Introduction
The study of the Riemann zeta function ζ ( s ) \zeta(s) ζ ( s ) plays an important role in understanding the distribution of the prime numbers. At the heart of this lies the important observation of Riemann that the Prime Number Theorem is equivalent to showing that ζ ( s ) \zeta(s) ζ ( s ) has no zeros on the line ℜ ( s ) = 1 \Re(s)=1 ℜ ( s ) = 1 . The well-known Riemann Hypothesis in fact asserts that all the zeros of ζ ( s ) \zeta(s) ζ ( s ) in the critical strip 0 < ℜ ( s ) < 1 0<\Re(s)<1 0 < ℜ ( s ) < 1 lie on the line ℜ ( s ) = 1 2 \Re(s)=\frac{1}{2} ℜ ( s ) = 2 1 . One of the major advances towards this conjecture was made by Hardy [9 ] who showed that ζ ( s ) \zeta(s) ζ ( s ) has infinitely many zeros on the critical line ℜ ( s ) = 1 2 \Re(s)=\frac{1}{2} ℜ ( s ) = 2 1 . A key point in the proof of this results is the introduction of Hardy’s function for ζ ( s ) \zeta(s) ζ ( s ) : a real-valued function Z ( t ) Z(t) Z ( t ) of the real valuable of t t t , whose zeros are exactly the zeros of ζ ( s ) \zeta(s) ζ ( s ) on the critical line ℜ ( s ) = 1 2 \Re(s)=\frac{1}{2} ℜ ( s ) = 2 1 . A similar argument applies to Dirichlet L L L -functions L ( s , χ ) L(s,\chi) L ( s , χ ) .
In a different direction, the Riemann zeta function can be viewed as a prototype of a more general class of L L L -functions. Chandrasekharan and Narasimhan [4 ] showed that the zeta functions of ideal classes in quadratic fields possess infinitely many zeros on the critical line by constructing a suitable analogue of Hardy’s function. Hafner [7 , 8 ] showed that the L L L -functions attached to holomorphic cusp forms and Maass wave forms for the full modular group with trivial character have infinitely many zeros on the critical line. More precisely, he proved that N o ( T ) ≫ T log T N_o(T)\gg T\log T N o ( T ) ≫ T log T , where N o ( T ) N_o(T) N o ( T ) counts the zeros on the critical line with imaginary part in ( 0 , T ] (0,T] ( 0 , T ] . Later, Rezvyakova [31 ] established the corresponding result for L L L -functions associated with automorphic cusp forms on Hecke congruence subgroups of arbitrary integral weight. See the necessary definitions in [13 ] .
Wilton [37 ] studied the additively twisted L L L -function of the modular discriminant Δ ( z ) \Delta(z) Δ ( z ) , defined by
ρ ( s ; p , q ) = ∑ n ≥ 1 τ ( n ) n s e 2 π i n p / q . \rho(s;p,q)=\sum_{n\ge1}\frac{\tau(n)}{n^s}e^{2\pi inp/q}. ρ ( s ; p , q ) = n ≥ 1 ∑ n s τ ( n ) e 2 π in p / q .
2020 Mathematics Subject Classification. Primary 11M06; Secondary 11M41, 11M26.
Key words and phrases. Selberg class, L L L -functions, Hardy’s function, Resonance method, Cusp forms. He proved that, for rational p / q p/q p / q satisfying p 2 ≡ 1 ( m o d q ) p^2 \equiv1 \pmod{q} p 2 ≡ 1 ( mod q ) , this function has infinitely many zeros on the line ℜ ( s ) = 6 \Re(s)=6 ℜ ( s ) = 6 , where τ ( n ) \tau(n) τ ( n ) denotes the n n n -th Fourier coefficient of Δ ( z ) \Delta(z) Δ ( z ) . His proof used a variant of Hardy’s method together with the Hardy-type function
H α ( t , n ) = t n R ( t ; p , q ) e α t , R ( t ; p , q ) = ( q 2 π ) 6 + i t Γ ( 6 + i t ) ρ ( 6 + i t ; p , q ) , H_{\alpha}(t,n)=t^nR(t;p,q)e^{\alpha t},\qquad R(t;p,q)=\left(\frac{q}{2\pi}\right)^{6+it}\Gamma(6+it)\rho(6+it;p,q), H α ( t , n ) = t n R ( t ; p , q ) e α t , R ( t ; p , q ) = ( 2 π q ) 6 + i t Γ ( 6 + i t ) ρ ( 6 + i t ; p , q ) ,
where − π / 2 < α < π / 2 -\pi/2<\alpha<\pi/2 − π /2 < α < π /2 . Later, Kim [20 ] generalized this result to cusp forms of integral and half-integral weights of higher level. In [24 ] , Meher, Pujahari, and Kotyada studied the zeros of L L L -functions associated with cusp forms. For half-integral weight k + 1 / 2 k+1/2 k + 1/2 cusp forms of level 4 4 4 , they showed that infinitely many zeros lie on the critical line ℜ ( s ) = k + 1 / 2 2 \Re(s)=\frac{k+1/2}{2} ℜ ( s ) = 2 k + 1/2 . They also obtained a similar result for integral weight k k k cusp forms of level N N N , where infinitely many zeros lie on the critical line ℜ ( s ) = k 2 \Re(s)=\frac{k}{2} ℜ ( s ) = 2 k . This result was later generalized by Meher, Pujahari, and Shankhadhar [25 ] to half-integral weight cusp forms of level 4 N 2 4N^2 4 N 2 .
Many of the L L L -functions considered above, such as those attached to integral weight cusp forms and Dedekind zeta functions, belong to a broader axiomatic framework known as the Selberg class S \mathcal{S} S , introduced by Selberg [32 ] in 1992. This class consists of functions satisfying a set of analytic properties, including a functional equation of the type that will be essential for constructing an appropriate analogue of Hardy’s function. We now recall the precise definition of the Selberg class.
Definition 1.1. (The Selberg Class). The Selberg class S \mathcal{S} S is the collection of meromorphic functions F ( s ) F(s) F ( s ) for which the following conditions hold.
Dirichlet series - The function F ( s ) F(s) F ( s ) admits a Dirichlet series representation
F ( s ) = ∑ n = 1 ∞ a F ( n ) n s , F(s)=\sum_{n=1}^{\infty}\frac{a_F(n)}{n^s}, F ( s ) = n = 1 ∑ ∞ n s a F ( n ) , which converges absolutely for ℜ ( s ) > 1 \Re(s)>1 ℜ ( s ) > 1 . Also, the leading coefficient a F ( 1 ) a_F(1) a F ( 1 ) is normalized as a F ( 1 ) = 1 a_F(1)=1 a F ( 1 ) = 1 .
Analytic continuation - For some integer k ≥ 0 k\geq0 k ≥ 0 , the function ( s − 1 ) k F ( s ) (s-1)^kF(s) ( s − 1 ) k F ( s ) is entire and of finite order.
Functional equation - Choose Q > 0 Q>0 Q > 0 , α i > 0 \alpha_i>0 α i > 0 , β i ∈ C \beta_i\in\mathbb{C} β i ∈ C , and ω ∈ C \omega\in\mathbb{C} ω ∈ C with ℜ ( β i ) ≥ 0 \Re(\beta_i)\geq0 ℜ ( β i ) ≥ 0 and ∣ ω ∣ = 1 |\omega|=1 ∣ ω ∣ = 1 . Set
Φ ( s ) : = Q s ∏ i = 1 m Γ ( α i s + β i ) F ( s ) . (1) \Phi(s):=Q^s\prod_{i=1}^{m}\Gamma(\alpha_i s+\beta_i)F(s). \tag*{(1)} Φ ( s ) := Q s i = 1 ∏ m Γ ( α i s + β i ) F ( s ) . (1) Then Φ ( s ) \Phi(s) Φ ( s ) satisfies the functional equation
Φ ( s ) = ω Φ ( 1 − s ‾ ) ‾ . (2) \Phi(s)=\omega\overline{\Phi(1-\overline{s})}. \tag*{(2)} Φ ( s ) = ω Φ ( 1 − s ) . (2) Euler product - For ℜ ( s ) > 1 \Re(s)>1 ℜ ( s ) > 1 , the function F ( s ) F(s) F ( s ) has the Euler product
F ( s ) = ∏ p p r i m e F p ( s ) . (3) F(s)=\prod_{p\ \mathrm{prime}}F_p(s). \tag*{(3)} F ( s ) = p prime ∏ F p ( s ) . (3) where
log F p ( s ) = ∑ k = 1 ∞ b F ( p k ) p k s \log F_p(s)=\sum_{k=1}^{\infty}\frac{b_F(p^k)}{p^{ks}} log F p ( s ) = k = 1 ∑ ∞ p k s b F ( p k ) with b F ( p k ) = O ( p k θ ) b_F(p^k)=O(p^{k\theta}) b F ( p k ) = O ( p k θ ) for some θ < 1 2 \theta<\frac{1}{2} θ < 2 1 .
(v) Ramanujan hypothesis - For any ϵ > 0 \epsilon> 0 ϵ > 0 ,
a F ( n ) = O ϵ ( n ϵ ) . a_F(n) = O_{\epsilon}(n^{\epsilon}). a F ( n ) = O ϵ ( n ϵ ) .
We further use a stronger version of axiom (iv), known as the polynomial Euler product:
(iv′) Polynomial Euler product - for ℜ ( s ) > 1 \Re(s) > 1 ℜ ( s ) > 1 we have
F ( s ) = ∏ p p r i m e ∏ i = 1 r ( 1 − ν i ( p ) p s ) − 1 (4) F(s) = \prod_{p\ {\rm prime}} \prod_{i=1}^{r} \left(1-\frac{\nu_i(p)}{p^s}\right)^{-1} \tag*{(4)} F ( s ) = p prime ∏ i = 1 ∏ r ( 1 − p s ν i ( p ) ) − 1 (4)
where ν i ( p ) \nu_i(p) ν i ( p ) are complex numbers.
Let S p o l y \mathcal{S}_{\mathrm{poly}} S poly denote the subclass of S \mathcal{S} S comprising all L L L -functions that satisfy axiom (iv′). This additional assumption is not expected to exclude any important L L L -functions arising in number theory. In fact, all presently known examples in the Selberg class are believed to satisfy (iv′), at least under certain widely accepted conjectures. For our purposes, the main consequence of (iv′) is that, together with the Ramanujan hypothesis (v), it gives ∣ ν i ( p ) ∣ ≤ 1 |\nu_i(p)| \leq1 ∣ ν i ( p ) ∣ ≤ 1 (see [34 Lemma 2.2] ). Therefore, we have
a F ( p ) = b F ( p ) = ∑ i = 1 r ν i ( p ) and ∣ a F ( p ) ∣ ≤ r . a_F(p) = b_F(p) = \sum_{i=1}^{r}\nu_i(p) \quad\text{and} \quad|a_F(p)| \leq r. a F ( p ) = b F ( p ) = i = 1 ∑ r ν i ( p ) and ∣ a F ( p ) ∣ ≤ r .
The Euler product structure gives the multiplicativity of the coefficients, namely, a F ( m n ) = a F ( m ) a F ( n ) a_F(mn) = a_F(m)a_F(n) a F ( mn ) = a F ( m ) a F ( n ) whenever ( m , n ) = 1 (m,n) = 1 ( m , n ) = 1 . In addition, every Euler factor admits a Dirichlet series representation
F p ( s ) = ∑ k = 0 ∞ a F ( p k ) p k s , F_p(s) = \sum_{k=0}^{\infty}\frac{a_F(p^k)}{p^{ks}}, F p ( s ) = k = 0 ∑ ∞ p k s a F ( p k ) ,
and this series converges absolutely in the half-plane ℜ ( s ) > 0 \Re(s) > 0 ℜ ( s ) > 0 . Moreover, F p ( s ) F_p(s) F p ( s ) has no zeros for ℜ ( s ) > θ \Re(s) > \theta ℜ ( s ) > θ , where θ \theta θ is the quantity appearing in (3) .
The constants appearing in the functional equation (1.1) depend on F F F . Since the duplication formula for the Γ \Gamma Γ -function may give different forms of the functional equation, these constants are not unique. However, the quantity
d F : = 2 ∑ i = 1 m α i d_F := 2\sum_{i=1}^{m}\alpha_i d F := 2 i = 1 ∑ m α i
is a well-defined invariant called the degree of F F F . It is conjectured that d F d_F d F is always an integer. Henceforth, we will denote the members of S \mathcal{S} S having degree d d d by S d \mathcal{S}_d S d . We also define the ξ \xi ξ -invariant and the conductor by
ξ F : = 2 ∑ j = 1 m ( β j − 1 2 ) and q F : = ( 2 π ) d F Q 2 ∏ j = 1 m α j 2 α j , (5) \xi_F := 2\sum_{j=1}^{m}\left(\beta_j-\frac{1}{2}\right) \quad\text{and} \quad q_F := (2\pi)^{d_F}Q^2\prod_{j=1}^{m}\alpha_j^{2\alpha_j}, \tag*{(5)} ξ F := 2 j = 1 ∑ m ( β j − 2 1 ) and q F := ( 2 π ) d F Q 2 j = 1 ∏ m α j 2 α j , (5)
respectively.
Kaczorowski and Perelli [15 , 16 ] studied the Selberg class S \mathcal{S} S in the degree range 0 < d F < 2 0 < d_F < 2 0 < d F < 2 and obtained a complete description of the functions that occur in this range. They showed that the possible functions are the Riemann zeta function ζ ( s ) \zeta(s) ζ ( s ) , the Dirichlet L L L -functions L ( s , χ ) L(s,\chi) L ( s , χ ) associated with primitive characters χ \chi χ , and their imaginary shifts L ( s + i θ , χ ) L(s+i\theta,\chi) L ( s + i θ , χ ) , where θ ∈ R \theta\in\mathbb{R} θ ∈ R . Also note that for non-primitive character χ \chi χ (mod q q q ) with q ≠ 1 q \ne1 q = 1 , L ( s , χ ) ∉ S L(s,\chi) \notin\mathcal{S} L ( s , χ ) ∈ / S . Recently, in [17 , 18 ] , they classified the L L L -functions of degree 2 and conductor 1. They showed that every F ∈ S F \in\mathcal{S} F ∈ S of degree 2 and conductor 1 is an automorphic L L L -function. The known examples of degree 2 functions include products of two degree 1 functions, such as ζ ( s ) \zeta(s) ζ ( s ) and L ( s , χ 1 ) L ( s , χ 2 ) L(s,\chi_1)L(s,\chi_2) L ( s , χ 1 ) L ( s , χ 2 ) ; normalized L L L -functions associated with holomorphic newforms; the Dedekind zeta-function ζ K ( s ) \zeta_K(s) ζ K ( s ) associated with a quadratic number field K K K ; and the imaginary shifts of any of the above functions which are entire. Normalized L L L -functions associated with non-holomorphic newforms are expected to lie in S, but the Ramanujan hypothesis has not yet been verified, the best current upper bound for the Fourier coefficients is n 7 / 64 + ϵ n^{7/64+\epsilon} n 7/64 + ϵ due to Kim and Sarnak [19 ] . Assuming the Ramanujan hypothesis holds, we believe these L L L -functions belong to S. In these cases, infinitely many zeros are known to be on the critical line. However, for degree 2, we do not yet know whether the known list contains all possible cases. Thus, the problem is still open in general. Under some suitable assumptions, Mukhopadhyay, Srinivas, and Rajkumar [28 ] proved that degree-two functions in the Selberg class have infinitely many zeros on the critical line. Their result applies, in particular, to the L L L -functions L f ( s ) L_f(s) L f ( s ) attached to cuspidal modular forms f f f of level N \mathcal{N} N when N \mathcal{N} N is not a perfect square, and also to the products L ( s , χ 1 ) L ( s , χ 2 ) L(s,\chi_1)L(s,\chi_2) L ( s , χ 1 ) L ( s , χ 2 ) with primitive Dirichlet characters χ 1 \chi_1 χ 1 (mod q 1 q_1 q 1 ) and χ 2 \chi_2 χ 2 (mod q 2 q_2 q 2 ), whenever q 1 q 2 q_1q_2 q 1 q 2 is not a perfect square.
The functional equation (2) can be written as
F ( s ) = H ( s ) F ( 1 − s ˉ ) ‾ , (6) F(s)=H(s)\overline{F(1-\bar{s})}, \tag*{(6)} F ( s ) = H ( s ) F ( 1 − s ˉ ) , (6)
where
H ( s ) : = ω Q 1 − 2 s ∏ i = 1 m Γ ( α i ( 1 − s ) + β i ‾ ) Γ ( α i s + β i ) . (7) H(s):=\omega Q^{1-2s}\prod_{i=1}^{m}\frac{\Gamma(\alpha_i(1-s)+\overline{\beta_i})}{\Gamma(\alpha_i s+\beta_i)}. \tag*{(7)} H ( s ) := ω Q 1 − 2 s i = 1 ∏ m Γ ( α i s + β i ) Γ ( α i ( 1 − s ) + β i ) . (7)
This allows us to define the analogue of Hardy’s function for F ∈ S F\in\mathrm{S} F ∈ S as
Z F ( t ) : = F ( 1 / 2 + i t ) H ( 1 / 2 + i t ) − 1 / 2 . (8) Z_F(t):=F(1/2+it)H(1/2+it)^{-1/2}. \tag*{(8)} Z F ( t ) := F ( 1/2 + i t ) H ( 1/2 + i t ) − 1/2 . (8)
Using (6) , we have ∣ H ( 1 / 2 + i t ) ∣ = 1 |H(1/2+it)|=1 ∣ H ( 1/2 + i t ) ∣ = 1 , and hence Z F ( t ) ‾ = Z F ( t ) \overline{Z_F(t)}=Z_F(t) Z F ( t ) = Z F ( t ) . Thus, Z F ( t ) Z_F(t) Z F ( t ) is real valued for real t t t , and its zeros correspond exactly to the zeros of F ( 1 / 2 + i t ) F(1/2+it) F ( 1/2 + i t ) .
Let us define
Z F + ( t ) : = max ( 0 , Z F ( t ) ) , Z_F^+(t):=\max(0,Z_F(t)), Z F + ( t ) := max ( 0 , Z F ( t )) ,
Z F − ( t ) : = max ( 0 , − Z F ( t ) ) . Z_F^-(t):=\max(0,-Z_F(t)). Z F − ( t ) := max ( 0 , − Z F ( t )) .
The second author [22 ] established Ω ± \Omega_{\pm} Ω ± results for Hardy’s function Z ( t ) Z(t) Z ( t ) corresponding to the Riemann zeta function ζ ( s ) \zeta(s) ζ ( s ) using the resonator constructed by Bondarenko and Seip [2 ] . Specifically, it was shown that,
max T 3 / 4 ≤ t ≤ T ± Z ( t ) ≫ exp ( C log T log log log T log log T ) , \max_{T^{3/4}\leq t\leq T}\pm Z(t)\gg\exp\left(C\sqrt{\frac{\log T\log\log\log T}{\log\log T}}\right), T 3/4 ≤ t ≤ T max ± Z ( t ) ≫ exp ( C log log T log T log log log T ) ,
holds with the constant C = 1 / 2 − ϵ C=1/2-\epsilon C = 1/2 − ϵ for every ϵ > 0 \epsilon>0 ϵ > 0 .
It was also remarked in the paper that the interval [ T 3 / 4 , T ] [T^{3/4},T] [ T 3/4 , T ] for t t t can be improved to [ T , 2 T ] [T,2T] [ T , 2 T ] using Soundararajan’s resonance method [33 ] , but at the cost of a weaker result. More precisely, the bound
max T ≤ t ≤ 2 T ± Z ( t ) ≫ exp ( C log T log log T ) , (9) \max_{T\leq t\leq2T}\pm Z(t)\gg\exp\left(C\sqrt{\frac{\log T}{\log\log T}}\right), \tag*{(9)} T ≤ t ≤ 2 T max ± Z ( t ) ≫ exp ( C log log T log T ) , (9)
holds with the constant C = 1 / 2 − ϵ C=1/2-\epsilon C = 1/2 − ϵ for every ϵ > 0 \epsilon>0 ϵ > 0 . In a related direction, Gonek and Ivić [6 ] and Mawia [23 ] studied the interval of distribution of positive and negative values of Hardy’s function. Further large-value results using the resonance method were obtained by Aistleitner and Pańkowski [1 ] for L L L -functions in the Selberg class, and by Peyrot [29 ] for Hecke–Maass L L L -functions on the critical line with prescribed argument.
In this paper, we extend the above Ω ± \Omega_{\pm} Ω ± result for Hardy’s function to L L L -functions of degree 1 and certain L L L -functions of degree 2 using the resonance method. Here and throughout, the notation Z F ± ( t ) Z_F^{\pm}(t) Z F ± ( t ) means that the corresponding estimate holds separately for both Z F + ( t ) Z_F^{+}(t) Z F + ( t ) and Z F − ( t ) Z_F^{-}(t) Z F − ( t ) . We now state the results of this paper.
Theorem 1.2. Let F ( s ) F(s) F ( s ) be an element of the Selberg class S \mathcal{S} S of degree 1. Then, for sufficiently large T T T ,
max T ≤ t ≤ 2 T Z F ± ( t ) ≫ exp ( ( 1 2 − ϵ ) log T log log T ) , \max_{T\leq t\leq2T} Z_F^{\pm}(t) \gg\exp\left(\left(\frac{1}{2}-\epsilon\right)\sqrt{\frac{\log T}{\log\log T}}\right), T ≤ t ≤ 2 T max Z F ± ( t ) ≫ exp ( ( 2 1 − ϵ ) log log T log T ) , where ϵ > 0 \epsilon>0 ϵ > 0 is arbitrarily small.
Before stating our second theorem, we introduce the following hypothesis for degree-two functions.
Hypothesis 1. Let F ( s ) ∈ S p o l y F(s)\in\mathcal{S}_{\mathrm{poly}} F ( s ) ∈ S poly have degree 2 and a F ( n ) a_F(n) a F ( n ) be the Dirichlet coefficients of F F F . There exists γ < 1 \gamma<1 γ < 1 and γ 1 ∈ R \gamma_1\in\mathbb{R} γ 1 ∈ R such that
∑ n ≤ T a F ( n ) e 2 π i n x ≪ T γ ( log T ) γ 1 , \sum_{n\leq T}a_F(n)e^{2\pi i n x}\ll T^\gamma(\log T)^{\gamma_1}, n ≤ T ∑ a F ( n ) e 2 π in x ≪ T γ ( log T ) γ 1 , uniformly for all large T T T and x ∈ R x\in\mathbb{R} x ∈ R , with implied constant depends only on F F F . Here the values of γ \gamma γ and γ 1 \gamma_1 γ 1 depend on F ( s ) F(s) F ( s ) .
Theorem 1.3. Let F ( s ) = ∑ n = 1 ∞ a F ( n ) n − s ∈ S p o l y F(s)=\sum_{n=1}^{\infty}a_F(n)n^{-s}\in\mathcal{S}_{\mathrm{poly}} F ( s ) = ∑ n = 1 ∞ a F ( n ) n − s ∈ S poly be a function of degree 2 satisfying Hypothesis 1, with ℑ ξ F = 0 \Im\xi_F=0 ℑ ξ F = 0 , where ξ F \xi_F ξ F is the ξ \xi ξ -invariant defined in (5) . Moreover, assume that F ( s ) F(s) F ( s ) satisfies the following Selberg normality conjecture,
∑ p ≤ y ∣ a F ( p ) ∣ 2 = ( κ + o ( 1 ) ) y log y , ( κ > 0 ) . (10) \sum_{p\leq y}|a_F(p)|^2=(\kappa+o(1))\frac{y}{\log y},\qquad(\kappa>0). \tag*{(10)} p ≤ y ∑ ∣ a F ( p ) ∣ 2 = ( κ + o ( 1 )) log y y , ( κ > 0 ) . (10) Then for sufficiently large T T T ,
max T ≤ t ≤ 2 T Z F ± ( t ) ≫ exp ( C F log T log log T ) , \max_{T\leq t\leq2T}Z_F^{\pm}(t)\gg\exp\left(C_F\sqrt{\frac{\log T}{\log\log T}}\right), T ≤ t ≤ 2 T max Z F ± ( t ) ≫ exp ( C F log log T log T ) , where the constant C F C_F C F depends on F ( s ) F(s) F ( s ) . More precisely, for arbitrarily small ϵ > 0 \epsilon>0 ϵ > 0 , we have
C F = κ ( 1 − γ ) − ϵ . (11) C_F=\sqrt{\kappa(1-\gamma)}-\epsilon. \tag*{(11)} C F = κ ( 1 − γ ) − ϵ . (11)
Remark 1.4. The condition ℑ ξ F = 0 \Im\xi_F=0 ℑ ξ F = 0 is imposed to avoid the additional factor t − i ℑ ξ F t^{-i\Im\xi_F} t − i ℑ ξ F in (13) , whose presence makes the direct application of Lemma 2.4 to (39) more complicated. This condition does not restrict the result when F ( s ) F(s) F ( s ) is regular at s = 1 s=1 s = 1 . Indeed, by considering the shift G ( s ) = F ( s + i τ ) G(s)=F(s+i\tau) G ( s ) = F ( s + i τ ) with τ = − ℑ ξ F / d F \tau=-\Im\xi_F/d_F τ = − ℑ ξ F / d F , we can achieve the condition ℑ ξ G = 0 \Im\xi_G=0 ℑ ξ G = 0 . With a compatible choice of the square root, Z G ( t ) = Z F ( t + τ ) Z_G(t)=Z_F(t+\tau) Z G ( t ) = Z F ( t + τ ) . Therefore, applying the same argument to G G G on the translated interval [ T − τ , 2 T − τ ] [T-\tau,2T-\tau] [ T − τ , 2 T − τ ] yields the conclusion of the above theorem for the original function F F F on [ T , 2 T ] [T,2T] [ T , 2 T ] .
Remark 1.5. Parseval’s identity and (10) give
∫ 0 1 ∣ ∑ n ≤ Y a F ( n ) e 2 π i n x ∣ 2 d x = ∑ n ≤ Y ∣ a F ( n ) ∣ 2 ≥ ∑ p ≤ Y ∣ a F ( p ) ∣ 2 ∼ κ Y log Y . \int_0^1\left|\sum_{n\leq Y}a_F(n)e^{2\pi i n x}\right|^2dx=\sum_{n\leq Y}|a_F(n)|^2\geq\sum_{p\leq Y}|a_F(p)|^2\sim\kappa\frac{Y}{\log Y}. ∫ 0 1 n ≤ Y ∑ a F ( n ) e 2 π in x 2 d x = n ≤ Y ∑ ∣ a F ( n ) ∣ 2 ≥ p ≤ Y ∑ ∣ a F ( p ) ∣ 2 ∼ κ log Y Y .
On the other hand, Hypothesis 1 implies
∫ 0 1 ∣ ∑ n ≤ Y a F ( n ) e 2 π i n x ∣ 2 d x ≪ Y 2 γ ( log Y ) 2 γ 1 . \int_0^1 \left|\sum_{n\le Y} a_F(n)e^{2\pi i n x}\right|^2 dx \ll Y^{2\gamma}(\log Y)^{2\gamma_1}. ∫ 0 1 n ≤ Y ∑ a F ( n ) e 2 π in x 2 d x ≪ Y 2 γ ( log Y ) 2 γ 1 .
Comparing these two estimates, we must have γ ≥ 1 / 2 \gamma\ge1/2 γ ≥ 1/2 . Therefore, under Hypothesis 1, it follows that 1 / 2 ≤ γ < 1 1/2 \le\gamma< 1 1/2 ≤ γ < 1 .
As a special case of Theorem 1.3 , we get the following theorems and give a sketch of their proofs, as the arguments are similar.
Theorem 1.6. Let F ( s ) F(s) F ( s ) be one of the following functions:
(i) the L L L -function associated with a normalized primitive holomorphic Hecke cusp eigenform of even weight k ≥ 2 k \ge2 k ≥ 2 , level N N N and nebentypus χ \chi χ .
(ii) assuming the Ramanujan hypothesis, the L L L -function associated with a normalized primitive Hecke–Maass cusp form of weight 0 0 0 , level N N N and nebentypus χ \chi χ , with Laplace eigenvalue 1 / 4 + r 2 1/4+r^2 1/4 + r 2 and r ∈ R r \in\mathbb{R} r ∈ R .
Then for sufficiently large T T T ,
max T ≤ t ≤ 2 T Z F ± ( t ) ≫ exp ( C F log T log log T ) , \max_{T\le t\le2T} Z_F^{\pm}(t) \gg\exp\left(C_F\sqrt{\frac{\log T}{\log\log T}}\right), T ≤ t ≤ 2 T max Z F ± ( t ) ≫ exp ( C F log log T log T ) , where the constant C F C_F C F depends on F ( s ) F(s) F ( s ) , and can be computed using (11) .
