \mathrm{GL}_3 GL 3 families. Building on the kernel estimates of Blomer–Buttcane [9 ] , Huang–Liu–Xu [25 ] used a suitable mollifier to prove a positive harmonic non-vanishing proportion for forms of level one in generic short spectral intervals.
In the level aspect, Blomer–Buttcane–Maga [10 ] developed the Kuznetsov trace formula of the congruence subgroup and its applications, while Blomer–Comtat [11 ] supplied local and spectral input for prime level averages. Tsuzuki [62 ] proved simultaneous non-vanishing in families whose conductors have a fixed ramified part and an additional varying prime.
Motivated by these developments, we study non-vanishing of central standard L L L -values of degree three for some families of cuspidal automorphics of G L 3 \mathrm{GL}_3 GL 3 . We state the following non-vanishing theorem in § 1.1 and the underlying mollified moment formulas in § 1.2.
The family and the non-vanishing theorem
Let p p p be an odd prime and define
Γ 0 ( p ) : = { γ = ( γ i j ) 1 ≤ i , j ≤ 3 ∈ S L 3 ( Z ) : γ 31 ≡ γ 32 ≡ 0 ( m o d p ) } . \Gamma_0(p):=\{\gamma=(\gamma_{ij})_{1\le i,j\le3}\in\mathrm{SL}_3(\mathbb{Z}):\gamma_{31}\equiv\gamma_{32}\equiv0\pmod p\}. Γ 0 ( p ) := { γ = ( γ ij ) 1 ≤ i , j ≤ 3 ∈ SL 3 ( Z ) : γ 31 ≡ γ 32 ≡ 0 ( mod p )} .
We denote its index in S L 3 ( Z ) \mathrm{SL}_3(\mathbb{Z}) SL 3 ( Z ) by
P : = [ S L 3 ( Z ) : Γ 0 ( p ) ] = p 2 + p + 1. P:=[\mathrm{SL}_3(\mathbb{Z}):\Gamma_0(p)]=p^2+p+1. P := [ SL 3 ( Z ) : Γ 0 ( p )] = p 2 + p + 1.
Let B ( p ) \mathcal{B}(p) B ( p ) denote the family of cuspidal automorphic representations π \pi π on G L 3 \mathrm{GL}_3 GL 3 of conductor p p p with trivial central character and spherical archimedean component. We normalize the Fourier coefficients by A π ( 1 , 1 ) = 1 A_\pi(1,1)=1 A π ( 1 , 1 ) = 1 . Write μ π \mu_\pi μ π for the archimedean spectral parameters and L ( s , π ) L(s,\pi) L ( s , π ) for the standard L L L -function. In § 2.1, we specify the normalization and the local description at p p p .
Fix the test function h h h as in (2.19). We define
H h : = ∫ Re ( μ ) = 0 h ( μ ) spec ( μ ) d μ > 0. \mathcal{H}_h:=\int_{\operatorname{Re}(\mu)=0}h(\mu)\operatorname{spec}(\mu)\,\mathrm{d}\mu>0. H h := ∫ Re ( μ ) = 0 h ( μ ) spec ( μ ) d μ > 0.
The density spec ( μ ) d μ \operatorname{spec}(\mu)\,\mathrm{d}\mu spec ( μ ) d μ is specified in § 3.2. Here the finite adjoint L L L -function is defined for Re ( s ) > 1 \operatorname{Re}(s)>1 Re ( s ) > 1 as in [17 § 2.4] .
L ( s , π , Ad ) : = L ( s , π × π ~ ) ζ ( s ) = ∏ q ( 1 − q − s ) L q ( s , π × π ~ ) . L(s,\pi,\operatorname{Ad}):=\frac{L(s,\pi\times\widetilde{\pi})}{\zeta(s)}=\prod_q(1-q^{-s})L_q(s,\pi\times\widetilde{\pi}). L ( s , π , Ad ) := ζ ( s ) L ( s , π × π ) = q ∏ ( 1 − q − s ) L q ( s , π × π ) .
The total harmonic mass satisfies, as p → ∞ p \to\infty p → ∞ through primes (see § 6.1 ),
c c u s p ∑ π ∈ B ( p ) h ( μ π ) L ( 1 , π , A d ) = P H h + O h , ε ( p 1 + ε ) . c_{\mathrm{cusp}}\sum_{\pi\in B(p)}\frac{h(\mu_\pi)}{L(1,\pi,\mathrm{Ad})}=PH_h+O_{h,\varepsilon}(p^{1+\varepsilon}). c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) = P H h + O h , ε ( p 1 + ε ) .
The positive constant c c u s p c_{\mathrm{cusp}} c cusp is fixed in § 3.3 .
Theorem 1.1. Let h h h be the fixed test function. For every ε > 0 \varepsilon>0 ε > 0 and all sufficiently large primes p p p , we have
∑ π ∈ B ( p ) L ( 1 / 2 , π ) ≠ 0 h ( μ π ) L ( 1 , π , A d ) ≥ ( 1 2 − ε ) ∑ π ∈ B ( p ) h ( μ π ) L ( 1 , π , A d ) . \sum_{\substack{\pi\in B(p)\\L(1/2,\pi)\ne0}}\frac{h(\mu_\pi)}{L(1,\pi,\mathrm{Ad})}\geq\left(\frac{1}{2}-\varepsilon\right)\sum_{\pi\in B(p)}\frac{h(\mu_\pi)}{L(1,\pi,\mathrm{Ad})}. π ∈ B ( p ) L ( 1/2 , π ) = 0 ∑ L ( 1 , π , Ad ) h ( μ π ) ≥ ( 2 1 − ε ) π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) .
The asymptotic for the total harmonic mass shows that the non-vanishing subfamily has harmonic mass of order p 2 p^2 p 2 . The theorem provides a counterpart in the level aspect to the spectral result of Huang–Liu–Xu [25 Theorem 1.1 and Remark 1.2] .
The factor L ( 1 , π , A d ) − 1 L(1,\pi,\mathrm{Ad})^{-1} L ( 1 , π , Ad ) − 1 is the harmonic weight arising from the Kuznetsov formula. Removing it would require uniform control of an adjoint L L L -value approximation inside the mollified moments, together with the additional Hecke indices it introduces. For a treatment at level one, see [17 §§ 2.4 and 3] . Pointwise bounds for the adjoint values alone do not preserve the proportion 1 / 2 1/2 1/2 . We therefore retain the harmonic weight throughout.
Remark 1. The present paper treats the level aspect with the archimedean test function fixed. A natural continuation is to establish a hybrid non-vanishing theorem in which the prime level p p p and the archimedean spectral parameter T T T tend to infinity simultaneously. More precisely, one would seek a positive lower bound for the harmonic proportion of non-vanishing in the corresponding family, uniformly in a suitable joint range of p p p and T T T . Such a result would combine the level aspect studied here with the spectral aspect considered by Huang–Liu–Xu. Its proof would require hybrid estimates for the prime-level Kuznetsov formula and the associated mollified moments, with uniform control of the oldform and Eisenstein contributions as both parameters grow. We leave this problem for future work.
Mollified moments
Fix a square-free positive integer Q Q Q and 0 < η < 1 / 2 0<\eta<1/2 0 < η < 1/2 , independently of p p p . For primes p ∤ Q p\nmid Q p ∤ Q , put L = p η L=p^\eta L = p η and define
x ℓ : = μ ( ℓ ) ∏ q ∣ ℓ q q + 1 log ( L / ℓ ) log L 1 ℓ ≤ L , ( ℓ , Q ) = 1 . (1) x_\ell:=\mu(\ell)\prod_{q\mid\ell}\frac{q}{q+1}\frac{\log(L/\ell)}{\log L}\mathbf{1}_{\ell\leq L,\;(\ell,Q)=1}. \tag*{(1)} x ℓ := μ ( ℓ ) q ∣ ℓ ∏ q + 1 q log L log ( L / ℓ ) 1 ℓ ≤ L , ( ℓ , Q ) = 1 . (1)
Here μ ( n ) \mu(n) μ ( n ) denotes the Möbius function. These coefficients satisfy x 1 = 1 x_1=1 x 1 = 1 and ∣ x ℓ ∣ ≤ 1 |x_\ell|\leq1 ∣ x ℓ ∣ ≤ 1 . These x ℓ x_\ell x ℓ vanish unless ℓ \ell ℓ is square-free and ( ℓ , Q ) = 1 (\ell,Q)=1 ( ℓ , Q ) = 1 . We use them to define the mollifier
M π : = ∏ q ∣ Q ( 1 − A π ( 1 , q ) q + A π ( q , 1 ) q − 1 q 3 / 2 ) ∑ ℓ ≤ L x ℓ A π ( 1 , ℓ ) ℓ . (2) M_\pi:=\prod_{q\mid Q}\left(1-\frac{A_\pi(1,q)}{\sqrt{q}}+\frac{A_\pi(q,1)}{q}-\frac{1}{q^{3/2}}\right)\sum_{\ell\leq L}\frac{x_\ell A_\pi(1,\ell)}{\sqrt{\ell}}. \tag*{(2)} M π := q ∣ Q ∏ ( 1 − q A π ( 1 , q ) + q A π ( q , 1 ) − q 3/2 1 ) ℓ ≤ L ∑ ℓ x ℓ A π ( 1 , ℓ ) . (2)
The finite product consists of the inverse of Euler factors at primes dividing Q Q Q . Since L < p L<p L < p , all mollifier indices are prime to p p p . The mollified harmonic first and second moments are
M 1 ( p ) : = c c u s p ∑ π ∈ B ( p ) h ( μ π ) L ( 1 / 2 , π ) L ( 1 , π , A d ) M π , (3) \mathcal{M}_1(p):=c_{\mathrm{cusp}}\sum_{\pi\in B(p)}\frac{h(\mu_\pi)L(1/2,\pi)}{L(1,\pi,\mathrm{Ad})}M_\pi, \tag*{(3)} M 1 ( p ) := c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) L ( 1/2 , π ) M π , (3)
and
M 2 ( p ) : = c c u s p ∑ π ∈ B ( p ) h ( μ π ) ∣ L ( 1 / 2 , π ) ∣ 2 L ( 1 , π , Ad ) ∣ M π ∣ 2 . (4) \mathcal{M}_2(p) := c_{\mathrm{cusp}} \sum_{\pi\in\mathcal{B}(p)} \frac{h(\mu_\pi)|L(1/2,\pi)|^2}{L(1,\pi,\operatorname{Ad})}|M_\pi|^2. \tag*{(4)} M 2 ( p ) := c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) ∣ L ( 1/2 , π ) ∣ 2 ∣ M π ∣ 2 . (4)
Theorem 1.2. For every fixed square-free Q Q Q and 0 < η < 1 / 2 0 < \eta< 1/2 0 < η < 1/2 , the mollifier in (2) satisfies, as p → ∞ p \to\infty p → ∞ through primes with p ∤ Q p \nmid Q p ∤ Q ,
M 1 ( p ) = P H h + O h , Q , η ( p 3 / 2 ) , \mathcal{M}_1(p) = P\mathcal{H}_h + O_{h,Q,\eta}(p^{3/2}), M 1 ( p ) = P H h + O h , Q , η ( p 3/2 ) , and
M 2 ( p ) = ζ ( 2 ) ∏ q ∣ Q ( 1 − q − 2 ) ( 1 + η − 1 ) P H h + O h , Q , η ( p 2 log ( 2 + log p ) log p ) . \mathcal{M}_2(p) = \zeta(2)\prod_{q\mid Q}(1-q^{-2})(1+\eta^{-1})P\mathcal{H}_h + O_{h,Q,\eta}\left(\frac{p^2\log(2+\log p)}{\log p}\right). M 2 ( p ) = ζ ( 2 ) q ∣ Q ∏ ( 1 − q − 2 ) ( 1 + η − 1 ) P H h + O h , Q , η ( log p p 2 log ( 2 + log p ) ) .
The test function and the parameters Q , η Q,\eta Q , η are fixed before p p p tends to infinity. The finite inverse Euler factors replace the arithmetic factor ζ ( 2 ) \zeta(2) ζ ( 2 ) by ζ ( 2 ) ∏ q ∣ Q ( 1 − q − 2 ) \zeta(2)\prod_{q\mid Q}(1-q^{-2}) ζ ( 2 ) ∏ q ∣ Q ( 1 − q − 2 ) , which approaches one as Q Q Q includes sufficiently many fixed small primes.
To improve the lower bound obtained from these two moments alone, we use a construction with two pieces inspired by Michel and VanderKam [46 ] and define
M ^ π : = M π + ϵ ( π ) γ μ π ( 1 / 2 ) γ − μ π ( 1 / 2 ) M π ‾ . (5) \widehat{M}_\pi:= M_\pi+ \epsilon(\pi)\frac{\gamma_{\mu_\pi}(1/2)}{\gamma_{-\mu_\pi}(1/2)}\overline{M_\pi}. \tag*{(5)} M π := M π + ϵ ( π ) γ − μ π ( 1/2 ) γ μ π ( 1/2 ) M π . (5)
Here ϵ ( π ) \epsilon(\pi) ϵ ( π ) and γ μ \gamma_\mu γ μ are normalized as in (13) and (12) .
The functional equation gives L ( 1 / 2 , π ) M ^ π = 2 Re ( L ( 1 / 2 , π ) M π ) L(1/2,\pi)\widehat{M}_\pi= 2\operatorname{Re}(L(1/2,\pi)M_\pi) L ( 1/2 , π ) M π = 2 Re ( L ( 1/2 , π ) M π ) . The second moment for the mollifier with two pieces also involves the second moment without complex conjugation, which supplies its mixed term,
M 1 , 1 ( p ) : = c c u s p ∑ π ∈ B ( p ) h ( μ π ) L ( 1 / 2 , π ) 2 L ( 1 , π , Ad ) M π 2 . (6) \mathcal{M}_{1,1}(p) := c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(\mu_\pi)L(1/2,\pi)^2}{L(1,\pi,\operatorname{Ad})}M_\pi^2. \tag*{(6)} M 1 , 1 ( p ) := c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) L ( 1/2 , π ) 2 M π 2 . (6)
We refer to M 1 , 1 ( p ) \mathcal{M}_{1,1}(p) M 1 , 1 ( p ) as the mixed moment below.
Theorem 1.3. For every fixed square-free Q Q Q and 0 < η < 1 / 2 0 < \eta< 1/2 0 < η < 1/2 , as p → ∞ p \to\infty p → ∞ through primes, we have
M 1 , 1 ( p ) = P H h ∏ q ∣ Q ( 1 − 2 q ( q + 1 ) 2 + 1 q ( q + 1 ) 2 ) + O h , Q , η ( p 2 log p ) . (7) \mathcal{M}_{1,1}(p) = P\mathcal{H}_h\prod_{q\mid Q}\left(1-\frac{2}{\sqrt{q}(q+1)^2}+\frac{1}{q(q+1)^2}\right)+O_{h,Q,\eta}\left(\frac{p^2}{\log p}\right). \tag*{(7)} M 1 , 1 ( p ) = P H h q ∣ Q ∏ ( 1 − q ( q + 1 ) 2 2 + q ( q + 1 ) 2 1 ) + O h , Q , η ( log p p 2 ) . (7) The Euler product converges absolutely and lies in ( 0 , 1 ] (0,1] ( 0 , 1 ] .
The role of the construction with two pieces is visible in the moment constants. As Q Q Q ranges through products of increasingly many small primes, the arithmetic factor in Theorem 1.2 and the Euler product in Theorem 1.3 both tend to one. In this limit, the constants in the lower bounds from the mollifiers with one and two pieces are governed, respectively, by
1 1 + η − 1 ⟶ 1 3 , 2 ( 1 + η − 1 ) + 1 ⟶ 1 2 , η → ( 1 / 2 ) − . \frac{1}{1+\eta^{-1}} \longrightarrow\frac{1}{3}, \qquad\frac{2}{(1+\eta^{-1})+1} \longrightarrow\frac{1}{2}, \qquad\eta\to(1/2)^-. 1 + η − 1 1 ⟶ 3 1 , ( 1 + η − 1 ) + 1 2 ⟶ 2 1 , η → ( 1/2 ) − .
The second expression uses the mixed moment through (6.3). For each prescribed error, we first fix Q Q Q , then fix η < 1 / 2 \eta< 1/2 η < 1/2 , and finally take p p p sufficiently large. Together with the total harmonic mass, these three moments prove Theorem 1.1 , as shown in § 6.
Difficulties in the level aspect
First, ramification at p p p changes the dual Fourier indices produced by the functional equation. These terms require local coefficient identities at p p p , beyond the unramified Hecke relations. We use the local description in [11 Lemma 1] to put them in a form suitable for the Kuznetsov formula.
Second, the Kuznetsov formula also contains oldforms and Eisenstein series, whose contributions must be controlled uniformly in p p p . We sum the three oldvectors before taking absolute values to preserve cancellation in the inverse Gram matrix. The maximal Eisenstein new contribution is controlled by harmonic coefficient averages on G L 2 \mathrm{GL}_2 GL 2 .
Finally, the mollifier length L = p η L = p^\eta L = p η makes the Fourier twists grow with p p p . We need estimates for the first moment that are uniform for r ≤ Q p η r \le Qp^\eta r ≤ Q p η and s ≤ Q s \le Q s ≤ Q , and estimates for the second moment that are uniform for a b ≤ Q 2 p 2 η ab \le Q^2p^{2\eta} ab ≤ Q 2 p 2 η , with all indices prime to p p p . These ranges must accommodate every fixed 0 < η < 1 / 2 0 < \eta< 1/2 0 < η < 1/2 to obtain the proportion 1 / 2 − ε 1/2 - \varepsilon 1/2 − ε .
Outline of the proof
We first prove the uniform twisted moment formulas in Theorems 4.1 and 4.9 by the Kuznetsov formula at prime level.
We then insert the Hecke expansion of the mollifier to obtain M 1 ( p ) \mathcal{M}_1(p) M 1 ( p ) and M 2 ( p ) \mathcal{M}_2(p) M 2 ( p ) . Retaining the full factor log ( p / ( a b ) ) \log(p/(ab)) log ( p / ( ab )) in the diagonal contribution to the second moment produces both quadratic forms in (96) . Their evaluation gives the main term in Theorem 1.2 .
The mixed moment is evaluated separately using the asymmetric approximate functional equation of Proposition 2.8 . After combining the Fourier coefficients, we apply the Kuznetsov formula. The first diagonal gives the Euler product in Theorem 1.3 , while the dual diagonal is absent. The kernel bounds and spectral estimates control the remaining terms.
Finally, the central functional equation combines the three moments through (6.3). The Cauchy–Schwarz inequality and the total harmonic mass asymptotic give the required nonvanishing proportion.
Plan of this paper
§ 2 fixes the coefficient conventions and proves the approximate functional equations and weight estimates. § 3 gives the Kuznetsov formula at prime level and its normalizations. §§ 4 and 5 evaluate the twisted and mollified moments, respectively. § 6 completes the proof of Theorem 1.1 .
Notation
Throughout the paper, ε \varepsilon ε and A A A denote arbitrarily small and arbitrarily large positive real numbers, respectively. Their values may differ at different occurrences. The Mellin contours and the remaining auxiliary parameters are then fixed before p p p varies. In a truncation estimate, the positive exponent is fixed before the level varies and before the Mellin line for the tail is selected. We write e ( z ) = exp ( 2 π i z ) e(z) = \exp(2\pi iz) e ( z ) = exp ( 2 π i z ) , f ≪ g f \ll g f ≪ g or f = O ( g ) f = O(g) f = O ( g ) for ∣ f ∣ ≤ C g |f| \le Cg ∣ f ∣ ≤ C g with an implied positive constant, and f ≪ g f \ll g f ≪ g when both f ≪ g f \ll g f ≪ g and g ≪ f g \ll f g ≪ f hold. The relation f ∼ g f \sim g f ∼ g means f − g = o ( g ) f - g = o(g) f − g = o ( g ) . The symbols ( m , n ) (m,n) ( m , n ) , μ ( n ) \mu(n) μ ( n ) and τ ( n ) \tau(n) τ ( n ) denote the greatest common divisor, the Möbius function and the divisor function, respectively. In products indexed by q q q , the index runs over primes. For a prime q q q , ord q n \operatorname{ord}_q n ord q n denotes the exponent of q q q in n n n . A superscript ( p ) (p) ( p ) on an L L L -function indicates that its Euler factor at p p p is omitted. We use the Euclidean norm ∥ μ ∥ = ( ∑ j ∣ μ j ∣ 2 ) 1 / 2 \lVert\mu\rVert= \left(\sum_j |\mu_j|^2\right)^{1/2} ∥ μ ∥ = ( ∑ j ∣ μ j ∣ 2 ) 1/2 .
Preliminaries
GL3 L-functions
We work with the notations Γ 0 ( p ) \Gamma_0(p) Γ 0 ( p ) , P = p 2 + p + 1 P = p^2 + p + 1 P = p 2 + p + 1 and B ( p ) \mathcal{B}(p) B ( p ) from the introduction, with p p p an odd prime. The finite Fourier coefficients and local bases follow Blomer–Comtat [11 ] . The archimedean parameters use the convention of [17 § 1.3] , with standard gamma factor ∏ j Γ R ( s − μ j ) \prod_j \Gamma_{\mathbb{R}}(s-\mu_j) ∏ j Γ R ( s − μ j ) .
Fourier coefficients
For every prime q ≠ p q \ne p q = p , the Fourier coefficient A π ( q k , q ℓ ) A_\pi(q^k,q^\ell) A π ( q k , q ℓ ) is given by
det ( α q k + ℓ + 2 β q k + ℓ + 2 γ q k + ℓ + 2 α q k + 1 β q k + 1 γ q k + 1 1 1 1 ) / det ( α q 2 β q 2 γ q 2 α q β q γ q 1 1 1 ) , k , ℓ ≥ 0 , (8) \det\begin{pmatrix}
\alpha_q^{k+\ell+2} & \beta_q^{k+\ell+2} & \gamma_q^{k+\ell+2} \\
\alpha_q^{k+1} & \beta_q^{k+1} & \gamma_q^{k+1} \\
1 & 1 & 1
\end{pmatrix}
\bigg/ \det\begin{pmatrix}
\alpha_q^2 & \beta_q^2 & \gamma_q^2 \\
\alpha_q & \beta_q & \gamma_q \\
1 & 1 & 1
\end{pmatrix},
\qquad k,\ell\ge0, \tag*{(8)} det α q k + ℓ + 2 α q k + 1 1 β q k + ℓ + 2 β q k + 1 1 γ q k + ℓ + 2 γ q k + 1 1 / det α q 2 α q 1 β q 2 β q 1 γ q 2 γ q 1 , k , ℓ ≥ 0 , (8)
in terms of the Satake parameters α q \alpha_q α q , β q \beta_q β q , γ q \gamma_q γ q of π q \pi_q π q . At coinciding parameters, the quotient is interpreted by its Schur polynomial continuation. Since π \pi π is unitary with trivial central character, we have
α q β q γ q = 1 and { α q − 1 , β q − 1 , γ q − 1 } = { α q ‾ , β q ‾ , γ q ‾ } . \alpha_q\beta_q\gamma_q = 1 \quad\text{and} \quad\{\alpha_q^{-1},\beta_q^{-1},\gamma_q^{-1}\} = \{\overline{\alpha_q},\overline{\beta_q},\overline{\gamma_q}\}. α q β q γ q = 1 and { α q − 1 , β q − 1 , γ q − 1 } = { α q , β q , γ q } .
The bounds towards the Ramanujan conjecture of Kim–Sarnak [33 ] and Blomer–Brumley [8 Theorem 1] give
max { ∣ α q ∣ , ∣ β q ∣ , ∣ γ q ∣ } ≤ q θ 3 , θ 3 = 5 14 . \max\{|\alpha_q|,|\beta_q|,|\gamma_q|\} \le q^{\theta_3}, \qquad\theta_3 = \frac{5}{14}. max { ∣ α q ∣ , ∣ β q ∣ , ∣ γ q ∣ } ≤ q θ 3 , θ 3 = 14 5 .
For q = p q=p q = p , we know π p ≃ ∣ ⋅ ∣ p − 2 ρ × ∣ ⋅ ∣ p ρ St 2 \pi_p \simeq|\cdot|_p^{-2\rho} \times|\cdot|_p^\rho\operatorname{St}_2 π p ≃ ∣ ⋅ ∣ p − 2 ρ × ∣ ⋅ ∣ p ρ St 2 from the proof of [11 Lemma 1] . Furthermore, we have the following lemma.
Lemma 2.1. Let π ∈ B ( p ) \pi\in\mathcal{B}(p) π ∈ B ( p ) . Then there is ρ ∈ i R \rho\in i\mathbb{R} ρ ∈ i R such that the parameters of its standard local L L L -factor are
( α p , β p , γ p ) = ( p − 1 / 2 − ρ , p 2 ρ , 0 ) . (\alpha_p,\beta_p,\gamma_p) = (p^{-1/2-\rho},p^{2\rho},0). ( α p , β p , γ p ) = ( p − 1/2 − ρ , p 2 ρ , 0 ) .
Proof. See [11 Lemma 1] . □
The Fourier coefficients A π ( p k , p ℓ ) A_\pi(p^k,p^\ell) A π ( p k , p ℓ ) are still given by (8) with these local L L L -parameters.
Lemma 2.2. For k , ℓ ≥ 0 k,\ell\ge0 k , ℓ ≥ 0 we have
A π ( p k , p ℓ ) = ( α p β p ) k ∑ j = 0 ℓ α p j β p ℓ − j = p k ( − 1 / 2 + ρ ) A π ( 1 , p ℓ ) . (9) A_\pi(p^k,p^\ell) = (\alpha_p\beta_p)^k \sum_{j=0}^{\ell}\alpha_p^j\beta_p^{\ell-j} = p^{k(-1/2+\rho)}A_\pi(1,p^\ell). \tag*{(9)} A π ( p k , p ℓ ) = ( α p β p ) k j = 0 ∑ ℓ α p j β p ℓ − j = p k ( − 1/2 + ρ ) A π ( 1 , p ℓ ) . (9)
In particular, A π ( p , 1 ) = p − 1 / 2 + ρ A_\pi(p,1) = p^{-1/2+\rho} A π ( p , 1 ) = p − 1/2 + ρ .
Proof. Setting γ p = 0 \gamma_p = 0 γ p = 0 in (8) gives A π ( p k , p ℓ ) = ( α p β p ) k ( α p ℓ + 1 − β p ℓ + 1 ) / ( α p − β p ) A_\pi(p^k,p^\ell) = (\alpha_p\beta_p)^k(\alpha_p^{\ell+1}-\beta_p^{\ell+1})/(\alpha_p-\beta_p) A π ( p k , p ℓ ) = ( α p β p ) k ( α p ℓ + 1 − β p ℓ + 1 ) / ( α p − β p ) . This proves (9) . □
We extend these local coefficients multiplicatively to all positive integers m , n m,n m , n .
A π ~ ( m , n ) = A π ( m , n ) . (10) A_{\widetilde{\pi}}(m,n) = A_\pi(m,n). \tag*{(10)} A π ( m , n ) = A π ( m , n ) . (10)
In particular,
A π ( m , n ) = A π ( n , m ) , ( m n , p ) = 1. (11) A_\pi(m,n) = A_\pi(n,m), \qquad(mn,p)=1. \tag*{(11)} A π ( m , n ) = A π ( n , m ) , ( mn , p ) = 1. (11)
We will use, uniformly in π \pi π ,
∣ A π ( m , n ) ∣ ≪ ε ( m n ) θ 3 + ε . |A_\pi(m,n)| \ll_\varepsilon(mn)^{\theta_3+\varepsilon}. ∣ A π ( m , n ) ∣ ≪ ε ( mn ) θ 3 + ε .
At p p p this follows with exponent zero from (9) . At unramified primes, apply the preceding Ramanujan bound to the Schur polynomial in (8) and use the divisor estimate.
The standard L L L -function and its root number
For Re ( s ) > 1 \operatorname{Re}(s)>1 Re ( s ) > 1 ,
L ( s , π ) = ∑ n ≥ 1 A π ( 1 , n ) n s = ∏ q ∏ ξ ∈ { α q , β q , γ q } ( 1 − ξ q − s ) − 1 . L(s,\pi)=\sum_{n\geq1}\frac{A_\pi(1,n)}{n^s}=\prod_q\prod_{\xi\in\{\alpha_q,\beta_q,\gamma_q\}}(1-\xi q^{-s})^{-1}. L ( s , π ) = n ≥ 1 ∑ n s A π ( 1 , n ) = q ∏ ξ ∈ { α q , β q , γ q } ∏ ( 1 − ξ q − s ) − 1 .
The contragredient L L L -function L ( s , π ~ ) L(s,\widetilde{\pi}) L ( s , π ) is defined analogously. Let μ = μ π \mu=\mu_\pi μ = μ π denote the parameters in the Kuznetsov convention, so ∑ j μ j = 0 \sum_j\mu_j=0 ∑ j μ j = 0 , ∣ Re ( μ j ) ∣ ≤ θ 3 |\operatorname{Re}(\mu_j)|\leq\theta_3 ∣ Re ( μ j ) ∣ ≤ θ 3 and { μ j ~ } = { − μ j } \{\widetilde{\mu_j}\}=\{-\mu_j\} { μ j } = { − μ j } . Define
γ μ ( s ) : = ∏ j = 1 3 Γ R ( s − μ j ) , Γ R ( s ) = π − s / 2 Γ ( s / 2 ) . (12) \gamma_\mu(s):=\prod_{j=1}^{3}\Gamma_{\mathbb{R}}(s-\mu_j),\qquad\Gamma_{\mathbb{R}}(s)=\pi^{-s/2}\Gamma(s/2). \tag*{(12)} γ μ ( s ) := j = 1 ∏ 3 Γ R ( s − μ j ) , Γ R ( s ) = π − s /2 Γ ( s /2 ) . (12)
Then the completed L L L -function is
Λ ( s , π ) : = p s / 2 γ μ ( s ) L ( s , π ) , \Lambda(s,\pi):=p^{s/2}\gamma_\mu(s)L(s,\pi), Λ ( s , π ) := p s /2 γ μ ( s ) L ( s , π ) ,
which is entire and satisfies the functional equation
Λ ( s , π ) = ϵ ( π ) Λ ( 1 − s , π ~ ) , μ π ~ = − μ π . (13) \Lambda(s,\pi)=\epsilon(\pi)\Lambda(1-s,\widetilde{\pi}),\qquad\mu_{\widetilde{\pi}}=-\mu_\pi. \tag*{(13)} Λ ( s , π ) = ϵ ( π ) Λ ( 1 − s , π ) , μ π = − μ π . (13)
The conductor at height t t t is
C ( π , t ) ≍ p ∏ j = 1 3 ( 1 + ∣ t − Im ( μ j ) ∣ ) . C(\pi,t)\asymp p\prod_{j=1}^{3}(1+|t-\operatorname{Im}(\mu_j)|). C ( π , t ) ≍ p j = 1 ∏ 3 ( 1 + ∣ t − Im ( μ j ) ∣ ) .
Fix the standard additive character ψ = ∏ v ψ v \psi=\prod_v\psi_v ψ = ∏ v ψ v of Q \ A \mathbb{Q}\backslash\mathbb{A} Q \ A and the corresponding self-dual Haar measures. The global root number is
ϵ ( π ) = ∏ v ϵ ( 1 2 , π v , ψ v ) . \epsilon(\pi)=\prod_v\epsilon\left(\frac{1}{2},\pi_v,\psi_v\right). ϵ ( π ) = v ∏ ϵ ( 2 1 , π v , ψ v ) .
Lemma 2.3 (Root number at prime conductor). We have
ϵ ( s , π p , ψ p ) = − p − ρ − ( s − 1 / 2 ) , (14) \epsilon(s,\pi_p,\psi_p)=-p^{-\rho-(s-1/2)}, \tag*{(14)} ϵ ( s , π p , ψ p ) = − p − ρ − ( s − 1/2 ) , (14) ϵ ( π ) = − p − ρ = − p 1 / 2 A π ( p , 1 ) , (15) \epsilon(\pi)=-p^{-\rho}=-p^{1/2}A_\pi(p,1), \tag*{(15)} ϵ ( π ) = − p − ρ = − p 1/2 A π ( p , 1 ) , (15) ϵ ( π ) A π ~ ( 1 , n ) = − p 1 / 2 A π ( p , n ) ( n ≥ 1 ) . (16) \epsilon(\pi)A_{\widetilde{\pi}}(1,n)=-p^{1/2}A_\pi(p,n)\qquad(n\geq1). \tag*{(16)} ϵ ( π ) A π ( 1 , n ) = − p 1/2 A π ( p , n ) ( n ≥ 1 ) . (16) In particular , (16) also holds when p ∣ n p\mid n p ∣ n .
Proof. Let D : = WD ( π p ) D:=\operatorname{WD}(\pi_p) D := WD ( π p ) be the associated Weil–Deligne representation. Explicitly, D = ( r , N ) D=(r,N) D = ( r , N ) acts on V = ⟨ e 0 , e 1 , e 2 ⟩ V=\langle e_0,e_1,e_2\rangle V = ⟨ e 0 , e 1 , e 2 ⟩ by
r ( w ) e 0 = ∣ w ∣ − 2 ρ e 0 , N e 0 = 0 , r(w)e_0=|w|^{-2\rho}e_0,\qquad Ne_0=0, r ( w ) e 0 = ∣ w ∣ − 2 ρ e 0 , N e 0 = 0 ,
r ( w ) e 1 = ∣ w ∣ ρ − 1 / 2 e 1 , N e 1 = e 2 , r(w)e_1=|w|^{\rho-1/2}e_1,\qquad Ne_1=e_2, r ( w ) e 1 = ∣ w ∣ ρ − 1/2 e 1 , N e 1 = e 2 ,
r ( w ) e 2 = ∣ w ∣ ρ + 1 / 2 e 2 , N e 2 = 0. r(w)e_2=|w|^{\rho+1/2}e_2,\qquad Ne_2=0. r ( w ) e 2 = ∣ w ∣ ρ + 1/2 e 2 , N e 2 = 0.
Let I p I_p I p be the inertia subgroup and choose a geometric Frobenius Φ p \Phi_p Φ p with ∣ Φ p ∣ = p − 1 |\Phi_p|=p^{-1} ∣ Φ p ∣ = p − 1 . Since the Weil representation r r r is unramified, we have
V I p = V , ( ker N ) I p = ⟨ e 0 , e 2 ⟩ , V I p / ( ker N ) I p = ⟨ e 1 ⟩ . V^{I_p}=V,\qquad(\ker N)^{I_p}=\langle e_0,e_2\rangle,\qquad V^{I_p}/(\ker N)^{I_p}=\langle e_1\rangle. V I p = V , ( ker N ) I p = ⟨ e 0 , e 2 ⟩ , V I p / ( ker N ) I p = ⟨ e 1 ⟩ .
The Frobenius Φ p \Phi_p Φ p acts on ⟨ e 1 ⟩ \langle e_1\rangle ⟨ e 1 ⟩ by
r ( Φ p ) e 1 = p 1 / 2 − ρ e 1 . r(\Phi_p)e_1=p^{1/2-\rho}e_1. r ( Φ p ) e 1 = p 1/2 − ρ e 1 .
The conductor formula gives
a ( D ) = a ( r ) + dim V I p − dim ( ker N ) I p = 0 + 3 − 2 = 1. a(D)=a(r)+\dim V^{I_p}-\dim(\ker N)^{I_p}=0+3-2=1. a ( D ) = a ( r ) + dim V I p − dim ( ker N ) I p = 0 + 3 − 2 = 1.
Hence Deligne’s monodromy correction formula is
ϵ ( s , D , ψ p ) = ϵ ( s , r , ψ p ) det ( − p − s r ( Φ p ) ∣ V I p / ( ker N ) I p ) . \epsilon(s,D,\psi_p)=\epsilon(s,r,\psi_p)\det\left(-p^{-s}r(\Phi_p)\mid V^{I_p}/(\ker N)^{I_p}\right). ϵ ( s , D , ψ p ) = ϵ ( s , r , ψ p ) det ( − p − s r ( Φ p ) ∣ V I p / ( ker N ) I p ) .
see [20 § 8.12] and [61 § 4.1.6, pp. 20–21] . Since r r r is a direct sum of unramified characters and ψ p \psi_p ψ p has conductor Z p \mathbb{Z}_p Z p , we have ϵ ( s , r , ψ p ) = 1 \epsilon(s,r,\psi_p)=1 ϵ ( s , r , ψ p ) = 1 . Consequently,
ϵ ( s , π p , ψ p ) = − p − ρ − ( s − 1 / 2 ) \epsilon(s,\pi_p,\psi_p)=-p^{-\rho-(s-1/2)} ϵ ( s , π p , ψ p ) = − p − ρ − ( s − 1/2 )
which proves (14) .
At each finite prime ℓ ≠ p \ell\ne p ℓ = p , both π ℓ \pi_\ell π ℓ and ψ ℓ \psi_\ell ψ ℓ are unramified, and hence ϵ ( 1 / 2 , π ℓ , ψ ℓ ) = 1 \epsilon(1/2,\pi_\ell,\psi_\ell)=1 ϵ ( 1/2 , π ℓ , ψ ℓ ) = 1 . At infinity, the trivial central character and O 3 = S O 3 ⋅ { ± I 3 } O_3=SO_3\cdot\{\pm I_3\} O 3 = S O 3 ⋅ { ± I 3 } imply that an S O 3 SO_3 S O 3 -fixed vector is also O 3 O_3 O 3 -fixed. The archimedean Langlands parameter is therefore a sum of even characters of R × \mathbb{R}^{\times} R × , each of which has epsilon factor 1. See [34 Eqs. (3.7)–(3.8), p. 404] . It follows that
ϵ ( π ) = ϵ ( 1 2 , π p , ψ p ) = − p − ρ . (17) \epsilon(\pi)=\epsilon\left(\frac{1}{2},\pi_p,\psi_p\right)=-p^{-\rho}. \tag*{(17)} ϵ ( π ) = ϵ ( 2 1 , π p , ψ p ) = − p − ρ . (17)
The coefficients A π ( 1 , n ) A_\pi(1,n) A π ( 1 , n ) are the Dirichlet coefficients of the standard L L L -function, so the compatible normalization for π ~ \widetilde{\pi} π gives
A π ~ ( 1 , n ) = A π ( 1 , n ) ‾ ( n ≥ 1 ) . A_{\widetilde{\pi}}(1,n)=\overline{A_\pi(1,n)}\qquad(n\ge1). A π ( 1 , n ) = A π ( 1 , n ) ( n ≥ 1 ) .
The (17) and Lemma 2.2 imply
− p 1 / 2 A π ( p , 1 ) ‾ = − p 1 / 2 p − 1 / 2 − ρ = − p − ρ = ϵ ( π ) , -p^{1/2}\overline{A_\pi(p,1)}=-p^{1/2}p^{-1/2-\rho}=-p^{-\rho}=\epsilon(\pi), − p 1/2 A π ( p , 1 ) = − p 1/2 p − 1/2 − ρ = − p − ρ = ϵ ( π ) ,
which proves (15) . Using (9) at p p p , multiplicativity at the other primes and (10) , we obtain
− p 1 / 2 A π ( p , n ) ‾ = − p 1 / 2 A π ( p , 1 ) ‾ A π ( 1 , n ) ‾ = ϵ ( π ) A π ~ ( 1 , n ) . \begin{aligned}
-p^{1/2}\overline{A_\pi(p,n)}=-p^{1/2}\overline{A_\pi(p,1)}\overline{A_\pi(1,n)} \\
&=\epsilon(\pi)A_{\widetilde{\pi}}(1,n).
\end{aligned} − p 1/2 A π ( p , n ) = − p 1/2 A π ( p , 1 ) A π ( 1 , n ) = ϵ ( π ) A π ( 1 , n ) .
This proves (16) for every n ≥ 1 n\ge1 n ≥ 1 . □
Stirling’s formula and bounds for the zeta function
Stirling’s formula [48 § 8.4, Eq. (4.03)] gives, for ∣ arg ( s ) ∣ ≤ π − ε \lvert\arg(s)\rvert\le\pi-\varepsilon ∣ arg ( s )∣ ≤ π − ε , ∣ s ∣ ≫ 1 \lvert s\rvert\gg1 ∣ s ∣ ≫ 1 , any ε > 0 \varepsilon>0 ε > 0 and any integer $K_1\ge 0,
log Γ ( s ) = ( s − 1 2 ) log s − s + 1 2 log ( 2 π ) + ∑ j = 1 K 1 B 2 j 2 j ( 2 j − 1 ) s 2 j − 1 + O K 1 , ε ( 1 ∣ s ∣ 2 K 1 + 1 ) , \log\Gamma(s)=\left(s-\frac{1}{2}\right)\log s-s+\frac{1}{2}\log(2\pi)+\sum_{j=1}^{K_1}\frac{B_{2j}}{2j(2j-1)s^{2j-1}}+O_{K_1,\varepsilon}\left(\frac{1}{\lvert s\rvert^{2K_1+1}}\right), log Γ ( s ) = ( s − 2 1 ) log s − s + 2 1 log ( 2 π ) + j = 1 ∑ K 1 2 j ( 2 j − 1 ) s 2 j − 1 B 2 j + O K 1 , ε ( ∣ s ∣ 2 K 1 + 1 1 ) ,
where B j B_j B j is the j j j -th Bernoulli number. Thus, for s = σ + i t s=\sigma+it s = σ + i t with fixed σ \sigma σ and ∣ t ∣ ≥ 1 \lvert t\rvert\ge1 ∣ t ∣ ≥ 1 , we have
∣ Γ ( σ + i t ) ∣ = 2 π ∣ t ∣ σ − 1 / 2 e − π ∣ t ∣ / 2 ( 1 + O ( ∣ t ∣ − 1 ) ) , \lvert\Gamma(\sigma+it)\rvert=\sqrt{2\pi}\lvert t\rvert^{\sigma-1/2}e^{-\pi\lvert t\rvert/2}\left(1+O(\lvert t\rvert^{-1})\right), ∣ Γ ( σ + i t )∣ = 2 π ∣ t ∣ σ − 1/2 e − π ∣ t ∣ /2 ( 1 + O (∣ t ∣ − 1 ) ) ,
and
Γ ′ Γ ( σ + i t ) = log ( σ + i t ) − 1 2 ( σ + i t ) + O ( ∣ t ∣ − 2 ) . \frac{\Gamma'}{\Gamma}(\sigma+it)=\log(\sigma+it)-\frac{1}{2(\sigma+it)}+O(\lvert t\rvert^{-2}). Γ Γ ′ ( σ + i t ) = log ( σ + i t ) − 2 ( σ + i t ) 1 + O (∣ t ∣ − 2 ) .
Consequently,
∣ Γ R ( σ + i t ) ∣ ≍ ( 1 + ∣ t ∣ ) σ / 2 − 1 / 2 e − π ∣ t ∣ / 4 ( ∣ t ∣ ≥ 1 ) . \lvert\Gamma_{\mathbb{R}}(\sigma+it)\rvert\asymp(1+\lvert t\rvert)^{\sigma/2-1/2}e^{-\pi\lvert t\rvert/4}\qquad(\lvert t\rvert\ge1). ∣ Γ R ( σ + i t )∣ ≍ ( 1 + ∣ t ∣ ) σ /2 − 1/2 e − π ∣ t ∣ /4 (∣ t ∣ ≥ 1 ) .
Fix a ≥ 0 a \ge0 a ≥ 0 , c 0 ≤ c 1 c_0 \le c_1 c 0 ≤ c 1 and δ > 0 \delta> 0 δ > 0 . For c 0 ≤ Re ( u ) ≤ c 1 c_0 \le\operatorname{Re}(u) \le c_1 c 0 ≤ Re ( u ) ≤ c 1 , assume that each argument 1 / 2 + u ∓ μ j 1/2+u\mp\mu_j 1/2 + u ∓ μ j stays at distance at least δ \delta δ from { 0 , − 2 , − 4 , … } \{0,-2,-4,\ldots\} { 0 , − 2 , − 4 , … } . Stirling’s formula and the triangle inequality give
∣ γ ± μ ( 1 / 2 + u ) γ μ ( 1 / 2 ) ∣ ≪ a , c 0 , c 1 , δ ( 1 + ∥ μ ∥ + ∣ Im ( u ) ∣ ) C e 3 π ∣ Im ( u ) ∣ / 4 , ∣ Re ( μ j ) ∣ ≤ a , (18) \left|\frac{\gamma_{\pm\mu}(1/2+u)}{\gamma_\mu(1/2)}\right| \ll_{a,c_0,c_1,\delta} (1+\|\mu\|+|\operatorname{Im}(u)|)^C e^{3\pi|\operatorname{Im}(u)|/4}, \qquad|\operatorname{Re}(\mu_j)|\le a, \tag*{(18)} γ μ ( 1/2 ) γ ± μ ( 1/2 + u ) ≪ a , c 0 , c 1 , δ ( 1 + ∥ μ ∥ + ∣ Im ( u ) ∣ ) C e 3 π ∣ Im ( u ) ∣/4 , ∣ Re ( μ j ) ∣ ≤ a , (18)
where C = C ( a , c 0 , c 1 ) C=C(a,c_0,c_1) C = C ( a , c 0 , c 1 ) , and 1 / γ μ ( 1 / 2 ) 1/\gamma_\mu(1/2) 1/ γ μ ( 1/2 ) is interpreted by its entire continuation.
We also record two bounds for the zeta function. The integral formula in [19 § 4, Eq. (7), pp. 32–33] gives
∣ ( s − 1 ) ζ ( s ) ∣ ≤ ∣ s ∣ + ∣ s ( s − 1 ) ∣ Re ( s ) ≪ ( 1 + ∣ s ∣ ) 2 , Re ( s ) ≥ 1 / 2 , (19) |(s-1)\zeta(s)|\le|s|+\frac{|s(s-1)|}{\operatorname{Re}(s)}\ll(1+|s|)^2,\qquad\operatorname{Re}(s)\ge1/2, \tag*{(19)} ∣ ( s − 1 ) ζ ( s ) ∣ ≤ ∣ s ∣ + Re ( s ) ∣ s ( s − 1 ) ∣ ≪ ( 1 + ∣ s ∣ ) 2 , Re ( s ) ≥ 1/2 , (19)
with continuous extension at s = 1 s=1 s = 1 . The convexity estimate [27 § 5.2, Eqs. (5.19)–(5.20), p. 100] gives
ζ ( 5 / 8 + i t ) ≪ ε ( 1 + ∣ t ∣ ) 3 / 16 + ε , t ∈ R . (20) \zeta(5/8+it)\ll_\varepsilon(1+|t|)^{3/16+\varepsilon},\qquad t\in\mathbb{R}. \tag*{(20)} ζ ( 5/8 + i t ) ≪ ε ( 1 + ∣ t ∣ ) 3/16 + ε , t ∈ R . (20)
Approximate functional equations
Fix an integer J ≥ 2 J\ge2 J ≥ 2 and put
P J ( u , μ ) : = ∏ j = 1 3 ∏ ℓ = 0 J ∏ ϵ = ± 1 ( ( 1 2 + 2 ℓ + ϵ μ j ) 2 − u 2 ) . (21) P_J(u,\mu):=\prod_{j=1}^{3}\prod_{\ell=0}^{J}\prod_{\epsilon=\pm1}\left(\left(\frac{1}{2}+2\ell+\epsilon\mu_j\right)^2-u^2\right). \tag*{(21)} P J ( u , μ ) := j = 1 ∏ 3 ℓ = 0 ∏ J ϵ = ± 1 ∏ ( ( 2 1 + 2 ℓ + ϵ μ j ) 2 − u 2 ) . (21)
and
G J ( u , μ ) : = e u 2 P J ( u , μ ) P J ( 0 , μ ) . (22) G_J(u,\mu):=e^{u^2}\frac{P_J(u,\mu)}{P_J(0,\mu)}. \tag*{(22)} G J ( u , μ ) := e u 2 P J ( 0 , μ ) P J ( u , μ ) . (22)
On the generic spherical unitary spectrum occurring in Theorem 3.1 , the parameters satisfy ∣ Re ( μ j ) ∣ ≤ 5 / 14 < 1 / 2 |\operatorname{Re}(\mu_j)|\le5/14<1/2 ∣ Re ( μ j ) ∣ ≤ 5/14 < 1/2 . Hence P J ( 0 , μ ) ≠ 0 P_J(0,\mu)\ne0 P J ( 0 , μ ) = 0 on this spectrum. The function G J G_J G J is even in u u u , invariant under μ ↦ − μ \mu\mapsto-\mu μ ↦ − μ and under the Weyl group, and G J ( 0 , μ ) = 1 G_J(0,\mu)=1 G J ( 0 , μ ) = 1 .
The standard L-function
Write U μ + = V μ U_\mu^+=V_\mu U μ + = V μ and U μ − = W μ U_\mu^-=W_\mu U μ − = W μ . For y > 0 y>0 y > 0 and c > 1 c>1 c > 1 , we define
U μ ± ( y ) : = 1 2 π i ∫ ( c ) y − u G J ( u , μ ) γ ± μ ( 1 / 2 + u ) γ μ ( 1 / 2 ) d u u . U_\mu^\pm(y):=\frac{1}{2\pi i}\int_{(c)}y^{-u}G_J(u,\mu)\frac{\gamma_{\pm\mu}(1/2+u)}{\gamma_\mu(1/2)}\frac{du}{u}. U μ ± ( y ) := 2 π i 1 ∫ ( c ) y − u G J ( u , μ ) γ μ ( 1/2 ) γ ± μ ( 1/2 + u ) u d u .
Proposition 2.4. For π ∈ B ( p ) \pi\in\mathcal{B}(p) π ∈ B ( p ) , the central value L ( 1 / 2 , π ) L(1/2,\pi) L ( 1/2 , π ) satisfies the following approximate functional equation
L ( 1 / 2 , π ) = ∑ n ≥ 1 A π ( 1 , n ) n V μ π ( n / p ) − p ∑ n ≥ 1 A π ( p , n ) ‾ n W μ π ( n / p ) . (23) L(1/2,\pi)=\sum_{n\ge1}\frac{A_\pi(1,n)}{\sqrt{n}}V_{\mu_\pi}(n/\sqrt{p})-\sqrt{p}\sum_{n\ge1}\frac{\overline{A_\pi(p,n)}}{\sqrt{n}}W_{\mu_\pi}(n/\sqrt{p}). \tag*{(23)} L ( 1/2 , π ) = n ≥ 1 ∑ n A π ( 1 , n ) V μ π ( n / p ) − p n ≥ 1 ∑ n A π ( p , n ) W μ π ( n / p ) . (23)
Proof. For μ = μ π \mu=\mu_\pi μ = μ π and c > 1 c>1 c > 1 , consider
1 2 π i ∫ ( c ) Λ ( 1 / 2 + u , π ) p 1 / 4 γ μ ( 1 / 2 ) G J ( u , μ ) d u u . \frac{1}{2\pi i}\int_{(c)}\frac{\Lambda(1/2+u,\pi)}{p^{1/4}\gamma_\mu(1/2)}G_J(u,\mu)\frac{du}{u}. 2 π i 1 ∫ ( c ) p 1/4 γ μ ( 1/2 ) Λ ( 1/2 + u , π ) G J ( u , μ ) u d u .
Moving the contour to Re ( u ) = − c \operatorname{Re}(u)=-c Re ( u ) = − c gives the residue L ( 1 / 2 , π ) L(1/2,\pi) L ( 1/2 , π ) . The functional equation (13) and (16) then yield (23) after expanding the Dirichlet series. □
We record the following estimates for U μ ± ( y ) U_\mu^\pm(y) U μ ± ( y ) .
Lemma 2.5. Let J ≥ 2 J \ge2 J ≥ 2 , A > 0 A > 0 A > 0 and j ∈ Z ≥ 0 j \in\mathbb{Z}_{\ge0} j ∈ Z ≥ 0 . On the unitary automorphic spectrum with ∣ Re ( μ j ) ∣ ≤ 5 / 14 |\operatorname{Re}(\mu_j)| \le5/14 ∣ Re ( μ j ) ∣ ≤ 5/14 , for every y > 0 y > 0 y > 0 ,
∣ y j d j d y j U μ ± ( y ) ∣ ≪ A , j , J ( 1 + ∥ μ ∥ ) C ( 1 + y ) − A , (24) \left|y^j \frac{\mathrm{d}^j}{\mathrm{d}y^j}U_\mu^\pm(y)\right| \ll_{A,j,J} (1+\|\mu\|)^C(1+y)^{-A}, \tag*{(24)} y j d y j d j U μ ± ( y ) ≪ A , j , J ( 1 + ∥ μ ∥ ) C ( 1 + y ) − A , (24) where C = C ( A , j , J ) C=C(A,j,J) C = C ( A , j , J ) . For 0 < y ≤ 1 0<y\le1 0 < y ≤ 1 ,
U μ ± ( y ) = γ ± μ ( 1 / 2 ) γ μ ( 1 / 2 ) + O J ( y 2 ( 1 + ∥ μ ∥ ) C ) , (25) U_\mu^\pm(y)=\frac{\gamma_{\pm\mu}(1/2)}{\gamma_\mu(1/2)}+O_J\left(y^2(1+\|\mu\|)^C\right), \tag*{(25)} U μ ± ( y ) = γ μ ( 1/2 ) γ ± μ ( 1/2 ) + O J ( y 2 ( 1 + ∥ μ ∥ ) C ) , (25) with a suitable C = C ( J ) C=C(J) C = C ( J ) . The ratio in the main term has modulus one.
Proof. Since ∣ Re ( μ k ) ∣ ≤ 5 / 14 |\operatorname{Re}(\mu_k)|\le5/14 ∣ Re ( μ k ) ∣ ≤ 5/14 , P J ( 0 , μ ) P_J(0,\mu) P J ( 0 , μ ) is bounded away from zero, and the factors in (21) cancel the numerator gamma poles in Re ( u ) ≥ − 2 \operatorname{Re}(u)\ge-2 Re ( u ) ≥ − 2 . The gamma recurrence and (18) give, for u = σ + i v u=\sigma+iv u = σ + i v with − 2 ≤ σ ≤ A + 2 -2\le\sigma\le A+2 − 2 ≤ σ ≤ A + 2 ,
∣ G J ( u , μ ) γ ± μ ( 1 / 2 + u ) γ μ ( 1 / 2 ) ∣ ≪ A , J ( 1 + ∥ μ ∥ + ∣ v ∣ ) C e − v 2 + 3 π ∣ v ∣ / 4 . \left|G_J(u,\mu)\frac{\gamma_{\pm\mu}(1/2+u)}{\gamma_\mu(1/2)}\right|\ll_{A,J}(1+\|\mu\|+|v|)^C e^{-v^2+3\pi|v|/4}. G J ( u , μ ) γ μ ( 1/2 ) γ ± μ ( 1/2 + u ) ≪ A , J ( 1 + ∥ μ ∥ + ∣ v ∣ ) C e − v 2 + 3 π ∣ v ∣/4 . This estimate justifies the contour shifts and differentiation under the integral below. For 0 < y ≤ 1 0<y\le1 0 < y ≤ 1 , move to Re ( u ) = − 2 \operatorname{Re}(u)=-2 Re ( u ) = − 2 . The residue at zero is γ ± μ ( 1 / 2 ) / γ μ ( 1 / 2 ) \gamma_{\pm\mu}(1/2)/\gamma_\mu(1/2) γ ± μ ( 1/2 ) / γ μ ( 1/2 ) , which gives (25) . For j ≥ 1 j\ge1 j ≥ 1 , differentiating y − u y^{-u} y − u introduces the factor ( − 1 ) j u ( u + 1 ) ⋯ ( u + j − 1 ) (-1)^ju(u+1)\cdots(u+j-1) ( − 1 ) j u ( u + 1 ) ⋯ ( u + j − 1 ) after multiplication by y j y^j y j . This cancels the pole at zero, so the same shift gives
∣ y j d j d y j U μ ± ( y ) ∣ ≪ j , J y 2 ( 1 + ∥ μ ∥ ) C . \left|y^j \frac{\mathrm{d}^j}{\mathrm{d}y^j}U_\mu^\pm(y)\right|\ll_{j,J} y^2(1+\|\mu\|)^C. y j d y j d j U μ ± ( y ) ≪ j , J y 2 ( 1 + ∥ μ ∥ ) C . Unitarity gives γ − μ ( 1 / 2 ) = γ μ ( 1 / 2 ) ‾ \gamma_{-\mu}(1/2)=\overline{\gamma_\mu(1/2)} γ − μ ( 1/2 ) = γ μ ( 1/2 ) , so the ratio in the main term has modulus one. This proves (24) for y ≤ 1 y\le1 y ≤ 1 . For y ≥ 1 y\ge1 y ≥ 1 , moving to Re ( u ) = A + 1 \operatorname{Re}(u)=A+1 Re ( u ) = A + 1 bounds the differentiated integral by O A , j , J ( y − A − 1 ( 1 + ∥ μ ∥ ) C ) O_{A,j,J}(y^{-A-1}(1+\|\mu\|)^C) O A , j , J ( y − A − 1 ( 1 + ∥ μ ∥ ) C ) , completing the proof. □
The product of the standard and dual L L L -functions
For the second moment, all mollified moments, and non-vanishing, we use the fixed test function
h ( μ ) : = exp ( ∑ j = 1 3 μ j 2 ) ∏ j = 1 3 cos 4 ( π μ j ) ∏ 1 ≤ i < j ≤ 3 cos 2 ( π ( μ i − μ j ) 2 ) . (26) h(\mu):=\exp\left(\sum_{j=1}^{3}\mu_j^2\right)\prod_{j=1}^{3}\cos^4(\pi\mu_j)\prod_{1\le i<j\le3}\cos^2\left(\frac{\pi(\mu_i-\mu_j)}{2}\right). \tag*{(26)} h ( μ ) := exp ( j = 1 ∑ 3 μ j 2 ) j = 1 ∏ 3 cos 4 ( π μ j ) 1 ≤ i < j ≤ 3 ∏ cos 2 ( 2 π ( μ i − μ j ) ) . (26)
The test h h h is independent of p p p , η \eta η and J J J . It is even, Weyl invariant and rapidly decreasing on every finite tube. Unitarity makes the multiset of parameters invariant under μ j ↦ − μ j \mu_j\mapsto-\mu_j μ j ↦ − μ j , so ∑ j μ j 2 \sum_j\mu_j^2 ∑ j μ j 2 is real and its exponential is positive. The coordinate cosine factors pair into absolute-value powers. The difference factors occur in conjugate pairs or as squares of real cosines. Thus h h h is non-negative on the unitary spectrum. On Re ( μ ) = 0 \operatorname{Re}(\mu)=0 Re ( μ ) = 0 , all cosine factors are positive hyperbolic cosines, so h > 0 h>0 h > 0 . Its zeros on μ j ∈ 1 2 + Z \mu_j\in\frac{1}{2}+\mathbb{Z} μ j ∈ 2 1 + Z and μ i − μ j ∈ 2 Z + 1 \mu_i-\mu_j\in2\mathbb{Z}+1 μ i − μ j ∈ 2 Z + 1 have orders at least four and two, respectively. For a 0 ≥ 10 a_0\ge10 a 0 ≥ 10 and an integer J ≥ ⌈ a 0 ⌉ + 4 J\ge\lceil a_0\rceil+4 J ≥ ⌈ a 0 ⌉ + 4 , the quotient
h 0 ( μ ) : = h ( μ ) P J ( 0 , μ ) (27) h_0(\mu):=\frac{h(\mu)}{P_J(0,\mu)} \tag*{(27)} h 0 ( μ ) := P J ( 0 , μ ) h ( μ ) (27)
belongs to H ~ ( a 0 + 1 ) \widetilde{\mathcal{H}}(a_0+1) H ( a 0 + 1 ) . The test class, the tubes Λ a \Lambda_a Λ a and the coordinate derivatives used below are defined in § 3.1.
