Introduction
Let χ \chi χ (mod q q q ) be a primitive Dirichlet character. In this paper, we investigate what appears to be a deep connection between the distribution of the nontrivial zeros of the Dirichlet L L L -function L ( s , χ ) L(s,\chi) L ( s , χ ) , and the Gauss-type sum
S ( χ ) = ∑ n = 1 2 q χ ( n ) e ( − n 2 2 q ) , (1) S(\chi)=\sum_{n=1}^{2q}\chi(n)e\left(-\frac{n^2}{2q}\right), \tag*{(1)} S ( χ ) = n = 1 ∑ 2 q χ ( n ) e ( − 2 q n 2 ) , (1)
where as usual we let e ( x ) = e 2 π i x e(x)=e^{2\pi i x} e ( x ) = e 2 π i x . Notice that χ \chi χ is periodic (mod q q q ) and the additive character is periodic (mod 2 q 2q 2 q ), so S ( χ ) S(\chi) S ( χ ) is a quadratic Gauss sum (mod 2 q 2q 2 q ) twisted by an imprimitive character. In a natural sense, we prove that S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 for 100% of primitive characters. However, we prove that there is an infinite (but zero-density) family of characters such that S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 . Proposition 1.7 gives a complete classification of this thin family of Dirichlet characters.
In Theorem 1.1 , we show that the integral of the Z Z Z -function, Z χ ( t ) Z_\chi(t) Z χ ( t ) , associated to the character χ \chi χ has a main term when S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 that is not present if S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 . Moreover, when S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 it appears that there is a significant and previously unobserved bias in the gaps between the nontrivial zeros of L ( s , χ ) L(s,\chi) L ( s , χ ) . Experimentally, we give numerous examples of this newly discovered bias.
The Hardy Z-function
The sum S ( χ ) S(\chi) S ( χ ) naturally arises when considering the average of Z χ ( t ) Z_\chi(t) Z χ ( t ) . To describe this result, let us first recall how to define the Hardy Z Z Z -function. The functional equation for the Riemann zeta-function, ζ ( s ) \zeta(s) ζ ( s ) , is given by
Γ R ( s ) ζ ( s ) = Γ R ( 1 − s ) ζ ( 1 − s ) . \Gamma_{\mathbb{R}}(s)\zeta(s)=\Gamma_{\mathbb{R}}(1-s)\zeta(1-s). Γ R ( s ) ζ ( s ) = Γ R ( 1 − s ) ζ ( 1 − s ) .
where Γ R ( s ) = π − s / 2 Γ ( s / 2 ) \Gamma_{\mathbb{R}}(s)=\pi^{-s/2}\Gamma(s/2) Γ R ( s ) = π − s /2 Γ ( s /2 ) . The Hardy Z Z Z -function is defined by
Z ( t ) = ( Γ R ( 1 2 + i t ) Γ R ( 1 2 − i t ) ) 1 / 2 ζ ( 1 2 + i t ) . Z(t)=\left(\frac{\Gamma_{\mathbb{R}}\left(\frac{1}{2}+it\right)}{\Gamma_{\mathbb{R}}\left(\frac{1}{2}-it\right)}\right)^{1/2}\zeta\left(\frac{1}{2}+it\right). Z ( t ) = ( Γ R ( 2 1 − i t ) Γ R ( 2 1 + i t ) ) 1/2 ζ ( 2 1 + i t ) .
By the functional equation, we observe that Z ( t ) ‾ = Z ( t ) \overline{Z(t)}=Z(t) Z ( t ) = Z ( t ) , so Z ( t ) Z(t) Z ( t ) is real-valued for real t t t . Moreover, ∣ Z ( t ) ∣ = ∣ ζ ( 1 2 + i t ) ∣ |Z(t)|=\left|\zeta\left(\frac{1}{2}+it\right)\right| ∣ Z ( t ) ∣ = ζ ( 2 1 + i t ) , so the zeros of Z ( t ) Z(t) Z ( t ) correspond to zeros of ζ ( s ) \zeta(s) ζ ( s ) on the critical line, Re ( s ) = 1 / 2 \operatorname{Re}(s)=1/2 Re ( s ) = 1/2 , and Z ( t ) Z(t) Z ( t ) changes sign when t t t is an ordinate of a zero of ζ ( s ) \zeta(s) ζ ( s ) on the critical line of odd multiplicity. This makes Z ( t ) Z(t) Z ( t ) useful for understanding the distribution of the nontrivial zeros of ζ ( s ) \zeta(s) ζ ( s ) . For example, if we define the integral
F ( T ) = ∫ 0 T Z ( t ) d t , (2) F(T)=\int_0^T Z(t)\,dt, \tag*{(2)} F ( T ) = ∫ 0 T Z ( t ) d t , (2)
then the estimate F ( T ) = o ( T ) F(T)=o(T) F ( T ) = o ( T ) , as T → ∞ T\to\infty T → ∞ , implies Hardy’s theorem that infinitely many nontrivial zeros of the zeta function are on the critical line (see [33 Section 10.5] ).
One might expect a great deal of cancellation in the integral F ( T ) F(T) F ( T ) , as T → ∞ T\to\infty T → ∞ , since Z ( t ) Z(t) Z ( t ) oscillates in sign. Indeed, Titchmarsh [33 Eq. 10.5.1] proved that F ( T ) = O ( T 7 / 8 ) F(T)=O(T^{7/8}) F ( T ) = O ( T 7/8 ) . Ivić [16 ] dramatically improved this by showing that F ( T ) = O ε ( T 1 / 4 + ε ) F(T)=O_{\varepsilon}(T^{1/4+\varepsilon}) F ( T ) = O ε ( T 1/4 + ε ) for all ε > 0 \varepsilon>0 ε > 0 . In the same paper, Ivić conjectured that in fact
F ( T ) = O ( T 1 / 4 ) and F ( T ) = Ω ± ( T 1 / 4 ) . (3) F(T)=O(T^{1/4})\quad\text{and}\quad F(T)=\Omega_{\pm}(T^{1/4}). \tag*{(3)} F ( T ) = O ( T 1/4 ) and F ( T ) = Ω ± ( T 1/4 ) . (3)
Ivić’s conjecture has been independently established by Korolev [25 , 26 ] and Jutila [20 , 21 ] in a rather precise way. Let θ 0 = min ( ∣ { T / 2 π } − 1 4 ∣ , ∣ { T / 2 π } − 3 4 ∣ ) \theta_0=\min\left(\left|\{\sqrt{T/2\pi}\}-\frac{1}{4}\right|,\left|\{\sqrt{T/2\pi}\}-\frac{3}{4}\right|\right) θ 0 = min ( { T /2 π } − 4 1 , { T /2 π } − 4 3 ) . Jutila [21 ] proved that if θ 0 ≠ 0 \theta_0\ne0 θ 0 = 0 , then
∫ 0 T Z ( t ) d t = ( T 2 π ) 1 / 4 ( − 1 ) ⌊ T / 2 π ⌋ K ( { T / 2 π } ) + O ( T 1 / 6 log T ) + O ( min { T 1 / 4 , T 1 / 8 θ 0 − 3 / 4 } ) . \int_0^T Z(t)\,dt=\left(\frac{T}{2\pi}\right)^{1/4}(-1)^{\lfloor\sqrt{T/2\pi}\rfloor}K\left(\{\sqrt{T/2\pi}\}\right)+O(T^{1/6}\log T)+O\left(\min\{T^{1/4},T^{1/8}\theta_0^{-3/4}\}\right). ∫ 0 T Z ( t ) d t = ( 2 π T ) 1/4 ( − 1 ) ⌊ T /2 π ⌋ K ( { T /2 π } ) + O ( T 1/6 log T ) + O ( min { T 1/4 , T 1/8 θ 0 − 3/4 } ) .
where K ( x ) = 2 π K(x)=2\pi K ( x ) = 2 π for 1 4 < x < 3 4 \frac{1}{4}<x<\frac{3}{4} 4 1 < x < 4 3 , and K ( x ) = 0 K(x)=0 K ( x ) = 0 for 0 ≤ x < 1 4 0\le x<\frac{1}{4} 0 ≤ x < 4 1 and for 3 4 < x < 1 \frac{3}{4}<x<1 4 3 < x < 1 (see Figure 1 ). He also proved that if θ 0 = 0 \theta_0=0 θ 0 = 0 , then the same formula holds with error term O ( T 1 / 6 log T ) O(T^{1/6}\log T) O ( T 1/6 log T ) , where K ( 1 4 ) = 4 π 3 K\left(\frac{1}{4}\right)=\frac{4\pi}{3} K ( 4 1 ) = 3 4 π and K ( 3 4 ) = 2 π 3 K\left(\frac{3}{4}\right)=\frac{2\pi}{3} K ( 4 3 ) = 3 2 π .
Figure 1. A plot of 1 2 π F F ( 2 π T 2 ) \frac{1}{2\pi\sqrt{F}}F(2\pi T^2) 2 π F 1 F ( 2 π T 2 ) for the function F ( T ) F(T) F ( T ) defined in (2) .
The Z-function associated to a Dirichlet L L L -function
Let χ \chi χ (mod q q q ) be a primitive Dirichlet character. Then L ( s , χ ) L(s,\chi) L ( s , χ ) satisfies the functional equation, in asymmetric form,
L ( s , χ ) = G χ ( s ) L ( 1 − s , χ ‾ ) , where G χ ( s ) = ε χ q 1 / 2 − s Γ R ( 1 − s + a ) Γ R ( s + a ) . L(s,\chi)=\mathcal{G}_{\chi}(s)L(1-s,\overline{\chi}),\qquad\text{where}\qquad\mathcal{G}_{\chi}(s)=\varepsilon_{\chi}q^{1/2-s}\frac{\Gamma_{\mathbb{R}}(1-s+a)}{\Gamma_{\mathbb{R}}(s+a)}. L ( s , χ ) = G χ ( s ) L ( 1 − s , χ ) , where G χ ( s ) = ε χ q 1/2 − s Γ R ( s + a ) Γ R ( 1 − s + a ) .
Here, a ∈ { 0 , 1 } a\in\{0,1\} a ∈ { 0 , 1 } is such that χ ( − 1 ) = ( − 1 ) a \chi(-1)=(-1)^a χ ( − 1 ) = ( − 1 ) a and ε χ = i − a τ ( χ ) / q \varepsilon_{\chi}=i^{-a}\tau(\chi)/\sqrt{q} ε χ = i − a τ ( χ ) / q , where
τ ( χ ) = ∑ n = 1 q χ ( n ) e ( n q ) \tau(\chi)=\sum_{n=1}^{q}\chi(n)e\left(\frac{n}{q}\right) τ ( χ ) = n = 1 ∑ q χ ( n ) e ( q n )
is the classical Gauss sum satisfying ∣ τ ( χ ) ∣ = q |\tau(\chi)|=\sqrt{q} ∣ τ ( χ ) ∣ = q , so that ∣ ε χ ∣ = 1 |\varepsilon_{\chi}|=1 ∣ ε χ ∣ = 1 . In analogy with the definition of Z ( t ) Z(t) Z ( t ) for the Riemann zeta-function, we define
Z χ ( t ) = G χ ( 1 2 + i t ) − 1 / 2 L ( 1 2 + i t , χ ) . Z_{\chi}(t)=\mathcal{G}_{\chi}\left(\frac{1}{2}+it\right)^{-1/2}L\left(\frac{1}{2}+it,\chi\right). Z χ ( t ) = G χ ( 2 1 + i t ) − 1/2 L ( 2 1 + i t , χ ) .
By the functional equation, it again follows that Z χ ( t ) ‾ = Z χ ( t ) \overline{Z_{\chi}(t)}=Z_{\chi}(t) Z χ ( t ) = Z χ ( t ) , so Z χ ( t ) Z_{\chi}(t) Z χ ( t ) is real-valued for real t t t . Moreover, ∣ Z χ ( t ) ∣ = ∣ L ( 1 2 + i t , χ ) ∣ |Z_{\chi}(t)|=|L(\frac{1}{2}+it,\chi)| ∣ Z χ ( t ) ∣ = ∣ L ( 2 1 + i t , χ ) ∣ , so the zeros of Z χ ( t ) Z_{\chi}(t) Z χ ( t ) correspond to zeros of L ( s , χ ) L(s,\chi) L ( s , χ ) on the critical line and Z χ ( t ) Z_{\chi}(t) Z χ ( t ) changes sign when t t t is an ordinate of a zero of L ( s , χ ) L(s,\chi) L ( s , χ ) on the critical line of odd multiplicity. A well-known conjecture of Chowla [4 ] is that L ( 1 2 , χ ) ≠ 0 L(\frac{1}{2},\chi)\ne0 L ( 2 1 , χ ) = 0 when χ \chi χ is a primitive quadratic character. It is now generally believed that L ( 1 2 , χ ) ≠ 0 L(\frac{1}{2},\chi)\ne0 L ( 2 1 , χ ) = 0 for all primitive characters χ \chi χ , which would imply that Z χ ( 0 ) ≠ 0 Z_{\chi}(0)\ne0 Z χ ( 0 ) = 0 . When Z χ ( 0 ) ≠ 0 Z_{\chi}(0)\ne0 Z χ ( 0 ) = 0 , we choose a branch of G χ ( s ) − 1 / 2 \mathcal{G}_{\chi}(s)^{-1/2} G χ ( s ) − 1/2 so that Z χ ( 0 ) > 0 Z_{\chi}(0)>0 Z χ ( 0 ) > 0 . For quadratic characters, this means that Z χ ( 0 ) = L ( 1 / 2 , χ ) > 0 Z_{\chi}(0)=L(1/2,\chi)>0 Z χ ( 0 ) = L ( 1/2 , χ ) > 0 assuming Chowla’s conjecture and GRH. The usual square root for the Hardy Z Z Z -function is chosen so that Z ( 0 ) = ζ ( 1 / 2 ) < 0 Z(0)=\zeta(1/2)<0 Z ( 0 ) = ζ ( 1/2 ) < 0 . In both cases, the value of the Z Z Z -function at 0 is chosen to agree with the central value of the L L L -function.
Work of Rudnick and Sarnak [30 ] suggests that n n n -level correlations of the ordinates of the nontrivial zeros of an L L L -function associated to any cuspidal automorphic representation of G L ( m ) \mathrm{GL}(m) GL ( m ) over Q \mathbb{Q} Q should correspond to n n n -level correlations of eigenvalues of random matrices in the Gaussian unitary ensemble (GUE). This led Keating and Snaith [23 , 24 ] to suggest that the value distribution of any principal automorphic L L L -function on its critical line is related to that of the characteristic polynomials of random unitary matrices. An explicit description of this correspondence in terms of Z Z Z -functions is given by Conrey, Farmer, Keating, Rubinstein, and Snaith in [5 Section 1] . Hence, using the random matrix model for L L L -functions, one might expect an analogue of Ivić’s conjecture for F ( T ) F(T) F ( T ) in (3) to hold for the integral average of Z χ ( t ) Z_{\chi}(t) Z χ ( t ) for any primitive character χ \chi χ . In other words, is it true that
∫ 0 T Z χ ( t ) d t = O q ( T 1 / 4 ) and ∫ 0 T Z χ ( t ) d t = Ω ± ( T 1 / 4 ) ? \int_0^T Z_{\chi}(t)\,\mathrm{d}t=O_q(T^{1/4})\qquad\text{and}\qquad\int_0^T Z_{\chi}(t)\,\mathrm{d}t=\Omega_{\pm}(T^{1/4})? ∫ 0 T Z χ ( t ) d t = O q ( T 1/4 ) and ∫ 0 T Z χ ( t ) d t = Ω ± ( T 1/4 )?
When S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 , the first estimate fails (and the absolute value of the integral is ≫ q T 3 / 4 \gg_q T^{3/4} ≫ q T 3/4 ).
Theorem 1.1. Let χ \chi χ be a primitive Dirichlet character (mod q q q ). As T → ∞ T\to\infty T → ∞ , we have
∫ 0 T Z χ ( t ) d t = C χ T 3 / 4 + O q ( T 1 / 4 log T ) , (4) \int_0^T Z_{\chi}(t)\,\mathrm{d}t=C_{\chi}T^{3/4}+O_q(T^{1/4}\log T), \tag*{(4)} ∫ 0 T Z χ ( t ) d t = C χ T 3/4 + O q ( T 1/4 log T ) , (4) where
C χ = 2 3 / 4 π 1 / 4 e ( 1 / 16 ) 3 q 1 / 2 τ ( χ ) 1 / 2 S ( χ ) (5) C_{\chi}=\frac{2^{3/4}\pi^{1/4}e(1/16)}{3q^{1/2}\tau(\chi)^{1/2}}S(\chi) \tag*{(5)} C χ = 3 q 1/2 τ ( χ ) 1/2 2 3/4 π 1/4 e ( 1/16 ) S ( χ ) (5) and S ( χ ) S(\chi) S ( χ ) is defined in (1) . Moreover, there are infinitely many χ \chi χ for which $S(\chi)\ne 0.