The Ramanujan hypothesis in part (ii) of Theorem 1.6 can be removed at the cost of a smaller constant. More precisely, we have the following theorem.
Theorem 1.7. Let F ( s ) F(s) F ( s ) be the L L L -function associated with a normalized primitive Hecke–Maass cusp form of weight 0 0 0 , level N N N and nebentypus χ \chi χ , with Laplace eigenvalue 1 / 4 + r 2 1/4+r^2 1/4 + r 2 , where r ∈ R r \in\mathbb{R} r ∈ R . Then, unconditionally, for sufficiently large T T T ,
max T ≤ t ≤ 2 T Z F ± ( t ) ≫ exp ( ( 25 71 − ϵ ) log T log log T ) , \max_{T\le t\le2T} Z_F^{\pm}(t) \gg\exp\left(\left(\sqrt{\frac{25}{71}}-\epsilon\right)\sqrt{\frac{\log T}{\log\log T}}\right), T ≤ t ≤ 2 T max Z F ± ( t ) ≫ exp ( ( 71 25 − ϵ ) log log T log T ) , where ϵ > 0 \epsilon>0 ϵ > 0 is arbitrarily small.
It is well known that the Dedekind zeta function ζ K ( s ) = ∑ n = 1 ∞ a K ( n ) n − s \zeta_K(s)=\sum_{n=1}^{\infty}a_K(n)n^{-s} ζ K ( s ) = ∑ n = 1 ∞ a K ( n ) n − s attached to a quadratic field K K K has degree 2. However, it does not satisfy Hypothesis 1 uniformly for all x ∈ R x\in\mathbb{R} x ∈ R , since for x = 0 x=0 x = 0 ,
∑ n ≤ T a K ( n ) ∼ r K T , \sum_{n\le T}a_K(n)\sim r_KT, n ≤ T ∑ a K ( n ) ∼ r K T ,
where r K r_K r K denotes the residue of ζ K ( s ) \zeta_K(s) ζ K ( s ) at s = 1 s=1 s = 1 . Therefore, the corresponding omega result for ζ K ( s ) \zeta_K(s) ζ K ( s ) does not follow as a special case of Theorem 1.3 .
We also consider the degree 2 function F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) F(s)=L(s,\chi_1)L(s,\chi_2) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) , where χ 1 \chi_1 χ 1 and χ 2 \chi_2 χ 2 are primitive Dirichlet characters modulo q 1 q_1 q 1 and q 2 q_2 q 2 , respectively, with q 1 q 2 q_1q_2 q 1 q 2 not a perfect square. For this family also, the uniform exponential sum estimate required in our general argument is not directly available in the required form. Therefore, we treat these two families separately in the following theorem.
Theorem 1.8. Let F F F be one of the following functions:
(i) F ( s ) = ζ K ( s ) F(s)=\zeta_K(s) F ( s ) = ζ K ( s ) , where K K K is any quadratic field;
(ii) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) F(s)=L(s,\chi_1)L(s,\chi_2) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) , where χ 1 \chi_1 χ 1 and χ 2 \chi_2 χ 2 are primitive Dirichlet characters modulo q 1 q_1 q 1 and q 2 q_2 q 2 , respectively, with q 1 q 2 q_1q_2 q 1 q 2 not a perfect square.
Then for sufficiently large T T T ,
max T ≤ t ≤ 2 T Z F ± ( t ) ≫ exp ( ( 1 2 − ϵ ) log T log log T ) , \max_{T\le t\le2T} Z_F^{\pm}(t) \gg\exp\left(\left(\frac{1}{\sqrt{2}}-\epsilon\right)\sqrt{\frac{\log T}{\log\log T}}\right), T ≤ t ≤ 2 T max Z F ± ( t ) ≫ exp ( ( 2 1 − ϵ ) log log T log T ) ,
where ϵ > 0 \epsilon>0 ϵ > 0 is an arbitrarily small number.
Strategy of the paper
In this paper, we try to implement the philosophy that if an L L L -function has infinitely many zeros on the critical line and produces large values there by means of a resonator, then, together with suitable cancellation in the corresponding signed resonated moment, one can obtain large positive and negative values of the associated Hardy function of comparable magnitude. This cancellation is essential, since large absolute values and infinitely many zeros on the critical line alone do not guarantee large values of both signs. Accordingly, we exclude cases such as ζ 2 ( s ) \zeta^2(s) ζ 2 ( s ) and L 2 ( s , χ ) L^2(s,\chi) L 2 ( s , χ ) , for which the associated Hardy functions do not exhibit both positive and negative values.
To work on this idea, a natural approach is to use the second author’s argument [22 ] for the Riemann zeta function, with some modifications. This idea works well for Dirichlet L L L -functions but fails for degree 2 L L L -functions in the Selberg class due to the large error in the Voronoi summation formula. In particular, the proposition of [22 ] does not apply here directly. So, we modify it in our new Proposition 2.12 and show that the integral ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t \int_T^{2T} Z_F(t)|R(t)|^2\,dt ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t is small by making a shift in the contour. This idea is well known and has been used in [4 , 24 , 28 ] to show the existence of infinitely many zeros for degree 2 L L L -functions. However, the computation in our case requires uniform exponential-sum estimates involving the Dirichlet coefficients a F ( n ) a_F(n) a F ( n ) of the L L L -function (see (50) ). These estimates are readily available for L L L -functions associated with holomorphic cusp form and Maass cusp form. For the Dedekind zeta function of a quadratic field and for L ( s , χ 1 ) L ( s , χ 2 ) L(s,\chi_1)L(s,\chi_2) L ( s , χ 1 ) L ( s , χ 2 ) , we establish the required exponential-sum estimate in Section 5.
The method is not restricted to functions in the Selberg class and can also be applied to certain L L L -functions outside this class, provided that the corresponding Hardy function and the required exponential-sum estimates are available. We discuss such extensions in Section 6.
Remark 1.9. For L L L -functions of degree 3 or higher, a major difficulty in extending our method is obtaining a sufficiently strong upper bound for ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t \int_T^{2T} Z_F(t)|R(t)|^2\,dt ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t . In particular, the techniques used in this paper do not appear to provide the required cancellation in this integral for higher-degree L L L -functions. Thus, extending our results to degrees greater than 2 remains a challenging problem.
Preliminaries
Asymptotic and order estimates
We know from the Stirling formula for the Γ \Gamma Γ -function that in any fixed vertical strip − ∞ < a ≤ σ ≤ b < ∞ -\infty<a\le\sigma\le b<\infty − ∞ < a ≤ σ ≤ b < ∞ , we have as t → ∞ t\to\infty t → ∞ :
Γ ( σ + i t ) = 2 π t σ − 1 2 + i t e − π 2 t e π i 2 ( σ − 1 2 ) − i t ( 1 + O ( 1 t ) ) . \Gamma(\sigma+it)=\sqrt{2\pi}\,t^{\sigma-\frac{1}{2}+it}e^{-\frac{\pi}{2}t}e^{\frac{\pi i}{2}\left(\sigma-\frac{1}{2}\right)-it}\left(1+O\left(\frac{1}{t}\right)\right). Γ ( σ + i t ) = 2 π t σ − 2 1 + i t e − 2 π t e 2 π i ( σ − 2 1 ) − i t ( 1 + O ( t 1 ) ) .
To apply this to the ratio of Gamma factors in (1.7), let s = σ + i t s=\sigma+it s = σ + i t . For each j = 1 , … , m j=1,\ldots,m j = 1 , … , m , the ratio expands as:
Γ ( α j ( 1 − s ) + β j ‾ ) Γ ( α j s + β j ) = ( α j t + ℑ β j ) α j ( 1 − 2 σ ) − 2 i ( α j t + ℑ β j ) e − π 2 i ( α j + 2 ℜ β j − 1 ) e 2 i ( α j t + ℑ β j ) ( 1 + O ( 1 t ) ) = ( α j t ) α j ( 1 − 2 σ ) − 2 i ( α j t + ℑ β j ) e − 2 i α j t e − π 2 i ( α j + 2 ℜ β j − 1 ) e 2 i α j t e 2 i ℑ β j ( 1 + O ( 1 t ) ) = α j α j ( 1 − 2 s ) − 2 i ℑ β j t α j ( 1 − 2 s ) − 2 i ℑ β j e − π 2 i ( α j + 2 ℜ β j − 1 ) e 2 i α j t ( 1 + O ( 1 t ) ) . (12) \begin{aligned}
\frac{\Gamma(\alpha_j(1-s)+\overline{\beta_j})}{\Gamma(\alpha_js+\beta_j)}
&=(\alpha_jt+\Im\beta_j)^{\alpha_j(1-2\sigma)-2i(\alpha_jt+\Im\beta_j)}e^{-\frac{\pi}{2}i(\alpha_j+2\Re\beta_j-1)}e^{2i(\alpha_jt+\Im\beta_j)}\left(1+O\left(\frac{1}{t}\right)\right)\\
&=(\alpha_jt)^{\alpha_j(1-2\sigma)-2i(\alpha_jt+\Im\beta_j)}e^{-2i\alpha_jt}e^{-\frac{\pi}{2}i(\alpha_j+2\Re\beta_j-1)}e^{2i\alpha_jt}e^{2i\Im\beta_j}\left(1+O\left(\frac{1}{t}\right)\right)\\
&=\alpha_j^{\alpha_j(1-2s)-2i\Im\beta_j}t^{\alpha_j(1-2s)-2i\Im\beta_j}e^{-\frac{\pi}{2}i(\alpha_j+2\Re\beta_j-1)}e^{2i\alpha_jt}\left(1+O\left(\frac{1}{t}\right)\right). \tag*{(12)}
\end{aligned} Γ ( α j s + β j ) Γ ( α j ( 1 − s ) + β j ) = ( α j t + ℑ β j ) α j ( 1 − 2 σ ) − 2 i ( α j t + ℑ β j ) e − 2 π i ( α j + 2ℜ β j − 1 ) e 2 i ( α j t + ℑ β j ) ( 1 + O ( t 1 ) ) = ( α j t ) α j ( 1 − 2 σ ) − 2 i ( α j t + ℑ β j ) e − 2 i α j t e − 2 π i ( α j + 2ℜ β j − 1 ) e 2 i α j t e 2 i ℑ β j ( 1 + O ( t 1 ) ) = α j α j ( 1 − 2 s ) − 2 i ℑ β j t α j ( 1 − 2 s ) − 2 i ℑ β j e − 2 π i ( α j + 2ℜ β j − 1 ) e 2 i α j t ( 1 + O ( t 1 ) ) . (12)
after simplification and by using
( 1 + ℑ β j α j t ) − 2 i α j t = e − 2 i ℑ β j ( 1 + O ( 1 t ) ) . \left(1+\frac{\Im\beta_j}{\alpha_jt}\right)^{-2i\alpha_jt}=e^{-2i\Im\beta_j}\left(1+O\left(\frac{1}{t}\right)\right). ( 1 + α j t ℑ β j ) − 2 i α j t = e − 2 i ℑ β j ( 1 + O ( t 1 ) ) .
Applying this to the full product for H ( s ) H(s) H ( s ) , we have as t → ∞ t\to\infty t → ∞ :
H ( s ) = ω Q 1 − 2 s t ∑ j = 1 m α j ( 1 − 2 s ) − 2 i ∑ j = 1 m ℑ β j e − π 2 i ∑ j = 1 m ( α j + 2 ℜ β j − 1 ) × e 2 i t ∑ j = 1 m α j ∏ j = 1 m α j α j ( 1 − 2 s ) − 2 i ℑ β j ( 1 + O ( 1 t ) ) . \begin{aligned}
H(s)&=\omega Q^{1-2s}t^{\sum_{j=1}^{m}\alpha_j(1-2s)-2i\sum_{j=1}^{m}\Im\beta_j}e^{-\frac{\pi}{2}i\sum_{j=1}^{m}(\alpha_j+2\Re\beta_j-1)}\\
&\quad\times e^{2it\sum_{j=1}^{m}\alpha_j}\prod_{j=1}^{m}\alpha_j^{\alpha_j(1-2s)-2i\Im\beta_j}\left(1+O\left(\frac{1}{t}\right)\right).
\end{aligned} H ( s ) = ω Q 1 − 2 s t ∑ j = 1 m α j ( 1 − 2 s ) − 2 i ∑ j = 1 m ℑ β j e − 2 π i ∑ j = 1 m ( α j + 2ℜ β j − 1 ) × e 2 i t ∑ j = 1 m α j j = 1 ∏ m α j α j ( 1 − 2 s ) − 2 i ℑ β j ( 1 + O ( t 1 ) ) .
Recall the degree d F : = 2 ∑ j = 1 m α j d_F:=2\sum_{j=1}^{m}\alpha_j d F := 2 ∑ j = 1 m α j , the ξ \xi ξ -invariant ξ F : = 2 ∑ j = 1 m ( β j − 1 2 ) \xi_F:=2\sum_{j=1}^{m}\left(\beta_j-\frac{1}{2}\right) ξ F := 2 ∑ j = 1 m ( β j − 2 1 ) , and the conductor q F : = ( 2 π ) d F Q 2 ∏ j = 1 m α j 2 α j q_F:=(2\pi)^{d_F}Q^2\prod_{j=1}^{m}\alpha_j^{2\alpha_j} q F := ( 2 π ) d F Q 2 ∏ j = 1 m α j 2 α j . We simplify the above with d F d_F d F and ξ F \xi_F ξ F , and rewrite H ( s ) H(s) H ( s ) as
H ( s ) = ω o ( Q o t d F / 2 ) 1 − 2 s t − i ℑ ξ F e i d F t ( 1 + O ( 1 t ) ) , (13) H(s)=\omega_o\left(Q_ot^{d_F/2}\right)^{1-2s}t^{-i\Im\xi_F}e^{id_Ft}\left(1+O\left(\frac{1}{t}\right)\right), \tag*{(13)} H ( s ) = ω o ( Q o t d F /2 ) 1 − 2 s t − i ℑ ξ F e i d F t ( 1 + O ( t 1 ) ) , (13)
where ω o = ω e − π i 2 ( d F / 2 + ℜ ξ F ) ∏ j = 1 m α j − 2 i ℑ β j \omega_o=\omega e^{-\frac{\pi i}{2}(d_F/2+\Re\xi_F)}\prod_{j=1}^{m}\alpha_j^{-2i\Im\beta_j} ω o = ω e − 2 π i ( d F /2 + ℜ ξ F ) ∏ j = 1 m α j − 2 i ℑ β j with ∣ ω o ∣ = 1 |\omega_o|=1 ∣ ω o ∣ = 1 and Q o = Q ∏ j = 1 m α j α j > 0 Q_o=Q\prod_{j=1}^{m}\alpha_j^{\alpha_j}>0 Q o = Q ∏ j = 1 m α j α j > 0 . Also note that Q o = q F ( 2 π ) − d F / 2 Q_o=\sqrt{q_F}(2\pi)^{-d_F/2} Q o = q F ( 2 π ) − d F /2 .
In this article, we restrict ourselves to L L L -functions in the Selberg class of degree 2 with ℑ ξ F = 0 \Im\xi_F=0 ℑ ξ F = 0 . More specifically, we investigate Hardy’s Z Z Z -function associated to the Dedekind zeta function over a quadratic field, the L L L -function associated with normalized Hecke primitive eigen cusp forms of weight k k k and level N N N , and the L L L -function associated with a Maass cusp form of weight 0 and level N N N corresponding to an eigenvalue 1 / 4 + r 2 1/4+r^2 1/4 + r 2 of the hyperbolic Laplacian. Thus, for L L L -functions of degree 2 satisfying ℑ ξ F = 0 \Im\xi_F=0 ℑ ξ F = 0 , we have
H ( s ) = ω 1 ( Q o t ) 1 − 2 σ − 2 i t e 2 i t ( 1 + O ( 1 t ) ) , H(s)=\omega_1(Q_ot)^{1-2\sigma-2it}e^{2it}\left(1+O\left(\frac{1}{t}\right)\right), H ( s ) = ω 1 ( Q o t ) 1 − 2 σ − 2 i t e 2 i t ( 1 + O ( t 1 ) ) ,
where ω 1 = ω e − π i 2 ( 1 + ℜ ξ F ) ∏ i = 1 m α i − 2 i ℑ β i \omega_1=\omega e^{-\frac{\pi i}{2}(1+\Re\xi_F)}\prod_{i=1}^{m}\alpha_i^{-2i\Im\beta_i} ω 1 = ω e − 2 π i ( 1 + ℜ ξ F ) ∏ i = 1 m α i − 2 i ℑ β i , with ∣ ω 1 ∣ = 1 |\omega_1|=1 ∣ ω 1 ∣ = 1 and Q o = q F / 2 π Q_o=\sqrt{q_F}/2\pi Q o = q F /2 π .
Therefore,
H − 1 / 2 ( s ) = ω 2 ( Q o t ) σ − 1 / 2 + i t e − i t ( 1 + O ( 1 t ) ) , (14) H^{-1/2}(s)=\omega_2(Q_ot)^{\sigma-1/2+it}e^{-it}\left(1+O\left(\frac{1}{t}\right)\right), \tag*{(14)} H − 1/2 ( s ) = ω 2 ( Q o t ) σ − 1/2 + i t e − i t ( 1 + O ( t 1 ) ) , (14)
where ω 2 = ( ω 1 ) − 1 / 2 \omega_2=(\omega_1)^{-1/2} ω 2 = ( ω 1 ) − 1/2 with ∣ ω 2 ∣ = 1 |\omega_2|=1 ∣ ω 2 ∣ = 1 .
Now to find the order of Hardy’s function and order of its analytic extension in a neighbourhood, we need the following convexity bound for F ( s ) F(s) F ( s ) .
Lemma 2.1. Let F ( s ) ∈ S 2 F(s)\in\mathcal{S}_2 F ( s ) ∈ S 2 . Then for any constant δ > 0 \delta>0 δ > 0 , the L L L -function F ( s ) F(s) F ( s ) satisfies
F ( σ + i t ) = O ( t 1 − σ + δ ) for − δ ≤ σ ≤ 1 + δ . F(\sigma+it)=O\left(t^{1-\sigma+\delta}\right)\qquad\text{for}\qquad-\delta\leq\sigma\leq1+\delta. F ( σ + i t ) = O ( t 1 − σ + δ ) for − δ ≤ σ ≤ 1 + δ .
Proof. Using Stirling’s formula for Γ ( s ) \Gamma(s) Γ ( s ) , from the functional equation (6) , we have for δ > 0 \delta> 0 δ > 0 ,
F ( − δ + i t ) ≪ t 1 + 2 δ . F(-\delta+ it) \ll t^{1+2\delta}. F ( − δ + i t ) ≪ t 1 + 2 δ .
Since the Dirichlet series F ( s ) = ∑ n = 1 ∞ a F ( n ) n s F(s) = \sum_{n=1}^{\infty} \frac{a_F(n)}{n^s} F ( s ) = ∑ n = 1 ∞ n s a F ( n ) converges absolutely in the half-plane ℜ ( s ) > 1 \Re(s) > 1 ℜ ( s ) > 1 , it follows that, for any δ > 0 \delta> 0 δ > 0 ,
F ( 1 + δ + i t ) ≪ 1. F(1 + \delta+ it) \ll1. F ( 1 + δ + i t ) ≪ 1.
Then applying the appropriate Phragmén–Lindelöf [14 Theorem 5.53] argument, we get the required result. □ \square □
Using (14) together with Lemma 2.1, we obtain that for constant δ > 0 \delta> 0 δ > 0 and − δ ≤ σ ≤ 1 + δ -\delta\leq\sigma\leq1+\delta − δ ≤ σ ≤ 1 + δ ,
F ( σ + i t ) H − 1 2 ( σ + i t ) = O ( t 1 2 + δ ) , as t → ∞ . (15) F(\sigma+ it)H^{-\frac{1}{2}}(\sigma+ it) = O\left(t^{\frac{1}{2}+\delta}\right), \quad\text{as } t \to\infty. \tag*{(15)} F ( σ + i t ) H − 2 1 ( σ + i t ) = O ( t 2 1 + δ ) , as t → ∞. (15)
We now turn to certain exponential sums and integral estimates, which are the main tools we need. These estimates will be used both in preparing our auxiliary results and in the final proof of our main theorems.
Exponential sums and integral estimates
We now state the following results on exponential integrals, which will be useful in proving our main results.
Lemma 2.2. [35 Lemma 4.3] Let f ( x ) f(x) f ( x ) and g ( x ) g(x) g ( x ) be real functions such that f ′ ( x ) / g ( x ) f'(x)/g(x) f ′ ( x ) / g ( x ) is monotonic and ∣ f ′ ( x ) / g ( x ) ∣ ≥ m > 0 |f'(x)/g(x)| \geq m > 0 ∣ f ′ ( x ) / g ( x ) ∣ ≥ m > 0 , throughout the interval [ a , b ] [a,b] [ a , b ] . Then
∣ ∫ a b g ( x ) e i f ( x ) d x ∣ ≤ 4 m . \left|\int_a^b g(x)e^{if(x)}\,dx\right| \leq\frac{4}{m}. ∫ a b g ( x ) e i f ( x ) d x ≤ m 4 .
Lemma 2.3. [11 Lemma 2.3] Let f ( x ) f(x) f ( x ) be a real twice differentiable function such that ∣ f ′ ′ ( x ) ∣ ≥ λ 2 > 0 |f''(x)| \geq\lambda_2 > 0 ∣ f ′′ ( x ) ∣ ≥ λ 2 > 0 for x ∈ [ a , b ] x \in[a,b] x ∈ [ a , b ] . Let g ( x ) g(x) g ( x ) be a positive monotonic function such that ∣ g ( x ) ∣ ≤ G |g(x)| \leq G ∣ g ( x ) ∣ ≤ G on [ a , b ] [a,b] [ a , b ] . Then
∣ ∫ a b g ( x ) e i f ( x ) d x ∣ ≤ 8 G λ 2 . \left|\int_a^b g(x)e^{if(x)}\,dx\right| \leq\frac{8G}{\sqrt{\lambda_2}}. ∫ a b g ( x ) e i f ( x ) d x ≤ λ 2 8 G .
Lemma 2.4. [30 Lemma 4.3] Let f ( x ) f(x) f ( x ) be a real function in [ a , b ] [a,b] [ a , b ] , which is three times continuously differentiable, such that f ′ ( x ) f'(x) f ′ ( x ) is monotonic, f ′ ( c ) = 0 f'(c)=0 f ′ ( c ) = 0 for a < c < b a<c<b a < c < b , ∣ f ′ ′ ( x ) ∣ ≍ λ 2 |f''(x)|\asymp\lambda_2 ∣ f ′′ ( x ) ∣ ≍ λ 2 and 0 < ∣ f ′ ′ ′ ( x ) ∣ ≪ λ 3 0<|f'''(x)|\ll\lambda_3 0 < ∣ f ′′′ ( x ) ∣ ≪ λ 3 . Let g ( x ) g(x) g ( x ) be a real function which is two times continuously differentiable, with ∣ g ( x ) ∣ ≤ G |g(x)|\leq G ∣ g ( x ) ∣ ≤ G , ∣ g ′ ( x ) ∣ ≤ G 1 |g'(x)|\leq G_1 ∣ g ′ ( x ) ∣ ≤ G 1 for x ∈ [ a , b ] x\in[a,b] x ∈ [ a , b ] . Suppose that the functions f ′ 6 ( x ) − 8 f ′ ′ ( c ) f ′ ′ 2 ( x ) ( f ( x ) − f ( c ) ) 3 f'^{6}(x)-8f''(c)f''^{2}(x)(f(x)-f(c))^{3} f ′ 6 ( x ) − 8 f ′′ ( c ) f ′′ 2 ( x ) ( f ( x ) − f ( c ) ) 3 and g ′ ′ ( x ) g''(x) g ′′ ( x ) have only a bounded number of zeros in [ a , b ] [a,b] [ a , b ] . Then,
∫ a b g ( x ) e i f ( x ) d x = g ( c ) ( 2 π ) 1 / 2 e ± i π / 4 + i f ( c ) ∣ f ′ ′ ( c ) ∣ 1 / 2 + O ( G 1 λ 2 ) + O ( G ( b − a ) λ 3 2 λ 2 3 ) + O ( G λ 3 λ 2 2 ) + O ( G min ( 1 ∣ f ′ ( a ) ∣ + 1 ∣ f ′ ( b ) ∣ , λ 2 − 1 / 2 ) ) . \begin{aligned}
\int_a^b g(x)e^{if(x)}\,dx
={}&g(c)(2\pi)^{1/2}\frac{e^{\pm i\pi/4+if(c)}}{|f''(c)|^{1/2}}
+O\left(\frac{G_1}{\lambda_2}\right)
+O\left(\frac{G(b-a)\lambda_3^2}{\lambda_2^3}\right)\\
&+O\left(\frac{G\lambda_3}{\lambda_2^2}\right)
+O\left(G\min\left(\frac{1}{|f'(a)|}+\frac{1}{|f'(b)|},\lambda_2^{-1/2}\right)\right).
\end{aligned} ∫ a b g ( x ) e i f ( x ) d x = g ( c ) ( 2 π ) 1/2 ∣ f ′′ ( c ) ∣ 1/2 e ± iπ /4 + i f ( c ) + O ( λ 2 G 1 ) + O ( λ 2 3 G ( b − a ) λ 3 2 ) + O ( λ 2 2 G λ 3 ) + O ( G min ( ∣ f ′ ( a ) ∣ 1 + ∣ f ′ ( b ) ∣ 1 , λ 2 − 1/2 ) ) .
Here, the sign in e ± i π / 4 e^{\pm i\pi/4} e ± iπ /4 is chosen as positive or negative according as f ′ ′ f'' f ′′ is positive or negative.
We next state a bound for exponential sums involving the coefficients of integral weight cusp forms or Maass cusp forms.
Lemma 2.5. Let f f f be either a normalized primitive holomorphic Hecke cusp eigenform of even weight k ≥ 2 k \ge2 k ≥ 2 , or a normalized primitive Hecke–Maass cusp form of weight 0 0 0 with Laplace eigenvalue 1 / 4 + r 2 1/4+r^2 1/4 + r 2 , where r ∈ R r \in\mathbb{R} r ∈ R . In each case, let f f f have level N N N and nebentypus χ \chi χ . If λ f ( n ) \lambda_f(n) λ f ( n ) denotes the normalized Fourier coefficient of f f f , then
∑ n ≤ T λ f ( n ) e 2 π i n x ≪ T 1 / 2 log ( 2 T ) , \sum_{n \le T} \lambda_f(n)e^{2\pi inx} \ll T^{1/2}\log(2T), n ≤ T ∑ λ f ( n ) e 2 π in x ≪ T 1/2 log ( 2 T ) , for T ≥ 1 T \ge1 T ≥ 1 and x ∈ R x \in\mathbb{R} x ∈ R , where the implied constant depends only on f f f .
Proof. The result follows from [13 Lemma 5.3] and [12 Theorem 8.1] . □
We now present the cases needed in this paper for which the Selberg normality conjecture (10) is known to hold.
Lemma 2.6. [10 Theorem 2.2] Let f f f be either a normalized primitive holomorphic Hecke cusp eigenform of even weight k ≥ 2 k \ge2 k ≥ 2 , or a normalized primitive Hecke–Maass cusp form of weight 0 0 0 with Laplace eigenvalue 1 / 4 + r 2 1/4+r^2 1/4 + r 2 , where r ∈ R r \in\mathbb{R} r ∈ R . In each case, let f f f have level N N N and nebentypus χ \chi χ . If λ f ( n ) \lambda_f(n) λ f ( n ) denotes the normalized Fourier coefficient of f f f , then
∑ p ≤ x ∣ λ f ( p ) ∣ 2 = ( 1 + o ( 1 ) ) x log x . \sum_{p \le x} |\lambda_f(p)|^2 = (1+o(1))\frac{x}{\log x}. p ≤ x ∑ ∣ λ f ( p ) ∣ 2 = ( 1 + o ( 1 )) log x x .
Lemma 2.7. Let f f f and λ f ( n ) \lambda_f(n) λ f ( n ) be as in Lemma 2.6. Then we have
∑ p ≤ x ∣ λ f ( p ) ∣ 4 ≪ x log x , \sum_{p \le x} |\lambda_f(p)|^4 \ll\frac{x}{\log x}, p ≤ x ∑ ∣ λ f ( p ) ∣ 4 ≪ log x x , where the implied constant depends on f f f .