For y > 0 y>0 y > 0 and c > 1 c>1 c > 1 , we define
U μ ( y ) : = 1 2 π i ∫ ( c ) y − u G J ( u , μ ) γ μ ( 1 / 2 + u ) γ − μ ( 1 / 2 + u ) γ μ ( 1 / 2 ) γ − μ ( 1 / 2 ) d u u . (28) U_{\mu}(y):=\frac{1}{2\pi i}\int_{(c)}y^{-u}G_J(u,\mu)\frac{\gamma_{\mu}(1/2+u)\gamma_{-\mu}(1/2+u)}{\gamma_{\mu}(1/2)\gamma_{-\mu}(1/2)}\frac{du}{u}. \tag*{(28)} U μ ( y ) := 2 π i 1 ∫ ( c ) y − u G J ( u , μ ) γ μ ( 1/2 ) γ − μ ( 1/2 ) γ μ ( 1/2 + u ) γ − μ ( 1/2 + u ) u d u . (28)
Proposition 2.6. For π ∈ B ( p ) \pi\in\mathcal{B}(p) π ∈ B ( p ) , the absolute square ∣ L ( 1 / 2 , π ) ∣ 2 |L(1/2,\pi)|^2 ∣ L ( 1/2 , π ) ∣ 2 satisfies the following approximate functional equation
∣ L ( 1 / 2 , π ) ∣ 2 = 2 ∑ m , n ≥ 1 A π ( 1 , m ) ‾ A π ( 1 , n ) m n U μ π ( m n / p ) . (29) |L(1/2,\pi)|^2=2\sum_{m,n\geq1}\frac{\overline{A_{\pi}(1,m)}A_{\pi}(1,n)}{\sqrt{mn}}U_{\mu_{\pi}}(mn/p). \tag*{(29)} ∣ L ( 1/2 , π ) ∣ 2 = 2 m , n ≥ 1 ∑ mn A π ( 1 , m ) A π ( 1 , n ) U μ π ( mn / p ) . (29) Proof. For μ = μ π \mu=\mu_{\pi} μ = μ π and c > 1 c>1 c > 1 , consider
1 2 π i ∫ ( c ) Λ ( 1 / 2 + u , π ) Λ ( 1 / 2 + u , π ~ ) p 1 / 2 γ μ ( 1 / 2 ) γ − μ ( 1 / 2 ) G J ( u , μ ) d u u . \frac{1}{2\pi i}\int_{(c)}\frac{\Lambda(1/2+u,\pi)\Lambda(1/2+u,\widetilde{\pi})}{p^{1/2}\gamma_{\mu}(1/2)\gamma_{-\mu}(1/2)}G_J(u,\mu)\frac{du}{u}. 2 π i 1 ∫ ( c ) p 1/2 γ μ ( 1/2 ) γ − μ ( 1/2 ) Λ ( 1/2 + u , π ) Λ ( 1/2 + u , π ) G J ( u , μ ) u d u . The same contour argument as in the proof of Proposition 2.4 gives (29) . □
We next estimate U μ ( y ) U_{\mu}(y) U μ ( y ) and the associated spectral test functions
H t ( μ ) : = h ( μ ) U μ ( t / p ) , for t ≥ 1. H_t(\mu):=h(\mu)U_{\mu}(t/p),\qquad\text{for }t\geq1. H t ( μ ) := h ( μ ) U μ ( t / p ) , for t ≥ 1.
Lemma 2.7. Fix a 0 ≥ 10 a_0\geq10 a 0 ≥ 10 and an integer J ≥ ⌈ a 0 ⌉ + 4 J\geq\lceil a_0\rceil+4 J ≥ ⌈ a 0 ⌉ + 4 . For every fixed B > 0 B>0 B > 0 , on the unitary automorphic spectrum with ∣ Re ( μ j ) ∣ ≤ 5 / 14 |\operatorname{Re}(\mu_j)|\leq5/14 ∣ Re ( μ j ) ∣ ≤ 5/14 ,
∣ U μ ( y ) ∣ ≪ B , J ( 1 + ∥ μ ∥ ) C ( 1 + y ) − B (30) |U_{\mu}(y)|\ll_{B,J}(1+\|\mu\|)^C(1+y)^{-B} \tag*{(30)} ∣ U μ ( y ) ∣ ≪ B , J ( 1 + ∥ μ ∥ ) C ( 1 + y ) − B (30) for a suitable C C C . For every c > 0 c>0 c > 0 , M ≥ 0 M\geq0 M ≥ 0 and integer k ≥ 0 k\geq0 k ≥ 0 , H t H_t H t belongs to H ~ ( a 0 ) \widetilde{\mathcal{H}}(a_0) H ( a 0 ) and satisfies
sup μ ∈ Λ a 0 ∣ α ∣ ≤ k ( 1 + ∥ μ ∥ ) M ∣ ∂ μ α H t ( μ ) ∣ ≪ c , M , k , h , J ( p / t ) c . (31) \sup_{\substack{\mu\in\Lambda_{a_0}\\|\alpha|\leq k}}(1+\|\mu\|)^M|\partial_{\mu}^{\alpha}H_t(\mu)|\ll_{c,M,k,h,J}(p/t)^c. \tag*{(31)} μ ∈ Λ a 0 ∣ α ∣ ≤ k sup ( 1 + ∥ μ ∥ ) M ∣ ∂ μ α H t ( μ ) ∣ ≪ c , M , k , h , J ( p / t ) c . (31) At zeros of P J ( 0 , μ ) P_J(0,\mu) P J ( 0 , μ ) , H t H_t H t is understood by holomorphic continuation.
Proof. Write
R μ ( u ) : = γ μ ( 1 / 2 + u ) γ − μ ( 1 / 2 + u ) γ μ ( 1 / 2 ) γ − μ ( 1 / 2 ) . R_{\mu}(u):=\frac{\gamma_{\mu}(1/2+u)\gamma_{-\mu}(1/2+u)}{\gamma_{\mu}(1/2)\gamma_{-\mu}(1/2)}. R μ ( u ) := γ μ ( 1/2 ) γ − μ ( 1/2 ) γ μ ( 1/2 + u ) γ − μ ( 1/2 + u ) . The factors of P J P_J P J in (21) cancel the numerator gamma poles crossed below. The gamma recurrence and (18) justify the contour shifts in (28) . For y ≥ 1 y\geq1 y ≥ 1 , move to Re ( u ) = B \operatorname{Re}(u)=B Re ( u ) = B . For 0 < y ≤ 1 0<y\leq1 0 < y ≤ 1 , move to Re ( u ) = − 1 / 4 \operatorname{Re}(u)=-1/4 Re ( u ) = − 1/4 , crossing only the pole at zero with residue 1 1 1 . Thus
U μ ( y ) = 1 + O J ( y 1 / 4 ( 1 + ∥ μ ∥ ) C ) , 0 < y ≤ 1 , U_{\mu}(y)=1+O_J\left(y^{1/4}(1+\|\mu\|)^C\right),\qquad0<y\leq1, U μ ( y ) = 1 + O J ( y 1/4 ( 1 + ∥ μ ∥ ) C ) , 0 < y ≤ 1 , and (30) follows.
For the test estimate, (22) gives
h ( μ ) G J ( u , μ ) R μ ( u ) = h ( μ ) P J ( 0 , μ ) e u 2 P J ( u , μ ) R μ ( u ) . h(\mu)G_J(u,\mu)R_{\mu}(u)=\frac{h(\mu)}{P_J(0,\mu)}e^{u^2}P_J(u,\mu)R_{\mu}(u). h ( μ ) G J ( u , μ ) R μ ( u ) = P J ( 0 , μ ) h ( μ ) e u 2 P J ( u , μ ) R μ ( u ) . The quotient in (27) is holomorphic and rapidly decreasing on Λ a 0 + 1 \Lambda_{a_0+1} Λ a 0 + 1 . For fixed Re ( u ) = c > 0 \operatorname{Re}(u)=c>0 Re ( u ) = c > 0 , the gamma recurrence makes the displayed multiplier holomorphic on this tube and preserves Weyl invariance and the zeros in (37) below. Applying (18) and Cauchy’s integral formula on this larger tube gives, for u = c + i v u=c+iv u = c + i v ,
sup μ ∈ Λ a 0 ∣ α ∣ ≤ k ( 1 + ∥ μ ∥ ) M ∣ ∂ μ α ( h ( μ ) G J ( u , μ ) R μ ( u ) ) ∣ ≪ a 0 , c , M , k , h , J ( 1 + ∣ v ∣ ) C e − v 2 + 3 π ∣ v ∣ / 2 . \sup_{\substack{\mu\in\Lambda_{a_0}\\|\alpha|\leq k}}(1+\|\mu\|)^M\left|\partial_{\mu}^{\alpha}\left(h(\mu)G_J(u,\mu)R_{\mu}(u)\right)\right|\ll_{a_0,c,M,k,h,J}(1+|v|)^C e^{-v^2+3\pi|v|/2}. μ ∈ Λ a 0 ∣ α ∣ ≤ k sup ( 1 + ∥ μ ∥ ) M ∂ μ α ( h ( μ ) G J ( u , μ ) R μ ( u ) ) ≪ a 0 , c , M , k , h , J ( 1 + ∣ v ∣ ) C e − v 2 + 3 π ∣ v ∣/2 .
Integrating on Re ( u ) = c \operatorname{Re}(u)=c Re ( u ) = c with y = t / p y=t/p y = t / p now gives (31) and H t ∈ H ~ ( a 0 ) H_t\in\widetilde{\mathcal{H}}(a_0) H t ∈ H ( a 0 ) .
The square of the standard L L L -function
For y > 0 y>0 y > 0 and c > 1 c>1 c > 1 , we define
V μ ± ( y ) : = 1 2 π i ∫ ( c ) y − u G J ( u , μ ) 2 ( γ ± μ ( 1 / 2 + u ) γ μ ( 1 / 2 ) ) 2 d u u . (32) \mathcal{V}^{\pm}_{\mu}(y):=\frac{1}{2\pi i}\int_{(c)}y^{-u}G_J(u,\mu)^2\left(\frac{\gamma_{\pm\mu}(1/2+u)}{\gamma_\mu(1/2)}\right)^2\frac{du}{u}. \tag*{(32)} V μ ± ( y ) := 2 π i 1 ∫ ( c ) y − u G J ( u , μ ) 2 ( γ μ ( 1/2 ) γ ± μ ( 1/2 + u ) ) 2 u d u . (32)
Proposition 2.8. For π ∈ B ( p ) \pi\in\mathcal{B}(p) π ∈ B ( p ) and L > 0 L>0 L > 0 , the square L ( 1 / 2 , π ) 2 L(1/2,\pi)^2 L ( 1/2 , π ) 2 satisfies the following asymmetric approximate functional equation
L ( 1 / 2 , π ) 2 = ∑ m , n ≥ 1 A π ( 1 , m ) A π ( 1 , n ) m n V μ π + ( m n p L 2 ) + ϵ ( π ) 2 ∑ m , n ≥ 1 A π ( 1 , m ) A π ( 1 , n ) ‾ m n V μ π − ( m n L 2 p ) . (33) L(1/2,\pi)^2=\sum_{m,n\geq1}\frac{A_\pi(1,m)A_\pi(1,n)}{\sqrt{mn}}\mathcal{V}^{+}_{\mu_\pi}\left(\frac{mn}{pL^2}\right)+\epsilon(\pi)^2\sum_{m,n\geq1}\frac{\overline{A_\pi(1,m)A_\pi(1,n)}}{\sqrt{mn}}\mathcal{V}^{-}_{\mu_\pi}\left(\frac{mnL^2}{p}\right). \tag*{(33)} L ( 1/2 , π ) 2 = m , n ≥ 1 ∑ mn A π ( 1 , m ) A π ( 1 , n ) V μ π + ( p L 2 mn ) + ϵ ( π ) 2 m , n ≥ 1 ∑ mn A π ( 1 , m ) A π ( 1 , n ) V μ π − ( p mn L 2 ) . (33)
Proof. For μ = μ π \mu=\mu_\pi μ = μ π and c > 1 c>1 c > 1 , consider
1 2 π i ∫ ( c ) Λ ( 1 / 2 + u , π ) 2 p 1 / 2 + u γ μ ( 1 / 2 ) 2 G J ( u , μ ) 2 L 2 u d u u . \frac{1}{2\pi i}\int_{(c)}\frac{\Lambda(1/2+u,\pi)^2}{p^{1/2+u}\gamma_\mu(1/2)^2}G_J(u,\mu)^2L^{2u}\frac{du}{u}. 2 π i 1 ∫ ( c ) p 1/2 + u γ μ ( 1/2 ) 2 Λ ( 1/2 + u , π ) 2 G J ( u , μ ) 2 L 2 u u d u .
The same contour argument as in the proof of Proposition 2.4 gives (33) .
For y > 0 y>0 y > 0 , define the corresponding spectral test functions by
K μ ± ( y ) : = h ( μ ) V μ ± ( y ) , K^{\pm}_{\mu}(y):=h(\mu)\mathcal{V}^{\pm}_{\mu}(y), K μ ± ( y ) := h ( μ ) V μ ± ( y ) ,
where h h h is the fixed test function in (26) . At zeros of P J ( 0 , μ ) P_J(0,\mu) P J ( 0 , μ ) , these products are defined by holomorphic continuation, as justified in Lemma 2.9. The following bounds apply to V μ ± ( y ) \mathcal{V}^{\pm}_{\mu}(y) V μ ± ( y ) and K μ ± ( y ) K^{\pm}_{\mu}(y) K μ ± ( y ) .
Lemma 2.9. On the unitary automorphic spectrum with ∣ Re ( μ j ) ∣ ≤ 5 / 14 \lvert\operatorname{Re}(\mu_j)\rvert\leq5/14 ∣ Re ( μ j )∣ ≤ 5/14 , for every B > 0 B>0 B > 0 and some C > 0 C>0 C > 0 , uniformly for y > 0 y>0 y > 0 ,
∣ V μ ± ( y ) ∣ ≪ B , J ( 1 + ∥ μ ∥ ) C ( 1 + y ) − B . (34) \left\lvert\mathcal{V}^{\pm}_{\mu}(y)\right\rvert\ll_{B,J}(1+\lVert\mu\rVert)^C(1+y)^{-B}. \tag*{(34)} V μ ± ( y ) ≪ B , J ( 1 + ∥ μ ∥ ) C ( 1 + y ) − B . (34) For 0 < y ≤ 1 0<y\leq1 0 < y ≤ 1 and every fixed 0 < c < 1 / 4 0<c<1/4 0 < c < 1/4 ,
V μ ± ( y ) = ( γ ± μ ( 1 / 2 ) γ μ ( 1 / 2 ) ) 2 + O J , c ( y c ( 1 + ∥ μ ∥ ) C ) . (35) \mathcal{V}^{\pm}_{\mu}(y)=\left(\frac{\gamma_{\pm\mu}(1/2)}{\gamma_\mu(1/2)}\right)^2+O_{J,c}\left(y^c(1+\lVert\mu\rVert)^C\right). \tag*{(35)} V μ ± ( y ) = ( γ μ ( 1/2 ) γ ± μ ( 1/2 ) ) 2 + O J , c ( y c ( 1 + ∥ μ ∥ ) C ) . (35) Let a 0 ≥ 10 a_0\geq10 a 0 ≥ 10 , J ≥ ⌈ a 0 ⌉ + 4 J\geq\lceil a_0\rceil+4 J ≥ ⌈ a 0 ⌉ + 4 , and let h h h be the fixed test function in (26) . For every y > 0 y>0 y > 0 , the functions μ ↦ K μ ± ( y ) \mu\mapsto K^{\pm}_{\mu}(y) μ ↦ K μ ± ( y ) belong to H ~ ( a 0 ) \widetilde{\mathcal{H}}(a_0) H ( a 0 ) . For every fixed c > 0 c>0 c > 0 , M ≥ 0 M\geq0 M ≥ 0 and integer k ≥ 0 k\geq0 k ≥ 0 , they satisfy
sup μ ∈ Λ a 0 ∣ α ∣ ≤ k ( 1 + ∥ μ ∥ ) M ∣ ∂ μ α K μ ± ( y ) ∣ ≪ h , J , a 0 , c , M , k y − c . (36) \sup_{\substack{\mu\in\Lambda_{a_0}\\ \lvert\alpha\rvert\leq k}}(1+\lVert\mu\rVert)^M\left\lvert\partial^\alpha_\mu K^{\pm}_{\mu}(y)\right\rvert\ll_{h,J,a_0,c,M,k}y^{-c}. \tag*{(36)} μ ∈ Λ a 0 ∣ α ∣ ≤ k sup ( 1 + ∥ μ ∥ ) M ∂ μ α K μ ± ( y ) ≪ h , J , a 0 , c , M , k y − c . (36)
Proof. By (21) , P J ( u , μ ) 2 P_J(u,\mu)^2 P J ( u , μ ) 2 cancels the numerator gamma poles with their full multiplicities in Re ( u ) ≥ − 2 \operatorname{Re}(u)\geq-2 Re ( u ) ≥ − 2 . The contour argument in the proof of Lemma 2.5 applies to the squared multiplier in (32) . Move to Re ( u ) = B + 1 \operatorname{Re}(u)=B+1 Re ( u ) = B + 1 for y ≥ 1 y\geq1 y ≥ 1 and to Re ( u ) = − c \operatorname{Re}(u)=-c Re ( u ) = − c for 0 < y ≤ 1 0<y\leq1 0 < y ≤ 1 , where 0 < c < 1 / 4 0<c<1/4 0 < c < 1/4 . The latter shift crosses only the pole at zero, with residue ( γ ± μ ( 1 / 2 ) / γ μ ( 1 / 2 ) ) 2 (\gamma_{\pm\mu}(1/2)/\gamma_\mu(1/2))^2 ( γ ± μ ( 1/2 ) / γ μ ( 1/2 ) ) 2 of modulus one. This proves (34) and (35) .
For the test estimate, write
h ( μ ) G J ( u , μ ) 2 ( γ ± μ ( 1 / 2 + u ) γ μ ( 1 / 2 ) ) 2 = h ( μ ) P J ( 0 , μ ) 2 e 2 u P J ( u , μ ) 2 γ ± μ ( 1 / 2 + u ) 2 γ μ ( 1 / 2 ) 2 . h(\mu)G_J(u,\mu)^2\left(\frac{\gamma_{\pm\mu}(1/2+u)}{\gamma_\mu(1/2)}\right)^2=\frac{h(\mu)}{P_J(0,\mu)^2}e^{2u}P_J(u,\mu)^2\frac{\gamma_{\pm\mu}(1/2+u)^2}{\gamma_\mu(1/2)^2}. h ( μ ) G J ( u , μ ) 2 ( γ μ ( 1/2 ) γ ± μ ( 1/2 + u ) ) 2 = P J ( 0 , μ ) 2 h ( μ ) e 2 u P J ( u , μ ) 2 γ μ ( 1/2 ) 2 γ ± μ ( 1/2 + u ) 2 .
The denominator P J ( 0 , μ ) 2 P_J(0,\mu)^2 P J ( 0 , μ ) 2 has zeros of order four on each of its half-integer coordinate hyperplanes. The factors cos 4 ( π μ j ) \cos^4(\pi\mu_j) cos 4 ( π μ j ) in h h h cancel these zeros with their full multiplicities. Thus h ( μ ) / P J ( 0 , μ ) 2 h(\mu)/P_J(0,\mu)^2 h ( μ ) / P J ( 0 , μ ) 2 is holomorphic and rapidly decreasing on finite tubes. The gamma recurrence and J ≥ ⌈ a 0 ⌉ + 4 J\ge\lceil a_0\rceil+4 J ≥ ⌈ a 0 ⌉ + 4 make the displayed multiplier holomorphic on Λ a 0 + 1 \Lambda_{a_0+1} Λ a 0 + 1 for Re ( u ) > 0 \operatorname{Re}(u)>0 Re ( u ) > 0 and preserve Weyl invariance and the zeros in (37) . Stirling’s formula and Cauchy estimates in the proof of Lemma 2.7 now apply with the decay factor e − 2 v 2 + 3 π ∣ v ∣ / 2 e^{-2v^2+3\pi|v|/2} e − 2 v 2 + 3 π ∣ v ∣/2 , where u = c + i v u=c+iv u = c + i v . Integrating on Re ( u ) = c \operatorname{Re}(u)=c Re ( u ) = c proves (36) . □
The GL(3) Kuznetsov formula at prime level
We use the finite Fourier coefficients and local bases of Blomer–Comtat [11 §§ 1.4, 4 and 5] , the archimedean parameter and kernel conventions of [17 §§ 1.3 and 4] . The harmonic weights use full finite adjoint values. The geometric side is given in [10 Theorem 6] . For the archimedean estimates we use [13 Lemmas 2.4–2.6] , in arXiv:2401.04037v2.
Kuznetsov test functions
Let Λ a = { μ ∈ C 3 : μ 1 + μ 2 + μ 3 = 0 , ∣ Re ( μ j ) ∣ ≤ a } \Lambda_a=\{\mu\in\mathbb{C}^3:\mu_1+\mu_2+\mu_3=0,\ |\operatorname{Re}(\mu_j)|\le a\} Λ a = { μ ∈ C 3 : μ 1 + μ 2 + μ 3 = 0 , ∣ Re ( μ j ) ∣ ≤ a } . Following [11 Definition 1] , H ~ ( a ) \widetilde{\mathcal H}(a) H ( a ) consists of the test functions that are Weyl group invariant, holomorphic on a neighborhood of Λ a \Lambda_a Λ a , rapidly decreasing there and vanishing on
μ i − μ j = n , n ∈ 2 Z + 1 , ∣ n ∣ ≤ a , 1 ≤ i < j ≤ 3. (37) \mu_i-\mu_j=n,\qquad n\in2\mathbb{Z}+1,\qquad|n|\le a,\qquad1\le i<j\le3. \tag*{(37)} μ i − μ j = n , n ∈ 2 Z + 1 , ∣ n ∣ ≤ a , 1 ≤ i < j ≤ 3. (37)
Rapid decrease means that sup Λ a ( 1 + ∥ μ ∥ ) M ∣ h ( μ ) ∣ < ∞ \sup_{\Lambda_a}(1+\|\mu\|)^M|h(\mu)|<\infty sup Λ a ( 1 + ∥ μ ∥ ) M ∣ h ( μ ) ∣ < ∞ for every M ≥ 0 M\ge0 M ≥ 0 . We use μ 1 , μ 2 \mu_1,\mu_2 μ 1 , μ 2 as independent coordinates and set μ 3 = − μ 1 − μ 2 \mu_3=-\mu_1-\mu_2 μ 3 = − μ 1 − μ 2 .
For α = ( α 1 , α 2 ) ∈ Z ≥ 0 2 \alpha=(\alpha_1,\alpha_2)\in\mathbb{Z}_{\ge0}^2 α = ( α 1 , α 2 ) ∈ Z ≥ 0 2 , write
∣ α ∣ = α 1 + α 2 , ∂ α f ( μ ) : = ∂ α 1 + α 2 f ( μ 1 , μ 2 , − μ 1 − μ 2 ) ∂ μ 1 α 1 ∂ μ 2 α 2 . |\alpha|=\alpha_1+\alpha_2,\qquad\partial^\alpha f(\mu):=\frac{\partial^{\alpha_1+\alpha_2}f(\mu_1,\mu_2,-\mu_1-\mu_2)}{\partial\mu_1^{\alpha_1}\partial\mu_2^{\alpha_2}}. ∣ α ∣ = α 1 + α 2 , ∂ α f ( μ ) := ∂ μ 1 α 1 ∂ μ 2 α 2 ∂ α 1 + α 2 f ( μ 1 , μ 2 , − μ 1 − μ 2 ) .
The fixed test function for the moments is given in (26) .
Archimedean kernels and integral transforms
On Re ( μ ) = 0 \operatorname{Re}(\mu)=0 Re ( μ ) = 0 the spectral density is [17 § 4]
spec ( μ ) : = ∏ ( i , j ) ∈ { ( 1 , 2 ) , ( 2 , 3 ) , ( 3 , 1 ) } ( μ i − μ j ) tan ( π 2 ( μ i − μ j ) ) . \operatorname{spec}(\mu):=\prod_{(i,j)\in\{(1,2),(2,3),(3,1)\}}(\mu_i-\mu_j)\tan\left(\frac{\pi}{2}(\mu_i-\mu_j)\right). spec ( μ ) := ( i , j ) ∈ {( 1 , 2 ) , ( 2 , 3 ) , ( 3 , 1 )} ∏ ( μ i − μ j ) tan ( 2 π ( μ i − μ j ) ) .
For the archimedean transforms, write μ j = i t j \mu_j=it_j μ j = i t j , t 3 = − t 1 − t 2 t_3=-t_1-t_2 t 3 = − t 1 − t 2 , and d μ = d μ 1 d μ 2 = − d t 1 d t 2 \mathrm{d}\mu=\mathrm{d}\mu_1\,\mathrm{d}\mu_2=-\mathrm{d}t_1\,\mathrm{d}t_2 d μ = d μ 1 d μ 2 = − d t 1 d t 2 . Then spec ( μ ) d μ \operatorname{spec}(\mu)\mathrm{d}\mu spec ( μ ) d μ is non-negative and is positive away from the Weyl chamber walls. Since ∣ tan ( i x ) ∣ = ∣ tanh x ∣ ≤ 1 |\tan(ix)|=|\tanh x|\le1 ∣ tan ( i x ) ∣ = ∣ tanh x ∣ ≤ 1 for x ∈ R x\in\mathbb{R} x ∈ R ,
∣ spec ( μ ) ∣ ≪ ( 1 + ∥ μ ∥ ) 3 ( Re ( μ ) = 0 ) . |\operatorname{spec}(\mu)|\ll(1+\|\mu\|)^3\qquad(\operatorname{Re}(\mu)=0). ∣ spec ( μ ) ∣ ≪ ( 1 + ∥ μ ∥ ) 3 ( Re ( μ ) = 0 ) .
Consequently, for every fixed C ≥ 0 C\ge0 C ≥ 0 and rapidly decreasing h h h ,
∫ Re ( μ ) = 0 ∣ h ( μ ) ∣ ( 1 + ∥ μ ∥ ) C spec ( μ ) d μ ≪ h , C 1. (38) \int_{\operatorname{Re}(\mu)=0}|h(\mu)|(1+\|\mu\|)^C\operatorname{spec}(\mu)\,\mathrm{d}\mu\ll_{h,C}1. \tag*{(38)} ∫ Re ( μ ) = 0 ∣ h ( μ ) ∣ ( 1 + ∥ μ ∥ ) C spec ( μ ) d μ ≪ h , C 1. (38)
The following kernels are 256 π 5 256\pi^5 256 π 5 times those in [17 Eqs. (10), (12), (16)–(18)] , consistently with the spectral constants and the diagonal term (52) . For ϵ ∈ { ± 1 } \epsilon\in\{\pm1\} ϵ ∈ { ± 1 } , we set
G ~ ϵ ( z , μ ) : = π − 3 z 2 { ∏ j = 1 3 Γ ( ( z − μ j ) / 2 ) Γ ( ( 1 − z + μ j ) / 2 ) + ϵ i ∏ j = 1 3 Γ ( ( 1 + z − μ j ) / 2 ) Γ ( ( 2 − z + μ j ) / 2 ) } . \widetilde{G}^{\epsilon}(z,\mu) := \frac{\pi^{-3z}}{2}\left\{\prod_{j=1}^{3}\frac{\Gamma((z-\mu_j)/2)}{\Gamma((1-z+\mu_j)/2)}+\epsilon i\prod_{j=1}^{3}\frac{\Gamma((1+z-\mu_j)/2)}{\Gamma((2-z+\mu_j)/2)}\right\}. G ϵ ( z , μ ) := 2 π − 3 z { j = 1 ∏ 3 Γ (( 1 − z + μ j ) /2 ) Γ (( z − μ j ) /2 ) + ϵ i j = 1 ∏ 3 Γ (( 2 − z + μ j ) /2 ) Γ (( 1 + z − μ j ) /2 ) } .
For ϵ 1 , ϵ 2 ∈ { ± 1 } \epsilon_1,\epsilon_2 \in\{\pm1\} ϵ 1 , ϵ 2 ∈ { ± 1 } and z = ( z 1 , z 2 ) z=(z_1,z_2) z = ( z 1 , z 2 ) , we set
G s y m ϵ 1 , ϵ 2 ( z , μ ) : = π 5 / 2 4 ∑ d 1 , d 2 , d 3 ∈ { 0 , 1 } d 3 ≡ d 1 + d 2 ( m o d 2 ) ϵ 1 d 1 ϵ 2 d 2 ( − 1 ) d 1 d 2 Γ ( ( 1 + d 3 − z 1 − z 2 ) / 2 ) Γ ( ( d 3 + z 1 + z 2 ) / 2 ) G_{\mathrm{sym}}^{\epsilon_1,\epsilon_2}(z,\mu) := \frac{\pi^{5/2}}{4}\sum_{\substack{d_1,d_2,d_3\in\{0,1\}\\d_3\equiv d_1+d_2\pmod2}}\epsilon_1^{d_1}\epsilon_2^{d_2}(-1)^{d_1d_2}\frac{\Gamma((1+d_3-z_1-z_2)/2)}{\Gamma((d_3+z_1+z_2)/2)} G sym ϵ 1 , ϵ 2 ( z , μ ) := 4 π 5/2 d 1 , d 2 , d 3 ∈ { 0 , 1 } d 3 ≡ d 1 + d 2 ( mod 2 ) ∑ ϵ 1 d 1 ϵ 2 d 2 ( − 1 ) d 1 d 2 Γ (( d 3 + z 1 + z 2 ) /2 ) Γ (( 1 + d 3 − z 1 − z 2 ) /2 )
× ∏ j = 1 3 Γ ( ( d 1 + z 1 − μ j ) / 2 ) Γ ( ( d 2 + z 2 + μ j ) / 2 ) Γ ( ( 1 + d 1 − z 1 + μ j ) / 2 ) Γ ( ( 1 + d 2 − z 2 − μ j ) / 2 ) . {}\times\prod_{j=1}^{3}\frac{\Gamma((d_1+z_1-\mu_j)/2)\Gamma((d_2+z_2+\mu_j)/2)}{\Gamma((1+d_1-z_1+\mu_j)/2)\Gamma((1+d_2-z_2-\mu_j)/2)}. × j = 1 ∏ 3 Γ (( 1 + d 1 − z 1 + μ j ) /2 ) Γ (( 1 + d 2 − z 2 − μ j ) /2 ) Γ (( d 1 + z 1 − μ j ) /2 ) Γ (( d 2 + z 2 + μ j ) /2 ) .
The kernels are defined by the Barnes integrals
K w 4 ( y ; μ ) : = ∫ C ∣ y ∣ − z G ~ sgn y ( z , μ ) d z 2 π i . K_{w_4}(y;\mu):=\int_{\mathcal{C}} |y|^{-z}\widetilde{G}^{\operatorname{sgn} y}(z,\mu)\frac{dz}{2\pi i}. K w 4 ( y ; μ ) := ∫ C ∣ y ∣ − z G sgn y ( z , μ ) 2 π i d z .
K w 6 s y m ( ( y 1 , y 2 ) ; μ ) : = ∫ C 1 ∫ C 2 ( π 2 ∣ y 1 ∣ ) − z 1 ( π 2 ∣ y 2 ∣ ) − z 2 G s y m sgn ( y 1 ) , sgn ( y 2 ) ( z , μ ) d z 1 d z 2 ( 2 π i ) 2 . K_{w_6}^{\mathrm{sym}}((y_1,y_2);\mu):=\int_{\mathcal{C}_1}\int_{\mathcal{C}_2}(\pi^2|y_1|)^{-z_1}(\pi^2|y_2|)^{-z_2}G_{\mathrm{sym}}^{\operatorname{sgn}(y_1),\operatorname{sgn}(y_2)}(z,\mu)\frac{dz_1dz_2}{(2\pi i)^2}. K w 6 sym (( y 1 , y 2 ) ; μ ) := ∫ C 1 ∫ C 2 ( π 2 ∣ y 1 ∣ ) − z 1 ( π 2 ∣ y 2 ∣ ) − z 2 G sym sgn ( y 1 ) , sgn ( y 2 ) ( z , μ ) ( 2 π i ) 2 d z 1 d z 2 .
We use the Barnes contours of [13 § 2.4] , separating the poles, avoiding poles on the imaginary axis, and deforming the tails for convergence. For y , y 1 , y 2 ≠ 0 y,y_1,y_2\ne0 y , y 1 , y 2 = 0 define
Φ w 4 , h ( y ) : = ∫ Re ( μ ) = 0 h ( μ ) K w 4 ( y ; μ ) spec ( μ ) d μ , \Phi_{w_4,h}(y):=\int_{\operatorname{Re}(\mu)=0}h(\mu)K_{w_4}(y;\mu)\operatorname{spec}(\mu)\,d\mu, Φ w 4 , h ( y ) := ∫ Re ( μ ) = 0 h ( μ ) K w 4 ( y ; μ ) spec ( μ ) d μ ,
Φ w 5 , h ( y ) : = ∫ Re ( μ ) = 0 h ( μ ) K w 4 ( − y ; − μ ) spec ( μ ) d μ , \Phi_{w_5,h}(y):=\int_{\operatorname{Re}(\mu)=0}h(\mu)K_{w_4}(-y;-\mu)\operatorname{spec}(\mu)\,d\mu, Φ w 5 , h ( y ) := ∫ Re ( μ ) = 0 h ( μ ) K w 4 ( − y ; − μ ) spec ( μ ) d μ ,
Φ w 6 , h ( y 1 , y 2 ) : = ∫ Re ( μ ) = 0 h ( μ ) K w 6 s y m ( ( y 1 , y 2 ) ; μ ) spec ( μ ) d μ . (39) \Phi_{w_6,h}(y_1,y_2):=\int_{\operatorname{Re}(\mu)=0}h(\mu)K_{w_6}^{\mathrm{sym}}((y_1,y_2);\mu)\operatorname{spec}(\mu)\,d\mu. \tag*{(39)} Φ w 6 , h ( y 1 , y 2 ) := ∫ Re ( μ ) = 0 h ( μ ) K w 6 sym (( y 1 , y 2 ) ; μ ) spec ( μ ) d μ . (39)
We omit the subscript h h h when the test function is fixed.
Spectral decomposition
For q ∈ { 1 , p } q\in\{1,p\} q ∈ { 1 , p } , let
X q = Γ 0 ( q ) \ S L 3 ( R ) / S O 3 . X_q=\Gamma_0(q)\backslash\mathrm{SL}_3(\mathbb{R})/\mathrm{SO}_3. X q = Γ 0 ( q ) \ SL 3 ( R ) / SO 3 .
We equip X q X_q X q with its invariant probability measure d ν q d\nu_q d ν q and use the inner product
⟨ f , g ⟩ q = ∫ X q f ( z ) g ( z ) ‾ d ν q ( z ) . \langle f,g\rangle_q=\int_{X_q}f(z)\overline{g(z)}\,d\nu_q(z). ⟨ f , g ⟩ q = ∫ X q f ( z ) g ( z ) d ν q ( z ) .
The spectral side of the non-degenerate Kuznetsov formula contains the cuspidal, minimal Eisenstein, and maximal Eisenstein contributions described in [11 Lemma 16] .
Cuspidal spectrum
Let B ( 1 ) \mathcal{B}(1) B ( 1 ) denote the set of level-one spherical cuspidal automorphic representations of G L 3 \mathrm{GL}_3 GL 3 with trivial central character. For each π ∈ B ( 1 ) \pi\in\mathcal{B}(1) π ∈ B ( 1 ) , choose the Hecke-normalized spherical eigenform f ∈ π f \in\pi f ∈ π with A π ( 1 , 1 ) = 1 A_\pi(1,1)=1 A π ( 1 , 1 ) = 1 . The space of oldforms generated by f f f has dimension three, with basis T 0 f , T 1 f , T 2 f T_0f,T_1f,T_2f T 0 f , T 1 f , T 2 f given in [11 Proposition 5] . Let A π ( i ) ( m , n ) A_\pi^{(i)}(m,n) A π ( i ) ( m , n ) be their Fourier coefficients respectively. We have
A π ( 0 ) ( m , n ) = A π ( m , n ) , A_\pi^{(0)}(m,n)=A_\pi(m,n), A π ( 0 ) ( m , n ) = A π ( m , n ) ,
A π ( 1 ) ( m , n ) = p ( A π ( m / p , p n ) + A π ( m , n / p ) ) , A_\pi^{(1)}(m,n)=\sqrt{p}\left(A_\pi(m/p,pn)+A_\pi(m,n/p)\right), A π ( 1 ) ( m , n ) = p ( A π ( m / p , p n ) + A π ( m , n / p ) ) ,
A π ( 2 ) ( m , n ) = p A π ( m / p , n ) . (40) A_\pi^{(2)}(m,n)=pA_\pi(m/p,n). \tag*{(40)} A π ( 2 ) ( m , n ) = p A π ( m / p , n ) . (40)
Coefficients with nonintegral indices are zero. Put λ = A π ( 1 , p ) \lambda=A_\pi(1,p) λ = A π ( 1 , p ) , so A π ( p , 1 ) = λ ‾ A_\pi(p,1)=\overline{\lambda} A π ( p , 1 ) = λ and set
a p : = p ( p + 1 ) P , b p : = p P . a_p:=\frac{\sqrt{p}(p+1)}{P},\qquad b_p:=\frac{p}{P}. a p := P p ( p + 1 ) , b p := P p .
By [11 Proposition 5] , we have
Q π : = ( ⟨ T i f , T j f ⟩ p ∥ f ∥ 2 ) 0 ≤ i , j ≤ 2 = ( 1 P a p λ b p λ ‾ a p λ ‾ 1 + b p ∣ λ ∣ 2 a p λ ‾ b p λ a p λ ‾ 1 ) . (41) Q_\pi:=\left(\frac{\langle T_i f,T_j f\rangle_p}{\lVert f\rVert^2}\right)_{0\leq i,j\leq2}
=
\begin{pmatrix}
\frac{1}{P} & a_p\lambda& b_p\overline{\lambda}\\
a_p\overline{\lambda} & 1+b_p\lvert\lambda\rvert^2 & a_p\overline{\lambda}\\
b_p\lambda& a_p\overline{\lambda} & 1
\end{pmatrix}. \tag*{(41)} Q π := ( ∥ f ∥ 2 ⟨ T i f , T j f ⟩ p ) 0 ≤ i , j ≤ 2 = P 1 a p λ b p λ a p λ 1 + b p ∣ λ ∣ 2 a p λ b p λ a p λ 1 . (41)
Gram–Schmidt orthogonalization gives an orthonormal basis S 0 f , S 1 f , S 2 f S_0f,S_1f,S_2f S 0 f , S 1 f , S 2 f as in [11 Corollary 4] . Let B π ( j ) ∥ f ∥ − 1 B_\pi^{(j)}\lVert f\rVert^{-1} B π ( j ) ∥ f ∥ − 1 be the Fourier coefficients of S 0 f S_0f S 0 f , S 1 f S_1f S 1 f and S 2 f S_2f S 2 f respectively. Then
B π ( 0 ) ( m , n ) = A π ( 0 ) ( m , n ) , B_\pi^{(0)}(m,n)=A_\pi^{(0)}(m,n), B π ( 0 ) ( m , n ) = A π ( 0 ) ( m , n ) ,
B π ( j ) ( m , n ) = ( 1 + c j j ) A π ( j ) ( m , n ) + ∑ i = 0 j − 1 c j i A π ( i ) ( m , n ) , j = 1 , 2. (42) B_\pi^{(j)}(m,n)=(1+c_{jj})A_\pi^{(j)}(m,n)+\sum_{i=0}^{j-1}c_{ji}A_\pi^{(i)}(m,n),\qquad j=1,2. \tag*{(42)} B π ( j ) ( m , n ) = ( 1 + c j j ) A π ( j ) ( m , n ) + i = 0 ∑ j − 1 c j i A π ( i ) ( m , n ) , j = 1 , 2. (42)
Here c j i ≪ p θ 3 − 1 / 2 c_{ji}\ll p^{\theta_3-1/2} c j i ≪ p θ 3 − 1/2 and Q π = I + O ( p θ 3 − 1 / 2 ) Q_\pi=I+O(p^{\theta_3-1/2}) Q π = I + O ( p θ 3 − 1/2 ) , with θ 3 = 5 / 14 \theta_3=5/14 θ 3 = 5/14 .
Let π ∈ B ( p ) \pi\in\mathcal{B}(p) π ∈ B ( p ) , and let F π F_\pi F π be the corresponding normalized newform. Define
D π ( u ) : = ∑ m 1 , m 2 ≥ 1 ∣ A π ( m 1 , m 2 ) ∣ 2 m 1 2 u m 2 2 u . D_\pi(u):=\sum_{m_1,m_2\geq1}\frac{\lvert A_\pi(m_1,m_2)\rvert^2}{m_1^{2u}m_2^{2u}}. D π ( u ) := m 1 , m 2 ≥ 1 ∑ m 1 2 u m 2 2 u ∣ A π ( m 1 , m 2 ) ∣ 2 .
At primes ℓ ≠ p \ell\ne p ℓ = p , (8) gives
D π , ℓ ( u ) = L ℓ ( u , π ℓ × π ~ ℓ ) ζ ℓ ( 3 u ) . D_{\pi,\ell}(u)=\frac{L_\ell(u,\pi_\ell\times\widetilde{\pi}_\ell)}{\zeta_\ell(3u)}. D π , ℓ ( u ) = ζ ℓ ( 3 u ) L ℓ ( u , π ℓ × π ℓ ) .
At ℓ = p \ell=p ℓ = p , Lemmas 2.1 and 2.2 give
D π , p ( u ) = L p ( u , π p , Ad ) . (43) D_{\pi,p}(u)=L_p(u,\pi_p,\operatorname{Ad}). \tag*{(43)} D π , p ( u ) = L p ( u , π p , Ad ) . (43)
It follows that
D π ( u ) = L ( u , π × π ~ ) ζ ( 3 u ) 1 − p − u 1 − p − 3 u . (44) D_\pi(u)=\frac{L(u,\pi\times\widetilde{\pi})}{\zeta(3u)}\frac{1-p^{-u}}{1-p^{-3u}}. \tag*{(44)} D π ( u ) = ζ ( 3 u ) L ( u , π × π ) 1 − p − 3 u 1 − p − u . (44)
Since
L ( u , π × π ~ ) = ζ ( u ) L ( u , π , Ad ) , L(u,\pi\times\widetilde{\pi})=\zeta(u)L(u,\pi,\operatorname{Ad}), L ( u , π × π ) = ζ ( u ) L ( u , π , Ad ) ,
taking the residue at u = 1 u=1 u = 1 gives
Res u = 1 D π ( u ) = 1 − p − 1 1 − p − 3 L ( 1 , π , Ad ) ζ ( 3 ) = χ p − 1 L ( 1 , π , Ad ) ζ ( 3 ) . (45) \operatorname{Res}_{u=1}D_{\pi}(u)=\frac{1-p^{-1}}{1-p^{-3}}\frac{L(1,\pi,\operatorname{Ad})}{\zeta(3)}
=\chi_p^{-1}\frac{L(1,\pi,\operatorname{Ad})}{\zeta(3)}. \tag*{(45)} Res u = 1 D π ( u ) = 1 − p − 3 1 − p − 1 ζ ( 3 ) L ( 1 , π , Ad ) = χ p − 1 ζ ( 3 ) L ( 1 , π , Ad ) . (45)
By the Rankin–Selberg norm formula in [37 Section 4, proof of Theorem 4.1] , together with the essential-vector pairing formula in [30 Theorem 3.5] , we obtain
∥ F π ∥ 2 = C 3 , ∞ ( μ π ) χ p − 1 L ( 1 , π , Ad ) . (46) \lVert F_{\pi}\rVert^2=C_{3,\infty}(\mu_{\pi})\chi_p^{-1}L(1,\pi,\operatorname{Ad}). \tag*{(46)} ∥ F π ∥ 2 = C 3 , ∞ ( μ π ) χ p − 1 L ( 1 , π , Ad ) . (46)
Here L ( 1 , π , Ad ) L(1,\pi,\operatorname{Ad}) L ( 1 , π , Ad ) is the full finite adjoint value, C 3 , ∞ ( μ π ) C_{3,\infty}(\mu_{\pi}) C 3 , ∞ ( μ π ) is the factor occurring in the level-one norm formula and χ p : = P p 2 = 1 + 1 p + 1 p 2 \chi_p:=\frac{P}{p^2}=1+\frac{1}{p}+\frac{1}{p^2} χ p := p 2 P = 1 + p 1 + p 2 1 . After applying the archimedean normalization fixed above, taking the reciprocal of (46) gives the arithmetic weight
c c u s p χ p L ( 1 , π , Ad ) \frac{c_{\mathrm{cusp}}\chi_p}{L(1,\pi,\operatorname{Ad})} L ( 1 , π , Ad ) c cusp χ p
in the cuspidal new spectrum.
For π ∈ B ( 1 ) \pi\in\mathcal{B}(1) π ∈ B ( 1 ) , after orthogonalizing the three oldvectors, their common arithmetic weight is
c c u s p L ( 1 , π , Ad ) . \frac{c_{\mathrm{cusp}}}{L(1,\pi,\operatorname{Ad})}. L ( 1 , π , Ad ) c cusp .
The minimal Eisenstein spectrum
Now we consider the minimal Eisenstein series. Let s = ( s 1 , s 2 , s 3 ) ∈ ( i R ) 3 \mathbf{s}=(s_1,s_2,s_3)\in(i\mathbb{R})^3 s = ( s 1 , s 2 , s 3 ) ∈ ( i R ) 3 , s 3 = − s 1 − s 2 s_3=-s_1-s_2 s 3 = − s 1 − s 2 . The local Langlands parameters are
α q = q s 1 , β q = q s 2 , γ q = q s 3 . (47) \alpha_q=q^{s_1},\qquad\beta_q=q^{s_2},\qquad\gamma_q=q^{s_3}. \tag*{(47)} α q = q s 1 , β q = q s 2 , γ q = q s 3 . (47)
Define A s ( m , n ) A_{\mathbf{s}}(m,n) A s ( m , n ) by (8) at each prime q q q and by multiplicativity. In particular,
A s ( 1 , n ) = ∑ a b c = n a s 1 b s 2 c s 3 . A_{\mathbf{s}}(1,n)=\sum_{abc=n}a^{s_1}b^{s_2}c^{s_3}. A s ( 1 , n ) = ab c = n ∑ a s 1 b s 2 c s 3 .
For each fixed inducing parameter s \mathbf{s} s , the local space of level-p p p invariant sections has dimension three [11 Proposition 6] . Thus the minimal Eisenstein spectrum is parameterized by j ∈ { 0 , 1 , 2 } j\in\{0,1,2\} j ∈ { 0 , 1 , 2 } and ( s 1 , s 2 ) ∈ ( i R ) 2 (s_1,s_2)\in(i\mathbb{R})^2 ( s 1 , s 2 ) ∈ ( i R ) 2 , with s 3 = − s 1 − s 2 s_3=-s_1-s_2 s 3 = − s 1 − s 2 . We define A s ( j ) A_{\mathbf{s}}^{(j)} A s ( j ) by (40) and B s ( j ) B_{\mathbf{s}}^{(j)} B s ( j ) by (42) . The orthogonalization coefficients satisfy c j i ≪ p − 1 / 2 c_{ji}\ll p^{-1/2} c j i ≪ p − 1/2 by the same proposition.
In [11 Lemma 16] , the arithmetic weight is
c min ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 . \frac{c_{\min}}{\prod_{1\leq i<k\leq3}\left|\zeta(1+s_i-s_k)\right|^2}. ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 c m i n .
The maximal Eisenstein spectrum
For q ∈ { 1 , p } q\in\{1,p\} q ∈ { 1 , p } , let B 2 ( q ) B_2(q) B 2 ( q ) denote the set of Hecke-normalized even Maass newforms of level q q q and trivial central character. For ϕ ∈ B 2 ( q ) \phi\in B_2(q) ϕ ∈ B 2 ( q ) , write λ ϕ ( n ) \lambda_{\phi}(n) λ ϕ ( n ) for its Hecke eigenvalues, with λ ϕ ( 1 ) = 1 \lambda_{\phi}(1)=1 λ ϕ ( 1 ) = 1 , and ( τ ϕ , − τ ϕ ) (\tau_{\phi},-\tau_{\phi}) ( τ ϕ , − τ ϕ ) for its archimedean parameters.
By [8 Theorem 1] , we have
τ ϕ = i t ϕ ∈ i R ∪ [ − θ 2 , θ 2 ] , θ 2 = 7 64 . \tau_{\phi}=it_{\phi}\in i\mathbb{R}\cup[-\theta_2,\theta_2],\qquad\theta_2=\frac{7}{64}. τ ϕ = i t ϕ ∈ i R ∪ [ − θ 2 , θ 2 ] , θ 2 = 64 7 .
Its gamma factor is Γ R ( w + τ ϕ ) Γ R ( w − τ ϕ ) \Gamma_{\mathbb{R}}(w+\tau_\phi)\Gamma_{\mathbb{R}}(w-\tau_\phi) Γ R ( w + τ ϕ ) Γ R ( w − τ ϕ ) . For s ∈ i R s\in i\mathbb{R} s ∈ i R , the induced spectral parameters are μ = ( s + τ ϕ , s − τ ϕ , − 2 s ) \mu=(s+\tau_\phi,s-\tau_\phi,-2s) μ = ( s + τ ϕ , s − τ ϕ , − 2 s ) . At each prime ℓ \ell ℓ , let ( α ϕ , ℓ , β ϕ , ℓ ) (\alpha_{\phi,\ell},\beta_{\phi,\ell}) ( α ϕ , ℓ , β ϕ , ℓ ) be the inverse roots of the local standard L L L -factor of ϕ ℓ \phi_\ell ϕ ℓ , allowing a zero inverse root at a ramified prime. For ϕ ∈ B 2 ( p ) \phi\in B_2(p) ϕ ∈ B 2 ( p ) , we have
( α ϕ , p , β ϕ , p ) = ( λ ϕ ( p ) , 0 ) , λ ϕ ( p ) = ± p − 1 / 2 . (\alpha_{\phi,p},\beta_{\phi,p})=(\lambda_\phi(p),0),\qquad\lambda_\phi(p)=\pm p^{-1/2}. ( α ϕ , p , β ϕ , p ) = ( λ ϕ ( p ) , 0 ) , λ ϕ ( p ) = ± p − 1/2 .
As in [11 Eq. (1.9)] , we define the inverse roots of the induced local standard L L L -factor by
( α ℓ , β ℓ , γ ℓ ) = ( ℓ s α ϕ , ℓ , ℓ s β ϕ , ℓ , ℓ − 2 s ) . (\alpha_\ell,\beta_\ell,\gamma_\ell)=(\ell^s\alpha_{\phi,\ell},\ell^s\beta_{\phi,\ell},\ell^{-2s}). ( α ℓ , β ℓ , γ ℓ ) = ( ℓ s α ϕ , ℓ , ℓ s β ϕ , ℓ , ℓ − 2 s ) .
Then we define A ϕ , s ( m , n ) A_{\phi,s}(m,n) A ϕ , s ( m , n ) by (8) and by multiplicativity. In particular,
A ϕ , s ( 1 , n ) = ∑ a b = n λ ϕ ( a ) a s b − 2 s . A_{\phi,s}(1,n)=\sum_{ab=n}\lambda_\phi(a)a^s b^{-2s}. A ϕ , s ( 1 , n ) = ab = n ∑ λ ϕ ( a ) a s b − 2 s .
If ϕ ∈ B 2 ( 1 ) \phi\in B_2(1) ϕ ∈ B 2 ( 1 ) , we define the Fourier coefficients A ϕ , s ( j ) A_{\phi,s}^{(j)} A ϕ , s ( j ) and B ϕ , s ( j ) B_{\phi,s}^{(j)} B ϕ , s ( j ) obtained from (40) and (42) . By [11 Proposition 7] , the space of maximal generic unitary Eisenstein series is parameterized by
( ϕ , s ) (\phi,s) ( ϕ , s ) with ϕ ∈ B 2 ( p ) \phi\in B_2(p) ϕ ∈ B 2 ( p ) . The associated Eisenstein series have Fourier coefficients A ϕ , s A_{\phi,s} A ϕ , s .
( j , ϕ , s ) (j,\phi,s) ( j , ϕ , s ) with j ∈ { 0 , 1 , 2 } j\in\{0,1,2\} j ∈ { 0 , 1 , 2 } and ϕ ∈ B 2 ( 1 ) \phi\in B_2(1) ϕ ∈ B 2 ( 1 ) . The associated Eisenstein series have Fourier coefficients B ϕ , s ( j ) B_{\phi,s}^{(j)} B ϕ , s ( j ) with the constant c j i ≪ p θ 2 − 1 / 2 c_{ji}\ll p^{\theta_2-1/2} c j i ≪ p θ 2 − 1/2 .
For ϕ ∈ B 2 ( p ) \phi\in B_2(p) ϕ ∈ B 2 ( p ) , the norm here is taken with respect to probability measure on Γ 0 ( p ) \ H \Gamma_0(p)\backslash\mathbb{H} Γ 0 ( p ) \ H . Its volume for the hyperbolic measure d x d y / y 2 \mathrm{d}x\,\mathrm{d}y/y^2 d x d y / y 2 is π ( p + 1 ) / 3 \pi(p+1)/3 π ( p + 1 ) /3 . In the same archimedean Fourier normalization at levels one and p p p , the Rankin–Selberg norm identity [26 Lemma 4.6, Eq. (4.8)] therefore gives
∥ ϕ ∥ 2 = 3 π ( p + 1 ) ∫ Γ 0 ( p ) \ H ∣ ϕ ( x + i y ) ∣ 2 d x d y y 2 = C 2 , ∞ ( τ ϕ ) p p + 1 L ( 1 , ϕ , Ad ) . \begin{aligned}
\|\phi\|^2=\frac{3}{\pi(p+1)}\int_{\Gamma_0(p)\backslash\mathbb{H}}|\phi(x+iy)|^2\frac{\mathrm{d}x\,\mathrm{d}y}{y^2} \\
&=C_{2,\infty}(\tau_\phi)\frac{p}{p+1}L(1,\phi,\operatorname{Ad}).
\end{aligned} ∥ ϕ ∥ 2 = π ( p + 1 ) 3 ∫ Γ 0 ( p ) \ H ∣ ϕ ( x + i y ) ∣ 2 y 2 d x d y = C 2 , ∞ ( τ ϕ ) p + 1 p L ( 1 , ϕ , Ad ) .
where C 2 , ∞ ( τ ϕ ) C_{2,\infty}(\tau_\phi) C 2 , ∞ ( τ ϕ ) is the same archimedean factor as in the level-one formula. Let Φ ϕ , s \Phi_{\phi,s} Φ ϕ , s be the corresponding inducing section, normalized so that its component on the nontrivial double coset is ϕ \phi ϕ . Hence [11 Eq. (3.6) and Lemma 10] gives
∥ Φ ϕ , s ∥ 2 = p ( p + 1 ) P ∥ ϕ ∥ 2 = C 2 , ∞ ( τ ϕ ) p ( p + 1 ) P p p + 1 L ( 1 , ϕ , Ad ) = C 2 , ∞ ( τ ϕ ) χ p − 1 L ( 1 , ϕ , Ad ) . \begin{aligned}
\|\Phi_{\phi,s}\|^2=\frac{p(p+1)}{P}\|\phi\|^2 \\
&=C_{2,\infty}(\tau_\phi)\frac{p(p+1)}{P}\frac{p}{p+1}L(1,\phi,\operatorname{Ad}) \\
&=C_{2,\infty}(\tau_\phi)\chi_p^{-1}L(1,\phi,\operatorname{Ad}).
\end{aligned} ∥ Φ ϕ , s ∥ 2 = P p ( p + 1 ) ∥ ϕ ∥ 2 = C 2 , ∞ ( τ ϕ ) P p ( p + 1 ) p + 1 p L ( 1 , ϕ , Ad ) = C 2 , ∞ ( τ ϕ ) χ p − 1 L ( 1 , ϕ , Ad ) .
Consequently, the maximal Eisenstein new spectrum carries the arithmetic weight
c max χ p L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 s , ϕ ) ∣ 2 . \frac{c_{\max}\chi_p}{L(1,\phi,\operatorname{Ad})|L(1+3s,\phi)|^2}. L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 s , ϕ ) ∣ 2 c m a x χ p .
Kloosterman sums and level conditions
For D 1 ∣ D 2 D_1\mid D_2 D 1 ∣ D 2 , as in [11 § 5.2] , we define
S ~ ( n 1 , n 2 , m 1 , D 1 , D 2 ) : = ∑ C 1 ( m o d D 1 ) , C 2 ( m o d D 2 ) ( C 1 , D 1 ) = ( C 2 , D 2 / D 1 ) = 1 e ( n 2 C ‾ 1 C 2 + n 1 C 1 D 1 + m 1 C ‾ 2 D 2 / D 1 ) . (48) \widetilde{S}(n_1,n_2,m_1,D_1,D_2):=
\sum_{\substack{C_1\ (\mathrm{mod}\ D_1),\ C_2\ (\mathrm{mod}\ D_2)\\(C_1,D_1)=(C_2,D_2/D_1)=1}}
e\left(\frac{n_2\overline{C}_1C_2+n_1C_1}{D_1}+\frac{m_1\overline{C}_2}{D_2/D_1}\right). \tag*{(48)} S ( n 1 , n 2 , m 1 , D 1 , D 2 ) := C 1 ( mod D 1 ) , C 2 ( mod D 2 ) ( C 1 , D 1 ) = ( C 2 , D 2 / D 1 ) = 1 ∑ e ( D 1 n 2 C 1 C 2 + n 1 C 1 + D 2 / D 1 m 1 C 2 ) . (48)
Here C ‾ 1 \overline{C}_1 C 1 is an inverse modulo D 1 D_1 D 1 , and C ‾ 2 \overline{C}_2 C 2 is an inverse modulo D 2 / D 1 D_2/D_1 D 2 / D 1 . For p ∣ D 1 , D 2 p\mid D_1,D_2 p ∣ D 1 , D 2 , the sum for the long Weyl element is
S ( p ) ( n 1 , n 2 , m 1 , m 2 , D 1 , D 2 ) : = ∑ B 1 , C 1 ( mod D 1 ) , B 2 , C 2 ( mod D 2 ) D 1 C 2 + B 1 B 2 + D 2 C 1 ≡ 0 ( mod D 1 D 2 ) ( B j , C j , D j ) = 1 ( j = 1 , 2 ) , p ∣ B 1 e ( n 1 B 1 + m 1 ( Y 1 D 2 − Z 1 B 2 ) D 1 + n 2 B 2 + m 2 ( Y 2 D 1 − Z 2 B 1 ) D 2 ) . (49) S^{(p)}(n_1,n_2,m_1,m_2,D_1,D_2)
:=\sum_{\substack{B_1,C_1\ (\operatorname{mod} D_1),\ B_2,C_2\ (\operatorname{mod} D_2)\\ D_1C_2+B_1B_2+D_2C_1\equiv0\ (\operatorname{mod} D_1D_2)\\ (B_j,C_j,D_j)=1\ (j=1,2),\ p\mid B_1}}
e\left(\frac{n_1B_1+m_1(Y_1D_2-Z_1B_2)}{D_1}+\frac{n_2B_2+m_2(Y_2D_1-Z_2B_1)}{D_2}\right). \tag*{(49)} S ( p ) ( n 1 , n 2 , m 1 , m 2 , D 1 , D 2 ) := B 1 , C 1 ( mod D 1 ) , B 2 , C 2 ( mod D 2 ) D 1 C 2 + B 1 B 2 + D 2 C 1 ≡ 0 ( mod D 1 D 2 ) ( B j , C j , D j ) = 1 ( j = 1 , 2 ) , p ∣ B 1 ∑ e ( D 1 n 1 B 1 + m 1 ( Y 1 D 2 − Z 1 B 2 ) + D 2 n 2 B 2 + m 2 ( Y 2 D 1 − Z 2 B 1 ) ) . (49)
where Y j B j + Z j C j ≡ 1 ( m o d D j ) Y_jB_j+Z_jC_j\equiv1\pmod{D_j} Y j B j + Z j C j ≡ 1 ( mod D j ) . The sum is independent of the permissible choices of Y j , Z j Y_j,Z_j Y j , Z j . For the sum in (48) , the required bound is [11 Lemma 14(d)] :
S ~ ( n 1 , n 2 , m 1 , D 1 , D 2 ) ≪ ε ( ( m 1 , D 2 / D 1 ) D 1 2 , ( n 1 , n 2 , D 1 ) D 2 ) ( D 1 D 2 ) ε . (50) \widetilde{S}(n_1,n_2,m_1,D_1,D_2)\ll_\varepsilon\left((m_1,D_2/D_1)D_1^2,(n_1,n_2,D_1)D_2\right)(D_1D_2)^\varepsilon. \tag*{(50)} S ( n 1 , n 2 , m 1 , D 1 , D 2 ) ≪ ε ( ( m 1 , D 2 / D 1 ) D 1 2 , ( n 1 , n 2 , D 1 ) D 2 ) ( D 1 D 2 ) ε . (50)
We recall the Ramanujan sum
r p ( a ) = ∑ x ( mod p ) ⋆ e ( a x / p ) . r_p(a)=\sum_{x\ (\operatorname{mod} p)}^\star e(ax/p). r p ( a ) = x ( mod p ) ∑ ⋆ e ( a x / p ) .
and the classical Kloosterman sum
S ( a , b , c ) = ∑ x ( mod c ) ⋆ e ( ( a x + b x ‾ ) / c ) . S(a,b,c)=\sum_{x\ (\operatorname{mod} c)}^\star e((ax+b\overline{x})/c). S ( a , b , c ) = x ( mod c ) ∑ ⋆ e (( a x + b x ) / c ) .