In Theorem 1.1, τ ( χ ) 1 / 2 \tau(\chi)^{1/2} τ ( χ ) 1/2 denotes the square root determined by the branch of G χ ( s ) − 1 / 2 G_\chi(s)^{-1/2} G χ ( s ) − 1/2 used to define Z χ ( t ) Z_\chi(t) Z χ ( t ) . In particular, if L ( 1 2 , χ ) ≠ 0 L(\frac{1}{2},\chi) \ne0 L ( 2 1 , χ ) = 0 , then it is the square root for which Z χ ( 0 ) > 0 Z_\chi(0) > 0 Z χ ( 0 ) > 0 .
(i) Since Z χ ( t ) Z_\chi(t) Z χ ( t ) is real-valued, we see that the constant C χ C_\chi C χ defined in Theorem 1.1 is real. This is not immediate from the definition of C χ C_\chi C χ and we give a direct proof of this fact in Appendix A.
(ii) Notice that C χ ≠ 0 C_\chi\ne0 C χ = 0 if and only if S ( χ ) ≠ 0 S(\chi) \ne0 S ( χ ) = 0 . In Section 5, we give another evaluation of C χ C_\chi C χ in the case where it is nonzero, and speculate on its distribution.
(iii) We classify the primitive characters χ \chi χ for which S ( χ ) ≠ 0 S(\chi) \ne0 S ( χ ) = 0 in Proposition 1.7 and then use this classification to count the number of such χ \chi χ in Theorem 1.5 and Corollary 1.6.
(iv) We have not tried to optimize the error term in (1.4). Instead we keep our proof relatively simple and we focus our attention on classifying the characters for which S ( χ ) ≠ 0 S(\chi) \ne0 S ( χ ) = 0 . With more work, we can show that the error term in (4) is O q ( T 1 / 4 ) O_q(T^{1/4}) O q ( T 1/4 ) and Ω ± ( T 1 / 4 ) \Omega_{\pm}(T^{1/4}) Ω ± ( T 1/4 ) , as discussed at the end of this subsection.
(v) There are other situations where averages of L L L -functions over families have lower-order main terms that have not been predicted using random matrix theory models. For example, the work of Diaconu [6 ] and Diaconu and Whitehead [7 ] has shown that there is a lower-order main term in the third moment of quadratic Dirichlet L L L -functions over both number fields and function fields. Also, Florea [10 ] has shown that there is a lower-order main term in the first moment of quadratic Dirichlet L L L -functions over function fields (see also the recent work of Čech [2 ] and of Gao and Zhao [11 ] ). Similar lower-order main terms appear in averages of Hecke L L L -functions attached to ℓ \ell ℓ th order residue symbols. For example, see Hamdar [15 Remark 1.3 and Conjecture 1.5] .
Example 1.2. Let χ 8 \chi_8 χ 8 and χ − 8 \chi_{-8} χ − 8 be the even and odd primitive characters (mod 8), respectively. Then S ( χ 8 ) = 8 e ( − 1 / 16 ) S(\chi_8)=8e(-1/16) S ( χ 8 ) = 8 e ( − 1/16 ) and S ( χ − 8 ) = 0 S(\chi_{-8})=0 S ( χ − 8 ) = 0 . The character χ 8 \chi_8 χ 8 (mod 8) is the character of smallest modulus for which S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 . For these characters, Theorem 1.1 gives
∫ 0 T Z χ 8 ( t ) d t = 2 3 / 2 π 1 / 4 3 T 3 / 4 + O ( T 1 / 4 log T ) and ∫ 0 T Z χ − 8 ( t ) d t = O ( T 1 / 4 log T ) . \int_0^T Z_{\chi_8}(t)\,\mathrm{d}t=\frac{2^{3/2}\pi^{1/4}}{3}T^{3/4}+O(T^{1/4}\log T)\quad\text{and}\quad\int_0^T Z_{\chi_{-8}}(t)\,\mathrm{d}t=O(T^{1/4}\log T). ∫ 0 T Z χ 8 ( t ) d t = 3 2 3/2 π 1/4 T 3/4 + O ( T 1/4 log T ) and ∫ 0 T Z χ − 8 ( t ) d t = O ( T 1/4 log T ) . These integrals are plotted, as a function of T T T , in Figure 2 .
Figure 2. F χ ( T ) : = ∫ 0 T Z χ ( t ) d t F_\chi(T) := \int_0^T Z_\chi(t)\,dt F χ ( T ) := ∫ 0 T Z χ ( t ) d t for the odd (top) and even (middle) primitive characters mod 8, and the even primitive character (bottom) with the main term subtracted. Here C χ 8 ( 5 , ⋅ ) = 2 3 / 2 π 1 / 4 3 = 1.255195 … C_{\chi_8(5,\cdot)} = \frac{2^{3/2}\pi^{1/4}}{3} = 1.255195\ldots C χ 8 ( 5 , ⋅ ) = 3 2 3/2 π 1/4 = 1.255195 … . In this and other examples we use the LMFDB notation for Dirichlet Characters. See [27 ] .
Example 1.3. Let χ 1 \chi_1 χ 1 be the even primitive character (mod 16) with χ 1 ( 5 ) = − i \chi_1(5)=-i χ 1 ( 5 ) = − i , and let χ 2 , χ 3 , \chi_2,\chi_3, χ 2 , χ 3 , and χ 4 \chi_4 χ 4 denote the other 3 primitive characters (mod 16). Then S ( χ 1 ) = 16 e ( − 1 / 32 ) S(\chi_1)=16e(-1/32) S ( χ 1 ) = 16 e ( − 1/32 ) and S ( χ i ) = 0 S(\chi_i)=0 S ( χ i ) = 0 for i = 2 , 3 , 4 i=2,3,4 i = 2 , 3 , 4 . For these characters, Theorem 1.1 gives
∫ 0 T Z χ 1 ( t ) d t = 2 7 / 4 π 1 / 4 3 T 3 / 4 + O ( T 1 / 4 log T ) and ∫ 0 T Z χ i ( t ) d t = O ( T 1 / 4 log T ) , \int_0^T Z_{\chi_1}(t)\,\mathrm{d}t=\frac{2^{7/4}\pi^{1/4}}{3}T^{3/4}+O(T^{1/4}\log T)\quad\text{and}\quad\int_0^T Z_{\chi_i}(t)\,\mathrm{d}t=O(T^{1/4}\log T), ∫ 0 T Z χ 1 ( t ) d t = 3 2 7/4 π 1/4 T 3/4 + O ( T 1/4 log T ) and ∫ 0 T Z χ i ( t ) d t = O ( T 1/4 log T ) , for i = 2 , 3 , 4 i=2,3,4 i = 2 , 3 , 4 . These integrals are plotted, as a function of T T T , in Figure 4 .
The four primitive characters (mod 16) come in two conjugate pairs. This example shows that there are primitive characters χ \chi χ for which S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 and S ( χ ‾ ) = 0 S(\overline{\chi})=0 S ( χ ) = 0 . By the functional equation, we have
∫ 0 T Z χ ( t ) d t = ∫ − T 0 Z χ ‾ ( t ) d t . \int_0^T Z_{\chi}(t)\,\mathrm{d}t=\int_{-T}^0 Z_{\overline{\chi}}(t)\,\mathrm{d}t. ∫ 0 T Z χ ( t ) d t = ∫ − T 0 Z χ ( t ) d t . Therefore, for the even character χ 1 \chi_1 χ 1 (mod 16) with χ 1 ( 5 ) = − i \chi_1(5)=-i χ 1 ( 5 ) = − i in Example 1.3, we have
∫ 0 T Z χ 1 ( t ) d t = 2 7 / 4 π 1 / 4 3 T 3 / 4 + O ( T 1 / 4 log T ) and ∫ − T 0 Z χ 1 ( t ) d t = O ( T 1 / 4 log T ) . \int_0^T Z_{\chi_1}(t)\,\mathrm{d}t=\frac{2^{7/4}\pi^{1/4}}{3}T^{3/4}+O(T^{1/4}\log T)\quad\text{and}\quad\int_{-T}^0 Z_{\chi_1}(t)\,\mathrm{d}t=O(T^{1/4}\log T). ∫ 0 T Z χ 1 ( t ) d t = 3 2 7/4 π 1/4 T 3/4 + O ( T 1/4 log T ) and ∫ − T 0 Z χ 1 ( t ) d t = O ( T 1/4 log T ) . So the asymptotic behavior of these integrals, as functions of T T T , is not symmetric about T = 0 T=0 T = 0 .
Figure 4. F χ ( T ) : = ∫ 0 T Z χ ( t ) d t F_\chi(T) := \int_0^T Z_\chi(t)\,\mathrm{d}t F χ ( T ) := ∫ 0 T Z χ ( t ) d t for the four primitive characters (mod 16) and the four primitive characters (mod 88) for which C χ ≠ 0 C_\chi\ne0 C χ = 0 . Here χ 16 ( 3 , ⋅ ) \chi_{16}(3,\cdot) χ 16 ( 3 , ⋅ ) and χ 16 ( 11 , ⋅ ) \chi_{16}(11,\cdot) χ 16 ( 11 , ⋅ ) are the conjugate pair of odd characters (χ 16 ( 3 , 5 ) = − i \chi_{16}(3,5)=-i χ 16 ( 3 , 5 ) = − i ); characters χ 16 ( 5 , ⋅ ) \chi_{16}(5,\cdot) χ 16 ( 5 , ⋅ ) and χ 16 ( 13 , ⋅ ) \chi_{16}(13,\cdot) χ 16 ( 13 , ⋅ ) are the even characters (χ 16 ( 5 , 5 ) = i \chi_{16}(5,5)=i χ 16 ( 5 , 5 ) = i ). The characters (mod 88) can be distinguished by χ 88 ( 5 , 13 ) = e ( 9 / 10 ) \chi_{88}(5,13)=e(9/10) χ 88 ( 5 , 13 ) = e ( 9/10 ) , χ 88 ( 37 , 13 ) = e ( 7 / 10 ) \chi_{88}(37,13)=e(7/10) χ 88 ( 37 , 13 ) = e ( 7/10 ) , χ 88 ( 53 , 13 ) = e ( 1 / 10 ) \chi_{88}(53,13)=e(1/10) χ 88 ( 53 , 13 ) = e ( 1/10 ) , and χ 88 ( 69 , 13 ) = e ( 3 / 10 ) \chi_{88}(69,13)=e(3/10) χ 88 ( 69 , 13 ) = e ( 3/10 ) .
Example 1.4. There are 9 primitive even characters (mod 88). These are the quadratic character and 8 characters of order 10, which can be distinguished by the values of χ ( 13 ) \chi(13) χ ( 13 ) . There is one Galois orbit for which χ ( 13 ) = e ( a / 5 ) \chi(13)=e(a/5) χ ( 13 ) = e ( a /5 ) , for a = 1 , 2 , 3 , 4 a=1,2,3,4 a = 1 , 2 , 3 , 4 and another Galois orbit for which χ ( 13 ) = e ( a / 10 ) \chi(13)=e(a/10) χ ( 13 ) = e ( a /10 ) for a = 1 , 3 , 7 , 9 a=1,3,7,9 a = 1 , 3 , 7 , 9 . In the first orbit, C χ C_\chi C χ is 0. In the second orbit, C χ C_\chi C χ takes the values 1.076 … , − 0.2616 … , 1.353 … , 0.861 … 1.076\ldots,-0.2616\ldots,1.353\ldots,0.861\ldots 1.076 … , − 0.2616 … , 1.353 … , 0.861 … for a = 1 , 3 , 7 , 9 a=1,3,7,9 a = 1 , 3 , 7 , 9 . This is the smallest modulus for which there is a character χ \chi χ with C χ < 0 C_\chi<0 C χ < 0 . (See Figure 4 .)
We can prove an appropriately modified version of Ivić’s conjecture in (1.3) and show that
∫ 0 T Z χ ( t ) d t − C χ T 3 / 4 = { O q ( T 1 / 4 ) , and Ω ± ( T 1 / 4 ) , \int_{0}^{T} Z_{\chi}(t)\,\mathrm{d}t-C_{\chi}T^{3/4}=
\begin{cases}
O_{q}(T^{1/4}), & \text{and}\\
\Omega_{\pm}(T^{1/4}),
\end{cases} ∫ 0 T Z χ ( t ) d t − C χ T 3/4 = { O q ( T 1/4 ) , Ω ± ( T 1/4 ) , and
where C χ C_{\chi} C χ is the constant defined in (5) . Such a result is quite technical and is not the focus of this paper. For example, for χ 8 \chi_{8} χ 8 and χ − 8 \chi_{-8} χ − 8 , the even and odd characters (mod 8) of Example 1.2 , define
K χ 8 ( u ) = − π ( { u } − 1 2 ) and K χ − 8 ( u ) = − π 2 ( − 1 ) ⌊ u + 1 2 ⌋ . \mathcal{K}_{\chi_{8}}(u)=-\pi\left(\{u\}-\frac{1}{2}\right)
\qquad\text{and}\qquad
\mathcal{K}_{\chi_{-8}}(u)=-\frac{\pi}{2}(-1)^{\lfloor u+\frac{1}{2}\rfloor}. K χ 8 ( u ) = − π ( { u } − 2 1 ) and K χ − 8 ( u ) = − 2 π ( − 1 ) ⌊ u + 2 1 ⌋ .
Then we can prove that, for any fixed δ > 0 \delta>0 δ > 0 ,
∫ 0 T Z χ 8 ( t ) d t = 2 3 / 2 π 1 / 4 3 T 3 / 4 + ( 4 T π ) 1 / 4 K χ 8 ( T / π ) + O δ ( T 1 / 6 log T ) \int_{0}^{T} Z_{\chi_{8}}(t)\,\mathrm{d}t=
\frac{2^{3/2}\pi^{1/4}}{3}T^{3/4}
+\left(\frac{4T}{\pi}\right)^{1/4}\mathcal{K}_{\chi_{8}}\left(\sqrt{T/\pi}\right)
+O_{\delta}(T^{1/6}\log T) ∫ 0 T Z χ 8 ( t ) d t = 3 2 3/2 π 1/4 T 3/4 + ( π 4 T ) 1/4 K χ 8 ( T / π ) + O δ ( T 1/6 log T )
when T / π \sqrt{T/\pi} T / π is at distance at least δ \delta δ from the integers, and
∫ 0 T Z χ − 8 ( t ) d t = ( 4 T π ) 1 / 4 K χ − 8 ( T / π ) + O δ ( T 1 / 6 log T ) \int_{0}^{T} Z_{\chi_{-8}}(t)\,\mathrm{d}t=
\left(\frac{4T}{\pi}\right)^{1/4}\mathcal{K}_{\chi_{-8}}\left(\sqrt{T/\pi}\right)
+O_{\delta}(T^{1/6}\log T) ∫ 0 T Z χ − 8 ( t ) d t = ( π 4 T ) 1/4 K χ − 8 ( T / π ) + O δ ( T 1/6 log T )
when T / π \sqrt{T/\pi} T / π is at distance at least δ \delta δ from Z + 1 2 \mathbb{Z}+\frac{1}{2} Z + 2 1 . Normalized versions of these integrals are plotted in Figure 3 . Similar results, for all primitive characters, will appear in a companion paper.
Figure 3. 2 π T 1 / 2 ( F χ ( π T 2 ) − C χ ( π T 2 ) 3 / 4 ) \frac{\sqrt{2}}{\pi T^{1/2}}\left(F_{\chi}(\pi T^2)-C_{\chi}(\pi T^2)^{3/4}\right) π T 1/2 2 ( F χ ( π T 2 ) − C χ ( π T 2 ) 3/4 ) for the odd (top) and even primitive characters mod 8, where F χ ( T ) : = ∫ 0 T Z χ ( t ) d t F_{\chi}(T):=\int_0^T Z_{\chi}(t)\,\mathrm{d}t F χ ( T ) := ∫ 0 T Z χ ( t ) d t . For the odd character C χ = 0 C_{\chi}=0 C χ = 0 , while for the even character, C χ 8 ( 5 , ⋅ ) = 2 3 / 2 π 1 / 4 3 = 1.255195 … C_{\chi_8(5,\cdot)}=\frac{2^{3/2}\pi^{1/4}}{3}=1.255195\ldots C χ 8 ( 5 , ⋅ ) = 3 2 3/2 π 1/4 = 1.255195 …
A bias in the gaps between zeros of L ( s , χ ) L(s,\chi) L ( s , χ ) on the critical line when S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0
If S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 , then Theorem 1.1 gives
∫ 0 T Z χ ( t ) d t = C χ T 3 / 4 + O q ( T 1 / 4 log T ) \int_{0}^{T} Z_{\chi}(t)\,\mathrm{d}t=C_{\chi}T^{3/4}+O_{q}(T^{1/4}\log T) ∫ 0 T Z χ ( t ) d t = C χ T 3/4 + O q ( T 1/4 log T )
with C χ ≠ 0 C_{\chi}\ne0 C χ = 0 . The integral on the left-hand side is real, so C χ C_{\chi} C χ is real and nonzero. If C χ > 0 C_{\chi}>0 C χ > 0 (resp. C χ < 0 C_{\chi}<0 C χ < 0 ), then Z χ ( t ) Z_{\chi}(t) Z χ ( t ) has a bias towards positive (resp. negative) values. There are a number of reasons that this might happen. For instance, Z χ ( t ) Z_{\chi}(t) Z χ ( t ) might get larger in the positive direction than it does in the negative direction. Or it might be the case that the measure of the set where Z χ ( t ) > 0 Z_{\chi}(t)>0 Z χ ( t ) > 0 is larger than the measure of the set where Z χ ( t ) < 0 Z_{\chi}(t)<0 Z χ ( t ) < 0 . Though the first possibility is quite interesting, we experimentally investigated the second possibility. Since Z χ ( t ) Z_{\chi}(t) Z χ ( t ) conjecturally changes sign at every zero, it is natural to study the spacings between consecutive zeros.