Proof. For holomorphic Hecke cusp forms, we have ∣ λ f ( p ) ∣ ≤ 2 |\lambda_f(p)| \le2 ∣ λ f ( p ) ∣ ≤ 2 for p ∤ N p \nmid N p ∤ N . Hence, the result follows in this case.
Now, for Hecke–Maass cusp forms, f f f is either dihedral or non-dihedral. If f f f is dihedral, then ∣ λ f ( p ) ∣ ≤ 2 |\lambda_f(p)| \le2 ∣ λ f ( p ) ∣ ≤ 2 for p ∤ N p \nmid N p ∤ N , and the result follows.
Now assume that f f f is non-dihedral. Let L ( s , Ad ( f ) ) L(s,\operatorname{Ad}(f)) L ( s , Ad ( f )) denote the adjoint square L L L -function of f f f with the Euler factors at primes dividing N N N omitted, defined for ℜ ( s ) > 1 \Re(s)>1 ℜ ( s ) > 1 by
L ( s , Ad ( f ) ) : = ∑ n ≥ 1 A f ( n ) n s , L(s,\operatorname{Ad}(f)) := \sum_{n \ge1}\frac{A_f(n)}{n^s}, L ( s , Ad ( f )) := n ≥ 1 ∑ n s A f ( n ) ,
where
A f ( n ) = { ∑ k 2 ∣ n λ f ( n 2 k 4 ) χ ( n k 2 ) , if ( n , N ) = 1 , 0 , otherwise . A_f(n)=
\begin{cases}
\sum_{k^2\mid n}\lambda_f\left(\frac{n^2}{k^4}\right)\chi\left(\frac{n}{k^2}\right), & \text{if }(n,N)=1,\\
0, & \text{otherwise}.
\end{cases} A f ( n ) = { ∑ k 2 ∣ n λ f ( k 4 n 2 ) χ ( k 2 n ) , 0 , if ( n , N ) = 1 , otherwise .
For p ∤ N p \nmid N p ∤ N , we have
A f ( p ) = λ f ( p 2 ) χ ( p ) ‾ = ∣ λ f ( p ) ∣ 2 − 1 , A_f(p)=\lambda_f(p^2)\overline{\chi(p)}=|\lambda_f(p)|^2-1, A f ( p ) = λ f ( p 2 ) χ ( p ) = ∣ λ f ( p ) ∣ 2 − 1 ,
where the last equality follows from the Hecke relations λ f ( p 2 ) χ ( p ) ‾ = ∣ λ f ( p ) ∣ 2 − 1 \lambda_f(p^2)\overline{\chi(p)}=|\lambda_f(p)|^2-1 λ f ( p 2 ) χ ( p ) = ∣ λ f ( p ) ∣ 2 − 1 . Since f f f is non-dihedral, its adjoint lift is a self-contragredient cuspidal automorphic representation of G L 3 \mathrm{GL}_3 GL 3 . Hence, by [21 Lemma 3.2] and partial summation, we have
∑ p ≤ x ∣ A f ( p ) ∣ 2 ≪ x log x . \sum_{p \le x}|A_f(p)|^2\ll\frac{x}{\log x}. p ≤ x ∑ ∣ A f ( p ) ∣ 2 ≪ log x x .
Therefore,
∑ p ≤ x p ∤ N ∣ λ f ( p ) ∣ 4 = ∑ p ≤ x p ∤ N ( 1 + A f ( p ) ) 2 ≪ ∑ p ≤ x p ∤ N ( 1 + ∣ A f ( p ) ∣ 2 ) ≪ x log x . \sum_{\substack{p\le x\\p\nmid N}} |\lambda_f(p)|^4
= \sum_{\substack{p\le x\\p\nmid N}} (1+A_f(p))^2
\ll\sum_{\substack{p\le x\\p\nmid N}} (1+|A_f(p)|^2)
\ll\frac{x}{\log x}. p ≤ x p ∤ N ∑ ∣ λ f ( p ) ∣ 4 = p ≤ x p ∤ N ∑ ( 1 + A f ( p ) ) 2 ≪ p ≤ x p ∤ N ∑ ( 1 + ∣ A f ( p ) ∣ 2 ) ≪ log x x .
Since the primes dividing N N N contribute O ( 1 ) O(1) O ( 1 ) , the desired result follows. □
Lemma 2.8. [26 Lemma 5.2] Let ζ K ( s ) = ∑ n = 1 ∞ a K ( n ) n − s \zeta_K(s)=\sum_{n=1}^{\infty}a_K(n)n^{-s} ζ K ( s ) = ∑ n = 1 ∞ a K ( n ) n − s be the Dedekind zeta function associated to any quadratic field K K K . Then
∑ p ≤ x a K 2 ( p ) = ( 2 + o ( 1 ) ) x log x . \sum_{p\le x} a_K^2(p)=(2+o(1))\frac{x}{\log x}. p ≤ x ∑ a K 2 ( p ) = ( 2 + o ( 1 )) log x x .
Lemma 2.9. Let F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) F(s)=L(s,\chi_1)L(s,\chi_2) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) , where χ 1 \chi_1 χ 1 and χ 2 \chi_2 χ 2 are primitive Dirichlet characters modulo q 1 q_1 q 1 and q 2 q_2 q 2 , respectively, with q 1 q 2 q_1q_2 q 1 q 2 not a perfect square. Then
∑ p ≤ x ∣ a F ( p ) ∣ 2 = ( 2 + o ( 1 ) ) x log x . \sum_{p\le x}|a_F(p)|^2=(2+o(1))\frac{x}{\log x}. p ≤ x ∑ ∣ a F ( p ) ∣ 2 = ( 2 + o ( 1 )) log x x . Proof. Since a F ( p ) = χ 1 ( p ) + χ 2 ( p ) a_F(p)=\chi_1(p)+\chi_2(p) a F ( p ) = χ 1 ( p ) + χ 2 ( p ) , we have
∣ a F ( p ) ∣ 2 = ∣ χ 1 ( p ) ∣ 2 + ∣ χ 2 ( p ) ∣ 2 + χ 1 ( p ) χ 2 ( p ) ‾ + χ 1 ( p ) ‾ χ 2 ( p ) . |a_F(p)|^2=|\chi_1(p)|^2+|\chi_2(p)|^2+\chi_1(p)\overline{\chi_2(p)}+\overline{\chi_1(p)}\chi_2(p). ∣ a F ( p ) ∣ 2 = ∣ χ 1 ( p ) ∣ 2 + ∣ χ 2 ( p ) ∣ 2 + χ 1 ( p ) χ 2 ( p ) + χ 1 ( p ) χ 2 ( p ) . View both characters modulo lcm ( q 1 , q 2 ) \operatorname{lcm}(q_1,q_2) lcm ( q 1 , q 2 ) . Since q 1 q 2 q_1q_2 q 1 q 2 is not a perfect square, we have q 1 ≠ q 2 q_1\ne q_2 q 1 = q 2 ; hence, by primitivity, χ 1 ≠ χ 2 \chi_1\ne\chi_2 χ 1 = χ 2 , and therefore χ 1 χ 2 ‾ \chi_1\overline{\chi_2} χ 1 χ 2 is nonprincipal. By the prime number theorem for Dirichlet characters, the two cross terms contribute o ( x / log x ) o(x/\log x) o ( x / log x ) , while each diagonal term contributes ( 1 + o ( 1 ) ) x / log x (1+o(1))x/\log x ( 1 + o ( 1 )) x / log x . This proves the result. □
Resonance method
The resonator R ( t ) R(t) R ( t ) constructed by Soundararajan in [33 ] has the form of a Dirichlet polynomial
R ( t ) = ∑ m ≤ N r ( m ) m − i t with N ≤ T α , α = 1 − ϵ , ϵ > 0. (16) R(t)=\sum_{m\le N}r(m)m^{-it}\quad\text{with}\quad N\le T^\alpha,\ \alpha=1-\epsilon,\ \epsilon>0. \tag*{(16)} R ( t ) = m ≤ N ∑ r ( m ) m − i t with N ≤ T α , α = 1 − ϵ , ϵ > 0. (16)
Choose a smooth function Φ \Phi Φ supported on [ 1 , 2 ] [1,2] [ 1 , 2 ] with 0 ≤ Φ ( t ) ≤ 1 0\le\Phi(t)\le1 0 ≤ Φ ( t ) ≤ 1 , and assume that Φ ( t ) = 1 \Phi(t)=1 Φ ( t ) = 1 for t ∈ ( 5 / 4 , 7 / 4 ) t\in(5/4,7/4) t ∈ ( 5/4 , 7/4 ) . Define its Fourier transform by
Φ ^ ( y ) = ∫ R Φ ( t ) e − i t y d t . \widehat{\Phi}(y)=\int_{\mathbb{R}}\Phi(t)e^{-ity}\,dt. Φ ( y ) = ∫ R Φ ( t ) e − i t y d t .
Then, for any fixed integer A ≥ 1 A\ge1 A ≥ 1 , it satisfies the decay estimate Φ ^ ( y ) ≪ ∣ y ∣ − A \widehat{\Phi}(y)\ll|y|^{-A} Φ ( y ) ≪ ∣ y ∣ − A .
For a small 0 < η < 1 0<\eta<1 0 < η < 1 . Put
L 2 = ( 1 − η ) log N log log N κ , M = exp ( ( log L ) 2 ) , (17) L^2=\frac{(1-\eta)\log N\log\log N}{\kappa},\qquad M=\exp\bigl((\log L)^2\bigr), \tag*{(17)} L 2 = κ ( 1 − η ) log N log log N , M = exp ( ( log L ) 2 ) , (17)
and define the nonnegative multiplicative function f f f , supported on squarefree integers, by
f ( p ) = { L p log p , L 2 ≤ p ≤ M , 0 , otherwise , (18) f(p)=
\begin{cases}
\dfrac{L}{\sqrt{p}\log p}, & L^2\le p\le M,\\
0, & \text{otherwise},
\end{cases} \tag*{(18)} f ( p ) = ⎩ ⎨ ⎧ p log p L , 0 , L 2 ≤ p ≤ M , otherwise , (18)
with f ( p j ) = 0 f(p^j)=0 f ( p j ) = 0 for j ≥ 2 j\ge2 j ≥ 2 . Set r ( n ) = a F ( n ) f ( n ) r(n)=a_F(n)f(n) r ( n ) = a F ( n ) f ( n ) .
Lemma 2.10. Let F ( s ) = ∑ n = 1 ∞ a F ( n ) n − s F(s)=\sum_{n=1}^{\infty}a_F(n)n^{-s} F ( s ) = ∑ n = 1 ∞ a F ( n ) n − s , where a F ( n ) a_F(n) a F ( n ) is multiplicative and a F ( 1 ) = 1 a_F(1)=1 a F ( 1 ) = 1 . Suppose that the coefficients a F ( n ) a_F(n) a F ( n ) satisfy (1.10) and
∑ p ≤ x ∣ a F ( p ) ∣ 2 ≪ x log x . (19) \sum_{p\le x}|a_F(p)|^2 \ll\frac{x}{\log x}. \tag*{(19)} p ≤ x ∑ ∣ a F ( p ) ∣ 2 ≪ log x x . (19) Then, for the preceding resonator,
1 ∑ n ≤ N ∣ r ( n ) ∣ 2 ∑ n k ≤ N a F ( k ) r ( n ) r ( n k ) ‾ k ≥ exp ( ( κ ( 1 − η ) + o ( 1 ) ) log N log log N ) . \frac{1}{\sum_{n\le N}|r(n)|^2}\sum_{nk\le N}\frac{a_F(k)r(n)\overline{r(nk)}}{\sqrt{k}}\ge\exp\left(\left(\sqrt{\kappa(1-\eta)}+o(1)\right)\sqrt{\frac{\log N}{\log\log N}}\right). ∑ n ≤ N ∣ r ( n ) ∣ 2 1 nk ≤ N ∑ k a F ( k ) r ( n ) r ( nk ) ≥ exp ( ( κ ( 1 − η ) + o ( 1 ) ) log log N log N ) . In particular, this bound holds for F ∈ S p o l y F\in\mathcal{S}_{\mathrm{poly}} F ∈ S poly satisfies (1.10).
Proof. Since the nonnegative multiplicative function f f f is supported on squarefree integers and a F ( n ) a_F(n) a F ( n ) is multiplicative, we have
∑ n k ≤ N a F ( k ) r ( n ) r ( n k ) ‾ k = ∑ n k ≤ N ( n , k ) = 1 ∣ a F ( k ) ∣ 2 ∣ a F ( n ) ∣ 2 f ( k ) f ( n ) 2 k = : S N , \sum_{nk\le N}\frac{a_F(k)r(n)\overline{r(nk)}}{\sqrt{k}}=\sum_{\substack{nk\le N\\(n,k)=1}}\frac{|a_F(k)|^2|a_F(n)|^2f(k)f(n)^2}{\sqrt{k}}=:S_N, nk ≤ N ∑ k a F ( k ) r ( n ) r ( nk ) = nk ≤ N ( n , k ) = 1 ∑ k ∣ a F ( k ) ∣ 2 ∣ a F ( n ) ∣ 2 f ( k ) f ( n ) 2 =: S N , which is a non-negative real sum. Removing the restriction n k ≤ N nk\le N nk ≤ N , let
P = ∏ p ( 1 + ∣ a F ( p ) ∣ 2 f ( p ) 2 + ∣ a F ( p ) ∣ 2 f ( p ) p ) = ∏ p ( 1 + u p + v p ) , D = ∏ p ( 1 + ∣ a F ( p ) ∣ 2 f ( p ) 2 ) = ∏ p ( 1 + u p ) . (20) \begin{aligned}
P&=\prod_p\left(1+|a_F(p)|^2f(p)^2+\frac{|a_F(p)|^2f(p)}{\sqrt p}\right)=\prod_p(1+u_p+v_p),\\
D&=\prod_p(1+|a_F(p)|^2f(p)^2)=\prod_p(1+u_p). \tag*{(20)}
\end{aligned} P D = p ∏ ( 1 + ∣ a F ( p ) ∣ 2 f ( p ) 2 + p ∣ a F ( p ) ∣ 2 f ( p ) ) = p ∏ ( 1 + u p + v p ) , = p ∏ ( 1 + ∣ a F ( p ) ∣ 2 f ( p ) 2 ) = p ∏ ( 1 + u p ) . (20) where
u p = ∣ a F ( p ) ∣ 2 f ( p ) 2 , v p = ∣ a F ( p ) ∣ 2 f ( p ) p . u_p=|a_F(p)|^2f(p)^2,\qquad v_p=\frac{|a_F(p)|^2f(p)}{\sqrt p}. u p = ∣ a F ( p ) ∣ 2 f ( p ) 2 , v p = p ∣ a F ( p ) ∣ 2 f ( p ) . Note that ∑ n ≤ N ∣ r ( n ) ∣ 2 ≤ D \sum_{n\le N}|r(n)|^2\le D ∑ n ≤ N ∣ r ( n ) ∣ 2 ≤ D and set ρ = ( log L ) − 3 \rho=(\log L)^{-3} ρ = ( log L ) − 3 . Rankin's trick gives
0 ≤ P − S N ≤ N − ρ ∏ p ( 1 + ( u p + v p ) p ρ ) . 0\le P-S_N\le N^{-\rho}\prod_p(1+(u_p+v_p)p^\rho). 0 ≤ P − S N ≤ N − ρ p ∏ ( 1 + ( u p + v p ) p ρ ) . Hence, we obtain
P − S N P ≤ exp ( − ρ log N + ∑ p ( u p + v p ) ( p ρ − 1 ) ) . \frac{P-S_N}{P}\le\exp\left(-\rho\log N+\sum_p(u_p+v_p)(p^\rho-1)\right). P P − S N ≤ exp ( − ρ log N + p ∑ ( u p + v p ) ( p ρ − 1 ) ) . Since ρ log M = ( log L ) − 1 = o ( 1 ) \rho\log M=(\log L)^{-1}=o(1) ρ log M = ( log L ) − 1 = o ( 1 ) , we have p ρ − 1 = ( 1 + o ( 1 ) ) ρ log p p^\rho-1=(1+o(1))\rho\log p p ρ − 1 = ( 1 + o ( 1 )) ρ log p . Partial summation gives
∑ L 2 ≤ p ≤ M ∣ a F ( p ) ∣ 2 p log p = ( κ + o ( 1 ) ) ( 1 log ( L 2 ) − 1 log M ) = κ + o ( 1 ) 2 log ( L ) . \sum_{L^2\le p\le M}\frac{|a_F(p)|^2}{p\log p}=(\kappa+o(1))\left(\frac{1}{\log(L^2)}-\frac{1}{\log M}\right)=\frac{\kappa+o(1)}{2\log(L)}. L 2 ≤ p ≤ M ∑ p log p ∣ a F ( p ) ∣ 2 = ( κ + o ( 1 )) ( log ( L 2 ) 1 − log M 1 ) = 2 log ( L ) κ + o ( 1 ) . Consequently,
∑ p ∣ a F ( p ) ∣ 2 ( f ( p ) 2 + f ( p ) p ) log p = ( 1 − η + o ( 1 ) ) log N . \sum_p|a_F(p)|^2\left(f(p)^2+\frac{f(p)}{\sqrt p}\right)\log p=(1-\eta+o(1))\log N. p ∑ ∣ a F ( p ) ∣ 2 ( f ( p ) 2 + p f ( p ) ) log p = ( 1 − η + o ( 1 )) log N . Thus
P − S N P ≤ exp ( − ( η + o ( 1 ) ) ρ log N ) = o ( 1 ) . \frac{P-S_N}{P}\le\exp\left(-(\eta+o(1))\rho\log N\right)=o(1). P P − S N ≤ exp ( − ( η + o ( 1 )) ρ log N ) = o ( 1 ) .
Hence S N ≥ ( 1 − o ( 1 ) ) P S_N \ge(1-o(1))P S N ≥ ( 1 − o ( 1 )) P . Now
log P D = ∑ p log ( 1 + v p 1 + u p ) = ∑ p ( v p 1 + u p + O ( ( v p 1 + u p ) 2 ) ) . \log\frac{P}{D}=\sum_p \log\left(1+\frac{v_p}{1+u_p}\right)=\sum_p\left(\frac{v_p}{1+u_p}+O\left(\left(\frac{v_p}{1+u_p}\right)^2\right)\right). log D P = p ∑ log ( 1 + 1 + u p v p ) = p ∑ ( 1 + u p v p + O ( ( 1 + u p v p ) 2 ) ) .
Since
( v p 1 + u p ) 2 ≪ v p 2 and v p 1 + u p = v p + O ( u p v p ) , \left(\frac{v_p}{1+u_p}\right)^2\ll v_p^2 \qquad\text{and}\qquad\frac{v_p}{1+u_p}=v_p+O(u_pv_p), ( 1 + u p v p ) 2 ≪ v p 2 and 1 + u p v p = v p + O ( u p v p ) ,
we get
log P D = ∑ p v p + O ( ∑ p ( u p v p + v p 2 ) ) . \log\frac{P}{D}=\sum_p v_p+O\left(\sum_p(u_pv_p+v_p^2)\right). log D P = p ∑ v p + O ( p ∑ ( u p v p + v p 2 ) ) .
Using (2.7), we have obtained
log P D = L ∑ L 2 ≤ p ≤ M ∣ a F ( p ) ∣ 2 p log p + O ( L 3 ∑ L 2 ≤ p ≤ M ∣ a F ( p ) ∣ 4 p 2 ( log p ) 3 + L 2 ∑ L 2 ≤ p ≤ M ∣ a F ( p ) ∣ 4 p 2 ( log p ) 2 ) . (21) \log\frac{P}{D}=L\sum_{L^2\le p\le M}\frac{|a_F(p)|^2}{p\log p}+O\left(L^3\sum_{L^2\le p\le M}\frac{|a_F(p)|^4}{p^2(\log p)^3}+L^2\sum_{L^2\le p\le M}\frac{|a_F(p)|^4}{p^2(\log p)^2}\right). \tag*{(21)} log D P = L L 2 ≤ p ≤ M ∑ p log p ∣ a F ( p ) ∣ 2 + O L 3 L 2 ≤ p ≤ M ∑ p 2 ( log p ) 3 ∣ a F ( p ) ∣ 4 + L 2 L 2 ≤ p ≤ M ∑ p 2 ( log p ) 2 ∣ a F ( p ) ∣ 4 . (21)
Using (1.10), (2.6), and partial summation, the main term of the above expression is
( κ + o ( 1 ) ) L 2 log L . (\kappa+o(1))\frac{L}{2\log L}. ( κ + o ( 1 )) 2 log L L .
For the error term, we use the bound (2.8). Indeed, partial summation on (2.8) gives
∑ y ≤ p ∣ a F ∣ 4 p 2 ( log p ) j ≪ 1 y ( log y ) j + 1 ( j = 1 , 2 ) , \sum_{y\le p}\frac{|a_F|^4}{p^2(\log p)^j}\ll\frac{1}{y(\log y)^{j+1}}\qquad(j=1,2), y ≤ p ∑ p 2 ( log p ) j ∣ a F ∣ 4 ≪ y ( log y ) j + 1 1 ( j = 1 , 2 ) ,
so that the error term in (2.10) is
≪ L ( log L ) 4 + 1 ( log L ) 3 = o ( L log L ) . \ll\frac{L}{(\log L)^4}+\frac{1}{(\log L)^3}=o\left(\frac{L}{\log L}\right). ≪ ( log L ) 4 L + ( log L ) 3 1 = o ( log L L ) .
Hence
log P D = ( κ + o ( 1 ) ) L 2 log L = ( κ ( 1 − η ) + o ( 1 ) ) log N log log N , \begin{aligned}
\log\frac{P}{D}=(\kappa+o(1))\frac{L}{2\log L} \\
&=\left(\sqrt{\kappa(1-\eta)}+o(1)\right)\sqrt{\frac{\log N}{\log\log N}},
\end{aligned} log D P = ( κ + o ( 1 )) 2 log L L = ( κ ( 1 − η ) + o ( 1 ) ) log log N log N ,
by using (2.6). This completes the proof.
The proof of Theorem 1.2 is based on the following fundamental observation, which is analogous to the Proposition following Theorem 1 in [22 ] . We do not include the proof here since it follows in exactly the same way as done in [22 ] .
Proposition 2.11. Let F F F be an element of the Selberg class S \mathcal{S} S and Z F ( t ) Z_F(t) Z F ( t ) denote Hardy’s Z Z Z -function corresponding to F F F . Suppose there exists a non-negative function K ( t ) K(t) K ( t ) and a function A ( T ) A(T) A ( T ) such that
A ( T ) ∫ T 2 T K ( t ) d t ≪ ∫ T 2 T F ( 1 / 2 + i t ) K ( t ) d t , A(T)\int_T^{2T}K(t)\,dt\ll\int_T^{2T}F(1/2+it)K(t)\,dt, A ( T ) ∫ T 2 T K ( t ) d t ≪ ∫ T 2 T F ( 1/2 + i t ) K ( t ) d t , and
∫ T 2 T Z F ( t ) K ( t ) d t = o ( A ( T ) ∫ T 2 T K ( t ) d t ) , \int_T^{2T}Z_F(t)K(t)\,dt=o\left(A(T)\int_T^{2T}K(t)\,dt\right), ∫ T 2 T Z F ( t ) K ( t ) d t = o ( A ( T ) ∫ T 2 T K ( t ) d t ) ,
then
max T ≤ t ≤ 2 T Z F + ( t ) , max T ≤ t ≤ 2 T Z F − ( t ) ≫ A ( T ) . \max_{T\le t\le2T} Z_F^+(t),\ \max_{T\le t\le2T} Z_F^-(t) \gg A(T). T ≤ t ≤ 2 T max Z F + ( t ) , T ≤ t ≤ 2 T max Z F − ( t ) ≫ A ( T ) .
The proof of Theorem 1.6 is based on the following proposition, which is similar to the one given above.
Proposition 2.12. Let F F F be an element of the Selberg class S \mathbb{S} S and Z F ( t ) Z_F(t) Z F ( t ) denote Hardy’s Z Z Z -function corresponding to F F F . Suppose there is a function A ( T ) A(T) A ( T ) such that
A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≪ ∫ T 2 T F ( 1 / 2 + i t ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t , A(T)\int_T^{2T}|R(t)|^2\Phi(t/T)\,\mathrm{d}t \ll\int_T^{2T}F(1/2+\mathrm{i}t)|R(t)|^2\Phi(t/T)\,\mathrm{d}t, A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≪ ∫ T 2 T F ( 1/2 + i t ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t , and
∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t = o ( A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 d t ) , \int_T^{2T}Z_F(t)|R(t)|^2\,\mathrm{d}t=o\left(A(T)\int_T^{2T}|R(t)|^2\,\mathrm{d}t\right), ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t = o ( A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 d t ) , then
max T ≤ t ≤ 2 T Z F + ( t ) , max T ≤ t ≤ 2 T Z F − ( t ) ≫ A ( T ) . \max_{T\le t\le2T} Z_F^+(t),\ \max_{T\le t\le2T} Z_F^-(t) \gg A(T). T ≤ t ≤ 2 T max Z F + ( t ) , T ≤ t ≤ 2 T max Z F − ( t ) ≫ A ( T ) . where R ( t ) R(t) R ( t ) , Φ ( t ) \Phi(t) Φ ( t ) are as defined above Proposition 2.11 and N = ⌊ T α ⌋ N=\lfloor T^\alpha\rfloor N = ⌊ T α ⌋ with 0 < α < 1 0<\alpha<1 0 < α < 1 .
Proof. To prove the claim, observe that there exists a constant C 1 > 0 C_1>0 C 1 > 0 for which
C 1 A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≤ ∫ T 2 T ∣ Z F ( t ) ∣ ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≤ ∫ T 2 T ∣ Z F ( t ) ∣ ∣ R ( t ) ∣ 2 d t . (22) C_1A(T)\int_T^{2T}|R(t)|^2\Phi(t/T)\,\mathrm{d}t\le\int_T^{2T}|Z_F(t)||R(t)|^2\Phi(t/T)\,\mathrm{d}t\le\int_T^{2T}|Z_F(t)||R(t)|^2\,\mathrm{d}t. \tag*{(22)} C 1 A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≤ ∫ T 2 T ∣ Z F ( t ) ∣∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≤ ∫ T 2 T ∣ Z F ( t ) ∣∣ R ( t ) ∣ 2 d t . (22)
Since R ( t ) = ∑ n ≤ N r ( n ) n − i t R(t)=\sum_{n\le N}r(n)n^{-\mathrm{i}t} R ( t ) = ∑ n ≤ N r ( n ) n − i t , by the mean-value theorem for Dirichlet polynomials (see [27 Theorem 6.1] ), we have
∫ T 2 T ∣ R ( t ) ∣ 2 d t = ( T + O ( N ) ) ∑ n ≤ N ∣ r ( n ) ∣ 2 = T ( 1 + O ( T α − 1 ) ) ∑ n ≤ N ∣ r ( n ) ∣ 2 ≪ T ∑ n ≤ N ∣ r ( n ) ∣ 2 . (23) \begin{aligned}
\int_T^{2T}|R(t)|^2\,\mathrm{d}t
&=(T+O(N))\sum_{n\le N}|r(n)|^2\\
&=T(1+O(T^{\alpha-1}))\sum_{n\le N}|r(n)|^2\\
&\ll T\sum_{n\le N}|r(n)|^2. \tag*{(23)}
\end{aligned} ∫ T 2 T ∣ R ( t ) ∣ 2 d t = ( T + O ( N )) n ≤ N ∑ ∣ r ( n ) ∣ 2 = T ( 1 + O ( T α − 1 )) n ≤ N ∑ ∣ r ( n ) ∣ 2 ≪ T n ≤ N ∑ ∣ r ( n ) ∣ 2 . (23)
where we have used N ≤ T α N\le T^\alpha N ≤ T α with 0 < α < 1 0<\alpha<1 0 < α < 1 . From [33 Eq. (2)] , we have
∫ T 2 T ∣ R ( t ) ∣ 2 d t ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≪ T ∑ n ≤ N ∣ r ( n ) ∣ 2 T Φ ^ ( 0 ) ( 1 + O ( 1 / T ) ) ∑ n ≤ N ∣ r ( n ) ∣ 2 ≪ 1. \frac{\int_T^{2T}|R(t)|^2\,\mathrm{d}t}{\int_T^{2T}|R(t)|^2\Phi(t/T)\,\mathrm{d}t}
\ll\frac{ T\sum_{n\le N}|r(n)|^2}{ T\widehat{\Phi}(0)(1+O(1/T))\sum_{n\le N}|r(n)|^2}\ll1. ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ∫ T 2 T ∣ R ( t ) ∣ 2 d t ≪ T Φ ( 0 ) ( 1 + O ( 1/ T )) ∑ n ≤ N ∣ r ( n ) ∣ 2 T ∑ n ≤ N ∣ r ( n ) ∣ 2 ≪ 1.