The first two identities below are [11 Lemmas 14(b) and 15(a)] . The third corrects the parameter order in [11 Lemma 15(b)] .
S ( p ) ( n 1 , n 2 , m 1 , m 2 , p , p ) = p − 1 + r p ( m 1 ) r p ( n 2 ) , S^{(p)}(n_1,n_2,m_1,m_2,p,p)=p-1+r_p(m_1)r_p(n_2), S ( p ) ( n 1 , n 2 , m 1 , m 2 , p , p ) = p − 1 + r p ( m 1 ) r p ( n 2 ) ,
S ( p ) ( n 1 , n 2 , m 1 , m 2 , p , p 2 ) = r p ( m 1 ) S ( n 2 , p m 2 , p 2 ) , S^{(p)}(n_1,n_2,m_1,m_2,p,p^2)=r_p(m_1)S(n_2,pm_2,p^2), S ( p ) ( n 1 , n 2 , m 1 , m 2 , p , p 2 ) = r p ( m 1 ) S ( n 2 , p m 2 , p 2 ) ,
S ( p ) ( n 1 , n 2 , m 1 , m 2 , p 2 , p ) = r p ( n 2 ) S ( m 1 , p n 1 , p 2 ) . S^{(p)}(n_1,n_2,m_1,m_2,p^2,p)=r_p(n_2)S(m_1,pn_1,p^2). S ( p ) ( n 1 , n 2 , m 1 , m 2 , p 2 , p ) = r p ( n 2 ) S ( m 1 , p n 1 , p 2 ) .
To verify the last identity, we set D 1 = p 2 D_1=p^2 D 1 = p 2 , D 2 = p D_2=p D 2 = p and write B 1 = p b B_1=pb B 1 = p b . The defining congruence becomes
p C 2 + b B 2 + C 1 ≡ 0 ( m o d p 2 ) . pC_2+bB_2+C_1\equiv0\pmod{p^2}. p C 2 + b B 2 + C 1 ≡ 0 ( mod p 2 ) .
Primitivity forces C 1 C_1 C 1 , b b b and B 2 B_2 B 2 to be units modulo p p p . Choose Y 1 = 0 Y_1=0 Y 1 = 0 , Z 1 = C 1 − 1 ( m o d p 2 ) Z_1=C_1^{-1}\pmod{p^2} Z 1 = C 1 − 1 ( mod p 2 ) , and set x = − B 2 C 1 − 1 ( m o d p 2 ) x=-B_2C_1^{-1}\pmod{p^2} x = − B 2 C 1 − 1 ( mod p 2 ) . Then
x − 1 ≡ b + p C 2 B 2 − 1 ( m o d p 2 ) , B 1 ≡ p x − 1 ( m o d p 2 ) . x^{-1}\equiv b+pC_2B_2^{-1}\pmod{p^2},\qquad B_1\equiv px^{-1}\pmod{p^2}. x − 1 ≡ b + p C 2 B 2 − 1 ( mod p 2 ) , B 1 ≡ p x − 1 ( mod p 2 ) .
For each fixed unit B 2 ( m o d p ) B_2\pmod p B 2 ( mod p ) , the map ( b , C 2 ) ↦ x (b,C_2)\mapsto x ( b , C 2 ) ↦ x is a bijection onto ( Z / p 2 Z ) × (\mathbb{Z}/p^2\mathbb{Z})^\times ( Z / p 2 Z ) × . The entire m 2 m_2 m 2 part of the phase is an integer because p ∣ B 1 p\mid B_1 p ∣ B 1 . The remaining sums therefore separate as
S ( p ) ( n 1 , n 2 , m 1 , m 2 , p 2 , p ) = ∑ B 2 ( mod p ) ⋆ e ( n 2 B 2 / p ) ∑ x ( mod p 2 ) ⋆ e ( m 1 x + p n 1 x − 1 p 2 ) = r p ( n 2 ) S ( m 1 , p n 1 , p 2 ) . \begin{aligned}
S^{(p)}(n_1,n_2,m_1,m_2,p^2,p)
&=\sum_{B_2\ (\operatorname{mod} p)}^\star e(n_2B_2/p)
\sum_{x\ (\operatorname{mod} p^2)}^\star e\left(\frac{m_1x+pn_1x^{-1}}{p^2}\right)\\
&=r_p(n_2)S(m_1,pn_1,p^2).
\end{aligned} S ( p ) ( n 1 , n 2 , m 1 , m 2 , p 2 , p ) = B 2 ( mod p ) ∑ ⋆ e ( n 2 B 2 / p ) x ( mod p 2 ) ∑ ⋆ e ( p 2 m 1 x + p n 1 x − 1 ) = r p ( n 2 ) S ( m 1 , p n 1 , p 2 ) .
The complete Kuznetsov formula at prime level
We use the corrected kernel constants in the published version of Buttcane–Zhou [17 Section 4] . With the μ \mu μ -coordinates and the contour measure d μ = d μ 1 d μ 2 \mathrm{d}\mu=\mathrm{d}\mu_1\mathrm{d}\mu_2 d μ = d μ 1 d μ 2 fixed above, the level-one diagonal term in that normalization is
δ m 1 , n 1 δ m 2 , n 2 256 π 5 ∫ Re μ = 0 h ( μ ) spec ( μ ) d μ . \frac{\delta_{m_1,n_1}\delta_{m_2,n_2}}{256\pi^5}\int_{\operatorname{Re}\mu=0}h(\mu)\operatorname{spec}(\mu)\,\mathrm{d}\mu. 256 π 5 δ m 1 , n 1 δ m 2 , n 2 ∫ Re μ = 0 h ( μ ) spec ( μ ) d μ .
We multiply the level-one Kuznetsov formula by 256 π 5 256\pi^5 256 π 5 , so that the coefficient of the diagonal integral becomes one. The same scalar is incorporated into the spectral constants c c u s p c_{\mathrm{cusp}} c cusp , c min c_{\min} c m i n , c max c_{\max} c m a x and into the kernels K w 4 K_{w_4} K w 4 and K w 6 s y m K_{w_6}^{\mathrm{sym}} K w 6 sym . The kernel formulas in § 3.2 are written in this rescaled normalization, and the integral transforms Φ w 4 , h \Phi_{w_4,h} Φ w 4 , h , Φ w 5 , h \Phi_{w_5,h} Φ w 5 , h and Φ w 6 , h \Phi_{w_6,h} Φ w 6 , h are defined using these kernels without any additional rescaling.
Fix positive integers m 1 , m 2 , n 1 , n 2 m_1,m_2,n_1,n_2 m 1 , m 2 , n 1 , n 2 . With the coefficients and normalizations above, define the five spectral terms by
C n e w : = c c u s p ∑ π ∈ B ( p ) h ( μ π ) A π ( m 1 , m 2 ) ‾ A π ( n 1 , n 2 ) L ( 1 , π , Ad ) , \mathcal{C}_{\mathrm{new}}:=c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(\mu_\pi)\overline{A_\pi(m_1,m_2)}A_\pi(n_1,n_2)}{L(1,\pi,\operatorname{Ad})}, C new := c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) A π ( m 1 , m 2 ) A π ( n 1 , n 2 ) ,
C o l d : = c c u s p ∑ π ∈ B ( 1 ) ∑ j = 0 2 h ( μ π ) B π ( j ) ( m 1 , m 2 ) ‾ B π ( j ) ( n 1 , n 2 ) L ( 1 , π , Ad ) , \mathcal{C}_{\mathrm{old}}:=c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(1)}\sum_{j=0}^{2}\frac{h(\mu_\pi)\overline{B_\pi^{(j)}(m_1,m_2)}B_\pi^{(j)}(n_1,n_2)}{L(1,\pi,\operatorname{Ad})}, C old := c cusp π ∈ B ( 1 ) ∑ j = 0 ∑ 2 L ( 1 , π , Ad ) h ( μ π ) B π ( j ) ( m 1 , m 2 ) B π ( j ) ( n 1 , n 2 ) ,
E min : = c min ∑ j = 0 2 ∫ Re ( s 1 ) = 0 ∫ Re ( s 2 ) = 0 h ( s ) B s ( j ) ( m 1 , m 2 ) ‾ B s ( j ) ( n 1 , n 2 ) ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 d s 1 d s 2 , \mathcal{E}_{\min}:=c_{\min}\sum_{j=0}^{2}\int_{\operatorname{Re}(s_1)=0}\int_{\operatorname{Re}(s_2)=0}\frac{h(s)\overline{B_s^{(j)}(m_1,m_2)}B_s^{(j)}(n_1,n_2)}{\prod_{1\le i<k\le3}\lvert\zeta(1+s_i-s_k)\rvert^2}\,\mathrm{d}s_1\,\mathrm{d}s_2, E m i n := c m i n j = 0 ∑ 2 ∫ Re ( s 1 ) = 0 ∫ Re ( s 2 ) = 0 ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 h ( s ) B s ( j ) ( m 1 , m 2 ) B s ( j ) ( n 1 , n 2 ) d s 1 d s 2 ,
E max , n e w : = c max ∑ ϕ ∈ B 2 ( p ) ∫ Re ( s ) = 0 h ( s + τ ϕ , s − τ ϕ , − 2 s ) A ϕ , s ( m 1 , m 2 ) ‾ A ϕ , s ( n 1 , n 2 ) L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 s , ϕ ) ∣ 2 d s , \mathcal{E}_{\max,\mathrm{new}}:=c_{\max}\sum_{\phi\in\mathcal{B}_2(p)}\int_{\operatorname{Re}(s)=0}\frac{h(s+\tau_\phi,s-\tau_\phi,-2s)\overline{A_{\phi,s}(m_1,m_2)}A_{\phi,s}(n_1,n_2)}{L(1,\phi,\operatorname{Ad})\lvert L(1+3s,\phi)\rvert^2}\,\mathrm{d}s, E m a x , new := c m a x ϕ ∈ B 2 ( p ) ∑ ∫ Re ( s ) = 0 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 s , ϕ ) ∣ 2 h ( s + τ ϕ , s − τ ϕ , − 2 s ) A ϕ , s ( m 1 , m 2 ) A ϕ , s ( n 1 , n 2 ) d s ,
E max , o l d : = c max ∑ ϕ ∈ B 2 ( 1 ) ∑ j = 0 2 ∫ Re ( s ) = 0 h ( s + τ ϕ , s − τ ϕ , − 2 s ) B ϕ , s ( j ) ( m 1 , m 2 ) ‾ B ϕ , s ( j ) ( n 1 , n 2 ) L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 s , ϕ ) ∣ 2 d s . \mathcal{E}_{\max,\mathrm{old}}:=c_{\max}\sum_{\phi\in\mathcal{B}_2(1)}\sum_{j=0}^{2}\int_{\operatorname{Re}(s)=0}\frac{h(s+\tau_\phi,s-\tau_\phi,-2s)\overline{B_{\phi,s}^{(j)}(m_1,m_2)}B_{\phi,s}^{(j)}(n_1,n_2)}{L(1,\phi,\operatorname{Ad})\lvert L(1+3s,\phi)\rvert^2}\,\mathrm{d}s. E m a x , old := c m a x ϕ ∈ B 2 ( 1 ) ∑ j = 0 ∑ 2 ∫ Re ( s ) = 0 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 s , ϕ ) ∣ 2 h ( s + τ ϕ , s − τ ϕ , − 2 s ) B ϕ , s ( j ) ( m 1 , m 2 ) B ϕ , s ( j ) ( n 1 , n 2 ) d s .
Here d s \mathrm{d}s d s and d s 1 d s 2 \mathrm{d}s_1\mathrm{d}s_2 d s 1 d s 2 denote positive spectral measures, with the fixed normalizing factors included in c max c_{\max} c m a x and c min c_{\min} c m i n respectively.
Theorem 3.1. For h ∈ H ~ ( 10 ) h\in\widetilde{\mathcal{H}}(10) h ∈ H ( 10 ) and prime p p p , the Kuznetsov formula is
χ p C n e w + C o l d + E min + χ p E max , n e w + E max , o l d = Δ + S 4 + S 5 + S 6 , (51) \chi_p\mathcal{C}_{\mathrm{new}}+\mathcal{C}_{\mathrm{old}}+\mathcal{E}_{\min}+\chi_p\mathcal{E}_{\max,\mathrm{new}}+\mathcal{E}_{\max,\mathrm{old}}=\Delta+S_4+S_5+S_6, \tag*{(51)} χ p C new + C old + E m i n + χ p E m a x , new + E m a x , old = Δ + S 4 + S 5 + S 6 , (51) where
Δ = P δ m 1 = n 1 δ m 2 = n 2 ∫ Re ( μ ) = 0 h ( μ ) spec ( μ ) d μ , (52) \Delta=P\delta_{m_1=n_1}\delta_{m_2=n_2}\int_{\operatorname{Re}(\mu)=0}h(\mu)\operatorname{spec}(\mu)\,\mathrm{d}\mu, \tag*{(52)} Δ = P δ m 1 = n 1 δ m 2 = n 2 ∫ Re ( μ ) = 0 h ( μ ) spec ( μ ) d μ , (52) S 4 = P ∑ ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p D 2 ∣ D 1 m 2 D 1 = n 1 D 2 S ~ ( − ϵ n 2 , m 2 , m 1 , D 2 , D 1 ) D 1 D 2 Φ w 4 ( ϵ m 1 m 2 n 2 D 1 D 2 ) , S_4=P\sum_{\epsilon=\pm1}\sum_{\substack{D_1,D_2\ge1\\pD_2\mid D_1\\m_2D_1=n_1D_2}}\frac{\widetilde{S}(-\epsilon n_2,m_2,m_1,D_2,D_1)}{D_1D_2}\Phi_{w_4}\left(\frac{\epsilon m_1m_2n_2}{D_1D_2}\right), S 4 = P ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p D 2 ∣ D 1 m 2 D 1 = n 1 D 2 ∑ D 1 D 2 S ( − ϵ n 2 , m 2 , m 1 , D 2 , D 1 ) Φ w 4 ( D 1 D 2 ϵ m 1 m 2 n 2 ) ,
S 5 = P ∑ ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 ∣ D 2 m 1 D 2 = n 2 D 1 2 S ~ ( ϵ n 1 , m 1 , m 2 , D 1 , D 2 ) D 1 D 2 Φ w 5 ( ϵ n 1 m 1 m 2 D 1 D 2 ) . S_5=P\sum_{\epsilon=\pm1}\sum_{\substack{D_1,D_2\geq1\\p\mid D_1\mid D_2\\m_1D_2=n_2D_1^2}}\frac{\widetilde{S}(\epsilon n_1,m_1,m_2,D_1,D_2)}{D_1D_2}\Phi_{w_5}\left(\frac{\epsilon n_1m_1m_2}{D_1D_2}\right). S 5 = P ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 ∣ D 2 m 1 D 2 = n 2 D 1 2 ∑ D 1 D 2 S ( ϵ n 1 , m 1 , m 2 , D 1 , D 2 ) Φ w 5 ( D 1 D 2 ϵ n 1 m 1 m 2 ) .
S 6 = P ∑ ϵ 1 , ϵ 2 = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , p ∣ D 2 S ( p ) ( ϵ 2 n 2 , ϵ 1 n 1 , m 1 , m 2 , D 1 , D 2 ) D 1 D 2 Φ w 6 ( − ϵ 2 m 1 n 2 D 2 D 1 2 , − ϵ 1 m 2 n 1 D 1 D 2 2 ) . S_6=P\sum_{\epsilon_1,\epsilon_2=\pm1}\sum_{\substack{D_1,D_2\geq1\\p\mid D_1,\ p\mid D_2}}\frac{S^{(p)}(\epsilon_2n_2,\epsilon_1n_1,m_1,m_2,D_1,D_2)}{D_1D_2}\Phi_{w_6}\left(-\frac{\epsilon_2m_1n_2D_2}{D_1^2},-\frac{\epsilon_1m_2n_1D_1}{D_2^2}\right). S 6 = P ϵ 1 , ϵ 2 = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , p ∣ D 2 ∑ D 1 D 2 S ( p ) ( ϵ 2 n 2 , ϵ 1 n 1 , m 1 , m 2 , D 1 , D 2 ) Φ w 6 ( − D 1 2 ϵ 2 m 1 n 2 D 2 , − D 2 2 ϵ 1 m 2 n 1 D 1 ) .
Proof. Combining the discussion in § 3.3.1, § 3.3.3 and [11 Lemma 16] , we can get the trace formula. □
Remark 2. Since χ p = 1 + O ( p − 1 ) \chi_p=1+O(p^{-1}) χ p = 1 + O ( p − 1 ) , conversion to full finite adjoint values changes the moment main terms within the error terms stated below.
Remark 3. In the moment calculations below, we insert the appropriate approximate functional equation from § 2.3 and combine the coefficient products by the Hecke relations before applying Theorem 3.1 . If a dual sum is reindexed by π ↦ π ~ \pi\mapsto\widetilde{\pi} π ↦ π , the family and the full adjoint value are unchanged, and the full spectral test, including its weight function, is evaluated at − μ -\mu − μ . The diagonal contribution is determined by equality of the two resulting Fourier index pairs as in (52) .
Bounds for the Kuznetsov transforms
We recall the following kernel bounds from [13 Lemmas 2.4, 2.5 and 2.6(a)] .
Lemma 3.2 (Kuznetsov kernel bounds). Let A , ε > 0 A,\varepsilon>0 A , ε > 0 and j , j 1 , j 2 ∈ Z ≥ 0 j,j_1,j_2\in\mathbb{Z}_{\geq0} j , j 1 , j 2 ∈ Z ≥ 0 . For sufficiently large a 0 a_0 a 0 and h ∈ H ( a 0 ) h\in\mathcal{H}(a_0) h ∈ H ( a 0 ) , the following bounds hold for y , y 1 , y 2 ∈ R ∖ { 0 } y,y_1,y_2\in\mathbb{R}\setminus\{0\} y , y 1 , y 2 ∈ R ∖ { 0 } .
(1) Short Weyl kernels:
∣ y ∣ j ∣ d j d y j Φ w i ( y ) ∣ ≪ A , j , ε , h ∣ y ∣ j / 3 min { ∣ y ∣ A , ∣ y ∣ − 1 / 6 + ε } , i = 4 , 5. (53) |y|^j\left|\frac{\mathrm{d}^j}{\mathrm{d}y^j}\Phi_{w_i}(y)\right|\ll_{A,j,\varepsilon,h}|y|^{j/3}\min\{|y|^A,|y|^{-1/6+\varepsilon}\},\qquad i=4,5. \tag*{(53)} ∣ y ∣ j d y j d j Φ w i ( y ) ≪ A , j , ε , h ∣ y ∣ j /3 min { ∣ y ∣ A , ∣ y ∣ − 1/6 + ε } , i = 4 , 5. (53) (2) Derivatives of the long Weyl kernel:
∣ y 1 j 1 y 2 j 2 ∂ j 1 + j 2 Φ w 6 ( y 1 , y 2 ) ∂ y 1 j 1 ∂ y 2 j 2 ∣ ≪ j 1 , j 2 , h ( 1 + ∣ y 1 ∣ 1 / 2 + ∣ y 1 ∣ 1 / 3 ∣ y 2 ∣ 1 / 6 ) j 1 ( 1 + ∣ y 2 ∣ 1 / 2 + ∣ y 2 ∣ 1 / 3 ∣ y 1 ∣ 1 / 6 ) j 2 . \left|y_1^{j_1}y_2^{j_2}\frac{\partial^{j_1+j_2}\Phi_{w_6}(y_1,y_2)}{\partial y_1^{j_1}\partial y_2^{j_2}}\right|\ll_{j_1,j_2,h}\left(1+|y_1|^{1/2}+|y_1|^{1/3}|y_2|^{1/6}\right)^{j_1}\left(1+|y_2|^{1/2}+|y_2|^{1/3}|y_1|^{1/6}\right)^{j_2}. y 1 j 1 y 2 j 2 ∂ y 1 j 1 ∂ y 2 j 2 ∂ j 1 + j 2 Φ w 6 ( y 1 , y 2 ) ≪ j 1 , j 2 , h ( 1 + ∣ y 1 ∣ 1/2 + ∣ y 1 ∣ 1/3 ∣ y 2 ∣ 1/6 ) j 1 ( 1 + ∣ y 2 ∣ 1/2 + ∣ y 2 ∣ 1/3 ∣ y 1 ∣ 1/6 ) j 2 . (3) Long Weyl kernel at small arguments:
Φ w 6 ( y 1 , y 2 ) ≪ A , ε , h min { ∣ y 1 y 2 2 ∣ , ∣ y 1 2 y 2 ∣ } A ( 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ) ε . (54) \Phi_{w_6}(y_1,y_2)\ll_{A,\varepsilon,h}\min\{|y_1y_2^2|,|y_1^2y_2|\}^A(1+|y_1|+|y_2|)^\varepsilon. \tag*{(54)} Φ w 6 ( y 1 , y 2 ) ≪ A , ε , h min { ∣ y 1 y 2 2 ∣ , ∣ y 1 2 y 2 ∣ } A ( 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ) ε . (54) (4) Global bound for the long Weyl kernel:
Φ w 6 ( y 1 , y 2 ) ≪ h ( 2 + ∣ log ∣ y 1 ∣ ∣ + ∣ log ∣ y 2 ∣ ∣ ) 3 ( 1 + ∣ y 1 ∣ ) − 1 / 4 ( 1 + ∣ y 2 ∣ ) − 1 / 4 . \Phi_{w_6}(y_1,y_2)\ll_h(2+|\log|y_1||+|\log|y_2||)^3(1+|y_1|)^{-1/4}(1+|y_2|)^{-1/4}. Φ w 6 ( y 1 , y 2 ) ≪ h ( 2 + ∣ log ∣ y 1 ∣∣ + ∣ log ∣ y 2 ∣∣ ) 3 ( 1 + ∣ y 1 ∣ ) − 1/4 ( 1 + ∣ y 2 ∣ ) − 1/4 .
Lemma 3.3 (Mean squares of Fourier coefficients). Fix 0 < δ < 1 0<\delta<1 0 < δ < 1 and C ≥ 0 C\geq0 C ≥ 0 , and let h h h be fixed and rapidly decreasing on the unitary spectrum. Uniformly for positive integers a , b a,b a , b with ( a b , p ) = 1 (ab,p)=1 ( ab , p ) = 1 and a b ≤ p 1 − δ ab\leq p^{1-\delta} ab ≤ p 1 − δ ,
∑ π ∈ B ( p ) ∣ h ( μ π ) ∣ ( 1 + ∥ μ π ∥ ) C ∣ A π ( a , b ) ∣ 2 L ( 1 , π , Ad ) ≪ h , C , δ p 2 . (55) \sum_{\pi\in\mathcal{B}(p)}\frac{|h(\mu_\pi)|(1+\|\mu_\pi\|)^C|A_\pi(a,b)|^2}{L(1,\pi,\operatorname{Ad})}\ll_{h,C,\delta}p^2. \tag*{(55)} π ∈ B ( p ) ∑ L ( 1 , π , Ad ) ∣ h ( μ π ) ∣ ( 1 + ∥ μ π ∥ ) C ∣ A π ( a , b ) ∣ 2 ≪ h , C , δ p 2 . (55) :::`
and
∑ π ∈ B ( p ) ∣ h ( μ π ) ∣ ( 1 + ∥ μ π ∥ ) C ∣ A π ( a , b ) ∣ L ( 1 , π , Ad ) ≪ h , C , δ p 2 . (56) \sum_{\pi\in B(p)} \frac{|h(\mu_\pi)|(1+\|\mu_\pi\|)^C|A_\pi(a,b)|}{L(1,\pi,\operatorname{Ad})}\ll_{h,C,\delta}p^2. \tag*{(56)} π ∈ B ( p ) ∑ L ( 1 , π , Ad ) ∣ h ( μ π ) ∣ ( 1 + ∥ μ π ∥ ) C ∣ A π ( a , b ) ∣ ≪ h , C , δ p 2 . (56)
Proof. Choose A > max ( 4 / 3 , 2 / ( 3 δ ) ) A>\max(4/3,2/(3\delta)) A > max ( 4/3 , 2/ ( 3 δ )) and take K K K sufficiently large for the corresponding kernel estimates. For R ≥ 1 R\ge1 R ≥ 1 , we set
g R ( μ ) : = exp ( R − 2 ∑ j = 1 3 μ j 2 ) ∏ 1 ≤ i < j ≤ 3 ∏ k = 0 K ( ( μ i − μ j ) 2 − ( 2 k + 1 ) 2 ) 2 . g_R(\mu):=\exp\left(R^{-2}\sum_{j=1}^{3}\mu_j^2\right)\prod_{1\le i<j\le3}\prod_{k=0}^{K}\left((\mu_i-\mu_j)^2-(2k+1)^2\right)^2. g R ( μ ) := exp ( R − 2 j = 1 ∑ 3 μ j 2 ) 1 ≤ i < j ≤ 3 ∏ k = 0 ∏ K ( ( μ i − μ j ) 2 − ( 2 k + 1 ) 2 ) 2 .
For g R g_R g R , the bounds (53) and (54) used below have implied constants O δ ( R N ) O_\delta(R^N) O δ ( R N ) , for some N N N independent of R R R , p p p . On the unitary spectrum, μ \mu μ and − μ ‾ -\overline{\mu} − μ lie in the same Weyl orbit, so g R ≥ 0 g_R\ge0 g R ≥ 0 . The bound ∣ Re ( μ j ) ∣ ≤ 5 / 14 |\operatorname{Re}(\mu_j)|\le5/14 ∣ Re ( μ j ) ∣ ≤ 5/14 also gives g R ( μ ) ≫ K 1 g_R(\mu)\gg_K 1 g R ( μ ) ≫ K 1 when ∥ μ ∥ ≤ R \|\mu\|\le R ∥ μ ∥ ≤ R . Apply Theorem 3.1 with indices ( a , b , a , b ) (a,b,a,b) ( a , b , a , b ) and the test function g R g_R g R . Put g = ( a , b ) g=(a,b) g = ( a , b ) , a ′ = a / g a'=a/g a ′ = a / g and b ′ = b / g b'=b/g b ′ = b / g . The short modulus pairs are
( D 1 , D 2 ) = ( p 2 a ′ b ′ k 2 , p b ′ k ) or ( p a ′ k , p 2 a ′ b ′ k 2 ) , k ≥ 1. (D_1,D_2)=(p^2a'b'k^2,pb'k)\quad\text{or}\quad(pa'k,p^2a'b'k^2),\qquad k\ge1. ( D 1 , D 2 ) = ( p 2 a ′ b ′ k 2 , p b ′ k ) or ( p a ′ k , p 2 a ′ b ′ k 2 ) , k ≥ 1.
Both kernel arguments have modulus g 3 / ( p 3 k 3 ) g^3/(p^3k^3) g 3 / ( p 3 k 3 ) . Since g 2 ≤ a b < p g^2\le ab<p g 2 ≤ ab < p , the trivial Kloosterman bound and (53) give
∣ S 4 ∣ + ∣ S 5 ∣ ≪ δ p 2 R N ( g / p ) 3 A ∑ k ≥ 1 k − 3 A ≪ δ p 2 R N . |S_4|+|S_5|\ll_\delta p^2R^N(g/p)^{3A}\sum_{k\ge1}k^{-3A}\ll_\delta p^2R^N. ∣ S 4 ∣ + ∣ S 5 ∣ ≪ δ p 2 R N ( g / p ) 3 A k ≥ 1 ∑ k − 3 A ≪ δ p 2 R N .
For the long term,
min { ∣ y 1 2 y 2 ∣ , ∣ y 1 y 2 2 ∣ } = ( a b ) 3 max ( D 1 , D 2 ) 3 , and 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ≤ 3 max ( D 1 , D 2 ) p . \min\{|y_1^2y_2|,|y_1y_2^2|\}=\frac{(ab)^3}{\max(D_1,D_2)^3},\qquad\text{and}\qquad1+|y_1|+|y_2|\le\frac{3\max(D_1,D_2)}{p}. min { ∣ y 1 2 y 2 ∣ , ∣ y 1 y 2 2 ∣ } = max ( D 1 , D 2 ) 3 ( ab ) 3 , and 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ≤ p 3 max ( D 1 , D 2 ) .
Apply (54) with decay exponent A + 1 A+1 A + 1 and growth exponent one. Using a b ≤ p ≤ max ( D 1 , D 2 ) ab\le p\le\max(D_1,D_2) ab ≤ p ≤ max ( D 1 , D 2 ) gives
Φ w 0 ( y 1 , y 2 ) ≪ δ R N ( a b ) 3 A max ( D 1 , D 2 ) − 3 A . \Phi_{w_0}(y_1,y_2)\ll_\delta R^N(ab)^{3A}\max(D_1,D_2)^{-3A}. Φ w 0 ( y 1 , y 2 ) ≪ δ R N ( ab ) 3 A max ( D 1 , D 2 ) − 3 A .
Thus
∣ S 6 ∣ ≪ δ p 2 R N ( a b ) 3 A ∑ p ∣ D 1 , D 2 D 1 D 2 max ( D 1 , D 2 ) − 3 A |S_6|\ll_\delta p^2R^N(ab)^{3A}\sum_{p\mid D_1,D_2}D_1D_2\max(D_1,D_2)^{-3A} ∣ S 6 ∣ ≪ δ p 2 R N ( ab ) 3 A p ∣ D 1 , D 2 ∑ D 1 D 2 max ( D 1 , D 2 ) − 3 A
≪ δ R N p 4 − 3 A ( a b ) 3 A ≤ R N p 4 − 3 A δ ≪ δ p 2 R N . \ll_\delta R^Np^{4-3A}(ab)^{3A}\le R^Np^{4-3A\delta}\ll_\delta p^2R^N. ≪ δ R N p 4 − 3 A ( ab ) 3 A ≤ R N p 4 − 3 A δ ≪ δ p 2 R N .
The diagonal term is O δ ( p 2 R N ) O_\delta(p^2R^N) O δ ( p 2 R N ) . Positivity therefore yields
∑ π ∈ B ( p ) ∥ μ π ∥ ≤ R ∣ A π ( a , b ) ∣ 2 L ( 1 , π , Ad ) ≪ δ p 2 R N . \sum_{\substack{\pi\in B(p)\\ \|\mu_\pi\|\le R}}\frac{|A_\pi(a,b)|^2}{L(1,\pi,\operatorname{Ad})}\ll_\delta p^2R^N. π ∈ B ( p ) ∥ μ π ∥ ≤ R ∑ L ( 1 , π , Ad ) ∣ A π ( a , b ) ∣ 2 ≪ δ p 2 R N .
Dyadic decomposition and rapid decrease of h h h give (55) . The Cauchy–Schwarz inequality and the case a = b = 1 a=b=1 a = b = 1 of (55) give (56) . □
Compatibility with the Kuznetsov test class
We first work with a general test function. Fix a tube width a ≥ 10 a \ge10 a ≥ 10 , an integer J ≥ ⌈ a ⌉ + 4 J \ge\lceil a\rceil+ 4 J ≥ ⌈ a ⌉ + 4 and h 0 ∈ H ~ ( a + 1 ) h_0 \in\widetilde{\mathcal H}(a+1) h 0 ∈ H ( a + 1 ) . We set
h ( μ ) = P J ( 0 , μ ) h 0 ( μ ) , (57) h(\mu)=P_J(0,\mu)h_0(\mu), \tag*{(57)} h ( μ ) = P J ( 0 , μ ) h 0 ( μ ) , (57)
F u ± ( μ ) = h 0 ( ± μ ) e u 2 P J ( u , μ ) γ μ ( 1 / 2 + u ) γ ± μ ( 1 / 2 ) . F_u^{\pm}(\mu)=h_0(\pm\mu)e^{u^2}P_J(u,\mu)\frac{\gamma_\mu(1/2+u)}{\gamma_{\pm\mu}(1/2)}. F u ± ( μ ) = h 0 ( ± μ ) e u 2 P J ( u , μ ) γ ± μ ( 1/2 ) γ μ ( 1/2 + u ) .
Here P J P_J P J is given by (21) , and F u − F_u^{-} F u − is the Mellin multiplier of h ( − μ ) W − μ h(-\mu)W_{-\mu} h ( − μ ) W − μ .
Remark 4. The fixed test function in (26) is a special case of (57) obtained by taking h 0 h_0 h 0 as in (27) . In this specialization, h h h remains fixed and h 0 h_0 h 0 depends on J J J .
Lemma 3.4 (Uniform admissibility). For every fixed C > 0 C>0 C > 0 and − 2 ≤ Re ( u ) ≤ C -2\le\operatorname{Re}(u)\le C − 2 ≤ Re ( u ) ≤ C , the functions F u ± F_u^{\pm} F u ± belong to H ~ ( a ) \widetilde{\mathcal H}(a) H ( a ) . For all M , k , R ≥ 0 M,k,R\ge0 M , k , R ≥ 0 ,
sup μ ∈ Λ a ∣ α ∣ ≤ k ( 1 + ∥ μ ∥ ) M ∣ ∂ μ α F u ± ( μ ) ∣ ≪ h 0 , a , J , C , M , k , R ( 1 + ∣ Im ( u ) ∣ ) − R . (58) \sup_{\substack{\mu\in\Lambda_a\\|\alpha|\le k}}(1+\|\mu\|)^M\left|\partial_\mu^\alpha F_u^{\pm}(\mu)\right|\ll_{h_0,a,J,C,M,k,R}(1+|\operatorname{Im}(u)|)^{-R}. \tag*{(58)} μ ∈ Λ a ∣ α ∣ ≤ k sup ( 1 + ∥ μ ∥ ) M ∂ μ α F u ± ( μ ) ≪ h 0 , a , J , C , M , k , R ( 1 + ∣ Im ( u ) ∣ ) − R . (58)
Proof. The gamma recurrence gives
P J ( u , μ ) γ μ ( 1 / 2 + u ) = ( 2 π ) 3 ( J + 1 ) P J ( u , μ ) ∏ j = 1 3 ∏ ℓ = 0 J ( 1 / 2 + u − μ j + 2 ℓ ) γ μ ( 1 / 2 + u + 2 ( J + 1 ) ) . P_J(u,\mu)\gamma_\mu(1/2+u)=(2\pi)^{3(J+1)}\frac{P_J(u,\mu)}{\prod_{j=1}^{3}\prod_{\ell=0}^{J}(1/2+u-\mu_j+2\ell)}\gamma_\mu(1/2+u+2(J+1)). P J ( u , μ ) γ μ ( 1/2 + u ) = ( 2 π ) 3 ( J + 1 ) ∏ j = 1 3 ∏ ℓ = 0 J ( 1/2 + u − μ j + 2 ℓ ) P J ( u , μ ) γ μ ( 1/2 + u + 2 ( J + 1 )) .
By (21) , the quotient is a polynomial. For μ ∈ Λ a + 1 \mu\in\Lambda_{a+1} μ ∈ Λ a + 1 and Re ( u ) ≥ − 2 \operatorname{Re}(u)\ge-2 Re ( u ) ≥ − 2 , the shifted gamma arguments have real parts at least 2 J − a − 1 / 2 > 0 2J-a-1/2>0 2 J − a − 1/2 > 0 . Thus F u ± F_u^{\pm} F u ± are holomorphic on a neighborhood of Λ a \Lambda_a Λ a . And they are Weyl invariant and keep the structural zeros of h 0 h_0 h 0 .
For u = σ + i v u=\sigma+iv u = σ + i v with − 2 ≤ σ ≤ C -2\le\sigma\le C − 2 ≤ σ ≤ C , (18) applied to the shifted product gives
∣ F u ± ( μ ) ∣ ≪ a , J , C ∣ h 0 ( ± μ ) ∣ ( 1 + ∥ μ ∥ + ∣ v ∣ ) N e − v 2 + 3 π ∣ v ∣ / 4 , |F_u^{\pm}(\mu)|\ll_{a,J,C}|h_0(\pm\mu)|(1+\|\mu\|+|v|)^N e^{-v^2+3\pi|v|/4}, ∣ F u ± ( μ ) ∣ ≪ a , J , C ∣ h 0 ( ± μ ) ∣ ( 1 + ∥ μ ∥ + ∣ v ∣ ) N e − v 2 + 3 π ∣ v ∣/4 ,
for some fixed N N N . Rapid decrease of h 0 h_0 h 0 on Λ a + 1 \Lambda_{a+1} Λ a + 1 , the Gaussian decay in v v v and Cauchy’s integral formula on the larger tube give (58) . □ \square □
For n ≥ 1 n\ge1 n ≥ 1 and c > 0 c>0 c > 0 , define the weighted test functions by
H n ± ( μ ) : = h ( ± μ ) U ± ± ( n / p ) = 1 2 π i ∫ ( c ) ( p n ) u F u ± ( μ ) d u u . (59) H_n^{\pm}(\mu):=h(\pm\mu)U_{\pm}^{\pm}(n/\sqrt p)=\frac{1}{2\pi i}\int_{(c)}\left(\frac{\sqrt p}{n}\right)^uF_u^{\pm}(\mu)\frac{du}{u}. \tag*{(59)} H n ± ( μ ) := h ( ± μ ) U ± ± ( n / p ) = 2 π i 1 ∫ ( c ) ( n p ) u F u ± ( μ ) u d u . (59)
Here U ± ± U_{\pm}^{\pm} U ± ± are as in § 2.3, and the integral gives the holomorphic continuation of the products. Lemma 3.4 allows any positive Mellin line and gives H n ± ∈ H ~ ( a ) H_n^{\pm}\in\widetilde{\mathcal H}(a) H n ± ∈ H ( a ) with
sup μ ∈ Λ a ∣ α ∣ ≤ k ( 1 + ∥ μ ∥ ) M ∣ ∂ μ α H n ± ( μ ) ∣ ≪ c ( p / n ) c , (60) \sup_{\substack{\mu\in\Lambda_a\\|\alpha|\le k}}(1+\|\mu\|)^M\left|\partial_\mu^\alpha H_n^{\pm}(\mu)\right|\ll_c(\sqrt p/n)^c, \tag*{(60)} μ ∈ Λ a ∣ α ∣ ≤ k sup ( 1 + ∥ μ ∥ ) M ∂ μ α H n ± ( μ ) ≪ c ( p / n ) c , (60)
for every fixed M , k ≥ 0 M,k\ge0 M , k ≥ 0 .
Remark 5. For fixed exponents in (53) and (54) , their implied constants are bounded by a fixed finite sum of weighted suprema of the test function and its derivatives on a sufficiently wide fixed tube. The tube width and derivative orders depend only on the prescribed exponents. Indeed, the proofs of [13 Lemmas 2.4 and 2.5] use finitely many fixed contour shifts and residues. Stirling’s formula and Cauchy estimates on a slightly larger tube control the resulting integrals and residues by these suprema.
Consequently, (60) and (31) give the kernel bounds for H n ± H_n^{\pm} H n ± and H t H_t H t with respective additional factors ( p / n ) c (\sqrt p/n)^c ( p / n ) c and ( p / t ) c (p/t)^c ( p / t ) c , for every fixed c > 0 c>0 c > 0 . Similarly, (36) gives the bounds for K μ ± ( y ) K_{\mu}^{\pm}(y) K μ ± ( y ) with a factor y − c y^{-c} y − c . Taking c c c sufficiently large gives arbitrary decay in p p p when n > p 1 / 2 + ε n > p^{1/2+\varepsilon} n > p 1/2 + ε for fixed ε > 0 \varepsilon> 0 ε > 0 .
Harmonic twisted moments
In this section we evaluate the harmonic twisted first and second moments, uniformly in the twist ranges needed for mollification.
The harmonic twisted first moment
For a general test h h h as in (57) , define the harmonic twisted first moment and the corresponding spectral integral by
M h ( r , s , p ) : = c c u s p ∑ π ∈ B ( p ) A π ( r , s ) ‾ L ( 1 / 2 , π ) L ( 1 , π , Ad ) h ( μ π ) , (61) M_h(r,s,p) := c_{\mathrm{cusp}} \sum_{\pi\in\mathcal{B}(p)} \frac{\overline{A_\pi(r,s)}L(1/2,\pi)}{L(1,\pi,\operatorname{Ad})}h(\mu_\pi), \tag*{(61)} M h ( r , s , p ) := c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) A π ( r , s ) L ( 1/2 , π ) h ( μ π ) , (61)
and
H h : = ∫ Re ( μ ) = 0 h ( μ ) spec ( μ ) d μ . (62) \mathcal{H}_h := \int_{\operatorname{Re}(\mu)=0} h(\mu)\operatorname{spec}(\mu)\,d\mu. \tag*{(62)} H h := ∫ Re ( μ ) = 0 h ( μ ) spec ( μ ) d μ . (62)
Theorem 4.1. Fix a positive integer Q Q Q and 0 < η < 1 / 2 0 < \eta< 1/2 0 < η < 1/2 . Let a 0 ≥ 10 a_0 \ge10 a 0 ≥ 10 be sufficiently large in terms of η \eta η , and fix an integer J ≥ ⌈ a 0 ⌉ + 4 J \ge\lceil a_0\rceil+ 4 J ≥ ⌈ a 0 ⌉ + 4 and h 0 ∈ H ~ ( a 0 + 1 ) h_0 \in\widetilde{\mathcal{H}}(a_0+1) h 0 ∈ H ( a 0 + 1 ) , independently of p p p . Set h ( μ ) = P J ( 0 , μ ) h 0 ( μ ) h(\mu) = P_J(0,\mu)h_0(\mu) h ( μ ) = P J ( 0 , μ ) h 0 ( μ ) . Then, for every ε > 0 \varepsilon> 0 ε > 0 , as p → ∞ p \to\infty p → ∞ through primes and uniformly for
1 ≤ r ≤ Q p η , 1 ≤ s ≤ Q , ( r s , p ) = 1 , 1 \le r \le Qp^\eta,\qquad1 \le s \le Q,\qquad(rs,p)=1, 1 ≤ r ≤ Q p η , 1 ≤ s ≤ Q , ( r s , p ) = 1 , we have
M h ( r , s , p ) = δ r = 1 s ( p 2 + p + 1 ) H h + O Q , η , h 0 , J , ε ( p 5 / 4 + ε ) . (63) M_h(r,s,p) = \frac{\delta_{r=1}}{\sqrt{s}}(p^2+p+1)\mathcal{H}_h + O_{Q,\eta,h_0,J,\varepsilon}(p^{5/4+\varepsilon}). \tag*{(63)} M h ( r , s , p ) = s δ r = 1 ( p 2 + p + 1 ) H h + O Q , η , h 0 , J , ε ( p 5/4 + ε ) . (63)
Note that the fixed test (26) satisfies these hypotheses for every 0 < η < 1 / 2 0 < \eta< 1/2 0 < η < 1/2 , by § 3.7 .
Substituting (23) into (61) , we have
M h ( r , s , p ) = c c u s p ∑ π ∈ B ( p ) h ( μ π ) A π ( r , s ) ‾ L ( 1 , π , Ad ) ∑ n ≥ 1 A π ( 1 , n ) ‾ n V μ π ( n / p ) − p c c u s p ∑ π ∈ B ( p ) h ( μ π ) A π ( r , s ) ‾ L ( 1 , π , Ad ) ∑ n ≥ 1 A π ( p , n ) ‾ n W μ π ( n / p ) = M h , 1 ( r , s , p ) + M h , 2 ( r , s , p ) , say . (64) \begin{aligned}
M_h(r,s,p) &= c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(\mu_\pi)\overline{A_\pi(r,s)}}{L(1,\pi,\operatorname{Ad})}\sum_{n\ge1}\frac{\overline{A_\pi(1,n)}}{\sqrt{n}}V_{\mu_\pi}(n/\sqrt{p})\\
&\quad-\sqrt{p}\,c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(\mu_\pi)\overline{A_\pi(r,s)}}{L(1,\pi,\operatorname{Ad})}\sum_{n\ge1}\frac{\overline{A_\pi(p,n)}}{\sqrt{n}}W_{\mu_\pi}(n/\sqrt{p})\\
&= M_{h,1}(r,s,p)+M_{h,2}(r,s,p),\quad\text{say}. \tag*{(64)}
\end{aligned} M h ( r , s , p ) = c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) A π ( r , s ) n ≥ 1 ∑ n A π ( 1 , n ) V μ π ( n / p ) − p c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) A π ( r , s ) n ≥ 1 ∑ n A π ( p , n ) W μ π ( n / p ) = M h , 1 ( r , s , p ) + M h , 2 ( r , s , p ) , say . (64)
The family B ( p ) \mathcal{B}(p) B ( p ) and the full adjoint value are invariant under contragredience. Reindexing M h , 2 ( r , s , p ) M_{h,2}(r,s,p) M h , 2 ( r , s , p ) in (64) by π ↦ π ~ \pi\mapsto\widetilde{\pi} π ↦ π , and using (10) , (11) , (13) and ( r s , p ) = 1 (rs,p)=1 ( r s , p ) = 1 , we obtain
M h , 2 ( r , s , p ) = − p c c u s p ∑ π ∈ B ( p ) h ( − μ π ) A π ( s , r ) ‾ L ( 1 , π , Ad ) ∑ n ≥ 1 A π ( p , n ) ‾ n W − μ π ( n / p ) . M_{h,2}(r,s,p)=-\sqrt{p}\,c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(-\mu_\pi)\overline{A_\pi(s,r)}}{L(1,\pi,\operatorname{Ad})}\sum_{n\ge1}\frac{\overline{A_\pi(p,n)}}{\sqrt{n}}W_{-\mu_\pi}(n/\sqrt{p}). M h , 2 ( r , s , p ) = − p c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( − μ π ) A π ( s , r ) n ≥ 1 ∑ n A π ( p , n ) W − μ π ( n / p ) .
With H n ± H_n^{\pm} H n ± as in (59) , for any fixed B > 0 B>0 B > 0 , truncation at n ≤ p 1 / 2 + ε n\le p^{1/2+\varepsilon} n ≤ p 1/2 + ε yields
M h ( r , s , p ) = c c u s p ∑ n ≤ p 1 / 2 + ε 1 n ∑ π ∈ B ( p ) A π ( r , s ) ‾ A π ( 1 , n ) ‾ H n + ( μ π ) L ( 1 , π , Ad ) − p c c u s p ∑ n ≤ p 1 / 2 + ε 1 n ∑ π ∈ B ( p ) A π ( s , r ) ‾ A π ( p , n ) ‾ H n − ( μ π ) L ( 1 , π , Ad ) + O Q , η , h 0 , J , ε , B ( p − B ) . \begin{aligned}
M_h(r,s,p) &= c_{\mathrm{cusp}}\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\sum_{\pi\in\mathcal{B}(p)}\frac{\overline{A_\pi(r,s)}\overline{A_\pi(1,n)}H_n^+(\mu_\pi)}{L(1,\pi,\operatorname{Ad})}\\
&\quad-\sqrt{p}\,c_{\mathrm{cusp}}\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\sum_{\pi\in\mathcal{B}(p)}\frac{\overline{A_\pi(s,r)}\overline{A_\pi(p,n)}H_n^-(\mu_\pi)}{L(1,\pi,\operatorname{Ad})}+O_{Q,\eta,h_0,J,\varepsilon,B}(p^{-B}).
\end{aligned} M h ( r , s , p ) = c cusp n ≤ p 1/2 + ε ∑ n 1 π ∈ B ( p ) ∑ L ( 1 , π , Ad ) A π ( r , s ) A π ( 1 , n ) H n + ( μ π ) − p c cusp n ≤ p 1/2 + ε ∑ n 1 π ∈ B ( p ) ∑ L ( 1 , π , Ad ) A π ( s , r ) A π ( p , n ) H n − ( μ π ) + O Q , η , h 0 , J , ε , B ( p − B ) .
Apply (51) with all four indices equal to 1 and the positive test g R g_R g R from the proof of Lemma 3.3. All five spectra satisfy ∣ Re ( μ j ) ∣ ≤ 5 / 14 |\operatorname{Re}(\mu_j)| \le5/14 ∣ Re ( μ j ) ∣ ≤ 5/14 , so g R ≥ 0 g_R \ge0 g R ≥ 0 and g R ≫ 1 g_R \gg1 g R ≫ 1 on ∥ μ ∥ ≤ R \|\mu\| \le R ∥ μ ∥ ≤ R in every sector. The new coefficients at ( 1 , 1 ) (1,1) ( 1 , 1 ) equal 1, while the three oldvectors associated with each level one representation satisfy ∑ j = 0 2 ∣ B Π ( j ) ( 1 , 1 ) ∣ 2 ≥ ∣ B Π ( 0 ) ( 1 , 1 ) ∣ 2 = 1 \sum_{j=0}^{2}|B_{\Pi}^{(j)}(1,1)|^2 \ge|B_{\Pi}^{(0)}(1,1)|^2 = 1 ∑ j = 0 2 ∣ B Π ( j ) ( 1 , 1 ) ∣ 2 ≥ ∣ B Π ( 0 ) ( 1 , 1 ) ∣ 2 = 1 . Positivity and the same geometric estimates therefore bound the arithmetic mass of each sector on ∥ μ ∥ ≤ R \|\mu\| \le R ∥ μ ∥ ≤ R by O ( p 2 R N ) O(p^2R^N) O ( p 2 R N ) , counting each old representation once. Dyadic decomposition and rapid decrease of h h h give harmonic mass O h , C ( p 2 ) O_{h,C}(p^2) O h , C ( p 2 ) in each sector with weight ( ∣ h ( μ ) ∣ + ∣ h ( − μ ) ∣ ) ( 1 + ∥ μ ∥ ) C (|h(\mu)|+|h(-\mu)|)(1+\|\mu\|)^C ( ∣ h ( μ ) ∣ + ∣ h ( − μ ) ∣ ) ( 1 + ∥ μ ∥ ) C , for every fixed C ≥ 0 C \ge0 C ≥ 0 . By (40) –(42) , the unramified coefficient bounds and (9) ,
∣ B Π ( j ) ( 1 , n ) ∣ + ∣ B Π ( j ) ( p , n ) ∣ ≪ ε p 1 + ε n θ 3 + ε , 0 ≤ j ≤ 2. (65) |B_{\Pi}^{(j)}(1,n)|+|B_{\Pi}^{(j)}(p,n)|\ll_{\varepsilon}p^{1+\varepsilon}n^{\theta_3+\varepsilon},\qquad0\le j\le2. \tag*{(65)} ∣ B Π ( j ) ( 1 , n ) ∣ + ∣ B Π ( j ) ( p , n ) ∣ ≪ ε p 1 + ε n θ 3 + ε , 0 ≤ j ≤ 2. (65)
in the three level one sectors. The same bound holds for the new coefficients A π A_{\pi} A π and A ϕ , z A_{\phi,z} A ϕ , z . At ( r , s ) (r,s) ( r , s ) and ( s , r ) (s,r) ( s , r ) , the twist coefficient is bounded by ( r s ) θ 3 + ε ≪ Q p η ( θ 3 + ε ) (rs)^{\theta_3+\varepsilon}\ll_Q p^{\eta(\theta_3+\varepsilon)} ( r s ) θ 3 + ε ≪ Q p η ( θ 3 + ε ) uniformly in the stated range. Taking the decay exponent in (24) sufficiently large makes the contribution of n > p 1 / 2 + ε n>p^{1/2+\varepsilon} n > p 1/2 + ε , including the dual factor p \sqrt{p} p , of size O Q , η , h 0 , J , ε , B ( p − B ) O_{Q,\eta,h_0,J,\varepsilon,B}(p^{-B}) O Q , η , h 0 , J , ε , B ( p − B ) for any B > 0 B>0 B > 0 . These bounds also give absolute convergence and justify interchanging the sums and spectral integrals.
Apply Theorem 3.1 with indices ( r , s , 1 , n ) (r,s,1,n) ( r , s , 1 , n ) and ( s , r , p , n ) (s,r,p,n) ( s , r , p , n ) and tests H n + H_n^+ H n + and H n − H_n^- H n − , respectively. Admissibility follows from Lemma 3.4. The diagonal contribution is
D : = P ∑ n ≤ p 1 / 2 + ε 1 n ∫ Re ( μ ) = 0 ( δ r = 1 δ s = n H n + ( μ ) − p δ s = p δ r = n H n − ( μ ) ) spec ( μ ) d μ . (66) \mathcal{D}:=P\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\int_{\operatorname{Re}(\mu)=0}\left(\delta_{r=1}\delta_{s=n}H_n^+(\mu)-\sqrt{p}\,\delta_{s=p}\delta_{r=n}H_n^-(\mu)\right)\operatorname{spec}(\mu)\,d\mu. \tag*{(66)} D := P n ≤ p 1/2 + ε ∑ n 1 ∫ Re ( μ ) = 0 ( δ r = 1 δ s = n H n + ( μ ) − p δ s = p δ r = n H n − ( μ ) ) spec ( μ ) d μ . (66)
With the transforms in (39) , define the short Weyl contributions by
S 4 + : = P ∑ n ≤ p 1 / 2 + ε 1 n ∑ ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p D 2 ∣ D 1 s D 1 = D 2 2 S ~ ( − ϵ n , s , r , D 2 , D 1 ) D 1 D 2 Φ w 4 , H n + ( ϵ r s n D 1 D 2 ) , \mathcal{S}_4^+:=P\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\sum_{\epsilon=\pm1}\sum_{\substack{D_1,D_2\ge1\\pD_2\mid D_1\\sD_1=D_2^2}}\frac{\widetilde{S}(-\epsilon n,s,r,D_2,D_1)}{D_1D_2}\Phi_{w_4,H_n^+}\left(\frac{\epsilon rsn}{D_1D_2}\right), S 4 + := P n ≤ p 1/2 + ε ∑ n 1 ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p D 2 ∣ D 1 s D 1 = D 2 2 ∑ D 1 D 2 S ( − ϵ n , s , r , D 2 , D 1 ) Φ w 4 , H n + ( D 1 D 2 ϵr s n ) ,
S 4 − : = P p ∑ n ≤ p 1 / 2 + ε 1 n ∑ ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p D 2 ∣ D 1 r D 1 = p D 2 2 S ~ ( − ϵ n , r , s , D 2 , D 1 ) D 1 D 2 Φ w 4 , H n − ( ϵ r s n D 1 D 2 ) , \mathcal{S}_4^-:=P\sqrt{p}\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\sum_{\epsilon=\pm1}\sum_{\substack{D_1,D_2\ge1\\pD_2\mid D_1\\rD_1=pD_2^2}}\frac{\widetilde{S}(-\epsilon n,r,s,D_2,D_1)}{D_1D_2}\Phi_{w_4,H_n^-}\left(\frac{\epsilon rsn}{D_1D_2}\right), S 4 − := P p n ≤ p 1/2 + ε ∑ n 1 ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p D 2 ∣ D 1 r D 1 = p D 2 2 ∑ D 1 D 2 S ( − ϵ n , r , s , D 2 , D 1 ) Φ w 4 , H n − ( D 1 D 2 ϵr s n ) ,
S 5 + : = P ∑ n ≤ p 1 / 2 + ε 1 n ∑ ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 ∣ D 2 r D 2 = n D 1 2 S ~ ( ϵ , r , s , D 1 , D 2 ) D 1 D 2 Φ w 5 , H n + ( ϵ r s D 1 D 2 ) , \mathcal{S}_5^+:=P\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\sum_{\epsilon=\pm1}\sum_{\substack{D_1,D_2\ge1\\p\mid D_1\mid D_2\\rD_2=nD_1^2}}\frac{\widetilde{S}(\epsilon,r,s,D_1,D_2)}{D_1D_2}\Phi_{w_5,H_n^+}\left(\frac{\epsilon rs}{D_1D_2}\right), S 5 + := P n ≤ p 1/2 + ε ∑ n 1 ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 ∣ D 2 r D 2 = n D 1 2 ∑ D 1 D 2 S ( ϵ , r , s , D 1 , D 2 ) Φ w 5 , H n + ( D 1 D 2 ϵr s ) ,
S 5 − : = P p ∑ n ≤ p 1 / 2 + ε 1 n ∑ ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 ∣ D 2 s D 2 = n D 1 2 S ~ ( ϵ p , s , r , D 1 , D 2 ) D 1 D 2 Φ w 5 , H n − ( ϵ p r s D 1 D 2 ) . \mathcal{S}_5^-:=P\sqrt{p}\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\sum_{\epsilon=\pm1}\sum_{\substack{D_1,D_2\ge1\\p\mid D_1\mid D_2\\sD_2=nD_1^2}}\frac{\widetilde{S}(\epsilon p,s,r,D_1,D_2)}{D_1D_2}\Phi_{w_5,H_n^-}\left(\frac{\epsilon prs}{D_1D_2}\right). S 5 − := P p n ≤ p 1/2 + ε ∑ n 1 ϵ = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 ∣ D 2 s D 2 = n D 1 2 ∑ D 1 D 2 S ( ϵ p , s , r , D 1 , D 2 ) Φ w 5 , H n − ( D 1 D 2 ϵ p r s ) .