Since it is conjectured that L ( 1 2 , χ ) ≠ 0 L\left(\frac{1}{2},\chi\right)\ne0 L ( 2 1 , χ ) = 0 for all primitive characters χ \chi χ , we may assume that Z χ ( 0 ) > 0 Z_{\chi}(0)>0 Z χ ( 0 ) > 0 . We further assume the generalized Riemann hypothesis (GRH) for L ( s , χ ) L(s,\chi) L ( s , χ ) and that the zeros of L ( s , χ ) L(s,\chi) L ( s , χ ) are all simple. Then the zeros of Z χ ( t ) Z_{\chi}(t) Z χ ( t ) correspond precisely to the imaginary parts of the zeros of L ( s , χ ) L(s,\chi) L ( s , χ ) on the critical line and Z χ ( t ) Z_{\chi}(t) Z χ ( t ) changes sign at every zero.
With these assumptions, we let 0 < γ 1 < γ 2 < γ 3 < ⋯ 0<\gamma_{1}<\gamma_{2}<\gamma_{3}<\cdots 0 < γ 1 < γ 2 < γ 3 < ⋯ denote the sequence of zeros of Z χ ( t ) Z_{\chi}(t) Z χ ( t ) listed in increasing order, and let g n = γ n + 1 − γ n g_{n}=\gamma_{n+1}-\gamma_{n} g n = γ n + 1 − γ n denote the size of the n th n^{\text{th}} n th gap in this sequence. Then Z χ ( t ) < 0 Z_{\chi}(t)<0 Z χ ( t ) < 0 when t ∈ ( γ 2 n − 1 , γ 2 n ) t\in(\gamma_{2n-1},\gamma_{2n}) t ∈ ( γ 2 n − 1 , γ 2 n ) and Z χ ( t ) > 0 Z_{\chi}(t)>0 Z χ ( t ) > 0 when t ∈ ( γ 2 n , γ 2 n + 1 ) t\in(\gamma_{2n},\gamma_{2n+1}) t ∈ ( γ 2 n , γ 2 n + 1 ) . So the size of the gap g n g_{n} g n between consecutive zeros corresponds to the measure of the set where Z χ ( t ) < 0 Z_{\chi}(t)<0 Z χ ( t ) < 0 when n n n is odd and to the measure of the set where Z χ ( t ) > 0 Z_{\chi}(t)>0 Z χ ( t ) > 0 when n n n is even. If C χ > 0 C_{\chi}>0 C χ > 0 (resp. C χ < 0 C_{\chi}<0 C χ < 0 ), then Z χ ( t ) Z_{\chi}(t) Z χ ( t ) has a positive (resp. negative) bias and we want to know to what extent that is reflected in the sizes of the gaps g n g_{n} g n where n n n is even (resp. odd).
Here is a striking way to see a bias towards even or odd gaps in the zeros of certain L L L -functions. Define the normalized gaps g ~ n = γ n + 1 log ( γ n + 1 ) − γ n log ( γ n ) \widetilde{g}_n=\gamma_{n+1}\log(\gamma_{n+1})-\gamma_n\log(\gamma_n) g n = γ n + 1 log ( γ n + 1 ) − γ n log ( γ n ) , so that g ~ n \widetilde{g}_n g n is (asymptotically) constant on average. Take some large enough parameter X X X and consider the two averages
A 0 ( n ) = 1 X ∑ 0 ≤ k < X g ~ 2 n + 2 k and A 1 ( n ) = 1 X ∑ 0 ≤ k < X g ~ 2 n + 2 k + 1 (6) A_0(n)=\frac{1}{X}\sum_{0\leq k<X}\widetilde{g}_{2n+2k}
\quad\text{and}\quad
A_1(n)=\frac{1}{X}\sum_{0\leq k<X}\widetilde{g}_{2n+2k+1} \tag*{(6)} A 0 ( n ) = X 1 0 ≤ k < X ∑ g 2 n + 2 k and A 1 ( n ) = X 1 0 ≤ k < X ∑ g 2 n + 2 k + 1 (6)
as functions of n n n . Furthermore, to define something that averages to 1, define
A ~ 0 ( n ) = 2 A 0 ( n ) A 0 ( n ) + A 1 ( n ) and A ~ 1 ( n ) = 2 A 1 ( n ) A 0 ( n ) + A 1 ( n ) . \widetilde{A}_0(n)=\frac{2A_0(n)}{A_0(n)+A_1(n)}
\quad\text{and}\quad
\widetilde{A}_1(n)=\frac{2A_1(n)}{A_0(n)+A_1(n)}. A 0 ( n ) = A 0 ( n ) + A 1 ( n ) 2 A 0 ( n ) and A 1 ( n ) = A 0 ( n ) + A 1 ( n ) 2 A 1 ( n ) .
A priori, we should expect these averages to be very close to the mean value of the gaps over the whole range, so significant deviation from the mean is perhaps already a curious feature. In Figure 5 , we see (for X = 25000 X=25000 X = 25000 ) that the first of these averages is significantly larger than the second for n n n up to 1,500,000 (that is, over the first 3 million zeros) for the even primitive character (mod 8); for the odd character, there is no such bias.
Figure 5. Running averages of 25000 consecutive normalized zero gaps for the odd (top) and even (bottom) primitive characters (mod 8). See the discussion around Equation (6) .
In Figure 6 , we plot the same averages for the zeros of the L L L -functions associated to the even primitive characters (mod 16). Let χ 1 \chi_1 χ 1 (mod 16) be the character discussed in Example 1.3 , and let χ ‾ 1 \overline{\chi}_1 χ 1 be its conjugate. In this case, S ( χ 1 ) ≠ 0 S(\chi_1)\ne0 S ( χ 1 ) = 0 and S ( χ ‾ 1 ) = 0 S(\overline{\chi}_1)=0 S ( χ 1 ) = 0 , and we observe that there appears to be a dramatic bias in the average size of the gaps g n g_n g n depending on whether n n n is even or odd for the zeros of L ( s , χ 1 ) L(s,\chi_1) L ( s , χ 1 ) but not for the zeros of L ( s , χ ‾ 1 ) L(s,\overline{\chi}_1) L ( s , χ 1 ) .
Figure 6. Running averages of 25000 consecutive normalized zero gaps for the complex conjugate pair of primitive even characters (mod 16) of Example 1.3. Top: χ 16 ( 5 , 3 ) = − i \chi_{16}(5,3)=-i χ 16 ( 5 , 3 ) = − i . Bottom: χ 16 ( 13 , 3 ) = i \chi_{16}(13,3)=i χ 16 ( 13 , 3 ) = i . See the discussion around Equation (6) for a precise description.
The choice of parameters for Figure 5 and Figure 6 is somewhat arbitrary; however, studying bias in gaps is the same as studying a bias in measure in the following sense. Following Gonek and
Ivić [13 ] , who studied the distribution of positive and negative values of the Hardy Z Z Z -function Z ( t ) Z(t) Z ( t ) , we define
M χ ± ( T ) = meas { t ∈ [ 0 , T ] : ± Z χ ( t ) > 0 } . (7) M_\chi^\pm(T)=\operatorname{meas}\{t\in[0,T]:\ \pm Z_\chi(t)>0\}. \tag*{(7)} M χ ± ( T ) = meas { t ∈ [ 0 , T ] : ± Z χ ( t ) > 0 } . (7)
Then, we have
M χ + ( T ) − M χ − ( T ) = ∑ n ≥ 1 0 < γ 2 n + 1 ≤ T ( g 2 n − g 2 n − 1 ) + O ( 1 ) . M_\chi^+(T)-M_\chi^-(T)=\sum_{\substack{n\ge1\\0<\gamma_{2n+1}\le T}}(g_{2n}-g_{2n-1})+O(1). M χ + ( T ) − M χ − ( T ) = n ≥ 1 0 < γ 2 n + 1 ≤ T ∑ ( g 2 n − g 2 n − 1 ) + O ( 1 ) .
In Figure 7 , we plot M χ 8 + ( T ) − M χ 8 − ( T ) M_{\chi_8}^+(T)-M_{\chi_8}^-(T) M χ 8 + ( T ) − M χ 8 − ( T ) and M χ − 8 + ( T ) − M χ − 8 − ( T ) M_{\chi_{-8}}^+(T)-M_{\chi_{-8}}^-(T) M χ − 8 + ( T ) − M χ − 8 − ( T ) for the two primitive characters (mod 8) discussed in Example 1.2. For χ 8 \chi_8 χ 8 , which has S ( χ 8 ) ≠ 0 S(\chi_8)\ne0 S ( χ 8 ) = 0 , there is a clear bias in measure for positive values of Z χ 8 ( t ) Z_{\chi_8}(t) Z χ 8 ( t ) which is not present for the character χ − 8 \chi_{-8} χ − 8 , where S ( χ − 8 ) = 0 S(\chi_{-8})=0 S ( χ − 8 ) = 0 . In Figure 8 , we plot M χ 1 + ( T ) − M χ 1 − ( T ) M_{\chi_1}^+(T)-M_{\chi_1}^-(T) M χ 1 + ( T ) − M χ 1 − ( T ) and M χ ‾ 1 + ( T ) − M χ ‾ 1 − ( T ) M_{\overline{\chi}_1}^+(T)-M_{\overline{\chi}_1}^-(T) M χ 1 + ( T ) − M χ 1 − ( T ) for the even character χ 1 \chi_1 χ 1 (mod 16) discussed in Example 1.3. For χ 1 \chi_1 χ 1 , which has S ( χ 1 ) ≠ 0 S(\chi_1)\ne0 S ( χ 1 ) = 0 , there is a clear bias in measure for positive values of Z χ 1 ( t ) Z_{\chi_1}(t) Z χ 1 ( t ) which is not present for its conjugate character χ ‾ 1 \overline{\chi}_1 χ 1 , where S ( χ ‾ 1 ) = 0 S(\overline{\chi}_1)=0 S ( χ 1 ) = 0 .
Figure 7. The difference between the amount of time that Z χ ( t ) Z_{\chi}(t) Z χ ( t ) spends positive and negative for the odd (top) and even (bottom) primitive characters (mod 8).
Figure 8. The difference between the amount of time that Z χ ( t ) Z_{\chi}(t) Z χ ( t ) spends positive and negative, for the even character (mod 16) and its complex conjugate, discussed in Example 1.3. Top: χ 16 ( 5 , 3 ) = − i \chi_{16}(5,3)=-i χ 16 ( 5 , 3 ) = − i . Bottom: χ 16 ( 13 , 3 ) = i \chi_{16}(13,3)=i χ 16 ( 13 , 3 ) = i .
Letting M ± ( T ) = meas { t ∈ [ 0 , T ] : ± Z ( t ) > 0 } M^\pm(T)=\operatorname{meas}\{t\in[0,T]:\ \pm Z(t)>0\} M ± ( T ) = meas { t ∈ [ 0 , T ] : ± Z ( t ) > 0 } , Gonek and Ivić [13 ] have conjectured that M ± ( T ) ∼ T 2 M^\pm(T)\sim\frac{T}{2} M ± ( T ) ∼ 2 T as T → ∞ T\to\infty T → ∞ and they proved some interesting partial results toward this conjecture. Likewise, we conjecture that M χ ± ( T ) ∼ T 2 M_\chi^\pm(T)\sim\frac{T}{2} M χ ± ( T ) ∼ 2 T as T → ∞ T\to\infty T → ∞ for every primitive character χ \chi χ . However,
when S ( χ ) ≠ 0 S(\chi) \ne0 S ( χ ) = 0 , our numerical data suggest that M χ + ( T ) − M χ − ( T ) M^+_\chi(T)-M^-_\chi(T) M χ + ( T ) − M χ − ( T ) tends to + ∞ +\infty + ∞ if C χ > 0 C_\chi>0 C χ > 0 and to − ∞ -\infty − ∞ if C χ < 0 C_\chi<0 C χ < 0 . This is discussed more in Section 5.
Experimentally, we have found another thin family of Dirichlet characters χ \chi χ with odd modulus for which S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 but there still seems to be a bias in gaps and a bias in measure as described above. It seems these biases can be explained via discrete moments. For instance, assuming GRH for Dirichlet L L L -functions, we have proved asymptotic formulae for the discrete moments
∑ 0 < γ χ ≤ T Z χ ′ ( γ χ ) and ∑ 0 < λ χ ≤ T Z χ ( λ χ ) , \sum_{0<\gamma_\chi\le T} Z'_\chi(\gamma_\chi)
\qquad\text{and}\qquad
\sum_{0<\lambda_\chi\le T} Z_\chi(\lambda_\chi), 0 < γ χ ≤ T ∑ Z χ ′ ( γ χ ) and 0 < λ χ ≤ T ∑ Z χ ( λ χ ) ,
where γ χ \gamma_\chi γ χ and λ χ \lambda_\chi λ χ run over the zeros of Z χ ( t ) Z_\chi(t) Z χ ( t ) and Z χ ′ ( t ) Z'_\chi(t) Z χ ′ ( t ) , respectively. A new Gauss-type sum arises in the main terms of these asymptotic formulae; it vanishes for primitive characters of even modulus and can be nonzero for even characters of odd modulus.
Perhaps surprisingly, the Riemann zeta-function is in this family. For the usual Hardy Z Z Z -function, Z ( t ) Z(t) Z ( t ) , assuming RH, for every ε > 0 \varepsilon>0 ε > 0 , our results give
∑ 0 < γ ≤ T Z ′ ( γ ) = − 2 3 / 2 log 2 sin ( π / 8 ) 3 ( T 2 π ) 3 / 4 ( log T 2 π + 2 γ 0 − 5 log 2 − 4 3 ) + O ε ( T 1 / 2 + ε ) . \sum_{0<\gamma\le T} Z'(\gamma)=-\frac{2^{3/2}\log2\sin(\pi/8)}{3}\left(\frac{T}{2\pi}\right)^{3/4}\left(\log\frac{T}{2\pi}+2\gamma_0-5\log2-\frac{4}{3}\right)+O_\varepsilon\left(T^{1/2+\varepsilon}\right). 0 < γ ≤ T ∑ Z ′ ( γ ) = − 3 2 3/2 log 2 sin ( π /8 ) ( 2 π T ) 3/4 ( log 2 π T + 2 γ 0 − 5 log 2 − 3 4 ) + O ε ( T 1/2 + ε ) .
where γ 0 \gamma_0 γ 0 is Euler’s constant, and
∑ 0 < λ ≤ T Z ( λ ) = − 2 5 / 2 ( e − 2 ) log 2 cos ( π / 8 ) 3 ( T 2 π ) 3 / 4 + O ( T 3 / 4 log T ) , \sum_{0<\lambda\le T} Z(\lambda)=-\frac{2^{5/2}(e-2)\log2\cos(\pi/8)}{3}\left(\frac{T}{2\pi}\right)^{3/4}+O\left(\frac{T^{3/4}}{\log T}\right), 0 < λ ≤ T ∑ Z ( λ ) = − 3 2 5/2 ( e − 2 ) log 2 cos ( π /8 ) ( 2 π T ) 3/4 + O ( log T T 3/4 ) ,
where γ \gamma γ and λ \lambda λ run over the zeros of Z ( t ) Z(t) Z ( t ) and Z ′ ( t ) Z'(t) Z ′ ( t ) , respectively. Both discrete averages have negative leading coefficients and this is reflected in the experimental bias in gaps and bias in measure for the Hardy Z Z Z -function (see Figure 9 and Figure 10 ). For the second sum, we obtain an asymptotic expansion to arbitrarily many orders in inverse powers of log T \log T log T . We will present results for all primitive characters, together with their proofs, in a companion paper.
Figure 9. Running averages of 25000 consecutive normalized gaps of ordinates of zeros of ζ ( s ) \zeta(s) ζ ( s ) , the Riemann zeta-function. See the discussion around Equation (6) . The blue curve corresponds to the intervals where Z ( t ) < 0 Z(t)<0 Z ( t ) < 0 .
Figure 10. Plot of M + ( T ) − M − ( T ) M^{+}(T)-M^{-}(T) M + ( T ) − M − ( T ) , where M ± ( T ) = meas { t ∈ [ 0 , T ] : ± Z ( t ) > 0 } M^{\pm}(T)=\operatorname{meas}\{t\in[0,T]:\pm Z(t)>0\} M ± ( T ) = meas { t ∈ [ 0 , T ] : ± Z ( t ) > 0 } and Z ( t ) Z(t) Z ( t ) is the Hardy Z Z Z -function. This plot was made using approximately 100 billion zeros of ζ ( s ) \zeta(s) ζ ( s ) , computed by D. Platt [29 ] .