Hence, there exists a constant C 2 > 0 C_2>0 C 2 > 0 such that
∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≥ C 2 ∫ T 2 T ∣ R ( t ) ∣ 2 d t . \int_T^{2T}|R(t)|^2\Phi(t/T)\,\mathrm{d}t\ge C_2\int_T^{2T}|R(t)|^2\,\mathrm{d}t. ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≥ C 2 ∫ T 2 T ∣ R ( t ) ∣ 2 d t .
Assume for contradiction that the claim is false and that
max T ≤ t ≤ 2 T Z F − ( t ) ≤ C 1 C 2 A ( T ) 3 . \max_{T\le t\le2T}Z_F^-(t)\le\frac{C_1C_2A(T)}{3}. T ≤ t ≤ 2 T max Z F − ( t ) ≤ 3 C 1 C 2 A ( T ) .
Then
∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t = ∫ T 2 T ∣ Z F ( t ) ∣ ∣ R ( t ) ∣ 2 d t − ∫ T 2 T 2 Z F − ( t ) ∣ R ( t ) ∣ 2 d t ≥ C 1 A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t − ∫ T 2 T 2 Z F − ( t ) ∣ R ( t ) ∣ 2 d t ≥ C 1 C 2 A ( T ) 3 ∫ T 2 T ∣ R ( t ) ∣ 2 d t . \begin{aligned}
\int_T^{2T}Z_F(t)|R(t)|^2\,\mathrm{d}t
&=\int_T^{2T}|Z_F(t)||R(t)|^2\,\mathrm{d}t-\int_T^{2T}2Z_F^-(t)|R(t)|^2\,\mathrm{d}t\\
&\ge C_1A(T)\int_T^{2T}|R(t)|^2\Phi(t/T)\,\mathrm{d}t-\int_T^{2T}2Z_F^-(t)|R(t)|^2\,\mathrm{d}t\\
&\ge\frac{C_1C_2A(T)}{3}\int_T^{2T}|R(t)|^2\,\mathrm{d}t.
\end{aligned} ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t = ∫ T 2 T ∣ Z F ( t ) ∣∣ R ( t ) ∣ 2 d t − ∫ T 2 T 2 Z F − ( t ) ∣ R ( t ) ∣ 2 d t ≥ C 1 A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t − ∫ T 2 T 2 Z F − ( t ) ∣ R ( t ) ∣ 2 d t ≥ 3 C 1 C 2 A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 d t .
which contradicts the fact that ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t \int_T^{2T} Z_F(t)|R(t)|^2\,dt ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t is o ( A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 d t ) o\left(A(T)\int_T^{2T}|R(t)|^2\,dt\right) o ( A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 d t ) . This proves our claim for Z F − ( t ) Z_F^-(t) Z F − ( t ) . Similarly we can argue for Z F + ( t ) Z_F^+(t) Z F + ( t ) .
Proof of Theorem 1.2
Let F F F be an element of the Selberg class of degree 1. We know that the class S 1 S_1 S 1 consists of the Riemann zeta function ζ ( s ) \zeta(s) ζ ( s ) and the shifted Dirichlet L L L -functions L θ ( s , χ ) : = L ( s + i θ , χ ) L_\theta(s,\chi):=L(s+i\theta,\chi) L θ ( s , χ ) := L ( s + i θ , χ ) . Since the desired result is already known for the Riemann zeta function by (9) , it remains to consider the latter case in the proof of Theorem 1.2.
Let L ( s , χ ) L(s,\chi) L ( s , χ ) denote the Dirichlet L L L -function associated with a primitive Dirichlet character χ \chi χ of conductor q q q . From the functional equation for L ( s , χ ) L(s,\chi) L ( s , χ ) , we have
L ( s + i θ , χ ) = ψ ( s + i θ , χ ) L ( 1 − s − i θ , χ ‾ ) , (24) L(s+i\theta,\chi)=\psi(s+i\theta,\chi)L(1-s-i\theta,\overline{\chi}), \tag*{(24)} L ( s + i θ , χ ) = ψ ( s + i θ , χ ) L ( 1 − s − i θ , χ ) , (24)
with
ψ ( s , χ ) = m ( χ ) ( π q ) s − 1 2 Γ ( ( 1 − s + α ) / 2 ) Γ ( ( s + α ) / 2 ) , \psi(s,\chi)=\mathfrak{m}(\chi)\left(\frac{\pi}{q}\right)^{s-\frac{1}{2}}\frac{\Gamma\left((1-s+\alpha)/2\right)}{\Gamma\left((s+\alpha)/2\right)}, ψ ( s , χ ) = m ( χ ) ( q π ) s − 2 1 Γ ( ( s + α ) /2 ) Γ ( ( 1 − s + α ) /2 ) ,
where α = ( 1 − χ ( − 1 ) ) / 2 \alpha=(1-\chi(-1))/2 α = ( 1 − χ ( − 1 )) /2 measures the parity of the character χ \chi χ , m ( χ ) = τ ( χ ) / ( i α q ) \mathfrak{m}(\chi)=\tau(\chi)/(i^\alpha\sqrt{q}) m ( χ ) = τ ( χ ) / ( i α q ) and τ ( χ ) = ∑ a ( mod q ) χ ( a ) e ( a / q ) \tau(\chi)=\sum_{a\;(\operatorname{mod} q)}\chi(a)e(a/q) τ ( χ ) = ∑ a ( mod q ) χ ( a ) e ( a / q ) is the Gauss sum; m ( χ ) \mathfrak{m}(\chi) m ( χ ) is the root number. Comparing (3.1) with (6) , we see that Hardy’s Z Z Z -function corresponding to L ( s + i θ , χ ) L(s+i\theta,\chi) L ( s + i θ , χ ) is
Z θ ( t , χ ) : = L ( 1 / 2 + i t + i θ , χ ) ψ ( 1 / 2 + i t + i θ , χ ) − 1 / 2 , (25) Z_\theta(t,\chi):=L(1/2+it+i\theta,\chi)\psi(1/2+it+i\theta,\chi)^{-1/2}, \tag*{(25)} Z θ ( t , χ ) := L ( 1/2 + i t + i θ , χ ) ψ ( 1/2 + i t + i θ , χ ) − 1/2 , (25)
that is, a translate of Hardy’s function Z ( t , χ ) Z(t,\chi) Z ( t , χ ) corresponding to L ( s , χ ) L(s,\chi) L ( s , χ ) . Thus for our purpose, it suffices to work with Z ( t , χ ) Z(t,\chi) Z ( t , χ ) . The following lemma gives the approximate functional equation for Z ( t , χ ) Z(t,\chi) Z ( t , χ ) .
Lemma 3.1. For fixed χ \chi χ and t → ∞ t\to\infty t → ∞ , we have
Z ( t , χ ) = 2 ℜ { c χ ∑ n ≤ q ∣ t ∣ / 2 π χ ( n ) n exp ( i t ( 1 2 log ( q t 2 π e ) − log n ) ) } + O ( ∣ t ∣ − 1 / 4 ) , (26) Z(t,\chi)=2\Re\left\{\mathfrak{c}_\chi\sum_{n\le\sqrt{q|t|/2\pi}}\frac{\chi(n)}{\sqrt{n}}\exp\left(it\left(\frac{1}{2}\log\left(\frac{qt}{2\pi e}\right)-\log n\right)\right)\right\}+O(|t|^{-1/4}), \tag*{(26)} Z ( t , χ ) = 2ℜ ⎩ ⎨ ⎧ c χ n ≤ q ∣ t ∣/2 π ∑ n χ ( n ) exp ( i t ( 2 1 log ( 2 π e q t ) − log n ) ) ⎭ ⎬ ⎫ + O ( ∣ t ∣ − 1/4 ) , (26) where c χ = m ( χ ) − 1 / 2 e ( π i α / 4 ) − ( π i / 8 ) \mathfrak{c}_\chi=\mathfrak{m}(\chi)^{-1/2}e^{(\pi i\alpha/4)-(\pi i/8)} c χ = m ( χ ) − 1/2 e ( π i α /4 ) − ( π i /8 ) .
Proof. From [3 Eq. (44)] with x = y = q ∣ t ∣ / 2 π x=y=\sqrt{q|t|/2\pi} x = y = q ∣ t ∣/2 π , we have
L ( 1 2 + i t , χ ) = ∑ n ≤ q ∣ t ∣ / 2 π χ ( n ) n 1 / 2 + i t + ψ ( 1 2 + i t , χ ) ∑ n ≤ q ∣ t ∣ / 2 π χ ( n ) ‾ n 1 / 2 − i t + O ( ∣ t ∣ − 1 / 4 ) . L\left(\frac{1}{2}+it,\chi\right)=\sum_{n\le\sqrt{q|t|/2\pi}}\frac{\chi(n)}{n^{1/2+it}}+\psi\left(\frac{1}{2}+it,\chi\right)\sum_{n\le\sqrt{q|t|/2\pi}}\frac{\overline{\chi(n)}}{n^{1/2-it}}+O(|t|^{-1/4}). L ( 2 1 + i t , χ ) = n ≤ q ∣ t ∣/2 π ∑ n 1/2 + i t χ ( n ) + ψ ( 2 1 + i t , χ ) n ≤ q ∣ t ∣/2 π ∑ n 1/2 − i t χ ( n ) + O ( ∣ t ∣ − 1/4 ) .
Therefore,
L ( 1 2 + i t , χ ) ψ ( 1 2 + i t , χ ) − 1 / 2 = ψ ( 1 2 + i t , χ ) − 1 / 2 ∑ n ≤ q ∣ t ∣ / 2 π χ ( n ) n 1 / 2 + i t L\left(\frac{1}{2}+it,\chi\right)\psi\left(\frac{1}{2}+it,\chi\right)^{-1/2}
=\psi\left(\frac{1}{2}+it,\chi\right)^{-1/2}\sum_{n\le\sqrt{q|t|/2\pi}}\frac{\chi(n)}{n^{1/2+it}} L ( 2 1 + i t , χ ) ψ ( 2 1 + i t , χ ) − 1/2 = ψ ( 2 1 + i t , χ ) − 1/2 n ≤ q ∣ t ∣/2 π ∑ n 1/2 + i t χ ( n )
+ ψ ( 1 2 + i t , χ ) 1 / 2 ∑ n ≤ q ∣ t ∣ / 2 π χ ( n ) ‾ n 1 / 2 − i t + O ( ∣ t ∣ − 1 / 4 ) . \qquad+\psi\left(\frac{1}{2}+it,\chi\right)^{1/2}\sum_{n\le\sqrt{q|t|/2\pi}}\frac{\overline{\chi(n)}}{n^{1/2-it}}+O(|t|^{-1/4}). + ψ ( 2 1 + i t , χ ) 1/2 n ≤ q ∣ t ∣/2 π ∑ n 1/2 − i t χ ( n ) + O ( ∣ t ∣ − 1/4 ) .
Using (2.1) for the ratio of Gamma factors in ψ ( 1 2 + i t , χ ) \psi\left(\frac{1}{2}+it,\chi\right) ψ ( 2 1 + i t , χ ) , the above equation completes the proof. □
Let us consider
I 1 = I 1 ( χ , R , T ) ≔ ∫ R ∣ Z ( t , χ ) ∣ 2 Φ ( t / T ) d t , (27) I_1=I_1(\chi,R,T)\coloneqq\int_{\mathbb{R}}|Z(t,\chi)|^2\Phi(t/T)\,\mathrm{d}t, \tag*{(27)} I 1 = I 1 ( χ , R , T ) : = ∫ R ∣ Z ( t , χ ) ∣ 2 Φ ( t / T ) d t , (27)
and
I 2 = I 2 ( χ , R , T ) ≔ ∫ R L ( 1 / 2 + i t , χ ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t . (28) I_2=I_2(\chi,R,T)\coloneqq\int_{\mathbb{R}}L(1/2+it,\chi)|R(t)|^2\Phi(t/T)\,\mathrm{d}t. \tag*{(28)} I 2 = I 2 ( χ , R , T ) : = ∫ R L ( 1/2 + i t , χ ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t . (28)
We now establish the following estimates for I 1 I_1 I 1 and I 2 I_2 I 2 .
Lemma 3.2. For sufficiently large T T T and N = ⌊ T 1 / 4 ⌋ N=\lfloor T^{1/4}\rfloor N = ⌊ T 1/4 ⌋ , we have
I 1 ≪ T ∑ n ≤ N ∣ r ( n ) ∣ 2 . I_1\ll T\sum_{n\leq N}|r(n)|^2. I 1 ≪ T n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Proof. The expression for M 1 M_1 M 1 following [1 Eq. (14)] gives
T − 1 / 4 ∫ R ∣ R ( t ) ∣ 2 Φ ( t / T ) d t = T 3 / 4 Φ ^ ( 0 ) ( 1 + O ( 1 / T ) ) ∑ n ≤ N ∣ r ( n ) ∣ 2 ≪ T 3 / 4 ∑ n ≤ N ∣ r ( n ) ∣ 2 . \begin{aligned}
T^{-1/4}\int_{\mathbb{R}}|R(t)|^2\Phi(t/T)\,\mathrm{d}t=T^{3/4}\widehat{\Phi}(0)(1+O(1/T))\sum_{n\leq N}|r(n)|^2 \\
&\ll T^{3/4}\sum_{n\leq N}|r(n)|^2.
\end{aligned} T − 1/4 ∫ R ∣ R ( t ) ∣ 2 Φ ( t / T ) d t = T 3/4 Φ ( 0 ) ( 1 + O ( 1/ T )) n ≤ N ∑ ∣ r ( n ) ∣ 2 ≪ T 3/4 n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Combining this with Lemma 3.1, one sees that it is enough to prove the desired upper bound for
∫ R c χ ∑ k ≤ q t / 2 π χ ( k ) k exp ( i t ( 1 2 log ( q t 2 π e ) − log k ) ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≪ ∑ n , m ≤ N ∣ r ( n ) r ( m ) ∣ ∑ k ≤ q T / π 1 k ∣ ∫ max ( T , 2 π k 2 / q ) 2 T exp ( i t ( 1 2 log ( q t 2 π e ) ) ) ( m n k ) i t Φ ( t / T ) d t ∣ . (29) \begin{aligned}
\int_{\mathbb{R}}c_{\chi}\sum_{k\leq\sqrt{qt/2\pi}}\frac{\chi(k)}{\sqrt{k}}\exp\left(it\left(\frac{1}{2}\log\left(\frac{qt}{2\pi e}\right)-\log k\right)\right)|R(t)|^2\Phi(t/T)\,\mathrm{d}t \\
&\ll\sum_{n,m\leq N}|r(n)r(m)|\sum_{k\leq\sqrt{qT/\pi}}\frac{1}{\sqrt{k}}\left|\int_{\max(T,2\pi k^2/q)}^{2T}\exp\left(it\left(\frac{1}{2}\log\left(\frac{qt}{2\pi e}\right)\right)\right)\left(\frac{m}{nk}\right)^{it}\Phi(t/T)\,\mathrm{d}t\right|. \tag*{(29)}
\end{aligned} ∫ R c χ k ≤ q t /2 π ∑ k χ ( k ) exp ( i t ( 2 1 log ( 2 π e q t ) − log k ) ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≪ n , m ≤ N ∑ ∣ r ( n ) r ( m ) ∣ k ≤ q T / π ∑ k 1 ∫ m a x ( T , 2 π k 2 / q ) 2 T exp ( i t ( 2 1 log ( 2 π e q t ) ) ) ( nk m ) i t Φ ( t / T ) d t . (29)
The same second-derivative estimate below holds uniformly on every subinterval of [ T , 2 T ] [T,2T] [ T , 2 T ] , including [ a k , 2 T ] [a_k,2T] [ a k , 2 T ] , where a k = max ( T , 2 π k 2 / q ) a_k=\max(T,2\pi k^2/q) a k = max ( T , 2 π k 2 / q ) . Assume that, for T ≤ y ≤ 2 T T\leq y\leq2T T ≤ y ≤ 2 T ,
I ( y ) = ∫ T y exp ( i t ( 1 2 log ( q t 2 π e ) ) ) ( m n k ) i t d t . \mathcal{I}(y)=\int_T^y\exp\left(it\left(\frac{1}{2}\log\left(\frac{qt}{2\pi e}\right)\right)\right)\left(\frac{m}{nk}\right)^{it}\,\mathrm{d}t. I ( y ) = ∫ T y exp ( i t ( 2 1 log ( 2 π e q t ) ) ) ( nk m ) i t d t .
Now, observing that
d 2 d t 2 ( t ( 1 2 log ( q t 2 π e ) + log ( m n k ) ) ) ≫ 1 t ≫ 1 T , \frac{d^2}{dt^2}\left(t\left(\frac{1}{2}\log\left(\frac{qt}{2\pi e}\right)+\log\left(\frac{m}{nk}\right)\right)\right)\gg\frac{1}{t}\gg\frac{1}{T}, d t 2 d 2 ( t ( 2 1 log ( 2 π e q t ) + log ( nk m ) ) ) ≫ t 1 ≫ T 1 ,
for T ≤ t ≤ y T\leq t\leq y T ≤ t ≤ y . So, we apply Lemma 2.3 to obtain I ( y ) ≪ T \mathcal{I}(y)\ll\sqrt{T} I ( y ) ≪ T . Applying integration by parts gives the following estimate, and replacing the lower endpoint T T T by a k a_k a k gives the same bound uniformly in k k k .
∫ T 2 T exp ( i t ( 1 2 log ( q t 2 π e ) ) ) ( m n k ) i t Φ ( t / T ) d t = [ Φ ( t / T ) I ( t ) ] T 2 T − 1 T ∫ T 2 T I ( t ) Φ ′ ( t / T ) d t ≪ T + T ∫ 1 2 ∣ Φ ′ ( x ) ∣ d x ≪ T , \begin{aligned}
\int_T^{2T}\exp\left(it\left(\frac{1}{2}\log\left(\frac{qt}{2\pi e}\right)\right)\right)\left(\frac{m}{nk}\right)^{it}\Phi(t/T)\,\mathrm{d}t \\
&=\left[\Phi(t/T)\mathcal{I}(t)\right]_T^{2T}-\frac{1}{T}\int_T^{2T}\mathcal{I}(t)\Phi'(t/T)\,\mathrm{d}t \\
&\ll\sqrt{T}+\sqrt{T}\int_1^2|\Phi'(x)|\,\mathrm{d}x\ll\sqrt{T},
\end{aligned} ∫ T 2 T exp ( i t ( 2 1 log ( 2 π e q t ) ) ) ( nk m ) i t Φ ( t / T ) d t = [ Φ ( t / T ) I ( t ) ] T 2 T − T 1 ∫ T 2 T I ( t ) Φ ′ ( t / T ) d t ≪ T + T ∫ 1 2 ∣ Φ ′ ( x ) ∣ d x ≪ T ,
where we have used the fact that Φ \Phi Φ is a smooth function compactly supported on [ 1 , 2 ] [1,2] [ 1 , 2 ] , and hence ∫ 1 2 ∣ Φ ′ ( x ) ∣ d x < ∞ \int_1^2 |\Phi'(x)|\,dx < \infty ∫ 1 2 ∣ Φ ′ ( x ) ∣ d x < ∞ . Therefore, from (29) , we have
I 1 ≪ T ∑ n , m ≤ N ∣ r ( n ) r ( m ) ∣ ∑ k ≤ q T / π 1 k ≪ N T 3 / 4 ∑ n ≤ N ∣ r ( n ) ∣ 2 ≪ T ∑ n ≤ N ∣ r ( n ) ∣ 2 , \begin{aligned}
I_1 \ll\sqrt{T}\sum_{n,m\le N}|r(n)r(m)|\sum_{k\le\sqrt{qT/\pi}}\frac{1}{\sqrt{k}} \\
&\ll NT^{3/4}\sum_{n\le N}|r(n)|^2 \ll T\sum_{n\le N}|r(n)|^2,
\end{aligned} I 1 ≪ T n , m ≤ N ∑ ∣ r ( n ) r ( m ) ∣ k ≤ q T / π ∑ k 1 ≪ N T 3/4 n ≤ N ∑ ∣ r ( n ) ∣ 2 ≪ T n ≤ N ∑ ∣ r ( n ) ∣ 2 ,
by using N = ⌊ T 1 / 4 ⌋ N=\lfloor T^{1/4}\rfloor N = ⌊ T 1/4 ⌋ . This completes the proof. □
Lemma 3.3. For sufficiently large T T T and N = ⌊ T α ⌋ N=\lfloor T^\alpha\rfloor N = ⌊ T α ⌋ , we have
I 2 ≫ T exp ( ( 1 − η + o ( 1 ) ) log N log log N ) ∑ n ≤ N ∣ r ( n ) ∣ 2 . I_2 \gg T\exp\left(\left(\sqrt{1-\eta}+o(1)\right)\sqrt{\frac{\log N}{\log\log N}}\right)\sum_{n\le N}|r(n)|^2. I 2 ≫ T exp ( ( 1 − η + o ( 1 ) ) log log N log N ) n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Proof. Choose X = T 1 + ϵ X=T^{1+\epsilon} X = T 1 + ϵ for some positive ϵ \epsilon ϵ . Using (14) of [1 ] , we have
I 2 : = ∫ R L ( 1 / 2 + i t , χ ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t I_2:=\int_{\mathbb{R}}L(1/2+it,\chi)|R(t)|^2\Phi(t/T)\,dt I 2 := ∫ R L ( 1/2 + i t , χ ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t
I 2 = T ∑ n , m ≤ N ∑ k ≤ T 1 + 2 ϵ χ ( k ) r ( n ) r ( m ) ‾ k e − k / X Φ ^ ( T log ( n k / m ) ) + O ( ( ∫ R ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ) ) . (30) I_2=T\sum_{n,m\le N}\sum_{k\le T^{1+2\epsilon}}\frac{\chi(k)r(n)\overline{r(m)}}{\sqrt{k}}e^{-k/X}\widehat{\Phi}(T\log(nk/m))+O\left(\left(\int_{\mathbb{R}}|R(t)|^2\Phi(t/T)\,dt\right)\right). \tag*{(30)} I 2 = T n , m ≤ N ∑ k ≤ T 1 + 2 ϵ ∑ k χ ( k ) r ( n ) r ( m ) e − k / X Φ ( T log ( nk / m )) + O ( ( ∫ R ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ) ) . (30)
The error term above can be shown to be (see (2) in [33 ] )
≪ T ∑ n ≤ N ∣ r ( n ) ∣ 2 . \ll T\sum_{n\le N}|r(n)|^2. ≪ T n ≤ N ∑ ∣ r ( n ) ∣ 2 .
For the contribution from off-diagonal terms n k ≠ m nk\ne m nk = m , we use ∣ log ( n k / m ) ∣ ≫ 1 N |\log(nk/m)|\gg\frac{1}{N} ∣ log ( nk / m ) ∣ ≫ N 1 to see that
Φ ^ ( T log ( n k m ) ) ≪ B ( T N ) − B . \widehat{\Phi}\left(T\log\left(\frac{nk}{m}\right)\right)\ll_B\left(\frac{T}{N}\right)^{-B}. Φ ( T log ( m nk ) ) ≪ B ( N T ) − B .
Since N ≤ T α N\le T^\alpha N ≤ T α , ( α < 1 ) (\alpha<1) ( α < 1 ) , we obtain that the off-diagonal contribution to (30) is
≪ N ∑ n ≤ N ∣ r ( n ) ∣ 2 ∑ k ≤ T 1 + 2 ϵ ∣ χ ( k ) ∣ k T − A ≪ T − A ∑ n ≤ N ∣ r ( n ) ∣ 2 , \ll N\sum_{n\le N}|r(n)|^2\sum_{k\le T^{1+2\epsilon}}\frac{|\chi(k)|}{\sqrt{k}}T^{-A}\ll T^{-A}\sum_{n\le N}|r(n)|^2, ≪ N n ≤ N ∑ ∣ r ( n ) ∣ 2 k ≤ T 1 + 2 ϵ ∑ k ∣ χ ( k ) ∣ T − A ≪ T − A n ≤ N ∑ ∣ r ( n ) ∣ 2 ,
for any A > 1 A>1 A > 1 . Hence,
I 2 = T Φ ^ ( 0 ) ∑ n k = m ≤ N χ ( k ) r ( n ) r ( m ) ‾ k e − k / X + O ( T ∑ n ≤ N ∣ r ( n ) ∣ 2 ) . I_2=T\widehat{\Phi}(0)\sum_{nk=m\le N}\frac{\chi(k)r(n)\overline{r(m)}}{\sqrt{k}}e^{-k/X}+O\left(T\sum_{n\le N}|r(n)|^2\right). I 2 = T Φ ( 0 ) nk = m ≤ N ∑ k χ ( k ) r ( n ) r ( m ) e − k / X + O ( T n ≤ N ∑ ∣ r ( n ) ∣ 2 ) .
As e − k / X ≥ 1 / 2 e^{-k/X}\ge1/2 e − k / X ≥ 1/2 for k < T 1 − ϵ k<T^{1-\epsilon} k < T 1 − ϵ and r ( n ) r ( n k ) ‾ = ∣ r ( n ) ∣ 2 r ( k ) ‾ 1 ( n , k ) = 1 = ∣ χ ( n ) ∣ 2 χ ( k ) ‾ f ( n ) 2 f ( k ) 1 ( n , k ) = 1 r(n)\overline{r(nk)}=|r(n)|^2\overline{r(k)}\mathbf{1}_{(n,k)=1}=|\chi(n)|^2\overline{\chi(k)}f(n)^2f(k)\mathbf{1}_{(n,k)=1} r ( n ) r ( nk ) = ∣ r ( n ) ∣ 2 r ( k ) 1 ( n , k ) = 1 = ∣ χ ( n ) ∣ 2 χ ( k ) f ( n ) 2 f ( k ) 1 ( n , k ) = 1 , applying Lemma 2.10, we have
∑ n k = m ≤ N χ ( k ) r ( n ) r ( m ) ‾ k e − k / X ≫ exp ( ( ( 1 − η + o ( 1 ) ) + o ( 1 ) ) log N log log N ) ∑ n ≤ N ∣ r ( n ) ∣ 2 . \sum_{nk=m\le N}\frac{\chi(k)r(n)\overline{r(m)}}{\sqrt{k}}e^{-k/X}\gg\exp\left(\left(\left(\sqrt{1-\eta}+o(1)\right)+o(1)\right)\sqrt{\frac{\log N}{\log\log N}}\right)\sum_{n\le N}|r(n)|^2. nk = m ≤ N ∑ k χ ( k ) r ( n ) r ( m ) e − k / X ≫ exp ( ( ( 1 − η + o ( 1 ) ) + o ( 1 ) ) log log N log N ) n ≤ N ∑ ∣ r ( n ) ∣ 2 .
This completes the proof. □
Proof of Theorem 1.2. Set N = ⌊ T 1 / 4 ⌋ N=\lfloor T^{1/4}\rfloor N = ⌊ T 1/4 ⌋ and given ϵ > 0 \epsilon>0 ϵ > 0 , choose η > 0 \eta>0 η > 0 sufficiently small that ( 1 − η ) / 2 > 1 / 2 − ϵ \sqrt{(1-\eta)/2}>1/2-\epsilon ( 1 − η ) /2 > 1/2 − ϵ . Then
log N log log N = ( 1 2 + o ( 1 ) ) log T log log T . \sqrt{\frac{\log N}{\log\log N}}=\left(\frac{1}{2}+o(1)\right)\sqrt{\frac{\log T}{\log\log T}}. log log N log N = ( 2 1 + o ( 1 ) ) log log T log T .