The long Weyl contributions are
S 6 + : = P ∑ n ≤ p 1 / 2 + ε 1 n ∑ ε 1 , ε 2 = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , p ∣ D 2 S ( p ) ( ε 2 n , ε 1 , r , s , D 1 , D 2 ) D 1 D 2 Φ w 6 , H n + ( − ε 2 r n D 2 D 1 2 , − ε 1 s D 1 D 2 2 ) , \mathcal{S}_6^+ := P \sum_{n \le p^{1/2+\varepsilon}} \frac{1}{\sqrt{n}} \sum_{\varepsilon_1,\varepsilon_2=\pm1} \sum_{\substack{D_1,D_2\ge1\\ p\mid D_1,\ p\mid D_2}} \frac{S^{(p)}(\varepsilon_2 n,\varepsilon_1,r,s,D_1,D_2)}{D_1D_2}\Phi_{w_6,H_n^+}\left(-\frac{\varepsilon_2rnD_2}{D_1^2},-\frac{\varepsilon_1sD_1}{D_2^2}\right), S 6 + := P n ≤ p 1/2 + ε ∑ n 1 ε 1 , ε 2 = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , p ∣ D 2 ∑ D 1 D 2 S ( p ) ( ε 2 n , ε 1 , r , s , D 1 , D 2 ) Φ w 6 , H n + ( − D 1 2 ε 2 r n D 2 , − D 2 2 ε 1 s D 1 ) ,
S 6 − : = P p ∑ n ≤ p 1 / 2 + ε 1 n ∑ ε 1 , ε 2 = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , p ∣ D 2 S ( p ) ( ε 2 n , ε 1 p , s , r , D 1 , D 2 ) D 1 D 2 Φ w 6 , H n − ( − ε 2 s n D 2 D 1 2 , − ε 1 p r D 1 D 2 2 ) . (67) \mathcal{S}_6^- := P\sqrt{p} \sum_{n \le p^{1/2+\varepsilon}} \frac{1}{\sqrt{n}} \sum_{\varepsilon_1,\varepsilon_2=\pm1} \sum_{\substack{D_1,D_2\ge1\\ p\mid D_1,\ p\mid D_2}} \frac{S^{(p)}(\varepsilon_2 n,\varepsilon_1p,s,r,D_1,D_2)}{D_1D_2}\Phi_{w_6,H_n^-}\left(-\frac{\varepsilon_2snD_2}{D_1^2},-\frac{\varepsilon_1prD_1}{D_2^2}\right). \tag*{(67)} S 6 − := P p n ≤ p 1/2 + ε ∑ n 1 ε 1 , ε 2 = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , p ∣ D 2 ∑ D 1 D 2 S ( p ) ( ε 2 n , ε 1 p , s , r , D 1 , D 2 ) Φ w 6 , H n − ( − D 1 2 ε 2 s n D 2 , − D 2 2 ε 1 p r D 1 ) . (67)
Thus S i − \mathcal{S}_i^- S i − includes the factor p \sqrt{p} p but not the minus sign in the moment expansion. The four remaining spectral contributions are
R o l d : = c c u s p ∑ n ≤ p 1 / 2 + ε 1 n ∑ π ∈ B ( 1 ) ∑ j = 0 2 1 L ( 1 , π , Ad ) ( B π ( j ) ( r , s ) ‾ B π ( j ) ( 1 , n ) H n + ( μ π ) − p B π ( j ) ( s , r ) ‾ B π ( j ) ( p , n ) H n − ( μ π ) ) , \mathcal{R}_{\mathrm{old}} := c_{\mathrm{cusp}} \sum_{n\le p^{1/2+\varepsilon}} \frac{1}{\sqrt{n}} \sum_{\pi\in\mathcal{B}(1)} \sum_{j=0}^{2} \frac{1}{L(1,\pi,\operatorname{Ad})}\left(\overline{B_\pi^{(j)}(r,s)}B_\pi^{(j)}(1,n)H_n^+(\mu_\pi)-\sqrt{p}\,\overline{B_\pi^{(j)}(s,r)}B_\pi^{(j)}(p,n)H_n^-(\mu_\pi)\right), R old := c cusp n ≤ p 1/2 + ε ∑ n 1 π ∈ B ( 1 ) ∑ j = 0 ∑ 2 L ( 1 , π , Ad ) 1 ( B π ( j ) ( r , s ) B π ( j ) ( 1 , n ) H n + ( μ π ) − p B π ( j ) ( s , r ) B π ( j ) ( p , n ) H n − ( μ π ) ) ,
R min : = c min ∑ n ≤ p 1 / 2 + ε 1 n ∑ j = 0 2 ∫ Re ( s 1 ) = 0 ∫ Re ( s 2 ) = 0 1 ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 ( B s ( j ) ( r , s ) ‾ B s ( j ) ( 1 , n ) H n + ( s ) − p B s ( j ) ( s , r ) ‾ B s ( j ) ( p , n ) H n − ( s ) ) d s 1 d s 2 , \mathcal{R}_{\min} := c_{\min}\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\sum_{j=0}^{2}\int_{\operatorname{Re}(s_1)=0}\int_{\operatorname{Re}(s_2)=0}\frac{1}{\prod_{1\le i<k\le3}|\zeta(1+s_i-s_k)|^2}\left(\overline{B_s^{(j)}(r,s)}B_s^{(j)}(1,n)H_n^+(s)-\sqrt{p}\,\overline{B_s^{(j)}(s,r)}B_s^{(j)}(p,n)H_n^-(s)\right)\,ds_1\,ds_2, R m i n := c m i n n ≤ p 1/2 + ε ∑ n 1 j = 0 ∑ 2 ∫ Re ( s 1 ) = 0 ∫ Re ( s 2 ) = 0 ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 1 ( B s ( j ) ( r , s ) B s ( j ) ( 1 , n ) H n + ( s ) − p B s ( j ) ( s , r ) B s ( j ) ( p , n ) H n − ( s ) ) d s 1 d s 2 ,
R max , n e w : = χ p c max ∑ n ≤ p 1 / 2 + ε 1 n ∑ ϕ ∈ B 2 ( p ) ∫ Re ( z ) = 0 1 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 ( A ϕ , z ( r , s ) ‾ A ϕ , z ( 1 , n ) H n + ( z + τ ϕ , z − τ ϕ , − 2 z ) − p A ϕ , z ( s , r ) ‾ A ϕ , z ( p , n ) H n − ( z + τ ϕ , z − τ ϕ , − 2 z ) ) d z , \mathcal{R}_{\max,\mathrm{new}} := \chi_p c_{\max}\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\sum_{\phi\in\mathcal{B}_2(p)}\int_{\operatorname{Re}(z)=0}\frac{1}{L(1,\phi,\operatorname{Ad})|L(1+3z,\phi)|^2}\left(\overline{A_{\phi,z}(r,s)}A_{\phi,z}(1,n)H_n^+(z+\tau_\phi,z-\tau_\phi,-2z)-\sqrt{p}\,\overline{A_{\phi,z}(s,r)}A_{\phi,z}(p,n)H_n^-(z+\tau_\phi,z-\tau_\phi,-2z)\right)\,dz, R m a x , new := χ p c m a x n ≤ p 1/2 + ε ∑ n 1 ϕ ∈ B 2 ( p ) ∑ ∫ Re ( z ) = 0 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 1 ( A ϕ , z ( r , s ) A ϕ , z ( 1 , n ) H n + ( z + τ ϕ , z − τ ϕ , − 2 z ) − p A ϕ , z ( s , r ) A ϕ , z ( p , n ) H n − ( z + τ ϕ , z − τ ϕ , − 2 z ) ) d z ,
R max , o l d : = c max ∑ n ≤ p 1 / 2 + ε 1 n ∑ ϕ ∈ B 2 ( 1 ) ∑ j = 0 2 ∫ Re ( z ) = 0 1 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 ( B ϕ , z ( j ) ( r , s ) ‾ B ϕ , z ( j ) ( 1 , n ) H n + ( z + τ ϕ , z − τ ϕ , − 2 z ) − p B ϕ , z ( j ) ( s , r ) ‾ B ϕ , z ( j ) ( p , n ) H n − ( z + τ ϕ , z − τ ϕ , − 2 z ) ) d z . \mathcal{R}_{\max,\mathrm{old}} := c_{\max}\sum_{n\le p^{1/2+\varepsilon}}\frac{1}{\sqrt{n}}\sum_{\phi\in\mathcal{B}_2(1)}\sum_{j=0}^{2}\int_{\operatorname{Re}(z)=0}\frac{1}{L(1,\phi,\operatorname{Ad})|L(1+3z,\phi)|^2}\left(\overline{B_{\phi,z}^{(j)}(r,s)}B_{\phi,z}^{(j)}(1,n)H_n^+(z+\tau_\phi,z-\tau_\phi,-2z)-\sqrt{p}\,\overline{B_{\phi,z}^{(j)}(s,r)}B_{\phi,z}^{(j)}(p,n)H_n^-(z+\tau_\phi,z-\tau_\phi,-2z)\right)\,dz. R m a x , old := c m a x n ≤ p 1/2 + ε ∑ n 1 ϕ ∈ B 2 ( 1 ) ∑ j = 0 ∑ 2 ∫ Re ( z ) = 0 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 1 ( B ϕ , z ( j ) ( r , s ) B ϕ , z ( j ) ( 1 , n ) H n + ( z + τ ϕ , z − τ ϕ , − 2 z ) − p B ϕ , z ( j ) ( s , r ) B ϕ , z ( j ) ( p , n ) H n − ( z + τ ϕ , z − τ ϕ , − 2 z ) ) d z .
The continuous integrals use the measures specified in § 3.5. Set
S : = ∑ i = 4 6 ( S i + − S i − ) , \mathcal{S}:=\sum_{i=4}^{6}\left(\mathcal{S}_i^+-\mathcal{S}_i^-\right), S := i = 4 ∑ 6 ( S i + − S i − ) ,
R : = R o l d + R min + R max , n e w + R max , o l d . \mathcal{R}:=\mathcal{R}_{\mathrm{old}}+\mathcal{R}_{\min}+\mathcal{R}_{\max,\mathrm{new}}+\mathcal{R}_{\max,\mathrm{old}}. R := R old + R m i n + R m a x , new + R m a x , old .
Then the trace expansion is
χ p M h ( r , s , p ) = D + S − R + O Q , η , h 0 , J , ε , B ( p − B ) . (68) \chi_p M_h(r,s,p)=\mathcal{D}+\mathcal{S}-\mathcal{R}+O_{Q,\eta,h_0,J,\varepsilon,B}\left(p^{-B}\right). \tag*{(68)} χ p M h ( r , s , p ) = D + S − R + O Q , η , h 0 , J , ε , B ( p − B ) . (68)
The diagonal term
The first diagonal in (66) requires r = 1 r=1 r = 1 and n = s n=s n = s , which is included for large p p p since s ≤ Q s\le Q s ≤ Q . The dual diagonal requires s = p s=p s = p and r = n r=n r = n , contrary to ( s , p ) = 1 (s,p)=1 ( s , p ) = 1 . Hence
D = δ r = 1 P s ∫ Re ( μ ) = 0 h ( μ ) V μ ( s / p ) spec ( μ ) d μ . \mathcal{D}=\frac{\delta_{r=1}P}{\sqrt{s}}\int_{\operatorname{Re}(\mu)=0}h(\mu)V_{\mu}(s/\sqrt{p})\operatorname{spec}(\mu)\,d\mu. D = s δ r = 1 P ∫ Re ( μ ) = 0 h ( μ ) V μ ( s / p ) spec ( μ ) d μ .
For sufficiently large p p p , s / p ≤ 1 s/\sqrt{p}\le1 s / p ≤ 1 uniformly for s ≤ Q s\le Q s ≤ Q , so (25) gives V μ ( s / p ) = 1 + O J ( s 2 p − 1 ( 1 + ∥ μ ∥ ) C ) V_{\mu}(s/\sqrt{p})=1+O_J(s^2p^{-1}(1+\lVert\mu\rVert)^C) V μ ( s / p ) = 1 + O J ( s 2 p − 1 ( 1 + ∥ μ ∥ ) C ) . Using (38) , we obtain
D = δ r = 1 P s H h + O Q , h , J ( p − 1 ) . (69) \mathcal{D}=\frac{\delta_{r=1}P}{\sqrt{s}}\mathcal{H}_h+O_{Q,h,J}(p^{-1}). \tag*{(69)} D = s δ r = 1 P H h + O Q , h , J ( p − 1 ) . (69)
The short Weyl elements
Lemma 4.2. The four short Weyl sums in J \mathcal{J} J are O Q , η , h 0 , J ( p − 1 ) O_{Q,\eta,h_0,J}(p^{-1}) O Q , η , h 0 , J ( p − 1 ) , uniformly in the range of Theorem 4.1 .
Proof. Choose a fixed decay exponent A A A sufficiently large in terms of η \eta η , and take a 0 a_0 a 0 large enough for the corresponding kernel estimates. Counting the summands in (48) gives ∣ S ~ ∣ ≤ D 1 D 2 |\widetilde{\mathcal{S}}|\le D_1D_2 ∣ S ∣ ≤ D 1 D 2 . Combining (60) with c = 2 ε c=2\varepsilon c = 2 ε and (53) , we obtain, for either sign,
∣ Φ w 4 , H ± ( y ) ∣ + ∣ Φ w 5 , H ± ( y ) ∣ ≪ h 0 , J , A , ε p ε ∣ y ∣ A . |\Phi_{w_4,H^\pm}(y)|+|\Phi_{w_5,H^\pm}(y)|\ll_{h_0,J,A,\varepsilon}p^\varepsilon|y|^A. ∣ Φ w 4 , H ± ( y ) ∣ + ∣ Φ w 5 , H ± ( y ) ∣ ≪ h 0 , J , A , ε p ε ∣ y ∣ A .
For ( m 1 , m 2 , n 1 , n 2 ) = ( r , s , 1 , n ) (m_1,m_2,n_1,n_2)=(r,s,1,n) ( m 1 , m 2 , n 1 , n 2 ) = ( r , s , 1 , n ) , the w 4 w_4 w 4 conditions s D 1 = D 2 2 sD_1=D_2^2 s D 1 = D 2 2 and p D 2 ∣ D 1 pD_2\mid D_1 p D 2 ∣ D 1 give
D 2 = s p k , D 1 = s p 2 k , ∣ y ∣ = r n s p 3 k 3 , k ≥ 1. D_2=sp^k,\qquad D_1=sp^{2k},\qquad|y|=\frac{rn}{sp^3k^3},\qquad k\ge1. D 2 = s p k , D 1 = s p 2 k , ∣ y ∣ = s p 3 k 3 r n , k ≥ 1.
For w 5 w_5 w 5 , put g = ( r , n ) g=(r,n) g = ( r , n ) , r ′ = r / g r'=r/g r ′ = r / g , n ′ = n / g n'=n/g n ′ = n / g . Writing D 2 = D 1 d D_2=D_1d D 2 = D 1 d gives r ′ d = n ′ D 1 r'd=n'D_1 r ′ d = n ′ D 1 . Since ( r ′ , n ′ ) = 1 (r',n')=1 ( r ′ , n ′ ) = 1 and ( r ′ , p ) = 1 (r',p)=1 ( r ′ , p ) = 1 , all solutions are
D 1 = p r ′ k , D 2 = p 2 r ′ n ′ k 2 , ∣ y ∣ = r s p 3 r ′ 2 n ′ k 3 . D_1=pr'k,\qquad D_2=p^2r'n'k^2,\qquad|y|=\frac{rs}{p^3r'^2n'k^3}. D 1 = p r ′ k , D 2 = p 2 r ′ n ′ k 2 , ∣ y ∣ = p 3 r ′2 n ′ k 3 r s .
For ( s , r , p , n ) (s,r,p,n) ( s , r , p , n ) , the w 4 w_4 w 4 conditions r D 1 = p D 2 2 rD_1=pD_2^2 r D 1 = p D 2 2 and p D 2 ∣ D 1 pD_2\mid D_1 p D 2 ∣ D 1 give
D 2 = r k , D 1 = p r k 2 , ∣ y ∣ = s n p r k 3 . D_2=rk,\qquad D_1=prk^2,\qquad|y|=\frac{sn}{prk^3}. D 2 = r k , D 1 = p r k 2 , ∣ y ∣ = p r k 3 s n .
For its w 5 w_5 w 5 term, put g = ( s , n ) g=(s,n) g = ( s , n ) , s ′ = s / g s'=s/g s ′ = s / g , n ′ = n / g n'=n/g n ′ = n / g to obtain
D 1 = p s ′ k , D 2 = p 2 s ′ n ′ k 2 , ∣ y ∣ = s r p 2 s ′ 2 n ′ k 3 . D_1=ps'k,\qquad D_2=p^2s'n'k^2,\qquad|y|=\frac{sr}{p^2s'^2n'k^3}. D 1 = p s ′ k , D 2 = p 2 s ′ n ′ k 2 , ∣ y ∣ = p 2 s ′2 n ′ k 3 sr .
Summing k − 3 A k^{-3A} k − 3 A , using s ≤ Q s\le Q s ≤ Q and r ′ 2 n ′ r'^2n' r ′2 n ′ , s ′ 2 n ′ ≥ 1 s'^2n'\ge1 s ′2 n ′ ≥ 1 , and then summing over n n n gives
∣ S 4 + ∣ ≪ Q , h 0 , J , A , ε p 2 + ε − 3 A r A p ( 1 / 2 + ε ) ( A + 1 / 2 ) , |\mathcal{S}_4^+|\ll_{Q,h_0,J,A,\varepsilon}p^{2+\varepsilon-3A}r_Ap^{(1/2+\varepsilon)(A+1/2)}, ∣ S 4 + ∣ ≪ Q , h 0 , J , A , ε p 2 + ε − 3 A r A p ( 1/2 + ε ) ( A + 1/2 ) ,
∣ S 5 + ∣ ≪ Q , h 0 , J , A , ε p 2 + ε − 3 A r A p 1 / 4 + ε / 2 , |\mathcal{S}_5^+|\ll_{Q,h_0,J,A,\varepsilon}p^{2+\varepsilon-3A}r_Ap^{1/4+\varepsilon/2}, ∣ S 5 + ∣ ≪ Q , h 0 , J , A , ε p 2 + ε − 3 A r A p 1/4 + ε /2 ,
∣ S 4 − ∣ ≪ Q , h 0 , J , A , ε p 5 / 2 + ε − A p ( 1 / 2 + ε ) ( A + 1 / 2 ) , |\mathcal{S}_4^-|\ll_{Q,h_0,J,A,\varepsilon}p^{5/2+\varepsilon-A}p^{(1/2+\varepsilon)(A+1/2)}, ∣ S 4 − ∣ ≪ Q , h 0 , J , A , ε p 5/2 + ε − A p ( 1/2 + ε ) ( A + 1/2 ) ,
∣ S 5 − ∣ ≪ Q , h 0 , J , A , ε p 5 / 2 + ε − 2 A r A p 1 / 4 + ε / 2 . |\mathcal{S}_5^-|\ll_{Q,h_0,J,A,\varepsilon}p^{5/2+\varepsilon-2A}r_Ap^{1/4+\varepsilon/2}. ∣ S 5 − ∣ ≪ Q , h 0 , J , A , ε p 5/2 + ε − 2 A r A p 1/4 + ε /2 .
Since r ≤ Q p η r\le Qp^\eta r ≤ Q p η and η < 1 / 2 \eta<1/2 η < 1/2 , a sufficiently large fixed A A A and the arbitrarily small ε \varepsilon ε make all four terms O Q , η , h 0 , J ( p − 1 ) O_{Q,\eta,h_0,J}(p^{-1}) O Q , η , h 0 , J ( p − 1 ) .
:::
The long Weyl element
Lemma 4.3. The two long Weyl sums in S \mathcal{S} S have total size O Q , η , h 0 , J ( p − 1 ) O_{Q,\eta,h_0,J}(p^{-1}) O Q , η , h 0 , J ( p − 1 ) , uniformly in the range of Theorem 4.1 .
Proof. The kernel arguments in (67) , with ( m 1 , m 2 , n 1 , n 2 ) = ( r , s , 1 , n ) (m_1,m_2,n_1,n_2)=(r,s,1,n) ( m 1 , m 2 , n 1 , n 2 ) = ( r , s , 1 , n ) or ( s , r , p , n ) (s,r,p,n) ( s , r , p , n ) , satisfy
∣ y 2 2 y 2 ∣ = m 1 2 m 2 n 1 n 2 2 D 1 3 , |y_2^2y_2|=\frac{m_1^2m_2n_1n_2^2}{D_1^3}, ∣ y 2 2 y 2 ∣ = D 1 3 m 1 2 m 2 n 1 n 2 2 ,
∣ y 1 y 2 2 ∣ = m 1 m 2 2 n 1 2 n 2 D 2 3 . |y_1y_2^2|=\frac{m_1m_2^2n_1^2n_2}{D_2^3}. ∣ y 1 y 2 2 ∣ = D 2 3 m 1 m 2 2 n 1 2 n 2 .
Hence
min ( ∣ y 2 2 y 2 ∣ , ∣ y 1 y 2 2 ∣ ) A ≤ ( m 1 m 2 n 1 n 2 ) 3 A / 2 ( D 1 D 2 ) 3 A / 2 . \min(|y_2^2y_2|,|y_1y_2^2|)^A\leq\frac{(m_1m_2n_1n_2)^{3A/2}}{(D_1D_2)^{3A/2}}. min ( ∣ y 2 2 y 2 ∣ , ∣ y 1 y 2 2 ∣ ) A ≤ ( D 1 D 2 ) 3 A /2 ( m 1 m 2 n 1 n 2 ) 3 A /2 .
For sufficiently large p p p , we have r , s , n < p r,s,n<p r , s , n < p uniformly in the stated range. At ( r , s , 1 , n ) (r,s,1,n) ( r , s , 1 , n ) and ( s , r , p , n ) (s,r,p,n) ( s , r , p , n ) , the conditions D 1 , D 2 ≥ p D_1,D_2\geq p D 1 , D 2 ≥ p therefore give
1 + ∣ y 1 ∣ + ∣ y 2 ∣ ≪ D 1 D 2 . 1+|y_1|+|y_2|\ll D_1D_2. 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ≪ D 1 D 2 .
By (60) with c = 2 ε c=2\varepsilon c = 2 ε , (54) applies to H n ± H_n^\pm H n ± with an additional factor p ε p^\varepsilon p ε . Counting the variables in (49) gives ∣ S ( p ) ∣ ≤ ( D 1 D 2 ) 2 |\mathcal{S}^{(p)}|\leq(D_1D_2)^2 ∣ S ( p ) ∣ ≤ ( D 1 D 2 ) 2 . For fixed n , ϵ 1 , ϵ 2 n,\epsilon_1,\epsilon_2 n , ϵ 1 , ϵ 2 , (54) gives
p ∣ S ( p ) ( ϵ 2 n 2 , ϵ 1 n 1 , m 1 , m 2 , D 1 , D 2 ) ∣ D 1 D 2 ∣ Φ w 6 , H n ± ( y 1 , y 2 ) ∣ ≪ h 0 , J , A , ε p 2 + ε ( m 1 m 2 n 1 n 2 ) 3 A / 2 ( D 1 D 2 ) 1 − 3 A / 2 + ε . p\frac{|\mathcal{S}^{(p)}(\epsilon_2n_2,\epsilon_1n_1,m_1,m_2,D_1,D_2)|}{D_1D_2}\left|\Phi_{w_6,H_n^\pm}(y_1,y_2)\right|
\ll_{h_0,J,A,\varepsilon}p^{2+\varepsilon}(m_1m_2n_1n_2)^{3A/2}(D_1D_2)^{1-3A/2+\varepsilon}. p D 1 D 2 ∣ S ( p ) ( ϵ 2 n 2 , ϵ 1 n 1 , m 1 , m 2 , D 1 , D 2 ) ∣ Φ w 6 , H n ± ( y 1 , y 2 ) ≪ h 0 , J , A , ε p 2 + ε ( m 1 m 2 n 1 n 2 ) 3 A /2 ( D 1 D 2 ) 1 − 3 A /2 + ε .
We first sum over D 1 , D 2 D_1,D_2 D 1 , D 2 with p ∣ D 1 , D 2 p\mid D_1,D_2 p ∣ D 1 , D 2 . For sufficiently large A A A , we have
∑ p ∣ D 1 , D 2 ( D 1 D 2 ) 1 − 3 A / 2 + ε = p 2 − 3 A + 2 ε ζ ( 3 A / 2 − 1 − ε ) 2 . \sum_{p\mid D_1,D_2}(D_1D_2)^{1-3A/2+\varepsilon}
=p^{2-3A+2\varepsilon}\zeta(3A/2-1-\varepsilon)^2. p ∣ D 1 , D 2 ∑ ( D 1 D 2 ) 1 − 3 A /2 + ε = p 2 − 3 A + 2 ε ζ ( 3 A /2 − 1 − ε ) 2 .
Since m 1 m 2 n 1 n 2 m_1m_2n_1n_2 m 1 m 2 n 1 n 2 equals r s n rsn r s n in the first piece and r s p n rspn r s p n in the dual piece, summing over n ≤ p 1 / 2 + ε n\leq p^{1/2+\varepsilon} n ≤ p 1/2 + ε with weight n − 1 / 2 n^{-1/2} n − 1/2 and over ϵ 1 , ϵ 2 = ± 1 \epsilon_1,\epsilon_2=\pm1 ϵ 1 , ϵ 2 = ± 1 , and including the additional factor p \sqrt{p} p in the dual term, gives
P 6 + ≪ Q , h 0 , J , A , ε p 4 + 3 ε − 3 A r 3 A / 2 p ( 1 / 2 + ε ) ( 3 A / 2 + 1 / 2 ) , \mathcal{P}_6^+\ll_{Q,h_0,J,A,\varepsilon}p^{4+3\varepsilon-3A}r^{3A/2}p^{(1/2+\varepsilon)(3A/2+1/2)}, P 6 + ≪ Q , h 0 , J , A , ε p 4 + 3 ε − 3 A r 3 A /2 p ( 1/2 + ε ) ( 3 A /2 + 1/2 ) ,
P 6 − ≪ Q , h 0 , J , A , ε p 9 / 2 + 3 ε − 3 A / 2 r 3 A / 2 p ( 1 / 2 + ε ) ( 3 A / 2 + 1 / 2 ) . \mathcal{P}_6^-\ll_{Q,h_0,J,A,\varepsilon}p^{9/2+3\varepsilon-3A/2}r^{3A/2}p^{(1/2+\varepsilon)(3A/2+1/2)}. P 6 − ≪ Q , h 0 , J , A , ε p 9/2 + 3 ε − 3 A /2 r 3 A /2 p ( 1/2 + ε ) ( 3 A /2 + 1/2 ) .
Since r ≤ Q p η r\leq Qp^\eta r ≤ Q p η and η < 1 / 2 \eta<1/2 η < 1/2 , taking A A A sufficiently large in terms of η \eta η gives the stated bound. □
The old and minimal Eisenstein spectra
The following lemma records the cancellation in the dual contribution obtained by summing over the three oldvectors.
Lemma 4.4. Let Π \Pi Π be a cuspidal, minimal Eisenstein, or maximal Eisenstein representation of level one. Suppose its unramified coefficients satisfy ∣ A Π ( m , n ) ∣ ≪ ε ( m n ) θ + ε |A_\Pi(m,n)|\ll_\varepsilon(mn)^{\theta+\varepsilon} ∣ A Π ( m , n ) ∣ ≪ ε ( mn ) θ + ε , with respectively θ = 5 / 14 , 0 , \theta=5/14,\,0, θ = 5/14 , 0 , or 7 / 64 7/64 7/64 . For ( m 1 m 2 , p ) = 1 (m_1m_2,p)=1 ( m 1 m 2 , p ) = 1 and n < p n<p n < p ,
∑ j = 0 2 B Π ( j ) ( m 1 , m 2 ) B Π ( j ) ( 1 , n ) = A Π ( m 1 , m 2 ) A Π ( 1 , n ) ( Q Π − 1 ) 00 , \sum_{j=0}^{2}B_\Pi^{(j)}(m_1,m_2)B_\Pi^{(j)}(1,n)=A_\Pi(m_1,m_2)A_\Pi(1,n)(Q_\Pi^{-1})_{00}, j = 0 ∑ 2 B Π ( j ) ( m 1 , m 2 ) B Π ( j ) ( 1 , n ) = A Π ( m 1 , m 2 ) A Π ( 1 , n ) ( Q Π − 1 ) 00 , ∑ j = 0 2 B Π ( j ) ( m 1 , m 2 ) ‾ B Π ( j ) ( p , n ) ≪ ∣ A Π ( m 1 , m 2 ) A Π ( 1 , n ) ∣ p θ − 1 . \sum_{j=0}^{2}\overline{B_\Pi^{(j)}(m_1,m_2)}B_\Pi^{(j)}(p,n)\ll|A_\Pi(m_1,m_2)A_\Pi(1,n)|p^{\theta-1}. j = 0 ∑ 2 B Π ( j ) ( m 1 , m 2 ) B Π ( j ) ( p , n ) ≪ ∣ A Π ( m 1 , m 2 ) A Π ( 1 , n ) ∣ p θ − 1 .
Also ( Q Π − 1 ) 00 ≪ 1 (Q_{\Pi}^{-1})_{00} \ll1 ( Q Π − 1 ) 00 ≪ 1 , uniformly as p → ∞ p \to\infty p → ∞ .
Proof. Put λ = A Π ( 1 , p ) \lambda= A_{\Pi}(1,p) λ = A Π ( 1 , p ) in (41) . Since Π \Pi Π is unramified at p p p , ∣ λ ∣ ≪ p θ |\lambda| \ll p^{\theta} ∣ λ ∣ ≪ p θ and θ < 1 / 2 \theta< 1/2 θ < 1/2 give
Q Π = I + O ( p θ − 1 / 2 ) , ( Q Π − 1 ) 00 = O ( 1 ) , ( Q Π − 1 ) 01 = O ( p θ − 1 / 2 ) . Q_{\Pi}=I+O(p^{\theta-1/2}),\qquad(Q_{\Pi}^{-1})_{00}=O(1),\qquad(Q_{\Pi}^{-1})_{01}=O(p^{\theta-1/2}). Q Π = I + O ( p θ − 1/2 ) , ( Q Π − 1 ) 00 = O ( 1 ) , ( Q Π − 1 ) 01 = O ( p θ − 1/2 ) .
including b p ∣ λ ∣ 2 = O ( p 2 θ − 1 ) b_p|\lambda|^2=O(p^{2\theta-1}) b p ∣ λ ∣ 2 = O ( p 2 θ − 1 ) . Thus, uniformly in Π \Pi Π ,
Let E \mathcal{E} E be the change of basis from the original oldvectors to the orthogonalized oldvectors in (42) . Since the inner product is linear in its first variable, E Q Π E ∗ = I \mathcal{E}Q_{\Pi}\mathcal{E}^{*}=I E Q Π E ∗ = I and E ∗ E = Q Π − 1 \mathcal{E}^{*}\mathcal{E}=Q_{\Pi}^{-1} E ∗ E = Q Π − 1 . For coefficient columns a , b a,b a , b , the sum over the three oldvectors is therefore
∑ j = 0 2 ( E a ) j ( E b ) j = a ∗ Q Π − 1 b . \sum_{j=0}^{2}(\mathcal{E}a)_{j}(\mathcal{E}b)_{j}=a^{*}Q_{\Pi}^{-1}b. j = 0 ∑ 2 ( E a ) j ( E b ) j = a ∗ Q Π − 1 b .
For ( m 1 m 2 , p ) = 1 (m_{1}m_{2},p)=1 ( m 1 m 2 , p ) = 1 and n < p n<p n < p , the coefficient columns for the original oldvectors at ( m 1 , m 2 ) (m_{1},m_{2}) ( m 1 , m 2 ) and ( 1 , n ) (1,n) ( 1 , n ) are A Π ( m 1 , m 2 ) e 0 A_{\Pi}(m_{1},m_{2})e_{0} A Π ( m 1 , m 2 ) e 0 and A Π ( 1 , n ) e 0 A_{\Pi}(1,n)e_{0} A Π ( 1 , n ) e 0 , respectively, where e 0 = ( 1 , 0 , 0 ) t e_{0}=(1,0,0)^{\mathsf{t}} e 0 = ( 1 , 0 , 0 ) t . This proves the first identity. At ( p , n ) (p,n) ( p , n ) , multiplicativity, unitarity and p ∤ n p\nmid n p ∤ n give
A Π ( p , n ) = λ ‾ A Π ( 1 , n ) , A_{\Pi}(p,n)=\overline{\lambda}A_{\Pi}(1,n), A Π ( p , n ) = λ A Π ( 1 , n ) ,
A Π ( 1 ) ( p , n ) = p λ A Π ( 1 , n ) , A_{\Pi}^{(1)}(p,n)=\sqrt{p}\lambda A_{\Pi}(1,n), A Π ( 1 ) ( p , n ) = p λ A Π ( 1 , n ) ,
A Π ( 2 ) ( p , n ) = p A Π ( 1 , n ) . A_{\Pi}^{(2)}(p,n)=pA_{\Pi}(1,n). A Π ( 2 ) ( p , n ) = p A Π ( 1 , n ) .
Thus the coefficient column is A Π ( 1 , n ) v A_{\Pi}(1,n)v A Π ( 1 , n ) v , with v = ( λ ‾ , p λ , p ) t v=(\overline{\lambda},\sqrt{p}\lambda,p)^{\mathsf{t}} v = ( λ , p λ , p ) t . The definitions of a p , b p a_p,b_p a p , b p give
1 − p b p = p − 1 / 2 a p , p − p a p = p − 1 / 2 b p . 1-pb_p=p^{-1/2}a_p,\qquad\sqrt{p}-pa_p=p^{-1/2}b_p. 1 − p b p = p − 1/2 a p , p − p a p = p − 1/2 b p .
Consequently
v − p Q Π e 2 = p − 1 / 2 ( a p λ ‾ , b p λ , 0 ) t , e 2 = ( 0 , 0 , 1 ) t . v-pQ_{\Pi}e_{2}=p^{-1/2}(a_p\overline{\lambda},b_p\lambda,0)^{\mathsf{t}},\qquad e_{2}=(0,0,1)^{\mathsf{t}}. v − p Q Π e 2 = p − 1/2 ( a p λ , b p λ , 0 ) t , e 2 = ( 0 , 0 , 1 ) t .
Since e 0 t e 2 = 0 e_{0}^{\mathsf{t}}e_{2}=0 e 0 t e 2 = 0 , this yields
e 0 t Q Π − 1 v = p − 1 / 2 ( a p λ ‾ ( Q Π − 1 ) 00 + b p λ ( Q Π − 1 ) 01 ) ≪ p θ − 1 + p 2 θ − 2 ≪ p θ − 1 . e_{0}^{\mathsf{t}}Q_{\Pi}^{-1}v=p^{-1/2}\left(a_p\overline{\lambda}(Q_{\Pi}^{-1})_{00}+b_p\lambda(Q_{\Pi}^{-1})_{01}\right)\ll p^{\theta-1}+p^{2\theta-2}\ll p^{\theta-1}. e 0 t Q Π − 1 v = p − 1/2 ( a p λ ( Q Π − 1 ) 00 + b p λ ( Q Π − 1 ) 01 ) ≪ p θ − 1 + p 2 θ − 2 ≪ p θ − 1 .
Taking absolute values before summing over the three oldvectors would lose this saving. This proves the second assertion. The cancellation holds in the weighted trace because H n ± H_n^{\pm} H n ± is common to all three oldvectors. □
Proposition 4.5. Uniformly in the range of Theorem 4.1 ,
∣ R o l d ∣ + ∣ R m i n ∣ + ∣ R m a x , o l d ∣ ≪ h 0 , J , ϵ ( r s ) 5 / 14 + ϵ p 3 / 7 + ϵ . |\mathcal{R}_{\mathrm{old}}|+|\mathcal{R}_{\mathrm{min}}|+|\mathcal{R}_{\mathrm{max,old}}|\ll_{h_0,J,\epsilon}(rs)^{5/14+\epsilon}p^{3/7+\epsilon}. ∣ R old ∣ + ∣ R min ∣ + ∣ R max , old ∣ ≪ h 0 , J , ϵ ( r s ) 5/14 + ϵ p 3/7 + ϵ .
Proof. Lemma 2.5 with j = 0 j=0 j = 0 gives, uniformly in the three level one sectors and n ≤ p 1 / 2 + ϵ n\le p^{1/2+\epsilon} n ≤ p 1/2 + ϵ ,
∣ H n + ( μ ) ∣ + ∣ H n − ( μ ) ∣ ≪ J ( ∣ h ( μ ) ∣ + ∣ h ( − μ ) ∣ ) ( 1 + ∥ μ ∥ ) C . |H_n^{+}(\mu)|+|H_n^{-}(\mu)|\ll_J (|h(\mu)|+|h(-\mu)|)(1+\|\mu\|)^C. ∣ H n + ( μ ) ∣ + ∣ H n − ( μ ) ∣ ≪ J ( ∣ h ( μ ) ∣ + ∣ h ( − μ ) ∣ ) ( 1 + ∥ μ ∥ ) C .
Apply Lemma 4.4 to ( r , s ) (r,s) ( r , s ) and ( s , r ) (s,r) ( s , r ) in the first and dual pieces. The twist coefficient is O ϵ ( ( r s ) θ + ϵ ) O_{\epsilon}((rs)^{\theta+\epsilon}) O ϵ (( r s ) θ + ϵ ) , and the first piece, apart from the rapidly decreasing spectral factor, is bounded by
( r s ) θ + ϵ ∑ n ≤ p 1 / 2 + ϵ n − 1 / 2 + θ + ϵ ≪ ϵ ( r s ) θ + ϵ p ( 1 / 2 + ϵ ) ( 1 / 2 + θ + ϵ ) . (rs)^{\theta+\epsilon}\sum_{n\le p^{1/2+\epsilon}}n^{-1/2+\theta+\epsilon}\ll_{\epsilon}(rs)^{\theta+\epsilon}p^{(1/2+\epsilon)(1/2+\theta+\epsilon)}. ( r s ) θ + ϵ n ≤ p 1/2 + ϵ ∑ n − 1/2 + θ + ϵ ≪ ϵ ( r s ) θ + ϵ p ( 1/2 + ϵ ) ( 1/2 + θ + ϵ ) .
By Lemma 4.4 , the additional factor in the dual contribution is p p θ − 1 = p θ − 1 / 2 ≤ 1 \sqrt{p}\,p^{\theta-1}=p^{\theta-1/2}\le1 p p θ − 1 = p θ − 1/2 ≤ 1 .
Brumley’s lower bounds for Rankin–Selberg residues and values [14 Theorems 3 and 5] , together with an estimate for 1 / ζ ( s ) 1/\zeta(s) 1/ ζ ( s ) [27 Theorem 8.29] , give, at level one and for Re ( s i ) = Re ( z ) = 0 \operatorname{Re}(s_i)=\operatorname{Re}(z)=0 Re ( s i ) = Re ( z ) = 0 ,
L ( 1 , Π , Ad ) − 1 ≪ ( 1 + ∥ μ Π ∥ ) C , Π ∈ B ( 1 ) , ∏ i < j ∣ ζ ( 1 + s i − s j ) ∣ − 2 ≪ ( 1 + ∥ s ∥ ) C , 1 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 ≪ ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ ) C , ϕ ∈ B 2 ( 1 ) . (70) \begin{aligned}
L(1,\Pi,\operatorname{Ad})^{-1} &\ll(1+\lVert\mu_\Pi\rVert)^C, && \Pi\in\mathcal{B}(1),\\
\prod_{i<j}\lvert\zeta(1+s_i-s_j)\rvert^{-2} &\ll(1+\lVert s\rVert)^C,\\
\frac{1}{L(1,\phi,\operatorname{Ad})\lvert L(1+3z,\phi)\rvert^2} &\ll(1+\lvert z\rvert+\lvert\tau_\phi\rvert)^C, && \phi\in\mathcal{B}_2(1). \tag*{(70)}
\end{aligned} L ( 1 , Π , Ad ) − 1 i < j ∏ ∣ ζ ( 1 + s i − s j ) ∣ − 2 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 1 ≪ ( 1 + ∥ μ Π ∥ ) C , ≪ ( 1 + ∥ s ∥ ) C , ≪ ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ ) C , Π ∈ B ( 1 ) , ϕ ∈ B 2 ( 1 ) . (70)
Here L ( 1 , Π , Ad ) = Res w = 1 L ( w , Π × Π ~ ) L(1,\Pi,\operatorname{Ad})=\operatorname{Res}_{w=1}L(w,\Pi\times\widetilde{\Pi}) L ( 1 , Π , Ad ) = Res w = 1 L ( w , Π × Π ) , and similarly for ϕ \phi ϕ . The standard L L L -value is covered by [14 Theorem 5] with degrees 2 and 1. These factors admit continuous extensions, with value zero when the parameters coincide. Müller’s Weyl law [47 Theorem 0.1] , with n = 3 n=3 n = 3 and n = 2 n=2 n = 2 , implies, for some fixed C C C and all R ≥ 1 R\ge1 R ≥ 1 ,
# { Π ∈ B ( 1 ) : ∥ μ Π ∥ ≤ R } + # { ϕ ∈ B 2 ( 1 ) : ∣ τ ϕ ∣ ≤ R } ≪ R C . (71) \#\{\Pi\in\mathcal{B}(1):\lVert\mu_\Pi\rVert\le R\}+\#\{\phi\in\mathcal{B}_2(1):\lvert\tau_\phi\rvert\le R\}\ll R^C. \tag*{(71)} # { Π ∈ B ( 1 ) : ∥ μ Π ∥ ≤ R } + # { ϕ ∈ B 2 ( 1 ) : ∣ τ ϕ ∣ ≤ R } ≪ R C . (71)
These bounds and rapid decrease of h h h make the remaining sums and integrals finite, independently of p p p . Taking θ = 5 / 14 , 0 , 7 / 64 \theta=5/14,0,7/64 θ = 5/14 , 0 , 7/64 , respectively, we obtain
R o l d ≪ h 0 , J , ε ( r s ) 5 / 14 + ε p 3 / 7 + ε , \mathcal{R}_{\mathrm{old}}\ll_{h_0,J,\varepsilon}(rs)^{5/14+\varepsilon}p^{3/7+\varepsilon}, R old ≪ h 0 , J , ε ( r s ) 5/14 + ε p 3/7 + ε ,
R min ≪ h 0 , J , ε ( r s ) ε p 1 / 4 + ε , \mathcal{R}_{\min}\ll_{h_0,J,\varepsilon}(rs)^\varepsilon p^{1/4+\varepsilon}, R m i n ≪ h 0 , J , ε ( r s ) ε p 1/4 + ε ,
and
R max , o l d ≪ h 0 , J , ε ( r s ) 7 / 64 + ε p 39 / 128 + ε . \mathcal{R}_{\max,\mathrm{old}}\ll_{h_0,J,\varepsilon}(rs)^{7/64+\varepsilon}p^{39/128+\varepsilon}. R m a x , old ≪ h 0 , J , ε ( r s ) 7/64 + ε p 39/128 + ε .
The stated bound follows from these estimates. □
The maximal Eisenstein new spectrum
We first bound the weighted Dirichlet series arising from the maximal Eisenstein new spectrum.
Lemma 4.6. Let ϕ ∈ B 2 ( p ) \phi\in\mathcal{B}_2(p) ϕ ∈ B 2 ( p ) , z ∈ i R z\in i\mathbb{R} z ∈ i R and μ = ( z + τ ϕ , z − τ ϕ , − 2 z ) \mu=(z+\tau_\phi,z-\tau_\phi,-2z) μ = ( z + τ ϕ , z − τ ϕ , − 2 z ) . For either sign, and for some C = C ( J , ε ) C=C(J,\varepsilon) C = C ( J , ε ) , we have
∑ n ≥ 1 A ϕ , z ( 1 , n ) n U ± μ ± ( n / p ) ≪ J , ε p 1 / 4 + ε ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ ) C . \sum_{n\ge1}\frac{A_{\phi,z}(1,n)}{\sqrt{n}}U^\pm_{\pm\mu}(n/\sqrt{p})\ll_{J,\varepsilon}p^{1/4+\varepsilon}(1+\lvert z\rvert+\lvert\tau_\phi\rvert)^C. n ≥ 1 ∑ n A ϕ , z ( 1 , n ) U ± μ ± ( n / p ) ≪ J , ε p 1/4 + ε ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ ) C .
Proof. Here τ ϕ ∈ i R ∪ [ − 7 / 64 , 7 / 64 ] \tau_\phi\in i\mathbb{R}\cup[-7/64,7/64] τ ϕ ∈ i R ∪ [ − 7/64 , 7/64 ] . The convolution formula gives, in a right half plane,
∑ n ≥ 1 A ϕ , z ( 1 , n ) n w = ∑ a , b ≥ 1 λ ϕ ( a ) a z b − 2 z ( a b ) w = L ( w − z , ϕ ) ζ ( w + 2 z ) . \sum_{n\ge1}\frac{A_{\phi,z}(1,n)}{n^w}
=\sum_{a,b\ge1}\frac{\lambda_\phi(a)a^z b^{-2z}}{(ab)^w}
=L(w-z,\phi)\zeta(w+2z). n ≥ 1 ∑ n w A ϕ , z ( 1 , n ) = a , b ≥ 1 ∑ ( ab ) w λ ϕ ( a ) a z b − 2 z = L ( w − z , ϕ ) ζ ( w + 2 z ) .
On the line Re ( u ) = c > 1 \operatorname{Re}(u)=c>1 Re ( u ) = c > 1 , absolute convergence justifies interchanging the integral and the n n n -sum. This gives
1 2 π i ∫ ( c ) p u / 2 G J ( u , μ ) γ μ ( 1 / 2 + u ) γ ± μ ( 1 / 2 ) L ( 1 / 2 + u − z , ϕ ) ζ ( 1 / 2 + u + 2 z ) d u u . \frac{1}{2\pi i}\int_{(c)}p^{u/2}G_J(u,\mu)\frac{\gamma_\mu(1/2+u)}{\gamma_{\pm\mu}(1/2)}L(1/2+u-z,\phi)\zeta(1/2+u+2z)\frac{du}{u}. 2 π i 1 ∫ ( c ) p u /2 G J ( u , μ ) γ ± μ ( 1/2 ) γ μ ( 1/2 + u ) L ( 1/2 + u − z , ϕ ) ζ ( 1/2 + u + 2 z ) u d u .
The denominator signs correspond to V μ V_\mu V μ and W − μ W_{-\mu} W − μ , respectively.
Shifting the integral to Re ( u ) = 1 / 8 \operatorname{Re}(u)=1/8 Re ( u ) = 1/8 , we find that the pole of ζ \zeta ζ at u = 1 / 2 − 2 z = 1 / 2 + μ 3 u=1/2-2z=1/2+\mu_3 u = 1/2 − 2 z = 1/2 + μ 3 is cancelled by the factor ( 1 / 2 + μ 3 ) 2 − u 2 (1/2+\mu_3)^2-u^2 ( 1/2 + μ 3 ) 2 − u 2 in P J P_J P J , defined by (21) . P J ( 0 , μ ) P_J(0,\mu) P J ( 0 , μ ) is non-zero since ∣ Re ( μ j ) ∣ < 1 / 2 \lvert\operatorname{Re}(\mu_j)\rvert<1/2 ∣ Re ( μ j )∣ < 1/2 , and
Re ( 1 / 2 + u − μ j ) ≥ 1 / 2 + 1 / 8 − 7 / 64 > 0. \operatorname{Re}(1/2+u-\mu_j)\ge1/2+1/8-7/64>0. Re ( 1/2 + u − μ j ) ≥ 1/2 + 1/8 − 7/64 > 0.
Thus no gamma pole or pole at u = 0 u=0 u = 0 is crossed. By (19) and (21) ,
∣ P J ( u , μ ) ζ ( 1 / 2 + u + 2 z ) ∣ ≪ J ( 1 + ∥ μ ∥ + ∣ u ∣ ) C J , Re ( u ) ≥ 1 / 8. |P_J(u,\mu)\zeta(1/2+u+2z)|\ll_J(1+\|\mu\|+|u|)^{C_J},\qquad\operatorname{Re}(u)\ge1/8. ∣ P J ( u , μ ) ζ ( 1/2 + u + 2 z ) ∣ ≪ J ( 1 + ∥ μ ∥ + ∣ u ∣ ) C J , Re ( u ) ≥ 1/8.
The G L 2 \mathrm{GL}_2 GL 2 L L L -function is entire. Its bound [27 Theorem 5.41] , together with (18) and the Gaussian e u 2 e^{u^2} e u 2 , makes the horizontal integrals tend to 0 0 0 , so the shift crosses no residue.
For u = 1 / 8 + i v u=1/8+iv u = 1/8 + i v , G L 2 \mathrm{GL}_2 GL 2 convexity [27 Theorem 5.41] , (20) and (18) give the following bounds, respectively, valid also for exceptional parameters,
L ( 5 / 8 + i v − z , ϕ ) ≪ ε p 3 / 16 + ε ( 1 + ∣ v ∣ + ∣ z ∣ + ∣ τ ϕ ∣ ) C , L(5/8+iv-z,\phi)\ll_\varepsilon p^{3/16+\varepsilon}(1+|v|+|z|+|\tau_\phi|)^C, L ( 5/8 + i v − z , ϕ ) ≪ ε p 3/16 + ε ( 1 + ∣ v ∣ + ∣ z ∣ + ∣ τ ϕ ∣ ) C ,
∣ ζ ( 5 / 8 + i v + 2 z ) ∣ ≪ ε ( 1 + ∣ v ∣ + ∣ z ∣ ) 3 / 16 + ε , |\zeta(5/8+iv+2z)|\ll_\varepsilon(1+|v|+|z|)^{3/16+\varepsilon}, ∣ ζ ( 5/8 + i v + 2 z ) ∣ ≪ ε ( 1 + ∣ v ∣ + ∣ z ∣ ) 3/16 + ε ,
and
∣ G J ( u , μ ) γ μ ( 1 / 2 + u ) γ ± μ ( 1 / 2 ) ∣ ≪ J ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ + ∣ v ∣ ) C J e − v 2 + 3 π ∣ v ∣ / 4 . \left|G_J(u,\mu)\frac{\gamma_\mu(1/2+u)}{\gamma_{\pm\mu}(1/2)}\right|\ll_J(1+|z|+|\tau_\phi|+|v|)^{C_J}e^{-v^2+3\pi|v|/4}. G J ( u , μ ) γ ± μ ( 1/2 ) γ μ ( 1/2 + u ) ≪ J ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ + ∣ v ∣ ) C J e − v 2 + 3 π ∣ v ∣/4 .
Integration over v v v and p 1 / 16 p 3 / 16 + ε = p 1 / 4 + ε p^{1/16}p^{3/16+\varepsilon}=p^{1/4+\varepsilon} p 1/16 p 3/16 + ε = p 1/4 + ε prove the estimate. Unitarity gives γ − μ ( 1 / 2 ) = γ μ ( 1 / 2 ) ‾ \gamma_{-\mu}(1/2)=\overline{\gamma_\mu(1/2)} γ − μ ( 1/2 ) = γ μ ( 1/2 ) , so the change of the central denominator has modulus 1 1 1 . □
We next bound the harmonic averages of the G L 2 \mathrm{GL}_2 GL 2 Hecke coefficients.
Lemma 4.7. For R ≥ 1 R\ge1 R ≥ 1 and ( k , p ) = 1 (k,p)=1 ( k , p ) = 1 , we have
∑ ϕ ∈ B 2 ( p ) ∣ τ ϕ ∣ ≤ R ∣ λ ϕ ( k ) ∣ 2 L ( 1 , ϕ , Ad ) ≪ p R 6 ( 1 + k 2 p 2 ) . (72) \sum_{\substack{\phi\in\mathcal{B}_2(p)\\|\tau_\phi|\le R}}\frac{|\lambda_\phi(k)|^2}{L(1,\phi,\operatorname{Ad})}\ll pR^6\left(1+\frac{k^2}{p^2}\right). \tag*{(72)} ϕ ∈ B 2 ( p ) ∣ τ ϕ ∣ ≤ R ∑ L ( 1 , ϕ , Ad ) ∣ λ ϕ ( k ) ∣ 2 ≪ p R 6 ( 1 + p 2 k 2 ) . (72) Let h h h be fixed and rapidly decreasing on the unitary spectrum. For every fixed C ≥ 0 C\ge0 C ≥ 0 and ε > 0 \varepsilon>0 ε > 0 , uniformly for ( a b m n , p ) = 1 (abmn,p)=1 ( abmn , p ) = 1 and with μ = ( z + τ ϕ , z − τ ϕ , − 2 z ) \mu=(z+\tau_\phi,z-\tau_\phi,-2z) μ = ( z + τ ϕ , z − τ ϕ , − 2 z ) , we have
∑ ϕ ∈ B 2 ( p ) 1 L ( 1 , ϕ , Ad ) ∫ Re ( z ) = 0 ( ∣ h ( μ ) ∣ + ∣ h ( − μ ) ∣ ) ( 1 + ∥ μ ∥ ) C ∣ L ( 1 + 3 z , ϕ ) ∣ 2 ∣ A ϕ , z ( a , b ) A ϕ , z ( m , n ) ∣ ∣ d z ∣ (73) \sum_{\phi\in\mathcal{B}_2(p)}\frac{1}{L(1,\phi,\operatorname{Ad})}\int_{\operatorname{Re}(z)=0}\frac{(|h(\mu)|+|h(-\mu)|)(1+\|\mu\|)^C}{|L(1+3z,\phi)|^2}|A_{\phi,z}(a,b)A_{\phi,z}(m,n)|\,|dz| \tag*{(73)} ϕ ∈ B 2 ( p ) ∑ L ( 1 , ϕ , Ad ) 1 ∫ Re ( z ) = 0 ∣ L ( 1 + 3 z , ϕ ) ∣ 2 ( ∣ h ( μ ) ∣ + ∣ h ( − μ ) ∣ ) ( 1 + ∥ μ ∥ ) C ∣ A ϕ , z ( a , b ) A ϕ , z ( m , n ) ∣ ∣ d z ∣ (73) ≪ h , C , ε p 1 + ε ( a b m n ) ε ( 1 + a b p ) ( 1 + m n p ) . \ll_{h,C,\varepsilon}p^{1+\varepsilon}(abmn)^\varepsilon\left(1+\frac{ab}{p}\right)\left(1+\frac{mn}{p}\right). ≪ h , C , ε p 1 + ε ( abmn ) ε ( 1 + p ab ) ( 1 + p mn ) .
Proof. For this proof only, rescale ϕ \phi ϕ to have norm 1 1 1 for the hyperbolic measure d x d y / y 2 dx\,dy/y^2 d x d y / y 2 on Γ 0 ( p ) \ H \Gamma_0(p)\backslash\mathbb{H} Γ 0 ( p ) \ H . This leaves its Hecke eigenvalues unchanged. Write ρ ϕ ( n ) \rho_\phi(n) ρ ϕ ( n ) for its coefficients with Whittaker factor y K i t ϕ ( 2 π ∣ n ∣ y ) \sqrt{y}K_{it_\phi}(2\pi|n|y) y K i t ϕ ( 2 π ∣ n ∣ y ) , where τ ϕ = i t ϕ \tau_\phi=it_\phi τ ϕ = i t ϕ . Then ρ ϕ ( k ) = ρ ϕ ( 1 ) λ ϕ ( k ) \rho_\phi(k)=\rho_\phi(1)\lambda_\phi(k) ρ ϕ ( k ) = ρ ϕ ( 1 ) λ ϕ ( k ) for ( k , p ) = 1 (k,p)=1 ( k , p ) = 1 . The Hecke relations, including λ ϕ ( p ) = ± p − 1 / 2 \lambda_\phi(p)=\pm p^{-1/2} λ ϕ ( p ) = ± p − 1/2 , give
∑ n ≥ 1 ∣ λ ϕ ( n ) ∣ 2 n w = ζ ( w ) L ( w , ϕ , Ad ) ζ ( 2 w ) ( 1 + p − w ) , Re ( w ) > 1. \sum_{n\ge1}\frac{|\lambda_\phi(n)|^2}{n^w}=\frac{\zeta(w)L(w,\phi,\operatorname{Ad})}{\zeta(2w)(1+p^{-w})},\qquad\operatorname{Re}(w)>1. n ≥ 1 ∑ n w ∣ λ ϕ ( n ) ∣ 2 = ζ ( 2 w ) ( 1 + p − w ) ζ ( w ) L ( w , ϕ , Ad ) , Re ( w ) > 1.
The Eisenstein series at infinity has residue 3 / ( π ( p + 1 ) ) 3/(\pi(p+1)) 3/ ( π ( p + 1 )) in [27 § 15.4] , and the Bessel Mellin formula [49 (10.32.19)] yields
∫ 0 ∞ K i t ϕ ( y ) 2 d y = π 2 4 cosh ( π t ϕ ) . \int_0^\infty K_{it_\phi}(y)^2\,dy=\frac{\pi^2}{4\cosh(\pi t_\phi)}. ∫ 0 ∞ K i t ϕ ( y ) 2 d y = 4 cosh ( π t ϕ ) π 2 .
Since ∣ ρ ϕ ( − n ) ∣ = ∣ ρ ϕ ( n ) ∣ |\rho_\phi(-n)|=|\rho_\phi(n)| ∣ ρ ϕ ( − n ) ∣ = ∣ ρ ϕ ( n ) ∣ , Rankin–Selberg unfolding and taking the residue at w = 1 w=1 w = 1 give
3 π ( p + 1 ) = 2 ∣ ρ ϕ ( 1 ) ∣ 2 ( 2 π ) − 1 π 2 4 cosh ( π t ϕ ) p L ( 1 , ϕ , Ad ) ( p + 1 ) ζ ( 2 ) . \frac{3}{\pi(p+1)}=2|\rho_\phi(1)|^2(2\pi)^{-1}\frac{\pi^2}{4\cosh(\pi t_\phi)}\frac{pL(1,\phi,\operatorname{Ad})}{(p+1)\zeta(2)}. π ( p + 1 ) 3 = 2∣ ρ ϕ ( 1 ) ∣ 2 ( 2 π ) − 1 4 cosh ( π t ϕ ) π 2 ( p + 1 ) ζ ( 2 ) p L ( 1 , ϕ , Ad ) .
Here the factor 2 comes from the positive and negative Fourier frequencies. Substituting ζ ( 2 ) = π 2 / 6 \zeta(2) = \pi^2/6 ζ ( 2 ) = π 2 /6 , we obtain
∣ ρ ϕ ( 1 ) ∣ 2 cosh ( π t ϕ ) = 2 p L ( 1 , ϕ , Ad ) . (74) \frac{|\rho_\phi(1)|^2}{\cosh(\pi t_\phi)} = \frac{2}{pL(1,\phi,\operatorname{Ad})}. \tag*{(74)} cosh ( π t ϕ ) ∣ ρ ϕ ( 1 ) ∣ 2 = p L ( 1 , ϕ , Ad ) 2 . (74)
and hence the conversion to the Kuznetsov weight is
∣ λ ϕ ( k ) ∣ 2 L ( 1 , ϕ , Ad ) = p 2 ∣ ρ ϕ ( k ) ∣ 2 cosh ( π t ϕ ) ( ( k , p ) = 1 ) . \frac{|\lambda_\phi(k)|^2}{L(1,\phi,\operatorname{Ad})} = \frac{p}{2}\frac{|\rho_\phi(k)|^2}{\cosh(\pi t_\phi)} \qquad((k,p)=1). L ( 1 , ϕ , Ad ) ∣ λ ϕ ( k ) ∣ 2 = 2 p cosh ( π t ϕ ) ∣ ρ ϕ ( k ) ∣ 2 (( k , p ) = 1 ) .
The Bessel integral converges for ∣ Im ( t ϕ ) ∣ < 1 / 2 |\operatorname{Im}(t_\phi)|<1/2 ∣ Im ( t ϕ ) ∣ < 1/2 , so these identities also hold for exceptional t ϕ ∈ i [ − 7 / 64 , 7 / 64 ] t_\phi\in i[-7/64,7/64] t ϕ ∈ i [ − 7/64 , 7/64 ] .
For R ≥ 1 R\ge1 R ≥ 1 , we take the even entire test function
b R ( t ) : = e − t 2 / R 2 ( t 2 + 1 / 4 ) 2 . b_R(t):=e^{-t^2/R^2}(t^2+1/4)^2. b R ( t ) := e − t 2 / R 2 ( t 2 + 1/4 ) 2 .