Random matrix models have played an important role in the study of L L L -functions for many years, so a natural question is how to construct a random model for these Z Z Z -functions that simultaneously accounts for their continuous and discrete moments. Such a model should reflect the arithmetic dependence of the main terms, including the different Gauss-type sums that arise in these asymptotic formulae. We plan to investigate this in our forthcoming paper on discrete moments.
Analysis of S ( χ ) S(\chi) S ( χ )
In Proposition 1.7 of this subsection, we give a complete classification of the primitive χ \chi χ (mod q q q ) for which
S ( χ ) = ∑ n = 1 2 q χ ( n ) e ( − n 2 2 q ) ≠ 0. S(\chi)=\sum_{n=1}^{2q}\chi(n)e\left(-\frac{n^2}{2q}\right)\ne0. S ( χ ) = n = 1 ∑ 2 q χ ( n ) e ( − 2 q n 2 ) = 0.
Note the following simple observations. Replacing n n n by − n -n − n in the sum, we see that S ( χ ) = χ ( − 1 ) S ( χ ) S(\chi)=\chi(-1)S(\chi) S ( χ ) = χ ( − 1 ) S ( χ ) . So S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 when χ \chi χ is odd. Also,
S ( χ ) = { ∑ n = 1 q + ∑ n = q + 1 2 q } χ ( n ) e ( − n 2 2 q ) = { 1 + e ( − q 2 ) } ∑ n = 1 q χ ( n ) e ( − n 2 2 q ) . S(\chi)=\left\{\sum_{n=1}^{q}+\sum_{n=q+1}^{2q}\right\}\chi(n)e\left(-\frac{n^2}{2q}\right)=\left\{1+e\left(-\frac{q}{2}\right)\right\}\sum_{n=1}^{q}\chi(n)e\left(-\frac{n^2}{2q}\right). S ( χ ) = { n = 1 ∑ q + n = q + 1 ∑ 2 q } χ ( n ) e ( − 2 q n 2 ) = { 1 + e ( − 2 q ) } n = 1 ∑ q χ ( n ) e ( − 2 q n 2 ) .
So S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 when q q q is odd. These two observations show that S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 for at least 75% of primitive characters. In general, we let
N ( q ) : = # { χ ( m o d q ) : χ primitive and S ( χ ) ≠ 0 } (8) N(q):=\#\{\chi\pmod q:\chi\text{ primitive and }S(\chi)\ne0\} \tag*{(8)} N ( q ) := # { χ ( mod q ) : χ primitive and S ( χ ) = 0 } (8)
denote the number of primitive characters mod q q q for which S ( χ ) S(\chi) S ( χ ) is nonzero. A consequence of Proposition 1.7 is the following exact formula for N ( q ) N(q) N ( q ) in terms of the prime factorization of q q q .
Theorem 1.5. Let q ∈ N q\in\mathbb{N} q ∈ N . If 8 8 8 does not divide q q q , then N ( q ) = 0 N(q)=0 N ( q ) = 0 . If 8 ∣ q 8\mid q 8 ∣ q , write
q = 2 m ( ∏ i p i ) ( ∏ j r j a j ) , q=2^m\left(\prod_i p_i\right)\left(\prod_j r_j^{a_j}\right), q = 2 m ( i ∏ p i ) ( j ∏ r j a j ) , where p i , r j p_i,r_j p i , r j are distinct odd primes, m ≥ 3 m\ge3 m ≥ 3 , and a j ≥ 2 a_j\ge2 a j ≥ 2 . Then, we have
N ( q ) = ⌊ 2 m − 5 ⌋ ∏ i ( p i − 3 2 ) ∏ j ( 1 4 r j a j − 2 ( r j − 1 ) 2 ) . (9) N(q)=\left\lfloor2^{m-5}\right\rfloor\prod_i\left(\frac{p_i-3}{2}\right)\prod_j\left(\frac{1}{4}r_j^{a_j-2}(r_j-1)^2\right). \tag*{(9)} N ( q ) = ⌊ 2 m − 5 ⌋ i ∏ ( 2 p i − 3 ) j ∏ ( 4 1 r j a j − 2 ( r j − 1 ) 2 ) . (9)
Using the Selberg–Delange method, we can compute the average of N ( q ) N(q) N ( q ) , as q q q varies.
Corollary 1.6. As Q → ∞ Q \to\infty Q → ∞ , we have
∑ q ≤ Q N ( q ) = A Q 2 log Q + O ( Q 2 ( log Q ) 3 / 2 ) , (10) \sum_{q \le Q} N(q) = \frac{\mathcal{A}Q^2}{\sqrt{\log Q}} + O\left(\frac{Q^2}{(\log Q)^{3/2}}\right), \tag*{(10)} q ≤ Q ∑ N ( q ) = log Q A Q 2 + O ( ( log Q ) 3/2 Q 2 ) , (10) where A > 0 \mathcal{A} > 0 A > 0 is a constant given explicitly in (3.2).
Since there are ≍ Q 2 \asymp Q^2 ≍ Q 2 primitive characters with modulus not exceeding Q Q Q , Corollary 1.6 implies that S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 for 100% of primitive characters. Thus, the constant in Theorem 1.1 satisfies C χ = 0 C_\chi=0 C χ = 0 for 100% of primitive characters and therefore we expect the analogue of Ivić’s conjecture in (3) to hold for 100% of primitive characters.
We now state our classification of the primitive χ \chi χ for which S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 . As seen in Corollary 1.6 , these characters form a set of density zero among all primitive characters.
Proposition 1.7 (Main Proposition). Let χ \chi χ be a primitive character modulo q q q , where q = ∏ p p e p q=\prod_p p^{e_p} q = ∏ p p e p is the factorization of q q q into prime powers and χ = ∏ p χ p \chi=\prod_p\chi_p χ = ∏ p χ p is the factorization of χ \chi χ into characters of prime-power modulus. Set α 2 \alpha_2 α 2 to be an integer such that χ 2 ( 5 ) = e ( α 2 / 2 e 2 − 2 ) \chi_2(5)=e(\alpha_2/2^{e_2-2}) χ 2 ( 5 ) = e ( α 2 / 2 e 2 − 2 ) , and for each other prime p p p set α p \alpha_p α p to be an integer such that χ p ( p + 1 ) = e ( α p / p e p − 1 ) \chi_p(p+1)=e(\alpha_p/p^{e_p-1}) χ p ( p + 1 ) = e ( α p / p e p − 1 ) . Then S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 if and only if all of the following hold:
(i) χ p \chi_p χ p is even for each p p p ;
(ii) For each p p p with e p > 1 e_p>1 e p > 1 , the residue α p q / p e p \alpha_pq/p^{e_p} α p q / p e p is a square modulo p p p ;
(iii) e 2 > 2 e_2>2 e 2 > 2 , and α 2 ≡ 3 q / 2 e 2 \alpha_2\equiv3q/2^{e_2} α 2 ≡ 3 q / 2 e 2 (mod $2^{e_2-2}$) if 3 ≤ e 2 ≤ 5 3\le e_2\le5 3 ≤ e 2 ≤ 5 or α 2 ≡ − q / 2 e 2 \alpha_2\equiv-q/2^{e_2} α 2 ≡ − q / 2 e 2 (mod 8) if $e_2\ge 6.
In this case, let ψ = ∏ p ∣ q ψ p \psi=\prod_{p\mid q}\psi_p ψ = ∏ p ∣ q ψ p be a character (mod 2 q 2q 2 q ) such that ψ 2 = χ \psi^2=\chi ψ 2 = χ . Then S ( χ ) = S e v e n S o d d S(\chi)=S_{\rm even}S_{\rm odd} S ( χ ) = S even S odd , where
S e v e n = { 2 e 2 ψ 2 ( q / 2 e 2 ) e ( − q / 2 e 2 + 1 ) , if e 2 < 6 , 4 ψ 2 ( − q / 2 e 2 ) τ ( ψ 2 ) , if e 2 ≥ 6 , (11) S_{\rm even} =
\begin{cases}
2^{e_2}\psi_2(q/2^{e_2})e(-q/2^{e_2+1}), & \text{if } e_2<6,\\
4\psi_2(-q/2^{e_2})\tau(\psi_2), & \text{if } e_2\ge6,
\end{cases} \tag*{(11)} S even = { 2 e 2 ψ 2 ( q / 2 e 2 ) e ( − q / 2 e 2 + 1 ) , 4 ψ 2 ( − q / 2 e 2 ) τ ( ψ 2 ) , if e 2 < 6 , if e 2 ≥ 6 , (11) and
S o d d = ∏ p ∣ q p o d d ψ p ( − 2 q / p e p ) ( τ ( ψ p ) + ϵ p ( − 2 q / p e p ) τ ( ϵ p ψ p ) ) = ∏ p ∥ q p o d d ψ p ( − 2 q / p e p ) ( τ ( ψ p ) + ϵ p ( − 2 q / p e p ) τ ( ϵ p ψ p ) ) ∏ p 2 ∣ q p o d d 2 ψ p ( − 2 q / p e p ) τ ( ψ p ) , (12) \begin{aligned}
S_{\rm odd}
&=\prod_{\substack{p\mid q\\p\ \mathrm{odd}}}\psi_p(-2q/p^{e_p})\left(\tau(\psi_p)+\epsilon_p(-2q/p^{e_p})\tau(\epsilon_p\psi_p)\right)\\
&=\prod_{\substack{p\parallel q\\p\ \mathrm{odd}}}\psi_p(-2q/p^{e_p})\left(\tau(\psi_p)+\epsilon_p(-2q/p^{e_p})\tau(\epsilon_p\psi_p)\right)
\prod_{\substack{p^2\mid q\\p\ \mathrm{odd}}}2\psi_p(-2q/p^{e_p})\tau(\psi_p), \tag*{(12)}
\end{aligned} S odd = p ∣ q p odd ∏ ψ p ( − 2 q / p e p ) ( τ ( ψ p ) + ϵ p ( − 2 q / p e p ) τ ( ϵ p ψ p ) ) = p ∥ q p odd ∏ ψ p ( − 2 q / p e p ) ( τ ( ψ p ) + ϵ p ( − 2 q / p e p ) τ ( ϵ p ψ p ) ) p 2 ∣ q p odd ∏ 2 ψ p ( − 2 q / p e p ) τ ( ψ p ) , (12) where ϵ p \epsilon_p ϵ p is the quadratic character (mod p p p ).
Remarks.
(i) Note, for example, that there is no even primitive character (mod 3), so N ( q ) = 0 N(q)=0 N ( q ) = 0 if q q q is divisible by 3 but not 9.
(ii) From the above evaluation of S ( χ ) S(\chi) S ( χ ) , we see that when S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 , the value ∣ S ( χ ) ∣ |S(\chi)| ∣ S ( χ ) ∣ is easy to compute whenever q q q is square-full. For example, ∣ S ( χ ) ∣ = 2 ω ( q ) + 3 / 2 q |S(\chi)|=2^{\omega(q)+3/2}\sqrt{q} ∣ S ( χ ) ∣ = 2 ω ( q ) + 3/2 q if q q q is square-full and e 2 ≥ 6 e_2\ge6 e 2 ≥ 6 , where as usual ω ( q ) \omega(q) ω ( q ) is the number of distinct prime divisors of q q q . On the other hand, for a character ψ p \psi_p ψ p (mod p p p ), the quantity
τ ( ψ p ) ± τ ( ϵ p ψ p ) \tau(\psi_p)\pm\tau(\epsilon_p\psi_p) τ ( ψ p ) ± τ ( ϵ p ψ p )
may be much smaller than 2 p 2\sqrt{p} 2 p . We discuss this in more detail in Section 5.
We conclude this subsection by describing a simple sub-family of characters χ \chi χ for which S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 which follows directly from Proposition 1.7 .
Example 1.8. Let q = 8 d q=8d q = 8 d , where d d d is an odd square-free number. Then the primitive characters χ \chi χ (mod q q q ) for which S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 are precisely those characters of the form χ = ψ 2 \chi=\psi^2 χ = ψ 2 , where ψ \psi ψ is a primitive character (mod 16 d 16d 16 d ).
To see why, note that part (i) of Proposition 1.7 implies that χ \chi χ is a product of the even primitive character (mod 8 8 8 ) times the product of even primitive characters (mod p p p ) for each prime p p p dividing d d d . When p p p is odd, even primitive characters (mod p p p ) are precisely the squares of non-quadratic primitive characters (mod p p p ). The even primitive character (mod 8 8 8 ) is the square of a primitive character (mod 16 16 16 ). Therefore, S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 if and only if χ \chi χ is the square of a primitive character (mod 16 d 16d 16 d ).
The Z-function of degree two L-functions
It is known that L L L -functions associated to cuspidal automorphic representations of GL ( 2 ) \operatorname{GL}(2) GL ( 2 ) over the rationals correspond to primitive holomorphic cusp forms and primitive Hecke–Maass cusp forms. In contrast to the thin family of Dirichlet L L L -functions associated to the characters described in Proposition 1.7 , we show that the integral of the Z Z Z -functions of these degree two L L L -functions never has a main term of size T 3 / 4 T^{3/4} T 3/4 .
Let f ∈ S k ( Γ 0 ( N ) , χ ) f\in S_k(\Gamma_0(N),\chi) f ∈ S k ( Γ 0 ( N ) , χ ) be a primitive holomorphic cusp form of weight k ≥ 1 k\ge1 k ≥ 1 , level N N N , and nebentypus χ \chi χ , and let
f ( z ) = ∑ n ≥ 1 λ f ( n ) n ( k − 1 ) / 2 e ( n z ) f(z)=\sum_{n\ge1}\lambda_f(n)n^{(k-1)/2}e(nz) f ( z ) = n ≥ 1 ∑ λ f ( n ) n ( k − 1 ) /2 e ( n z )
be its normalized Fourier expansion at the cusp ∞ \infty ∞ . Then the L L L -function L ( s , f ) = ∑ n ≥ 1 λ f ( n ) n − s L(s,f)=\sum_{n\ge1}\lambda_f(n)n^{-s} L ( s , f ) = ∑ n ≥ 1 λ f ( n ) n − s extends to an entire function and satisfies the functional equation, in asymmetric form,
L ( s , f ) = G f ( s ) L ( 1 − s , f ‾ ) , where G f ( s ) = ε f N 1 / 2 − s Γ R ( 1 − s + k − 1 2 ) Γ R ( 1 − s + k + 1 2 ) Γ R ( s + k − 1 2 ) Γ R ( s + k + 1 2 ) , L(s,f)=G_f(s)L(1-s,\overline{f}),\qquad\text{where}\qquad G_f(s)=\varepsilon_f N^{1/2-s}\frac{\Gamma_{\mathbb{R}}\left(1-s+\frac{k-1}{2}\right)\Gamma_{\mathbb{R}}\left(1-s+\frac{k+1}{2}\right)}{\Gamma_{\mathbb{R}}\left(s+\frac{k-1}{2}\right)\Gamma_{\mathbb{R}}\left(s+\frac{k+1}{2}\right)}, L ( s , f ) = G f ( s ) L ( 1 − s , f ) , where G f ( s ) = ε f N 1/2 − s Γ R ( s + 2 k − 1 ) Γ R ( s + 2 k + 1 ) Γ R ( 1 − s + 2 k − 1 ) Γ R ( 1 − s + 2 k + 1 ) ,
where ε f \varepsilon_f ε f is the root number of L ( s , f ) L(s,f) L ( s , f ) , and f ‾ ∈ S k ( Γ 0 ( N ) , χ ‾ ) \overline{f}\in S_k(\Gamma_0(N),\overline{\chi}) f ∈ S k ( Γ 0 ( N ) , χ ) is the dual form of f f f with Fourier coefficients satisfying λ f ‾ ( n ) = λ f ( n ) ‾ \lambda_{\overline{f}}(n)=\overline{\lambda_f(n)} λ f ( n ) = λ f ( n ) . Then we define
Z f ( t ) = G f ( 1 2 + i t ) − 1 / 2 L ( 1 2 + i t , f ) . Z_f(t)=G_f\left(\frac{1}{2}+it\right)^{-1/2}L\left(\frac{1}{2}+it,f\right). Z f ( t ) = G f ( 2 1 + i t ) − 1/2 L ( 2 1 + i t , f ) .
As before, Z f ( t ) Z_f(t) Z f ( t ) is real-valued for real t t t and ∣ Z f ( t ) ∣ = ∣ L ( 1 2 + i t , f ) ∣ \lvert Z_f(t)\rvert=\left\lvert L\left(\frac{1}{2}+it,f\right)\right\rvert ∣ Z f ( t )∣ = L ( 2 1 + i t , f ) .