Using Lemma 3.2, along with the estimate [33 Eq. (2)]
∫ R ∣ R ( t ) ∣ 2 Φ ( t / T ) d t = T Φ ^ ( 0 ) ( 1 + O ( 1 / T ) ) ∑ n ≤ N ∣ r ( n ) ∣ 2 (31) \int_{\mathbb{R}} |R(t)|^2\Phi(t/T)\,dt=T\widehat{\Phi}(0)(1+O(1/T))\sum_{n\le N}|r(n)|^2 \tag*{(31)} ∫ R ∣ R ( t ) ∣ 2 Φ ( t / T ) d t = T Φ ( 0 ) ( 1 + O ( 1/ T )) n ≤ N ∑ ∣ r ( n ) ∣ 2 (31)
we have
I 1 A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ≪ T ∑ n ≤ N ∣ r ( n ) ∣ 2 T ∑ n ≤ N ∣ r ( n ) ∣ 2 ( 1 + O ( 1 / T ) ) A ( T ) ≪ 1 A ( T ) → 0 as T → ∞ , \begin{aligned}
\frac{I_1}{A(T)\int_T^{2T}|R(t)|^2\Phi(t/T)\,dt}
\ll\frac{T\sum_{n\le N}|r(n)|^2}{T\sum_{n\le N}|r(n)|^2(1+O(1/T))A(T)} \\
&\ll\frac{1}{A(T)}\to0\quad\text{as }T\to\infty,
\end{aligned} A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t I 1 ≪ T ∑ n ≤ N ∣ r ( n ) ∣ 2 ( 1 + O ( 1/ T )) A ( T ) T ∑ n ≤ N ∣ r ( n ) ∣ 2 ≪ A ( T ) 1 → 0 as T → ∞ ,
where
A ( T ) = exp ( ( 1 2 − ϵ ) log T log log T ) . A(T)=\exp\left(\left(\frac{1}{2}-\epsilon\right)\sqrt{\frac{\log T}{\log\log T}}\right). A ( T ) = exp ( ( 2 1 − ϵ ) log log T log T ) .
Hence,
I 1 = o ( A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ) . (32) I_1=o\left(A(T)\int_T^{2T}|R(t)|^2\Phi(t/T)\,dt\right). \tag*{(32)} I 1 = o ( A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ) . (32)
Therefore, from Lemma 3.3 and equation (32) , we see that hypotheses of Proposition 2.11 are satisfied. Hence, the proof is complete. □
Proof of Theorem 1.3
We have
Z F ( t ) : = F ( 1 / 2 + i t ) H − 1 / 2 ( 1 / 2 + i t ) . Z_F(t):=F(1/2+it)H^{-1/2}(1/2+it). Z F ( t ) := F ( 1/2 + i t ) H − 1/2 ( 1/2 + i t ) .
Let us consider
I F = I F ( F , R , T ) : = ∫ R F ( 1 / 2 + i t ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t . I_F=I_F(F,R,T):=\int_{\mathbb{R}}F(1/2+it)|R(t)|^2\Phi(t/T)\,dt. I F = I F ( F , R , T ) := ∫ R F ( 1/2 + i t ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t .
and
J F : = ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t . (33) J_F:=\int_T^{2T}Z_F(t)|R(t)|^2\,dt. \tag*{(33)} J F := ∫ T 2 T Z F ( t ) ∣ R ( t ) ∣ 2 d t . (33)
We now prove the following estimates for I F I_F I F and J F J_F J F .
Lemma 4.1. Let F ( s ) ∈ S p o l y F(s)\in\mathcal{S}_{\mathrm{poly}} F ( s ) ∈ S poly be a function of degree 2 satisfying (1.10). Then for sufficiently large T T T and N = ⌊ T α ⌋ N=\lfloor T^\alpha\rfloor N = ⌊ T α ⌋ , we have
I F ≫ T exp ( ( κ ( 1 − η ) + o ( 1 ) ) log N log log N ) ∑ n ≤ N ∣ r ( n ) ∣ 2 . I_F\gg T\exp\left(\left(\sqrt{\kappa(1-\eta)}+o(1)\right)\sqrt{\frac{\log N}{\log\log N}}\right)\sum_{n\le N}|r(n)|^2. I F ≫ T exp ( ( κ ( 1 − η ) + o ( 1 ) ) log log N log N ) n ≤ N ∑ ∣ r ( n ) ∣ 2 . Proof. The proof of this lemma follows the same lines as that of Lemma 3.3, replacing r ( n ) = χ ( n ) f ( n ) r(n)=\chi(n)f(n) r ( n ) = χ ( n ) f ( n ) and X = T 1 + ϵ X=T^{1+\epsilon} X = T 1 + ϵ by r ( n ) = a F ( n ) f ( n ) r(n)=a_F(n)f(n) r ( n ) = a F ( n ) f ( n ) and X = T 2 + ϵ X=T^{2+\epsilon} X = T 2 + ϵ , respectively. □
Lemma 4.2. Let F ( s ) ∈ S p o l y F(s) \in\mathcal{S}_{\mathrm{poly}} F ( s ) ∈ S poly be a function of degree 2 2 2 satisfying Hypothesis 1 and ξ F = 0 \xi_F=0 ξ F = 0 . Then for sufficiently large T T T , we have
J F = o ( T ∑ n ≤ N ∣ r ( n ) ∣ 2 ) . J_F=o\left(T\sum_{n\leq N}|r(n)|^2\right). J F = o ( T n ≤ N ∑ ∣ r ( n ) ∣ 2 ) . where N = ⌊ T α ⌋ N=\lfloor T^\alpha\rfloor N = ⌊ T α ⌋ with 0 < α < 1 − γ 0<\alpha<1-\gamma 0 < α < 1 − γ .
Proof. Using the definition of R ( t ) R(t) R ( t ) in (2.5), we write
J F = ∑ m , n ≤ N r ( n ) r ( m ) ‾ K m , n (say) , J_F=\sum_{m,n\leq N}r(n)\overline{r(m)}K_{m,n}\qquad\text{(say)}, J F = m , n ≤ N ∑ r ( n ) r ( m ) K m , n (say) ,
where
K m , n : = ∫ T 2 T Z F ( t ) ( m n ) i t d t . K_{m,n}:=\int_T^{2T}Z_F(t)\left(\frac{m}{n}\right)^{it}\,dt. K m , n := ∫ T 2 T Z F ( t ) ( n m ) i t d t .
As Z F ( t ) ∈ R Z_F(t)\in\mathbb{R} Z F ( t ) ∈ R , we get K ‾ m , n = K n , m \overline{K}_{m,n}=K_{n,m} K m , n = K n , m . Therefore, the contribution to J F J_F J F from the non-diagonal terms for which m ≠ n m\ne n m = n is
∑ m > n r ( n ) r ( m ) ‾ K m , n + ∑ n > m r ( n ) r ( m ) ‾ K m , n = ∑ m > n r ( n ) r ( m ) ‾ K m , n + ∑ n > m r ( m ) r ( n ) ‾ K n , m = 2 ℜ ( ∑ m > n r ( n ) r ( m ) ‾ K m , n ) . \begin{aligned}
\sum_{m>n}r(n)\overline{r(m)}K_{m,n}+\sum_{n>m}r(n)\overline{r(m)}K_{m,n} \\
&=\sum_{m>n}r(n)\overline{r(m)}K_{m,n}+\sum_{n>m}r(m)\overline{r(n)}K_{n,m} \\
&=2\Re\left(\sum_{m>n}r(n)\overline{r(m)}K_{m,n}\right).
\end{aligned} m > n ∑ r ( n ) r ( m ) K m , n + n > m ∑ r ( n ) r ( m ) K m , n = m > n ∑ r ( n ) r ( m ) K m , n + n > m ∑ r ( m ) r ( n ) K n , m = 2ℜ ( m > n ∑ r ( n ) r ( m ) K m , n ) .
Hence,
J F = ∑ n ≤ N ∣ r ( n ) ∣ 2 K n , n + 2 ℜ ( ∑ 1 ≤ n < m ≤ N r ( n ) r ( m ) ‾ K m , n ) . J_F=\sum_{n\leq N}|r(n)|^2K_{n,n}+2\Re\left(\sum_{1\leq n<m\leq N}r(n)\overline{r(m)}K_{m,n}\right). J F = n ≤ N ∑ ∣ r ( n ) ∣ 2 K n , n + 2ℜ ( 1 ≤ n < m ≤ N ∑ r ( n ) r ( m ) K m , n ) .
Therefore,
∣ J F ∣ ≤ 2 ∑ 1 ≤ n ≤ m ≤ N ∣ r ( n ) r ( m ) ∣ ∣ K m , n ∣ |J_F|\leq2\sum_{1\leq n\leq m\leq N}|r(n)r(m)||K_{m,n}| ∣ J F ∣ ≤ 2 1 ≤ n ≤ m ≤ N ∑ ∣ r ( n ) r ( m ) ∣∣ K m , n ∣
≤ 2 ∑ 1 ≤ n ≤ m ≤ N ∣ r ( n ) r ( m ) ∣ m / n ∣ ∫ T 2 T F ( 1 2 + i t ) H − 1 2 ( 1 2 + i t ) ( m n ) 1 2 + i t d t ∣ . (34) \leq2\sum_{1\leq n\leq m\leq N}\frac{|r(n)r(m)|}{\sqrt{m/n}}\left|\int_T^{2T}F\left(\frac{1}{2}+it\right)H^{-\frac{1}{2}}\left(\frac{1}{2}+it\right)\left(\frac{m}{n}\right)^{\frac{1}{2}+it}\,dt\right|. \tag*{(34)} ≤ 2 1 ≤ n ≤ m ≤ N ∑ m / n ∣ r ( n ) r ( m ) ∣ ∫ T 2 T F ( 2 1 + i t ) H − 2 1 ( 2 1 + i t ) ( n m ) 2 1 + i t d t . (34)
We now turn to establishing an upper bound for the integral
I m , n = ∫ T 2 T F ( 1 2 + i t ) H − 1 2 ( 1 2 + i t ) ( m n ) 1 2 + i t d t = − i ∫ 1 2 + i T 1 2 + i 2 T F ( s ) H − 1 2 ( s ) ( m n ) s d s . \begin{aligned}
I_{m,n}=\int_T^{2T}F\left(\frac{1}{2}+it\right)H^{-\frac{1}{2}}\left(\frac{1}{2}+it\right)\left(\frac{m}{n}\right)^{\frac{1}{2}+it}\,dt \\
&=-i\int_{\frac{1}{2}+iT}^{\frac{1}{2}+i2T}F(s)H^{-\frac{1}{2}}(s)\left(\frac{m}{n}\right)^s\,ds.
\end{aligned} I m , n = ∫ T 2 T F ( 2 1 + i t ) H − 2 1 ( 2 1 + i t ) ( n m ) 2 1 + i t d t = − i ∫ 2 1 + i T 2 1 + i 2 T F ( s ) H − 2 1 ( s ) ( n m ) s d s .
We apply Cauchy’s theorem to estimate the integral
∫ R F ( s ) H − 1 2 ( s ) ( m n ) s d s , \int_{\mathcal{R}}F(s)H^{-\frac{1}{2}}(s)\left(\frac{m}{n}\right)^s\,ds, ∫ R F ( s ) H − 2 1 ( s ) ( n m ) s d s ,
where R \mathcal{R} R denotes the rectangle with vertices at 1 2 + i T \frac{1}{2}+iT 2 1 + i T , 1 + δ + i T 1+\delta+iT 1 + δ + i T , 1 + δ + i 2 T 1+\delta+i2T 1 + δ + i 2 T , and 1 2 + i 2 T \frac{1}{2}+i2T 2 1 + i 2 T . Here δ > 0 \delta>0 δ > 0 is a small constant to be specified later. By Cauchy’s theorem, we have
I m , n = − i ( J 1 + J 2 + J 3 ) , (35) I_{m,n}=-i(J_1+J_2+J_3), \tag*{(35)} I m , n = − i ( J 1 + J 2 + J 3 ) , (35)
where we have dropped the m , n m,n m , n suffix for each J i J_i J i for ease of notation, and
J 1 = ∫ 1 2 1 + δ F ( σ + i 2 T ) H − 1 2 ( σ + i 2 T ) ( m n ) σ + i 2 T d σ , J_1=\int_{\frac{1}{2}}^{1+\delta}F(\sigma+i2T)H^{-\frac{1}{2}}(\sigma+i2T)\left(\frac{m}{n}\right)^{\sigma+i2T}\,d\sigma, J 1 = ∫ 2 1 1 + δ F ( σ + i 2 T ) H − 2 1 ( σ + i 2 T ) ( n m ) σ + i 2 T d σ ,
J 2 = i ∫ T 2 T F ( 1 + δ + i t ) H − 1 2 ( 1 + δ + i t ) ( m n ) 1 + δ + i t d t , J_2=i\int_T^{2T}F(1+\delta+it)H^{-\frac{1}{2}}(1+\delta+it)\left(\frac{m}{n}\right)^{1+\delta+it}\,dt, J 2 = i ∫ T 2 T F ( 1 + δ + i t ) H − 2 1 ( 1 + δ + i t ) ( n m ) 1 + δ + i t d t ,
J 3 = ∫ 1 2 1 + δ F ( σ + i T ) H − 1 2 ( σ + i T ) ( m n ) σ + i T d σ . J_3=\int_{\frac{1}{2}}^{1+\delta}F(\sigma+iT)H^{-\frac{1}{2}}(\sigma+iT)\left(\frac{m}{n}\right)^{\sigma+iT}\,d\sigma. J 3 = ∫ 2 1 1 + δ F ( σ + i T ) H − 2 1 ( σ + i T ) ( n m ) σ + i T d σ .
We now first estimate J 3 J_3 J 3 . By using the estimate given in (15) , we have
J 3 ≪ T 1 2 + δ × { 1 if m = n , ( m n ) 1 + δ if m > n ≪ T 1 2 + δ ( m n ) 1 + δ . (36) J_3\ll T^{\frac{1}{2}+\delta}\times
\begin{cases}
1 & \text{if }m=n,\\
\left(\frac{m}{n}\right)^{1+\delta} & \text{if }m>n
\end{cases}
\ll T^{\frac{1}{2}+\delta}\left(\frac{m}{n}\right)^{1+\delta}. \tag*{(36)} J 3 ≪ T 2 1 + δ × { 1 ( n m ) 1 + δ if m = n , if m > n ≪ T 2 1 + δ ( n m ) 1 + δ . (36)
The bound for J 1 J_1 J 1 follows similarly to that for J 3 J_3 J 3 , giving the same estimate as in (36) . We now estimate J 2 J_2 J 2 . Because of the asymptotic formula (14) , J 2 J_2 J 2 is a constant multiple of
( m n ) 1 + δ ∫ T 2 T ∑ k = 1 ∞ a F ( k ) k 1 + δ + i t ( Q 0 t ) 1 2 + δ + i t e − i t ( m n ) i t ( 1 + O ( 1 / t ) ) d t . (37) \left(\frac{m}{n}\right)^{1+\delta}\int_T^{2T}\sum_{k=1}^{\infty}\frac{a_F(k)}{k^{1+\delta+it}}(Q_0t)^{\frac{1}{2}+\delta+it}e^{-it}\left(\frac{m}{n}\right)^{it}(1+O(1/t))\,dt. \tag*{(37)} ( n m ) 1 + δ ∫ T 2 T k = 1 ∑ ∞ k 1 + δ + i t a F ( k ) ( Q 0 t ) 2 1 + δ + i t e − i t ( n m ) i t ( 1 + O ( 1/ t )) d t . (37)
The error term in (37) gives
( m n ) 1 + δ ∫ T 2 T O ( t δ − 1 2 ) d t = O ( ( m n ) 1 + δ T 1 2 + δ ) . (38) \left(\frac{m}{n}\right)^{1+\delta}\int_T^{2T}O(t^{\delta-\frac{1}{2}})\,dt
=O\left(\left(\frac{m}{n}\right)^{1+\delta}T^{\frac{1}{2}+\delta}\right). \tag*{(38)} ( n m ) 1 + δ ∫ T 2 T O ( t δ − 2 1 ) d t = O ( ( n m ) 1 + δ T 2 1 + δ ) . (38)
Thus we need to estimate
( m n ) 1 + δ ∑ k = 1 ∞ a F ( k ) k 1 + δ ∫ T 2 T g ( t ) e i f ( t ) d t . (39) \left(\frac{m}{n}\right)^{1+\delta}\sum_{k=1}^{\infty}\frac{a_F(k)}{k^{1+\delta}}\int_T^{2T}g(t)e^{if(t)}\,dt. \tag*{(39)} ( n m ) 1 + δ k = 1 ∑ ∞ k 1 + δ a F ( k ) ∫ T 2 T g ( t ) e i f ( t ) d t . (39)
where g ( t ) = t 1 2 + δ g(t)=t^{\frac{1}{2}+\delta} g ( t ) = t 2 1 + δ and f ( t ) = t log ( Q 0 t m / n e k ) f(t)=t\log\left(\frac{Q_0tm/n}{ek}\right) f ( t ) = t log ( e k Q 0 t m / n ) . We now partition the sum in (39) according to the following intervals:
1 ≤ k ≤ Q 0 T m n − 1 , Q 0 T m n + 1 ≤ k ≤ 2 Q 0 T m n − 1 , 2 Q 0 T m n + 1 ≤ k , 1\leq k\leq Q_0T\frac{m}{n}-1,\qquad Q_0T\frac{m}{n}+1\leq k\leq2Q_0T\frac{m}{n}-1,\qquad2Q_0T\frac{m}{n}+1\leq k, 1 ≤ k ≤ Q 0 T n m − 1 , Q 0 T n m + 1 ≤ k ≤ 2 Q 0 T n m − 1 , 2 Q 0 T n m + 1 ≤ k ,
with S 1 S_1 S 1 , S 2 S_2 S 2 , and S 3 S_3 S 3 denoting the sums over these intervals, respectively. We will use the Lemma 2.2 to estimate S 1 , S 3 S_1,S_3 S 1 , S 3 and to estimate S 2 S_2 S 2 we will use Lemma 2.4. At most four integers are not covered by the intervals considered above. Any such k k k lies in the range T m n ≪ k ≪ T m n T\frac{m}{n}\ll k\ll T\frac{m}{n} T n m ≪ k ≪ T n m , and satisfies a F ( k ) ≪ k ϵ a_F(k)\ll k^\epsilon a F ( k ) ≪ k ϵ . Therefore, their contribution to (39) is
O ( ( m n ) 1 + δ ∣ a F ( k ) ∣ k 1 + δ T 1 2 + δ T ) = O ( ( m n ) 1 + δ k ϵ k 1 + δ T 1 2 + δ T ) = O ( T 1 2 + ϵ ( m n ) ϵ ) . (40) O\left(\left(\frac{m}{n}\right)^{1+\delta}\frac{|a_F(k)|}{k^{1+\delta}}T^{\frac{1}{2}+\delta}T\right)
=O\left(\left(\frac{m}{n}\right)^{1+\delta}\frac{k^\epsilon}{k^{1+\delta}}T^{\frac{1}{2}+\delta}T\right)
=O\left(T^{\frac{1}{2}+\epsilon}\left(\frac{m}{n}\right)^\epsilon\right). \tag*{(40)} O ( ( n m ) 1 + δ k 1 + δ ∣ a F ( k ) ∣ T 2 1 + δ T ) = O ( ( n m ) 1 + δ k 1 + δ k ϵ T 2 1 + δ T ) = O ( T 2 1 + ϵ ( n m ) ϵ ) . (40)
We next estimate the contribution from S 1 S_1 S 1 using Lemma 2.2. We note that f ′ ( t ) = log ( Q 0 t m / n k ) ≫ log ( Q 0 T m / n k ) f'(t)=\log\left(\frac{Q_0tm/n}{k}\right)\gg\log\left(\frac{Q_0Tm/n}{k}\right) f ′ ( t ) = log ( k Q 0 t m / n ) ≫ log ( k Q 0 T m / n ) and g ( t ) ≍ T 1 2 + δ g(t)\asymp T^{\frac{1}{2}+\delta} g ( t ) ≍ T 2 1 + δ . If necessary, we split [ T , 2 T ] [T,2T] [ T , 2 T ] into two subintervals so that the required monotonicity condition for f ′ / g f'/g f ′ / g is satisfied. Hence, Lemma 2.2 gives
S 1 ≪ ( m n ) 1 + δ ∑ 1 ≤ k ≤ T Q 0 m / n − 1 ∣ a F ( k ) ∣ k 1 + δ T 1 2 + δ log ( T Q 0 m / n k ) (41) S_1 \ll\left(\frac{m}{n}\right)^{1+\delta}\sum_{1\le k\le TQ_0^{m/n}-1}\frac{|a_F(k)|}{k^{1+\delta}}\frac{T^{\frac{1}{2}+\delta}}{\log\left(\frac{TQ_0^{m/n}}{k}\right)} \tag*{(41)} S 1 ≪ ( n m ) 1 + δ 1 ≤ k ≤ T Q 0 m / n − 1 ∑ k 1 + δ ∣ a F ( k ) ∣ log ( k T Q 0 m / n ) T 2 1 + δ (41)
We split the above sum once again into two parts:
1 ≤ k ≤ Q 0 T m / n 2 and Q 0 T m / n 2 < k ≤ Q 0 T m / n − 1 , 1\le k\le\frac{Q_0T^{m/n}}{2}\qquad\text{and}\qquad\frac{Q_0T^{m/n}}{2}<k\le Q_0T^{m/n}-1, 1 ≤ k ≤ 2 Q 0 T m / n and 2 Q 0 T m / n < k ≤ Q 0 T m / n − 1 ,
and denote them by S 11 S_{11} S 11 and S 12 S_{12} S 12 , respectively. We first consider S 11 S_{11} S 11 . For 1 ≤ k ≤ Q 0 T m / n 2 1\le k\le\frac{Q_0T^{m/n}}{2} 1 ≤ k ≤ 2 Q 0 T m / n , we have log ( Q 0 T m / n k ) ≥ log 2 \log\left(\frac{Q_0T^{m/n}}{k}\right)\ge\log2 log ( k Q 0 T m / n ) ≥ log 2 . Hence, we obtain
S 11 ≪ ( m n ) 1 + δ T 1 2 + δ ∑ 1 ≤ k ≤ Q 0 T m / n 2 ∣ a F ( k ) ∣ k 1 + δ ≪ ( m n ) 1 + δ T 1 2 + δ , (42) S_{11}\ll\left(\frac{m}{n}\right)^{1+\delta}T^{\frac{1}{2}+\delta}\sum_{1\le k\le\frac{Q_0T^{m/n}}{2}}\frac{|a_F(k)|}{k^{1+\delta}}\ll\left(\frac{m}{n}\right)^{1+\delta}T^{\frac{1}{2}+\delta}, \tag*{(42)} S 11 ≪ ( n m ) 1 + δ T 2 1 + δ 1 ≤ k ≤ 2 Q 0 T m / n ∑ k 1 + δ ∣ a F ( k ) ∣ ≪ ( n m ) 1 + δ T 2 1 + δ , (42)
as ∑ k = 1 ∞ a F ( k ) k − s \sum_{k=1}^{\infty}a_F(k)k^{-s} ∑ k = 1 ∞ a F ( k ) k − s converges absolutely for ℜ ( s ) > 1 \Re(s)>1 ℜ ( s ) > 1 .
Now, we estimate S 12 S_{12} S 12 , using the inequality log Q 0 T m / n k ≥ Q 0 T m / n − k Q 0 T m / n \log\frac{Q_0T^{m/n}}{k}\ge\frac{Q_0T^{m/n}-k}{Q_0T^{m/n}} log k Q 0 T m / n ≥ Q 0 T m / n Q 0 T m / n − k for Q 0 T m / n 2 < k ≤ Q 0 T m / n − 1 \frac{Q_0T^{m/n}}{2}<k\le Q_0T^{m/n}-1 2 Q 0 T m / n < k ≤ Q 0 T m / n − 1 and a F ( k ) = O ( k ϵ ) a_F(k)=O(k^\epsilon) a F ( k ) = O ( k ϵ ) , we have
S 12 ≪ ( m n ) 1 + δ T 1 2 + δ ∑ Q 0 T m / n 2 < k ≤ Q 0 T m / n − 1 ∣ a F ( k ) ∣ k 1 + δ Q 0 T m / n Q 0 T m / n − k ≪ ( m n ) 1 + ϵ T 1 2 + ϵ ∑ Q 0 T m / n 2 < k ≤ Q 0 T m / n − 1 1 Q 0 T m / n − k . (43) \begin{aligned}
S_{12}&\ll\left(\frac{m}{n}\right)^{1+\delta}T^{\frac{1}{2}+\delta}\sum_{\frac{Q_0T^{m/n}}{2}<k\le Q_0T^{m/n}-1}\frac{|a_F(k)|}{k^{1+\delta}}\frac{Q_0T^{m/n}}{Q_0T^{m/n}-k}\\
&\ll\left(\frac{m}{n}\right)^{1+\epsilon}T^{\frac{1}{2}+\epsilon}\sum_{\frac{Q_0T^{m/n}}{2}<k\le Q_0T^{m/n}-1}\frac{1}{Q_0T^{m/n}-k}. \tag*{(43)}
\end{aligned} S 12 ≪ ( n m ) 1 + δ T 2 1 + δ 2 Q 0 T m / n < k ≤ Q 0 T m / n − 1 ∑ k 1 + δ ∣ a F ( k ) ∣ Q 0 T m / n − k Q 0 T m / n ≪ ( n m ) 1 + ϵ T 2 1 + ϵ 2 Q 0 T m / n < k ≤ Q 0 T m / n − 1 ∑ Q 0 T m / n − k 1 . (43)
Estimating the sum in (43) , we observe that
∑ Q 0 T m / n 2 < k ≤ Q 0 T m / n − 1 1 Q 0 T m / n − k = ∑ 1 ≤ l < Q 0 T m / n 2 1 l + g ≪ log ( T m / n ) , \sum_{\frac{Q_0T^{m/n}}{2}<k\le Q_0T^{m/n}-1}\frac{1}{Q_0T^{m/n}-k}
=\sum_{1\le l<\frac{Q_0T^{m/n}}{2}}\frac{1}{l+g}\ll\log(Tm/n), 2 Q 0 T m / n < k ≤ Q 0 T m / n − 1 ∑ Q 0 T m / n − k 1 = 1 ≤ l < 2 Q 0 T m / n ∑ l + g 1 ≪ log ( T m / n ) ,
where g g g is the fractional part of Q 0 T m / n Q_0T^{m/n} Q 0 T m / n . Using this, from (43) , we have
S 12 ≪ ( m n ) 1 + ϵ T 1 2 + ϵ log ( T m / n ) . (44) S_{12}\ll\left(\frac{m}{n}\right)^{1+\epsilon}T^{\frac{1}{2}+\epsilon}\log(Tm/n). \tag*{(44)} S 12 ≪ ( n m ) 1 + ϵ T 2 1 + ϵ log ( T m / n ) . (44)
Therefore, from (42) and (44) , we obtain that
S 1 ≪ ( m n ) 1 + δ T 1 2 + δ + ( m n ) 1 + ϵ T 1 2 + ϵ log ( T m / n ) . (45) S_1\ll\left(\frac{m}{n}\right)^{1+\delta}T^{\frac{1}{2}+\delta}+\left(\frac{m}{n}\right)^{1+\epsilon}T^{\frac{1}{2}+\epsilon}\log(Tm/n). \tag*{(45)} S 1 ≪ ( n m ) 1 + δ T 2 1 + δ + ( n m ) 1 + ϵ T 2 1 + ϵ log ( T m / n ) . (45)
The term S 3 S_3 S 3 is handled in the same way as S 1 S_1 S 1 . In this case, we apply Lemma 2.2 together with the bound ∣ f ′ ( t ) ∣ ≫ log ( k 2 Q 0 T m / n ) |f'(t)|\gg\log\left(\frac{k}{2Q_0T^{m/n}}\right) ∣ f ′ ( t ) ∣ ≫ log ( 2 Q 0 T m / n k ) . Hence, by the same argument, we obtain
S 3 ≪ ( m n ) 1 + δ T 1 2 + δ + ( m n ) 1 + ϵ T 1 2 + ϵ log ( T m / n ) . (46) S_3\ll\left(\frac{m}{n}\right)^{1+\delta}T^{\frac{1}{2}+\delta}+\left(\frac{m}{n}\right)^{1+\epsilon}T^{\frac{1}{2}+\epsilon}\log(Tm/n). \tag*{(46)} S 3 ≪ ( n m ) 1 + δ T 2 1 + δ + ( n m ) 1 + ϵ T 2 1 + ϵ log ( T m / n ) . (46)
Finally, we estimate the sum S 2 S_2 S 2 by using Lemma 2.4. Here note that c = k / ( Q 0 m / n ) c=k/(Q_0m/n) c = k / ( Q 0 m / n ) , λ 2 = 1 / T \lambda_2=1/T λ 2 = 1/ T , λ 3 = 1 / T 2 \lambda_3=1/T^2 λ 3 = 1/ T 2 , G = T 1 / 2 + δ G=T^{1/2+\delta} G = T 1/2 + δ and G 1 = T − 1 / 2 + δ G_1=T^{-1/2+\delta} G 1 = T − 1/2 + δ . Hence, we have
S 2 = c 1 ∑ T Q 0 m n + 1 ≤ k ≤ 2 T Q 0 m n − 1 a F ( k ) e − i k Q 0 m / n + O ( ( m n ) 1 + δ T 1 2 + δ ∑ T Q 0 m n + 1 ≤ k ≤ 2 T Q 0 m n − 1 ∣ a F ( k ) ∣ k 1 + δ ) + O ( ( m n ) 1 + δ T 1 2 + δ ∑ T Q 0 m n + 1 ≤ k ≤ 2 T Q 0 m n − 1 ∣ a F ( k ) ∣ k 1 + δ ( 1 log ( k Q 0 T m / n ) + 1 log ( 2 Q 0 T m / n k ) ) ) . \begin{aligned}
S_2={}&c_1\sum_{TQ_0\frac{m}{n}+1\leq k\leq2TQ_0\frac{m}{n}-1}a_F(k)e^{-\frac{ik}{Q_0m/n}}+O\left(\left(\frac{m}{n}\right)^{1+\delta}T^{\frac12+\delta}\sum_{TQ_0\frac{m}{n}+1\leq k\leq2TQ_0\frac{m}{n}-1}\frac{|a_F(k)|}{k^{1+\delta}}\right)\\
&+O\left(\left(\frac{m}{n}\right)^{1+\delta}T^{\frac12+\delta}\sum_{TQ_0\frac{m}{n}+1\leq k\leq2TQ_0\frac{m}{n}-1}\frac{|a_F(k)|}{k^{1+\delta}}\left(\frac{1}{\log\left(\frac{k}{Q_0Tm/n}\right)}+\frac{1}{\log\left(\frac{2Q_0Tm/n}{k}\right)}\right)\right).