This is non-negative on the real and exceptional spectra and bounded below by a positive constant for ∣ τ ϕ ∣ ≤ R |\tau_\phi|\le R ∣ τ ϕ ∣ ≤ R . Its transforms [27 Eqs. (16.35) and (16.36)] are
g 0 , R = π − 2 ∫ R t b R ( t ) tanh ( π t ) d t ≪ R 6 , g_{0,R}=\pi^{-2}\int_{\mathbb{R}}t b_R(t)\tanh(\pi t)\,dt\ll R^6, g 0 , R = π − 2 ∫ R t b R ( t ) tanh ( π t ) d t ≪ R 6 ,
and
g R ( 2 ) ( x ) = 2 i π ∫ R J 2 i t ( x ) t b R ( t ) cosh ( π t ) d t . g_R^{(2)}(x)=\frac{2i}{\pi}\int_{\mathbb{R}}J_{2it}(x)\frac{t b_R(t)}{\cosh(\pi t)}\,dt. g R ( 2 ) ( x ) = π 2 i ∫ R J 2 i t ( x ) cosh ( π t ) t b R ( t ) d t .
Moving the second contour to t = v − i t=v-i t = v − i , the only crossed pole is − i / 2 -i/2 − i /2 , which can be cancelled by b R b_R b R . Poisson’s integral representation [49 Eq. (10.9.4)] gives, for x > 0 x>0 x > 0 ,
J 2 + 2 i v ( x ) = ( x / 2 ) 2 + 2 i v π Γ ( 5 / 2 + 2 i v ) ∫ − 1 1 e i x t ( 1 − t 2 ) 3 / 2 + 2 i v d t , J_{2+2iv}(x)=\frac{(x/2)^{2+2iv}}{\sqrt{\pi}\,\Gamma(5/2+2iv)}\int_{-1}^{1}e^{ixt}(1-t^2)^{3/2+2iv}\,dt, J 2 + 2 i v ( x ) = π Γ ( 5/2 + 2 i v ) ( x /2 ) 2 + 2 i v ∫ − 1 1 e i x t ( 1 − t 2 ) 3/2 + 2 i v d t ,
and
∣ J 2 + 2 i v ( x ) ∣ ≪ x 2 e π ∣ v ∣ ( 1 + ∣ v ∣ ) − 2 . |J_{2+2iv}(x)|\ll x^2e^{\pi|v|}(1+|v|)^{-2}. ∣ J 2 + 2 i v ( x ) ∣ ≪ x 2 e π ∣ v ∣ ( 1 + ∣ v ∣ ) − 2 .
Stirling’s formula and the bound
∣ e − t 2 / R 2 ∣ ≤ e 1 / R 2 e − ( Re ( t ) ) 2 / R 2 ( − 1 ≤ Im ( t ) ≤ 0 ) \left|e^{-t^2/R^2}\right|\le e^{1/R^2}e^{-(\operatorname{Re}(t))^2/R^2}\qquad(-1\le\operatorname{Im}(t)\le0) e − t 2 / R 2 ≤ e 1/ R 2 e − ( Re ( t ) ) 2 / R 2 ( − 1 ≤ Im ( t ) ≤ 0 )
justify the contour shift. Since cosh ( π ( v − i ) ) = − cosh ( π v ) \cosh(\pi(v-i))=-\cosh(\pi v) cosh ( π ( v − i )) = − cosh ( π v ) ,
∣ g R ( 2 ) ( x ) ∣ ≪ x 2 ∫ R ( 1 + ∣ v ∣ ) 3 e − v 2 / R 2 d v ≪ x 2 R 6 ( x > 0 ) . |g_R^{(2)}(x)|\ll x^2\int_{\mathbb{R}}(1+|v|)^3e^{-v^2/R^2}\,dv\ll x^2R^6\qquad(x>0). ∣ g R ( 2 ) ( x ) ∣ ≪ x 2 ∫ R ( 1 + ∣ v ∣ ) 3 e − v 2 / R 2 d v ≪ x 2 R 6 ( x > 0 ) .
Applying the G L 2 \mathrm{GL}_2 GL 2 Kuznetsov trace formula [27 Theorem 16.3, Eq. (16.34)] with both indices equal to k k k , the old and continuous spectra are non-negative. Using the trivial bound ∣ S ( k , k ; c ) ∣ ≤ c |S(k,k;c)|\le c ∣ S ( k , k ; c ) ∣ ≤ c , the geometric contribution is bounded by
∑ p ∣ c ∣ S ( k , k ; c ) ∣ c ∣ g R ( 2 ) ( 4 π k / c ) ∣ ≪ R 6 k 2 ∑ p ∣ c c − 2 ≪ R 6 k 2 p 2 . \sum_{p\mid c}\frac{|S(k,k;c)|}{c}|g_R^{(2)}(4\pi k/c)|\ll R^6k^2\sum_{p\mid c}c^{-2}\ll R^6\frac{k^2}{p^2}. p ∣ c ∑ c ∣ S ( k , k ; c ) ∣ ∣ g R ( 2 ) ( 4 π k / c ) ∣ ≪ R 6 k 2 p ∣ c ∑ c − 2 ≪ R 6 p 2 k 2 .
The identity (4.14) and the lower bound for b R b_R b R prove (4.12). In particular,
∑ ϕ ∈ B 2 ( p ) ∣ t ϕ ∣ ≤ R 1 L ( 1 , ϕ , Ad ) ≪ p R 6 . \sum_{\substack{\phi\in B_2(p)\\ |t_\phi|\le R}}\frac{1}{L(1,\phi,\operatorname{Ad})}\ll pR^6. ϕ ∈ B 2 ( p ) ∣ t ϕ ∣ ≤ R ∑ L ( 1 , ϕ , Ad ) 1 ≪ p R 6 .
The unramified Hecke eigenvalues of ϕ \phi ϕ are real, since its central character is trivial. The Hecke relations give
A ϕ , z ( a , b ) = ∑ d ∣ ( a , b ) μ ( d ) A ϕ , z ( 1 , a / d ) A ϕ , z ( 1 , b / d ) , A_{\phi,z}(a,b)=\sum_{d\mid(a,b)}\mu(d)A_{\phi,z}(1,a/d)A_{\phi,z}(1,b/d), A ϕ , z ( a , b ) = d ∣ ( a , b ) ∑ μ ( d ) A ϕ , z ( 1 , a / d ) A ϕ , z ( 1 , b / d ) ,
A ϕ , z ( 1 , n ) = ∑ u v = n λ ϕ ( u ) u z v − 2 z , A_{\phi,z}(1,n)=\sum_{uv=n}\lambda_{\phi}(u)u^{z}v^{-2z}, A ϕ , z ( 1 , n ) = uv = n ∑ λ ϕ ( u ) u z v − 2 z ,
λ ϕ ( u ) λ ϕ ( v ) = ∑ d ∣ ( u , v ) λ ϕ ( u v / d 2 ) . \lambda_{\phi}(u)\lambda_{\phi}(v)=\sum_{d\mid(u,v)}\lambda_{\phi}(uv/d^{2}). λ ϕ ( u ) λ ϕ ( v ) = d ∣ ( u , v ) ∑ λ ϕ ( uv / d 2 ) .
For z ∈ i R z\in i\mathbb{R} z ∈ i R , these express A ϕ , z ( a , b ) A_{\phi,z}(a,b) A ϕ , z ( a , b ) as a sum of at most τ ( a b ) 4 \tau(ab)^{4} τ ( ab ) 4 terms of the form c λ ϕ ( k ) c\lambda_{\phi}(k) c λ ϕ ( k ) , with ∣ c ∣ ≤ 1 |c|\leq1 ∣ c ∣ ≤ 1 and k ≤ a b k\leq ab k ≤ ab . The coefficients may depend on z z z , but the bound is uniform in z z z . The Cauchy–Schwarz inequality, the divisor bound τ ( n ) ≪ ε n ε \tau(n)\ll_{\varepsilon}n^{\varepsilon} τ ( n ) ≪ ε n ε (see [27 Eq. (1.81)] ) and (4.12) therefore give
∑ ϕ ∈ B 2 ( p ) ∣ τ ϕ ∣ ≤ R ∣ A ϕ , z ( a , b ) ∣ 2 L ( 1 , ϕ , Ad ) ≪ ε p R 6 ( a b ) ε ( 1 + ( a b ) 2 p 2 ) . \sum_{\substack{\phi\in B_{2}(p)\\ |\tau_{\phi}|\leq R}}\frac{|A_{\phi,z}(a,b)|^{2}}{L(1,\phi,\operatorname{Ad})}\ll_{\varepsilon}pR^{6}(ab)^{\varepsilon}\left(1+\frac{(ab)^{2}}{p^{2}}\right). ϕ ∈ B 2 ( p ) ∣ τ ϕ ∣ ≤ R ∑ L ( 1 , ϕ , Ad ) ∣ A ϕ , z ( a , b ) ∣ 2 ≪ ε p R 6 ( ab ) ε ( 1 + p 2 ( ab ) 2 ) .
Applying the Cauchy–Schwarz inequality again yields the corresponding product bound with factors ( a b m n ) ε ( 1 + a b / p ) ( 1 + m n / p ) (abmn)^{\varepsilon}(1+ab/p)(1+mn/p) ( abmn ) ε ( 1 + ab / p ) ( 1 + mn / p ) . Applying [?, Theorem C(2)] to the unitary twist ϕ ⊗ ∣ det ∣ 3 z \phi\otimes|\det|^{3z} ϕ ⊗ ∣ det ∣ 3 z , we obtain
∣ L ( 1 + 3 z , ϕ ) ∣ − 2 ≪ ε p ε ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ ) C ε . (75) |L(1+3z,\phi)|^{-2}\ll_{\varepsilon}p^{\varepsilon}(1+|z|+|\tau_{\phi}|)^{C_{\varepsilon}}. \tag*{(75)} ∣ L ( 1 + 3 z , ϕ ) ∣ − 2 ≪ ε p ε ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ ) C ε . (75)
Finally, ∥ μ ∥ 2 = 6 ∣ z ∣ 2 + 2 ∣ τ ϕ ∣ 2 \|\mu\|^{2}=6|z|^{2}+2|\tau_{\phi}|^{2} ∥ μ ∥ 2 = 6∣ z ∣ 2 + 2∣ τ ϕ ∣ 2 . Rapid decrease of h h h , integration in z z z , and dyadic decomposition in ∣ τ ϕ ∣ |\tau_{\phi}| ∣ τ ϕ ∣ prove (4.13). Omitting the factor ∣ L ( 1 + 3 z , ϕ ) ∣ − 2 |L(1+3z,\phi)|^{-2} ∣ L ( 1 + 3 z , ϕ ) ∣ − 2 , the case k = 1 k=1 k = 1 of (4.12) similarly gives, for every fixed C ≥ 0 C\geq0 C ≥ 0 ,
∑ ϕ ∈ B 2 ( p ) 1 L ( 1 , ϕ , Ad ) ∫ Re ( z ) = 0 ( ∣ h ( μ ) ∣ + ∣ h ( − μ ) ∣ ) ( 1 + ∥ μ ∥ ) C d z ≪ h , C p . (76) \sum_{\phi\in B_{2}(p)}\frac{1}{L(1,\phi,\operatorname{Ad})}\int_{\operatorname{Re}(z)=0}(|h(\mu)|+|h(-\mu)|)(1+\|\mu\|)^{C}\,dz\ll_{h,C}p. \tag*{(76)} ϕ ∈ B 2 ( p ) ∑ L ( 1 , ϕ , Ad ) 1 ∫ Re ( z ) = 0 ( ∣ h ( μ ) ∣ + ∣ h ( − μ ) ∣ ) ( 1 + ∥ μ ∥ ) C d z ≪ h , C p . (76)
This completes the proof. □
Proposition 4.8. Uniformly in the range of Theorem 4.1 ,
R max , n e w ≪ Q , η , h 0 , J , ε p 5 / 4 + ε . \mathcal{R}_{\max,\mathrm{new}}\ll_{Q,\eta,h_{0},J,\varepsilon}p^{5/4+\varepsilon}. R m a x , new ≪ Q , η , h 0 , J , ε p 5/4 + ε .
Proof. By § 3.3.3, the two non-zero standard parameters at p p p are p z λ ϕ ( p ) p^{z}\lambda_{\phi}(p) p z λ ϕ ( p ) and p − 2 z p^{-2z} p − 2 z , with λ ϕ ( p ) = ± p − 1 / 2 \lambda_{\phi}(p)=\pm p^{-1/2} λ ϕ ( p ) = ± p − 1/2 . Their product is p − z λ ϕ ( p ) p^{-z}\lambda_{\phi}(p) p − z λ ϕ ( p ) . The ramified Schur identity (2.1) gives, for every n ≥ 1 n\geq1 n ≥ 1 ,
A ϕ , z ( p , n ) = p − z λ ϕ ( p ) A ϕ , z ( 1 , n ) , A_{\phi,z}(p,n)=p^{-z}\lambda_{\phi}(p)A_{\phi,z}(1,n), A ϕ , z ( p , n ) = p − z λ ϕ ( p ) A ϕ , z ( 1 , n ) ,
and p ∣ p − z λ ϕ ( p ) ∣ = 1 \sqrt{p}|p^{-z}\lambda_{\phi}(p)|=1 p ∣ p − z λ ϕ ( p ) ∣ = 1 .
Thus Lemma 4.6 bounds both full sums, including the dual factor p \sqrt{p} p , after extracting h ( μ ) h(\mu) h ( μ ) or h ( − μ ) h(-\mu) h ( − μ ) . The truncation argument following (4.4), using Lemma 2.5 and (4.5), replaces the truncated sums by full sums with error O Q , η , h 0 , J , ε , B ( p − B ) O_{Q,\eta,h_{0},J,\varepsilon,B}(p^{-B}) O Q , η , h 0 , J , ε , B ( p − B ) , uniformly in r , s r,s r , s .
Apply (4.13) with one coefficient at ( 1 , 1 ) (1,1) ( 1 , 1 ) and the other at ( r , s ) (r,s) ( r , s ) or ( s , r ) (s,r) ( s , r ) . Together with Lemma 4.6, this gives
∣ R max , n e w ∣ ≪ h 0 , J , ε χ p p 5 / 4 + ε ( r s ) ε ( 1 + r s p ) . |\mathcal{R}_{\max,\mathrm{new}}|\ll_{h_{0},J,\varepsilon}\chi_{p}p^{5/4+\varepsilon}(rs)^{\varepsilon}\left(1+\frac{rs}{p}\right). ∣ R m a x , new ∣ ≪ h 0 , J , ε χ p p 5/4 + ε ( r s ) ε ( 1 + p r s ) .
Since r s ≤ Q 2 p η < p rs\leq Q^{2}p^{\eta}<p r s ≤ Q 2 p η < p for sufficiently large p p p and χ p ≪ 1 \chi_{p}\ll1 χ p ≪ 1 , we obtain the stated bound. □
Proof of Theorem 4.1
Combining (69) , Lemmas 4.2 and 4.3, and Propositions 4.5 and 4.8 in (68) , and using r s ≤ Q 2 p η rs \le Q^{2}p^{\eta} r s ≤ Q 2 p η , we obtain
χ p M h ( r , s , p ) = δ r = 1 P s H h + O Q , η , h 0 , J , ε ( p 5 / 4 + ε ) . \chi_{p}M_{h}(r,s,p)=\frac{\delta_{r=1}P}{\sqrt{s}}\mathcal{H}_{h}+O_{Q,\eta,h_{0},J,\varepsilon}\left(p^{5/4+\varepsilon}\right). χ p M h ( r , s , p ) = s δ r = 1 P H h + O Q , η , h 0 , J , ε ( p 5/4 + ε ) .
Indeed, the level one terms are O Q , η , h 0 , J , ε ( p 3 / 7 + 5 η / 14 + ε ) O_{Q,\eta,h_{0},J,\varepsilon}\left(p^{3/7+5\eta/14+\varepsilon}\right) O Q , η , h 0 , J , ε ( p 3/7 + 5 η /14 + ε ) and the diagonal error is O Q , h , J ( p ) O_{Q,h,J}(p) O Q , h , J ( p ) . Dividing by χ p = 1 + O ( p − 1 ) \chi_{p}=1+O(p^{-1}) χ p = 1 + O ( p − 1 ) changes the main term by O Q , h ( p ) O_{Q,h}(p) O Q , h ( p ) and proves (63) .
The harmonic twisted second moment
From this subsection onward, we use the fixed test (26) . For positive integers a , b a,b a , b , define the harmonic twisted second moment
N 2 ( a , b , p ) : = c c u s p ∑ π ∈ B ( p ) h ( μ π ) A π ( a , b ) ‾ ∣ L ( 1 / 2 , π ) ∣ 2 L ( 1 , π , Ad ) . \mathcal{N}_{2}(a,b,p):=c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(\mu_{\pi})\overline{A_{\pi}(a,b)}\lvert L(1/2,\pi)\rvert^{2}}{L(1,\pi,\operatorname{Ad})}. N 2 ( a , b , p ) := c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) A π ( a , b ) ∣ L ( 1/2 , π ) ∣ 2 .
We write N 2 ( p ) : = N 2 ( 1 , 1 , p ) \mathcal{N}_{2}(p):=\mathcal{N}_{2}(1,1,p) N 2 ( p ) := N 2 ( 1 , 1 , p ) for the untwisted moment. With H h \mathcal{H}_{h} H h as in (62) , we also set
κ ( μ ) : = γ μ ′ γ μ ( 1 / 2 ) + γ − μ ′ γ − μ ( 1 / 2 ) , \kappa(\mu):=\frac{\gamma'_{\mu}}{\gamma_{\mu}}(1/2)+\frac{\gamma'_{-\mu}}{\gamma_{-\mu}}(1/2), κ ( μ ) := γ μ γ μ ′ ( 1/2 ) + γ − μ γ − μ ′ ( 1/2 ) ,
and
K h : = ∫ Re ( μ ) = 0 h ( μ ) κ ( μ ) spec ( μ ) d μ . \mathcal{K}_{h}:=\int_{\operatorname{Re}(\mu)=0}h(\mu)\kappa(\mu)\operatorname{spec}(\mu)\,\mathrm{d}\mu. K h := ∫ Re ( μ ) = 0 h ( μ ) κ ( μ ) spec ( μ ) d μ .
The positivity and rapid decrease of h h h give H h > 0 \mathcal{H}_{h}>0 H h > 0 and absolute convergence of K h \mathcal{K}_{h} K h .
Theorem 4.9. Fix a positive integer Q Q Q and 0 < η < 1 / 2 0<\eta<1/2 0 < η < 1/2 , independently of p p p . Choose a 0 a_{0} a 0 sufficiently large in terms of η \eta η and J ≥ ⌈ a 0 ⌉ + 4 J\ge\lceil a_{0}\rceil+4 J ≥ ⌈ a 0 ⌉ + 4 , independently of p p p . Uniformly for positive integers a , b a,b a , b with ( a b , p ) = 1 (ab,p)=1 ( ab , p ) = 1 and a b ≤ Q 2 p 2 η ab\le Q^{2}p^{2\eta} ab ≤ Q 2 p 2 η ,
N 2 ( a , b , p ) = P a b ( ( 1 − p − 1 ) { H h ( log p a b + 2 γ ) + K h } + 2 log p p H h ) + O h , Q , η , ε ( p 7 / 4 ( a b ) − 1 / 4 + p 3 / 2 + ε ) . (77) \mathcal{N}_{2}(a,b,p)=\frac{P}{\sqrt{ab}}\left(\left(1-p^{-1}\right)\left\{\mathcal{H}_{h}\left(\log\frac{p}{ab}+2\gamma\right)+\mathcal{K}_{h}\right\}+\frac{2\log p}{p}\mathcal{H}_{h}\right)+O_{h,Q,\eta,\varepsilon}\left(p^{7/4}(ab)^{-1/4}+p^{3/2+\varepsilon}\right). \tag*{(77)} N 2 ( a , b , p ) = ab P ( ( 1 − p − 1 ) { H h ( log ab p + 2 γ ) + K h } + p 2 log p H h ) + O h , Q , η , ε ( p 7/4 ( ab ) − 1/4 + p 3/2 + ε ) . (77) Here γ \gamma γ is Euler’s constant, and the small exponents in the error can be chosen in terms of η \eta η . No condition ( a , b ) = 1 (a,b)=1 ( a , b ) = 1 is imposed. In particular, uniformly in this range,
N 2 ( a , b , p ) = P H h a b log p a b ( 1 + O h , Q , η ( 1 log p ) ) . \mathcal{N}_{2}(a,b,p)=\frac{P\mathcal{H}_{h}}{\sqrt{ab}}\log\frac{p}{ab}\left(1+O_{h,Q,\eta}\left(\frac{1}{\log p}\right)\right). N 2 ( a , b , p ) = ab P H h log ab p ( 1 + O h , Q , η ( log p 1 ) ) . For a = b = 1 a=b=1 a = b = 1 , taking Q = 1 Q=1 Q = 1 gives
N 2 ( p ) = P H h log p + P ( 2 γ H h + K h ) + O h , η , ε ( p 7 / 4 + ε ) . (78) \mathcal{N}_{2}(p)=P\mathcal{H}_{h}\log p+P(2\gamma\mathcal{H}_{h}+\mathcal{K}_{h})+O_{h,\eta,\varepsilon}\left(p^{7/4+\varepsilon}\right). \tag*{(78)} N 2 ( p ) = P H h log p + P ( 2 γ H h + K h ) + O h , η , ε ( p 7/4 + ε ) . (78)
Substituting (29) into the definition of N 2 ( a , b , p ) \mathcal{N}_{2}(a,b,p) N 2 ( a , b , p ) gives
N 2 ( a , b , p ) = 2 c c u s p ∑ π ∈ B ( p ) h ( μ π ) A π ( a , b ) ‾ L ( 1 , π , Ad ) ∑ m , n ≥ 1 A π ( 1 , m ) ‾ A π ( 1 , n ) m n U μ π ( m n / p ) . \mathcal{N}_{2}(a,b,p)=2c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(\mu_{\pi})\overline{A_{\pi}(a,b)}}{L(1,\pi,\operatorname{Ad})}\sum_{m,n\ge1}\frac{\overline{A_{\pi}(1,m)}A_{\pi}(1,n)}{\sqrt{mn}}U_{\mu_{\pi}}(mn/p). N 2 ( a , b , p ) = 2 c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) A π ( a , b ) m , n ≥ 1 ∑ mn A π ( 1 , m ) A π ( 1 , n ) U μ π ( mn / p ) .
By (9) , ∣ A π ( 1 , p k ) ∣ ≤ ( 1 − p − 1 / 2 ) − 1 \lvert A_{\pi}(1,p^{k})\rvert\le(1-p^{-1/2})^{-1} ∣ A π ( 1 , p k )∣ ≤ ( 1 − p − 1/2 ) − 1 for $k\ge0.
Writing m = p k m ′ m = p^{k}m' m = p k m ′ and n = p ℓ n ′ n = p^{\ell}n' n = p ℓ n ′ with ( m ′ n ′ , p ) = 1 (m'n',p) = 1 ( m ′ n ′ , p ) = 1 , the terms with k + ℓ ≥ 1 k+\ell\ge1 k + ℓ ≥ 1 satisfy, by (30) ,
∑ m , n ≥ 1 p ∣ m n ∣ A π ( 1 , m ) A π ( 1 , n ) ∣ m n ∣ μ π π ( m n / p ) ∣ \sum_{\substack{m,n\geq1\\p\mid mn}} \frac{\lvert A_{\pi}(1,m)A_{\pi}(1,n)\rvert}{\sqrt{mn}}\lvert\mu_{\pi\pi}(mn/p)\rvert m , n ≥ 1 p ∣ mn ∑ mn ∣ A π ( 1 , m ) A π ( 1 , n )∣ ∣ μ π π ( mn / p )∣
≪ B , J , ε ( 1 + ∥ μ π ∥ ) C ∑ k + ℓ ≥ 1 p − ( k + ℓ ) / 2 ∑ m ′ , n ′ ≥ 1 ( m ′ n ′ ) − 1 / 2 + θ 3 + ε ( 1 + p k + ℓ − 1 m ′ n ′ ) − B \ll_{B,J,\varepsilon} (1+\lVert\mu_{\pi}\rVert)^{C}\sum_{k+\ell\geq1}p^{-(k+\ell)/2}\sum_{m',n'\geq1}(m'n')^{-1/2+\theta_{3}+\varepsilon}(1+p^{k+\ell-1}m'n')^{-B} ≪ B , J , ε ( 1 + ∥ μ π ∥ ) C k + ℓ ≥ 1 ∑ p − ( k + ℓ ) /2 m ′ , n ′ ≥ 1 ∑ ( m ′ n ′ ) − 1/2 + θ 3 + ε ( 1 + p k + ℓ − 1 m ′ n ′ ) − B
≪ J , ε p − 1 / 2 ( 1 + ∥ μ π ∥ ) C . \ll_{J,\varepsilon}p^{-1/2}(1+\lVert\mu_{\pi}\rVert)^{C}. ≪ J , ε p − 1/2 ( 1 + ∥ μ π ∥ ) C .
Here B > 2 + 2 θ 3 + 2 ε B>2+2\theta_{3}+2\varepsilon B > 2 + 2 θ 3 + 2 ε is fixed. For sufficiently large p p p , we have a b ≤ Q 2 p 2 η ≤ p 1 / 2 + η ab\leq Q^{2}p^{2\eta}\leq p^{1/2+\eta} ab ≤ Q 2 p 2 η ≤ p 1/2 + η . Apply (56) with δ = 1 / 2 − η \delta=1/2-\eta δ = 1/2 − η . The total ramified contribution is therefore O h , Q , η , ε ( p 3 / 2 + ε ) O_{h,Q,\eta,\varepsilon}(p^{3/2+\varepsilon}) O h , Q , η , ε ( p 3/2 + ε ) , uniformly in a , b a,b a , b . For ( m n , p ) = 1 (mn,p)=1 ( mn , p ) = 1 , the Hecke relation gives
A π ( 1 , m ) ‾ A π ( 1 , n ) = ∑ r ∣ ( m , n ) A π ( m / r , n / r ) . \overline{A_{\pi}(1,m)}A_{\pi}(1,n)=\sum_{r\mid(m,n)}A_{\pi}(m/r,n/r). A π ( 1 , m ) A π ( 1 , n ) = r ∣ ( m , n ) ∑ A π ( m / r , n / r ) .
Replacing the indices by r m , r n rm,rn r m , r n , we obtain
N 2 ( a , b , p ) = 2 c c u s p ∑ r , m , n ≥ 1 ( r m n , p ) = 1 1 r m n ∑ π ∈ B ( p ) A π ( a , b ) A π ( m , n ) H r 2 m n ( μ π ) L ( 1 , π , Ad ) + O h , Q , η , ε ( p 3 / 2 + ε ) . \mathcal{N}_{2}(a,b,p)=2c_{\mathrm{cusp}}\sum_{\substack{r,m,n\geq1\\(rmn,p)=1}}\frac{1}{r\sqrt{mn}}\sum_{\pi\in\mathcal{B}(p)}\frac{A_{\pi}(a,b)A_{\pi}(m,n)H_{r^{2}mn}(\mu_{\pi})}{L(1,\pi,\operatorname{Ad})}+O_{h,Q,\eta,\varepsilon}(p^{3/2+\varepsilon}). N 2 ( a , b , p ) = 2 c cusp r , m , n ≥ 1 ( r mn , p ) = 1 ∑ r mn 1 π ∈ B ( p ) ∑ L ( 1 , π , Ad ) A π ( a , b ) A π ( m , n ) H r 2 mn ( μ π ) + O h , Q , η , ε ( p 3/2 + ε ) .
The weighted mass bound after (64) and (30) justify the interchanges and allow truncation at r 2 m n ≤ p 1 + ε r^{2}mn\leq p^{1+\varepsilon} r 2 mn ≤ p 1 + ε with error O h , Q , η , ε , B ( p − B ) O_{h,Q,\eta,\varepsilon,B}(p^{-B}) O h , Q , η , ε , B ( p − B ) , uniformly for a b ≤ Q 2 p 2 η ab\leq Q^{2}p^{2\eta} ab ≤ Q 2 p 2 η . Apply Theorem 3.1 with indices ( a , b , m , n ) (a,b,m,n) ( a , b , m , n ) and test H r 2 m n H_{r^{2}mn} H r 2 mn . Admissibility follows from Lemma 2.7 . Removing the truncation r 2 m n ≤ p 1 + ε r^{2}mn\leq p^{1+\varepsilon} r 2 mn ≤ p 1 + ε from the diagonal and the other spectral terms introduces an error of the same size. Define
D 2 ( a , b , p ) : = 2 P a b ∑ r ≥ 1 ( r , p ) = 1 1 r ∫ Re ( μ ) = 0 H r 2 a b ( μ ) spec ( μ ) d μ . (79) D_{2}(a,b,p):=\frac{2P}{\sqrt{ab}}\sum_{\substack{r\geq1\\(r,p)=1}}\frac{1}{r}\int_{\operatorname{Re}(\mu)=0}H_{r^{2}ab}(\mu)\operatorname{spec}(\mu)\,d\mu. \tag*{(79)} D 2 ( a , b , p ) := ab 2 P r ≥ 1 ( r , p ) = 1 ∑ r 1 ∫ Re ( μ ) = 0 H r 2 ab ( μ ) spec ( μ ) d μ . (79)
The three Weyl contributions are
S 4 ( a , b ) : = 2 P ∑ r , m , n ≥ 1 ( r m n , p ) = 1 r 2 m n ≤ p 1 + ε 1 r m n ∑ ε = ± 1 ∑ D 1 , D 2 ≥ 1 p D 2 ∣ D 1 b D 1 = m D 2 2 S ~ ( − ε n , b , a , D 2 , D 1 ) D 1 D 2 Φ w 4 , H r 2 m n ( ε a b n D 1 D 2 ) . (80) S_{4}(a,b):=2P\sum_{\substack{r,m,n\geq1\\(rmn,p)=1\\r^{2}mn\leq p^{1+\varepsilon}}}\frac{1}{r\sqrt{mn}}\sum_{\varepsilon=\pm1}\sum_{\substack{D_{1},D_{2}\geq1\\pD_{2}\mid D_{1}\\bD_{1}=mD_{2}^{2}}}\frac{\widetilde{S}(-\varepsilon n,b,a,D_{2},D_{1})}{D_{1}D_{2}}\Phi_{w_{4},H_{r^{2}mn}}\left(\frac{\varepsilon abn}{D_{1}D_{2}}\right). \tag*{(80)} S 4 ( a , b ) := 2 P r , m , n ≥ 1 ( r mn , p ) = 1 r 2 mn ≤ p 1 + ε ∑ r mn 1 ε = ± 1 ∑ D 1 , D 2 ≥ 1 p D 2 ∣ D 1 b D 1 = m D 2 2 ∑ D 1 D 2 S ( − ε n , b , a , D 2 , D 1 ) Φ w 4 , H r 2 mn ( D 1 D 2 ε abn ) . (80)
S 5 ( a , b ) : = 2 P ∑ r , m , n ≥ 1 ( r m n , p ) = 1 r 2 m n ≤ p 1 + ε 1 r m n ∑ ε = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 ∣ D 2 a D 2 = n D 1 2 S ~ ( ε m , a , b , D 1 , D 2 ) D 1 D 2 Φ w 5 , H r 2 m n ( ε m a b D 1 D 2 ) . (81) S_{5}(a,b):=2P\sum_{\substack{r,m,n\geq1\\(rmn,p)=1\\r^{2}mn\leq p^{1+\varepsilon}}}\frac{1}{r\sqrt{mn}}\sum_{\varepsilon=\pm1}\sum_{\substack{D_{1},D_{2}\geq1\\p\mid D_{1}\mid D_{2}\\aD_{2}=nD_{1}^{2}}}\frac{\widetilde{S}(\varepsilon m,a,b,D_{1},D_{2})}{D_{1}D_{2}}\Phi_{w_{5},H_{r^{2}mn}}\left(\frac{\varepsilon mab}{D_{1}D_{2}}\right). \tag*{(81)} S 5 ( a , b ) := 2 P r , m , n ≥ 1 ( r mn , p ) = 1 r 2 mn ≤ p 1 + ε ∑ r mn 1 ε = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 ∣ D 2 a D 2 = n D 1 2 ∑ D 1 D 2 S ( ε m , a , b , D 1 , D 2 ) Φ w 5 , H r 2 mn ( D 1 D 2 ε mab ) . (81)
S 6 ( a , b ) : = 2 P ∑ r , m , n ≥ 1 ( r m n , p ) = 1 r 2 m n ≤ p 1 + ε 1 r m n ∑ ε 1 , ε 2 = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , p ∣ D 2 S ( p ) ( ε 2 n , ε 1 m , a , b , D 1 , D 2 ) D 1 D 2 × Φ w 6 , H r 2 m n ( − ε 2 a n D 2 D 1 2 , − ε 1 b m D 1 D 2 2 ) . (82) \begin{aligned}
S_{6}(a,b):={}&2P\sum_{\substack{r,m,n\geq1\\(rmn,p)=1\\r^{2}mn\leq p^{1+\varepsilon}}}\frac{1}{r\sqrt{mn}}\sum_{\varepsilon_{1},\varepsilon_{2}=\pm1}\sum_{\substack{D_{1},D_{2}\geq1\\p\mid D_{1},\,p\mid D_{2}}}\frac{S^{(p)}(\varepsilon_{2}n,\varepsilon_{1}m,a,b,D_{1},D_{2})}{D_{1}D_{2}}\\
&\times\Phi_{w_{6},H_{r^{2}mn}}\left(-\frac{\varepsilon_{2}anD_{2}}{D_{1}^{2}},-\frac{\varepsilon_{1}bmD_{1}}{D_{2}^{2}}\right). \tag*{(82)}
\end{aligned} S 6 ( a , b ) := 2 P r , m , n ≥ 1 ( r mn , p ) = 1 r 2 mn ≤ p 1 + ε ∑ r mn 1 ε 1 , ε 2 = ± 1 ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , p ∣ D 2 ∑ D 1 D 2 S ( p ) ( ε 2 n , ε 1 m , a , b , D 1 , D 2 ) × Φ w 6 , H r 2 mn ( − D 1 2 ε 2 an D 2 , − D 2 2 ε 1 bm D 1 ) . (82)
The four remaining spectral contributions, with full index sums, are
R o l d ( a , b ) : = 2 c c u s p ∑ r , m , n ≥ 1 ( r m n , p ) = 1 1 r m n ∑ π ∈ B ( 1 ) ∑ j = 0 2 B π ( j ) ( a , b ) ‾ B π ( j ) ( m , n ) H r 2 m n ( μ π ) L ( 1 , π , Ad ) , \mathcal{R}_{\mathrm{old}}(a,b) := 2c_{\mathrm{cusp}} \sum_{\substack{r,m,n\ge1\\(rmn,p)=1}} \frac{1}{r\sqrt{mn}} \sum_{\pi\in\mathcal{B}(1)} \sum_{j=0}^{2} \frac{\overline{B_{\pi}^{(j)}(a,b)}B_{\pi}^{(j)}(m,n)H_{r2mn}(\mu_{\pi})}{L(1,\pi,\operatorname{Ad})}, R old ( a , b ) := 2 c cusp r , m , n ≥ 1 ( r mn , p ) = 1 ∑ r mn 1 π ∈ B ( 1 ) ∑ j = 0 ∑ 2 L ( 1 , π , Ad ) B π ( j ) ( a , b ) B π ( j ) ( m , n ) H r 2 mn ( μ π ) ,
R m i n ( a , b ) : = 2 c m i n ∑ r , m , n ≥ 1 ( r m n , p ) = 1 1 r m n ∑ j = 0 2 ∫ Re ( s 1 ) = 0 ∫ Re ( s 2 ) = 0 B s ( j ) ( a , b ) ‾ B s ( j ) ( m , n ) H r 2 m n ( s ) ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 d s 1 d s 2 , \mathcal{R}_{\mathrm{min}}(a,b) := 2c_{\mathrm{min}} \sum_{\substack{r,m,n\ge1\\(rmn,p)=1}} \frac{1}{r\sqrt{mn}} \sum_{j=0}^{2} \int_{\operatorname{Re}(s_1)=0}\int_{\operatorname{Re}(s_2)=0} \frac{\overline{B_s^{(j)}(a,b)}B_s^{(j)}(m,n)H_{r2mn}(s)}{\prod_{1\le i<k\le3}|\zeta(1+s_i-s_k)|^2}\,ds_1\,ds_2, R min ( a , b ) := 2 c min r , m , n ≥ 1 ( r mn , p ) = 1 ∑ r mn 1 j = 0 ∑ 2 ∫ Re ( s 1 ) = 0 ∫ Re ( s 2 ) = 0 ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 B s ( j ) ( a , b ) B s ( j ) ( m , n ) H r 2 mn ( s ) d s 1 d s 2 ,
R m a x , n e w ( a , b ) : = 2 χ p c m a x ∑ r , m , n ≥ 1 ( r m n , p ) = 1 1 r m n ∑ ϕ ∈ B 2 ( p ) ∫ Re ( z ) = 0 A ϕ , z ( a , b ) A ϕ , z ( m , n ) H r 2 m n ( z + τ ϕ , z − τ ϕ , − 2 z ) L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 d z , \mathcal{R}_{\mathrm{max,new}}(a,b) := 2\chi_p c_{\mathrm{max}} \sum_{\substack{r,m,n\ge1\\(rmn,p)=1}} \frac{1}{r\sqrt{mn}} \sum_{\phi\in\mathcal{B}_2(p)}\int_{\operatorname{Re}(z)=0} \frac{A_{\phi,z}(a,b)A_{\phi,z}(m,n)H_{r2mn}(z+\tau_\phi,z-\tau_\phi,-2z)}{L(1,\phi,\operatorname{Ad})|L(1+3z,\phi)|^2}\,dz, R max , new ( a , b ) := 2 χ p c max r , m , n ≥ 1 ( r mn , p ) = 1 ∑ r mn 1 ϕ ∈ B 2 ( p ) ∑ ∫ Re ( z ) = 0 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 A ϕ , z ( a , b ) A ϕ , z ( m , n ) H r 2 mn ( z + τ ϕ , z − τ ϕ , − 2 z ) d z ,
R m a x , o l d ( a , b ) : = 2 c m a x ∑ r , m , n ≥ 1 ( r m n , p ) = 1 1 r m n ∑ ϕ ∈ B 2 ( 1 ) ∑ j = 0 2 ∫ Re ( z ) = 0 B ϕ , z ( j ) ( a , b ) ‾ B ϕ , z ( j ) ( m , n ) H r 2 m n ( z + τ ϕ , z − τ ϕ , − 2 z ) L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 d z . (83) \mathcal{R}_{\mathrm{max,old}}(a,b) := 2c_{\mathrm{max}} \sum_{\substack{r,m,n\ge1\\(rmn,p)=1}} \frac{1}{r\sqrt{mn}} \sum_{\phi\in\mathcal{B}_2(1)}\sum_{j=0}^{2}\int_{\operatorname{Re}(z)=0} \frac{\overline{B_{\phi,z}^{(j)}(a,b)}B_{\phi,z}^{(j)}(m,n)H_{r2mn}(z+\tau_\phi,z-\tau_\phi,-2z)}{L(1,\phi,\operatorname{Ad})|L(1+3z,\phi)|^2}\,dz. \tag*{(83)} R max , old ( a , b ) := 2 c max r , m , n ≥ 1 ( r mn , p ) = 1 ∑ r mn 1 ϕ ∈ B 2 ( 1 ) ∑ j = 0 ∑ 2 ∫ Re ( z ) = 0 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 B ϕ , z ( j ) ( a , b ) B ϕ , z ( j ) ( m , n ) H r 2 mn ( z + τ ϕ , z − τ ϕ , − 2 z ) d z . (83)
The transforms and continuous measures are normalized as in § 3.5. The trace expansion is
χ p N 2 ( a , b , p ) = D 2 ( a , b , p ) + ∑ i = 4 6 S i ( a , b ) − R o l d ( a , b ) − R m i n ( a , b ) − R m a x , n e w ( a , b ) − R m a x , o l d ( a , b ) + O h , Q , η , ε ( p 3 / 2 + ε ) . (84) \chi_p\mathcal{N}_2(a,b,p)=\mathcal{D}_2(a,b,p)+\sum_{i=4}^{6}\mathcal{S}_i(a,b)-\mathcal{R}_{\mathrm{old}}(a,b)-\mathcal{R}_{\mathrm{min}}(a,b)
-\mathcal{R}_{\mathrm{max,new}}(a,b)-\mathcal{R}_{\mathrm{max,old}}(a,b)+O_{h,Q,\eta,\varepsilon}(p^{3/2+\varepsilon}). \tag*{(84)} χ p N 2 ( a , b , p ) = D 2 ( a , b , p ) + i = 4 ∑ 6 S i ( a , b ) − R old ( a , b ) − R min ( a , b ) − R max , new ( a , b ) − R max , old ( a , b ) + O h , Q , η , ε ( p 3/2 + ε ) . (84)
The diagonal term
The diagonal conditions are m = a m=a m = a , n = b n=b n = b . With R μ R_\mu R μ as defined in the proof of Lemma Lemma 2.7 , (79) and (28) give, by the Mellin inversion formula on Re ( u ) = c > 1 \operatorname{Re}(u)=c>1 Re ( u ) = c > 1 ,
2 ∑ r ≥ 1 ( r , p ) = 1 1 r U μ ( r 2 a b / p ) = 2 2 π i ∫ ( c ) ( a b p ) − u G J ( u , μ ) R μ ( u ) ζ ( p ) ( 1 + 2 u ) d u u , 2\sum_{\substack{r\ge1\\(r,p)=1}}\frac{1}{r}U_\mu(r^2ab/p)=\frac{2}{2\pi i}\int_{(c)}\left(\frac{ab}{p}\right)^{-u}G_J(u,\mu)R_\mu(u)\zeta^{(p)}(1+2u)\frac{du}{u}, 2 r ≥ 1 ( r , p ) = 1 ∑ r 1 U μ ( r 2 ab / p ) = 2 π i 2 ∫ ( c ) ( p ab ) − u G J ( u , μ ) R μ ( u ) ζ ( p ) ( 1 + 2 u ) u d u ,
where ζ ( p ) ( w ) = ( 1 − p − w ) ζ ( w ) \zeta^{(p)}(w)=(1-p^{-w})\zeta(w) ζ ( p ) ( w ) = ( 1 − p − w ) ζ ( w ) . For Re ( μ ) = 0 \operatorname{Re}(\mu)=0 Re ( μ ) = 0 , move the contour to Re ( u ) = − 1 / 4 \operatorname{Re}(u)=-1/4 Re ( u ) = − 1/4 , crossing only the double pole at u = 0 u=0 u = 0 . To compute its residue, first use the Laurent expansion of the zeta factor,
ζ ( p ) ( 1 + 2 u ) = 1 − p − 1 2 u + ( 1 − p − 1 ) γ + log p p + O p ( u ) . \zeta^{(p)}(1+2u)=\frac{1-p^{-1}}{2u}+(1-p^{-1})\gamma+\frac{\log p}{p}+O_p(u). ζ ( p ) ( 1 + 2 u ) = 2 u 1 − p − 1 + ( 1 − p − 1 ) γ + p log p + O p ( u ) .
The remaining factor is holomorphic and equals one at zero. Since G J G_J G J is even and R μ ′ ( 0 ) = κ ( μ ) R_\mu'(0)=\kappa(\mu) R μ ′ ( 0 ) = κ ( μ ) , its Taylor expansion is
( p a b ) u G J ( u , μ ) R μ ( u ) = 1 + u ( log p a b + κ ( μ ) ) + O p , a , b , μ , J ( u 2 ) . \left(\frac{p}{ab}\right)^uG_J(u,\mu)R_\mu(u)=1+u\left(\log\frac{p}{ab}+\kappa(\mu)\right)+O_{p,a,b,\mu,J}(u^2). ( ab p ) u G J ( u , μ ) R μ ( u ) = 1 + u ( log ab p + κ ( μ ) ) + O p , a , b , μ , J ( u 2 ) .
Multiplying these expansions by 2 / u 2/u 2/ u and applying the residue theorem gives
2 ∑ r ≥ 1 ( r , p ) = 1 1 r U μ ( r 2 a b / p ) = ( 1 − p − 1 ) ( log p a b + 2 γ + κ ( μ ) ) + 2 log p p + O J ( ( a b / p ) 1 / 4 ( 1 + ∥ μ ∥ ) C ) . 2\sum_{\substack{r\geq1\\(r,p)=1}}\frac{1}{r}U_\mu(r^2ab/p)=(1-p^{-1})\left(\log\frac{p}{ab}+2\gamma+\kappa(\mu)\right)+\frac{2\log p}{p}+O_J\left((ab/p)^{1/4}(1+\lVert\mu\rVert)^C\right). 2 r ≥ 1 ( r , p ) = 1 ∑ r 1 U μ ( r 2 ab / p ) = ( 1 − p − 1 ) ( log ab p + 2 γ + κ ( μ ) ) + p 2 log p + O J ( ( ab / p ) 1/4 ( 1 + ∥ μ ∥ ) C ) .
The error comes from the integral on Re ( u ) = − 1 / 4 \operatorname{Re}(u)=-1/4 Re ( u ) = − 1/4 , where ∣ 1 − p − 1 − 2 u ∣ ≤ 1 + p − 1 / 2 \lvert1-p^{-1-2u}\rvert\leq1+p^{-1/2} ∣ 1 − p − 1 − 2 u ∣ ≤ 1 + p − 1/2 . By (19) and (18) , for Re ( μ ) = 0 \operatorname{Re}(\mu)=0 Re ( μ ) = 0 ,
∣ G J ( − 1 / 4 + i v , μ ) R μ ( − 1 / 4 + i v ) ζ ( p ) ( 1 / 2 + 2 i v ) ∣ ≪ J ( 1 + ∥ μ ∥ + ∣ v ∣ ) C e − v 2 + 3 π ∣ v ∣ / 2 . \lvert G_J(-1/4+iv,\mu)R_\mu(-1/4+iv)\zeta^{(p)}(1/2+2iv)\rvert\ll_J(1+\lVert\mu\rVert+\lvert v\rvert)^C e^{-v^2+3\pi\lvert v\rvert/2}. ∣ G J ( − 1/4 + i v , μ ) R μ ( − 1/4 + i v ) ζ ( p ) ( 1/2 + 2 i v )∣ ≪ J ( 1 + ∥ μ ∥ + ∣ v ∣ ) C e − v 2 + 3 π ∣ v ∣ /2 .
This justifies the integral estimate and the horizontal limits. Integrating against h h h using (38) gives
D 2 ( a , b , p ) = P a b ( ( 1 − p − 1 ) { H h ( log p a b + 2 γ ) + K h } + 2 log p p H h ) + O h , η ( p 7 / 4 ( a b ) − 1 / 4 ) . (85) \mathcal{D}_2(a,b,p)=\frac{P}{\sqrt{ab}}\left((1-p^{-1})\left\{\mathcal{H}_h\left(\log\frac{p}{ab}+2\gamma\right)+\mathcal{K}_h\right\}+\frac{2\log p}{p}\mathcal{H}_h\right)+O_{h,\eta}\left(p^{7/4}(ab)^{-1/4}\right). \tag*{(85)} D 2 ( a , b , p ) = ab P ( ( 1 − p − 1 ) { H h ( log ab p + 2 γ ) + K h } + p 2 log p H h ) + O h , η ( p 7/4 ( ab ) − 1/4 ) . (85)
The short Weyl elements
Lemma 4.10. Uniformly in the twist range of Theorem 4.9 , S 4 ( a , b ) S_4(a,b) S 4 ( a , b ) and S 5 ( a , b ) S_5(a,b) S 5 ( a , b ) are O h , Q , η ( p − 1 ) O_{h,Q,\eta}(p^{-1}) O h , Q , η ( p − 1 ) .
Proof. Choose A A A sufficiently large in terms of η \eta η and take the tube width a 0 a_0 a 0 large enough for the kernel estimates. By (31) and (53) ,
∣ Φ w 4 , H t ( y ) ∣ + ∣ Φ w 5 , H t ( y ) ∣ ≪ h , η , ε p ε ∣ y ∣ A , t ≥ 1. \lvert\Phi_{w_4,H_t}(y)\rvert+\lvert\Phi_{w_5,H_t}(y)\rvert\ll_{h,\eta,\varepsilon}p^\varepsilon\lvert y\rvert^A,\qquad t\geq1. ∣ Φ w 4 , H t ( y )∣ + ∣ Φ w 5 , H t ( y )∣ ≪ h , η , ε p ε ∣ y ∣ A , t ≥ 1.
The modulus conditions in (80) and (81) give, respectively,
D 2 = p b k ( b , m ) , D 1 = p 2 m b k 2 ( b , m ) 2 , D_2=\frac{pbk}{(b,m)},\qquad D_1=\frac{p^2mbk^2}{(b,m)^2}, D 2 = ( b , m ) p bk , D 1 = ( b , m ) 2 p 2 mb k 2 ,
D 1 = p a k ( a , n ) , D 2 = p 2 n a k 2 ( a , n ) 2 , k ≥ 1. D_1=\frac{pak}{(a,n)},\qquad D_2=\frac{p^2nak^2}{(a,n)^2},\qquad k\geq1. D 1 = ( a , n ) p ak , D 2 = ( a , n ) 2 p 2 na k 2 , k ≥ 1.
In both cases D 1 D 2 ≥ p 3 k 3 D_1D_2\geq p^3k^3 D 1 D 2 ≥ p 3 k 3 . Hence, on r 2 m n ≤ p 1 + ε r^2mn\leq p^{1+\varepsilon} r 2 mn ≤ p 1 + ε , the kernel argument satisfies
∣ y ∣ ≤ a b p − 2 + ε k 3 . \lvert y\rvert\leq\frac{abp^{-2+\varepsilon}}{k^3}. ∣ y ∣ ≤ k 3 ab p − 2 + ε .
A direct computation gives
∑ r 2 m n ≤ p 1 + ε 1 r m n ≪ ε p 1 / 2 + ε . (86) \sum_{r^2mn\leq p^{1+\varepsilon}}\frac{1}{r\sqrt{mn}}\ll_\varepsilon p^{1/2+\varepsilon}. \tag*{(86)} r 2 mn ≤ p 1 + ε ∑ r mn 1 ≪ ε p 1/2 + ε . (86)
Using ∣ S ~ ∣ ≤ D 1 D 2 \lvert\widetilde{S}\rvert\leq D_1D_2 ∣ S ∣ ≤ D 1 D 2 , P ≪ p 2 P\ll p^2 P ≪ p 2 and ∑ k ≥ 1 k − 3 A ≪ A 1 \sum_{k\geq1}k^{-3A}\ll_A 1 ∑ k ≥ 1 k − 3 A ≪ A 1 , we obtain
∣ S 4 ( a , b ) ∣ + ∣ S 5 ( a , b ) ∣ ≪ h , η , ε p 5 / 2 + ( A + 2 ) ε ( a b ) A p − 2 A . \lvert S_4(a,b)\rvert+\lvert S_5(a,b)\rvert\ll_{h,\eta,\varepsilon}p^{5/2+(A+2)\varepsilon}(ab)^A p^{-2A}. ∣ S 4 ( a , b )∣ + ∣ S 5 ( a , b )∣ ≪ h , η , ε p 5/2 + ( A + 2 ) ε ( ab ) A p − 2 A .
Since a b ≤ Q 2 p 2 η ab\leq Q^2p^{2\eta} ab ≤ Q 2 p 2 η , Q Q Q is fixed and ε \varepsilon ε is arbitrarily small, the claim follows by our choice of A A A . □ \square □
The long Weyl element
Lemma 4.11. Uniformly in the twist range of Theorem 4.9 , S 6 ( a , b ) = O h , Q , η ( p − 1 ) S_6(a,b)=O_{h,Q,\eta}(p^{-1}) S 6 ( a , b ) = O h , Q , η ( p − 1 ) .
Proof. Use the same choice of A A A as in the preceding proof. (54) and (31) give
∣ Φ w 6 , H t ( y 1 , y 2 ) ∣ ≪ h , η , ε p ε min ( ∣ y 1 2 y 2 ∣ , ∣ y 1 y 2 2 ∣ ) A ( 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ) ε . \left|\Phi_{w_6,H_t}(y_1,y_2)\right|\ll_{h,\eta,\varepsilon}p^\varepsilon\min\left(\left|y_1^2y_2\right|,\left|y_1y_2^2\right|\right)^A(1+|y_1|+|y_2|)^\varepsilon. ∣ Φ w 6 , H t ( y 1 , y 2 ) ∣ ≪ h , η , ε p ε min ( y 1 2 y 2 , y 1 y 2 2 ) A ( 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ) ε .
The arguments in (82) satisfy
∣ y 1 2 y 2 ∣ = a 2 b m n 2 D 1 3 , ∣ y 1 y 2 2 ∣ = a b 2 m 2 n D 2 3 . \left|y_1^2y_2\right|=\frac{a^2bmn^2}{D_1^3},\qquad\left|y_1y_2^2\right|=\frac{ab^2m^2n}{D_2^3}. y 1 2 y 2 = D 1 3 a 2 bm n 2 , y 1 y 2 2 = D 2 3 a b 2 m 2 n .
Further, we have
min ( ∣ y 1 2 y 2 ∣ , ∣ y 1 y 2 2 ∣ ) A ≤ ( a b m n ) 3 A / 2 ( D 1 D 2 ) 3 A / 2 . \min\left(\left|y_1^2y_2\right|,\left|y_1y_2^2\right|\right)^A\leq\frac{(abmn)^{3A/2}}{(D_1D_2)^{3A/2}}. min ( y 1 2 y 2 , y 1 y 2 2 ) A ≤ ( D 1 D 2 ) 3 A /2 ( abmn ) 3 A /2 .
For sufficiently large p p p , we have a n , b m ≤ a b m n ≤ Q 2 p 2 η + 1 + ε < p 3 an,bm\leq abmn\leq Q^2p^{2\eta+1+\varepsilon}<p^3 an , bm ≤ abmn ≤ Q 2 p 2 η + 1 + ε < p 3 . Since D 1 , D 2 ≥ p D_1,D_2\geq p D 1 , D 2 ≥ p , this gives 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ≤ 3 D 1 D 2 1+|y_1|+|y_2|\leq3D_1D_2 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ≤ 3 D 1 D 2 . Using ∣ S ( p ) ∣ ≤ ( D 1 D 2 ) 2 |S^{(p)}|\leq(D_1D_2)^2 ∣ S ( p ) ∣ ≤ ( D 1 D 2 ) 2 and
∑ p ∣ D 1 , D 2 ( D 1 D 2 ) − 3 A / 2 + ε = p 2 − 3 A + 2 ε ζ ( 3 A / 2 − 1 − ε ) 2 , \sum_{p\mid D_1,D_2}(D_1D_2)^{-3A/2+\varepsilon}=p^{2-3A+2\varepsilon}\zeta(3A/2-1-\varepsilon)^2, p ∣ D 1 , D 2 ∑ ( D 1 D 2 ) − 3 A /2 + ε = p 2 − 3 A + 2 ε ζ ( 3 A /2 − 1 − ε ) 2 ,
and then (86) , gives
∣ S 6 ( a , b ) ∣ ≪ h , η , ε p 9 / 2 + ( 3 A / 2 + 4 ) ε ( a b ) 3 A / 2 p − 3 A / 2 . |S_6(a,b)|\ll_{h,\eta,\varepsilon}p^{9/2+(3A/2+4)\varepsilon}(ab)^{3A/2}p^{-3A/2}. ∣ S 6 ( a , b ) ∣ ≪ h , η , ε p 9/2 + ( 3 A /2 + 4 ) ε ( ab ) 3 A /2 p − 3 A /2 .
Since a b ≤ Q 2 p 2 η ab\leq Q^2p^{2\eta} ab ≤ Q 2 p 2 η , Q Q Q is fixed and η < 1 / 2 \eta<1/2 η < 1/2 , our choice of A A A and the arbitrarily small ε \varepsilon ε give the claimed bound. The sum over D 1 , D 2 D_1,D_2 D 1 , D 2 converges absolutely, completing the proof. □
The old and minimal Eisenstein spectra
By (40) –(42) , for ( a b m n , p ) = 1 (abmn,p)=1 ( abmn , p ) = 1 ,
∑ j = 0 2 B Π ( j ) ( a , b ) ‾ B Π ( j ) ( m , n ) = A Π ( a , b ) ‾ A Π ( m , n ) ( Q Π − 1 ‾ ) 00 , \sum_{j=0}^{2}\overline{B_{\Pi}^{(j)}(a,b)}B_{\Pi}^{(j)}(m,n)=\overline{A_{\Pi}(a,b)}A_{\Pi}(m,n)\left(\overline{Q_{\Pi}^{-1}}\right)_{00}, j = 0 ∑ 2 B Π ( j ) ( a , b ) B Π ( j ) ( m , n ) = A Π ( a , b ) A Π ( m , n ) ( Q Π − 1 ) 00 ,
where ( Q Π − 1 ‾ ) 00 ≪ 1 \left(\overline{Q_{\Pi}^{-1}}\right)_{00}\ll1 ( Q Π − 1 ) 00 ≪ 1 by Lemma 4.4 . For the coefficient sums below, fix θ ≥ 0 \theta\geq0 θ ≥ 0 and B > 2 + θ + ε B>2+\theta+\varepsilon B > 2 + θ + ε . Applying the divisor bound separately to r ≤ p r\leq\sqrt p r ≤ p and r > p r>\sqrt p r > p gives
∑ r , m , n ≥ 1 ( m n ) θ r m n ( 1 + r 2 m n p ) − B ≪ B , θ , ε p 1 / 2 + θ + ε ∑ r ≤ p r − 2 − 2 θ − 2 ε + p B ∑ r > p r − 1 − 2 B (87) \sum_{r,m,n\geq1}\frac{(mn)^\theta}{r\sqrt{mn}}\left(1+\frac{r^2mn}{p}\right)^{-B}
\ll_{B,\theta,\varepsilon}p^{1/2+\theta+\varepsilon}\sum_{r\leq\sqrt p}r^{-2-2\theta-2\varepsilon}+p^B\sum_{r>\sqrt p}r^{-1-2B} \tag*{(87)} r , m , n ≥ 1 ∑ r mn ( mn ) θ ( 1 + p r 2 mn ) − B ≪ B , θ , ε p 1/2 + θ + ε r ≤ p ∑ r − 2 − 2 θ − 2 ε + p B r > p ∑ r − 1 − 2 B (87)
≪ B , θ , ε p 1 / 2 + θ + ε . \ll_{B,\theta,\varepsilon}p^{1/2+\theta+\varepsilon}. ≪ B , θ , ε p 1/2 + θ + ε .
Proposition 4.12. Uniformly in the twist range, we have
R o l d ( a , b ) ≪ h , η , ε ( a b ) 5 / 14 + ε p 6 / 7 + ε , \mathcal{R}_{\mathrm{old}}(a,b)\ll_{h,\eta,\varepsilon}(ab)^{5/14+\varepsilon}p^{6/7+\varepsilon}, R old ( a , b ) ≪ h , η , ε ( ab ) 5/14 + ε p 6/7 + ε , R min ( a , b ) ≪ h , η , ε ( a b ) ε p 1 / 2 + ε , \mathcal{R}_{\min}(a,b)\ll_{h,\eta,\varepsilon}(ab)^\varepsilon p^{1/2+\varepsilon}, R m i n ( a , b ) ≪ h , η , ε ( ab ) ε p 1/2 + ε , and
R max , o l d ( a , b ) ≪ h , η , ε ( a b ) 7 / 64 + ε p 39 / 64 + ε . \mathcal{R}_{\max,\mathrm{old}}(a,b)\ll_{h,\eta,\varepsilon}(ab)^{7/64+\varepsilon}p^{39/64+\varepsilon}. R m a x , old ( a , b ) ≪ h , η , ε ( ab ) 7/64 + ε p 39/64 + ε .