Theorem 1.9. Let f ∈ S k ( Γ 0 ( N ) , χ ) f\in S_k(\Gamma_0(N),\chi) f ∈ S k ( Γ 0 ( N ) , χ ) be a primitive holomorphic cusp form. Then
∫ 0 T Z f ( t ) d t ≪ f T log T , \int_0^T Z_f(t)\,\mathrm{d}t\ll_f \sqrt{T}\log T, ∫ 0 T Z f ( t ) d t ≪ f T log T , as T → ∞ T\to\infty T → ∞ .
Let φ \varphi φ be a primitive Hecke–Maass cusp form of level N N N with nebentypus χ \chi χ which is an eigenfunction of the Laplace operator with eigenvalue λ = 1 4 + r 2 \lambda=\frac14+r^2 λ = 4 1 + r 2 , where r ∈ R r\in\mathbb{R} r ∈ R or i r ∈ [ 0 , 1 2 ) ir\in[0,\frac12) i r ∈ [ 0 , 2 1 ) . For z = x + i y z=x+iy z = x + i y , we write the Fourier expansion of φ ( z ) \varphi(z) φ ( z ) at ∞ \infty ∞ as
φ ( z ) = y ∑ n ∈ Z ∖ { 0 } λ φ ( n ) K i r ( 2 π ∣ n ∣ y ) e ( n x ) . \varphi(z)=\sqrt{y}\sum_{n\in\mathbb{Z}\setminus\{0\}}\lambda_\varphi(n)K_{ir}(2\pi\lvert n\rvert y)e(nx). φ ( z ) = y n ∈ Z ∖ { 0 } ∑ λ φ ( n ) K i r ( 2 π ∣ n ∣ y ) e ( n x ) .
Then the L L L -function L ( s , φ ) = ∑ n ≥ 1 λ φ ( n ) n − s L(s,\varphi)=\sum_{n\ge1}\lambda_\varphi(n)n^{-s} L ( s , φ ) = ∑ n ≥ 1 λ φ ( n ) n − s extends to an entire function and satisfies the functional equation, in asymmetric form,
L ( s , φ ) = G φ ( s ) L ( 1 − s , φ ‾ ) , where G φ ( s ) = ε φ N 1 / 2 − s Γ R ( 1 − s + δ + i r ) Γ R ( 1 − s + δ − i r ) Γ R ( s + δ + i r ) Γ R ( s + δ − i r ) , L(s,\varphi)=G_\varphi(s)L(1-s,\overline{\varphi}),\qquad\text{where}\qquad G_\varphi(s)=\varepsilon_\varphi N^{1/2-s}\frac{\Gamma_{\mathbb{R}}(1-s+\delta+ir)\Gamma_{\mathbb{R}}(1-s+\delta-ir)}{\Gamma_{\mathbb{R}}(s+\delta+ir)\Gamma_{\mathbb{R}}(s+\delta-ir)}, L ( s , φ ) = G φ ( s ) L ( 1 − s , φ ) , where G φ ( s ) = ε φ N 1/2 − s Γ R ( s + δ + i r ) Γ R ( s + δ − i r ) Γ R ( 1 − s + δ + i r ) Γ R ( 1 − s + δ − i r ) ,
where ε φ \varepsilon_\varphi ε φ is the root number of φ \varphi φ , δ = 0 \delta=0 δ = 0 if φ \varphi φ is even, δ = 1 \delta=1 δ = 1 if φ \varphi φ is odd, and L ( s , φ ‾ ) L(s,\overline{\varphi}) L ( s , φ ) is the L L L -function with coefficients λ φ ‾ ( n ) = λ φ ( n ) ‾ \lambda_{\overline{\varphi}}(n)=\overline{\lambda_\varphi(n)} λ φ ( n ) = λ φ ( n ) . Then we define
Z φ ( t ) = G φ ( 1 2 + i t ) − 1 / 2 L ( 1 2 + i t , φ ) . Z_\varphi(t)=\mathcal{G}_\varphi\left(\frac{1}{2}+it\right)^{-1/2}L\left(\frac{1}{2}+it,\varphi\right). Z φ ( t ) = G φ ( 2 1 + i t ) − 1/2 L ( 2 1 + i t , φ ) .
Theorem 1.10. Let φ \varphi φ be a primitive Hecke–Maass cusp form of level N N N with nebentypus χ \chi χ . Then, as T → ∞ T\to\infty T → ∞ , we have
∫ 0 T Z φ ( t ) d t ≪ φ , ε T 11 20 + ε \int_0^T Z_\varphi(t)\,\mathrm{d}t\ll_{\varphi,\varepsilon}T^{\frac{11}{20}+\varepsilon} ∫ 0 T Z φ ( t ) d t ≪ φ , ε T 20 11 + ε for any ε > 0 \varepsilon>0 ε > 0 .
Remarks.
(i) There is some hope to improve the upper bound on the right-hand side of Theorem 1.9 from ≪ f T log T \ll_f\sqrt{T}\log T ≪ f T log T to ≪ ε , f T 1 / 3 + ε \ll_{\varepsilon,f}T^{1/3+\varepsilon} ≪ ε , f T 1/3 + ε when the level N N N is square. We plan to investigate this possibility in a future work.
(ii) The difference in order of magnitude in upper bounds in Theorems 1.9 and 1.10 comes from the current knowledge of estimates for the coefficients of the L L L -functions in short intervals. See Equations (27) and (28) . In particular, the assumption of the Ramanujan–Petersson conjecture for the coefficients of L ( s , φ ) L(s,\varphi) L ( s , φ ) would allow us to replace the exponent 11 20 + ε \frac{11}{20}+\varepsilon 20 11 + ε by 1 2 + ε \frac{1}{2}+\varepsilon 2 1 + ε in Theorem 1.10 .
(iii) It seems difficult with known techniques to prove results analogous to Theorems 1.9 and 1.10 for degree 3 3 3 L L L -functions. A bound of o ( T ) o(T) o ( T ) , as T → ∞ T\to\infty T → ∞ , for the integral of the Z Z Z -function associated to an L L L -function on [ 0 , T ] [0,T] [ 0 , T ] would imply an analogue of Hardy’s theorem that there are infinitely many zeros of that L L L -function on the critical line. This is a notoriously difficult problem for principal L L L -functions of degree greater than 2.
Proof of Proposition 1.7
In this section, we study the structure of the sum S ( χ ) S(\chi) S ( χ ) and prove Proposition 1.7 .
Three lemmas for simplifying Gauss sums
Lemma 2.1. Let χ \chi χ be a primitive character (mod p a p^a p a ) for some a > 1 a>1 a > 1 , and ψ \psi ψ a character (mod p b p^b p b ) for some b ≤ a / 2 b\leq a/2 b ≤ a /2 . For any integer ℓ \ell ℓ in the range a / 2 ≤ ℓ ≤ a − b a/2\leq\ell\leq a-b a /2 ≤ ℓ ≤ a − b let α \alpha α be an integer such that
χ ( p ℓ + 1 ) = e ( α p a − ℓ ) . \chi(p^\ell+1)=e\left(\frac{\alpha}{p^{a-\ell}}\right). χ ( p ℓ + 1 ) = e ( p a − ℓ α ) . Then
τ ( ψ χ ) = ψ ( − α ) τ ( χ ) . \tau(\psi\chi)=\psi(-\alpha)\tau(\chi). τ ( ψ χ ) = ψ ( − α ) τ ( χ ) . Proof. We first note that ( m p ℓ + 1 ) ( n p ℓ + 1 ) = m n p 2 ℓ + ( m + n ) p ℓ + 1 (mp^\ell+1)(np^\ell+1)=mnp^{2\ell}+(m+n)p^\ell+1 ( m p ℓ + 1 ) ( n p ℓ + 1 ) = mn p 2 ℓ + ( m + n ) p ℓ + 1 , and 2 ℓ ≥ a 2\ell\geq a 2 ℓ ≥ a by hypothesis, so χ ( n p ℓ + 1 ) \chi(np^\ell+1) χ ( n p ℓ + 1 ) is an additive character as a function of n n n , and more specifically
χ ( n p ℓ + 1 ) = e ( n α p a − ℓ ) . \chi(np^\ell+1)=e\left(\frac{n\alpha}{p^{a-\ell}}\right). χ ( n p ℓ + 1 ) = e ( p a − ℓ n α ) . We now split the Gauss sum τ ( ψ χ ) \tau(\psi\chi) τ ( ψ χ ) into the various arithmetic progressions (mod p ℓ p^\ell p ℓ ), writing
τ ( ψ χ ) = ∑ n = 0 p a − ℓ − 1 ∑ m = 0 p ℓ − 1 ψ ( n p ℓ + m ) χ ( n p ℓ + m ) e ( n p ℓ + m p a ) . \tau(\psi\chi)=\sum_{n=0}^{p^{a-\ell}-1}\sum_{m=0}^{p^\ell-1}\psi(np^\ell+m)\chi(np^\ell+m)e\left(\frac{np^\ell+m}{p^a}\right). τ ( ψ χ ) = n = 0 ∑ p a − ℓ − 1 m = 0 ∑ p ℓ − 1 ψ ( n p ℓ + m ) χ ( n p ℓ + m ) e ( p a n p ℓ + m ) .
As we restrict to the case where ℓ ≥ b \ell\ge b ℓ ≥ b , ψ ( n p ℓ + m ) \psi(np^\ell+ m) ψ ( n p ℓ + m ) depends only on m m m , so we move it outside the inner sum and evaluate χ ( n p ℓ + m ) \chi(np^\ell+ m) χ ( n p ℓ + m ) as χ ( m ) \chi(m) χ ( m ) times an additive character, writing this sum as
τ ( ψ χ ) = ∑ m = 1 p ∤ m p ℓ − 1 ψ ( m ) χ ( m ) ∑ n = 1 p a − ℓ χ ( m ‾ n p ℓ + 1 ) e ( n p ℓ + m p a ) = ∑ m = 1 p ∤ m p ℓ − 1 ψ ( m ) χ ( m ) e ( m / p a ) ∑ n = 1 p a − ℓ e ( m ‾ n α + n p a − ℓ ) . \begin{aligned}
\tau(\psi\chi)=\sum_{\substack{m=1\\p\nmid m}}^{p^\ell-1}\psi(m)\chi(m)\sum_{n=1}^{p^{a-\ell}}\chi(\overline{m}np^\ell+1)e\left(\frac{np^\ell+m}{p^a}\right) \\
&=\sum_{\substack{m=1\\p\nmid m}}^{p^\ell-1}\psi(m)\chi(m)e(m/p^a)\sum_{n=1}^{p^{a-\ell}}e\left(\frac{\overline{m}n\alpha+n}{p^{a-\ell}}\right).
\end{aligned} τ ( ψ χ ) = m = 1 p ∤ m ∑ p ℓ − 1 ψ ( m ) χ ( m ) n = 1 ∑ p a − ℓ χ ( m n p ℓ + 1 ) e ( p a n p ℓ + m ) = m = 1 p ∤ m ∑ p ℓ − 1 ψ ( m ) χ ( m ) e ( m / p a ) n = 1 ∑ p a − ℓ e ( p a − ℓ m n α + n ) .
where m ‾ \overline{m} m denotes the inverse of m m m modulo p a p^a p a . We now see that the innermost sum vanishes unless m ≡ − α ( m o d p a − ℓ ) m\equiv-\alpha\pmod{p^{a-\ell}} m ≡ − α ( mod p a − ℓ ) , in which case it is exactly p a − ℓ p^{a-\ell} p a − ℓ , so we find that the sum is equal to
p a − ℓ ∑ m = 1 p ∤ m m ≡ − α ( m o d p a − ℓ ) p ℓ − 1 ψ ( m ) χ ( m ) e ( m / p a ) . p^{a-\ell}\sum_{\substack{m=1\\p\nmid m\\m\equiv-\alpha\;(\mathrm{mod}\;p^{a-\ell})}}^{p^\ell-1}\psi(m)\chi(m)e(m/p^a). p a − ℓ m = 1 p ∤ m m ≡ − α ( mod p a − ℓ ) ∑ p ℓ − 1 ψ ( m ) χ ( m ) e ( m / p a ) .
Finally, we use the fact that b ≤ a − ℓ b\le a-\ell b ≤ a − ℓ , so ψ ( m ) = ψ ( − α ) \psi(m)=\psi(-\alpha) ψ ( m ) = ψ ( − α ) when m ≡ − α ( m o d p a − ℓ ) m\equiv-\alpha\pmod{p^{a-\ell}} m ≡ − α ( mod p a − ℓ ) , so the whole sum is
ψ ( − α ) p a − ℓ ∑ m = 1 p ∤ m m ≡ − α ( m o d p a − ℓ ) p ℓ − 1 χ ( m ) e ( m / p a ) . \psi(-\alpha)p^{a-\ell}\sum_{\substack{m=1\\p\nmid m\\m\equiv-\alpha\;(\mathrm{mod}\;p^{a-\ell})}}^{p^\ell-1}\chi(m)e(m/p^a). ψ ( − α ) p a − ℓ m = 1 p ∤ m m ≡ − α ( mod p a − ℓ ) ∑ p ℓ − 1 χ ( m ) e ( m / p a ) .
Considering that this expression is equally valid when ψ \psi ψ is the trivial character finishes the proof.
We extract a few simpler variants of this lemma which allow us to consider only the values χ ( p + 1 ) \chi(p+1) χ ( p + 1 ) (when p p p is odd) and χ ( 5 ) \chi(5) χ ( 5 ) (when p = 2 p=2 p = 2 ).
Lemma 2.2. Let χ \chi χ be a primitive character (mod p a p^a p a ) for p > 2 p>2 p > 2 and some a > 1 a>1 a > 1 , and ψ \psi ψ a character mod p p p . Let α \alpha α be an integer such that
χ ( p + 1 ) = e ( α p a − 1 ) . \chi(p+1)=e\left(\frac{\alpha}{p^{a-1}}\right). χ ( p + 1 ) = e ( p a − 1 α ) . Then
τ ( ψ χ ) = ψ ( − α ) τ ( χ ) . \tau(\psi\chi)=\psi(-\alpha)\tau(\chi). τ ( ψ χ ) = ψ ( − α ) τ ( χ ) . Proof. From the previous lemma, we have
τ ( ψ χ ) = ψ ( − α ′ ) τ ( χ ) , \tau(\psi\chi)=\psi(-\alpha')\tau(\chi), τ ( ψ χ ) = ψ ( − α ′ ) τ ( χ ) , where χ ( p a − 1 + 1 ) = e ( α ′ / p ) \chi(p^{a-1}+1)=e(\alpha'/p) χ ( p a − 1 + 1 ) = e ( α ′ / p ) . We need only show that α ≡ α ′ ( m o d p ) \alpha\equiv\alpha'\pmod p α ≡ α ′ ( mod p ) . The congruence
( p + 1 ) p a − 2 ≡ p a − 1 + 1 ( m o d p a ) (p+1)^{p^{a-2}}\equiv p^{a-1}+1\pmod{p^a} ( p + 1 ) p a − 2 ≡ p a − 1 + 1 ( mod p a ) follows easily by expanding using the binomial theorem, from which we see that in fact χ ( p a − 1 + 1 ) = e ( α / p ) \chi(p^{a-1}+1)=e(\alpha/p) χ ( p a − 1 + 1 ) = e ( α / p ) . □
Lemma 2.3. Let χ \chi χ be a primitive character modulo 2 a 2^a 2 a , for a > 6 a>6 a > 6 , and let ϵ \epsilon ϵ be a character (mod 8 8 8 ). Let α \alpha α be an integer such that
χ ( 5 ) = e ( α 2 a − 2 ) . \chi(5)=e\left(\frac{\alpha}{2^{a-2}}\right). χ ( 5 ) = e ( 2 a − 2 α ) . Then
τ ( ϵ χ ) = ϵ ( α ) τ ( χ ) . \tau(\epsilon\chi)=\epsilon(\alpha)\tau(\chi). τ ( ϵχ ) = ϵ ( α ) τ ( χ ) . Proof. From Lemma 2.1, we have
τ ( ϵ χ ) = ϵ ( − α ′ ) τ ( χ ) , \tau(\epsilon\chi)=\epsilon(-\alpha')\tau(\chi), τ ( ϵχ ) = ϵ ( − α ′ ) τ ( χ ) , where χ ( 2 a − 3 + 1 ) = e ( α ′ / 8 ) \chi(2^{a-3}+1)=e(\alpha'/8) χ ( 2 a − 3 + 1 ) = e ( α ′ /8 ) . For ℓ ≥ 4 \ell\geq4 ℓ ≥ 4 , applying the binomial theorem to ( 2 2 + 1 ) 7 ⋅ 2 ℓ − 2 (2^2+1)^{7\cdot2^{\ell-2}} ( 2 2 + 1 ) 7 ⋅ 2 ℓ − 2 , we find the congruence 5 7 ⋅ 2 ℓ − 2 ≡ 2 ℓ + 1 ( m o d 2 ℓ + 3 ) 5^{7\cdot2^{\ell-2}}\equiv2^\ell+1\pmod{2^{\ell+3}} 5 7 ⋅ 2 ℓ − 2 ≡ 2 ℓ + 1 ( mod 2 ℓ + 3 ) . As the multiplicative order of 5 ( m o d 2 ℓ + 3 ) 5\pmod{2^{\ell+3}} 5 ( mod 2 ℓ + 3 ) is 2 ℓ + 1 2^{\ell+1} 2 ℓ + 1 , we find that the discrete logarithm of 2 ℓ + 1 2^\ell+1 2 ℓ + 1 with respect to 5 ( m o d 2 a ) 5\pmod{2^a} 5 ( mod 2 a ) must be congruent to 7 ⋅ 2 ℓ − 2 ( m o d 2 ℓ + 1 ) 7\cdot2^{\ell-2}\pmod{2^{\ell+1}} 7 ⋅ 2 ℓ − 2 ( mod 2 ℓ + 1 ) for any a ≥ ℓ + 3 a\geq\ell+3 a ≥ ℓ + 3 . This means that
χ ( 2 ℓ + 1 ) = χ ( 5 ) 7 ⋅ 2 ℓ − 2 + k 2 ℓ + 1 \chi(2^\ell+1)=\chi(5)^{7\cdot2^{\ell-2}+k2^{\ell+1}} χ ( 2 ℓ + 1 ) = χ ( 5 ) 7 ⋅ 2 ℓ − 2 + k 2 ℓ + 1 for some integer k k k . That is,
χ ( 2 ℓ + 1 ) = e ( 7 α + 8 k α 2 a − ℓ ) . \chi(2^\ell+1)=e\left(\frac{7\alpha+8k\alpha}{2^{a-\ell}}\right). χ ( 2 ℓ + 1 ) = e ( 2 a − ℓ 7 α + 8 k α ) . When ℓ = a − 3 \ell=a-3 ℓ = a − 3 , this is simply
χ ( 2 a − 3 + 1 ) = e ( 7 α 8 ) , \chi(2^{a-3}+1)=e\left(\frac{7\alpha}{8}\right), χ ( 2 a − 3 + 1 ) = e ( 8 7 α ) , and now Lemma 2.1 finishes the proof.