\end{aligned} S 2 = c 1 T Q 0 n m + 1 ≤ k ≤ 2 T Q 0 n m − 1 ∑ a F ( k ) e − Q 0 m / n ik + O ( n m ) 1 + δ T 2 1 + δ T Q 0 n m + 1 ≤ k ≤ 2 T Q 0 n m − 1 ∑ k 1 + δ ∣ a F ( k ) ∣ + O ( n m ) 1 + δ T 2 1 + δ T Q 0 n m + 1 ≤ k ≤ 2 T Q 0 n m − 1 ∑ k 1 + δ ∣ a F ( k ) ∣ log ( Q 0 T m / n k ) 1 + log ( k 2 Q 0 T m / n ) 1 .
where c 1 c_1 c 1 is a constant. The absolute convergence of ∑ k = 1 ∞ a F ( k ) k − s \sum_{k=1}^{\infty}a_F(k)k^{-s} ∑ k = 1 ∞ a F ( k ) k − s for R ( s ) > 1 \mathfrak{R}(s)>1 R ( s ) > 1 shows that the first error term in S 2 S_2 S 2 is O ( ( m n ) 1 + δ T 1 / 2 + δ ) O\left(\left(\frac{m}{n}\right)^{1+\delta}T^{1/2+\delta}\right) O ( ( n m ) 1 + δ T 1/2 + δ ) . The second error term can be treated in the same way as S 12 S_{12} S 12 , which gives O ( ( m n ) 1 + ϵ T 1 / 2 + ϵ log ( T m / n ) ) O\left(\left(\frac{m}{n}\right)^{1+\epsilon}T^{1/2+\epsilon}\log(Tm/n)\right) O ( ( n m ) 1 + ϵ T 1/2 + ϵ log ( T m / n ) ) . Thus, it remains only to bound the exponential sum occurring in S 2 S_2 S 2 . We have
S 2 = c 1 ∑ T Q 0 m n + 1 ≤ k ≤ 2 T Q 0 m n − 1 a F ( k ) e − i k Q 0 m / n + O ( ( m n ) 1 + δ T 1 2 + δ ) + O ( ( m n ) 1 + ϵ T 1 2 + ϵ log ( T m / n ) ) . (47) S_2=c_1\sum_{TQ_0\frac{m}{n}+1\leq k\leq2TQ_0\frac{m}{n}-1}a_F(k)e^{-\frac{ik}{Q_0m/n}}+O\left(\left(\frac{m}{n}\right)^{1+\delta}T^{\frac12+\delta}\right)+O\left(\left(\frac{m}{n}\right)^{1+\epsilon}T^{\frac12+\epsilon}\log(Tm/n)\right). \tag*{(47)} S 2 = c 1 T Q 0 n m + 1 ≤ k ≤ 2 T Q 0 n m − 1 ∑ a F ( k ) e − Q 0 m / n ik + O ( ( n m ) 1 + δ T 2 1 + δ ) + O ( ( n m ) 1 + ϵ T 2 1 + ϵ log ( T m / n ) ) . (47)
Hence, using the estimate of S 1 S_1 S 1 , S 2 S_2 S 2 , and S 3 S_3 S 3 , we get
J 2 ≪ ∣ ∑ T Q 0 m n + 1 ≤ k ≤ 2 T Q 0 m n − 1 a F ( k ) e − i k Q 0 m / n ∣ + ( m n ) 1 + δ T 1 2 + δ + ( m n ) 1 + ϵ T 1 2 + ϵ log ( T m / n ) . (48) J_2\ll\left|\sum_{TQ_0\frac{m}{n}+1\leq k\leq2TQ_0\frac{m}{n}-1}a_F(k)e^{-\frac{ik}{Q_0m/n}}\right|+\left(\frac{m}{n}\right)^{1+\delta}T^{\frac12+\delta}+\left(\frac{m}{n}\right)^{1+\epsilon}T^{\frac12+\epsilon}\log(Tm/n). \tag*{(48)} J 2 ≪ T Q 0 n m + 1 ≤ k ≤ 2 T Q 0 n m − 1 ∑ a F ( k ) e − Q 0 m / n ik + ( n m ) 1 + δ T 2 1 + δ + ( n m ) 1 + ϵ T 2 1 + ϵ log ( T m / n ) . (48)
Using the estimate of J 1 J_1 J 1 , J 2 J_2 J 2 and J 3 J_3 J 3 , from (4.3) we have
I m , n ≪ ∣ ∑ T Q 0 m n + 1 ≤ k ≤ 2 T Q 0 m n − 1 a F ( k ) e − i k Q 0 m / n ∣ + ( m n ) 1 + δ T 1 2 + δ + ( m n ) 1 + ϵ T 1 2 + ϵ log ( T m / n ) . (49) I_{m,n}\ll\left|\sum_{TQ_0\frac{m}{n}+1\leq k\leq2TQ_0\frac{m}{n}-1}a_F(k)e^{-\frac{ik}{Q_0m/n}}\right|+\left(\frac{m}{n}\right)^{1+\delta}T^{\frac12+\delta}+\left(\frac{m}{n}\right)^{1+\epsilon}T^{\frac12+\epsilon}\log(Tm/n). \tag*{(49)} I m , n ≪ T Q 0 n m + 1 ≤ k ≤ 2 T Q 0 n m − 1 ∑ a F ( k ) e − Q 0 m / n ik + ( n m ) 1 + δ T 2 1 + δ + ( n m ) 1 + ϵ T 2 1 + ϵ log ( T m / n ) . (49)
Therefore, from (4.2), with Q 0 = q F / ( 2 π ) Q_0=\sqrt{q_F}/(2\pi) Q 0 = q F / ( 2 π ) , we obtain
J F ≪ ∑ 1 ≤ n ≤ N ∣ r ( m ) r ( n ) ∣ m / n ( ∣ ∑ T Q 0 m n + 1 ≤ k ≤ 2 T Q 0 m n − 1 a F ( k ) e − 2 π i k n m q F ∣ + ( m n ) 1 + δ T 1 2 + δ + ( m n ) 1 + ϵ T 1 2 + ϵ log ( T m / n ) ) . (50) J_F\ll\sum_{1\leq n\leq N}\frac{|r(m)r(n)|}{\sqrt{m/n}}\left(\left|\sum_{TQ_0\frac{m}{n}+1\leq k\leq2TQ_0\frac{m}{n}-1}a_F(k)e^{-\frac{2\pi ikn}{m\sqrt{q_F}}}\right|+\left(\frac{m}{n}\right)^{1+\delta}T^{\frac12+\delta}+\left(\frac{m}{n}\right)^{1+\epsilon}T^{\frac12+\epsilon}\log(Tm/n)\right). \tag*{(50)} J F ≪ 1 ≤ n ≤ N ∑ m / n ∣ r ( m ) r ( n ) ∣ T Q 0 n m + 1 ≤ k ≤ 2 T Q 0 n m − 1 ∑ a F ( k ) e − m q F 2 π ik n + ( n m ) 1 + δ T 2 1 + δ + ( n m ) 1 + ϵ T 2 1 + ϵ log ( T m / n ) . (50)
Next, we use Hypothesis 1 to estimate the exponential sum in (4.18). We have
∑ T Q 0 m n + 1 ≤ k ≤ 2 T Q 0 m n − 1 a F ( k ) e − 2 π i k n m q F ≪ ( T m n ) γ ( log ( T m / n ) ) γ 1 . \sum_{TQ_0\frac{m}{n}+1\leq k\leq2TQ_0\frac{m}{n}-1}a_F(k)e^{-\frac{2\pi ikn}{m\sqrt{q_F}}}\ll\left(T\frac{m}{n}\right)^\gamma(\log(Tm/n))^{\gamma_1}. T Q 0 n m + 1 ≤ k ≤ 2 T Q 0 n m − 1 ∑ a F ( k ) e − m q F 2 π ik n ≪ ( T n m ) γ ( log ( T m / n ) ) γ 1 .
Hence, from (4.18), we obtain
J F ≪ ∑ 1 ≤ n ≤ m ≤ N ∣ r ( m ) r ( n ) ∣ ( ( m n ) 1 2 + ϵ T 1 2 + ϵ log ( T m / n ) + ( m n ) 1 2 + δ T 1 2 + δ + T γ ( m n ) γ − 1 2 ( log ( T m / n ) ) γ 1 ) ≪ ∑ 1 ≤ n ≤ m ≤ N ∣ r ( m ) r ( n ) ∣ ( T γ ( m n ) γ − 1 2 ( log ( T m / n ) ) γ 1 + ( m n ) 1 2 + δ T 1 2 + δ ) , (51) \begin{aligned}
J_F \ll\sum_{1\le n\le m\le N} |r(m)r(n)| \left( \left(\frac{m}{n}\right)^{\frac{1}{2}+\epsilon} T^{\frac{1}{2}+\epsilon} \log(Tm/n) + \left(\frac{m}{n}\right)^{\frac{1}{2}+\delta} T^{\frac{1}{2}+\delta} + T^\gamma\left(\frac{m}{n}\right)^{\gamma-\frac{1}{2}} (\log(Tm/n))^{\gamma_1} \right) \\
&\ll\sum_{1\le n\le m\le N} |r(m)r(n)| \left( T^\gamma\left(\frac{m}{n}\right)^{\gamma-\frac{1}{2}} (\log(Tm/n))^{\gamma_1} + \left(\frac{m}{n}\right)^{\frac{1}{2}+\delta} T^{\frac{1}{2}+\delta} \right), \tag*{(51)}
\end{aligned} J F ≪ 1 ≤ n ≤ m ≤ N ∑ ∣ r ( m ) r ( n ) ∣ ( ( n m ) 2 1 + ϵ T 2 1 + ϵ log ( T m / n ) + ( n m ) 2 1 + δ T 2 1 + δ + T γ ( n m ) γ − 2 1 ( log ( T m / n ) ) γ 1 ) ≪ 1 ≤ n ≤ m ≤ N ∑ ∣ r ( m ) r ( n ) ∣ ( T γ ( n m ) γ − 2 1 ( log ( T m / n ) ) γ 1 + ( n m ) 2 1 + δ T 2 1 + δ ) , (51)
where we use the fact that, for 0 < ϵ < δ 0<\epsilon<\delta 0 < ϵ < δ ,
T 1 2 + ϵ ( m n ) 1 2 + ϵ log ( T m / n ) ≪ T 1 2 + δ ( m n ) 1 2 + δ . T^{\frac{1}{2}+\epsilon}\left(\frac{m}{n}\right)^{\frac{1}{2}+\epsilon}\log(Tm/n) \ll T^{\frac{1}{2}+\delta}\left(\frac{m}{n}\right)^{\frac{1}{2}+\delta}. T 2 1 + ϵ ( n m ) 2 1 + ϵ log ( T m / n ) ≪ T 2 1 + δ ( n m ) 2 1 + δ .
Now, define
S a = ∑ 1 ≤ n ≤ m ≤ N ∣ r ( m ) r ( n ) ∣ ( m n ) a . S_a=\sum_{1\le n\le m\le N}|r(m)r(n)|\left(\frac{m}{n}\right)^a. S a = 1 ≤ n ≤ m ≤ N ∑ ∣ r ( m ) r ( n ) ∣ ( n m ) a .
By the Cauchy–Schwarz inequality,
S a ≤ ( ∑ m ≤ N m 2 a ) 1 / 2 ( ∑ n ≤ N n − 2 a ) 1 / 2 ∑ n ≤ N ∣ r ( n ) ∣ 2 . S_a\le\left(\sum_{m\le N}m^{2a}\right)^{1/2}\left(\sum_{n\le N}n^{-2a}\right)^{1/2}\sum_{n\le N}|r(n)|^2. S a ≤ ( m ≤ N ∑ m 2 a ) 1/2 ( n ≤ N ∑ n − 2 a ) 1/2 n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Consequently,
(4.20) \qquad S a ≪ N ∑ n ≤ N ∣ r ( n ) ∣ 2 ( ∣ a ∣ < 1 / 2 ) , S 1 / 2 + δ ≪ N 1 + δ ∑ n ≤ N ∣ r ( n ) ∣ 2 ( δ > 0 ) . S_a\ll N\sum_{n\le N}|r(n)|^2\quad(|a|<1/2),\qquad S_{1/2+\delta}\ll N^{1+\delta}\sum_{n\le N}|r(n)|^2\quad(\delta>0). S a ≪ N ∑ n ≤ N ∣ r ( n ) ∣ 2 ( ∣ a ∣ < 1/2 ) , S 1/2 + δ ≪ N 1 + δ ∑ n ≤ N ∣ r ( n ) ∣ 2 ( δ > 0 ) .
Using (4.20) in (4.19), we obtain
J F ≪ ( T γ N ( log T ) c + T 1 2 + δ N 1 + δ ) ∑ n ≤ N ∣ r ( n ) ∣ 2 , c = max ( γ 1 , 0 ) . (52) J_F\ll\left(T^\gamma N(\log T)^c+T^{\frac{1}{2}+\delta}N^{1+\delta}\right)\sum_{n\le N}|r(n)|^2,\qquad c=\max(\gamma_1,0). \tag*{(52)} J F ≪ ( T γ N ( log T ) c + T 2 1 + δ N 1 + δ ) n ≤ N ∑ ∣ r ( n ) ∣ 2 , c = max ( γ 1 , 0 ) . (52)
Choose N = [ T α ] N=[T^\alpha] N = [ T α ] with any fixed α < 1 − γ \alpha<1-\gamma α < 1 − γ , and choose
0 < δ < 1 / 2 − α 1 + α . 0<\delta<\frac{1/2-\alpha}{1+\alpha}. 0 < δ < 1 + α 1/2 − α .
Then α < 1 / 2 \alpha<1/2 α < 1/2 , γ + α < 1 / 2 \gamma+\alpha<1/2 γ + α < 1/2 , and 1 / 2 + δ + α ( 1 + δ ) < 1 1/2+\delta+\alpha(1+\delta)<1 1/2 + δ + α ( 1 + δ ) < 1 . Thus, from (4.21), we get J F = o ( T ∑ n ≤ N ∣ r ( n ) ∣ 2 ) J_F=o\left(T\sum_{n\le N}|r(n)|^2\right) J F = o ( T ∑ n ≤ N ∣ r ( n ) ∣ 2 ) . This completes the proof. □ \square □
Proof of Theorem 1.3. Set N = [ T α ] N=[T^\alpha] N = [ T α ] , with α < 1 − γ \alpha<1-\gamma α < 1 − γ , and given ϵ > 0 \epsilon>0 ϵ > 0 , choose η > 0 \eta>0 η > 0 sufficiently small that κ ( 1 − η ) > κ − ϵ \sqrt{\kappa(1-\eta)}>\sqrt{\kappa}-\epsilon κ ( 1 − η ) > κ − ϵ . Then
log N log log N = ( α + o ( 1 ) ) log T log log T . \sqrt{\frac{\log N}{\log\log N}}=(\sqrt{\alpha}+o(1))\sqrt{\frac{\log T}{\log\log T}}. log log N log N = ( α + o ( 1 )) log log T log T .
Thus, Lemmas 4.1 and 4.2 give
I F ≫ A ( T ) T ∑ n ≤ N ∣ r ( n ) ∣ 2 , J F = o ( T ∑ n ≤ N ∣ r ( n ) ∣ 2 ) , I_F\gg A(T)T\sum_{n\le N}|r(n)|^2,\qquad J_F=o\left(T\sum_{n\le N}|r(n)|^2\right), I F ≫ A ( T ) T n ≤ N ∑ ∣ r ( n ) ∣ 2 , J F = o ( T n ≤ N ∑ ∣ r ( n ) ∣ 2 ) ,
where
A ( T ) = exp ( ( κ ( 1 − γ ) − ϵ ) log T log log T ) . A(T)=\exp\left((\sqrt{\kappa(1-\gamma)}-\epsilon)\sqrt{\frac{\log T}{\log\log T}}\right). A ( T ) = exp ( ( κ ( 1 − γ ) − ϵ ) log log T log T ) .
Since ∣ Z F ( t ) ∣ = ∣ F ( 1 / 2 + i t ) ∣ |Z_F(t)|=|F(1/2+it)| ∣ Z F ( t ) ∣ = ∣ F ( 1/2 + i t ) ∣ and 0 ≤ Φ ≤ 1 0\leq\Phi\leq1 0 ≤ Φ ≤ 1 , we have
∫ T 2 T ∣ Z F ( t ) ∣ ∣ R ( t ) ∣ 2 d t ≥ ∣ I F ∣ ≫ A ( T ) T ∑ n ≤ N ∣ r ( n ) ∣ 2 . \int_T^{2T}|Z_F(t)||R(t)|^2\,dt\geq|I_F|\gg A(T)T\sum_{n\leq N}|r(n)|^2. ∫ T 2 T ∣ Z F ( t ) ∣∣ R ( t ) ∣ 2 d t ≥ ∣ I F ∣ ≫ A ( T ) T n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Also, by (23) ,
∫ T 2 T ∣ R ( t ) ∣ 2 d t ∼ T ∑ n ≤ N ∣ r ( n ) ∣ 2 . \int_T^{2T}|R(t)|^2\,dt\sim T\sum_{n\leq N}|r(n)|^2. ∫ T 2 T ∣ R ( t ) ∣ 2 d t ∼ T n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Using
2 ∫ T 2 T Z F + ( t ) ∣ R ( t ) ∣ 2 d t = ∫ T 2 T ∣ Z F ( t ) ∣ ∣ R ( t ) ∣ 2 d t + J F , 2\int_T^{2T}Z_F^+(t)|R(t)|^2\,dt=\int_T^{2T}|Z_F(t)||R(t)|^2\,dt+J_F, 2 ∫ T 2 T Z F + ( t ) ∣ R ( t ) ∣ 2 d t = ∫ T 2 T ∣ Z F ( t ) ∣∣ R ( t ) ∣ 2 d t + J F ,
we obtain
∫ T 2 T Z F + ( t ) ∣ R ( t ) ∣ 2 d t ≫ A ( T ) T ∑ n ≤ N ∣ r ( n ) ∣ 2 . \int_T^{2T}Z_F^+(t)|R(t)|^2\,dt\gg A(T)T\sum_{n\leq N}|r(n)|^2. ∫ T 2 T Z F + ( t ) ∣ R ( t ) ∣ 2 d t ≫ A ( T ) T n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Dividing by ∫ T 2 T ∣ R ( t ) ∣ 2 d t \int_T^{2T}|R(t)|^2\,dt ∫ T 2 T ∣ R ( t ) ∣ 2 d t , we get
max T ≤ t ≤ 2 T Z F + ( t ) ≫ exp ( ( κ ( 1 − γ ) − ϵ ) log T log log T ) . \max_{T\leq t\leq2T}Z_F^+(t)\gg\exp\left(\left(\sqrt{\kappa(1-\gamma)}-\epsilon\right)\sqrt{\frac{\log T}{\log\log T}}\right). T ≤ t ≤ 2 T max Z F + ( t ) ≫ exp ( ( κ ( 1 − γ ) − ϵ ) log log T log T ) .
The same argument gives the corresponding result for Z F − ( t ) Z_F^-(t) Z F − ( t ) . This completes the proof. □
Proof of Theorem 1.6. The proof of this theorem is similar to that of Theorem 1.3. From Lemmas 2.5 and 2.6, we have γ = 1 / 2 \gamma=1/2 γ = 1/2 and κ = 1 \kappa=1 κ = 1 . Hence, the constant C F C_F C F is 1 / 2 − ϵ 1/\sqrt{2}-\epsilon 1/ 2 − ϵ , where ϵ > 0 \epsilon>0 ϵ > 0 is arbitrarily small. □
The unconditional Hecke–Maass case
We now remove the Ramanujan hypothesis from the Hecke–Maass case of Theorem 1.6. Let θ = 7 / 64 \theta=7/64 θ = 7/64 . By the Kim–Sarnak bound, for every ϵ > 0 \epsilon>0 ϵ > 0 ,
a F ( n ) ≪ n θ + ϵ . (53) a_F(n)\ll n^{\theta+\epsilon}. \tag*{(53)} a F ( n ) ≪ n θ + ϵ . (53)
Moreover, by Rankin–Selberg theory, the Dirichlet series associated with F ( s ) F(s) F ( s ) converges absolutely in the half-plane ℜ ( s ) > 1 \Re(s)>1 ℜ ( s ) > 1 .
Proof of Theorem 1.7. In the proof of Lemma 4.1, the off-diagonal contribution is estimated using the rapid decay of the Fourier transform Φ ^ \widehat{\Phi} Φ . In the present case, we use the same argument together with (53) . Since
Φ ^ ( y ) ≪ B ∣ y ∣ − B \widehat{\Phi}(y)\ll_B |y|^{-B} Φ ( y ) ≪ B ∣ y ∣ − B
for any fixed B > 0 B>0 B > 0 , we choose B B B sufficiently large so that the bound (53) is sufficient to obtain the required off-diagonal estimate. Hence, we have
I F = T Φ ^ ( 0 ) ∑ n k = m ≤ N a F ( k ) r ( n ) r ( m ) ‾ k e − k / X + O ( T ∑ n ≤ N ∣ r ( n ) ∣ 2 ) . I_F=T\widehat{\Phi}(0)\sum_{nk=m\leq N}\frac{a_F(k)r(n)\overline{r(m)}}{\sqrt{k}}e^{-k/X}+O\left(T\sum_{n\leq N}|r(n)|^2\right). I F = T Φ ( 0 ) nk = m ≤ N ∑ k a F ( k ) r ( n ) r ( m ) e − k / X + O ( T n ≤ N ∑ ∣ r ( n ) ∣ 2 ) .
As e − k / X ≥ 1 / 2 e^{-k/X}\geq1/2 e − k / X ≥ 1/2 for k < T 1 − ϵ k<T^{1-\epsilon} k < T 1 − ϵ and a F ( k ) r ( n ) r ( n k ) ‾ = ∣ a F ( k ) ∣ 2 ∣ a F ( n ) ∣ 2 f ( n ) 2 f ( k ) a_F(k)r(n)\overline{r(nk)}=|a_F(k)|^2|a_F(n)|^2f(n)^2f(k) a F ( k ) r ( n ) r ( nk ) = ∣ a F ( k ) ∣ 2 ∣ a F ( n ) ∣ 2 f ( n ) 2 f ( k ) , applying Lemmas 2.7, 2.10 and we proof Lemma 4.1, we have
I F ≫ T exp ( ( κ ( 1 − η ) + o ( 1 ) ) log N log log N ) ∑ n ≤ N ∣ r ( n ) ∣ 2 . I_F\gg T\exp\left(\left(\sqrt{\kappa(1-\eta)}+o(1)\right)\sqrt{\frac{\log N}{\log\log N}}\right)\sum_{n\leq N}|r(n)|^2. I F ≫ T exp ( ( κ ( 1 − η ) + o ( 1 ) ) log log N log N ) n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Now, proceeding as in the proof of Lemma 4.2, and using (4.22) in place of the Ramanujan hypothesis, we obtain, by choosing δ > 0 \delta> 0 δ > 0 sufficiently small, for
N = ⌊ T α ⌋ with α < 1 − 2 θ 2 ( 1 + θ ) , N = \lfloor T^{\alpha} \rfloor\quad\text{with} \quad\alpha< \frac{1-2\theta}{2(1+\theta)}, N = ⌊ T α ⌋ with α < 2 ( 1 + θ ) 1 − 2 θ ,
that
J F = o ( T ∑ n ≤ N ∣ r ( n ) ∣ 2 ) . J_F = o\left(T\sum_{n\leq N}|r(n)|^2\right). J F = o ( T n ≤ N ∑ ∣ r ( n ) ∣ 2 ) .
Hence, the argument used in the proof of Theorem 1.3 carries over to the present case with the above modifications. This completes the proof of Theorem 1.7.
Proof of Theorem 1.8
We recall a few basic facts about the Dedekind zeta function. Let K K K be a number field. For ℜ ( s ) > 1 \Re(s)>1 ℜ ( s ) > 1 , its Dedekind zeta function is defined by
ζ K ( s ) = ∑ a ≠ 0 1 N ( a ) s = ∑ n = 1 ∞ a K ( n ) n s . (54) \zeta_K(s)=\sum_{\mathfrak{a}\ne0}\frac{1}{N(\mathfrak{a})^s}=\sum_{n=1}^{\infty}\frac{a_K(n)}{n^s}. \tag*{(54)} ζ K ( s ) = a = 0 ∑ N ( a ) s 1 = n = 1 ∑ ∞ n s a K ( n ) . (54)
where the first sum runs over the non-zero integral ideals a \mathfrak{a} a of K K K , N ( a ) N(\mathfrak{a}) N ( a ) denotes the absolute norm of a \mathfrak{a} a , and a K ( n ) a_K(n) a K ( n ) is the number of integral ideals of K K K having norm n n n .
Now, let K = Q ( D K ) K=\mathbb{Q}(\sqrt{D_K}) K = Q ( D K ) be a quadratic field, where D K D_K D K is the fundamental discriminant of K K K . In this case, the Dedekind zeta function factors as
ζ K ( s ) = ζ ( s ) L ( s , χ D K ) , \zeta_K(s)=\zeta(s)L(s,\chi_{D_K}), ζ K ( s ) = ζ ( s ) L ( s , χ D K ) ,
where
χ D K ( n ) = ( D K n ) \chi_{D_K}(n)=\left(\frac{D_K}{n}\right) χ D K ( n ) = ( n D K )
is the primitive quadratic Dirichlet character associated with K K K . Comparing the Dirichlet series coefficients gives
a K ( n ) = ∑ d ∣ n χ D K ( d ) . (55) a_K(n)=\sum_{d\mid n}\chi_{D_K}(d). \tag*{(55)} a K ( n ) = d ∣ n ∑ χ D K ( d ) . (55)
We also consider the product F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) F(s)=L(s,\chi_1)L(s,\chi_2) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) , where χ 1 \chi_1 χ 1 and χ 2 \chi_2 χ 2 are primitive Dirichlet characters modulo q 1 q_1 q 1 and q 2 q_2 q 2 , respectively, and q 1 q 2 q_1q_2 q 1 q 2 is not a perfect square. Writing F ( s ) = ∑ n = 1 ∞ a F ( n ) n − s F(s)=\sum_{n=1}^{\infty}a_F(n)n^{-s} F ( s ) = ∑ n = 1 ∞ a F ( n ) n − s , we have
a F ( n ) = ∑ u v = n χ 1 ( u ) χ 2 ( v ) . a_F(n)=\sum_{uv=n}\chi_1(u)\chi_2(v). a F ( n ) = uv = n ∑ χ 1 ( u ) χ 2 ( v ) .
To prove Theorem 1.8, we need a uniform upper bound for the exponential sum appearing in (4.18). However, for the Dedekind zeta function ζ K ( s ) \zeta_K(s) ζ K ( s ) over a quadratic field K K K , as well as for the product L ( s , χ 1 ) L ( s , χ 2 ) L(s,\chi_1)L(s,\chi_2) L ( s , χ 1 ) L ( s , χ 2 ) , a bound of the required strength is not currently available. For ζ K ( s ) \zeta_K(s) ζ K ( s ) , the known estimates are of the weaker form o ( T ) o(T) o ( T ) for irrational x x x , as shown by Chandrasekharan and Narasimhan [4 ] . This is not sufficient for our argument, since we require an upper bound involving a power of T T T strictly smaller than 1. To obtain such an estimate, we use the following lemma.