Proof. By (70) and (71) , rapid decrease of h h h makes the level one spectral sums and integrals with any fixed polynomial weight bounded independently of p p p . The twist contributes ( a b ) θ + ε (ab)^{\theta+\varepsilon} ( ab ) θ + ε , while (30) and (87) bound the remaining index sums by p 1 / 2 + θ + ε p^{1/2+\theta+\varepsilon} p 1/2 + θ + ε . Taking θ = 5 / 14 , 0 , 7 / 64 \theta=5/14,0,7/64 θ = 5/14 , 0 , 7/64 for the old cuspidal, minimal and maximal old spectra proves the three estimates. □
The maximal Eisenstein new spectrum
Proposition 4.13. Uniformly in the twist range,
R m a x , n e w ( a , b ) ≪ h , Q , η , ε p 3 / 2 + ε . \mathcal{R}_{\mathrm{max,new}}(a,b)\ll_{h,Q,\eta,\varepsilon}p^{3/2+\varepsilon}. R max , new ( a , b ) ≪ h , Q , η , ε p 3/2 + ε .
Proof. Apply (73) in (83) . For sufficiently large p p p , we have a b ≤ Q 2 p 2 η < p ab\le Q^2p^{2\eta}<p ab ≤ Q 2 p 2 η < p . Together with χ p ≪ 1 \chi_p\ll1 χ p ≪ 1 and 1 + m n / p ≤ 1 + r 2 m n / p 1+mn/p\le1+r^2mn/p 1 + mn / p ≤ 1 + r 2 mn / p , (30) gives
∣ R m a x , n e w ( a , b ) ∣ ≪ h , η , B , ε p 1 + ε ( a b ) ε ∑ r , m , n ≥ 1 ( m n ) ε r m n ( 1 + r 2 m n p ) − B + 1 \left|\mathcal{R}_{\mathrm{max,new}}(a,b)\right|\ll_{h,\eta,B,\varepsilon}p^{1+\varepsilon}(ab)^\varepsilon\sum_{r,m,n\ge1}\frac{(mn)^\varepsilon}{r\sqrt{mn}}\left(1+\frac{r^2mn}{p}\right)^{-B+1} ∣ R max , new ( a , b ) ∣ ≪ h , η , B , ε p 1 + ε ( ab ) ε r , m , n ≥ 1 ∑ r mn ( mn ) ε ( 1 + p r 2 mn ) − B + 1
≪ h , Q , η , ε p 3 / 2 + ε . \ll_{h,Q,\eta,\varepsilon}p^{3/2+\varepsilon}. ≪ h , Q , η , ε p 3/2 + ε .
The last step follows from (87) with θ = ε \theta=\varepsilon θ = ε and B B B sufficiently large. □
Proof of Theorem 4.9
Proof of Theorem 4.9. Combine (85) , Lemmas 4.10 and 4.11, and Propositions 4.12 and 4.13 in (84) . For sufficiently large p p p , a b ≤ Q 2 p 2 η < p ab\le Q^2p^{2\eta}<p ab ≤ Q 2 p 2 η < p . The level one remainders are O h , Q , η , ε ( p 17 / 14 + ε ) O_{h,Q,\eta,\varepsilon}(p^{17/14+\varepsilon}) O h , Q , η , ε ( p 17/14 + ε ) and are absorbed by p 3 / 2 + ε p^{3/2+\varepsilon} p 3/2 + ε . Dividing by χ p = 1 + O ( p − 1 ) \chi_p=1+O(p^{-1}) χ p = 1 + O ( p − 1 ) changes the diagonal main term by O h , Q , η ( p log p / a b ) O_{h,Q,\eta}(p\log p/\sqrt{ab}) O h , Q , η ( p log p / ab ) , which is absorbed by p 7 / 4 ( a b ) − 1 / 4 p^{7/4}(ab)^{-1/4} p 7/4 ( ab ) − 1/4 . This proves (77) .
For the uniform asymptotic, a b ≤ Q 2 p 2 η ab\le Q^2p^{2\eta} ab ≤ Q 2 p 2 η gives
a b p 2 ( p 7 / 4 ( a b ) − 1 / 4 + p 3 / 2 + ε ) ≤ Q 1 / 2 p − 1 / 4 + η / 2 + Q p − 1 / 2 + η + ε = o Q , η ( 1 ) . \frac{\sqrt{ab}}{p^2}\left(p^{7/4}(ab)^{-1/4}+p^{3/2+\varepsilon}\right)\le Q^{1/2}p^{-1/4+\eta/2}+Qp^{-1/2+\eta+\varepsilon}=o_{Q,\eta}(1). p 2 ab ( p 7/4 ( ab ) − 1/4 + p 3/2 + ε ) ≤ Q 1/2 p − 1/4 + η /2 + Q p − 1/2 + η + ε = o Q , η ( 1 ) .
Here η < 1 / 2 \eta<1/2 η < 1/2 and ε \varepsilon ε is arbitrarily small. Since H h > 0 \mathcal{H}_h>0 H h > 0 and
log p a b ≥ ( 1 − 2 η ) log p − 2 log Q ≫ Q , η log p , \log\frac{p}{ab}\ge(1-2\eta)\log p-2\log Q\gg_{Q,\eta}\log p, log ab p ≥ ( 1 − 2 η ) log p − 2 log Q ≫ Q , η log p ,
the constant term and the finite Euler corrections have relative size O h , Q , η ( 1 / log p ) O_{h,Q,\eta}(1/\log p) O h , Q , η ( 1/ log p ) . The stated asymptotic follows. Taking Q = 1 Q=1 Q = 1 and a = b = 1 a=b=1 a = b = 1 gives (78) .
Mollified harmonic moments
We evaluate the mollified harmonic first and second moments using the uniform twisted moment formulas of § 4. We also compute the holomorphic mixed moment associated with M ^ π \widehat{M}_{\pi} M π in (5) . These estimates will be combined in § 6 to prove the non-vanishing theorem.
Mollified harmonic first moment
Recall the mollified harmonic first moment M 1 ( p ) M_1(p) M 1 ( p ) defined in (3) .
Theorem 5.1. Fix a square-free integer Q ≥ 1 Q\ge1 Q ≥ 1 and 0 < η < 1 / 2 0<\eta<1/2 0 < η < 1/2 , independently of p p p , and put L = p η L=p^\eta L = p η . As p → ∞ p\to\infty p → ∞ through primes p ∤ Q p\nmid Q p ∤ Q , we have
M 1 ( p ) = P H h + O h , Q , η , ε ( p 5 / 4 + ε L 1 / 2 ) M_1(p)=P\mathcal{H}_h+O_{h,Q,\eta,\varepsilon}\left(p^{5/4+\varepsilon}L^{1/2}\right) M 1 ( p ) = P H h + O h , Q , η , ε ( p 5/4 + ε L 1/2 ) = P H h + O h , Q , η ( p 3 / 2 ) . (88) =P\mathcal{H}_h+O_{h,Q,\eta}(p^{3/2}). \tag*{(88)} = P H h + O h , Q , η ( p 3/2 ) . (88)
Proof. The finite product in (2) expands as
∏ q ∣ Q ( 1 − A π ( 1 , q ) q + A π ( q , 1 ) q − 1 q 3 / 2 ) = ∑ u v w ∣ Q μ ( u ) μ ( w ) u v w 3 / 2 A π ( v , u ) , \prod_{q\mid Q}\left(1-\frac{A_\pi(1,q)}{\sqrt{q}}+\frac{A_\pi(q,1)}{q}-\frac{1}{q^{3/2}}\right)=\sum_{uvw\mid Q}\frac{\mu(u)\mu(w)}{\sqrt{u}v w^{3/2}}A_\pi(v,u), q ∣ Q ∏ ( 1 − q A π ( 1 , q ) + q A π ( q , 1 ) − q 3/2 1 ) = uv w ∣ Q ∑ u v w 3/2 μ ( u ) μ ( w ) A π ( v , u ) ,
where u , v , w u,v,w u , v , w are positive and pairwise coprime. Since p ∤ Q p\nmid Q p ∤ Q , ( ℓ , Q ) = 1 (\ell,Q)=1 ( ℓ , Q ) = 1 on the support of x ℓ x_\ell x ℓ , and ℓ ≤ L < p \ell\le L<p ℓ ≤ L < p , multiplicativity and (11) give A π ( v , u ) A π ( 1 , ℓ ) = A π ( v , u ℓ ) = A π ( u ℓ , v ) A_\pi(v,u)A_\pi(1,\ell)=A_\pi(v,u\ell)=A_\pi(u\ell,v) A π ( v , u ) A π ( 1 , ℓ ) = A π ( v , u ℓ ) = A π ( u ℓ , v ) . Recalling the definition of M h M_h M h in (61) , we obtain
M 1 ( p ) = ∑ u v w ∣ Q μ ( u ) μ ( w ) u v w 3 / 2 ∑ ℓ ≤ L x ℓ ℓ M h ( u ℓ , v , p ) . (89) M_1(p)=\sum_{uvw\mid Q}\frac{\mu(u)\mu(w)}{\sqrt{u}v w^{3/2}}\sum_{\ell\le L}\frac{x_\ell}{\sqrt{\ell}}M_h(u\ell,v,p). \tag*{(89)} M 1 ( p ) = uv w ∣ Q ∑ u v w 3/2 μ ( u ) μ ( w ) ℓ ≤ L ∑ ℓ x ℓ M h ( u ℓ , v , p ) . (89)
Here u ℓ ≤ Q p η u\ell\le Qp^\eta u ℓ ≤ Q p η and v ≤ Q v\le Q v ≤ Q . So Theorem 4.1 applies uniformly to every term in (89) . Its main term requires u ℓ = 1 u\ell=1 u ℓ = 1 , hence u = ℓ = 1 u=\ell=1 u = ℓ = 1 . Since x 1 = 1 x_1=1 x 1 = 1 , it follows that
M 1 ( p ) = P H h ∑ v w ∣ Q μ ( w ) ( v w ) 3 / 2 + O h , Q , η , ε ( p 5 / 4 + ε ∑ ℓ ≤ L ∣ x ℓ ∣ ℓ ) . M_1(p)=P H_h\sum_{vw\mid Q}\frac{\mu(w)}{(vw)^{3/2}}+O_{h,Q,\eta,\varepsilon}\left(p^{5/4+\varepsilon}\sum_{\ell\le L}\frac{|x_\ell|}{\sqrt{\ell}}\right). M 1 ( p ) = P H h v w ∣ Q ∑ ( v w ) 3/2 μ ( w ) + O h , Q , η , ε ( p 5/4 + ε ℓ ≤ L ∑ ℓ ∣ x ℓ ∣ ) .
Finally,
∑ v w ∣ Q μ ( w ) ( v w ) 3 / 2 = ∏ q ∣ Q ( 1 + q − 3 / 2 − q − 3 / 2 ) = 1 , \sum_{vw\mid Q}\frac{\mu(w)}{(vw)^{3/2}}=\prod_{q\mid Q}(1+q^{-3/2}-q^{-3/2})=1, v w ∣ Q ∑ ( v w ) 3/2 μ ( w ) = q ∣ Q ∏ ( 1 + q − 3/2 − q − 3/2 ) = 1 ,
and ∣ x ℓ ∣ ≤ 1 |x_\ell|\le1 ∣ x ℓ ∣ ≤ 1 gives ∑ ℓ < L ∣ x ℓ ∣ / ℓ ≪ L 1 / 2 \sum_{\ell<L}|x_\ell|/\sqrt{\ell}\ll L^{1/2} ∑ ℓ < L ∣ x ℓ ∣/ ℓ ≪ L 1/2 . This proves the first estimate in (88) . Since 5 / 4 + η / 2 < 3 / 2 5/4+\eta/2<3/2 5/4 + η /2 < 3/2 , the arbitrarily small ε \varepsilon ε gives the second estimate. □
Mollified harmonic second moment
Recall the mollified harmonic second moment M 2 ( p ) M_2(p) M 2 ( p ) defined in (4) .
Theorem 5.2. Fix a square-free integer Q ≥ 1 Q\ge1 Q ≥ 1 and 0 < η < 1 / 2 0<\eta<1/2 0 < η < 1/2 , and put L = p η L=p^\eta L = p η . As p → ∞ p\to\infty p → ∞ through primes p ∤ Q p\nmid Q p ∤ Q , we have
M 2 ( p ) = ζ ( 2 ) ∏ q ∣ Q ( 1 − q − 2 ) ( 1 + η − 1 ) P H h + O h , Q , η ( p 2 log ( 2 + log p ) log p ) . (90) M_2(p)=\zeta(2)\prod_{q\mid Q}(1-q^{-2})(1+\eta^{-1})P H_h+O_{h,Q,\eta}\left(\frac{p^2\log(2+\log p)}{\log p}\right). \tag*{(90)} M 2 ( p ) = ζ ( 2 ) q ∣ Q ∏ ( 1 − q − 2 ) ( 1 + η − 1 ) P H h + O h , Q , η ( log p p 2 log ( 2 + log p ) ) . (90)
Hecke expansion and the twisted second moment
The finite Euler factors in (2) have the exact Hecke expansion
∏ q ∣ Q ∣ 1 − A π ( 1 , q ) q + A π ( q , 1 ) q − 1 q 3 / 2 ∣ 2 = ∑ u v ∣ Q 2 c Q ( u , v ) A π ( u , v ) ‾ . (91) \prod_{q\mid Q}\left|1-\frac{A_\pi(1,q)}{\sqrt{q}}+\frac{A_\pi(q,1)}{q}-\frac{1}{q^{3/2}}\right|^2=\sum_{uv\mid Q^2}c_Q(u,v)\overline{A_\pi(u,v)}. \tag*{(91)} q ∣ Q ∏ 1 − q A π ( 1 , q ) + q A π ( q , 1 ) − q 3/2 1 2 = uv ∣ Q 2 ∑ c Q ( u , v ) A π ( u , v ) . (91)
Here c Q ( u , v ) = ∏ q ∣ Q c q ( ord q u , ord q v ) c_Q(u,v)=\prod_{q\mid Q}c_q(\operatorname{ord}_q u,\operatorname{ord}_q v) c Q ( u , v ) = ∏ q ∣ Q c q ( ord q u , ord q v ) for u v ∣ Q 2 uv\mid Q^2 uv ∣ Q 2 , where
c q ( i , j ) : = { ( 1 − q − 3 / 2 ) 2 + q − 1 + q − 2 , ( i , j ) = ( 0 , 0 ) , ( 1 − q − 3 / 2 ) ( q − 1 − q − 1 / 2 ) − q − 3 / 2 , ( i , j ) = ( 1 , 0 ) , ( 0 , 1 ) , q − 1 + q − 2 , ( i , j ) = ( 1 , 1 ) , − q − 3 / 2 , ( i , j ) = ( 2 , 0 ) , ( 0 , 2 ) , 0 , otherwise . c_q(i,j):=
\begin{cases}
(1-q^{-3/2})^2+q^{-1}+q^{-2}, & (i,j)=(0,0),\\
(1-q^{-3/2})(q^{-1}-q^{-1/2})-q^{-3/2}, & (i,j)=(1,0),(0,1),\\
q^{-1}+q^{-2}, & (i,j)=(1,1),\\
-q^{-3/2}, & (i,j)=(2,0),(0,2),\\
0, & \text{otherwise}.
\end{cases} c q ( i , j ) := ⎩ ⎨ ⎧ ( 1 − q − 3/2 ) 2 + q − 1 + q − 2 , ( 1 − q − 3/2 ) ( q − 1 − q − 1/2 ) − q − 3/2 , q − 1 + q − 2 , − q − 3/2 , 0 , ( i , j ) = ( 0 , 0 ) , ( i , j ) = ( 1 , 0 ) , ( 0 , 1 ) , ( i , j ) = ( 1 , 1 ) , ( i , j ) = ( 2 , 0 ) , ( 0 , 2 ) , otherwise .
These coefficients follow from the unramified Hecke relations A π ( 1 , q ) A π ( q , 1 ) = A π ( q , q ) + 1 A_\pi(1,q)A_\pi(q,1)=A_\pi(q,q)+1 A π ( 1 , q ) A π ( q , 1 ) = A π ( q , q ) + 1 and A π ( 1 , q ) 2 = A π ( 1 , q 2 ) + A π ( q , 1 ) A_\pi(1,q)^2=A_\pi(1,q^2)+A_\pi(q,1) A π ( 1 , q ) 2 = A π ( 1 , q 2 ) + A π ( q , 1 ) , together with their conjugates. In particular,
∑ u v ∣ Q 2 c Q ( u , v ) u v = ∏ q ∣ Q ( c q ( 0 , 0 ) + 2 c q ( 1 , 0 ) q + c q ( 1 , 1 ) q + 2 c q ( 2 , 0 ) q ) = ∏ q ∣ Q ( 1 − q − 1 ) . (92) \sum_{uv\mid Q^2}\frac{c_Q(u,v)}{\sqrt{uv}}=\prod_{q\mid Q}\left(c_q(0,0)+\frac{2c_q(1,0)}{\sqrt{q}}+\frac{c_q(1,1)}{q}+\frac{2c_q(2,0)}{q}\right)=\prod_{q\mid Q}(1-q^{-1}). \tag*{(92)} uv ∣ Q 2 ∑ uv c Q ( u , v ) = q ∣ Q ∏ ( c q ( 0 , 0 ) + q 2 c q ( 1 , 0 ) + q c q ( 1 , 1 ) + q 2 c q ( 2 , 0 ) ) = q ∣ Q ∏ ( 1 − q − 1 ) . (92)
Recall that the coefficients x ℓ x_\ell x ℓ are real and supported on ℓ ≤ L \ell\le L ℓ ≤ L with ( ℓ , Q ) = 1 (\ell,Q)=1 ( ℓ , Q ) = 1 . The unramified Hecke relation and (11) give
∣ ∑ ℓ ≤ L x ℓ A π ( 1 , ℓ ) ℓ ∣ 2 = ∑ d , ℓ 1 , ℓ 2 ≥ 1 x d ℓ 1 x d ℓ 2 d ℓ 1 ℓ 2 A π ( ℓ 1 , ℓ 2 ) ‾ . \left|\sum_{\ell\le L}\frac{x_\ell A_\pi(1,\ell)}{\sqrt{\ell}}\right|^2=\sum_{d,\ell_1,\ell_2\ge1}\frac{x_{d\ell_1}x_{d\ell_2}}{d\sqrt{\ell_1\ell_2}}\overline{A_\pi(\ell_1,\ell_2)}. ℓ ≤ L ∑ ℓ x ℓ A π ( 1 , ℓ ) 2 = d , ℓ 1 , ℓ 2 ≥ 1 ∑ d ℓ 1 ℓ 2 x d ℓ 1 x d ℓ 2 A π ( ℓ 1 , ℓ 2 ) .
Here d ℓ 1 , d ℓ 2 ≤ L d\ell_1,d\ell_2\le L d ℓ 1 , d ℓ 2 ≤ L are the original mollifier indices. Since ( ℓ 1 ℓ 2 , u v ) = 1 (\ell_1\ell_2,uv)=1 ( ℓ 1 ℓ 2 , uv ) = 1 in every non-zero term, multiplicativity and (91) yield
M 2 ( p ) = ∑ d , ℓ 1 , ℓ 2 ≥ 1 x d ℓ 1 x d ℓ 2 d ℓ 1 ℓ 2 ∑ u v ∣ Q 2 c Q ( u , v ) N 2 ( u ℓ 1 , v ℓ 2 , p ) . (93) \mathcal{M}_2(p)=\sum_{d,\ell_1,\ell_2\ge1}\frac{x_{d\ell_1}x_{d\ell_2}}{d\sqrt{\ell_1\ell_2}}\sum_{uv\mid Q^2}c_Q(u,v)\mathcal{N}_2(u\ell_1,v\ell_2,p). \tag*{(93)} M 2 ( p ) = d , ℓ 1 , ℓ 2 ≥ 1 ∑ d ℓ 1 ℓ 2 x d ℓ 1 x d ℓ 2 uv ∣ Q 2 ∑ c Q ( u , v ) N 2 ( u ℓ 1 , v ℓ 2 , p ) . (93)
Every non-zero term in (93) satisfies u v ℓ 1 ℓ 2 ≤ Q 2 L 2 = Q 2 p 2 η uv\ell_1\ell_2\le Q^2L^2=Q^2p^{2\eta} uv ℓ 1 ℓ 2 ≤ Q 2 L 2 = Q 2 p 2 η and ( u v ℓ 1 ℓ 2 , p ) = 1 (uv\ell_1\ell_2,p)=1 ( uv ℓ 1 ℓ 2 , p ) = 1 . Thus Theorem 4.9 applies uniformly, with the same fixed Q Q Q , η \eta η and h h h . Substituting (77) into (93) , we obtain
M 2 ( p ) = P ∑ d , ℓ 1 , ℓ 2 ≥ 1 x d ℓ 1 x d ℓ 2 d ℓ 1 ℓ 2 ∑ u v ∣ Q 2 c Q ( u , v ) u v × ( ( 1 − p − 1 ) { H h ( log p u v ℓ 1 ℓ 2 + 2 γ ) + K h } + 2 log p p H h ) + O h , Q , η , ε ( p 7 / 4 ∑ d , ℓ 1 , ℓ 2 ≥ 1 ∣ x d ℓ 1 x d ℓ 2 ∣ d ( ℓ 1 ℓ 2 ) 3 / 4 + p 3 / 2 + ε ∑ d , ℓ 1 , ℓ 2 ≥ 1 ∣ x d ℓ 1 x d ℓ 2 ∣ d ℓ 1 ℓ 2 ) . (94) \begin{aligned}
\mathcal{M}_2(p)={}&P\sum_{d,\ell_1,\ell_2\ge1}\frac{x_{d\ell_1}x_{d\ell_2}}{d\ell_1\ell_2}\sum_{uv\mid Q^2}\frac{c_Q(u,v)}{\sqrt{uv}}\\
&\times\left((1-p^{-1})\left\{\mathcal{H}_h\left(\log\frac{p}{uv\ell_1\ell_2}+2\gamma\right)+K_h\right\}+\frac{2\log p}{p}\mathcal{H}_h\right)\\
&+O_{h,Q,\eta,\varepsilon}\left(p^{7/4}\sum_{d,\ell_1,\ell_2\ge1}\frac{|x_{d\ell_1}x_{d\ell_2}|}{d(\ell_1\ell_2)^{3/4}}+p^{3/2+\varepsilon}\sum_{d,\ell_1,\ell_2\ge1}\frac{|x_{d\ell_1}x_{d\ell_2}|}{d\sqrt{\ell_1\ell_2}}\right). \tag*{(94)}
\end{aligned} M 2 ( p ) = P d , ℓ 1 , ℓ 2 ≥ 1 ∑ d ℓ 1 ℓ 2 x d ℓ 1 x d ℓ 2 uv ∣ Q 2 ∑ uv c Q ( u , v ) × ( ( 1 − p − 1 ) { H h ( log uv ℓ 1 ℓ 2 p + 2 γ ) + K h } + p 2 log p H h ) + O h , Q , η , ε p 7/4 d , ℓ 1 , ℓ 2 ≥ 1 ∑ d ( ℓ 1 ℓ 2 ) 3/4 ∣ x d ℓ 1 x d ℓ 2 ∣ + p 3/2 + ε d , ℓ 1 , ℓ 2 ≥ 1 ∑ d ℓ 1 ℓ 2 ∣ x d ℓ 1 x d ℓ 2 ∣ . (94)
In the error terms, the u , v u,v u , v -sums are absorbed into the implied constant, since Q Q Q is fixed. For 0 < β < 1 0<\beta<1 0 < β < 1 , the coefficients multiplying the twisted moments satisfy
∑ d , ℓ 1 , ℓ 2 ≥ 1 ∣ x d ℓ 1 x d ℓ 2 ∣ d ( ℓ 1 ℓ 2 ) β ≤ ∑ d ≤ L 1 d ( ∑ ℓ 1 ≤ L / d ℓ 1 − β ) 2 ≪ β L 2 ( 1 − β ) ∑ d ≥ 1 d − 3 + 2 β ≪ β L 2 ( 1 − β ) . (95) \sum_{d,\ell_1,\ell_2\ge1}\frac{|x_{d\ell_1}x_{d\ell_2}|}{d(\ell_1\ell_2)^\beta}\le\sum_{d\le L}\frac{1}{d}\left(\sum_{\ell_1\le L/d}\ell_1^{-\beta}\right)^2\ll_\beta L^{2(1-\beta)}\sum_{d\ge1}d^{-3+2\beta}\ll_\beta L^{2(1-\beta)}. \tag*{(95)} d , ℓ 1 , ℓ 2 ≥ 1 ∑ d ( ℓ 1 ℓ 2 ) β ∣ x d ℓ 1 x d ℓ 2 ∣ ≤ d ≤ L ∑ d 1 ℓ 1 ≤ L / d ∑ ℓ 1 − β 2 ≪ β L 2 ( 1 − β ) d ≥ 1 ∑ d − 3 + 2 β ≪ β L 2 ( 1 − β ) . (95)
Taking β = 3 / 4 \beta=3/4 β = 3/4 and 1 / 2 1/2 1/2 in (95) bounds the total error in (94) by
O h , Q , η , ε ( p 7 / 4 L 1 / 2 + p 3 / 2 + ε L ) . O_{h,Q,\eta,\varepsilon}\left(p^{7/4}L^{1/2}+p^{3/2+\varepsilon}L\right). O h , Q , η , ε ( p 7/4 L 1/2 + p 3/2 + ε L ) .
Arithmetic quadratic forms
It remains to evaluate the arithmetic main term, which is expressed through
Q 0 ( L ) : = ∑ d , ℓ 1 , ℓ 2 ≥ 1 x d ℓ 1 x d ℓ 2 d ℓ 1 ℓ 2 , Q_0(L):=\sum_{d,\ell_1,\ell_2\ge1}\frac{x_{d\ell_1}x_{d\ell_2}}{d\ell_1\ell_2}, Q 0 ( L ) := d , ℓ 1 , ℓ 2 ≥ 1 ∑ d ℓ 1 ℓ 2 x d ℓ 1 x d ℓ 2 ,
Q 1 ( L ) : = − ∑ d , ℓ 1 , ℓ 2 ≥ 1 x d ℓ 1 x d ℓ 2 log ( ℓ 1 ℓ 2 ) d ℓ 1 ℓ 2 . (96) Q_1(L):=-\sum_{d,\ell_1,\ell_2\ge1}\frac{x_{d\ell_1}x_{d\ell_2}\log(\ell_1\ell_2)}{d\ell_1\ell_2}. \tag*{(96)} Q 1 ( L ) := − d , ℓ 1 , ℓ 2 ≥ 1 ∑ d ℓ 1 ℓ 2 x d ℓ 1 x d ℓ 2 log ( ℓ 1 ℓ 2 ) . (96)
All non-zero terms have ( d ℓ 1 ℓ 2 , Q ) = 1 (d\ell_1\ell_2,Q)=1 ( d ℓ 1 ℓ 2 , Q ) = 1 . Using (92) and (96) , we therefore obtain
M 2 ( p ) = P ( 1 − p − 1 ) ∏ q ∣ Q ( 1 − q − 1 ) { H h ( ( log p + 2 γ ) Q 0 ( L ) + Q 1 ( L ) ) + K h Q 0 ( L ) } − P ( 1 − p − 1 ) H h Q 0 ( L ) ∑ u v ∣ Q 2 c Q ( u , v ) log ( u v ) u v + 2 P log p p H h Q 0 ( L ) ∏ q ∣ Q ( 1 − q − 1 ) + O h , Q , η , ε ( p 7 / 4 L 1 / 2 + p 3 / 2 + ε L ) . (97) \begin{aligned}
M_2(p)={}&P(1-p^{-1})\prod_{q\mid Q}(1-q^{-1})\{\mathcal{H}_h((\log p+2\gamma)Q_0(L)+Q_1(L))+\mathcal{K}_hQ_0(L)\}\\
&-P(1-p^{-1})\mathcal{H}_hQ_0(L)\sum_{uv\mid Q^2}\frac{c_Q(u,v)\log(uv)}{\sqrt{uv}}\\
&+\frac{2P\log p}{p}\mathcal{H}_hQ_0(L)\prod_{q\mid Q}(1-q^{-1})\\
&+O_{h,Q,\eta,\varepsilon}\left(p^{7/4}L^{1/2}+p^{3/2+\varepsilon}L\right). \tag*{(97)}
\end{aligned} M 2 ( p ) = P ( 1 − p − 1 ) q ∣ Q ∏ ( 1 − q − 1 ) { H h (( log p + 2 γ ) Q 0 ( L ) + Q 1 ( L )) + K h Q 0 ( L )} − P ( 1 − p − 1 ) H h Q 0 ( L ) uv ∣ Q 2 ∑ uv c Q ( u , v ) log ( uv ) + p 2 P log p H h Q 0 ( L ) q ∣ Q ∏ ( 1 − q − 1 ) + O h , Q , η , ε ( p 7/4 L 1/2 + p 3/2 + ε L ) . (97)
Lemma 5.3. Let x ℓ x_\ell x ℓ be in (1) and the quadratic forms in (96) . For fixed square-free Q Q Q and uniformly for L ≥ 2 L\ge2 L ≥ 2 , we have
Q 0 ( L ) = ζ ( 2 ) log L ∏ q ∣ Q ( 1 + q − 1 ) + O Q ( log ( 2 + log L ) ( log L ) 2 ) , (98) Q_0(L)=\frac{\zeta(2)}{\log L}\prod_{q\mid Q}(1+q^{-1})+O_Q\left(\frac{\log(2+\log L)}{(\log L)^2}\right), \tag*{(98)} Q 0 ( L ) = log L ζ ( 2 ) q ∣ Q ∏ ( 1 + q − 1 ) + O Q ( ( log L ) 2 log ( 2 + log L ) ) , (98)
Q 1 ( L ) = ζ ( 2 ) ∏ q ∣ Q ( 1 + q − 1 ) + O Q ( log ( 2 + log L ) log L ) . Q_1(L)=\zeta(2)\prod_{q\mid Q}(1+q^{-1})+O_Q\left(\frac{\log(2+\log L)}{\log L}\right). Q 1 ( L ) = ζ ( 2 ) q ∣ Q ∏ ( 1 + q − 1 ) + O Q ( log L log ( 2 + log L ) ) .
In particular, Q 0 ( L ) ≥ 0 Q_0(L)\ge0 Q 0 ( L ) ≥ 0 , Q 0 ( L ) ≪ Q 1 / log L Q_0(L)\ll_Q 1/\log L Q 0 ( L ) ≪ Q 1/ log L , and ∣ Q 1 ( L ) ∣ ≪ Q 1 |Q_1(L)|\ll_Q 1 ∣ Q 1 ( L ) ∣ ≪ Q 1 .
Proof. Let
F ( X ) : = ∑ n ≤ X μ ( n ) n log X n . F(X):=\sum_{n\le X}\frac{\mu(n)}{n}\log\frac{X}{n}. F ( X ) := n ≤ X ∑ n μ ( n ) log n X .
Perron’s formula [27 Proposition 5.54] expresses F ( X ) F(X) F ( X ) as the integral of X s / ( s 2 ζ ( 1 + s ) ) X^s/(s^2\zeta(1+s)) X s / ( s 2 ζ ( 1 + s )) . Shifting the contour using the zero-free region and the bound of 1 / ζ ( s ) 1/\zeta(s) 1/ ζ ( s ) in [27 Theorem 8.29] , and taking the residue at s = 0 s=0 s = 0 , gives, for some c > 0 c>0 c > 0 ,
F ( X ) = 1 + O ( e − c log X ) , F ( X ) ≪ 1 ( X ≥ 1 ) . (99) F(X)=1+O\left(e^{-c\sqrt{\log X}}\right),\qquad F(X)\ll1\quad(X\ge1). \tag*{(99)} F ( X ) = 1 + O ( e − c l o g X ) , F ( X ) ≪ 1 ( X ≥ 1 ) . (99)
For d ≥ 1 d\ge1 d ≥ 1 put
F d ( X ) : = ∑ ℓ 1 ≤ X ( ℓ 1 , d Q ) = 1 μ ( ℓ 1 ) ℓ 1 ∏ q ∣ ℓ 1 q q + 1 log X ℓ 1 , F_d(X):=\sum_{\substack{\ell_1\le X\\(\ell_1,dQ)=1}}\frac{\mu(\ell_1)}{\ell_1}\prod_{q\mid\ell_1}\frac{q}{q+1}\log\frac{X}{\ell_1}, F d ( X ) := ℓ 1 ≤ X ( ℓ 1 , d Q ) = 1 ∑ ℓ 1 μ ( ℓ 1 ) q ∣ ℓ 1 ∏ q + 1 q log ℓ 1 X ,
T d ( X ) : = ∑ ℓ 1 ≤ X ( ℓ 1 , d Q ) = 1 μ ( ℓ 1 ) log ℓ 1 ℓ 1 ∏ q ∣ ℓ 1 q q + 1 log X ℓ 1 . (100) T_d(X):=\sum_{\substack{\ell_1\le X\\(\ell_1,dQ)=1}}\frac{\mu(\ell_1)\log\ell_1}{\ell_1}\prod_{q\mid\ell_1}\frac{q}{q+1}\log\frac{X}{\ell_1}. \tag*{(100)} T d ( X ) := ℓ 1 ≤ X ( ℓ 1 , d Q ) = 1 ∑ ℓ 1 μ ( ℓ 1 ) log ℓ 1 q ∣ ℓ 1 ∏ q + 1 q log ℓ 1 X . (100)
Define the non-negative multiplicative function k d k_d k d by k d ( 1 ) = 1 k_d(1)=1 k d ( 1 ) = 1 and, for primes q q q and j ≥ 1 j\ge1 j ≥ 1 ,
k d ( q j ) : = { 1 , q ∣ d Q , ( q + 1 ) − 1 , q ∤ d Q . k_d(q^j):=
\begin{cases}
1, & q\mid dQ,\\
(q+1)^{-1}, & q\nmid dQ.
\end{cases} k d ( q j ) := { 1 , ( q + 1 ) − 1 , q ∣ d Q , q ∤ d Q .
By the Euler product, a direct calculation gives
F d ( X ) = ∑ e ≤ X k d ( e ) e F ( X / e ) . (101) F_d(X)=\sum_{e\le X}\frac{k_d(e)}{e}F(X/e). \tag*{(101)} F d ( X ) = e ≤ X ∑ e k d ( e ) F ( X / e ) . (101)
and
∑ e ≥ 1 k d ( e ) e = ζ ( 2 ) ∏ q ∣ d Q ( 1 + q − 1 ) . (102) \sum_{e\ge1}\frac{k_d(e)}{e}=\zeta(2)\prod_{q\mid dQ}(1+q^{-1}). \tag*{(102)} e ≥ 1 ∑ e k d ( e ) = ζ ( 2 ) q ∣ d Q ∏ ( 1 + q − 1 ) . (102)
The Euler product for ∑ e ≥ 1 k d ( e ) e − s \sum_{e\ge1} k_d(e)e^{-s} ∑ e ≥ 1 k d ( e ) e − s converges absolutely for Re ( s ) > 0 \operatorname{Re}(s)>0 Re ( s ) > 0 . Logarithmic differentiation at s = 1 s=1 s = 1 yields
∑ e ≥ 1 k d ( e ) log e e = ζ ( 2 ) ∏ q ∣ d Q ( 1 + q − 1 ) ( ∑ q ∣ d Q log q q − 1 + ∑ q ∤ d Q log q q ( q − 1 ) ) ≪ ∏ q ∣ d Q ( 1 + q − 1 ) ( 1 + ∑ q ∣ d Q log q q − 1 ) . (103) \begin{aligned}
\sum_{e\ge1}\frac{k_d(e)\log e}{e}=\zeta(2)\prod_{q\mid dQ}(1+q^{-1})\left(\sum_{q\mid dQ}\frac{\log q}{q-1}+\sum_{q\nmid dQ}\frac{\log q}{q(q-1)}\right) \\
&\ll\prod_{q\mid dQ}(1+q^{-1})\left(1+\sum_{q\mid dQ}\frac{\log q}{q-1}\right). \tag*{(103)}
\end{aligned} e ≥ 1 ∑ e k d ( e ) log e = ζ ( 2 ) q ∣ d Q ∏ ( 1 + q − 1 ) q ∣ d Q ∑ q − 1 log q + q ∤ d Q ∑ q ( q − 1 ) log q ≪ q ∣ d Q ∏ ( 1 + q − 1 ) 1 + q ∣ d Q ∑ q − 1 log q . (103)
Since k d ( e ) ≥ 0 k_d(e)\ge0 k d ( e ) ≥ 0 , (99) –(101) also imply F d ( X ) ≪ ∏ q ∣ d Q ( 1 + q − 1 ) F_d(X)\ll\prod_{q\mid dQ}(1+q^{-1}) F d ( X ) ≪ ∏ q ∣ d Q ( 1 + q − 1 ) .
For X ≥ 2 X\ge2 X ≥ 2 , we subtract the sum in (102) from (101) , split at e = X e=\sqrt{X} e = X , and obtain
F d ( X ) − ∑ e ≥ 1 k d ( e ) e = ∑ e ≤ X k d ( e ) e ( F ( X / e ) − 1 ) + ∑ X < e ≤ X k d ( e ) e F ( X / e ) − ∑ e > X k d ( e ) e . F_d(X)-\sum_{e\ge1}\frac{k_d(e)}{e}=\sum_{e\le\sqrt{X}}\frac{k_d(e)}{e}\left(F(X/e)-1\right)+\sum_{\sqrt{X}<e\le X}\frac{k_d(e)}{e}F(X/e)-\sum_{e>\sqrt{X}}\frac{k_d(e)}{e}. F d ( X ) − e ≥ 1 ∑ e k d ( e ) = e ≤ X ∑ e k d ( e ) ( F ( X / e ) − 1 ) + X < e ≤ X ∑ e k d ( e ) F ( X / e ) − e > X ∑ e k d ( e ) .
By (99) , ∣ F ( X / e ) − 1 ∣ ≪ exp ( − c ( log X ) / 2 ) \lvert F(X/e)-1\rvert\ll\exp(-c\sqrt{(\log X)/2}) ∣ F ( X / e ) − 1 ∣ ≪ exp ( − c ( log X ) /2 ) for e ≤ X e\le\sqrt{X} e ≤ X and ∣ F ( X / e ) ∣ ≪ 1 \lvert F(X/e)\rvert\ll1 ∣ F ( X / e )∣ ≪ 1 for X < e ≤ X \sqrt{X}<e\le X X < e ≤ X . Since k d ( e ) ≥ 0 k_d(e)\ge0 k d ( e ) ≥ 0 , it follows that
∣ F d ( X ) − ∑ e ≥ 1 k d ( e ) e ∣ ≪ exp ( − c ( log X ) / 2 ) ∑ e ≥ 1 k d ( e ) e + 2 log X ∑ e ≥ 1 k d ( e ) log e e . \left\lvert F_d(X)-\sum_{e\ge1}\frac{k_d(e)}{e}\right\rvert\ll\exp(-c\sqrt{(\log X)/2})\sum_{e\ge1}\frac{k_d(e)}{e}+\frac{2}{\log X}\sum_{e\ge1}\frac{k_d(e)\log e}{e}. F d ( X ) − e ≥ 1 ∑ e k d ( e ) ≪ exp ( − c ( log X ) /2 ) e ≥ 1 ∑ e k d ( e ) + log X 2 e ≥ 1 ∑ e k d ( e ) log e .
Inserting (102) and (103) , and using exp ( − c ( log X ) / 2 ) ≪ ( 1 + log X ) − 1 \exp(-c\sqrt{(\log X)/2})\ll(1+\log X)^{-1} exp ( − c ( log X ) /2 ) ≪ ( 1 + log X ) − 1 , we obtain
F d ( X ) = ζ ( 2 ) ∏ q ∣ d Q ( 1 + q − 1 ) + O ( ∏ q ∣ d Q ( 1 + q − 1 ) 1 + log X ( 1 + ∑ q ∣ d Q log q q − 1 ) ) . (104) F_d(X)=\zeta(2)\prod_{q\mid dQ}(1+q^{-1})+O\left(\frac{\prod_{q\mid dQ}(1+q^{-1})}{1+\log X}\left(1+\sum_{q\mid dQ}\frac{\log q}{q-1}\right)\right). \tag*{(104)} F d ( X ) = ζ ( 2 ) q ∣ d Q ∏ ( 1 + q − 1 ) + O 1 + log X ∏ q ∣ d Q ( 1 + q − 1 ) 1 + q ∣ d Q ∑ q − 1 log q . (104)
For 1 ≤ X < 2 1\le X<2 1 ≤ X < 2 , we have F d ( X ) = log X F_d(X)=\log X F d ( X ) = log X , so (104) holds for all X ≥ 1 X\ge1 X ≥ 1 . By the definitions of F d F_d F d and T d T_d T d in (100) , and then substituting (104) for F d F_d F d , we have
T d ( X ) = ( log X ) F d ( X ) − 2 ∫ 1 X F d ( t ) d t t = − ζ ( 2 ) ( log X ) ∏ q ∣ d Q ( 1 + q − 1 ) + O ( log ( 2 + log X ) ∏ q ∣ d Q ( 1 + q − 1 ) ( 1 + ∑ q ∣ d Q log q q − 1 ) ) . (105) \begin{aligned}
T_d(X)=(\log X)F_d(X)-2\int_1^X F_d(t)\frac{\mathrm{d}t}{t} \\
&=-\zeta(2)(\log X)\prod_{q\mid dQ}(1+q^{-1})+O\left(\log(2+\log X)\prod_{q\mid dQ}(1+q^{-1})\left(1+\sum_{q\mid dQ}\frac{\log q}{q-1}\right)\right). \tag*{(105)}
\end{aligned} T d ( X ) = ( log X ) F d ( X ) − 2 ∫ 1 X F d ( t ) t d t = − ζ ( 2 ) ( log X ) q ∣ d Q ∏ ( 1 + q − 1 ) + O log ( 2 + log X ) q ∣ d Q ∏ ( 1 + q − 1 ) 1 + q ∣ d Q ∑ q − 1 log q . (105)
Since x ℓ x_\ell x ℓ is supported on square-free integers, x d ℓ 1 x d ℓ 2 = 0 x_{d\ell_1}x_{d\ell_2}=0 x d ℓ 1 x d ℓ 2 = 0 unless ( d , ℓ 1 ℓ 2 ) = 1 (d,\ell_1\ell_2)=1 ( d , ℓ 1 ℓ 2 ) = 1 . Substituting (1) into (96) and using (100) , we obtain
Q 0 ( L ) = 1 ( log L ) 2 ∑ d ≤ L ( d , Q ) = 1 μ ( d ) 2 d ∏ q ∣ d ( q q + 1 ) 2 F d ( L / d ) 2 , Q_0(L)=\frac{1}{(\log L)^2}\sum_{\substack{d\le L\\(d,Q)=1}}\frac{\mu(d)^2}{d}\prod_{q\mid d}\left(\frac{q}{q+1}\right)^2F_d(L/d)^2, Q 0 ( L ) = ( log L ) 2 1 d ≤ L ( d , Q ) = 1 ∑ d μ ( d ) 2 q ∣ d ∏ ( q + 1 q ) 2 F d ( L / d ) 2 ,
Q 1 ( L ) = − 2 ( log L ) 2 ∑ d ≤ L ( d , Q ) = 1 μ ( d ) 2 d ∏ q ∣ d ( q q + 1 ) 2 F d ( L / d ) T d ( L / d ) . (106) Q_1(L)=-\frac{2}{(\log L)^2}\sum_{\substack{d\le L\\(d,Q)=1}}\frac{\mu(d)^2}{d}\prod_{q\mid d}\left(\frac{q}{q+1}\right)^2F_d(L/d)T_d(L/d). \tag*{(106)} Q 1 ( L ) = − ( log L ) 2 2 d ≤ L ( d , Q ) = 1 ∑ d μ ( d ) 2 q ∣ d ∏ ( q + 1 q ) 2 F d ( L / d ) T d ( L / d ) . (106)
The factor 2 2 2 in the second identity follows from log ( ℓ 1 ℓ 2 ) = log ℓ 1 + log ℓ 2 \log(\ell_1\ell_2)=\log\ell_1+\log\ell_2 log ( ℓ 1 ℓ 2 ) = log ℓ 1 + log ℓ 2 and symmetry in ℓ 1 , ℓ 2 \ell_1,\ell_2 ℓ 1 , ℓ 2 . The first identity also shows that Q 0 ( L ) ≥ 0 Q_0(L)\ge0 Q 0 ( L ) ≥ 0 .
It remains to evaluate the main terms and bound the errors in (106) . For $(d,Q)=1,
∏ q ∣ d Q ( 1 + q − 1 ) ∏ q ∣ d q q + 1 = ∏ q ∣ Q ( 1 + q − 1 ) . \prod_{q\mid dQ}(1+q^{-1})\prod_{q\mid d}\frac{q}{q+1}=\prod_{q\mid Q}(1+q^{-1}). q ∣ d Q ∏ ( 1 + q − 1 ) q ∣ d ∏ q + 1 q = q ∣ Q ∏ ( 1 + q − 1 ) .
Substituting (104) and (105) into (106) , and using ∣ F d ( X ) ∣ ≪ ∏ q ∣ d Q ( 1 + q − 1 ) |F_d(X)|\ll\prod_{q\mid dQ}(1+q^{-1}) ∣ F d ( X ) ∣ ≪ ∏ q ∣ d Q ( 1 + q − 1 ) , gives
Q 0 ( L ) = ζ ( 2 ) 2 ( log L ) 2 ∏ q ∣ Q ( 1 + q − 1 ) 2 ∑ d ≤ L ( d , Q ) = 1 μ ( d ) 2 d + O Q ( 1 ( log L ) 2 ∑ d ≤ L 1 + ∑ q ∣ d q log q q − 1 d ( 1 + log ( L / d ) ) ) , (107) Q_0(L)=\frac{\zeta(2)^2}{(\log L)^2}\prod_{q\mid Q}(1+q^{-1})^2\sum_{\substack{d\le L\\(d,Q)=1}}\frac{\mu(d)^2}{d}
+O_Q\left(\frac{1}{(\log L)^2}\sum_{d\le L}\frac{1+\sum_{q\mid dq}\frac{\log q}{q-1}}{d(1+\log(L/d))}\right), \tag*{(107)} Q 0 ( L ) = ( log L ) 2 ζ ( 2 ) 2 q ∣ Q ∏ ( 1 + q − 1 ) 2 d ≤ L ( d , Q ) = 1 ∑ d μ ( d ) 2 + O Q ( ( log L ) 2 1 d ≤ L ∑ d ( 1 + log ( L / d )) 1 + ∑ q ∣ d q q − 1 l o g q ) , (107)
and
Q 1 ( L ) = 2 ζ ( 2 ) 2 ( log L ) 2 ∏ q ∣ Q ( 1 + q − 1 ) 2 ∑ d ≤ L ( d , Q ) = 1 μ ( d ) 2 d log ( L / d ) Q_1(L)=\frac{2\zeta(2)^2}{(\log L)^2}\prod_{q\mid Q}(1+q^{-1})^2\sum_{\substack{d\le L\\(d,Q)=1}}\frac{\mu(d)^2}{d}\log(L/d) Q 1 ( L ) = ( log L ) 2 2 ζ ( 2 ) 2 q ∣ Q ∏ ( 1 + q − 1 ) 2 d ≤ L ( d , Q ) = 1 ∑ d μ ( d ) 2 log ( L / d )
+ O Q ( 1 ( log L ) 2 ∑ d ≤ L 1 + ∑ q ∣ d log q q − 1 d log ( 2 + log ( L / d ) ) ) . (108) \qquad+O_Q\left(\frac{1}{(\log L)^2}\sum_{d\le L}\frac{1+\sum_{q\mid d}\frac{\log q}{q-1}}{d}\log(2+\log(L/d))\right). \tag*{(108)} + O Q ( ( log L ) 2 1 d ≤ L ∑ d 1 + ∑ q ∣ d q − 1 l o g q log ( 2 + log ( L / d )) ) . (108)
In these error terms, we used μ ( d ) 2 ≤ 1 \mu(d)^2\le1 μ ( d ) 2 ≤ 1 and
1 + ∑ q ∣ d Q log q q − 1 ≪ Q 1 + ∑ q ∣ d log q q − 1 . 1+\sum_{q\mid dQ}\frac{\log q}{q-1}\ll_Q1+\sum_{q\mid d}\frac{\log q}{q-1}. 1 + q ∣ d Q ∑ q − 1 log q ≪ Q 1 + q ∣ d ∑ q − 1 log q .
We first evaluate the sums in the main terms of (107) and (108) . Using 1 ( m , Q ) = 1 = ∑ e ∣ ( m , Q ) μ ( e ) 1_{(m,Q)=1}=\sum_{e\mid(m,Q)}\mu(e) 1 ( m , Q ) = 1 = ∑ e ∣ ( m , Q ) μ ( e ) , we have, for y ≥ 1 y\ge1 y ≥ 1 ,
∑ m ≤ y ( m , Q ) = 1 1 m = ∑ e ∣ Q μ ( e ) e ∑ n ≤ y / e 1 n = ∏ q ∣ Q ( 1 − q − 1 ) log y + O Q ( 1 ) . \sum_{\substack{m\le y\\(m,Q)=1}}\frac{1}{m}
=\sum_{e\mid Q}\frac{\mu(e)}{e}\sum_{n\le y/e}\frac{1}{n}
=\prod_{q\mid Q}(1-q^{-1})\log y+O_Q(1). m ≤ y ( m , Q ) = 1 ∑ m 1 = e ∣ Q ∑ e μ ( e ) n ≤ y / e ∑ n 1 = q ∣ Q ∏ ( 1 − q − 1 ) log y + O Q ( 1 ) .
Since μ ( d ) 2 = ∑ e 2 ∣ d μ ( e ) \mu(d)^2=\sum_{e^2\mid d}\mu(e) μ ( d ) 2 = ∑ e 2 ∣ d μ ( e ) , the above gives
∑ d ≤ x ( d , Q ) = 1 μ ( d ) 2 d = ∑ e ≤ x ( e , Q ) = 1 μ ( e ) e 2 ∑ m ≤ x / e 2 ( m , Q ) = 1 1 m = log x ζ ( 2 ) ∏ q ∣ Q ( 1 + q − 1 ) − 1 + O Q ( 1 ) . (109) \sum_{\substack{d\le x\\(d,Q)=1}}\frac{\mu(d)^2}{d}
=\sum_{\substack{e\le\sqrt{x}\\(e,Q)=1}}\frac{\mu(e)}{e^2}
\sum_{\substack{m\le x/e^2\\(m,Q)=1}}\frac{1}{m}
=\frac{\log x}{\zeta(2)}\prod_{q\mid Q}(1+q^{-1})^{-1}+O_Q(1). \tag*{(109)} d ≤ x ( d , Q ) = 1 ∑ d μ ( d ) 2 = e ≤ x ( e , Q ) = 1 ∑ e 2 μ ( e ) m ≤ x / e 2 ( m , Q ) = 1 ∑ m 1 = ζ ( 2 ) log x q ∣ Q ∏ ( 1 + q − 1 ) − 1 + O Q ( 1 ) . (109)
This evaluates the sum in the main term of Q 0 ( L ) Q_0(L) Q 0 ( L ) . For Q 1 ( L ) Q_1(L) Q 1 ( L ) , the additional factor log ( L / d ) \log(L/d) log ( L / d ) is handled by log ( L / d ) = ∫ d L d t / t \log(L/d)=\int_d^L dt/t log ( L / d ) = ∫ d L d t / t . Thus
∑ d ≤ L ( d , Q ) = 1 μ ( d ) 2 d log ( L / d ) = ∫ 1 L ( ∑ d ≤ t ( d , Q ) = 1 μ ( d ) 2 d ) d t t = ( log L ) 2 2 ζ ( 2 ) ∏ q ∣ Q ( 1 + q − 1 ) − 1 + O Q ( log L ) . (110) \sum_{\substack{d\le L\\(d,Q)=1}}\frac{\mu(d)^2}{d}\log(L/d)=\int_1^L\left(\sum_{\substack{d\le t\\(d,Q)=1}}\frac{\mu(d)^2}{d}\right)\frac{dt}{t}=\frac{(\log L)^2}{2\zeta(2)}\prod_{q\mid Q}(1+q^{-1})^{-1}+O_Q(\log L). \tag*{(110)} d ≤ L ( d , Q ) = 1 ∑ d μ ( d ) 2 log ( L / d ) = ∫ 1 L d ≤ t ( d , Q ) = 1 ∑ d μ ( d ) 2 t d t = 2 ζ ( 2 ) ( log L ) 2 q ∣ Q ∏ ( 1 + q − 1 ) − 1 + O Q ( log L ) . (110)
Substituting (109) and (110) into the main terms of (107) and (108) gives ζ ( 2 ) ∏ q ∣ Q ( 1 + q − 1 ) / log L \zeta(2)\prod_{q\mid Q}(1+q^{-1})/\log L ζ ( 2 ) ∏ q ∣ Q ( 1 + q − 1 ) / log L for Q 0 ( L ) Q_0(L) Q 0 ( L ) and ζ ( 2 ) ∏ q ∣ Q ( 1 + q − 1 ) \zeta(2)\prod_{q\mid Q}(1+q^{-1}) ζ ( 2 ) ∏ q ∣ Q ( 1 + q − 1 ) for Q 1 ( L ) Q_1(L) Q 1 ( L ) , with respective errors O Q ( ( log L ) − 2 ) O_Q((\log L)^{-2}) O Q (( log L ) − 2 ) and O Q ( ( log L ) − 1 ) O_Q((\log L)^{-1}) O Q (( log L ) − 1 ) .
We now bound the two explicit error sums in (107) and (108) . To control the factor 1 + ∑ q ∣ d ( log q ) / ( q − 1 ) 1+\sum_{q\mid d}(\log q)/(q-1) 1 + ∑ q ∣ d ( log q ) / ( q − 1 ) in both errors, we first sum it over d ≤ y d\le y d ≤ y , where y ≥ 1 y\ge1 y ≥ 1 :
∑ d ≤ y ( 1 + ∑ q ∣ d log q q − 1 ) = ⌊ y ⌋ + ∑ q ≤ y log q q − 1 ⌊ y q ⌋ ≤ y ( 1 + ∑ q ≤ y log q q ( q − 1 ) ) ≪ y . \sum_{d\le y}\left(1+\sum_{q\mid d}\frac{\log q}{q-1}\right)=\lfloor y\rfloor+\sum_{q\le y}\frac{\log q}{q-1}\left\lfloor\frac{y}{q}\right\rfloor\le y\left(1+\sum_{q\le y}\frac{\log q}{q(q-1)}\right)\ll y. d ≤ y ∑ 1 + q ∣ d ∑ q − 1 log q = ⌊ y ⌋ + q ≤ y ∑ q − 1 log q ⌊ q y ⌋ ≤ y ( 1 + q ≤ y ∑ q ( q − 1 ) log q ) ≪ y .
For the error in Q 0 ( L ) Q_0(L) Q 0 ( L ) , we apply Abel’s summation formula and use
∫ 1 L d t t ( 1 + log ( L / t ) ) = log ( 1 + log L ) . \int_1^L\frac{dt}{t(1+\log(L/t))}=\log(1+\log L). ∫ 1 L t ( 1 + log ( L / t )) d t = log ( 1 + log L ) .
For the error in Q 1 ( L ) Q_1(L) Q 1 ( L ) , we use log ( 2 + log ( L / d ) ) ≤ log ( 2 + log L ) \log(2+\log(L/d))\le\log(2+\log L) log ( 2 + log ( L / d )) ≤ log ( 2 + log L ) and
∑ d ≤ L 1 d ( 1 + ∑ q ∣ d log q q − 1 ) ≪ 1 + ∫ 1 L d t t ≪ log L . \sum_{d\le L}\frac{1}{d}\left(1+\sum_{q\mid d}\frac{\log q}{q-1}\right)\ll1+\int_1^L\frac{dt}{t}\ll\log L. d ≤ L ∑ d 1 1 + q ∣ d ∑ q − 1 log q ≪ 1 + ∫ 1 L t d t ≪ log L .
Consequently, the errors in Q 0 ( L ) Q_0(L) Q 0 ( L ) and Q 1 ( L ) Q_1(L) Q 1 ( L ) satisfy, respectively,
1 ( log L ) 2 ∑ d ≤ L 1 + ∑ q ∣ d log q q − 1 d ( 1 + log ( L / d ) ) ≪ log ( 2 + log L ) ( log L ) 2 , \frac{1}{(\log L)^2}\sum_{d\le L}\frac{1+\sum_{q\mid d}\frac{\log q}{q-1}}{d(1+\log(L/d))}\ll\frac{\log(2+\log L)}{(\log L)^2}, ( log L ) 2 1 d ≤ L ∑ d ( 1 + log ( L / d )) 1 + ∑ q ∣ d q − 1 l o g q ≪ ( log L ) 2 log ( 2 + log L ) ,
and
1 ( log L ) 2 ∑ d ≤ L 1 + ∑ q ∣ d log q q − 1 d log ( 2 + log ( L / d ) ) ≪ log ( 2 + log L ) log L . \frac{1}{(\log L)^2}\sum_{d\le L}\frac{1+\sum_{q\mid d}\frac{\log q}{q-1}}{d}\log(2+\log(L/d))\ll\frac{\log(2+\log L)}{\log L}. ( log L ) 2 1 d ≤ L ∑ d 1 + ∑ q ∣ d q − 1 l o g q log ( 2 + log ( L / d )) ≪ log L log ( 2 + log L ) .
Combining these bounds with the main terms above in (107) and (108) proves (98) .
For any real coefficients supported on ( ℓ , Q ) = 1 (\ell,Q)=1 ( ℓ , Q ) = 1 and ℓ ≤ L \ell\le L ℓ ≤ L , the definition gives
Q 0 ( L ) = ∑ d ≤ L ( d , Q ) = 1 d ( ∑ n ≤ L d ∣ n x n n ) 2 . Q_0(L)=\sum_{\substack{d\le L\\(d,Q)=1}}d\left(\sum_{\substack{n\le L\\d\mid n}}\frac{x_n}{n}\right)^2. Q 0 ( L ) = d ≤ L ( d , Q ) = 1 ∑ d n ≤ L d ∣ n ∑ n x n 2 .
If x 1 = 1 x_1=1 x 1 = 1 , the Möbius inversion formula and the Cauchy–Schwarz inequality imply
1 = ( ∑ d ≤ L ( d , Q ) = 1 μ ( d ) ∑ n ≤ L d ∣ n x n n ) 2 ≤ ( ∑ d ≤ L ( d , Q ) = 1 μ ( d ) 2 d ) Q 0 ( L ) . 1=\left(\sum_{\substack{d\le L\\(d,Q)=1}}\mu(d)\sum_{\substack{n\le L\\d\mid n}}\frac{x_n}{n}\right)^2\le\left(\sum_{\substack{d\le L\\(d,Q)=1}}\frac{\mu(d)^2}{d}\right)Q_0(L). 1 = d ≤ L ( d , Q ) = 1 ∑ μ ( d ) n ≤ L d ∣ n ∑ n x n 2 ≤ d ≤ L ( d , Q ) = 1 ∑ d μ ( d ) 2 Q 0 ( L ) .