An auxiliary sum
We will evaluate S ( χ ) S(\chi) S ( χ ) by relating it to the sum
S 2 ( ψ , r ) = ∑ n = 1 q 2 ψ ( n 2 ) e ( r n 2 q 2 ) . S_2(\psi,r)=\sum_{n=1}^{q_2}\psi(n^2)e\left(\frac{rn^2}{q_2}\right). S 2 ( ψ , r ) = n = 1 ∑ q 2 ψ ( n 2 ) e ( q 2 r n 2 ) .
where ψ \psi ψ is a primitive character (mod q 2 q_2 q 2 ) and r r r is coprime to q 2 q_2 q 2 . Via a standard application of the Chinese Remainder Theorem, we can factor S 2 ( ψ , r ) S_2(\psi,r) S 2 ( ψ , r ) as
S 2 ( ψ , r ) = ∏ p ψ p ( q 2 / p e p ) 2 S 2 ( ψ p , r q 2 / p e p ) . (13) S_2(\psi,r)=\prod_p\psi_p\left(q_2/p^{e_p}\right)^2S_2\left(\psi_p,rq_2/p^{e_p}\right). \tag*{(13)} S 2 ( ψ , r ) = p ∏ ψ p ( q 2 / p e p ) 2 S 2 ( ψ p , r q 2 / p e p ) . (13)
where q 2 = ∏ p p e p q_2=\prod_p p^{e_p} q 2 = ∏ p p e p and ψ = ∏ p ψ p \psi=\prod_p\psi_p ψ = ∏ p ψ p . As one might already guess from examining the formulae in Proposition 1.7, the relationship between S ( χ ) S(\chi) S ( χ ) and S 2 ( ψ , r ) S_2(\psi,r) S 2 ( ψ , r ) comes from the fact that when χ \chi χ is a primitive even character (mod q q q ) and q q q is even and there is a primitive character ψ \psi ψ (mod 2 q 2q 2 q ) such that ψ 2 = χ \psi^2=\chi ψ 2 = χ , we will have S ( χ ) = S 2 ( ψ , − 1 ) S(\chi)=S_2(\psi,-1) S ( χ ) = S 2 ( ψ , − 1 ) for this character ψ \psi ψ (mod q 2 q_2 q 2 ), where q 2 = 2 q q_2=2q q 2 = 2 q . If q q q is odd or χ \chi χ does not have a square root then we will see that it is easy to show that S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 , so anytime S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 we will be able to evaluate it this way.
It is straightforward to rewrite S 2 S_2 S 2 as a sum of Gauss sums. A sum over the quadratic characters (mod q 2 q_2 q 2 ) will detect squares, and count them with the correct multiplicity, and thus
S 2 ( ψ , r ) = ∑ n = 1 q 2 ( ∑ ξ 2 = χ 0 ξ ( n ) ) ψ ( n ) e ( r n q 2 ) = ∑ ξ 2 = χ 0 ∑ n = 1 q 2 ξ ψ ( n ) e ( r n q 2 ) = ∑ ξ 2 = χ 0 ξ ψ ‾ ( r ) τ ( ξ ψ ) . S_2(\psi,r)=\sum_{n=1}^{q_2}\left(\sum_{\substack{\xi^2=\chi_0}}\xi(n)\right)\psi(n)e\left(\frac{rn}{q_2}\right)=\sum_{\xi^2=\chi_0}\sum_{n=1}^{q_2}\xi\psi(n)e\left(\frac{rn}{q_2}\right)=\sum_{\xi^2=\chi_0}\overline{\xi\psi}(r)\tau(\xi\psi). S 2 ( ψ , r ) = n = 1 ∑ q 2 ξ 2 = χ 0 ∑ ξ ( n ) ψ ( n ) e ( q 2 r n ) = ξ 2 = χ 0 ∑ n = 1 ∑ q 2 ξ ψ ( n ) e ( q 2 r n ) = ξ 2 = χ 0 ∑ ξ ψ ( r ) τ ( ξ ψ ) .
This formula does not seem particularly elucidating, however, and we find it better to use the factorization and do this one prime at a time. Determining when the sum vanishes still comes down to studying Gauss sums, but only for prime power moduli.
Even prime powers
When q = 2 a q = 2^{a} q = 2 a is an even prime power (and a a a is large enough), the quadratic characters are precisely the characters (mod 8). Studying Gauss sums in this case will lead us to the following evaluation.
Proposition 2.4. Let a , r ∈ Z + a,r \in\mathbb{Z}^{+} a , r ∈ Z + with r r r odd. Let ψ 2 \psi_{2} ψ 2 be a primitive character (mod 2 a 2^{a} 2 a ) and α \alpha α be an integer such that ψ 2 ( 5 ) = e ( α / 2 a − 2 ) \psi_{2}(5) = e(\alpha/2^{a-2}) ψ 2 ( 5 ) = e ( α / 2 a − 2 ) . Then we have the following:
(a) If (i) a < 5 a < 5 a < 5 , or (ii) a = 5 a = 5 a = 5 and r ≡ α ( m o d 4 ) r \equiv\alpha\pmod{4} r ≡ α ( mod 4 ) , or (iii) a = 6 a = 6 a = 6 and r ≡ 5 α ( m o d 8 ) r \equiv5\alpha\pmod{8} r ≡ 5 α ( mod 8 ) , then
S 2 ( ψ 2 , r ) = 2 a − 1 e ( r / 2 a ) . S_{2}(\psi_{2},r) = 2^{a-1}e(r/2^{a}). S 2 ( ψ 2 , r ) = 2 a − 1 e ( r / 2 a ) . (b) If a > 6 a > 6 a > 6 and r ≡ α ( m o d 8 ) r \equiv\alpha\pmod{8} r ≡ α ( mod 8 ) , then S 2 ( ψ 2 , r ) = 4 ψ ‾ 2 ( r ) τ ( ψ 2 ) S_{2}(\psi_{2},r) = 4\overline{\psi}_{2}(r)\tau(\psi_{2}) S 2 ( ψ 2 , r ) = 4 ψ 2 ( r ) τ ( ψ 2 ) .
In all other cases, S 2 ( ψ 2 , r ) = 0 S_{2}(\psi_{2},r) = 0 S 2 ( ψ 2 , r ) = 0 .
Proof. There are only a small number of cases where a ≤ 6 a \leq6 a ≤ 6 , so this can be done using a small computer computation which we do not include. Otherwise, as described above, this sum is
S 2 ( ψ 2 , r ) = ∑ ξ ( m o d 8 ) ∑ n = 1 2 a ξ ψ 2 ( n ) e ( r n / 2 a ) = ∑ ξ ( m o d 8 ) ξ ψ ‾ 2 ( r ) τ ( ξ ψ 2 ) . S_{2}(\psi_{2},r) = \sum_{\substack{\xi\;(\bmod\;8)}}\sum_{n=1}^{2^{a}}\xi\psi_{2}(n)e(rn/2^{a}) = \sum_{\xi\;(\bmod\;8)}\overline{\xi\psi}_{2}(r)\tau(\xi\psi_{2}). S 2 ( ψ 2 , r ) = ξ ( mod 8 ) ∑ n = 1 ∑ 2 a ξ ψ 2 ( n ) e ( r n / 2 a ) = ξ ( mod 8 ) ∑ ξ ψ 2 ( r ) τ ( ξ ψ 2 ) .
It now follows from Lemma 2.3 that
S 2 ( ψ 2 , r ) = ψ ‾ 2 ( r ) τ ( ψ 2 ) ∑ ξ ( m o d 8 ) ξ ‾ ( r ) ξ ( α ) . S_{2}(\psi_{2},r) = \overline{\psi}_{2}(r)\tau(\psi_{2})\sum_{\xi\;(\bmod\;8)}\overline{\xi}(r)\xi(\alpha). S 2 ( ψ 2 , r ) = ψ 2 ( r ) τ ( ψ 2 ) ξ ( mod 8 ) ∑ ξ ( r ) ξ ( α ) .
By the orthogonality of characters (mod 8), this equals 4 ψ ‾ 2 ( r ) τ ( ψ 2 ) 4\overline{\psi}_{2}(r)\tau(\psi_{2}) 4 ψ 2 ( r ) τ ( ψ 2 ) if r ≡ α ( m o d 8 ) r \equiv\alpha\pmod{8} r ≡ α ( mod 8 ) and 0 otherwise. This completes the proof of the proposition. □
Odd prime powers
For any odd prime power, there is exactly one quadratic character, and thus the sum for characters to such moduli is just
S 2 ( ψ p , r ) = ψ ‾ p ( r ) ( τ ( ψ p ) + ϵ p ( r ) τ ( ϵ p ψ p ) ) . S_{2}(\psi_{p},r) = \overline{\psi}_{p}(r)\left(\tau(\psi_{p})+\epsilon_{p}(r)\tau(\epsilon_{p}\psi_{p})\right). S 2 ( ψ p , r ) = ψ p ( r ) ( τ ( ψ p ) + ϵ p ( r ) τ ( ϵ p ψ p ) ) .
The next proposition says that this sum does not vanish when the modulus of the character is a prime, and describes when it does vanish for higher powers.
Proposition 2.5. Given a ∈ Z + a \in\mathbb{Z}^{+} a ∈ Z + , let ψ p \psi_{p} ψ p be a primitive character modulo p a p^{a} p a , for a prime p > 2 p > 2 p > 2 and ( r , p ) = 1 (r,p) = 1 ( r , p ) = 1 . If a = 1 a = 1 a = 1 , then S 2 ( ψ p , r ) ≠ 0 S_{2}(\psi_{p},r) \ne0 S 2 ( ψ p , r ) = 0 . Otherwise, let α \alpha α be an integer such that ψ p ( p + 1 ) = e ( α / p a − 1 ) \psi_{p}(p+1) = e(\alpha/p^{a-1}) ψ p ( p + 1 ) = e ( α / p a − 1 ) . Then
S 2 ( ψ p , r ) = ( 1 + ( − α r p ) ) ψ ‾ p ( r ) τ ( ψ p ) . S_{2}(\psi_{p},r) = \left(1+\left(\frac{-\alpha r}{p}\right)\right)\overline{\psi}_{p}(r)\tau(\psi_{p}). S 2 ( ψ p , r ) = ( 1 + ( p − α r ) ) ψ p ( r ) τ ( ψ p ) .
Proof. We consider
S 2 ( ψ p , r ) = ψ ‾ p ( r ) ( τ ( ψ p ) + ϵ p ( r ) τ ( ϵ p ψ p ) ) . (14) S_{2}(\psi_{p},r)=\overline{\psi}_{p}(r)\left(\tau(\psi_{p})+\epsilon_{p}(r)\tau(\epsilon_{p}\psi_{p})\right). \tag*{(14)} S 2 ( ψ p , r ) = ψ p ( r ) ( τ ( ψ p ) + ϵ p ( r ) τ ( ϵ p ψ p ) ) . (14)
In the case a = 1 a = 1 a = 1 , both ψ p \psi_{p} ψ p and ϵ p \epsilon_{p} ϵ p are primitive characters (mod p p p ). In this case, we have two proofs that τ ( ψ p ) ≠ ± τ ( ϵ p ψ p ) \tau(\psi_{p}) \ne\pm\tau(\epsilon_{p}\psi_{p}) τ ( ψ p ) = ± τ ( ϵ p ψ p ) which implies that S 2 ( ψ p , r ) ≠ 0 S_{2}(\psi_{p},r) \ne0 S 2 ( ψ p , r ) = 0 . If ψ p = ϵ p \psi_{p} = \epsilon_{p} ψ p = ϵ p , then ∣ τ ( ψ p ) ∣ = p |\tau(\psi_{p})| = \sqrt{p} ∣ τ ( ψ p ) ∣ = p while τ ( ψ p ϵ p ) = − 1 \tau(\psi_{p}\epsilon_{p}) = -1 τ ( ψ p ϵ p ) = − 1 , so it is clear that S 2 ( ψ p , r ) ≠ 0 S_{2}(\psi_{p},r) \ne0 S 2 ( ψ p , r ) = 0 . Therefore, we may assume that ψ p ≠ ϵ p \psi_{p} \ne\epsilon_{p} ψ p = ϵ p .
One short and indirect proof relies on a beautiful generalization of the Chowla–Mordell theorem due to Evans. The Chowla–Mordell theorem states that if χ \chi χ is a primitive character (mod p p p ), then τ ( χ ) / p \tau(\chi)/\sqrt{p} τ ( χ ) / p is a root of unity if and only if χ \chi χ is quadratic. If χ 1 \chi_{1} χ 1 and χ 2 \chi_{2} χ 2 are primitive characters (mod p p p ), then Evans [9 Theorem 1] proved that τ ( χ 1 ) / τ ( χ 2 ) \tau(\chi_{1})/\tau(\chi_{2}) τ ( χ 1 ) / τ ( χ 2 ) is a root of unity if and only if χ 1 = χ 2 \chi_{1} = \chi_{2} χ 1 = χ 2 . If ψ p ≠ ϵ p \psi_{p} \ne\epsilon_{p} ψ p = ϵ p , then ψ p \psi_{p} ψ p and ϵ p ψ p \epsilon_{p}\psi_{p} ϵ p ψ p are distinct primitive characters (mod p p p ). Therefore, Evans’ theorem implies that τ ( ψ p ) / τ ( ϵ p ψ p ) ≠ ± 1 \tau(\psi_{p})/\tau(\epsilon_{p}\psi_{p}) \ne\pm1 τ ( ψ p ) / τ ( ϵ p ψ p ) = ± 1 . Hence S 2 ( ψ p , r ) ≠ 0 S_{2}(\psi_{p},r) \ne0 S 2 ( ψ p , r ) = 0 . We also provide a direct proof in the case a = 1 a=1 a = 1 using properties of Jacobi sums and cyclotomic fields. Again, since ψ p ≠ ϵ p \psi_p \ne\epsilon_p ψ p = ϵ p , we know that ψ p \psi_p ψ p and ϵ p ψ p \epsilon_p\psi_p ϵ p ψ p are distinct primitive characters (mod p p p ). We can express the Gauss sums τ ( ψ p ) \tau(\psi_p) τ ( ψ p ) and τ ( ϵ p ψ p ) \tau(\epsilon_p\psi_p) τ ( ϵ p ψ p ) in terms of the Jacobi sum
J ( ψ p , ϵ p ) = ∑ n = 1 p ψ p ( n ) ϵ p ( 1 − n ) = τ ( ψ p ) τ ( ϵ p ) τ ( ϵ p ψ p ) . (15) J(\psi_p,\epsilon_p)=\sum_{n=1}^{p}\psi_p(n)\epsilon_p(1-n)=\frac{\tau(\psi_p)\tau(\epsilon_p)}{\tau(\epsilon_p\psi_p)}. \tag*{(15)} J ( ψ p , ϵ p ) = n = 1 ∑ p ψ p ( n ) ϵ p ( 1 − n ) = τ ( ϵ p ψ p ) τ ( ψ p ) τ ( ϵ p ) . (15)
Evaluating the Gauss sum for the quadratic character and rearranging, we find that
τ ( ϵ p ψ p ) τ ( ψ p ) = ± p J ( ψ p , ϵ p ) . \frac{\tau(\epsilon_p\psi_p)}{\tau(\psi_p)}=\frac{\sqrt{\pm p}}{J(\psi_p,\epsilon_p)}. τ ( ψ p ) τ ( ϵ p ψ p ) = J ( ψ p , ϵ p ) ± p .