Lemma 5.1. [36 Lemma 2.2] . Let Y Y Y , Z Z Z , and α 0 \alpha_0 α 0 be real numbers such that Y , Z ≥ 1 Y,Z\geq1 Y , Z ≥ 1 , and suppose that
∣ α 0 − a q ∣ ≤ 1 q 2 , \left|\alpha_0-\frac{a}{q}\right|\leq\frac{1}{q^2}, α 0 − q a ≤ q 2 1 ,
where ( a , q ) = 1 (a,q)=1 ( a , q ) = 1 . Then
∑ x ≤ Y min ( Y Z x , ∥ α x ∥ − 1 ) ≪ ( Y Z q + Y + q ) log ( 2 Y q ) , \sum_{x\le Y}\min\left(\frac{YZ}{x},\|\alpha x\|^{-1}\right)\ll\left(\frac{YZ}{q}+Y+q\right)\log(2Yq), x ≤ Y ∑ min ( x Y Z , ∥ α x ∥ − 1 ) ≪ ( q Y Z + Y + q ) log ( 2 Y q ) ,
where ∥ y ∥ \|y\| ∥ y ∥ denotes distance to the nearest integer.
We next establish an estimate for the exponential sums corresponding to the Dedekind zeta function ζ K ( s ) \zeta_K(s) ζ K ( s ) over a quadratic field K K K and to the product L ( s , χ 1 ) L ( s , χ 2 ) L(s,\chi_1)L(s,\chi_2) L ( s , χ 1 ) L ( s , χ 2 ) .
Lemma 5.2. Let F F F be one of the following:
(i) F ( s ) = ζ K ( s ) F(s)=\zeta_K(s) F ( s ) = ζ K ( s ) , where K = Q ( D K ) K=\mathbb{Q}(\sqrt{D_K}) K = Q ( D K ) is any quadratic field with K ≠ Q ( i ) K\ne\mathbb{Q}(i) K = Q ( i ) , in which case q F = ∣ D K ∣ q_F=|D_K| q F = ∣ D K ∣ and a F ( n ) = a K ( n ) a_F(n)=a_K(n) a F ( n ) = a K ( n ) is given by (5.2);
(ii) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) F(s)=L(s,\chi_1)L(s,\chi_2) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) , where χ 1 \chi_1 χ 1 and χ 2 \chi_2 χ 2 are primitive Dirichlet characters modulo q 1 q_1 q 1 and q 2 q_2 q 2 , respectively, with q 1 q 2 q_1q_2 q 1 q 2 not a perfect square, in which case q F = q 1 q 2 q_F=q_1q_2 q F = q 1 q 2 and a F ( k ) = ∑ u v = k χ 1 ( u ) χ 2 ( v ) a_F(k)=\sum_{uv=k}\chi_1(u)\chi_2(v) a F ( k ) = ∑ uv = k χ 1 ( u ) χ 2 ( v ) .
Then
∑ k ≤ X a F ( k ) e ( − k n m q F ) ≪ m n X log ( 2 m n X ) , (56) \sum_{k\le X}a_F(k)e\left(-\frac{kn}{m\sqrt{q_F}}\right)\ll\sqrt{mnX}\log(2mnX), \tag*{(56)} k ≤ X ∑ a F ( k ) e ( − m q F k n ) ≪ mn X log ( 2 mn X ) , (56) uniformly for positive integers m , n m,n m , n and X ≥ 2 X\ge2 X ≥ 2 , where e ( x ) : = e 2 π i x e(x):=e^{2\pi ix} e ( x ) := e 2 π i x . Here, the implied constant depends only on F F F .
Proof. Set
β = 1 q F , θ = n m β , \beta=\frac{1}{\sqrt{q_F}},\qquad\theta=\frac{n}{m}\beta, β = q F 1 , θ = m n β ,
and
S m , n ( X ) : = ∑ k ≤ X a K ( k ) e ( − θ k ) . S_{m,n}(X):=\sum_{k\le X}a_K(k)e(-\theta k). S m , n ( X ) := k ≤ X ∑ a K ( k ) e ( − θ k ) .
We first consider case (i): F ( s ) = ζ K ( s ) F(s)=\zeta_K(s) F ( s ) = ζ K ( s ) . In this case, q F = ∣ D K ∣ = : d q_F=|D_K|=:d q F = ∣ D K ∣ =: d . Since K ≠ Q ( i ) K\ne\mathbb{Q}(i) K = Q ( i ) , the integer d d d is not a perfect square. Hence, β \beta β is a quadratic irrational. Therefore, there exists a constant c K > 0 c_K>0 c K > 0 such that
∣ β − u v ∣ ≥ c K v 2 ( u ∈ Z , v ≥ 1 ) \left|\beta-\frac{u}{v}\right|\ge\frac{c_K}{v^2}\qquad(u\in\mathbb{Z},\ v\ge1) β − v u ≥ v 2 c K ( u ∈ Z , v ≥ 1 )
Applying this with the rational number m a / ( n q ) ma/(nq) ma / ( n q ) , we obtain
∣ θ − a q ∣ = n m ∣ β − m a n q ∣ ≥ n m c K n 2 q 2 = c K m n q 2 . (57) \begin{aligned}
\left|\theta-\frac{a}{q}\right|=\frac{n}{m}\left|\beta-\frac{ma}{nq}\right| \\
&\ge\frac{n}{m}\frac{c_K}{n^2q^2}=\frac{c_K}{mnq^2}. \tag*{(57)}
\end{aligned} θ − q a = m n β − n q ma ≥ m n n 2 q 2 c K = mn q 2 c K . (57)
The fraction m a / ( n q ) ma/(nq) ma / ( n q ) need not be reduced; reducing it only strengthens the lower bound. We also need the same estimate for
d θ = n m d . d\theta=\frac{n}{m}\sqrt{d}. d θ = m n d .
Since d \sqrt{d} d is another fixed quadratic irrational, after decreasing c K c_K c K , we have
∣ ϑ − a q ∣ ≥ c K m n q 2 , ϑ ∈ { θ , d θ } . (58) \left|\vartheta-\frac{a}{q}\right|\ge\frac{c_K}{mnq^2},\qquad\vartheta\in\{\theta,d\theta\}. \tag*{(58)} ϑ − q a ≥ mn q 2 c K , ϑ ∈ { θ , d θ } . (58)
Now we choose
Q = ⌊ m n X ⌋ . Q=\left\lfloor\sqrt{mnX}\right\rfloor. Q = ⌊ mn X ⌋ .
For each of ϑ = θ , d θ \vartheta=\theta,d\theta ϑ = θ , d θ , Dirichlet’s approximation theorem gives its own coprime integers a , q a,q a , q satisfying
1 ≤ q ≤ Q , ∣ ϑ − a q ∣ ≤ 1 q Q . (59) 1\le q\le Q,\qquad\left|\vartheta-\frac{a}{q}\right|\le\frac{1}{qQ}. \tag*{(59)} 1 ≤ q ≤ Q , ϑ − q a ≤ q Q 1 . (59)
Combining this with (58) ,
c K m n q 2 ≤ 1 q Q , \frac{c_K}{mnq^2}\le\frac{1}{qQ}, mn q 2 c K ≤ q Q 1 ,
and therefore
c K Q m n ≤ q ≤ Q . (60) \frac{c_KQ}{mn}\le q\le Q. \tag*{(60)} mn c K Q ≤ q ≤ Q . (60)
In particular, we have
X q + q ≪ K m n X . (61) \frac{X}{q}+q\ll_K\sqrt{mnX}. \tag*{(61)} q X + q ≪ K mn X . (61)
Indeed,
X q ≤ m n X c K Q ≪ K m n X , q ≤ Q ≪ m n X . \frac{X}{q}\le\frac{mnX}{c_KQ}\ll_K\sqrt{mnX},\qquad q\le Q\ll\sqrt{mnX}. q X ≤ c K Q mn X ≪ K mn X , q ≤ Q ≪ mn X .
Using (55) , we have
S m , n ( X ) : = ∑ k ≤ X a K ( k ) e ( − θ k ) = ∑ u v ≤ X χ D K ( u ) e ( − θ u v ) . S_{m,n}(X):=\sum_{k\le X}a_K(k)e(-\theta k)=\sum_{uv\le X}\chi_{D_K}(u)e(-\theta uv). S m , n ( X ) := k ≤ X ∑ a K ( k ) e ( − θ k ) = uv ≤ X ∑ χ D K ( u ) e ( − θ uv ) .
Let H = ⌊ X ⌋ H=\lfloor\sqrt{X}\rfloor H = ⌊ X ⌋ . Every pair u v ≤ X uv\le X uv ≤ X has either u ≤ H u\le H u ≤ H , or u > H u>H u > H and v ≤ H v\le H v ≤ H . Thus, we have
S m , n ( X ) = ∑ u ≤ H χ D K ( u ) ∑ v ≤ X / u e ( − θ u v ) + ∑ v ≤ H ∑ H < u ≤ X / v χ D K ( u ) e ( − θ u v ) . (62) \begin{aligned}
S_{m,n}(X)
&=\sum_{u\le H}\chi_{D_K}(u)\sum_{v\le X/u}e(-\theta uv)\\
&\quad+\sum_{v\le H}\sum_{H<u\le X/v}\chi_{D_K}(u)e(-\theta uv). \tag*{(62)}
\end{aligned} S m , n ( X ) = u ≤ H ∑ χ D K ( u ) v ≤ X / u ∑ e ( − θ uv ) + v ≤ H ∑ H < u ≤ X / v ∑ χ D K ( u ) e ( − θ uv ) . (62)
For β o ∉ Z \beta_o\notin\mathbb{Z} β o ∈ / Z and an interval of L L L consecutive integers, the standard geometric-sum estimate gives
∣ ∑ ℓ = A + 1 A + L e ( β o ℓ ) ∣ ≪ min { L , ∥ β o ∥ − 1 } , (63) \left|\sum_{\ell=A+1}^{A+L}e(\beta_o\ell)\right|\ll\min\{L,\|\beta_o\|^{-1}\}, \tag*{(63)} ℓ = A + 1 ∑ A + L e ( β o ℓ ) ≪ min { L , ∥ β o ∥ − 1 } , (63)
where ∥ β o ∥ \|\beta_o\| ∥ β o ∥ denotes distance to the nearest integer (see [14 Eq. (8.6)] ). Since ∣ χ D K ( u ) ∣ ≤ 1 |\chi_{D_K}(u)|\le1 ∣ χ D K ( u ) ∣ ≤ 1 , using (63) , the first term in (62) is bounded by
∑ u ≤ H min ( X u , ∥ u θ ∥ − 1 ) . (64) \sum_{u\le H}\min\left(\frac{X}{u},\|u\theta\|^{-1}\right). \tag*{(64)} u ≤ H ∑ min ( u X , ∥ u θ ∥ − 1 ) . (64)
For the second term, fix v v v and split u u u into residue classes modulo d d d , we have
∑ H < u ≤ X / v χ D ( u ) e ( − θ u v ) = ∑ b m o d d χ D ( b ) e ( − θ v b ) ∑ ℓ ∈ Z H < b + d ℓ ≤ X / v e ( − d θ v ℓ ) . (65) \sum_{H<u\le X/v}\chi_D(u)e(-\theta uv)
=
\sum_{b\bmod d}\chi_D(b)e(-\theta vb)
\sum_{\substack{\ell\in\mathbb{Z}\\ H<b+d\ell\le X/v}}e(-d\theta v\ell). \tag*{(65)} H < u ≤ X / v ∑ χ D ( u ) e ( − θ uv ) = b mod d ∑ χ D ( b ) e ( − θ v b ) ℓ ∈ Z H < b + d ℓ ≤ X / v ∑ e ( − d θ v ℓ ) . (65)
The inner sum is geometric. Its number of terms is at most X / ( d v ) + 1 X/(dv)+1 X / ( d v ) + 1 , so
∣ ∑ H < u ≤ X / v χ D ( u ) e ( − θ u v ) ∣ ≪ K min ( X v , ∥ d v θ ∥ − 1 ) . (66) \left|\sum_{H<u\le X/v}\chi_D(u)e(-\theta uv)\right|\ll_K\min\left(\frac{X}{v},\lVert dv\theta\rVert^{-1}\right). \tag*{(66)} H < u ≤ X / v ∑ χ D ( u ) e ( − θ uv ) ≪ K min ( v X , ∥ d v θ ∥ − 1 ) . (66)
Here the factor arising from the d d d residue classes is absorbed into the constant depending on K K K . Define
V ( X , w ) = ∑ h ≤ H min ( X h , ∥ h w ∥ − 1 ) . V(X,w)=\sum_{h\le H}\min\left(\frac{X}{h},\lVert hw\rVert^{-1}\right). V ( X , w ) = h ≤ H ∑ min ( h X , ∥ h w ∥ − 1 ) .
Then equations (62) –(66) give
S m , n ( X ) ≪ K V ( X , θ ) + V ( X , d θ ) . (67) S_{m,n}(X)\ll_K V(X,\theta)+V(X,d\theta). \tag*{(67)} S m , n ( X ) ≪ K V ( X , θ ) + V ( X , d θ ) . (67)
The approximations constructed in (59) satisfy the hypothesis of Lemma 5.1 because
q ≤ Q ⇒ 1 q Q ≤ 1 q 2 . q\le Q\quad\Rightarrow\quad\frac{1}{qQ}\le\frac{1}{q^2}. q ≤ Q ⇒ q Q 1 ≤ q 2 1 .
Therefore, using Lemma 5.1, we have
∑ h ≤ H min ( X h , ∥ h α 0 ∥ − 1 ) ≪ ( X q + H + q ) log ( 2 X q ) . (68) \sum_{h\le H}\min\left(\frac{X}{h},\lVert h\alpha_0\rVert^{-1}\right)\ll\left(\frac{X}{q}+H+q\right)\log(2Xq). \tag*{(68)} h ≤ H ∑ min ( h X , ∥ h α 0 ∥ − 1 ) ≪ ( q X + H + q ) log ( 2 X q ) . (68)
By (61) and H ≤ X H\le\sqrt{X} H ≤ X ,
X q + H + q ≪ K m n X . \frac{X}{q}+H+q\ll_K\sqrt{mnX}. q X + H + q ≪ K mn X .
Also q ≪ m n X q\ll\sqrt{mnX} q ≪ mn X , so
log ( 2 X q ) ≪ log ( 2 m n X ) . \log(2Xq)\ll\log(2mnX). log ( 2 X q ) ≪ log ( 2 mn X ) .
Apply (68) separately to θ \theta θ and d θ d\theta d θ . We obtain
V ( X , θ ) , V ( X , d θ ) ≪ K m n X log ( 2 m n X ) . V(X,\theta),\ V(X,d\theta)\ll_K\sqrt{mnX}\log(2mnX). V ( X , θ ) , V ( X , d θ ) ≪ K mn X log ( 2 mn X ) .
Substituting this into (67) proves the lemma in case ( i ) (i) ( i ) .
We now consider case ( i i ) (ii) ( ii ) : F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) F(s)=L(s,\chi_1)L(s,\chi_2) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) . In this case, q F = q 1 q 2 q_F=q_1q_2 q F = q 1 q 2 . Since q 1 q 2 q_1q_2 q 1 q 2 is not a perfect square, β \beta β is a quadratic irrational. Since a F ( k ) = ∑ u v = k χ 1 ( u ) χ 2 ( v ) a_F(k)=\sum_{uv=k}\chi_1(u)\chi_2(v) a F ( k ) = ∑ uv = k χ 1 ( u ) χ 2 ( v ) , as in (62) , we have
S m , n ( X ) = ∑ u ≤ H χ 1 ( u ) ∑ v ≤ X / u χ 2 ( v ) e ( − θ u v ) + ∑ v < H χ 2 ( v ) ∑ H < u ≤ X / v χ 1 ( u ) e ( − θ u v ) . (69) \begin{aligned}
S_{m,n}(X)&=\sum_{u\le H}\chi_1(u)\sum_{v\le X/u}\chi_2(v)e(-\theta uv)\\
&\quad+\sum_{v<H}\chi_2(v)\sum_{H<u\le X/v}\chi_1(u)e(-\theta uv). \tag*{(69)}
\end{aligned} S m , n ( X ) = u ≤ H ∑ χ 1 ( u ) v ≤ X / u ∑ χ 2 ( v ) e ( − θ uv ) + v < H ∑ χ 2 ( v ) H < u ≤ X / v ∑ χ 1 ( u ) e ( − θ uv ) . (69)
Using the periodicity of χ 1 \chi_1 χ 1 and χ 2 \chi_2 χ 2 , we split the corresponding inner sums into residue classes modulo q 1 q_1 q 1 and q 2 q_2 q 2 , respectively. Each resulting sum is a geometric exponential sum of the same type as in case ( i ) (i) ( i ) . Since q 1 q_1 q 1 and q 2 q_2 q 2 are fixed, the same argument applied to (69) gives S m , n ( X ) ≪ χ 1 , χ 2 m n X log ( 2 m n X ) S_{m,n}(X)\ll_{\chi_1,\chi_2}\sqrt{mnX}\log(2mnX) S m , n ( X ) ≪ χ 1 , χ 2 mn X log ( 2 mn X ) . This completes the proof of the lemma in case ( i i ) (ii) ( ii ) . □ \square □
Lemma 5.3. Let K = Q ( i ) K=\mathbb{Q}(i) K = Q ( i ) be a quadratic field and a K ( n ) a_K(n) a K ( n ) be as defined in (55) . Then
∑ k ≤ X a K ( k ) e ( − k n m ∣ D K ∣ ) ≪ m ( m , n ) X + m 2 ( m , n ) 2 , (70) \sum_{k\le X}a_K(k)e\left(-\frac{kn}{m\sqrt{|D_K|}}\right)\ll\frac{m}{(m,n)}\sqrt{X}+\frac{m^2}{(m,n)^2}, \tag*{(70)} k ≤ X ∑ a K ( k ) e ( − m ∣ D K ∣ k n ) ≪ ( m , n ) m X + ( m , n ) 2 m 2 , (70) :::`
uniformly for positive integers m , n m,n m , n such that r ( m ) r ( n ) ≠ 0 r(m)r(n) \ne0 r ( m ) r ( n ) = 0 , for sufficiently large N N N , where r ( n ) r(n) r ( n ) is as defined in (16) , and X ≥ 2 X \ge2 X ≥ 2 . Here, the dependence of the implied constant is only on K K K .
Proof. For K = Q ( i ) K=\mathbb{Q}(i) K = Q ( i ) , we have D K = − 4 D_K=-4 D K = − 4 . Assume that
S m , n ( X ) = ∑ k ≤ X a K ( k ) e ( − k n 2 m ) . (71) S_{m,n}(X)=\sum_{k\le X}a_K(k)e\left(-\frac{kn}{2m}\right). \tag*{(71)} S m , n ( X ) = k ≤ X ∑ a K ( k ) e ( − 2 m k n ) . (71)
Here, we only need to consider the pairs m , n m,n m , n for which r ( m ) r ( n ) ≠ 0 r(m)r(n)\ne0 r ( m ) r ( n ) = 0 . Recall that r ( n ) = a K ( n ) f ( n ) r(n)=a_K(n)f(n) r ( n ) = a K ( n ) f ( n ) , where f f f is multiplicative defined in (18) . Since
L 2 = 1 2 log N log log N ⟶ ∞ , L^2=\frac{1}{2}\log N\log\log N\longrightarrow\infty, L 2 = 2 1 log N log log N ⟶ ∞ ,
we have L 2 > 2 L^2>2 L 2 > 2 for sufficiently large N N N . Hence f ( 2 ) = 0 f(2)=0 f ( 2 ) = 0 . Thus, if 2 ∣ n 2\mid n 2 ∣ n , then by the multiplicativity of f f f , f ( n ) = 0 f(n)=0 f ( n ) = 0 , and consequently r ( n ) = 0 r(n)=0 r ( n ) = 0 .
Therefore, for every non-zero contribution to the resonator sum, both m m m and n n n are odd. Let
g = ( m , n ) , m = g m 1 , n = g n 1 , g=(m,n),\qquad m=gm_1,\qquad n=gn_1, g = ( m , n ) , m = g m 1 , n = g n 1 ,
so that ( m 1 , n 1 ) = 1 (m_1,n_1)=1 ( m 1 , n 1 ) = 1 . Hence
n 2 m = n 1 2 m 1 \frac{n}{2m}=\frac{n_1}{2m_1} 2 m n = 2 m 1 n 1
is in lowest terms, where m 1 m_1 m 1 and n 1 n_1 n 1 are odd. Put
a = n 1 , q = 2 m 1 = 2 m ( m , n ) . a=n_1,\qquad q=2m_1=\frac{2m}{(m,n)}. a = n 1 , q = 2 m 1 = ( m , n ) 2 m .
Then ( a , q ) = 1 (a,q)=1 ( a , q ) = 1 and, since m 1 m_1 m 1 is odd, q ≡ 2 ( m o d 4 ) q\equiv2\pmod4 q ≡ 2 ( mod 4 ) . For K = Q ( i ) K=\mathbb{Q}(i) K = Q ( i ) , we have
a K ( k ) = 1 4 r 2 ( k ) , a_K(k)=\frac{1}{4}r_2(k), a K ( k ) = 4 1 r 2 ( k ) ,
where
r 2 ( k ) = # { ( u , v ) ∈ Z 2 : u 2 + v 2 = k } . r_2(k)=\#\{(u,v)\in\mathbb{Z}^2:u^2+v^2=k\}. r 2 ( k ) = # {( u , v ) ∈ Z 2 : u 2 + v 2 = k } .
Therefore, from (71) ,
4 S m , n ( X ) = ∑ k ≤ X r 2 ( k ) e ( − a k q ) = ∑ u , v ∈ Z 0 < u 2 + v 2 ≤ X e ( − a ( u 2 + v 2 ) q ) . (72) 4S_{m,n}(X)=\sum_{k\le X}r_2(k)e\left(-\frac{ak}{q}\right)=\sum_{\substack{u,v\in\mathbb{Z}\\0<u^2+v^2\le X}}e\left(-\frac{a(u^2+v^2)}{q}\right). \tag*{(72)} 4 S m , n ( X ) = k ≤ X ∑ r 2 ( k ) e ( − q ak ) = u , v ∈ Z 0 < u 2 + v 2 ≤ X ∑ e ( − q a ( u 2 + v 2 ) ) . (72)
We now split u u u and v v v into residue classes modulo q q q . For r , s ( m o d q ) r,s\pmod q r , s ( mod q ) , define
N r , s ( X ) = # { ( u , v ) ∈ Z 2 : u 2 + v 2 ≤ X , u ≡ r ( m o d q ) , v ≡ s ( m o d q ) } . N_{r,s}(X)=\#\{(u,v)\in\mathbb{Z}^2:u^2+v^2\le X,\ u\equiv r\pmod q,\ v\equiv s\pmod q\}. N r , s ( X ) = # {( u , v ) ∈ Z 2 : u 2 + v 2 ≤ X , u ≡ r ( mod q ) , v ≡ s ( mod q )} .
Hence, (72) gives
4 S m , n ( X ) = ∑ r , s m o d q e ( − a ( r 2 + s 2 ) q ) N r , s ( X ) + O ( 1 ) . (73) 4S_{m,n}(X)=\sum_{r,s\mathbin{\mathrm{mod}} q}e\left(-\frac{a(r^2+s^2)}{q}\right)N_{r,s}(X)+O(1). \tag*{(73)} 4 S m , n ( X ) = r , s mod q ∑ e ( − q a ( r 2 + s 2 ) ) N r , s ( X ) + O ( 1 ) . (73)
We next estimate N r , s ( X ) N_{r,s}(X) N r , s ( X ) . Write u = r + q ℓ u=r+q\ell u = r + q ℓ , v = s + q j v=s+qj v = s + q j , where ℓ , j ∈ Z \ell,j\in\mathbb{Z} ℓ , j ∈ Z . Then u 2 + v 2 ≤ X u^2+v^2\le X u 2 + v 2 ≤ X is equivalent to
( ℓ + r q ) 2 + ( j + s q ) 2 ≤ X q 2 . \left(\ell+\frac{r}{q}\right)^2+\left(j+\frac{s}{q}\right)^2\le\frac{X}{q^2}. ( ℓ + q r ) 2 + ( j + q s ) 2 ≤ q 2 X .
Thus, N r , s ( X ) N_{r,s}(X) N r , s ( X ) is the number of integral lattice points in a disk of radius R = X / q R=\sqrt{X}/q R = X / q and center ( − r / q , − s / q ) (-r/q,-s/q) ( − r / q , − s / q ) . The above region is a bounded region, whose boundary is a circle, hence a rectifiable curve.
Now we use the planar lattice-point estimate of Davenport [5 p. 179] . If R \mathcal{R} R is a bounded region in R 2 \mathbb{R}^{2} R 2 whose boundary is a rectifiable curve of length L ( R ) L(\mathcal{R}) L ( R ) , then
∣ N ( R ) − V ( R ) ∣ < 4 ( L ( R ) + 1 ) , \lvert N(\mathcal{R})-V(\mathcal{R})\rvert<4\bigl(L(\mathcal{R})+1\bigr), ∣ N ( R ) − V ( R )∣ < 4 ( L ( R ) + 1 ) ,
where N ( R ) N(\mathcal{R}) N ( R ) denotes the number of integral lattice points in R \mathcal{R} R and V ( R ) V(\mathcal{R}) V ( R ) denotes its area.
Here, we have
V ( R ) = π R 2 = π X q 2 and L ( R ) = 2 π R = 2 π X q . V(\mathcal{R})=\pi R^{2}=\frac{\pi X}{q^{2}}
\quad\text{and}\quad
L(\mathcal{R})=2\pi R=\frac{2\pi\sqrt{X}}{q}. V ( R ) = π R 2 = q 2 π X and L ( R ) = 2 π R = q 2 π X .
Hence, uniformly in r , s r,s r , s ,
N r , s ( X ) = π X q 2 + O ( X q + 1 ) . N_{r,s}(X)=\frac{\pi X}{q^{2}}+O\left(\frac{\sqrt{X}}{q}+1\right). N r , s ( X ) = q 2 π X + O ( q X + 1 ) .
Consequently, from (73) , we have
4 S m , n ( X ) = π X q 2 ∑ r , s m o d q e ( − a ( r 2 + s 2 ) q ) + O ( q X + q 2 ) . (74) 4S_{m,n}(X)=\frac{\pi X}{q^{2}}\sum_{r,s\ \mathrm{mod}\ q}e\left(\frac{-a(r^{2}+s^{2})}{q}\right)+O\left(q\sqrt{X}+q^{2}\right). \tag*{(74)} 4 S m , n ( X ) = q 2 π X r , s mod q ∑ e ( q − a ( r 2 + s 2 ) ) + O ( q X + q 2 ) . (74)
as there are q 2 q^{2} q 2 pairs ( r , s ) ( m o d q ) (r,s)\pmod q ( r , s ) ( mod q ) . Define
G ( a , q ) : = ∑ r m o d q e ( a r 2 q ) . G(a,q):=\sum_{r\ \mathrm{mod}\ q}e\left(\frac{ar^{2}}{q}\right). G ( a , q ) := r mod q ∑ e ( q a r 2 ) .
Since q ≡ 2 ( m o d 4 ) q\equiv2\pmod4 q ≡ 2 ( mod 4 ) , write q = 2 h q=2h q = 2 h , where h h h is odd. Moreover, a a a is odd. Pairing the terms corresponding to r r r and r + h r+h r + h , we have ( r + h ) 2 = r 2 + 2 r h + h 2 (r+h)^{2}=r^{2}+2rh+h^{2} ( r + h ) 2 = r 2 + 2 r h + h 2 . Thus, the terms cancel pairwise, and hence G ( a , q ) = 0 G(a,q)=0 G ( a , q ) = 0 , so G ( − a , q ) = 0 G(-a,q)=0 G ( − a , q ) = 0 , as well. Therefore, the main term of (74) vanishes, giving
S m , n ( X ) ≪ q X + q 2 , q = 2 m ( m , n ) . S_{m,n}(X)\ll q\sqrt{X}+q^{2},
\qquad q=\frac{2m}{(m,n)}. S m , n ( X ) ≪ q X + q 2 , q = ( m , n ) 2 m .
Hence, the proof is complete. □
Lemma 5.4. Let F F F be one of the following functions:
(i) F ( s ) = ζ K ( s ) F(s)=\zeta_K(s) F ( s ) = ζ K ( s ) , where K K K is any quadratic field;
(ii) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) F(s)=L(s,\chi_1)L(s,\chi_2) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) , where χ 1 \chi_1 χ 1 and χ 2 \chi_2 χ 2 are primitive Dirichlet characters modulo q 1 q_1 q 1 and q 2 q_2 q 2 , respectively, with q 1 q 2 q_1q_2 q 1 q 2 not a perfect square.