For fixed Q Q Q , (109) therefore gives
Q 0 ( L ) ≥ ( ∑ d ≤ L ( d , Q ) = 1 μ ( d ) 2 d ) − 1 = ζ ( 2 ) log L ∏ q ∣ Q ( 1 + q − 1 ) ( 1 + O Q ( 1 log L ) ) ( L → ∞ ) . Q_0(L) \ge\left(\sum_{\substack{d\le L\\(d,Q)=1}} \frac{\mu(d)^2}{d}\right)^{-1}
= \frac{\zeta(2)}{\log L}\prod_{q\mid Q}(1+q^{-1})\left(1+O_Q\left(\frac{1}{\log L}\right)\right)
\qquad(L\to\infty). Q 0 ( L ) ≥ d ≤ L ( d , Q ) = 1 ∑ d μ ( d ) 2 − 1 = log L ζ ( 2 ) q ∣ Q ∏ ( 1 + q − 1 ) ( 1 + O Q ( log L 1 ) ) ( L → ∞ ) .
By (98) , the coefficients in (1) attain this lower bound asymptotically. Thus they are asymptotically optimal for Q 0 ( L ) Q_0(L) Q 0 ( L ) among all real coefficients with x 1 = 1 x_1=1 x 1 = 1 supported on ℓ ≤ L \ell\le L ℓ ≤ L and ( ℓ , Q ) = 1 (\ell,Q)=1 ( ℓ , Q ) = 1 .
Proof of Theorem 5.2
Since log L = η log p \log L=\eta\log p log L = η log p , Lemma 5.3 gives
∏ q ∣ Q ( 1 − q − 1 ) ( ( log p + 2 γ ) Q 0 ( L ) + Q 1 ( L ) ) = ζ ( 2 ) ∏ q ∣ Q ( 1 − q − 2 ) ( 1 + η − 1 ) + O Q , η ( log ( 2 + log p ) log p ) . \prod_{q\mid Q}(1-q^{-1})\bigl((\log p+2\gamma)Q_0(L)+Q_1(L)\bigr)
= \zeta(2)\prod_{q\mid Q}(1-q^{-2})(1+\eta^{-1})+O_{Q,\eta}\left(\frac{\log(2+\log p)}{\log p}\right). q ∣ Q ∏ ( 1 − q − 1 ) ( ( log p + 2 γ ) Q 0 ( L ) + Q 1 ( L ) ) = ζ ( 2 ) q ∣ Q ∏ ( 1 − q − 2 ) ( 1 + η − 1 ) + O Q , η ( log p log ( 2 + log p ) ) .
Since Q Q Q is fixed and Q 0 ( L ) ≪ Q , η 1 / log p Q_0(L)\ll_{Q,\eta}1/\log p Q 0 ( L ) ≪ Q , η 1/ log p , the terms involving log ( u v ) \log(uv) log ( uv ) and K h \mathcal{K}_h K h in (97) are O h , Q , η ( p 2 / log p ) O_{h,Q,\eta}(p^2/\log p) O h , Q , η ( p 2 / log p ) . Replacing 1 − p − 1 1-p^{-1} 1 − p − 1 by 1 1 1 introduces an error O h , Q , η ( p ) O_{h,Q,\eta}(p) O h , Q , η ( p ) , and the term containing 2 log p / p 2\log p/p 2 log p / p satisfies the same bound. Moreover, L = p η L=p^\eta L = p η and η < 1 / 2 \eta<1/2 η < 1/2 give p 7 / 4 L 1 / 2 + p 3 / 2 + ϵ L ≪ η p 2 / log p p^{7/4}L^{1/2}+p^{3/2+\epsilon}L\ll_\eta p^2/\log p p 7/4 L 1/2 + p 3/2 + ϵ L ≪ η p 2 / log p , since ϵ \epsilon ϵ is arbitrarily small. Substituting these estimates into (97) proves (90) .
The holomorphic mixed moment
Recall the holomorphic mixed moment M 1 , 1 ( p ) \mathcal{M}_{1,1}(p) M 1 , 1 ( p ) defined in (6) . We use the same h h h , Q Q Q , η \eta η , L L L and the mollifier M π M_\pi M π defined above.
Theorem 5.4. For every fixed square-free Q Q Q and 0 < η < 1 / 2 0<\eta<1/2 0 < η < 1/2 , as p → ∞ p\to\infty p → ∞ through primes p ∤ Q p\nmid Q p ∤ Q , we have
M 1 , 1 ( p ) = P H h ∏ q ∣ Q ( 1 − 2 q ( q + 1 ) 2 + 1 q ( q + 1 ) 2 ) + O h , Q , η ( p 2 log p ) . (111) \mathcal{M}_{1,1}(p)=P\!H_h\prod_{q\mid Q}\left(1-\frac{2}{\sqrt{q}(q+1)^2}+\frac{1}{q(q+1)^2}\right)+O_{h,Q,\eta}\left(\frac{p^2}{\log p}\right). \tag*{(111)} M 1 , 1 ( p ) = P H h q ∣ Q ∏ ( 1 − q ( q + 1 ) 2 2 + q ( q + 1 ) 2 1 ) + O h , Q , η ( log p p 2 ) . (111) The Euler product converges absolutely, lies in ( 0 , 1 ] (0,1] ( 0 , 1 ] , and tends to 1 1 1 as Q Q Q ranges over products of all primes up to a bound tending to infinity.
Application of the Kuznetsov formula
By (33) , the moment takes the form
M 1 , 1 ( p ) = c c u s p ∑ π ∈ B ( p ) M π 2 L ( 1 , π , Ad ) ∑ m , n ≥ 1 A π ( 1 , m ) A π ( 1 , n ) m n K μ π + ( m n p L 2 ) + c c u s p ∑ π ∈ B ( p ) ϵ ( π ) 2 M π 2 L ( 1 , π , Ad ) ∑ m , n ≥ 1 A π ( 1 , m ) A π ( 1 , n ) ‾ m n K μ π − ( m n L 2 p ) . (112) \begin{aligned}
\mathcal{M}_{1,1}(p)
={}&c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}
\frac{M_\pi^2}{L(1,\pi,\operatorname{Ad})}
\sum_{m,n\ge1}\frac{A_\pi(1,m)A_\pi(1,n)}{\sqrt{mn}}
K_{\mu_\pi^+}\left(\frac{mn}{pL^2}\right)\\
&+c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}
\frac{\epsilon(\pi)^2M_\pi^2}{L(1,\pi,\operatorname{Ad})}
\sum_{m,n\ge1}\frac{\overline{A_\pi(1,m)A_\pi(1,n)}}{\sqrt{mn}}
K_{\mu_\pi^-}\left(\frac{mnL^2}{p}\right). \tag*{(112)}
\end{aligned} M 1 , 1 ( p ) = c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) M π 2 m , n ≥ 1 ∑ mn A π ( 1 , m ) A π ( 1 , n ) K μ π + ( p L 2 mn ) + c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) ϵ ( π ) 2 M π 2 m , n ≥ 1 ∑ mn A π ( 1 , m ) A π ( 1 , n ) K μ π − ( p mn L 2 ) . (112)
After fixing Q Q Q and η \eta η , choose the kernel decay exponents needed below, then take the auxiliary ϵ > 0 \epsilon>0 ϵ > 0 sufficiently small in terms of η \eta η and those exponents, and finally fix a 0 a_0 a 0 , J J J as required by Lemma 2.9. In particular, require 0 < ϵ < min { 2 η , 1 / 2 − η } 0<\epsilon<\min\{2\eta,1/2-\eta\} 0 < ϵ < min { 2 η , 1/2 − η } . All these choices precede p → ∞ p\to\infty p → ∞ . This auxiliary ϵ \epsilon ϵ is independent of the prescribed error in the final non-vanishing proportion.
We first justify truncating the two Dirichlet sums in (112) . The finite Euler factors in (2) are O Q ( 1 ) O_Q(1) O Q ( 1 ) by the coefficient bounds of § 2 . The Cauchy–Schwarz inequality and (55) , applied to A π ( 1 , ℓ ) A_\pi(1,\ell) A π ( 1 , ℓ ) with ℓ ≤ L < p 1 / 2 \ell\le L < p^{1/2} ℓ ≤ L < p 1/2 , give, for every fixed C ≥ 0 C \ge0 C ≥ 0 ,
c c u s p ∑ π ∈ B ( p ) h ( μ π ) ( 1 + ∥ μ π ∥ ) C ∣ M π ∣ 2 L ( 1 , π , Ad ) ≪ h , Q , C p 2 ( ∑ ℓ ≤ L ∣ x ℓ ∣ ℓ ) 2 ≪ h , Q , C p 2 L . c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(\mu_\pi)(1+\lVert\mu_\pi\rVert)^C|M_\pi|^2}{L(1,\pi,\operatorname{Ad})}\ll_{h,Q,C}p^2\left(\sum_{\ell\le L}\frac{|x_\ell|}{\sqrt{\ell}}\right)^2\ll_{h,Q,C}p^2L. c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) ( 1 + ∥ μ π ∥ ) C ∣ M π ∣ 2 ≪ h , Q , C p 2 ( ℓ ≤ L ∑ ℓ ∣ x ℓ ∣ ) 2 ≪ h , Q , C p 2 L .
Here we used ∣ x ℓ ∣ ≤ 1 |x_\ell|\le1 ∣ x ℓ ∣ ≤ 1 . Combining (34) with the coefficient bounds of § 2 and the preceding average, we may truncate the two Dirichlet sums in (112) at
m n ≤ p 1 + ε L 2 , m n ≤ p 1 + ε / L 2 , mn\le p^{1+\varepsilon}L^2,\qquad mn\le p^{1+\varepsilon}/L^2, mn ≤ p 1 + ε L 2 , mn ≤ p 1 + ε / L 2 ,
respectively, with total error O h , Q , η , ε , A ( p − A ) O_{h,Q,\eta,\varepsilon,A}(p^{-A}) O h , Q , η , ε , A ( p − A ) for every fixed A > 0 A>0 A > 0 . The same bounds give absolute convergence and justify interchanging the Dirichlet sums with the cuspidal sum. Since L = p η L=p^\eta L = p η and ε < 2 η \varepsilon<2\eta ε < 2 η , the truncated dual range satisfies
m n ≤ p 1 + ε / L 2 = p 1 − 2 η + ε < p . mn\le p^{1+\varepsilon}/L^2=p^{1-2\eta+\varepsilon}<p. mn ≤ p 1 + ε / L 2 = p 1 − 2 η + ε < p .
Hence ( m n , p ) = 1 (mn,p)=1 ( mn , p ) = 1 throughout the truncated dual sum. In the first sum, we retain the terms with p ∣ m n p\mid mn p ∣ mn . For example, m n = p mn=p mn = p gives the weight argument L − 2 < 1 L^{-2}<1 L − 2 < 1 , so (34) gives no decay in p p p for this term.
We next combine the Fourier coefficient products. The classical Hecke relations [23 § 2] imply
A π ( 1 , m ) A π ( 1 , n ) = ∑ r ∣ ( m , n ) A π ( r , m n / r 2 ) . (113) A_\pi(1,m)A_\pi(1,n)=\sum_{r\mid(m,n)}A_\pi(r,mn/r^2). \tag*{(113)} A π ( 1 , m ) A π ( 1 , n ) = r ∣ ( m , n ) ∑ A π ( r , mn / r 2 ) . (113)
This identity also holds at p p p , by the Schur polynomial identity (8) with γ p = 0 \gamma_p=0 γ p = 0 . For ( d b n , p ) = 1 (dbn,p)=1 ( d bn , p ) = 1 , taking complex conjugates in [11 Eq. (1.6)] and using (11) gives
A π ( d , b ) A π ( 1 , n ) = ∑ n 1 n 2 n 3 = n n 2 ∣ b , n 3 ∣ d A π ( d n 2 / n 3 , b n 1 / n 2 ) . (114) A_\pi(d,b)A_\pi(1,n)=\sum_{\substack{n_1n_2n_3=n\\ n_2\mid b,\ n_3\mid d}}A_\pi(dn_2/n_3,bn_1/n_2). \tag*{(114)} A π ( d , b ) A π ( 1 , n ) = n 1 n 2 n 3 = n n 2 ∣ b , n 3 ∣ d ∑ A π ( d n 2 / n 3 , b n 1 / n 2 ) . (114)
Expanding M π 2 M_\pi^2 M π 2 using (2) introduces indices ℓ 1 , ℓ 2 ≤ L \ell_1,\ell_2\le L ℓ 1 , ℓ 2 ≤ L with weights x ℓ 1 x ℓ 2 / ℓ 1 ℓ 2 x_{\ell_1}x_{\ell_2}/\sqrt{\ell_1\ell_2} x ℓ 1 x ℓ 2 / ℓ 1 ℓ 2 . For fixed m , n m,n m , n , applying (113) , (114) and their complex conjugates at unramified primes, we can express each product, including the finite Euler factors, as a finite sum of single Fourier coefficients A π ( u , v ) A_\pi(u,v) A π ( u , v ) . After summing over ℓ 1 \ell_1 ℓ 1 and ℓ 2 \ell_2 ℓ 2 , let b m , n ( u , v ) b_{m,n}(u,v) b m , n ( u , v ) be the total coefficient of A π ( u , v ) A_\pi(u,v) A π ( u , v ) . Thus
M π 2 A π ( 1 , m ) A π ( 1 , n ) = ∑ u , v ≥ 1 b m , n ( u , v ) A π ( u , v ) . (115) M_\pi^2A_\pi(1,m)A_\pi(1,n)=\sum_{u,v\ge1}b_{m,n}(u,v)A_\pi(u,v). \tag*{(115)} M π 2 A π ( 1 , m ) A π ( 1 , n ) = u , v ≥ 1 ∑ b m , n ( u , v ) A π ( u , v ) . (115)
Here m , n m,n m , n are the Dirichlet indices in (112) , and u , v u,v u , v are the Fourier indices in this expansion. The coefficients b m , n ( u , v ) b_{m,n}(u,v) b m , n ( u , v ) are real, depend on Q , L , m , n , u , v Q,L,m,n,u,v Q , L , m , n , u , v , and are independent of π \pi π . Since A π ( 1 , 1 ) = 1 A_\pi(1,1)=1 A π ( 1 , 1 ) = 1 , setting m = n = 1 m=n=1 m = n = 1 and writing the Fourier indices as ( d , b ) (d,b) ( d , b ) gives
M π 2 = ∑ d , b ≥ 1 b 1 , 1 ( d , b ) A π ( d , b ) . M_\pi^2=\sum_{d,b\ge1}b_{1,1}(d,b)A_\pi(d,b). M π 2 = d , b ≥ 1 ∑ b 1 , 1 ( d , b ) A π ( d , b ) .
We now bound the sum of the absolute coefficients in (115) and the Fourier indices that occur there. These bounds will be used to estimate the Weyl terms. For fixed ℓ 1 , ℓ 2 , m , n \ell_1,\ell_2,m,n ℓ 1 , ℓ 2 , m , n , the Hecke relations involve only divisor sums. The divisor bound shows that the number of resulting terms is O ε ( ( m n ℓ 1 ℓ 2 ) ε ) O_{\varepsilon}((mn\ell_1\ell_2)^{\varepsilon}) O ε (( mn ℓ 1 ℓ 2 ) ε ) counted with multiplicity. The finite Euler factors contribute a further constant depending only on Q Q Q . Summing the absolute weights of these terms and using ∣ x ℓ ∣ ≤ 1 |x_\ell|\le1 ∣ x ℓ ∣ ≤ 1 gives
∑ u , v ≥ 1 ∣ b m , n ( u , v ) ∣ ≪ Q , ε ∑ ℓ 1 , ℓ 2 ≤ L ∣ x ℓ 1 x ℓ 2 ∣ ℓ 1 ℓ 2 ( m n ℓ 1 ℓ 2 ) ε ≪ Q , ε ( m n L 2 ) ε L . (116) \sum_{u,v\ge1}|b_{m,n}(u,v)|\ll_{Q,\varepsilon}\sum_{\ell_1,\ell_2\le L}\frac{|x_{\ell_1}x_{\ell_2}|}{\sqrt{\ell_1\ell_2}}(mn\ell_1\ell_2)^{\varepsilon}\ll_{Q,\varepsilon}(mnL^2)^{\varepsilon}L. \tag*{(116)} u , v ≥ 1 ∑ ∣ b m , n ( u , v ) ∣ ≪ Q , ε ℓ 1 , ℓ 2 ≤ L ∑ ℓ 1 ℓ 2 ∣ x ℓ 1 x ℓ 2 ∣ ( mn ℓ 1 ℓ 2 ) ε ≪ Q , ε ( mn L 2 ) ε L . (116)
Next we bound u 2 v u^2v u 2 v , which controls the kernel arguments in the Weyl terms below. In (114) , multiplying A π ( d , b ) A_\pi(d,b) A π ( d , b ) by A π ( 1 , n ) A_\pi(1,n) A π ( 1 , n ) produces Fourier indices ( d n 2 / n 3 , b n 1 / n 2 ) (dn_2/n_3,bn_1/n_2) ( d n 2 / n 3 , b n 1 / n 2 ) . Since n 1 n 2 n 3 = n n_1n_2n_3=n n 1 n 2 n 3 = n ,
( d n 2 n 3 ) 2 b n 1 n 2 = d 2 b n n 3 3 ≤ d 2 b n . \left(\frac{dn_2}{n_3}\right)^2\frac{bn_1}{n_2}=\frac{d^2bn}{n_3^3}\le d^2bn. ( n 3 d n 2 ) 2 n 2 b n 1 = n 3 3 d 2 bn ≤ d 2 bn .
Iteration gives u 2 v ≤ m n ℓ 1 ℓ 2 u^2v\le mn\ell_1\ell_2 u 2 v ≤ mn ℓ 1 ℓ 2 at unramified primes. The same bound holds at p p p by (113) , since ( ℓ 1 ℓ 2 , p ) = 1 (\ell_1\ell_2,p)=1 ( ℓ 1 ℓ 2 , p ) = 1 . The finite Euler factors contribute only a constant depending on Q Q Q . Thus, for ℓ 1 , ℓ 2 ≤ L \ell_1,\ell_2\le L ℓ 1 , ℓ 2 ≤ L and m n ≤ p 1 + ε L 2 mn\le p^{1+\varepsilon}L^2 mn ≤ p 1 + ε L 2 , every non-zero b m , n ( u , v ) b_{m,n}(u,v) b m , n ( u , v ) satisfies
u 2 v ≪ Q m n L 2 ≤ p 1 + ε L 4 . (117) u^2v\ll_Q mnL^2\le p^{1+\varepsilon}L^4. \tag*{(117)} u 2 v ≪ Q mn L 2 ≤ p 1 + ε L 4 . (117)
Since A π ( 1 , 1 ) = 1 A_\pi(1,1)=1 A π ( 1 , 1 ) = 1 , the first sum has Kuznetsov indices ( 1 , 1 , u , v ) (1,1,u,v) ( 1 , 1 , u , v ) . For the second sum in (112) , reindex by π ↦ π ~ \pi\mapsto\widetilde{\pi} π ↦ π . The family and the full adjoint value are unchanged. By (10) , M π ~ = M π M_{\widetilde{\pi}}=M_\pi M π = M π , while (15) and (9) give ϵ ( π ~ ) 2 = p A π ( p 2 , 1 ) \epsilon(\widetilde{\pi})^2=pA_\pi(p^2,1) ϵ ( π ) 2 = p A π ( p 2 , 1 ) . Thus the Fourier coefficient product becomes
p M π ‾ 2 A π ( p 2 , 1 ) A π ( 1 , m ) A π ( 1 , n ) , (118) p\overline{M_\pi}^2A_\pi(p^2,1)A_\pi(1,m)A_\pi(1,n), \tag*{(118)} p M π 2 A π ( p 2 , 1 ) A π ( 1 , m ) A π ( 1 , n ) , (118)
with test K − μ − ( m n L 2 / p ) K^-_{-\mu}(mnL^2/p) K − μ − ( mn L 2 / p ) , since μ π ~ = − μ π \mu_{\widetilde{\pi}}=-\mu_\pi μ π = − μ π by (13) and h h h is even. Apply (113) to the last two factors in (118) . For each divisor e ∣ ( m , n ) e\mid(m,n) e ∣ ( m , n ) , write m = e m ′ m=em' m = e m ′ , n = e n ′ n=en' n = e n ′ and v = m ′ n ′ v=m'n' v = m ′ n ′ . The new coefficient is A π ( e , v ) A_\pi(e,v) A π ( e , v ) and the weight argument is e 2 v L 2 / p e^2vL^2/p e 2 v L 2 / p . Summing over the τ ( v ) \tau(v) τ ( v ) ordered factorizations m ′ n ′ = v m'n'=v m ′ n ′ = v , with m n = e v \sqrt{mn}=e\sqrt{v} mn = e v , gives the scalar weight τ ( v ) / ( e v ) \tau(v)/(e\sqrt{v}) τ ( v ) / ( e v ) . To expand M π ‾ 2 \overline{M_\pi}^2 M π 2 , take the conjugate of the m = n = 1 m=n=1 m = n = 1 case of (115) . Since b 1 , 1 ( d , b ) b_{1,1}(d,b) b 1 , 1 ( d , b ) is real and ( e v , p ) = 1 (ev,p)=1 ( e v , p ) = 1 , multiplicativity gives
p M π ‾ 2 A π ( p 2 , 1 ) A π ( e , v ) = p ∑ d , b ≥ 1 b 1 , 1 ( d , b ) A π ( d , b ) A π ( p 2 e , v ) ‾ . p\overline{M_\pi}^2A_\pi(p^2,1)A_\pi(e,v)=p\sum_{d,b\ge1}b_{1,1}(d,b)A_\pi(d,b)\overline{A_\pi(p^2e,v)}. p M π 2 A π ( p 2 , 1 ) A π ( e , v ) = p d , b ≥ 1 ∑ b 1 , 1 ( d , b ) A π ( d , b ) A π ( p 2 e , v ) .
The Fourier pair on the right has Kuznetsov indices ( d , b , p 2 e , v ) (d,b,p^2e,v) ( d , b , p 2 e , v ) . All mollifier indices are prime to p p p . Together with (117) for m = n = 1 m=n=1 m = n = 1 and the dual cutoff, this gives
d 2 b ≪ Q L 2 , e 2 v ≤ p 1 + ε / L 2 , ( d b e v , p ) = 1. (119) d^2b\ll_Q L^2,\qquad e^2v\le p^{1+\varepsilon}/L^2,\qquad(dbev,p)=1. \tag*{(119)} d 2 b ≪ Q L 2 , e 2 v ≤ p 1 + ε / L 2 , ( d b e v , p ) = 1. (119)
There is no dual diagonal, since the required equality d = p 2 e d=p^2e d = p 2 e contradicts ( d , p ) = 1 (d,p)=1 ( d , p ) = 1 .
To sum the short and long Weyl estimates below over all indices, we bound the sums of the absolute values of their coefficients. The first and dual sums have coefficients b m , n ( u , v ) / m n b_{m,n}(u,v)/\sqrt{mn} b m , n ( u , v ) / mn and p τ ( v ) b 1 , 1 ( d , b ) / ( e v ) p\tau(v)b_{1,1}(d,b)/(e\sqrt{v}) p τ ( v ) b 1 , 1 ( d , b ) / ( e v ) , respectively. Combining (116) with the divisor bound gives
∑ m , n ≥ 1 m n ≤ p 1 + ε L 2 1 m n ∑ u , v ≥ 1 ∣ b m , n ( u , v ) ∣ ≪ Q , ε p 1 / 2 + O ( ε ) L 2 + O ( ε ) , \sum_{\substack{m,n\ge1\\mn\le p^{1+\varepsilon}L^2}}\frac{1}{\sqrt{mn}}\sum_{u,v\ge1}|b_{m,n}(u,v)|\ll_{Q,\varepsilon}p^{1/2+O(\varepsilon)}L^{2+O(\varepsilon)}, m , n ≥ 1 mn ≤ p 1 + ε L 2 ∑ mn 1 u , v ≥ 1 ∑ ∣ b m , n ( u , v ) ∣ ≪ Q , ε p 1/2 + O ( ε ) L 2 + O ( ε ) ,
∑ e , v ≥ 1 ( e v , p ) = 1 e 2 v ≤ p 1 + ε / L 2 τ ( v ) e v ∑ d , b ≥ 1 ∣ b 1 , 1 ( d , b ) ∣ ≪ Q , ε p 1 / 2 + O ( ε ) . (120) \sum_{\substack{e,v\ge1\\(ev,p)=1\\e^2v\le p^{1+\varepsilon}/L^2}}\frac{\tau(v)}{e\sqrt{v}}\sum_{d,b\ge1}|b_{1,1}(d,b)|\ll_{Q,\varepsilon}p^{1/2+O(\varepsilon)}. \tag*{(120)} e , v ≥ 1 ( e v , p ) = 1 e 2 v ≤ p 1 + ε / L 2 ∑ e v τ ( v ) d , b ≥ 1 ∑ ∣ b 1 , 1 ( d , b ) ∣ ≪ Q , ε p 1/2 + O ( ε ) . (120)
The second bound excludes the factor p p p . Including it, the sum of absolute coefficients in the dual contribution is O Q , ε ( p 3 / 2 + O ( ε ) ) O_{Q,\varepsilon}(p^{3/2+O(\varepsilon)}) O Q , ε ( p 3/2 + O ( ε ) ) .
Up to an error O h , Q , η , ε ( p − 1 ) O_{h,Q,\eta,\varepsilon}(p^{-1}) O h , Q , η , ε ( p − 1 ) , applying the preceding coefficient identities to (112) with these cutoffs gives
M 1 , 1 ( p ) = c c u s p ∑ m , n , u , v ≥ 1 m n ≤ p 1 + ε L 2 ∑ π ∈ B ( p ) b m , n ( u , v ) A π ( u , v ) K μ + ( m n / ( p L 2 ) ) m n L ( 1 , π , Ad ) + p c c u s p ∑ e , v , d , b ≥ 1 ( e v , p ) = 1 e 2 v ≤ p 1 + ε / L 2 ∑ π ∈ B ( p ) τ ( v ) b 1 , 1 ( d , b ) A π ( d , b ) A π ( p 2 e , v ) K − μ − ( e 2 v L 2 / p ) e v v L ( 1 , π , Ad ) . (121) \begin{aligned}
\mathcal{M}_{1,1}(p)={}&c_{\mathrm{cusp}}\sum_{\substack{m,n,u,v\geq1\\ mn\leq p^{1+\varepsilon}L^2}}\sum_{\pi\in\mathcal{B}(p)}
\frac{b_{m,n}(u,v)A_\pi(u,v)K_{\mu}^{+}\left(mn/(pL^2)\right)}
{\sqrt{mn}L(1,\pi,\operatorname{Ad})}\\
&+pc_{\mathrm{cusp}}\sum_{\substack{e,v,d,b\geq1\\ (ev,p)=1\\ e^2v\leq p^{1+\varepsilon}/L^2}}\sum_{\pi\in\mathcal{B}(p)}
\frac{\tau(v)b_{1,1}(d,b)A_\pi(d,b)A_\pi(p^2e,v)K_{-\mu}^{-}\left(e^2vL^2/p\right)}
{ev\sqrt{v}L(1,\pi,\operatorname{Ad})}. \tag*{(121)}
\end{aligned} M 1 , 1 ( p ) = c cusp m , n , u , v ≥ 1 mn ≤ p 1 + ε L 2 ∑ π ∈ B ( p ) ∑ mn L ( 1 , π , Ad ) b m , n ( u , v ) A π ( u , v ) K μ + ( mn / ( p L 2 ) ) + p c cusp e , v , d , b ≥ 1 ( e v , p ) = 1 e 2 v ≤ p 1 + ε / L 2 ∑ π ∈ B ( p ) ∑ e v v L ( 1 , π , Ad ) τ ( v ) b 1 , 1 ( d , b ) A π ( d , b ) A π ( p 2 e , v ) K − μ − ( e 2 v L 2 / p ) . (121)
The tests in (121) are admissible by Lemma 2.9 and § 3.7 . Since the truncated Dirichlet sums and their Hecke expansions are finite, Theorem 3.1 applies term by term. We retain the length cutoffs in the Weyl terms. In the diagonal and the four remaining spectral contributions, we remove the length restrictions, retaining ( e v , p ) = 1 (ev,p)=1 ( e v , p ) = 1 in the dual sums. The additional terms have weight arguments m n / ( p L 2 ) > p ε mn/(pL^2)>p^\varepsilon mn / ( p L 2 ) > p ε or e 2 v L 2 / p > p ε e^2vL^2/p>p^\varepsilon e 2 v L 2 / p > p ε .
The truncation argument following (64) applies with (34) and (116) . For the maximal new spectrum, we first separate the powers of p p p in the local coefficients and use (73) . The resulting polynomial losses are absorbed by choosing the weight decay sufficiently large. Thus each added contribution is O h , Q , η , J , ε , A ( p − A ) O_{h,Q,\eta,J,\varepsilon,A}(p^{-A}) O h , Q , η , J , ε , A ( p − A ) for every fixed A > 0 A>0 A > 0 , with the cutoff exponent ε \varepsilon ε and the desired A A A fixed before choosing the decay exponent.
Taking A = 1 A=1 A = 1 , define the diagonal contribution by
D 1 , 1 : = P ∑ m , n ≥ 1 b m , n ( 1 , 1 ) m n ∫ Re ( μ ) = 0 K μ + ( m n / ( p L 2 ) ) spec ( μ ) d μ . (122) D_{1,1}:=P\sum_{m,n\geq1}\frac{b_{m,n}(1,1)}{\sqrt{mn}}\int_{\operatorname{Re}(\mu)=0}K_{\mu}^{+}\left(mn/(pL^2)\right)\operatorname{spec}(\mu)\,d\mu. \tag*{(122)} D 1 , 1 := P m , n ≥ 1 ∑ mn b m , n ( 1 , 1 ) ∫ Re ( μ ) = 0 K μ + ( mn / ( p L 2 ) ) spec ( μ ) d μ . (122)
For i = 4 , 5 , 6 i=4,5,6 i = 4 , 5 , 6 , the Weyl contributions are
S i , 1 , 1 : = ∑ m , n ≥ 1 m n ≤ p 1 + ε L 2 1 m n ∑ u , v ≥ 1 b m , n ( u , v ) S i ( 1 , 1 , u , v ; μ ↦ K μ + ( m n / ( p L 2 ) ) ) + p ∑ e , v ≥ 1 ( e v , p ) = 1 e 2 v ≤ p 1 + ε / L 2 τ ( v ) e v ∑ d , b ≥ 1 b 1 , 1 ( d , b ) S i ( d , b , p 2 e , v ; μ ↦ K − μ − ( e 2 v L 2 / p ) ) . \begin{aligned}
S_{i,1,1}:={}&\sum_{\substack{m,n\geq1\\ mn\leq p^{1+\varepsilon}L^2}}\frac{1}{\sqrt{mn}}\sum_{u,v\geq1}b_{m,n}(u,v)S_i\left(1,1,u,v;\mu\mapsto K_{\mu}^{+}\left(mn/(pL^2)\right)\right)\\
&+p\sum_{\substack{e,v\geq1\\ (ev,p)=1\\ e^2v\leq p^{1+\varepsilon}/L^2}}\frac{\tau(v)}{e\sqrt{v}}\sum_{d,b\geq1}b_{1,1}(d,b)S_i\left(d,b,p^2e,v;\mu\mapsto K_{-\mu}^{-}\left(e^2vL^2/p\right)\right).
\end{aligned} S i , 1 , 1 := m , n ≥ 1 mn ≤ p 1 + ε L 2 ∑ mn 1 u , v ≥ 1 ∑ b m , n ( u , v ) S i ( 1 , 1 , u , v ; μ ↦ K μ + ( mn / ( p L 2 ) ) ) + p e , v ≥ 1 ( e v , p ) = 1 e 2 v ≤ p 1 + ε / L 2 ∑ e v τ ( v ) d , b ≥ 1 ∑ b 1 , 1 ( d , b ) S i ( d , b , p 2 e , v ; μ ↦ K − μ − ( e 2 v L 2 / p ) ) .
Here S i S_i S i uses the displayed Fourier indices and spectral test, with the conventions of § 3.5 3.5 3.5 and the transforms in (39) . The four remaining spectral contributions are
R o l d , 1 , 1 : = c c u s p ∑ Π ∈ B ( 1 ) ∑ j = 0 2 1 L ( 1 , Π , Ad ) × ( ∑ m , n ≥ 1 1 m n ∑ u , v ≥ 1 b m , n ( u , v ) B Π ( j ) ( 1 , 1 ) ‾ B Π ( j ) ( u , v ) K μ + ( m n p L 2 ) + p ∑ e , v ≥ 1 ( e v , p ) = 1 τ ( v ) e v ∑ d , b ≥ 1 b 1 , 1 ( d , b ) B Π ( j ) ( d , b ) ‾ B Π ( j ) ( p 2 e , v ) K − μ − ( e 2 v L 2 p ) ) , \begin{aligned}
\mathcal{R}_{\mathrm{old},1,1} :={}& c_{\mathrm{cusp}} \sum_{\Pi\in\mathcal{B}(1)}\sum_{j=0}^{2}\frac{1}{L(1,\Pi,\operatorname{Ad})}\\
&\times\left(\sum_{m,n\geq1}\frac{1}{\sqrt{mn}}\sum_{u,v\geq1}b_{m,n}(u,v)\overline{B_{\Pi}^{(j)}(1,1)}B_{\Pi}^{(j)}(u,v)K_{\mu}^{+}\left(\frac{mn}{pL^{2}}\right)\right.\\
&\left.\qquad+p\sum_{\substack{e,v\geq1\\(ev,p)=1}}\frac{\tau(v)}{e\sqrt{v}}\sum_{d,b\geq1}b_{1,1}(d,b)\overline{B_{\Pi}^{(j)}(d,b)}B_{\Pi}^{(j)}(p^{2}e,v)K_{-\mu}^{-}\left(\frac{e^{2}vL^{2}}{p}\right)\right),
\end{aligned} R old , 1 , 1 := c cusp Π ∈ B ( 1 ) ∑ j = 0 ∑ 2 L ( 1 , Π , Ad ) 1 × ( m , n ≥ 1 ∑ mn 1 u , v ≥ 1 ∑ b m , n ( u , v ) B Π ( j ) ( 1 , 1 ) B Π ( j ) ( u , v ) K μ + ( p L 2 mn ) + p e , v ≥ 1 ( e v , p ) = 1 ∑ e v τ ( v ) d , b ≥ 1 ∑ b 1 , 1 ( d , b ) B Π ( j ) ( d , b ) B Π ( j ) ( p 2 e , v ) K − μ − ( p e 2 v L 2 ) ,
R min , 1 , 1 : = c min ∑ j = 0 2 ∫ Re ( s 1 ) = 0 ∫ Re ( s 2 ) = 0 1 ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 × ( ∑ m , n ≥ 1 1 m n ∑ u , v ≥ 1 b m , n ( u , v ) B s ( j ) ( 1 , 1 ) ‾ B s ( j ) ( u , v ) K s + ( m n p L 2 ) + p ∑ e , v ≥ 1 ( e v , p ) = 1 τ ( v ) e v ∑ d , b ≥ 1 b 1 , 1 ( d , b ) B s ( j ) ( d , b ) ‾ B s ( j ) ( p 2 e , v ) K − s − ( e 2 v L 2 p ) ) d s 1 d s 2 , \begin{aligned}
\mathcal{R}_{\min,1,1}:={}&c_{\min}\sum_{j=0}^{2}\int_{\operatorname{Re}(s_{1})=0}\int_{\operatorname{Re}(s_{2})=0}\frac{1}{\prod_{1\leq i<k\leq3}\left|\zeta(1+s_i-s_k)\right|^{2}}\\
&\times\left(\sum_{m,n\geq1}\frac{1}{\sqrt{mn}}\sum_{u,v\geq1}b_{m,n}(u,v)\overline{B_{s}^{(j)}(1,1)}B_{s}^{(j)}(u,v)K_{s}^{+}\left(\frac{mn}{pL^{2}}\right)\right.\\
&\left.\qquad+p\sum_{\substack{e,v\geq1\\(ev,p)=1}}\frac{\tau(v)}{e\sqrt{v}}\sum_{d,b\geq1}b_{1,1}(d,b)\overline{B_{s}^{(j)}(d,b)}B_{s}^{(j)}(p^{2}e,v)K_{-s}^{-}\left(\frac{e^{2}vL^{2}}{p}\right)\right)\,ds_{1}\,ds_{2},
\end{aligned} R m i n , 1 , 1 := c m i n j = 0 ∑ 2 ∫ Re ( s 1 ) = 0 ∫ Re ( s 2 ) = 0 ∏ 1 ≤ i < k ≤ 3 ∣ ζ ( 1 + s i − s k ) ∣ 2 1 × ( m , n ≥ 1 ∑ mn 1 u , v ≥ 1 ∑ b m , n ( u , v ) B s ( j ) ( 1 , 1 ) B s ( j ) ( u , v ) K s + ( p L 2 mn ) + p e , v ≥ 1 ( e v , p ) = 1 ∑ e v τ ( v ) d , b ≥ 1 ∑ b 1 , 1 ( d , b ) B s ( j ) ( d , b ) B s ( j ) ( p 2 e , v ) K − s − ( p e 2 v L 2 ) d s 1 d s 2 ,
and, with μ = ( z + τ ϕ , z − τ ϕ , − 2 z ) \mu=(z+\tau_{\phi},z-\tau_{\phi},-2z) μ = ( z + τ ϕ , z − τ ϕ , − 2 z ) ,
R max , n e w , 1 , 1 : = χ p c max ∑ ϕ ∈ B 2 ( p ) ∫ Re ( z ) = 0 1 L ( 1 , ϕ , Ad ) ‾ ∣ L ( 1 + 3 z , ϕ ) ∣ 2 × ( ∑ m , n ≥ 1 1 m n ∑ u , v ≥ 1 b m , n ( u , v ) A ϕ , z ( u , v ) K μ + ( m n p L 2 ) + p ∑ e , v ≥ 1 ( e v , p ) = 1 τ ( v ) e v ∑ d , b ≥ 1 b 1 , 1 ( d , b ) A ϕ , z ( d , b ) A ϕ , z ( p 2 e , v ) K − μ − ( e 2 v L 2 p ) ) d z , \begin{aligned}
\mathcal{R}_{\max,\mathrm{new},1,1}:={}&\chi_p c_{\max}\sum_{\phi\in\mathcal{B}_2(p)}\int_{\operatorname{Re}(z)=0}\frac{1}{\overline{L(1,\phi,\operatorname{Ad})}\left|L(1+3z,\phi)\right|^{2}}\\
&\times\left(\sum_{m,n\geq1}\frac{1}{\sqrt{mn}}\sum_{u,v\geq1}b_{m,n}(u,v)A_{\phi,z}(u,v)K_{\mu}^{+}\left(\frac{mn}{pL^{2}}\right)\right.\\
&\left.\qquad+p\sum_{\substack{e,v\geq1\\(ev,p)=1}}\frac{\tau(v)}{e\sqrt{v}}\sum_{d,b\geq1}b_{1,1}(d,b)A_{\phi,z}(d,b)A_{\phi,z}(p^{2}e,v)K_{-\mu}^{-}\left(\frac{e^{2}vL^{2}}{p}\right)\right)\,dz,
\end{aligned} R m a x , new , 1 , 1 := χ p c m a x ϕ ∈ B 2 ( p ) ∑ ∫ Re ( z ) = 0 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 1 × ( m , n ≥ 1 ∑ mn 1 u , v ≥ 1 ∑ b m , n ( u , v ) A ϕ , z ( u , v ) K μ + ( p L 2 mn ) + p e , v ≥ 1 ( e v , p ) = 1 ∑ e v τ ( v ) d , b ≥ 1 ∑ b 1 , 1 ( d , b ) A ϕ , z ( d , b ) A ϕ , z ( p 2 e , v ) K − μ − ( p e 2 v L 2 ) d z ,
R max , o l d , 1 , 1 : = c max ∑ ϕ ∈ B 2 ( 1 ) ∑ j = 0 2 ∫ Re ( z ) = 0 1 L ( 1 , ϕ , Ad ) ‾ ∣ L ( 1 + 3 z , ϕ ) ∣ 2 × ( ∑ m , n ≥ 1 1 m n ∑ u , v ≥ 1 b m , n ( u , v ) B ϕ , z ( j ) ( 1 , 1 ) ‾ B ϕ , z ( j ) ( u , v ) K μ + ( m n p L 2 ) + p ∑ e , v ≥ 1 ( e v , p ) = 1 τ ( v ) e v ∑ d , b ≥ 1 b 1 , 1 ( d , b ) B ϕ , z ( j ) ( d , b ) ‾ B ϕ , z ( j ) ( p 2 e , v ) K − μ − ( e 2 v L 2 p ) ) d z . \begin{aligned}
\mathcal{R}_{\max,\mathrm{old},1,1}:={}&c_{\max}\sum_{\phi\in\mathcal{B}_2(1)}\sum_{j=0}^{2}\int_{\operatorname{Re}(z)=0}\frac{1}{\overline{L(1,\phi,\operatorname{Ad})}\left|L(1+3z,\phi)\right|^{2}}\\
&\times\left(\sum_{m,n\geq1}\frac{1}{\sqrt{mn}}\sum_{u,v\geq1}b_{m,n}(u,v)\overline{B_{\phi,z}^{(j)}(1,1)}B_{\phi,z}^{(j)}(u,v)K_{\mu}^{+}\left(\frac{mn}{pL^{2}}\right)\right.\\
&\left.\qquad+p\sum_{\substack{e,v\geq1\\(ev,p)=1}}\frac{\tau(v)}{e\sqrt{v}}\sum_{d,b\geq1}b_{1,1}(d,b)\overline{B_{\phi,z}^{(j)}(d,b)}B_{\phi,z}^{(j)}(p^{2}e,v)K_{-\mu}^{-}\left(\frac{e^{2}vL^{2}}{p}\right)\right)\,dz.
\end{aligned} R m a x , old , 1 , 1 := c m a x ϕ ∈ B 2 ( 1 ) ∑ j = 0 ∑ 2 ∫ Re ( z ) = 0 L ( 1 , ϕ , Ad ) ∣ L ( 1 + 3 z , ϕ ) ∣ 2 1 × ( m , n ≥ 1 ∑ mn 1 u , v ≥ 1 ∑ b m , n ( u , v ) B ϕ , z ( j ) ( 1 , 1 ) B ϕ , z ( j ) ( u , v ) K μ + ( p L 2 mn ) + p e , v ≥ 1 ( e v , p ) = 1 ∑ e v τ ( v ) d , b ≥ 1 ∑ b 1 , 1 ( d , b ) B ϕ , z ( j ) ( d , b ) B ϕ , z ( j ) ( p 2 e , v ) K − μ − ( p e 2 v L 2 ) d z .
The continuous measures retain the conventions of § 3.5 . Thus
χ p M 1 , 1 ( p ) = D 1 , 1 + ∑ i = 4 6 S i , 1 , 1 − R o l d , 1 , 1 − R min , 1 , 1 − R max , n e w , 1 , 1 − R max , o l d , 1 , 1 + O h , Q , η , ε ( p − 1 ) . (123) \chi_p\mathcal{M}_{1,1}(p)=\mathcal{D}_{1,1}+\sum_{i=4}^{6}\mathcal{S}_{i,1,1}-\mathcal{R}_{\mathrm{old},1,1}-\mathcal{R}_{\min,1,1}
-\mathcal{R}_{\max,\mathrm{new},1,1}-\mathcal{R}_{\max,\mathrm{old},1,1}+O_{h,Q,\eta,\varepsilon}(p^{-1}). \tag*{(123)} χ p M 1 , 1 ( p ) = D 1 , 1 + i = 4 ∑ 6 S i , 1 , 1 − R old , 1 , 1 − R m i n , 1 , 1 − R m a x , new , 1 , 1 − R m a x , old , 1 , 1 + O h , Q , η , ε ( p − 1 ) . (123)
We estimate these contributions in the following subsections.
The diagonal term
By (122) , the first trace diagonal selects the coefficient of A π ( 1 , 1 ) A_\pi(1,1) A π ( 1 , 1 ) in (115) . We first compute the sum of these coefficients with ν μ + \nu_\mu^+ ν μ + replaced by 1, and then estimate the error in this replacement.
For q ∤ p Q q\nmid pQ q ∤ pQ , write
L q = L q ( 1 / 2 , π ) = ∑ j ≥ 0 A π ( 1 , q j ) q j / 2 . L_q=L_q(1/2,\pi)=\sum_{j\geq0}\frac{A_\pi(1,q^j)}{q^{j/2}}. L q = L q ( 1/2 , π ) = j ≥ 0 ∑ q j /2 A π ( 1 , q j ) .
By (113) ,
L q 2 = ∑ j , k ≥ 0 q − ( j + k ) / 2 ∑ r = 0 min ( j , k ) A π ( q r , q j + k − 2 r ) . L_q^2=\sum_{j,k\geq0}q^{-(j+k)/2}\sum_{r=0}^{\min(j,k)}A_\pi(q^r,q^{j+k-2r}). L q 2 = j , k ≥ 0 ∑ q − ( j + k ) /2 r = 0 ∑ m i n ( j , k ) A π ( q r , q j + k − 2 r ) .
The coefficients of A π ( 1 , 1 ) A_\pi(1,1) A π ( 1 , 1 ) , A π ( q , 1 ) A_\pi(q,1) A π ( q , 1 ) , A π ( 1 , q ) A_\pi(1,q) A π ( 1 , q ) and A π ( q 2 , 1 ) A_\pi(q^2,1) A π ( q 2 , 1 ) in this expansion are 1 1 1 , q − 1 q^{-1} q − 1 , 2 q − 1 / 2 2q^{-1/2} 2 q − 1/2 and q − 2 q^{-2} q − 2 , respectively. By (114) with n = q n=q n = q , multiplication by A π ( 1 , q ) A_\pi(1,q) A π ( 1 , q ) produces A π ( 1 , 1 ) A_\pi(1,1) A π ( 1 , 1 ) only from A π ( q , 1 ) A_\pi(q,1) A π ( q , 1 ) , and produces A π ( q , 1 ) A_\pi(q,1) A π ( q , 1 ) only from A π ( 1 , q ) A_\pi(1,q) A π ( 1 , q ) and A π ( q 2 , 1 ) A_\pi(q^2,1) A π ( q 2 , 1 ) . Thus the coefficients of A π ( 1 , 1 ) A_\pi(1,1) A π ( 1 , 1 ) in L q 2 L_q^2 L q 2 , L q 2 A π ( 1 , q ) L_q^2A_\pi(1,q) L q 2 A π ( 1 , q ) and L q 2 A π ( 1 , q ) 2 L_q^2A_\pi(1,q)^2 L q 2 A π ( 1 , q ) 2 are
1 , q − 1 , 2 q − 1 / 2 + q − 2 , 1,\qquad q^{-1},\qquad2q^{-1/2}+q^{-2}, 1 , q − 1 , 2 q − 1/2 + q − 2 ,
respectively. The second value comes from m = n = q m=n=q m = n = q . For the third, the contributing Dirichlet pairs are ( m , n ) = ( q , 1 ) , ( 1 , q ) , ( q 2 , q 2 ) (m,n)=(q,1),(1,q),(q^2,q^2) ( m , n ) = ( q , 1 ) , ( 1 , q ) , ( q 2 , q 2 ) , with weights q − 1 / 2 , q − 1 / 2 , q − 2 q^{-1/2},q^{-1/2},q^{-2} q − 1/2 , q − 1/2 , q − 2 .
We first compute the local diagonal with the logarithmic mollifier cutoffs removed, keeping ( ℓ , p Q ) = 1 (\ell,pQ)=1 ( ℓ , pQ ) = 1 . The absolutely convergent majorant below justifies this formal computation. The local mollifier factor at q ∤ p Q q\nmid pQ q ∤ pQ is then 1 − q q + 1 A π ( 1 , q ) / q 1-\frac{q}{q+1}A_\pi(1,q)/\sqrt q 1 − q + 1 q A π ( 1 , q ) / q . Expanding its square and using the three coefficients above, the coefficient of A π ( 1 , 1 ) A_\pi(1,1) A π ( 1 , 1 ) in the local product is
1 − 2 q q + 1 q − 3 / 2 + ( q q + 1 ) 2 ( 2 q − 3 / 2 + q − 3 ) = 1 − 2 q ( q + 1 ) 2 + 1 q ( q + 1 ) 2 . 1-\frac{2q}{q+1}q^{-3/2}+\left(\frac{q}{q+1}\right)^2(2q^{-3/2}+q^{-3})=1-\frac{2}{\sqrt q(q+1)^2}+\frac{1}{q(q+1)^2}. 1 − q + 1 2 q q − 3/2 + ( q + 1 q ) 2 ( 2 q − 3/2 + q − 3 ) = 1 − q ( q + 1 ) 2 2 + q ( q + 1 ) 2 1 .
For q ∣ Q q\mid Q q ∣ Q , the finite inverse Euler factors in (2) cancel L q 2 L_q^2 L q 2 exactly. At p p p , the mollifier has local factor 1. By (113) , A π ( 1 , p j ) A π ( 1 , p k ) A_\pi(1,p^j)A_\pi(1,p^k) A π ( 1 , p j ) A π ( 1 , p k ) contributes to A π ( 1 , 1 ) A_\pi(1,1) A π ( 1 , 1 ) only when j = k = 0 j=k=0 j = k = 0 , so the local diagonal factor at p p p is also 1.
We now justify removing the mollifier cutoffs. Taking absolute values before collecting the Hecke coefficients gives local diagonal factors 1 + O ( q − 3 / 2 ) 1+O(q^{-3/2}) 1 + O ( q − 3/2 ) for q ∤ p Q q\nmid pQ q ∤ pQ . Since ∑ q ( log q ) q − 3 / 2 < ∞ \sum_q(\log q)q^{-3/2}<\infty ∑ q ( log q ) q − 3/2 < ∞ , the corresponding absolute series weighted by log ℓ 1 + log ℓ 2 \log\ell_1+\log\ell_2 log ℓ 1 + log ℓ 2 also converges. The elementary bound
∣ 1 − 1 ℓ ≤ L log ( L / ℓ ) log L ∣ ≤ log ℓ log L , ℓ ≥ 1 , \left|1-1_{\ell\leq L}\frac{\log(L/\ell)}{\log L}\right|\leq\frac{\log\ell}{\log L},\qquad\ell\geq1, 1 − 1 ℓ ≤ L log L log ( L / ℓ ) ≤ log L log ℓ , ℓ ≥ 1 ,
applied to both mollifier indices therefore gives
∑ m , n ≥ 1 b m , n ( 1 , 1 ) m n = ∏ q ∤ p Q ( 1 − 2 q ( q + 1 ) 2 + 1 q ( q + 1 ) 2 ) + O Q ( 1 log L ) . \sum_{m,n\geq1}\frac{b_{m,n}(1,1)}{\sqrt{mn}}=\prod_{q\nmid pQ}\left(1-\frac{2}{\sqrt q(q+1)^2}+\frac{1}{q(q+1)^2}\right)+O_Q\left(\frac{1}{\log L}\right). m , n ≥ 1 ∑ mn b m , n ( 1 , 1 ) = q ∤ pQ ∏ ( 1 − q ( q + 1 ) 2 2 + q ( q + 1 ) 2 1 ) + O Q ( log L 1 ) .
Next fix 0 < c < 1 / 4 0<c<1/4 0 < c < 1/4 to estimate the replacement of V μ + \mathcal{V}_{\mu}^{+} V μ + by 1. Weighting the Dirichlet pairs by ( m n ) c (mn)^c ( mn ) c gives local absolute factors
1 + O ( q − 3 / 2 + 2 c + q − 3 / 2 + c + q − 3 + 4 c ) . 1+O\left(q^{-3/2+2c}+q^{-3/2+c}+q^{-3+4c}\right). 1 + O ( q − 3/2 + 2 c + q − 3/2 + c + q − 3 + 4 c ) .
The three powers of q c q^c q c correspond to m n = q 2 , q , q 4 mn=q^2,q,q^4 mn = q 2 , q , q 4 , respectively. These factors have an absolutely convergent product for 0 < c < 1 / 4 0<c<1/4 0 < c < 1/4 . Since the cutoff factors have absolute value at most 1, this proves, uniformly for L < p L<p L < p and p ∤ Q p\nmid Q p ∤ Q ,
∑ m , n ≥ 1 ∣ b m , n ( 1 , 1 ) ∣ m n ( m n ) c ≪ Q , c 1. \sum_{m,n\geq1}\frac{\lvert b_{m,n}(1,1)\rvert}{\sqrt{mn}}(mn)^c\ll_{Q,c}1. m , n ≥ 1 ∑ mn ∣ b m , n ( 1 , 1 )∣ ( mn ) c ≪ Q , c 1.
(2.28) and (2.27) imply, for all y > 0 y>0 y > 0 ,
∣ V μ + ( y ) − 1 ∣ ≪ J , c y c ( 1 + ∥ μ ∥ ) C . \left\lvert\mathcal{V}_{\mu}^{+}(y)-1\right\rvert\ll_{J,c}y^c(1+\lVert\mu\rVert)^C. V μ + ( y ) − 1 ≪ J , c y c ( 1 + ∥ μ ∥ ) C .
Combining these bounds with (3.2) in (5.35) yields
D 1 , 1 = P H h ∑ m , n ≥ 1 b m , n ( 1 , 1 ) m n + O h , Q , J , c ( p 2 ( p L 2 ) − c ) . D_{1,1}=P\mathcal{H}_h\sum_{m,n\geq1}\frac{b_{m,n}(1,1)}{\sqrt{mn}}+O_{h,Q,J,c}\left(p^2(pL^2)^{-c}\right). D 1 , 1 = P H h m , n ≥ 1 ∑ mn b m , n ( 1 , 1 ) + O h , Q , J , c ( p 2 ( p L 2 ) − c ) .
Finally, the factor at q = p q=p q = p in the unramified Euler product is
1 − 2 p ( p + 1 ) 2 + 1 p ( p + 1 ) 2 = 1 + O ( p − 5 / 2 ) . 1-\frac{2}{\sqrt{p}(p+1)^2}+\frac{1}{p(p+1)^2}=1+O(p^{-5/2}). 1 − p ( p + 1 ) 2 2 + p ( p + 1 ) 2 1 = 1 + O ( p − 5/2 ) .
Including it changes the main term by O h , Q ( p − 1 / 2 ) O_{h,Q}(p^{-1/2}) O h , Q ( p − 1/2 ) . Since L = p η L=p^\eta L = p η , we have 1 / log L = 1 / ( η log p ) 1/\log L=1/(\eta\log p) 1/ log L = 1/ ( η log p ) and ( p L 2 ) − c = p − c ( 1 + 2 η ) ≪ η , c 1 / log p (pL^2)^{-c}=p^{-c(1+2\eta)}\ll_{\eta,c}1/\log p ( p L 2 ) − c = p − c ( 1 + 2 η ) ≪ η , c 1/ log p . The preceding estimates therefore give
D 1 , 1 = P H h ∏ q ∣ Q ( 1 − 2 q ( q + 1 ) 2 + 1 q ( q + 1 ) 2 ) + O h , Q , η , J ( p 2 log p ) . (124) D_{1,1}=P\mathcal{H}_h\prod_{q\mid Q}\left(1-\frac{2}{\sqrt{q}(q+1)^2}+\frac{1}{q(q+1)^2}\right)+O_{h,Q,\eta,J}\left(\frac{p^2}{\log p}\right). \tag*{(124)} D 1 , 1 = P H h q ∣ Q ∏ ( 1 − q ( q + 1 ) 2 2 + q ( q + 1 ) 2 1 ) + O h , Q , η , J ( log p p 2 ) . (124)
The short Weyl elements
Lemma 5.5. For i = 4 , 5 i=4,5 i = 4 , 5 , the first and dual contributions to S i , 1 , 1 S_{i,1,1} S i , 1 , 1 are each O h , Q , η , J ( p − 1 ) O_{h,Q,\eta,J}(p^{-1}) O h , Q , η , J ( p − 1 ) .
Proof. Choose A A A sufficiently large in terms of η \eta η , and take a 0 , J a_0,J a 0 , J large enough for the corresponding kernel estimates. Since L = p η < p 1 / 2 L=p^\eta<p^{1/2} L = p η < p 1/2 , the weight arguments m n / ( p L 2 ) mn/(pL^2) mn / ( p L 2 ) and e 2 v L 2 / p e^{2v}L^2/p e 2 v L 2 / p are at least p − 2 p^{-2} p − 2 . By Remark 5, (2.29) with c = ε / 2 c=\varepsilon/2 c = ε /2 , and (3.17), for either test functions in (5.34),
∣ Φ w 4 ( y ) ∣ + ∣ Φ w 5 ( y ) ∣ ≪ A , h , J , ε p ε ∣ y ∣ A . \lvert\Phi_{w_4}(y)\rvert+\lvert\Phi_{w_5}(y)\rvert\ll_{A,h,J,\varepsilon}p^\varepsilon\lvert y\rvert^A. ∣ Φ w 4 ( y )∣ + ∣ Φ w 5 ( y )∣ ≪ A , h , J , ε p ε ∣ y ∣ A . For the first indices ( 1 , 1 , u , v ) (1,1,u,v) ( 1 , 1 , u , v ) , the w 4 w_4 w 4 modulus conditions are parametrized by D 2 = p k / ( p , u ) D_2=pk/(p,u) D 2 = p k / ( p , u ) and D 1 = u D 2 2 D_1=uD_2^2 D 1 = u D 2 2 . For w 5 w_5 w 5 , they give D 1 = p k D_1=pk D 1 = p k and D 2 = v D 1 2 D_2=vD_1^2 D 2 = v D 1 2 , with k ≥ 1 k\geq1 k ≥ 1 . Consequently,
∣ y w 4 ∣ = v ( p , u ) 3 u p 3 k 3 ≤ u 2 v p 3 k 3 , ∣ y w 5 ∣ = u v p 3 k 3 ≤ u 2 v p 3 k 3 . \lvert y_{w_4}\rvert=\frac{v(p,u)^3}{up^3k^3}\leq\frac{u^2v}{p^3k^3},\qquad
\lvert y_{w_5}\rvert=\frac{u}{vp^3k^3}\leq\frac{u^2v}{p^3k^3}. ∣ y w 4 ∣ = u p 3 k 3 v ( p , u ) 3 ≤ p 3 k 3 u 2 v , ∣ y w 5 ∣ = v p 3 k 3 u ≤ p 3 k 3 u 2 v . By (5.30), both arguments are ≪ Q p − 2 + ε L 4 k − 3 \ll_Q p^{-2+\varepsilon}L^4k^{-3} ≪ Q p − 2 + ε L 4 k − 3 . For the dual indices ( d , b , p 2 e , v ) (d,b,p^2e,v) ( d , b , p 2 e , v ) , using ( d b e v , p ) = 1 (dbev,p)=1 ( d b e v , p ) = 1 , the corresponding parametrizations are
w 4 : D 2 = b k ( b , e ) , D 1 = p 2 b e k 2 ( b , e ) 2 , w_4:\quad D_2=\frac{bk}{(b,e)},\qquad D_1=\frac{p^2bek^2}{(b,e)^2}, w 4 : D 2 = ( b , e ) bk , D 1 = ( b , e ) 2 p 2 b e k 2 , w 5 : D 1 = p d k ( d , v ) , D 2 = p 2 d v k 2 ( d , v ) 2 , k ≥ 1. w_5:\quad D_1=\frac{pdk}{(d,v)},\qquad D_2=\frac{p^2dvk^2}{(d,v)^2},\qquad k\geq1. w 5 : D 1 = ( d , v ) p d k , D 2 = ( d , v ) 2 p 2 d v k 2 , k ≥ 1.