The ratio on the right-hand side cannot be ± 1 \pm1 ± 1 because the Jacobi sum is an element of the cyclotomic field Q ( ζ p − 1 ) \mathbb{Q}(\zeta_{p-1}) Q ( ζ p − 1 ) and the elements ± p \sqrt{\pm p} ± p are not. Hence τ ( ψ p ) / τ ( ϵ p ψ p ) ≠ ± 1 \tau(\psi_p)/\tau(\epsilon_p\psi_p)\ne\pm1 τ ( ψ p ) / τ ( ϵ p ψ p ) = ± 1 and S 2 ( ψ p , r ) ≠ 0 S_2(\psi_p,r)\ne0 S 2 ( ψ p , r ) = 0 .
In the case where a > 1 a>1 a > 1 , we use Lemma 2.2 to simplify τ ( ϵ p ψ p ) \tau(\epsilon_p\psi_p) τ ( ϵ p ψ p ) , giving
S 2 ( ψ p , r ) = ψ p ‾ ( r ) ( τ ( ψ p ) + ϵ p ( − α r ) τ ( ψ p ) ) . S_2(\psi_p,r)=\overline{\psi_p}(r)\left(\tau(\psi_p)+\epsilon_p(-\alpha r)\tau(\psi_p)\right). S 2 ( ψ p , r ) = ψ p ( r ) ( τ ( ψ p ) + ϵ p ( − α r ) τ ( ψ p ) ) .
This completes the proof of the proposition. □
Remark. We can elaborate on the relationship between S 2 ( ψ p , r ) S_2(\psi_p,r) S 2 ( ψ p , r ) and Jacobi sums. Consider
S ± = ψ p ( 4 ) τ ( ϵ p ) ⋅ ( τ ( ψ p ) ± τ ( ϵ p ψ p ) ) 2 τ ( ψ p 2 ) . (16) S_{\pm}=\frac{\psi_p(4)}{\tau(\epsilon_p)}\cdot\frac{\left(\tau(\psi_p)\pm\tau(\epsilon_p\psi_p)\right)^2}{\tau(\psi_p^2)}. \tag*{(16)} S ± = τ ( ϵ p ) ψ p ( 4 ) ⋅ τ ( ψ p 2 ) ( τ ( ψ p ) ± τ ( ϵ p ψ p ) ) 2 . (16)
By [1 Theorem 2.1.4] , if p > 2 p>2 p > 2 , for nontrivial, non-quadratic ψ p \psi_p ψ p and quadratic ϵ p \epsilon_p ϵ p (mod p p p ), we have
ψ p ( 4 ) J ( ψ p , ψ p ) = J ( ψ p , ϵ p ) . \psi_p(4)J(\psi_p,\psi_p)=J(\psi_p,\epsilon_p). ψ p ( 4 ) J ( ψ p , ψ p ) = J ( ψ p , ϵ p ) .
Applying this identity and (15) in a short calculation, we obtain that
S ± = ψ p ( 4 ) τ ( ϵ p ) ⋅ τ ( ψ p ) 2 τ ( ψ p 2 ) ( 1 ± τ ( ϵ p ψ p ) τ ( ψ p ) ) 2 = ψ p ( 4 ) J ( ψ p , ψ p ) τ ( ϵ p ) ( 1 ± τ ( ϵ p ) J ( ψ p , ϵ p ) ) 2 = J ( ψ p , ϵ p ) τ ( ϵ p ) ⋅ ( 1 ± τ ( ϵ p ) J ( ψ p , ϵ p ) ) 2 = J ( ψ p , ϵ p ) τ ( ϵ p ) ± 2 + τ ( ϵ p ) J ( ψ p , ϵ p ) = 2 Re ( τ ( ψ p ) τ ( ϵ p ψ p ) ) ± 2. (17) \begin{aligned}
S_{\pm}
&=\frac{\psi_p(4)}{\tau(\epsilon_p)}\cdot\frac{\tau(\psi_p)^2}{\tau(\psi_p^2)}\left(1\pm\frac{\tau(\epsilon_p\psi_p)}{\tau(\psi_p)}\right)^2 \\
&=\frac{\psi_p(4)J(\psi_p,\psi_p)}{\tau(\epsilon_p)}\left(1\pm\frac{\tau(\epsilon_p)}{J(\psi_p,\epsilon_p)}\right)^2 \\
&=\frac{J(\psi_p,\epsilon_p)}{\tau(\epsilon_p)}\cdot\left(1\pm\frac{\tau(\epsilon_p)}{J(\psi_p,\epsilon_p)}\right)^2 \\
&=\frac{J(\psi_p,\epsilon_p)}{\tau(\epsilon_p)}\pm2+\frac{\tau(\epsilon_p)}{J(\psi_p,\epsilon_p)}
=2\operatorname{Re}\left(\frac{\tau(\psi_p)}{\tau(\epsilon_p\psi_p)}\right)\pm2. \tag*{(17)}
\end{aligned} S ± = τ ( ϵ p ) ψ p ( 4 ) ⋅ τ ( ψ p 2 ) τ ( ψ p ) 2 ( 1 ± τ ( ψ p ) τ ( ϵ p ψ p ) ) 2 = τ ( ϵ p ) ψ p ( 4 ) J ( ψ p , ψ p ) ( 1 ± J ( ψ p , ϵ p ) τ ( ϵ p ) ) 2 = τ ( ϵ p ) J ( ψ p , ϵ p ) ⋅ ( 1 ± J ( ψ p , ϵ p ) τ ( ϵ p ) ) 2 = τ ( ϵ p ) J ( ψ p , ϵ p ) ± 2 + J ( ψ p , ϵ p ) τ ( ϵ p ) = 2 Re ( τ ( ϵ p ψ p ) τ ( ψ p ) ) ± 2. (17)
By work of Katz [22 Chapter 9] , this ratio of Gauss sums becomes uniformly distributed on the unit circle as p → ∞ p\to\infty p → ∞ , where ψ p \psi_p ψ p runs over the non-principal, non-quadratic characters modulo p p p , which in turn suggests something about the distribution of the size of S 2 ( ψ p , r ) S_2(\psi_p,r) S 2 ( ψ p , r ) . We will return to this in Section 5.
For reference, we summarize what the results of Proposition 2.4 and Proposition 2.5 imply about nonvanishing of S 2 ( ψ , r ) S_2(\psi,r) S 2 ( ψ , r ) once we apply the Chinese Remainder Theorem factorization in (13) .
Proposition 2.6. Let ψ = ∏ p ψ p \psi=\prod_p\psi_p ψ = ∏ p ψ p be a primitive character (mod q 2 q_2 q 2 ), where q 2 = ∏ p p e p q_2=\prod_p p^{e_p} q 2 = ∏ p p e p and each factor ψ p \psi_p ψ p is a primitive character (mod p e p p^{e_p} p e p ), and let r r r be an integer with ( r , q 2 ) = 1 (r,q_2)=1 ( r , q 2 ) = 1 . Set β 2 \beta_2 β 2 to be an integer such that ψ 2 ( 5 ) = e ( β 2 / 2 e 2 − 2 ) \psi_2(5)=e(\beta_2/2^{e_2-2}) ψ 2 ( 5 ) = e ( β 2 / 2 e 2 − 2 ) , and for each odd prime p p p set β p \beta_p β p to be an integer such that ψ p ( p + 1 ) = e ( β p / p e p − 1 ) \psi_p(p+1)=e(\beta_p/p^{e_p-1}) ψ p ( p + 1 ) = e ( β p / p e p − 1 ) . Then S 2 ( ψ , r ) ≠ 0 S_2(\psi,r)\ne0 S 2 ( ψ , r ) = 0 if and only if
(a) (i) e 2 < 5 e_2<5 e 2 < 5 , or (ii) e 2 = 5 e_2=5 e 2 = 5 and β 2 ≡ r q 2 / 2 5 \beta_2\equiv rq_2/2^5 β 2 ≡ r q 2 / 2 5 (mod 4), or (iii) e 2 = 6 e_2=6 e 2 = 6 and β 2 ≡ 5 r q 2 / 2 6 \beta_2\equiv5rq_2/2^6 β 2 ≡ 5 r q 2 / 2 6 (mod 8), or (iv) e 2 > 6 e_2>6 e 2 > 6 and β 2 ≡ r q 2 / 2 e 2 \beta_2\equiv rq_2/2^{e_2} β 2 ≡ r q 2 / 2 e 2 (mod 8); and
(b) for each odd prime p p p such that p 2 p^2 p 2 divides q 2 q_2 q 2 , the residue − β p r q 2 / p e p -\beta_prq_2/p^{e_p} − β p r q 2 / p e p is a square modulo $p.
When the sum is nonzero, our proof will give an explicit evaluation in terms of Gauss sums. It is tedious to write out all of the cases when e 2 e_2 e 2 is small, but when e 2 > 6 e_2 > 6 e 2 > 6 we have
S 2 ( ψ , r ) = 2 ν ( q 2 ) + 1 ∏ p ∣ q 2 p 2 ∤ q 2 ψ ( q 2 r ‾ p e p ) ∏ p 2 ∣ q 2 τ ( ψ p ) ∏ p ∣ q 2 ( τ ( ψ p ) + ϵ p ( r q 2 p ) τ ( ϵ p ψ p ) ) . (18) S_2(\psi,r)=2^{\nu(q_2)+1}\prod_{\substack{p\mid q_2\\p^2\nmid q_2}}\psi\left(\frac{q_2\overline{r}}{p^{e_p}}\right)\prod_{p^2\mid q_2}\tau(\psi_p)\prod_{p\mid q_2}\left(\tau(\psi_p)+\epsilon_p\left(\frac{rq_2}{p}\right)\tau(\epsilon_p\psi_p)\right). \tag*{(18)} S 2 ( ψ , r ) = 2 ν ( q 2 ) + 1 p ∣ q 2 p 2 ∤ q 2 ∏ ψ ( p e p q 2 r ) p 2 ∣ q 2 ∏ τ ( ψ p ) p ∣ q 2 ∏ ( τ ( ψ p ) + ϵ p ( p r q 2 ) τ ( ϵ p ψ p ) ) . (18)
where ϵ p \epsilon_p ϵ p is the quadratic character (mod p p p ) and ν ( q 2 ) \nu(q_2) ν ( q 2 ) is the number of primes p p p dividing q 2 q_2 q 2 such that p 2 p^2 p 2 divides q 2 q_2 q 2 .
Evaluation of S ( χ ) S(\chi) S ( χ )
Let χ \chi χ (mod q q q ) be a primitive character. Recall that we have already seen that S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 if q q q is odd or χ \chi χ is odd.
Lemma 2.7. Let q ∈ N q\in\mathbb{N} q ∈ N be even and let χ \chi χ (mod q q q ) be a primitive character. Then
S ( χ ) = ∑ ψ ( m o d 2 q ) ψ 2 = χ ψ ‾ ( − 1 ) τ ( ψ ) . S(\chi)=\sum_{\substack{\psi\;(\mathrm{mod}\;2q)\\\psi^2=\chi}}\overline{\psi}(-1)\tau(\psi). S ( χ ) = ψ ( mod 2 q ) ψ 2 = χ ∑ ψ ( − 1 ) τ ( ψ ) . In particular, if S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 , then there is a primitive character ψ \psi ψ (mod 2 q 2q 2 q ) such that χ = ψ 2 \chi=\psi^2 χ = ψ 2 .
Proof. If ( a , b ) = 1 (a,b)=1 ( a , b ) = 1 , then
e ( a / b ) = 1 φ ( b ) ∑ ξ ( m o d b ) τ ( ξ ) ξ ‾ ( a ) . e(a/b)=\frac{1}{\varphi(b)}\sum_{\xi\;(\mathrm{mod}\;b)}\tau(\xi)\overline{\xi}(a). e ( a / b ) = φ ( b ) 1 ξ ( mod b ) ∑ τ ( ξ ) ξ ( a ) .
Since q q q is even, we have
S ( χ ) = ∑ n = 1 ( n , q ) = 1 2 q χ ( n ) e ( − n 2 2 q ) = ∑ n = 1 2 q χ ( n ) ( 1 φ ( 2 q ) ∑ ψ ( m o d 2 q ) ψ ‾ ( − n 2 ) τ ( ψ ) ) = ∑ ψ ( m o d 2 q ) ψ ‾ ( − 1 ) τ ( ψ ) ( 1 φ ( 2 q ) ∑ n = 1 2 q ( χ ⋅ ψ ‾ 2 ) ( n ) ) . \begin{aligned}
S(\chi)=\sum_{\substack{n=1\\(n,q)=1}}^{2q}\chi(n)e\left(\frac{-n^2}{2q}\right)
=\sum_{n=1}^{2q}\chi(n)\left(\frac{1}{\varphi(2q)}\sum_{\psi\;(\mathrm{mod}\;2q)}\overline{\psi}(-n^2)\tau(\psi)\right) \\
&=\sum_{\psi\;(\mathrm{mod}\;2q)}\overline{\psi}(-1)\tau(\psi)\left(\frac{1}{\varphi(2q)}\sum_{n=1}^{2q}(\chi\cdot\overline{\psi}^{\,2})(n)\right).
\end{aligned} S ( χ ) = n = 1 ( n , q ) = 1 ∑ 2 q χ ( n ) e ( 2 q − n 2 ) = n = 1 ∑ 2 q χ ( n ) φ ( 2 q ) 1 ψ ( mod 2 q ) ∑ ψ ( − n 2 ) τ ( ψ ) = ψ ( mod 2 q ) ∑ ψ ( − 1 ) τ ( ψ ) ( φ ( 2 q ) 1 n = 1 ∑ 2 q ( χ ⋅ ψ 2 ) ( n ) ) .
The inner sum over n n n equals zero unless χ = ψ 2 \chi=\psi^2 χ = ψ 2 , in which case it equals φ ( 2 q ) \varphi(2q) φ ( 2 q ) . Therefore
S ( χ ) = ∑ ψ ( m o d 2 q ) ψ 2 = χ ψ ‾ ( − 1 ) τ ( ψ ) . S(\chi)=\sum_{\substack{\psi\;(\mathrm{mod}\;2q)\\\psi^2=\chi}}\overline{\psi}(-1)\tau(\psi). S ( χ ) = ψ ( mod 2 q ) ψ 2 = χ ∑ ψ ( − 1 ) τ ( ψ ) .
It is straightforward to show that if χ \chi χ is primitive (mod q q q ) with q q q even and if ψ 2 = χ \psi^2=\chi ψ 2 = χ , then it must be the case that ψ \psi ψ is primitive (mod 2 q 2q 2 q ). This proves the lemma. □ \square □
We are now in a position to prove Proposition 1.7.
Proof of Proposition 1.7. Let q ∈ N q\in\mathbb{N} q ∈ N be even and write q = 2 e 2 ∏ p ∣ q p e p q=2^{e_2}\prod_{p\mid q}p^{e_p} q = 2 e 2 ∏ p ∣ q p e p , where the product runs over distinct odd primes p p p . If χ \chi χ (mod q q q ) is a primitive character with S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 , then Lemma 2.7 implies that there is a primitive character ψ \psi ψ (mod 2 q 2q 2 q ) with χ = ψ 2 \chi=\psi^2 χ = ψ 2 and therefore
S ( χ ) = S ( ψ 2 ) = S 2 ( ψ , − 1 ) . S(\chi)=S(\psi^2)=S_2(\psi,-1). S ( χ ) = S ( ψ 2 ) = S 2 ( ψ , − 1 ) .
Conversely, suppose that (i) and (iii) hold. Then χ \chi χ is the square of a character ψ \psi ψ (mod $2q$), since for odd p p p an even character (mod p e p p^{e_p} p e p ) is the square of a character (mod p e p p^{e_p} p e p ), and an even character (mod 2 e 2 2^{e_2} 2 e 2 ) is the square of a character (mod 2 e 2 + 1 2^{e_2+1} 2 e 2 + 1 ). As noted in the proof of Lemma 2.7, ψ \psi ψ is primitive, so again S ( χ ) = S 2 ( ψ , − 1 ) S(\chi)=S_2(\psi,-1) S ( χ ) = S 2 ( ψ , − 1 ) . Write ψ = ∏ p ψ p \psi=\prod_p\psi_p ψ = ∏ p ψ p and χ = ∏ p χ p \chi=\prod_p\chi_p χ = ∏ p χ p , the products running over all primes p p p . By unique factorization of characters, we have χ p = ψ p 2 \chi_p=\psi_p^2 χ p = ψ p 2 for each p p p dividing $q. For odd primes p p p , when e p ≥ 2 e_p \ge2 e p ≥ 2 , we define
χ p ( p + 1 ) = e ( α p p e p − 1 ) and ψ p ( p + 1 ) = e ( β p p e p − 1 ) . \chi_p(p+1)=e\left(\frac{\alpha_p}{p^{e_p-1}}\right) \quad\text{and}\quad\psi_p(p+1)=e\left(\frac{\beta_p}{p^{e_p-1}}\right). χ p ( p + 1 ) = e ( p e p − 1 α p ) and ψ p ( p + 1 ) = e ( p e p − 1 β p ) .