Then, for sufficiently large T T T , we have
J F ≪ T log T ∑ n ≤ N ∣ r ( n ) ∣ 2 , J_F\ll T\log T\sum_{n\le N}\lvert r(n)\rvert^{2}, J F ≪ T log T n ≤ N ∑ ∣ r ( n ) ∣ 2 , where J F J_F J F is defined in (33) and N = ⌊ T 1 / 4 ⌋ N=\lfloor T^{1/4}\rfloor N = ⌊ T 1/4 ⌋ .
Proof. From (50) , we have
J F ≪ ∑ 1 ≤ n ≤ m ≤ N ∣ r ( m ) r ( n ) ∣ m / n ( ∣ S m , n ∣ + ( m n ) 1 + δ T 1 2 + δ + ( m n ) 1 + ϵ T 1 2 + ϵ log ( T m / n ) ) . (75) J_F\ll\sum_{1\le n\le m\le N}\frac{\lvert r(m)r(n)\rvert}{\sqrt{m/n}}\left(\lvert S_{m,n}\rvert+\left(\frac{m}{n}\right)^{1+\delta}T^{\frac{1}{2}+\delta}+\left(\frac{m}{n}\right)^{1+\epsilon}T^{\frac{1}{2}+\epsilon}\log(Tm/n)\right). \tag*{(75)} J F ≪ 1 ≤ n ≤ m ≤ N ∑ m / n ∣ r ( m ) r ( n )∣ ( ∣ S m , n ∣ + ( n m ) 1 + δ T 2 1 + δ + ( n m ) 1 + ϵ T 2 1 + ϵ log ( T m / n ) ) . (75)
where
S m , n = ∑ T Q 0 m n + 1 ≤ k ≤ 2 T Q 0 m n − 1 a K ( k ) e ( − k n m ∣ D K ∣ ) . S_{m,n}=\sum_{TQ_0\frac{m}{n}+1\leq k\leq2TQ_0\frac{m}{n}-1}a_K(k)e\left(-\frac{kn}{m\sqrt{|D_K|}}\right). S m , n = T Q 0 n m + 1 ≤ k ≤ 2 T Q 0 n m − 1 ∑ a K ( k ) e ( − m ∣ D K ∣ k n ) .
We now divide the proof into two cases.
Case 1. Suppose that either F ( s ) = ζ K ( s ) F(s)=\zeta_K(s) F ( s ) = ζ K ( s ) with K ≠ Q ( i ) K\ne\mathbb{Q}(i) K = Q ( i ) , or F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) F(s)=L(s,\chi_1)L(s,\chi_2) F ( s ) = L ( s , χ 1 ) L ( s , χ 2 ) , where χ 1 \chi_1 χ 1 and χ 2 \chi_2 χ 2 are primitive Dirichlet characters modulo q 1 q_1 q 1 and q 2 q_2 q 2 , respectively, with q 1 q 2 q_1q_2 q 1 q 2 not a perfect square.
In both cases, using Lemma 5.2, we get
S m , n ≪ F m n U log ( 2 m n U ) , where U = Q 0 T m n . S_{m,n}\ll_F\sqrt{mnU}\log(2mnU),\qquad\text{where}\qquad U=Q_0T^{\frac{m}{n}}. S m , n ≪ F mn U log ( 2 mn U ) , where U = Q 0 T n m .
Thus, uniformly for m , n ≤ N m,n\leq N m , n ≤ N , we have
S m , n ≪ m T log ( 2 T N ) . S_{m,n}\ll m\sqrt{T}\log(2TN). S m , n ≪ m T log ( 2 T N ) .
Hence, using (4.20), (5.22) gives
J F ≪ ∑ 1 ≤ n ≤ m ≤ N ∣ r ( m ) r ( n ) ∣ ( T 1 2 ( m n ) 1 2 log ( 2 T N ) + ( m n ) 1 2 + δ T 1 2 + δ + ( m n ) 1 2 + ϵ T 1 2 + ϵ log ( T N ) ) . ≪ ( T 1 2 N 2 log ( 2 T N ) + T 1 2 + δ N 1 + δ ) ∑ n ≤ N ∣ r ( n ) ∣ 2 . (76) \begin{aligned}
J_F\ll\sum_{1\leq n\leq m\leq N}|r(m)r(n)|\left(T^{\frac12}(mn)^{\frac12}\log(2TN)+\left(\frac{m}{n}\right)^{\frac12+\delta}T^{\frac12+\delta}+\left(\frac{m}{n}\right)^{\frac12+\epsilon}T^{\frac12+\epsilon}\log(TN)\right). \\
&\ll\left(T^{\frac12}N^2\log(2TN)+T^{\frac12+\delta}N^{1+\delta}\right)\sum_{n\leq N}|r(n)|^2. \tag*{(76)}
\end{aligned} J F ≪ 1 ≤ n ≤ m ≤ N ∑ ∣ r ( m ) r ( n ) ∣ ( T 2 1 ( mn ) 2 1 log ( 2 T N ) + ( n m ) 2 1 + δ T 2 1 + δ + ( n m ) 2 1 + ϵ T 2 1 + ϵ log ( T N ) ) . ≪ ( T 2 1 N 2 log ( 2 T N ) + T 2 1 + δ N 1 + δ ) n ≤ N ∑ ∣ r ( n ) ∣ 2 . (76)
For N = ⌊ T 1 / 4 ⌋ N=\left\lfloor T^{1/4}\right\rfloor N = ⌊ T 1/4 ⌋ , we obtain
J F ≪ ( T log T + T 3 4 + 5 4 δ ) ∑ n ≤ N ∣ r ( n ) ∣ 2 . J_F\ll\left(T\log T+T^{\frac34+\frac54\delta}\right)\sum_{n\leq N}|r(n)|^2. J F ≪ ( T log T + T 4 3 + 4 5 δ ) n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Case 2: Let F ( s ) = ζ K ( s ) F(s)=\zeta_K(s) F ( s ) = ζ K ( s ) with K = Q ( i ) K=\mathbb{Q}(i) K = Q ( i ) . In this case, using Lemma 5.3, we get
S m , n ≪ m ( m , n ) T m n + m 2 ( m , n ) 2 . S_{m,n}\ll\frac{m}{(m,n)}\sqrt{T^{\frac{m}{n}}}+\frac{m^2}{(m,n)^2}. S m , n ≪ ( m , n ) m T n m + ( m , n ) 2 m 2 .
Proceeding as in Case 1, we obtain
J F ≪ ( T + T 3 4 + T 3 4 + 5 4 δ ) ∑ n ≤ N ∣ r ( n ) ∣ 2 . J_F\ll\left(T+T^{\frac34}+T^{\frac34+\frac54\delta}\right)\sum_{n\leq N}|r(n)|^2. J F ≪ ( T + T 4 3 + T 4 3 + 4 5 δ ) n ≤ N ∑ ∣ r ( n ) ∣ 2 .
In both cases, choosing any fixed 0 < δ < 1 / 5 0<\delta<1/5 0 < δ < 1/5 , we have
J F ≪ T log T ∑ n ≤ N ∣ r ( n ) ∣ 2 . J_F\ll T\log T\sum_{n\leq N}|r(n)|^2. J F ≪ T log T n ≤ N ∑ ∣ r ( n ) ∣ 2 .
This completes the proof. □
Proof of Theorem 1.8. Set N = ⌊ T 1 / 4 ⌋ N=\left\lfloor T^{1/4}\right\rfloor N = ⌊ T 1/4 ⌋ . Let
A ( T ) = exp ( C F log T log log T ) . A(T)=\exp\left(C_F\sqrt{\frac{\log T}{\log\log T}}\right). A ( T ) = exp ( C F log log T log T ) .
Using Lemmas 2.8 and 2.9, we have κ = 2 \kappa=2 κ = 2 in both cases. Since N = ⌊ T 1 / 4 ⌋ N=\left\lfloor T^{1/4}\right\rfloor N = ⌊ T 1/4 ⌋ , we get C F = 1 / 2 − ϵ C_F=1/\sqrt{2}-\epsilon C F = 1/ 2 − ϵ , where ϵ > 0 \epsilon>0 ϵ > 0 is arbitrarily small. By Lemma 5.4 and (23) , we have
J F A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 d t ≪ log T A ( T ) as T → ∞ , \frac{J_F}{A(T)\int_T^{2T}|R(t)|^2\,dt}\ll\frac{\log T}{A(T)}\quad\text{as}\quad T\to\infty, A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 d t J F ≪ A ( T ) log T as T → ∞ ,
it follows that
J F = o ( A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 d t ) . (77) J_F=o\left(A(T)\int_T^{2T}|R(t)|^2\,dt\right). \tag*{(77)} J F = o ( A ( T ) ∫ T 2 T ∣ R ( t ) ∣ 2 d t ) . (77)
Therefore, by Lemma 4.1 and (77) , the hypotheses of Proposition 2.12 are satisfied. This completes the proof.
Beyond Selberg class
In this section, we study large positive and negative values of the Hardy functions associated with certain L L L -functions that need not belong to the Selberg class.
Theorem 6.1. Let F 1 , … , F h F_1,\ldots,F_h F 1 , … , F h be distinct elements of S p o l y S_{\mathrm{poly}} S poly . Write
F j ( s ) = ∑ n = 1 ∞ a j ( n ) n s , 1 ≤ j ≤ h . F_j(s)=\sum_{n=1}^{\infty}\frac{a_j(n)}{n^s},\qquad1\leq j\leq h. F j ( s ) = n = 1 ∑ ∞ n s a j ( n ) , 1 ≤ j ≤ h . Assume strong version of Selberg’s orthonormality condition that
∑ p ≤ x a j ( p ) a ℓ ( p ) ‾ = { ( κ + o ( 1 ) ) x log x , j = ℓ , o ( x log x ) , j ≠ ℓ , (78) \sum_{p\leq x}a_j(p)\overline{a_\ell(p)}=
\begin{cases}
(\kappa+o(1))\dfrac{x}{\log x},&j=\ell,\\
o\left(\dfrac{x}{\log x}\right),&j\ne\ell,
\end{cases} \tag*{(78)} p ≤ x ∑ a j ( p ) a ℓ ( p ) = ⎩ ⎨ ⎧ ( κ + o ( 1 )) log x x , o ( log x x ) , j = ℓ , j = ℓ , (78) where κ > 0 \kappa>0 κ > 0 is fixed. Let β 1 , … , β h \beta_1,\ldots,\beta_h β 1 , … , β h be fixed complex numbers, not all zero, and put
F ( s ) = ∑ j = 1 h β j F j ( s ) . F(s)=\sum_{j=1}^{h}\beta_jF_j(s). F ( s ) = j = 1 ∑ h β j F j ( s ) . Then, for sufficiently large T T T ,
max T ≤ t ≤ 2 T ∣ F ( 1 2 + i t ) ∣ ≫ exp ( ( κ + o ( 1 ) ) log T log log T ) . \max_{T\leq t\leq2T}\left|F\left(\frac{1}{2}+it\right)\right|
\gg\exp\left((\sqrt{\kappa}+o(1))\sqrt{\frac{\log T}{\log\log T}}\right). T ≤ t ≤ 2 T max F ( 2 1 + i t ) ≫ exp ( ( κ + o ( 1 )) log log T log T ) .
Proof. We follow the method of Aistleitner and Pańkowski [1 Section 2] , and explain the modifications required for the linear combination.
Since the coefficients β 1 , … , β h \beta_1,\ldots,\beta_h β 1 , … , β h are not all zero, we may assume, without loss of generality, that β 1 ≠ 0 \beta_1\ne0 β 1 = 0 . We choose r ( n ) = a 1 ( n ) f ( n ) r(n)=a_1(n)f(n) r ( n ) = a 1 ( n ) f ( n ) , where f f f is defined in (18) . Since F ( s ) = ∑ j = 1 h β j F j ( s ) F(s)=\sum_{j=1}^{h}\beta_jF_j(s) F ( s ) = ∑ j = 1 h β j F j ( s ) , we have
I F = ∑ j = 1 h β j ∫ R F j ( 1 2 + i t ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t . I_F=\sum_{j=1}^{h}\beta_j\int_{\mathbb{R}}F_j\left(\frac{1}{2}+it\right)|R(t)|^2\Phi(t/T)\,dt. I F = j = 1 ∑ h β j ∫ R F j ( 2 1 + i t ) ∣ R ( t ) ∣ 2 Φ ( t / T ) d t .
Proceeding as in [1 Section 2] , we obtain
I F = T Φ ^ ( 0 ) ∑ j = 1 h β j M j + O ( T ∑ n ≤ N ∣ r ( n ) ∣ 2 ) , I_F=T\widehat{\Phi}(0)\sum_{j=1}^{h}\beta_jM_j+O\left(T\sum_{n\leq N}|r(n)|^2\right), I F = T Φ ( 0 ) j = 1 ∑ h β j M j + O ( T n ≤ N ∑ ∣ r ( n ) ∣ 2 ) ,
where
M j = ∑ n k ≤ N a j ( k ) r ( n ) r ( n k ) ‾ k e − k / X \mathcal{M}_j=\sum_{nk\le N}\frac{a_j(k)r(n)\overline{r(nk)}}{\sqrt{k}}e^{-k/X} M j = nk ≤ N ∑ k a j ( k ) r ( n ) r ( nk ) e − k / X
with X = T d + ϵ X=T^{d+\epsilon} X = T d + ϵ , where d = max 1 ≤ j ≤ h d F j d=\max_{1\le j\le h}d_{F_j} d = max 1 ≤ j ≤ h d F j and ϵ > 0 \epsilon>0 ϵ > 0 is small. Since f f f is supported on squarefree integers, the nonzero summands satisfy ( n , k ) = 1 (n,k)=1 ( n , k ) = 1 , and hence
a j ( k ) r ( n ) r ( n k ) ‾ = a j ( k ) a 1 ( k ) ‾ ∣ a 1 ( n ) ∣ 2 f ( n ) 2 f ( k ) . a_j(k)r(n)\overline{r(nk)}=a_j(k)\overline{a_1(k)}|a_1(n)|^2f(n)^2f(k). a j ( k ) r ( n ) r ( nk ) = a j ( k ) a 1 ( k ) ∣ a 1 ( n ) ∣ 2 f ( n ) 2 f ( k ) .
Thus, unlike the single L L L -function case, we need to estimate both the selected contribution M 1 \mathcal{M}_1 M 1 and the cross contributions M j \mathcal{M}_j M j , for j ≠ 1 j\ne1 j = 1 .
Put
D ∞ = ∏ p ( 1 + u p ) , P j = ∏ p ( 1 + u p + v j , p ) , D_\infty=\prod_p(1+u_p),\qquad P_j=\prod_p(1+u_p+v_{j,p}), D ∞ = p ∏ ( 1 + u p ) , P j = p ∏ ( 1 + u p + v j , p ) ,
where
u p = ∣ a 1 ( p ) ∣ 2 f ( p ) 2 , v j , p = a j ( p ) a 1 ( p ) ‾ f ( p ) p . u_p=|a_1(p)|^2f(p)^2,\qquad v_{j,p}=\frac{a_j(p)\overline{a_1(p)}f(p)}{\sqrt{p}}. u p = ∣ a 1 ( p ) ∣ 2 f ( p ) 2 , v j , p = p a j ( p ) a 1 ( p ) f ( p ) .
By using (6.1) and partial summation,
∑ L 2 ≤ p ≤ M a j ( p ) a 1 ( p ) ‾ p log p = { κ + o ( 1 ) 2 log L , j = 1 , o ( 1 log L ) , j ≠ 1. \sum_{L^2\le p\le M}\frac{a_j(p)\overline{a_1(p)}}{p\log p}=
\begin{cases}
\dfrac{\kappa+o(1)}{2\log L}, & j=1,\\[6pt]
o\left(\dfrac{1}{\log L}\right), & j\ne1.
\end{cases} L 2 ≤ p ≤ M ∑ p log p a j ( p ) a 1 ( p ) = ⎩ ⎨ ⎧ 2 log L κ + o ( 1 ) , o ( log L 1 ) , j = 1 , j = 1.
Consequently, the Euler-product calculation gives
P 1 D ∞ = exp ( ( κ ( 1 − η ) + o ( 1 ) ) log N log log N ) , \frac{P_1}{D_\infty}=\exp\left(\left(\sqrt{\kappa(1-\eta)}+o(1)\right)\sqrt{\frac{\log N}{\log\log N}}\right), D ∞ P 1 = exp ( ( κ ( 1 − η ) + o ( 1 ) ) log log N log N ) ,
whereas
∣ P j ∣ D ∞ = exp ( o ( log N log log N ) ) , 2 ≤ j ≤ h . \frac{|P_j|}{D_\infty}=\exp\left(o\left(\sqrt{\frac{\log N}{\log\log N}}\right)\right),\qquad2\le j\le h. D ∞ ∣ P j ∣ = exp ( o ( log log N log N ) ) , 2 ≤ j ≤ h .
Applying the same Rankin truncation argument as [1 ] , with absolute values for the cross terms, yields M j = P j + o ( D ∞ ) \mathcal{M}_j=P_j+o(D_\infty) M j = P j + o ( D ∞ ) and ∑ n ≤ N ∣ r ( n ) ∣ 2 = ( 1 + o ( 1 ) ) D ∞ \sum_{n\le N}|r(n)|^2=(1+o(1))D_\infty ∑ n ≤ N ∣ r ( n ) ∣ 2 = ( 1 + o ( 1 )) D ∞ . Thus M 1 \mathcal{M}_1 M 1 is positive and dominates all the cross contributions. Since β 1 ≠ 0 \beta_1\ne0 β 1 = 0 , we obtain
∣ I F ∣ ≥ T Φ ^ ( 0 ) ( ∣ β 1 ∣ M 1 − ∑ j = 2 h ∣ β j ∣ ∣ M j ∣ ) − o ( T ∑ n ≤ N ∣ r ( n ) ∣ 2 ) ≫ T exp ( ( κ ( 1 − η ) + o ( 1 ) ) log N log log N ) ∑ n ≤ N ∣ r ( n ) ∣ 2 . \begin{aligned}
|I_F|\ge T\widehat{\Phi}(0)\left(|\beta_1|\mathcal{M}_1-\sum_{j=2}^h|\beta_j||\mathcal{M}_j|\right)-o\left(T\sum_{n\le N}|r(n)|^2\right) \\
&\gg T\exp\left(\left(\sqrt{\kappa(1-\eta)}+o(1)\right)\sqrt{\frac{\log N}{\log\log N}}\right)\sum_{n\le N}|r(n)|^2.
\end{aligned} ∣ I F ∣ ≥ T Φ ( 0 ) ( ∣ β 1 ∣ M 1 − j = 2 ∑ h ∣ β j ∣∣ M j ∣ ) − o ( T n ≤ N ∑ ∣ r ( n ) ∣ 2 ) ≫ T exp ( ( κ ( 1 − η ) + o ( 1 ) ) log log N log N ) n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Hence, the inequality
max T ≤ t ≤ 2 T ∣ F ( 1 2 + i t ) ∣ ≥ ∣ I F ∣ ∫ R ∣ R ( t ) ∣ 2 Φ ( t / T ) d t \max_{T\le t\le2T}\left|F\left(\frac{1}{2}+it\right)\right|\ge\frac{|I_F|}{\int_{\mathbb{R}}|R(t)|^2\Phi(t/T)\,dt} T ≤ t ≤ 2 T max F ( 2 1 + i t ) ≥ ∫ R ∣ R ( t ) ∣ 2 Φ ( t / T ) d t ∣ I F ∣
give
max T ≤ t ≤ 2 T ∣ F ( 1 2 + i t ) ∣ ≫ exp ( ( κ α ( 1 − η ) + o ( 1 ) ) log T log log T ) . \max_{T\le t\le2T}\left|F\left(\frac{1}{2}+it\right)\right|\gg\exp\left(\left(\sqrt{\kappa\alpha(1-\eta)}+o(1)\right)\sqrt{\frac{\log T}{\log\log T}}\right). T ≤ t ≤ 2 T max F ( 2 1 + i t ) ≫ exp ( ( κ α ( 1 − η ) + o ( 1 ) ) log log T log T ) .
Choosing α < 1 \alpha< 1 α < 1 sufficiently close to 1 1 1 and η > 0 \eta> 0 η > 0 sufficiently small completes the proof. □
Theorem 6.2. Let F ( s ) F(s) F ( s ) be as in Theorem 6.1 . Suppose further that all the constituent L L L -functions F j F_j F j have the same degree d ∈ { 1 , 2 } d \in\{1,2\} d ∈ { 1 , 2 } and satisfy the functional equations
F j ( s ) = ω j H 0 ( s ) F j ( 1 − s ‾ ) ‾ , 1 ≤ j ≤ h , F_j(s)=\omega_j H_0(s)\overline{F_j(1-\overline{s})}, \qquad1 \leq j \leq h, F j ( s ) = ω j H 0 ( s ) F j ( 1 − s ) , 1 ≤ j ≤ h , where H 0 ( s ) H_0(s) H 0 ( s ) is common to all the L L L -functions F j F_j F j and ∣ ω j ∣ = 1 |\omega_j|=1 ∣ ω j ∣ = 1 . Assume that
β j ω j = ω β j ‾ , 1 ≤ j ≤ h , (79) \beta_j\omega_j=\omega\overline{\beta_j}, \qquad1 \leq j \leq h, \tag*{(79)} β j ω j = ω β j , 1 ≤ j ≤ h , (79) for some ω ∈ C \omega\in\mathbb{C} ω ∈ C with ∣ ω ∣ = 1 |\omega|=1 ∣ ω ∣ = 1 . If d = 2 d=2 d = 2 , suppose additionally that ℑ ξ F j = 0 \Im\xi_{F_j}=0 ℑ ξ F j = 0 and that each F j F_j F j satisfies Hypothesis 1, for 1 ≤ j ≤ h 1 \leq j \leq h 1 ≤ j ≤ h .
Let Z F ( t ) Z_F(t) Z F ( t ) denote the corresponding Hardy function. Then, for every ϵ > 0 \epsilon>0 ϵ > 0 and all sufficiently large T T T ,
max T ≤ t ≤ 2 T Z F ± ( t ) ≫ exp ( ( C F − ϵ ) log T log log T ) . \max_{T\leq t\leq2T} Z_F^{\pm}(t) \gg\exp\left((C_F-\epsilon)\sqrt{\frac{\log T}{\log\log T}}\right). T ≤ t ≤ 2 T max Z F ± ( t ) ≫ exp ( ( C F − ϵ ) log log T log T ) . where
C F = { 1 2 , d = 1 , κ ( 1 − γ ) , d = 2 , C_F=
\begin{cases}
\frac{1}{2}, & d=1,\\
\sqrt{\kappa(1-\gamma)}, & d=2,
\end{cases} C F = { 2 1 , κ ( 1 − γ ) , d = 1 , d = 2 , with κ > 0 \kappa>0 κ > 0 as in Theorem 6.1 and γ \gamma γ as in Hypothesis 1.
Proof. By the functional equations and the condition (79) , we obtain
F ( s ) = ω H 0 ( s ) F ( 1 − s ‾ ) ‾ . F(s)=\omega H_0(s)\overline{F(1-\overline{s})}. F ( s ) = ω H 0 ( s ) F ( 1 − s ) .
Consequently, the corresponding Hardy function
Z F ( t ) = ( ω H 0 ( 1 2 + i t ) ) − 1 / 2 F ( 1 2 + i t ) Z_F(t)=\left(\omega H_0\left(\frac{1}{2}+it\right)\right)^{-1/2}F\left(\frac{1}{2}+it\right) Z F ( t ) = ( ω H 0 ( 2 1 + i t ) ) − 1/2 F ( 2 1 + i t )
is real-valued, and ∣ Z F ( t ) ∣ = ∣ F ( 1 2 + i t ) ∣ |Z_F(t)|=\left|F\left(\frac{1}{2}+it\right)\right| ∣ Z F ( t ) ∣ = F ( 2 1 + i t ) .
Choose N ≤ T α N\leq T^\alpha N ≤ T α , with α < 1 \alpha<1 α < 1 , and the same resonator R ( t ) R(t) R ( t ) as in the proof of Theorem 6.1 . By the argument given there, we have
I F ≫ T exp ( ( κ ( 1 − η ) + o ( 1 ) ) log N log log N ) ∑ n ≤ N ∣ r ( n ) ∣ 2 . I_F\gg T\exp\left((\sqrt{\kappa(1-\eta)}+o(1))\sqrt{\frac{\log N}{\log\log N}}\right)\sum_{n\leq N}|r(n)|^2. I F ≫ T exp ( ( κ ( 1 − η ) + o ( 1 )) log log N log N ) n ≤ N ∑ ∣ r ( n ) ∣ 2 .
Moreover, since F F F is a finite linear combination of the L L L -functions F j F_j F j , the required upper bound follows by applying the corresponding estimates to each L L L -functions F j F_j F j . If d = 1 d=1 d = 1 , we proceed as in the proof of Theorem 1.2 . If d = 2 d=2 d = 2 , we proceed as in the proof of Theorem 1.3 . This completes the proof.
Remark 6.3. An example in degree 1 is given by linear combinations of distinct primitive Dirichlet L L L -functions having the same conductor and parity, since their functional-equation factors coincide. In this case, the conclusion of Theorem 6.2 holds with C F = 1 / 2 C_F=1/2 C F = 1/2 , provided the coefficients satisfy (79) .
Similarly, in degree 2, we may consider linear combinations of L L L -functions associated with distinct normalized primitive holomorphic Hecke eigenforms of the same weight and level, whose nebentypus characters need not be the same. In this case, the conclusion holds with C F = 1 / 2 C_F=1/\sqrt{2} C F = 1/ 2 , provided the coefficients satisfy (79) .
Example 6.4. Consider
g ( z ) = Δ ( z ) 2 g(z) = \Delta(z)^2 g ( z ) = Δ ( z ) 2 where Δ ( z ) \Delta(z) Δ ( z ) is the Ramanujan cusp form of weight 12. Then g ∈ S 24 ( SL 2 ( Z ) ) g \in S_{24}(\operatorname{SL}_2(\mathbb{Z})) g ∈ S 24 ( SL 2 ( Z )) , and its first Fourier coefficient is zero. Since this space has dimension two, we can write
g = c ( f 1 − f 2 ) , g = c(f_1-f_2), g = c ( f 1 − f 2 ) , where c c c is a nonzero real constant and f 1 , f 2 f_1, f_2 f 1 , f 2 are distinct normalized primitive holomorphic Hecke eigenforms of weight 24 and level one. Both associated L L L -functions have root number + 1 +1 + 1 . Thus, the condition β j ω j = ω β j \beta_j\omega_j=\omega\beta_j β j ω j = ω β j is satisfied with β 1 = c \beta_1=c β 1 = c , β 2 = − c \beta_2=-c β 2 = − c , and ω = 1 \omega=1 ω = 1 . Consequently,
F g ( s ) = L ( s + 23 2 , g ) = c ( F 1 ( s ) − F 2 ( s ) ) , F_g(s)=L\left(s+\frac{23}{2},g\right)=c(F_1(s)-F_2(s)), F g ( s ) = L ( s + 2 23 , g ) = c ( F 1 ( s ) − F 2 ( s )) , where F 1 F_1 F 1 and F 2 F_2 F 2 are the corresponding normalized L L L -functions. Therefore, Theorem 6.2 applies to F g F_g F g with C F = 1 / 2 C_F=1/\sqrt{2} C F = 1/ 2 .
This provides an example of a Dirichlet series satisfying a degree-two functional equation and the conclusion of Theorem 6.2 , but not belonging to the Selberg class, since its first Dirichlet coefficient is zero and its coefficients are not multiplicative.
Acknowledgments
A.M. Hossain is supported by the Institute Fellowship from IIT Kharagpur, India. K. Mahatab is supported by the ARG-MATRICS Grant (Grant No. ANRF/ARGM/2025/002540/MTR). Ahmad Mustafa Hossain, Department of Mathematics, Indian Institute of Technology Kharagpur, Kharagpur-721302, India. Email address: mostafa.24@kgpian.iitkgp.ac.in
Kamalakshya Mahatab, Department of Mathematics, Indian Institute of Technology Kharagpur, Kharagpur-721302, India. Email address: kamalakshya@maths.iitkgp.ac.in
Akshaa Vatwani, Department of Mathematics, Indian Institute of Technology Gandhinagar,, Palaj, Gandhinagar Gujarat-382355, India. Email address: akshaa.vatwani@iitgn.ac.in