Thus (119) gives
∣ y w 4 ∣ = d v ( b , e ) 3 p 2 b e k 3 ≪ Q p − 1 + ε k − 3 , and ∣ y w 5 ∣ = e b ( d , v ) 3 p d v k 3 ≪ Q L p − 1 / 2 + ε k − 3 . |y_{w_4}|=\frac{dv(b,e)^3}{p^2bek^3}\ll_Q p^{-1+\varepsilon}k^{-3},\qquad\text{and}\qquad|y_{w_5}|=\frac{eb(d,v)^3}{pdvk^3}\ll_Q Lp^{-1/2+\varepsilon}k^{-3}. ∣ y w 4 ∣ = p 2 b e k 3 d v ( b , e ) 3 ≪ Q p − 1 + ε k − 3 , and ∣ y w 5 ∣ = p d v k 3 e b ( d , v ) 3 ≪ Q L p − 1/2 + ε k − 3 .
Apply ∣ S ~ ∣ ≤ D 1 D 2 |\widetilde{S}|\le D_1D_2 ∣ S ∣ ≤ D 1 D 2 and the kernel bound above. In each of the four sums, the sum over k k k is bounded by ∑ k ≥ 1 k − 3 A ≪ A 1 \sum_{k\ge1}k^{-3A}\ll_A 1 ∑ k ≥ 1 k − 3 A ≪ A 1 . Using P ≍ p 2 P\asymp p^2 P ≍ p 2 and (120) , each first contribution is
≪ A , h , Q , J , ε p 5 / 2 + O ( ε ) L 2 + O ( ε ) ( p ε L 4 p 2 ) A . \ll_{A,h,Q,J,\varepsilon}p^{5/2+O(\varepsilon)}L^{2+O(\varepsilon)}\left(\frac{p^\varepsilon L^4}{p^2}\right)^A. ≪ A , h , Q , J , ε p 5/2 + O ( ε ) L 2 + O ( ε ) ( p 2 p ε L 4 ) A .
and each dual contribution, including its outside factor p p p , is
≪ A , h , Q , J , ε p 7 / 2 + O ( ε ) ( p − 1 + ε + L p − 1 / 2 + ε ) A . \ll_{A,h,Q,J,\varepsilon}p^{7/2+O(\varepsilon)}\left(p^{-1+\varepsilon}+Lp^{-1/2+\varepsilon}\right)^A. ≪ A , h , Q , J , ε p 7/2 + O ( ε ) ( p − 1 + ε + L p − 1/2 + ε ) A .
Since L = p η L=p^\eta L = p η and η < 1 / 2 \eta<1/2 η < 1/2 , both bounds are O h , Q , η , J ( p − 1 ) O_{h,Q,\eta,J}(p^{-1}) O h , Q , η , J ( p − 1 ) for A A A sufficiently large in terms of η \eta η , as ε \varepsilon ε is arbitrarily small. This proves the lemma. □
The long Weyl element
Lemma 5.6. The first and dual contributions to S 6 , 1 , 1 S_{6,1,1} S 6 , 1 , 1 are each O h , Q , η , J ( p − 1 ) O_{h,Q,\eta,J}(p^{-1}) O h , Q , η , J ( p − 1 ) .
Proof. Choose A > 1 A>1 A > 1 and the auxiliary parameters as in Lemma 5.5 . For the indices ( 1 , 1 , u , v ) (1,1,u,v) ( 1 , 1 , u , v ) from the first sum, we have
∣ y 1 2 y 2 ∣ = u w 2 D 1 3 , ∣ y 1 y 2 2 ∣ = u 2 v D 2 3 . |y_1^2y_2|=\frac{uw^2}{D_1^3},\qquad|y_1y_2^2|=\frac{u^2v}{D_2^3}. ∣ y 1 2 y 2 ∣ = D 1 3 u w 2 , ∣ y 1 y 2 2 ∣ = D 2 3 u 2 v .
By Remark 5 and (36) with c = 1 c=1 c = 1 , (54) with decay exponent A + 2 A+2 A + 2 and growth exponent one, and the bounds ∣ S ( p ) ∣ ≤ ( D 1 D 2 ) 2 |S^{(p)}|\le(D_1D_2)^2 ∣ S ( p ) ∣ ≤ ( D 1 D 2 ) 2 and 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ≤ 3 u 2 v D 1 D 2 1+|y_1|+|y_2|\le3u^2vD_1D_2 1 + ∣ y 1 ∣ + ∣ y 2 ∣ ≤ 3 u 2 v D 1 D 2 , the sum over D 1 , D 2 D_1,D_2 D 1 , D 2 for fixed m , n , u , v m,n,u,v m , n , u , v is bounded by
≪ A , h , J P p L 2 m n u 4 v 5 ( u 2 v ) A ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , D 2 D 1 − 4 D 2 2 − 3 A ≪ A , h , J p L 2 m n ( u 2 v ) 5 ( u 2 v p 3 ) A . \ll_{A,h,J}P\frac{pL^2}{mn}u^4v^5(u^2v)^A\sum_{\substack{D_1,D_2\ge1\\p\mid D_1,D_2}}D_1^{-4}D_2^{2-3A}
\ll_{A,h,J}\frac{pL^2}{mn}(u^2v)^5\left(\frac{u^2v}{p^3}\right)^A. ≪ A , h , J P mn p L 2 u 4 v 5 ( u 2 v ) A D 1 , D 2 ≥ 1 p ∣ D 1 , D 2 ∑ D 1 − 4 D 2 2 − 3 A ≪ A , h , J mn p L 2 ( u 2 v ) 5 ( p 3 u 2 v ) A .
By (117) and (120) , the total first contribution is
≪ A , h , Q , J , ε p C ( p ε L 4 p 2 ) A . \ll_{A,h,Q,J,\varepsilon}p^C\left(\frac{p^\varepsilon L^4}{p^2}\right)^A. ≪ A , h , Q , J , ε p C ( p 2 p ε L 4 ) A .
For the dual indices ( d , b , p 2 e , v ) (d,b,p^2e,v) ( d , b , p 2 e , v ) , twisted multiplicativity reduces the case ord p ( D 2 ) = 1 \operatorname{ord}_p(D_2)=1 ord p ( D 2 ) = 1 to local moduli ( p α , p ) (p^\alpha,p) ( p α , p ) , up to unit twists. The case α = 1 \alpha=1 α = 1 vanishes by the local identities in § 3.4 , since p ∤ d p\nmid d p ∤ d . For α ≥ 2 \alpha\ge2 α ≥ 2 , the same substitution B 1 = p β B_1=p^\beta B 1 = p β in (49) as in that subsection gives
S ( p ) ( ϵ 2 v , ϵ 1 p e , d , b , p α , p ) = r p ( ϵ 1 p e ) S ( d , ϵ 2 p v , p α ) = 0. S^{(p)}(\epsilon_2v,\epsilon_1p^e,d,b,p^\alpha,p)=r_p(\epsilon_1p^e)S(d,\epsilon_2pv,p^\alpha)=0. S ( p ) ( ϵ 2 v , ϵ 1 p e , d , b , p α , p ) = r p ( ϵ 1 p e ) S ( d , ϵ 2 p v , p α ) = 0.
The last equality follows by summing over x ↦ x + p α − 1 t x\mapsto x+p^{\alpha-1}t x ↦ x + p α − 1 t , t ( m o d p ) t\pmod p t ( mod p ) , using p ∤ d p\nmid d p ∤ d . Thus only p 2 ∣ D 2 p^2\mid D_2 p 2 ∣ D 2 contributes. The dual arguments and (119) give
∣ y 1 2 y 2 ∣ = p 2 d 2 b e v 2 D 1 3 , |y_1^2y_2|=\frac{p^2d^2bev^2}{D_1^3}, ∣ y 1 2 y 2 ∣ = D 1 3 p 2 d 2 b e v 2 ,
and
∣ y 1 y 2 2 ∣ = p 4 d b 2 e 2 v D 2 3 ≪ Q L 2 p − 1 + ε ( p 2 D 2 ) 3 . |y_1y_2^2|=\frac{p^4db^2e^2v}{D_2^3}\ll_Q L^2p^{-1+\varepsilon}\left(\frac{p^2}{D_2}\right)^3. ∣ y 1 y 2 2 ∣ = D 2 3 p 4 d b 2 e 2 v ≪ Q L 2 p − 1 + ε ( D 2 p 2 ) 3 .
The same summation over D 1 , D 2 D_1,D_2 D 1 , D 2 , now with p 2 ∣ D 2 p^2 \mid D_2 p 2 ∣ D 2 and test factor p / ( e 2 v L 2 ) p/(e^{2v}L^2) p / ( e 2 v L 2 ) from (36) , gives
≪ A , h , J p 9 L 2 d 5 b 3 e 4 v ( d b 2 e 2 v p 2 ) A . \ll_{A,h,J} \frac{p^9}{L^2}d^5b^3e^{4v}\left(\frac{db^2e^{2v}}{p^2}\right)^A. ≪ A , h , J L 2 p 9 d 5 b 3 e 4 v ( p 2 d b 2 e 2 v ) A .
By (119) and (120) , including the outside factor p p p , the total dual contribution is
≪ A , h , Q , J , ε p C ( L 2 p − 1 + ε ) A . \ll_{A,h,Q,J,\varepsilon} p^C\left(L^2p^{-1+\varepsilon}\right)^A. ≪ A , h , Q , J , ε p C ( L 2 p − 1 + ε ) A .
Here and above, C C C is independent of A A A . Since L = p η L=p^\eta L = p η and η < 1 / 2 \eta<1/2 η < 1/2 , taking A A A sufficiently large in terms of η \eta η makes both contributions O h , Q , η , J ( p − 1 ) O_{h,Q,\eta,J}(p^{-1}) O h , Q , η , J ( p − 1 ) , as ε \varepsilon ε is arbitrarily small. □
The old and minimal Eisenstein spectra
Proposition 5.7. For every ε > 0 \varepsilon>0 ε > 0 , we have
R o l d , 1 , 1 ≪ h , Q , η , J , ε p ε L 12 / 7 + ε , R_{\mathrm{old},1,1}\ll_{h,Q,\eta,J,\varepsilon}p^\varepsilon L^{12/7+\varepsilon}, R old , 1 , 1 ≪ h , Q , η , J , ε p ε L 12/7 + ε , R m i n , 1 , 1 ≪ h , Q , η , J , ε p ε L 1 + ε , R_{\mathrm{min},1,1}\ll_{h,Q,\eta,J,\varepsilon}p^\varepsilon L^{1+\varepsilon}, R min , 1 , 1 ≪ h , Q , η , J , ε p ε L 1 + ε , R m a x , o l d , 1 , 1 ≪ h , Q , η , J , ε p ε L 39 / 32 + ε . R_{\mathrm{max,old},1,1}\ll_{h,Q,\eta,J,\varepsilon}p^\varepsilon L^{39/32+\varepsilon}. R max , old , 1 , 1 ≪ h , Q , η , J , ε p ε L 39/32 + ε .
Proof. Let Π \Pi Π lie in the old cuspidal, minimal or maximal old spectrum, with θ = 5 / 14 , 0 , 7 / 64 \theta=5/14,0,7/64 θ = 5/14 , 0 , 7/64 , respectively. Define M Π M_\Pi M Π by (2) using the coefficients of Π \Pi Π . The coefficient bounds of § 2 give
∣ M Π ∣ 2 ≪ Q , ε ( ∑ ℓ ≤ L ℓ − 1 / 2 + θ + ε ) 2 ≪ Q , ε L 1 + 2 θ + 2 ε . (125) |M_\Pi|^2\ll_{Q,\varepsilon}\left(\sum_{\ell\le L}\ell^{-1/2+\theta+\varepsilon}\right)^2\ll_{Q,\varepsilon}L^{1+2\theta+2\varepsilon}. \tag*{(125)} ∣ M Π ∣ 2 ≪ Q , ε ( ℓ ≤ L ∑ ℓ − 1/2 + θ + ε ) 2 ≪ Q , ε L 1 + 2 θ + 2 ε . (125)
We compute the local series at p p p to recover the full squared L L L -function after summing over the oldvectors. Insert (32) and put s = 1 / 2 + u s=1/2+u s = 1/2 + u on the initial Mellin line Re ( u ) > 1 \operatorname{Re}(u)>1 Re ( u ) > 1 . Writing m = p k m ′ m=p^km' m = p k m ′ and n = p ℓ n ′ n=p^\ell n' n = p ℓ n ′ with ( m ′ n ′ , p ) = 1 (m'n',p)=1 ( m ′ n ′ , p ) = 1 , (113) gives the local series
T j ( s ) : = ∑ k , ℓ ≥ 0 p − ( k + ℓ ) s ∑ r = 0 min ( k , ℓ ) A Π ( j ) ( p r , p k + ℓ − 2 r ) , j = 0 , 1 , 2. (126) T_j(s):=\sum_{k,\ell\geq0}p^{-(k+\ell)s}\sum_{r=0}^{\min(k,\ell)}A_\Pi^{(j)}(p^r,p^{k+\ell-2r}),\qquad j=0,1,2. \tag*{(126)} T j ( s ) := k , ℓ ≥ 0 ∑ p − ( k + ℓ ) s r = 0 ∑ m i n ( k , ℓ ) A Π ( j ) ( p r , p k + ℓ − 2 r ) , j = 0 , 1 , 2. (126)
By (40) and (113) , (126) gives
T 0 ( s ) = L p ( s , Π ) 2 , T 1 ( s ) = 2 p 1 / 2 − s L p ( s , Π ) 2 , T 2 ( s ) = p 1 − 2 s L p ( s , Π ) 2 . (127) T_0(s)=L_p(s,\Pi)^2,\qquad T_1(s)=2p^{1/2-s}L_p(s,\Pi)^2,\qquad T_2(s)=p^{1-2s}L_p(s,\Pi)^2. \tag*{(127)} T 0 ( s ) = L p ( s , Π ) 2 , T 1 ( s ) = 2 p 1/2 − s L p ( s , Π ) 2 , T 2 ( s ) = p 1 − 2 s L p ( s , Π ) 2 . (127)
We now sum the first Dirichlet series over the three oldvectors. By (115) , the primes other than p p p contribute M Π 2 L ( p ) ( s , Π ) 2 M_\Pi^{2}L^{(p)}(s,\Pi)^2 M Π 2 L ( p ) ( s , Π ) 2 . The oldspace summation in the proof of Lemma 4.4 and (127) therefore give
∑ m , n ≥ 1 1 ( m n ) s ∑ d , b ≥ 1 b m , n ( d , b ) ∑ j = 0 2 B Π ( j ) ( 1 , 1 ) ‾ B Π ( j ) ( d , b ) = M Π 2 L ( p ) ( s , Π ) 2 ∑ j = 0 2 ( Q Π − 1 ) 0 j T j ( s ) = M Π 2 L ( s , Π ) 2 ( ( Q Π − 1 ) 00 + 2 p 1 / 2 − s ( Q Π − 1 ) 01 + p 1 − 2 s ( Q Π − 1 ) 02 ) . (128) \begin{aligned}
\sum_{m,n\geq1}\frac{1}{(mn)^s}\sum_{d,b\geq1}b_{m,n}(d,b)\sum_{j=0}^{2}\overline{B_\Pi^{(j)}(1,1)}B_\Pi^{(j)}(d,b) \\
&=M_\Pi^2L^{(p)}(s,\Pi)^2\sum_{j=0}^{2}(Q_\Pi^{-1})_{0j}T_j(s) \\
&=M_\Pi^2L(s,\Pi)^2\left((Q_\Pi^{-1})_{00}+2p^{1/2-s}(Q_\Pi^{-1})_{01}+p^{1-2s}(Q_\Pi^{-1})_{02}\right). \tag*{(128)}
\end{aligned} m , n ≥ 1 ∑ ( mn ) s 1 d , b ≥ 1 ∑ b m , n ( d , b ) j = 0 ∑ 2 B Π ( j ) ( 1 , 1 ) B Π ( j ) ( d , b ) = M Π 2 L ( p ) ( s , Π ) 2 j = 0 ∑ 2 ( Q Π − 1 ) 0 j T j ( s ) = M Π 2 L ( s , Π ) 2 ( ( Q Π − 1 ) 00 + 2 p 1/2 − s ( Q Π − 1 ) 01 + p 1 − 2 s ( Q Π − 1 ) 02 ) . (128)
Taking s = 1 / 2 + u s=1/2+u s = 1/2 + u in (128) , the factor in parentheses is O ( 1 ) O(1) O ( 1 ) on Re ( u ) = ε \operatorname{Re}(u)=\varepsilon Re ( u ) = ε by Lemma 4.4. For fixed Π \Pi Π , the Mellin integral multiplying h ( μ Π ) M Π 2 h(\mu_\Pi)M_\Pi^2 h ( μ Π ) M Π 2 in the first contribution is
1 2 π i ∫ ( c ) ( p L 2 ) u G J ( u , μ Π ) 2 ( γ μ Π ( 1 / 2 + u ) γ μ Π ( 1 / 2 ) ) 2 L ( 1 / 2 + u , Π ) 2 ( ( Q Π − 1 ) 00 + 2 p − u ( Q Π − 1 ) 01 + p − 2 u ( Q Π − 1 ) 02 ) d u u . (129) \frac{1}{2\pi i}\int_{(c)}(pL^2)^uG_J(u,\mu_\Pi)^2\left(\frac{\gamma_{\mu_\Pi}(1/2+u)}{\gamma_{\mu_\Pi}(1/2)}\right)^2L(1/2+u,\Pi)^2\left((Q_\Pi^{-1})_{00}+2p^{-u}(Q_\Pi^{-1})_{01}+p^{-2u}(Q_\Pi^{-1})_{02}\right)\frac{du}{u}. \tag*{(129)} 2 π i 1 ∫ ( c ) ( p L 2 ) u G J ( u , μ Π ) 2 ( γ μ Π ( 1/2 ) γ μ Π ( 1/2 + u ) ) 2 L ( 1/2 + u , Π ) 2 ( ( Q Π − 1 ) 00 + 2 p − u ( Q Π − 1 ) 01 + p − 2 u ( Q Π − 1 ) 02 ) u d u . (129)
For each of the three level one spectra, move the u u u -contour in (129) from Re ( u ) = c > 1 \operatorname{Re}(u)=c>1 Re ( u ) = c > 1 to Re ( u ) = ε \operatorname{Re}(u)=\varepsilon Re ( u ) = ε without a residue. By (21) , P J ( u , μ ) 2 P_J(u,\mu)^2 P J ( u , μ ) 2 cancels the Eisenstein poles with multiplicity. The level one convexity bound [27 Theorem 5.41] , (19) , (18) , (70) and (71) , together with the rapid decrease of h h h , bound the remaining spectral and Mellin integrals independently of p p p . Combining (125) with ∣ ( p L 2 ) u ∣ = ( p L 2 ) ε |(pL^2)^u|=(pL^2)^\varepsilon ∣ ( p L 2 ) u ∣ = ( p L 2 ) ε and rescaling ε \varepsilon ε , the first contribution is
≪ h , Q , η , J , ε p ε L 1 + 2 θ + ε . (130) \ll_{h,Q,\eta,J,\varepsilon}p^\varepsilon L^{1+2\theta+\varepsilon}. \tag*{(130)} ≪ h , Q , η , J , ε p ε L 1 + 2 θ + ε . (130)
For the dual contribution, the Fourier pairs are ( d , b ) (d,b) ( d , b ) and ( p 2 e , v ) (p^2e,v) ( p 2 e , v ) , with ( d b e v , p ) = 1 (dbev,p)=1 ( d b e v , p ) = 1 , by (119) . We first combine the outside factor p p p with the sum over the three oldvectors. Put λ = A Π ( 1 , p ) \lambda=A_\Pi(1,p) λ = A Π ( 1 , p ) and e 0 = ( 1 , 0 , 0 ) T e_0=(1,0,0)^{\mathsf T} e 0 = ( 1 , 0 , 0 ) T . The unramified Hecke relations give A Π ( p 2 , 1 ) = λ ˉ 2 − λ A_\Pi(p^2,1)=\bar{\lambda}^2-\lambda A Π ( p 2 , 1 ) = λ ˉ 2 − λ and A Π ( p , p ) = ∣ λ ∣ 2 − 1 A_\Pi(p,p)=|\lambda|^2-1 A Π ( p , p ) = ∣ λ ∣ 2 − 1 . Hence (40) gives the coefficient column at ( p 2 , 1 ) (p^2,1) ( p 2 , 1 ) as
( A Π ( 0 ) ( p 2 , 1 ) A Π ( 1 ) ( p 2 , 1 ) A Π ( 2 ) ( p 2 , 1 ) ) = ( λ ˉ 2 − λ p ( ∣ λ ∣ 2 − 1 ) p λ ) . \begin{pmatrix} A_\Pi^{(0)}(p^2,1) \\ A_\Pi^{(1)}(p^2,1) \\ A_\Pi^{(2)}(p^2,1) \end{pmatrix}
=
\begin{pmatrix} \bar{\lambda}^2-\lambda\\ \sqrt{p}(|\lambda|^2-1) \\ p\lambda\end{pmatrix}. A Π ( 0 ) ( p 2 , 1 ) A Π ( 1 ) ( p 2 , 1 ) A Π ( 2 ) ( p 2 , 1 ) = λ ˉ 2 − λ p ( ∣ λ ∣ 2 − 1 ) p λ .
Since ( d b e v , p ) = 1 (dbev,p)=1 ( d b e v , p ) = 1 , the coefficient columns at ( d , b ) (d,b) ( d , b ) and ( p 2 e , v ) (p^2e,v) ( p 2 e , v ) are A Π ( d , b ) e 0 A_\Pi(d,b)e_0 A Π ( d , b ) e 0 and A Π ( e , v ) A_\Pi(e,v) A Π ( e , v ) times the displayed column, respectively. The oldspace summation in the proof of Lemma 4.4 therefore gives
p ∑ j = 0 2 B Π ( j ) ( d , b ) ‾ B Π ( j ) ( p 2 e , v ) = A Π ( d , b ) A Π ( e , v ) p ∑ j = 0 2 ( Q Π − 1 ) 0 j A Π ( j ) ( p 2 , 1 ) . p\sum_{j=0}^{2}\overline{B_\Pi^{(j)}(d,b)}B_\Pi^{(j)}(p^2e,v)=A_\Pi(d,b)A_\Pi(e,v)p\sum_{j=0}^{2}(Q_\Pi^{-1})_{0j}A_\Pi^{(j)}(p^2,1). p j = 0 ∑ 2 B Π ( j ) ( d , b ) B Π ( j ) ( p 2 e , v ) = A Π ( d , b ) A Π ( e , v ) p j = 0 ∑ 2 ( Q Π − 1 ) 0 j A Π ( j ) ( p 2 , 1 ) .
Using (41) and the identities a p 2 = b p + b p 2 a_p^2=b_p+b_p^2 a p 2 = b p + b p 2 , p a p = ( p + 1 ) b p \sqrt{p}\,a_p=(p+1)b_p p a p = ( p + 1 ) b p and p ( p + 1 ) b p = p − b p p(p+1)b_p=p-b_p p ( p + 1 ) b p = p − b p , direct multiplication yields
p ∑ j = 0 2 ( Q Π − 1 ) 0 j A Π ( j ) ( p 2 , 1 ) = a p 2 λ ˉ 2 − b p λ ( 1 + b p ∣ λ ∣ 2 ) det ( Q Π ) = ( Q Π − 1 ) 20 . (131) p\sum_{j=0}^{2}(Q_\Pi^{-1})_{0j}A_\Pi^{(j)}(p^2,1)=\frac{a_p^2\bar{\lambda}^2-b_p\lambda(1+b_p|\lambda|^2)}{\det(Q_\Pi)}=(Q_\Pi^{-1})_{20}. \tag*{(131)} p j = 0 ∑ 2 ( Q Π − 1 ) 0 j A Π ( j ) ( p 2 , 1 ) = det ( Q Π ) a p 2 λ ˉ 2 − b p λ ( 1 + b p ∣ λ ∣ 2 ) = ( Q Π − 1 ) 20 . (131)
The last equality is the cofactor formula. Since a p 2 , b p ≪ p − 1 a_p^2,b_p\ll p^{-1} a p 2 , b p ≪ p − 1 , ∣ λ ∣ ≪ p θ |\lambda|\ll p^\theta ∣ λ ∣ ≪ p θ and det ( Q Π ) ≍ 1 \det(Q_\Pi)\asymp1 det ( Q Π ) ≍ 1 by Lemma 4.4, the combined local factor, including the outside p p p , satisfies
∣ ( Q Π − 1 ) 20 ∣ ≪ p 2 θ − 1 + p θ − 1 + p 3 θ − 2 ≪ p 2 θ − 1 . |(Q_\Pi^{-1})_{20}|\ll p^{2\theta-1}+p^{\theta-1}+p^{3\theta-2}\ll p^{2\theta-1}. ∣ ( Q Π − 1 ) 20 ∣ ≪ p 2 θ − 1 + p θ − 1 + p 3 θ − 2 ≪ p 2 θ − 1 .
The remaining dual indices are coprime to p p p . Thus the dual Dirichlet series is M Π ( p ) ( 1 / 2 + u , Π ) 2 ‾ \overline{M_\Pi^{(p)}(1/2+u,\Pi)^2} M Π ( p ) ( 1/2 + u , Π ) 2 multiplied by ( Q Π − 1 ) 20 (Q_\Pi^{-1})_{20} ( Q Π − 1 ) 20 . The removed Euler factors are uniformly bounded on Re ( u ) = ε \operatorname{Re}(u)=\varepsilon Re ( u ) = ε , since θ < 1 / 2 \theta<1/2 θ < 1/2 . Using V − μ Π V_{-\mu_\Pi} V − μ Π in (32) , the dual u u u -integral is
1 2 π i ∫ ( c ) ( p L 2 ) u G J ( u , μ Π ) 2 ( γ μ Π ( 1 / 2 + u ) γ − μ Π ( 1 / 2 ) ) 2 L ( p ) ( 1 / 2 + u , Π ) 2 ‾ d u u . (132) \frac{1}{2\pi i}\int_{(c)}\left(\frac{p}{L^2}\right)^uG_J(u,\mu_\Pi)^2\left(\frac{\gamma_{\mu_\Pi}(1/2+u)}{\gamma_{-\mu_\Pi}(1/2)}\right)^2\overline{L^{(p)}(1/2+u,\Pi)^2}\frac{du}{u}. \tag*{(132)} 2 π i 1 ∫ ( c ) ( L 2 p ) u G J ( u , μ Π ) 2 ( γ − μ Π ( 1/2 ) γ μ Π ( 1/2 + u ) ) 2 L ( p ) ( 1/2 + u , Π ) 2 u d u . (132)
Move the u u u -contour in (132) from Re ( u ) = c > 1 \operatorname{Re}(u)=c>1 Re ( u ) = c > 1 to Re ( u ) = ε \operatorname{Re}(u)=\varepsilon Re ( u ) = ε in the same way. Combining this with (125) and (131) gives
≪ h , Q , η , J , ε p 2 θ − 1 + ε L 1 + 2 θ + ε . \ll_{h,Q,\eta,J,\varepsilon}p^{2\theta-1+\varepsilon}L^{1+2\theta+\varepsilon}. ≪ h , Q , η , J , ε p 2 θ − 1 + ε L 1 + 2 θ + ε .
Since 2 θ − 1 < 0 2\theta-1<0 2 θ − 1 < 0 , the first contribution (130) dominates. Taking θ = 5 / 14 , 0 , 7 / 64 \theta=5/14,0,7/64 θ = 5/14 , 0 , 7/64 proves the three estimates. □ \square □
The maximal Eisenstein new spectrum
Proposition 5.8. For every ε > 0 \varepsilon>0 ε > 0 , the first and dual contributions to R max , n e w , 1 , 1 R_{\max,\mathrm{new},1,1} R m a x , new , 1 , 1 are each O h , Q , η , J , ε ( p 3 / 2 + ε L 1 + ε ) O_{h,Q,\eta,J,\varepsilon}(p^{3/2+\varepsilon}L^{1+\varepsilon}) O h , Q , η , J , ε ( p 3/2 + ε L 1 + ε ) .
Proof. Let Π \Pi Π be the maximal Eisenstein representation associated with ϕ ∈ B 2 ( p ) \phi\in B_2(p) ϕ ∈ B 2 ( p ) and z ∈ i R z\in i\mathbb{R} z ∈ i R in § 3.3.3 , with μ = ( z + τ ϕ , z − τ ϕ , − 2 z ) \mu=(z+\tau_\phi,z-\tau_\phi,-2z) μ = ( z + τ ϕ , z − τ ϕ , − 2 z ) . Define M Π M_\Pi M Π by (1.2) using the coefficients A ϕ , z A_{\phi,z} A ϕ , z . On the initial line Re ( u ) = c > 1 \operatorname{Re}(u)=c>1 Re ( u ) = c > 1 , (115) and § 3.3.3 give
∑ m , n ≥ 1 1 ( m n ) 1 / 2 + u ∑ d , b ≥ 1 b m , n ( d , b ) A ϕ , z ( d , b ) = M Π 2 L ( 1 / 2 + u − z , ϕ ) 2 ζ ( 1 / 2 + u + 2 z ) 2 . (133) \sum_{m,n\geq1}\frac{1}{(mn)^{1/2+u}}\sum_{d,b\geq1}b_{m,n}(d,b)A_{\phi,z}(d,b)
=M_\Pi^2L(1/2+u-z,\phi)^2\zeta(1/2+u+2z)^2. \tag*{(133)} m , n ≥ 1 ∑ ( mn ) 1/2 + u 1 d , b ≥ 1 ∑ b m , n ( d , b ) A ϕ , z ( d , b ) = M Π 2 L ( 1/2 + u − z , ϕ ) 2 ζ ( 1/2 + u + 2 z ) 2 . (133)
Insert (133) into the Mellin representation (32) . By (21) , the factor ( ( 1 / 2 + μ 3 ) 2 − u 2 ) 2 ((1/2+\mu_3)^2-u^2)^2 (( 1/2 + μ 3 ) 2 − u 2 ) 2 in P J ( u , μ ) 2 P_J(u,\mu)^2 P J ( u , μ ) 2 cancels the double zeta pole at u = 1 / 2 − 2 z u=1/2-2z u = 1/2 − 2 z . The G L 2 \mathrm{GL}_2 GL 2 L L L -function is entire, and the numerator gamma factors have no poles for Re ( u ) > 0 \operatorname{Re}(u)>0 Re ( u ) > 0 , since ∣ Re ( μ j ) ∣ ≤ 7 / 64 |\operatorname{Re}(\mu_j)|\leq7/64 ∣ Re ( μ j ) ∣ ≤ 7/64 . (18) and (19) , together with e 2 u 2 e^{2u^2} e 2 u 2 , make the horizontal integrals tend to zero. We may therefore move the u u u -contour to Re ( u ) = ε \operatorname{Re}(u)=\varepsilon Re ( u ) = ε without a residue.
On this line, the G L 2 \mathrm{GL}_2 GL 2 convexity bound [27 Theorem 5.41] gives
∣ L ( 1 / 2 + u − z , ϕ ) ∣ 2 ≪ ε p 1 / 2 + ε ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ + ∣ Im ( u ) ∣ ) C . |L(1/2+u-z,\phi)|^2\ll_\varepsilon p^{1/2+\varepsilon}(1+|z|+|\tau_\phi|+|\operatorname{Im}(u)|)^C. ∣ L ( 1/2 + u − z , ϕ ) ∣ 2 ≪ ε p 1/2 + ε ( 1 + ∣ z ∣ + ∣ τ ϕ ∣ + ∣ Im ( u ) ∣ ) C .
We next estimate the spectral average of ∣ M Π ∣ 2 |M_\Pi|^2 ∣ M Π ∣ 2 . The finite Euler product in (2) is O Q ( 1 ) O_Q(1) O Q ( 1 ) by the coefficient bounds of § 2 and the induced parameters of § 3.3.3 . Thus
∣ M Π ∣ 2 ≪ Q ∑ ℓ 1 , ℓ 2 ≤ L ∣ x ℓ 1 x ℓ 2 ∣ ℓ 1 ℓ 2 ∣ A ϕ , z ( 1 , ℓ 1 ) A ϕ , z ( 1 , ℓ 2 ) ∣ . |M_\Pi|^2\ll_Q\sum_{\ell_1,\ell_2\leq L}\frac{|x_{\ell_1}x_{\ell_2}|}{\sqrt{\ell_1\ell_2}}|A_{\phi,z}(1,\ell_1)A_{\phi,z}(1,\ell_2)|. ∣ M Π ∣ 2 ≪ Q ℓ 1 , ℓ 2 ≤ L ∑ ℓ 1 ℓ 2 ∣ x ℓ 1 x ℓ 2 ∣ ∣ A ϕ , z ( 1 , ℓ 1 ) A ϕ , z ( 1 , ℓ 2 ) ∣.
Apply (73) with ( a , b , m , n ) = ( 1 , ℓ 1 , 1 , ℓ 2 ) (a,b,m,n)=(1,\ell_1,1,\ell_2) ( a , b , m , n ) = ( 1 , ℓ 1 , 1 , ℓ 2 ) . Since ℓ i ≤ L < p \ell_i\leq L<p ℓ i ≤ L < p , all indices are coprime to p p p and ( 1 + ℓ 1 / p ) ( 1 + ℓ 2 / p ) ≪ 1 (1+\ell_1/p)(1+\ell_2/p)\ll1 ( 1 + ℓ 1 / p ) ( 1 + ℓ 2 / p ) ≪ 1 . For every fixed C ≥ 0 C\geq0 C ≥ 0 , using ∣ x ℓ ∣ ≤ 1 |x_\ell|\leq1 ∣ x ℓ ∣ ≤ 1 gives
∑ ϕ ∈ B 2 ( p ) 1 L ( 1 , ϕ , Ad ) ∫ Re ( z ) = 0 ∣ h ( μ ) ∣ ∣ M Π ∣ 2 ( 1 + ∥ μ ∥ ) C ∣ L ( 1 + 3 z , ϕ ) ∣ 2 ∣ d z ∣ ≪ h , Q , C , ε p 1 + ε ( ∑ ℓ ≤ L ∣ x ℓ ∣ ℓ − 1 / 2 + ε ) 2 ≪ h , Q , C , ε p 1 + ε L 1 + 2 ε . (134) \sum_{\phi\in B_2(p)}\frac{1}{L(1,\phi,\operatorname{Ad})}\int_{\operatorname{Re}(z)=0}\frac{|h(\mu)||M_\Pi|^2(1+\|\mu\|)^C}{|L(1+3z,\phi)|^2}|dz|
\ll_{h,Q,C,\varepsilon}p^{1+\varepsilon}\left(\sum_{\ell\leq L}|x_\ell|\ell^{-1/2+\varepsilon}\right)^2
\ll_{h,Q,C,\varepsilon}p^{1+\varepsilon}L^{1+2\varepsilon}. \tag*{(134)} ϕ ∈ B 2 ( p ) ∑ L ( 1 , ϕ , Ad ) 1 ∫ Re ( z ) = 0 ∣ L ( 1 + 3 z , ϕ ) ∣ 2 ∣ h ( μ ) ∣∣ M Π ∣ 2 ( 1 + ∥ μ ∥ ) C ∣ d z ∣ ≪ h , Q , C , ε p 1 + ε ( ℓ ≤ L ∑ ∣ x ℓ ∣ ℓ − 1/2 + ε ) 2 ≪ h , Q , C , ε p 1 + ε L 1 + 2 ε . (134)
The estimates (18) and (19) and the factor e 2 u 2 e^{2u^2} e 2 u 2 control the remaining integration in Im ( u ) \operatorname{Im}(u) Im ( u ) by a fixed polynomial in 1 + ∥ μ ∥ 1+\|\mu\| 1 + ∥ μ ∥ . Since ∣ ( p L 2 ) ε ∣ = ( p L 2 ) ε |(pL^2)^{\varepsilon}|=(pL^2)^\varepsilon ∣ ( p L 2 ) ε ∣ = ( p L 2 ) ε and χ p ≪ 1 \chi_p\ll1 χ p ≪ 1 , the convexity bound and (134) bound the first contribution by
≪ h , Q , η , J , ε p 1 / 2 + ε ( p L 2 ) ε p 1 + ε L 1 + 2 ε = p 3 / 2 + 3 ε L 1 + 4 ε . \ll_{h,Q,\eta,J,\varepsilon}p^{1/2+\varepsilon}(pL^2)^\varepsilon p^{1+\varepsilon}L^{1+2\varepsilon}=p^{3/2+3\varepsilon}L^{1+4\varepsilon}. ≪ h , Q , η , J , ε p 1/2 + ε ( p L 2 ) ε p 1 + ε L 1 + 2 ε = p 3/2 + 3 ε L 1 + 4 ε .
For the dual contribution, (8) and § 3.3.3 give
p A ϕ , z ( p 2 , 1 ) = p λ ϕ ( p ) 2 p − 2 z = p − 2 z , ∣ p − 2 z ∣ = 1. pA_{\phi,z}(p^2,1)=p\lambda_{\phi}(p)^2p^{-2z}=p^{-2z},\qquad\lvert p^{-2z}\rvert=1. p A ϕ , z ( p 2 , 1 ) = p λ ϕ ( p ) 2 p − 2 z = p − 2 z , ∣ p − 2 z ∣ = 1.
By (119) , all remaining dual indices are coprime to p p p . Using (115) with m = n = 1 m=n=1 m = n = 1 and (113) , the dual Dirichlet series is
p − 2 z M Π ‾ 2 L ( p ) ( 1 2 + u − z , ϕ ) 2 ζ ( p ) ( 1 2 + u + 2 z ) 2 . p^{-2z}\overline{M_{\Pi}}^2 L^{(p)}\left(\frac{1}{2}+u-z,\phi\right)^2\zeta^{(p)}\left(\frac{1}{2}+u+2z\right)^2. p − 2 z M Π 2 L ( p ) ( 2 1 + u − z , ϕ ) 2 ζ ( p ) ( 2 1 + u + 2 z ) 2 .
Compared with the full L L L -functions in (133) , the omitted Euler factors contribute
∣ ( 1 − λ ϕ ( p ) p − 1 / 2 − u + z ) 2 ( 1 − p − 1 / 2 − u − 2 z ) 2 ∣ ≤ ( 1 + p − 1 − ε ) 2 ( 1 + p − 1 / 2 − ε ) 2 ≪ 1 ( Re ( u ) = ε ) . \left|\left(1-\lambda_{\phi}(p)p^{-1/2-u+z}\right)^2\left(1-p^{-1/2-u-2z}\right)^2\right|
\leq\left(1+p^{-1-\varepsilon}\right)^2\left(1+p^{-1/2-\varepsilon}\right)^2\ll1\qquad(\operatorname{Re}(u)=\varepsilon). ( 1 − λ ϕ ( p ) p − 1/2 − u + z ) 2 ( 1 − p − 1/2 − u − 2 z ) 2 ≤ ( 1 + p − 1 − ε ) 2 ( 1 + p − 1/2 − ε ) 2 ≪ 1 ( Re ( u ) = ε ) .
For the dual weight V − μ \mathcal{V}_{-\mu} V − μ , the numerator gamma factor is again γ μ ( 1 / 2 + u ) 2 \gamma_{\mu}(1/2+u)^2 γ μ ( 1/2 + u ) 2 , by (32) . The same pole cancellation therefore permits the contour shift to Re ( u ) = ε \operatorname{Re}(u)=\varepsilon Re ( u ) = ε . Now ∣ M Π 2 ∣ = ∣ M Π ∣ 2 \lvert M_{\Pi}^2\rvert=\lvert M_{\Pi}\rvert^2 ∣ M Π 2 ∣ = ∣ M Π ∣ 2 , and the length factor is ( p / L 2 ) ε ≤ ( p L 2 ) ε (p/L^2)^\varepsilon\leq(pL^2)^\varepsilon ( p / L 2 ) ε ≤ ( p L 2 ) ε . The convexity bound and (134) , together with the preceding uniform bound for the omitted Euler factors, give
≪ h , Q , η , J , ε p 1 / 2 + ε ( p / L 2 ) ε p 1 + ε L 1 + 2 ε = p 3 / 2 + 3 ε L . \ll_{h,Q,\eta,J,\varepsilon}p^{1/2+\varepsilon}(p/L^2)^\varepsilon p^{1+\varepsilon}L^{1+2\varepsilon}=p^{3/2+3\varepsilon}L. ≪ h , Q , η , J , ε p 1/2 + ε ( p / L 2 ) ε p 1 + ε L 1 + 2 ε = p 3/2 + 3 ε L .
This proves the proposition. □
Proof of Theorem 5.4
Insert (124) , Lemmas 5.5 and 5.6, and Propositions 5.7 and 5.8 into (123) . Fix the auxiliary parameters in terms of η \eta η as required by the preceding lemmas. Since L = p η L=p^\eta L = p η , η < 1 / 2 \eta<1/2 η < 1/2 and ε \varepsilon ε is arbitrarily small, the total spectral remainder is
O h , Q , η , J , ε ( p 3 / 2 + ε L 1 + ε ) = O h , Q , η , J ( p 2 log p ) . O_{h,Q,\eta,J,\varepsilon}\left(p^{3/2+\varepsilon}L^{1+\varepsilon}\right)=O_{h,Q,\eta,J}\left(\frac{p^2}{\log p}\right). O h , Q , η , J , ε ( p 3/2 + ε L 1 + ε ) = O h , Q , η , J ( log p p 2 ) .
Division by χ p = P / p 2 \chi_p=P/p^2 χ p = P / p 2 replaces the diagonal factor P P P by p 2 p^2 p 2 . Replacing p 2 p^2 p 2 by P P P introduces an error O h , Q ( p ) O_{h,Q}(p) O h , Q ( p ) . The fixed auxiliary parameters are absorbed into the dependence on η \eta η . This proves (111) . Each Euler factor is positive and at most 1, and differs from 1 by O ( q − 5 / 2 ) O(q^{-5/2}) O ( q − 5/2 ) . The final assertions follow.
Proof of Theorem 1.1
We now combine the mollified moment estimates of § 5 with the total harmonic mass. The functional equation relates the moments of L ( 1 / 2 , π ) M π ‾ L(1/2,\pi)\overline{M_{\pi}} L ( 1/2 , π ) M π to those of L ( 1 / 2 , π ) M π L(1/2,\pi)M_{\pi} L ( 1/2 , π ) M π . The Cauchy–Schwarz inequality then gives the following lower bound for the harmonic non-vanishing proportion.
Theorem 6.1. Let h h h be the fixed non-negative test function in (26) . For every ε > 0 \varepsilon>0 ε > 0 and all sufficiently large primes p p p , we have
c c u s p ∑ π ∈ B ( p ) L ( 1 / 2 , π ) ≠ 0 h ( μ π ) L ( 1 , π , Ad ) ≥ ( 1 2 − ε ) c c u s p ∑ π ∈ B ( p ) h ( μ π ) L ( 1 , π , Ad ) . (135) c_{\mathrm{cusp}}\sum_{\substack{\pi\in\mathcal{B}(p)\\L(1/2,\pi)\ne0}}\frac{h(\mu_\pi)}{L(1,\pi,\operatorname{Ad})}
\geq\left(\frac{1}{2}-\varepsilon\right)c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(\mu_\pi)}{L(1,\pi,\operatorname{Ad})}. \tag*{(135)} c cusp π ∈ B ( p ) L ( 1/2 , π ) = 0 ∑ L ( 1 , π , Ad ) h ( μ π ) ≥ ( 2 1 − ε ) c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) . (135) In particular the non-vanishing harmonic mass is ≫ h p 2 \gg_h p^2 ≫ h p 2 .
The total harmonic mass
We first evaluate the total harmonic mass.
Lemma 6.2. For every ε > 0 \varepsilon> 0 ε > 0 , we have
c c u s p ∑ π ∈ B ( p ) h ( μ π ) L ( 1 , π , Ad ) = P H h + O h , ε ( p 1 + ε ) . (136) c_{\mathrm{cusp}} \sum_{\pi\in\mathcal{B}(p)} \frac{h(\mu_\pi)}{L(1,\pi,\operatorname{Ad})} = P\mathcal{H}_h + O_{h,\varepsilon}(p^{1+\varepsilon}). \tag*{(136)} c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) = P H h + O h , ε ( p 1 + ε ) . (136) In particular, this sum is ≫ h p 2 \gg_h p^2 ≫ h p 2 for sufficiently large p p p .
Proof. Apply Theorem 3.1 with Fourier indices ( m 1 , m 2 , n 1 , n 2 ) = ( 1 , 1 , 1 , 1 ) (m_1,m_2,n_1,n_2)=(1,1,1,1) ( m 1 , m 2 , n 1 , n 2 ) = ( 1 , 1 , 1 , 1 ) and test function h h h . The diagonal term is P H h P\mathcal{H}_h P H h . For the short Weyl terms, the modulus pairs are ( d 2 , d ) (d^2,d) ( d 2 , d ) and ( d , d 2 ) (d,d^2) ( d , d 2 ) with p ∣ d p\mid d p ∣ d . The kernel arguments have modulus d − 3 d^{-3} d − 3 . For the long Weyl term, we have ∣ y 1 2 y 2 ∣ = D 1 − 3 |y_1^2y_2|=D_1^{-3} ∣ y 1 2 y 2 ∣ = D 1 − 3 and ∣ y 1 y 2 2 ∣ = D 2 − 3 |y_1y_2^2|=D_2^{-3} ∣ y 1 y 2 2 ∣ = D 2 − 3 . Using the trivial Kloosterman bounds ∣ S ∣ ≤ D 1 D 2 |S| \leq D_1D_2 ∣ S ∣ ≤ D 1 D 2 and ∣ S ( p ) ∣ ≤ ( D 1 D 2 ) 2 |S^{(p)}| \leq(D_1D_2)^2 ∣ S ( p ) ∣ ≤ ( D 1 D 2 ) 2 , (53) and (54) give
∣ S 4 ∣ + ∣ S 5 ∣ ≪ h , A P ∑ d ≥ 1 p ∣ d d − 3 A ≪ h , A p 2 − 3 A , |S_4|+|S_5| \ll_{h,A} P\sum_{\substack{d\geq1\\p\mid d}}d^{-3A}\ll_{h,A}p^{2-3A}, ∣ S 4 ∣ + ∣ S 5 ∣ ≪ h , A P d ≥ 1 p ∣ d ∑ d − 3 A ≪ h , A p 2 − 3 A , and
∣ S 6 ∣ ≪ h , A , ε P ∑ D 1 , D 2 ≥ 1 p ∣ D 1 , D 2 ( D 1 D 2 ) 1 + ε max ( D 1 , D 2 ) − 3 A ≪ h , A , ε p 4 − 3 A + 2 ε |S_6|\ll_{h,A,\varepsilon}P\sum_{\substack{D_1,D_2\geq1\\p\mid D_1,D_2}}(D_1D_2)^{1+\varepsilon}\max(D_1,D_2)^{-3A}\ll_{h,A,\varepsilon}p^{4-3A+2\varepsilon} ∣ S 6 ∣ ≪ h , A , ε P D 1 , D 2 ≥ 1 p ∣ D 1 , D 2 ∑ ( D 1 D 2 ) 1 + ε max ( D 1 , D 2 ) − 3 A ≪ h , A , ε p 4 − 3 A + 2 ε for sufficiently large A A A . Taking A A A sufficiently large makes all three terms O h , ε ( p − 1 ) O_{h,\varepsilon}(p^{-1}) O h , ε ( p − 1 ) . For each level one sector, (40) and the oldspace summation in Lemma 4.4 give
∑ j = 0 2 ∣ B Π ( j ) ( 1 , 1 ) ∣ 2 = ( Q Π − 1 ) 00 ≪ 1. \sum_{j=0}^{2}|B_{\Pi}^{(j)}(1,1)|^2=(Q_{\Pi}^{-1})_{00}\ll1. j = 0 ∑ 2 ∣ B Π ( j ) ( 1 , 1 ) ∣ 2 = ( Q Π − 1 ) 00 ≪ 1. Together with (70) , (71) and the rapid decrease of h h h , this yields
∣ C o l d ∣ + ∣ E min ∣ + ∣ E max , o l d ∣ ≪ h 1. |C_{\mathrm{old}}|+|\mathcal{E}_{\min}|+|\mathcal{E}_{\max,\mathrm{old}}|\ll_h 1. ∣ C old ∣ + ∣ E m i n ∣ + ∣ E m a x , old ∣ ≪ h 1. The maximal new coefficient satisfies A ϕ , z ( 1 , 1 ) = 1 A_{\phi,z}(1,1)=1 A ϕ , z ( 1 , 1 ) = 1 . Taking ( a , b , m , n ) = ( 1 , 1 , 1 , 1 ) (a,b,m,n)=(1,1,1,1) ( a , b , m , n ) = ( 1 , 1 , 1 , 1 ) in (73) and χ p ≪ 1 \chi_p\ll1 χ p ≪ 1 , we have
χ p ∣ E max , n e w ∣ ≪ h , ε p 1 + ε . \chi_p|\mathcal{E}_{\max,\mathrm{new}}|\ll_{h,\varepsilon}p^{1+\varepsilon}. χ p ∣ E m a x , new ∣ ≪ h , ε p 1 + ε . Inserting these estimates into (51) , we obtain
χ p c c u s p ∑ π ∈ B ( p ) h ( μ π ) L ( 1 , π , Ad ) = P H h + O h , ε ( p 1 + ε ) . \chi_p c_{\mathrm{cusp}}\sum_{\pi\in\mathcal{B}(p)}\frac{h(\mu_\pi)}{L(1,\pi,\operatorname{Ad})}=P\mathcal{H}_h+O_{h,\varepsilon}(p^{1+\varepsilon}). χ p c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) = P H h + O h , ε ( p 1 + ε ) . Since P / χ p = p 2 P/\chi_p=p^2 P / χ p = p 2 and P − p 2 = p + 1 P-p^2=p+1 P − p 2 = p + 1 , division by χ p ≍ 1 \chi_p\asymp1 χ p ≍ 1 proves (136) . The final assertion follows from H h > 0 \mathcal{H}_h>0 H h > 0 .
Proof of Theorems 6.1 and 1.1
Proof. Fix a square-free Q Q Q and 0 < η < 1 / 2 0<\eta<1/2 0 < η < 1/2 , independently of p p p , and put L = p η L=p^\eta L = p η . By the functional equation from (13) and (5) ,
L ( 1 / 2 , π ) M ^ π = L ( 1 / 2 , π ) M π + L ( 1 / 2 , π ) M π ‾ = 2 Re ( L ( 1 / 2 , π ) M π ) . L(1/2,\pi)\widehat{M}_{\pi}=L(1/2,\pi)M_{\pi}+L(1/2,\pi)\overline{M_{\pi}}=2\operatorname{Re}(L(1/2,\pi)M_{\pi}). L ( 1/2 , π ) M π = L ( 1/2 , π ) M π + L ( 1/2 , π ) M π = 2 Re ( L ( 1/2 , π ) M π ) .
Thus
∣ L ( 1 / 2 , π ) M ^ π ∣ 2 = 2 ∣ L ( 1 / 2 , π ) M π ∣ 2 + 2 Re ( ( L ( 1 / 2 , π ) M π ) 2 ) . |L(1/2,\pi)\widehat{M}_{\pi}|^2=2|L(1/2,\pi)M_{\pi}|^2+2\operatorname{Re}\left((L(1/2,\pi)M_{\pi})^2\right). ∣ L ( 1/2 , π ) M π ∣ 2 = 2∣ L ( 1/2 , π ) M π ∣ 2 + 2 Re ( ( L ( 1/2 , π ) M π ) 2 ) .
Summing these identities with the factors c c u s p h ( μ π ) / L ( 1 , π , A d ) c_{\mathrm{cusp}}h(\mu_\pi)/L(1,\pi,\mathrm{Ad}) c cusp h ( μ π ) / L ( 1 , π , Ad ) gives
c c u s p ∑ π ∈ B ( p ) h ( μ π ) L ( 1 / 2 , π ) M ^ π L ( 1 , π , A d ) = 2 Re M 1 ( p ) , (137) c_{\mathrm{cusp}}\sum_{\pi\in B(p)}\frac{h(\mu_\pi)L(1/2,\pi)\widehat{M}_\pi}{L(1,\pi,\mathrm{Ad})}=2\operatorname{Re}\mathcal{M}_1(p), \tag*{(137)} c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) L ( 1/2 , π ) M π = 2 Re M 1 ( p ) , (137)
c c u s p ∑ π ∈ B ( p ) h ( μ π ) ∣ L ( 1 / 2 , π ) M ^ π ∣ 2 L ( 1 , π , A d ) = 2 M 2 ( p ) + 2 Re M 1 , 1 ( p ) . (138) c_{\mathrm{cusp}}\sum_{\pi\in B(p)}\frac{h(\mu_\pi)\lvert L(1/2,\pi)\widehat{M}_\pi\rvert^2}{L(1,\pi,\mathrm{Ad})}=2\mathcal{M}_2(p)+2\operatorname{Re}\mathcal{M}_{1,1}(p). \tag*{(138)} c cusp π ∈ B ( p ) ∑ L ( 1 , π , Ad ) h ( μ π ) ∣ L ( 1/2 , π ) M π ∣ 2 = 2 M 2 ( p ) + 2 Re M 1 , 1 ( p ) . (138)
For sufficiently large p p p , (88) gives Re M 1 ( p ) > 0 \operatorname{Re}\mathcal{M}_1(p)>0 Re M 1 ( p ) > 0 , so the second moment in (6.3) is positive. When L ( 1 / 2 , π ) = 0 L(1/2,\pi)=0 L ( 1/2 , π ) = 0 , we have L ( 1 / 2 , π ) M ^ π = 0 L(1/2,\pi)\widehat{M}_\pi=0 L ( 1/2 , π ) M π = 0 . Then the Cauchy–Schwarz inequality with the non-negative factors h ( μ π ) / L ( 1 , π , A d ) h(\mu_\pi)/L(1,\pi,\mathrm{Ad}) h ( μ π ) / L ( 1 , π , Ad ) yields
c c u s p ∑ π ∈ B ( p ) L ( 1 / 2 , π ) ≠ 0 h ( μ π ) L ( 1 , π , A d ) ≥ 2 { Re M 1 ( p ) } 2 M 2 ( p ) + Re M 1 , 1 ( p ) . (139) c_{\mathrm{cusp}}\sum_{\substack{\pi\in B(p)\\L(1/2,\pi)\ne0}}\frac{h(\mu_\pi)}{L(1,\pi,\mathrm{Ad})}\ge\frac{2\{\operatorname{Re}\mathcal{M}_1(p)\}^2}{\mathcal{M}_2(p)+\operatorname{Re}\mathcal{M}_{1,1}(p)}. \tag*{(139)} c cusp π ∈ B ( p ) L ( 1/2 , π ) = 0 ∑ L ( 1 , π , Ad ) h ( μ π ) ≥ M 2 ( p ) + Re M 1 , 1 ( p ) 2 { Re M 1 ( p ) } 2 . (139)
The Euler product in (111) lies in ( 0 , 1 ] (0,1] ( 0 , 1 ] . Hence (90) and (111) give
M 2 ( p ) + Re M 1 , 1 ( p ) ≤ P H h ( ζ ( 2 ) ( 1 + η − 1 ) ∏ q ∣ Q ( 1 − q − 2 ) + 1 ) + O h , Q , η ( p 2 log ( 2 + log p ) log p ) . \mathcal{M}_2(p)+\operatorname{Re}\mathcal{M}_{1,1}(p)\le P\!H_h\left(\zeta(2)(1+\eta^{-1})\prod_{q\mid Q}(1-q^{-2})+1\right)+O_{h,Q,\eta}\left(\frac{p^2\log(2+\log p)}{\log p}\right). M 2 ( p ) + Re M 1 , 1 ( p ) ≤ P H h ζ ( 2 ) ( 1 + η − 1 ) q ∣ Q ∏ ( 1 − q − 2 ) + 1 + O h , Q , η ( log p p 2 log ( 2 + log p ) ) .
Inserting this bound and (88) into (139) and dividing by the total harmonic mass in (136) , we then have
∑ π ∈ B ( p ) L ( 1 / 2 , π ) ≠ 0 h ( μ π ) L ( 1 , π , A d ) ∑ π ∈ B ( p ) h ( μ π ) L ( 1 , π , A d ) ≥ 2 ζ ( 2 ) ( 1 + η − 1 ) ∏ q ∣ Q ( 1 − q − 2 ) + 1 − O h , Q , η ( log ( 2 + log p ) log p ) . (140) \frac{\sum_{\substack{\pi\in B(p)\\L(1/2,\pi)\ne0}}\frac{h(\mu_\pi)}{L(1,\pi,\mathrm{Ad})}}{\sum_{\pi\in B(p)}\frac{h(\mu_\pi)}{L(1,\pi,\mathrm{Ad})}}\ge\frac{2}{\zeta(2)(1+\eta^{-1})\prod_{q\mid Q}(1-q^{-2})+1}-O_{h,Q,\eta}\left(\frac{\log(2+\log p)}{\log p}\right). \tag*{(140)} ∑ π ∈ B ( p ) L ( 1 , π , Ad ) h ( μ π ) ∑ π ∈ B ( p ) L ( 1/2 , π ) = 0 L ( 1 , π , Ad ) h ( μ π ) ≥ ζ ( 2 ) ( 1 + η − 1 ) ∏ q ∣ Q ( 1 − q − 2 ) + 1 2 − O h , Q , η ( log p log ( 2 + log p ) ) . (140)
Let ϵ > 0 \epsilon>0 ϵ > 0 be given. As Q Q Q runs through products of all primes up to a bound tending to infinity, the Euler product for ζ ( 2 ) \zeta(2) ζ ( 2 ) gives
ζ ( 2 ) ∏ q ∣ Q ( 1 − q − 2 ) ⟶ 1. \zeta(2)\prod_{q\mid Q}(1-q^{-2})\longrightarrow1. ζ ( 2 ) q ∣ Q ∏ ( 1 − q − 2 ) ⟶ 1.
Since 1 + η − 1 → 3 1+\eta^{-1}\to3 1 + η − 1 → 3 as η → ( 1 / 2 ) − \eta\to(1/2)^- η → ( 1/2 ) − , we may first choose such a fixed Q Q Q , and then a fixed 0 < η < 1 / 2 0<\eta<1/2 0 < η < 1/2 , so that
2 ζ ( 2 ) ( 1 + η − 1 ) ∏ q ∣ Q ( 1 − q − 2 ) + 1 > 1 2 − ϵ 2 . \frac{2}{\zeta(2)(1+\eta^{-1})\prod_{q\mid Q}(1-q^{-2})+1}>\frac{1}{2}-\frac{\epsilon}{2}. ζ ( 2 ) ( 1 + η − 1 ) ∏ q ∣ Q ( 1 − q − 2 ) + 1 2 > 2 1 − 2 ϵ .
With these choices fixed, take p > Q p>Q p > Q sufficiently large. Then p ∤ Q p\nmid Q p ∤ Q and the error in (140) is less than ϵ / 2 \epsilon/2 ϵ /2 . Multiplying (140) by the total harmonic mass gives (135) . The parameters Q , η Q,\eta Q , η depend only on ϵ \epsilon ϵ . Thus the lower bound holds for all sufficiently large primes, with a threshold depending only on h , ϵ h,\epsilon h , ϵ . Lemma 6.2 also gives the asserted lower bound ≫ h p 2 \gg_h p^2 ≫ h p 2 . This proves Theorem 6.1 . Cancelling the fixed positive constant c c u s p c_{\mathrm{cusp}} c cusp proves Theorem 1.1 . □
Acknowledgements The authors would like to thank Professors Qingfeng Sun and Zhining Wei for their help and encouragement.
References
The University of Hong Kong Shenzhen Institute of Research and Innovation, Shenzhen, Guangdong 518057, China
Email address: wh0315@mail.sdu.edu.cn
Tongji University, Shanghai 200092, China
Email address: slwang@tongji.edu.cn