As ψ p ( p + 1 ) 2 = χ p ( p + 1 ) \psi_p(p+1)^2=\chi_p(p+1) ψ p ( p + 1 ) 2 = χ p ( p + 1 ) , we see that there is only a single choice α p ≡ 2 β p ( m o d p ) \alpha_p\equiv2\beta_p\pmod{p} α p ≡ 2 β p ( mod p ) for odd p p p . In this case, part (b) of Proposition 2.6 implies part (ii) of Proposition 1.7 . When e p = 1 e_p=1 e p = 1 , we note that χ p \chi_p χ p is a square if and only if χ p \chi_p χ p is even which implies part (i) of Proposition 1.7 .
At the prime 2, we note that ψ 2 \psi_2 ψ 2 is a primitive character (mod 2 e 2 + 1 2^{e_2+1} 2 e 2 + 1 ) and we write
χ 2 ( 5 ) = e ( α 2 2 e 2 − 2 ) and ψ 2 ( 5 ) = e ( β 2 2 e 2 − 1 ) . \chi_2(5)=e\left(\frac{\alpha_2}{2^{e_2-2}}\right) \quad\text{and}\quad\psi_2(5)=e\left(\frac{\beta_2}{2^{e_2-1}}\right). χ 2 ( 5 ) = e ( 2 e 2 − 2 α 2 ) and ψ 2 ( 5 ) = e ( 2 e 2 − 1 β 2 ) .
Since χ 2 = ψ 2 2 \chi_2=\psi_2^2 χ 2 = ψ 2 2 , we have the possibilities β 2 ≡ α 2 ( m o d 2 e 2 − 1 ) \beta_2\equiv\alpha_2\pmod{2^{e_2-1}} β 2 ≡ α 2 ( mod 2 e 2 − 1 ) and β 2 ≡ α 2 + 2 e 2 − 2 ( m o d 2 e 2 − 1 ) \beta_2\equiv\alpha_2+2^{e_2-2}\pmod{2^{e_2-1}} β 2 ≡ α 2 + 2 e 2 − 2 ( mod 2 e 2 − 1 ) . If e 2 = 3 e_2=3 e 2 = 3 , then we are not concerned about the value of β 2 \beta_2 β 2 ; if e 2 = 4 e_2=4 e 2 = 4 , then the value of β 2 \beta_2 β 2 is determined (mod 4); if e 2 > 4 e_2>4 e 2 > 4 then the value of β 2 \beta_2 β 2 is determined (mod 8). In every case, we may apply part (a) of Proposition 2.6 to deduce part (iii) of Proposition 1.7 .
Finally, the expressions for S e v e n S_{\mathrm{even}} S even and S o d d S_{\mathrm{odd}} S odd in (11) and (12) , respectively, follow immediately from Proposition 2.4 and Proposition 2.5 . This completes the proof. □
Proof of Theorem 1.5 and Corollary 1.6
In this section, we count the number of primitive characters χ \chi χ for which S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 . Recall that N ( q ) N(q) N ( q ) is the number of primitive χ \chi χ (mod q q q ) such that S ( χ ) ≠ 0 S(\chi)\ne0 S ( χ ) = 0 . This is not a multiplicative function1 , but from the factorization of the sum using the Chinese Remainder Theorem in (13) , it is closely related to one.
As we have evaluated S ( χ ) S(\chi) S ( χ ) by relating it to S 2 ( ψ , r ) S_2(\psi,r) S 2 ( ψ , r ) , it will similarly be convenient to first count the number of primitive ψ \psi ψ (mod q q q ) such that S 2 ( ψ , r ) ≠ 0 S_2(\psi,r)\ne0 S 2 ( ψ , r ) = 0 , which we label N 2 ( q ) N_2(q) N 2 ( q ) . As we will see in a moment, this number does not depend on r r r as long as r r r is coprime to q q q , and once it is clear that this number does not depend on r r r , it is also clear that it is a multiplicative function from the factorization (13) .
Proof of Theorem 1.5. When q = 2 a q=2^a q = 2 a , a character ψ \psi ψ (mod q q q ) is determined by its values ψ ( − 1 ) = ± 1 \psi(-1)=\pm1 ψ ( − 1 ) = ± 1 and ψ ( 5 ) = e ( α / 2 a − 2 ) \psi(5)=e(\alpha/2^{a-2}) ψ ( 5 ) = e ( α / 2 a − 2 ) for the 2 a − 2 2^{a-2} 2 a − 2 different possible values of α \alpha α . The primitive characters are exactly those for which α \alpha α is odd, and clearly every residue class modulo 8 is covered the same number of times by the possible values of α \alpha α . As N 2 ( 2 a ) N_2(2^a) N 2 ( 2 a ) depends only on α \alpha α (mod 8) we see that it is independent of r r r and, examining Proposition 2.4 , we find that:
(i) N 2 ( 2 ) = 0 N_2(2)=0 N 2 ( 2 ) = 0 (there are no primitive characters);
(ii) for 2 ≤ a < 5 2\le a<5 2 ≤ a < 5 , we have N 2 ( 2 a ) = 2 a − 2 N_2(2^a)=2^{a-2} N 2 ( 2 a ) = 2 a − 2 (the sum is always nonzero);
(iii) N 2 ( 2 5 ) = 4 N_2(2^5)=4 N 2 ( 2 5 ) = 4 (there are 8 primitive characters, and the sum vanishes for half of them);
(iv) for a > 5 a>5 a > 5 , we have N 2 ( 2 a ) = 2 a − 4 N_2(2^a)=2^{a-4} N 2 ( 2 a ) = 2 a − 4 (there are 2 a − 2 2^{a-2} 2 a − 2 primitive characters and the sum is nonzero one quarter of the time).
Now suppose that q = p a q=p^a q = p a for an odd prime p p p . When a = 1 a=1 a = 1 , by Proposition 2.5 , the sum S 2 ( ψ , r ) S_2(\psi,r) S 2 ( ψ , r ) never vanishes, so N 2 ( p ) = p − 2 N_2(p)=p-2 N 2 ( p ) = p − 2 . When a ≥ 2 a\ge2 a ≥ 2 , the vanishing or nonvanishing of S 2 ( ψ , r ) S_2(\psi,r) S 2 ( ψ , r ) is determined by the congruence class of α ( m o d p ) \alpha\pmod{p} α ( mod p ) , where ψ ( p + 1 ) = e ( α / p a − 1 ) \psi(p+1)=e(\alpha/p^{a-1}) ψ ( p + 1 ) = e ( α / p a − 1 ) . For half of the α ( m o d p ) \alpha\pmod{p} α ( mod p ) the sum will vanish and for the other half it will not. The primitive characters are exactly those for which p ∤ α p\nmid\alpha p ∤ α , and, as the character group is cyclic, each value of α \alpha α occurs the same number of times. There are φ ( p a ) − φ ( p a − 1 ) \varphi(p^a)-\varphi(p^{a-1}) φ ( p a ) − φ ( p a − 1 ) primitive characters, and for half of them the sum does not vanish, so N 2 ( p a ) = p a − 2 ( p − 1 ) 2 / 2 N_2(p^a)=p^{a-2}(p-1)^2/2 N 2 ( p a ) = p a − 2 ( p − 1 ) 2 /2 when a ≥ 2 a\ge2 a ≥ 2 .
Now, if 8 divides q q q , then for every primitive character ψ \psi ψ (mod 2 q 2q 2 q ) such that no factor ψ p \psi_p ψ p is quadratic, the character ψ 2 \psi^2 ψ 2 has conductor q q q . So, we can compute N ( q ) N(q) N ( q ) by computing N 2 ( 2 q ) N_2(2q) N 2 ( 2 q ) and adjusting for the multiplicity of squares and each quadratic character (mod p p p ). For each odd prime p p p dividing q q q , this is a factor of 2, and there is another factor of 4 at the prime 2, so we find that
N ( q ) = ⌊ 2 m − 5 ⌋ ∏ p ∣ q ( p − 3 2 ) ∏ p k ∥ q p o d d , k > 1 ( p k − 2 ( p − 1 ) 2 4 ) , N(q)=\lfloor2^{m-5}\rfloor\prod_{p\mid q}\left(\frac{p-3}{2}\right)\prod_{\substack{p^k\parallel q\\ p\ \mathrm{odd},\ k>1}}\left(\frac{p^{k-2}(p-1)^2}{4}\right), N ( q ) = ⌊ 2 m − 5 ⌋ p ∣ q ∏ ( 2 p − 3 ) p k ∥ q p odd , k > 1 ∏ ( 4 p k − 2 ( p − 1 ) 2 ) ,
which proves (9) , and completes the proof of the theorem. □
We now use the Selberg–Delange method to deduce Corollary 1.6 from Theorem 1.5.
Proof of Corollary 1.6. Using Theorem 1.5, we can formally expand the Dirichlet series
F ( s ) : = ∑ q = 1 ∞ N ( q ) q s + 1 = ( 1 2 3 ( s + 1 ) + 1 2 4 ( s + 1 ) + ∑ a ≥ 5 2 a − 5 2 a ( s + 1 ) ) ∏ p ≠ 2 ( 1 + p − 3 2 p s + 1 + ∑ a ≥ 2 p a − 2 ( p − 1 ) 2 4 p a ( s + 1 ) ) = f ( s ) g ( s ) ∏ p ( 1 + 1 2 p s ) , \begin{aligned}
F(s):=\sum_{q=1}^{\infty}\frac{N(q)}{q^{s+1}}=\left(\frac{1}{2^{3(s+1)}}+\frac{1}{2^{4(s+1)}}+\sum_{a\ge5}\frac{2^{a-5}}{2^{a(s+1)}}\right)\prod_{p\ne2}\left(1+\frac{p-3}{2p^{s+1}}+\sum_{a\ge2}\frac{p^{a-2}(p-1)^2}{4p^{a(s+1)}}\right) \\
&=f(s)g(s)\prod_p\left(1+\frac{1}{2p^s}\right),
\end{aligned} F ( s ) := q = 1 ∑ ∞ q s + 1 N ( q ) = ( 2 3 ( s + 1 ) 1 + 2 4 ( s + 1 ) 1 + a ≥ 5 ∑ 2 a ( s + 1 ) 2 a − 5 ) p = 2 ∏ ( 1 + 2 p s + 1 p − 3 + a ≥ 2 ∑ 4 p a ( s + 1 ) p a − 2 ( p − 1 ) 2 ) = f ( s ) g ( s ) p ∏ ( 1 + 2 p s 1 ) ,
where
f ( s ) = ( 1 2 3 ( s + 1 ) + 1 2 4 ( s + 1 ) + 1 2 4 s + 5 ( 2 s − 1 ) ) ( 1 − 1 2 s + 2 + 1 2 s + 4 ( 2 s − 1 ) ) − 1 f(s)=\left(\frac{1}{2^{3(s+1)}}+\frac{1}{2^{4(s+1)}}+\frac{1}{2^{4s+5}(2^s-1)}\right)\left(1-\frac{1}{2^{s+2}}+\frac{1}{2^{s+4}(2^s-1)}\right)^{-1} f ( s ) = ( 2 3 ( s + 1 ) 1 + 2 4 ( s + 1 ) 1 + 2 4 s + 5 ( 2 s − 1 ) 1 ) ( 1 − 2 s + 2 1 + 2 s + 4 ( 2 s − 1 ) 1 ) − 1
and
g ( s ) = ∏ p ( 1 − 3 ( 2 p s + 1 ) p + ( p − 1 ) 2 2 p 2 ( p s − 1 ) ( 2 p s + 1 ) ) . g(s)=\prod_p\left(1-\frac{3}{(2p^s+1)p}+\frac{(p-1)^2}{2p^2(p^s-1)(2p^s+1)}\right). g ( s ) = p ∏ ( 1 − ( 2 p s + 1 ) p 3 + 2 p 2 ( p s − 1 ) ( 2 p s + 1 ) ( p − 1 ) 2 ) .
Note that f ( s ) ⋅ g ( s ) f(s)\cdot g(s) f ( s ) ⋅ g ( s ) is analytic, nonzero, and uniformly bounded for Re ( s ) ≥ 1 / 2 + ε \operatorname{Re}(s)\ge1/2+\varepsilon Re ( s ) ≥ 1/2 + ε . If we let
h ( s ) = ∏ p ( 1 − 1 4 p 2 s ) ( 1 + 1 2 p 2 s + p s − 1 ) , h(s)=\prod_p\left(1-\frac{1}{4p^{2s}}\right)\left(1+\frac{1}{2p^{2s}+p^s-1}\right), h ( s ) = p ∏ ( 1 − 4 p 2 s 1 ) ( 1 + 2 p 2 s + p s − 1 1 ) ,
then h ( s ) h(s) h ( s ) is analytic, nonzero, and uniformly bounded in the half-plane Re ( s ) ≥ 1 / 2 + ε \operatorname{Re}(s)\ge1/2+\varepsilon Re ( s ) ≥ 1/2 + ε , and satisfies
∏ p ( 1 + 1 2 p s ) 2 = ζ ( s ) ζ ( 2 s ) h ( s ) . \prod_p\left(1+\frac{1}{2p^s}\right)^2=\frac{\zeta(s)}{\zeta(2s)}h(s). p ∏ ( 1 + 2 p s 1 ) 2 = ζ ( 2 s ) ζ ( s ) h ( s ) .
Therefore, we have shown that
F ( s ) ζ ( s ) − 1 / 2 = f ( s ) g ( s ) h ( s ) 1 / 2 ζ ( 2 s ) − 1 / 2 , F(s)\zeta(s)^{-1/2}=f(s)g(s)h(s)^{1/2}\zeta(2s)^{-1/2}, F ( s ) ζ ( s ) − 1/2 = f ( s ) g ( s ) h ( s ) 1/2 ζ ( 2 s ) − 1/2 ,
and hence F ( s ) ζ ( s ) − 1 / 2 F(s)\zeta(s)^{-1/2} F ( s ) ζ ( s ) − 1/2 is analytic and uniformly bounded in the half-plane Re ( s ) ≥ 1 / 2 + ε \operatorname{Re}(s)\ge1/2+\varepsilon Re ( s ) ≥ 1/2 + ε . Thus, for Q ≥ 3 Q\ge3 Q ≥ 3 , using the Selberg–Delange method as stated in [32 Chapter II, Theorem 5.2] , we derive that
∑ q ≤ Q N ( q ) q = f ( 1 ) g ( 1 ) h ( 1 ) π ζ ( 2 ) Q log Q + O ( Q ( log Q ) 3 / 2 ) . \sum_{q\le Q}\frac{N(q)}{q}=f(1)g(1)\sqrt{\frac{h(1)}{\pi\zeta(2)}}\frac{Q}{\sqrt{\log Q}}+O\left(\frac{Q}{(\log Q)^{3/2}}\right). q ≤ Q ∑ q N ( q ) = f ( 1 ) g ( 1 ) π ζ ( 2 ) h ( 1 ) log Q Q + O ( ( log Q ) 3/2 Q ) .
Summation by parts implies that
∑ q ≤ Q N ( q ) = f ( 1 ) g ( 1 ) 2 h ( 1 ) π ζ ( 2 ) Q 2 log Q + O ( Q 2 ( log Q ) 3 / 2 ) . (19) \sum_{q\le Q}N(q)=\frac{f(1)g(1)}{2}\sqrt{\frac{h(1)}{\pi\zeta(2)}}\frac{Q^2}{\sqrt{\log Q}}+O\left(\frac{Q^2}{(\log Q)^{3/2}}\right). \tag*{(19)} q ≤ Q ∑ N ( q ) = 2 f ( 1 ) g ( 1 ) π ζ ( 2 ) h ( 1 ) log Q Q 2 + O ( ( log Q ) 3/2 Q 2 ) . (19)
which proves (1.10) with
A = f ( 1 ) g ( 1 ) 2 h ( 1 ) π ζ ( 2 ) . (20) \mathcal{A}=\frac{f(1)g(1)}{2}\sqrt{\frac{h(1)}{\pi\zeta(2)}}. \tag*{(20)} A = 2 f ( 1 ) g ( 1 ) π ζ ( 2 ) h ( 1 ) . (20)
This completes the proof of the corollary. □
Proofs of Theorem 1.1, Theorem 1.9, and Theorem 1.10
In this section, we prove Theorem 1.1, Theorem 1.9, and Theorem 1.10. Instead of three separate proofs, we estimate the integral of a Z Z Z -function associated to a generic L L L -function and then specialize to the three cases covered in these theorems.
L-functions
For d ∈ N d \in\mathbb{N} d ∈ N , following [19 Chapter 5] , we suppose that an L L L -function of degree d d d is given by the absolutely convergent Dirichlet series and Euler product
L ( s , f ) = ∑ n ≥ 1 λ f ( n ) n − s = ∏ p ( 1 − α 1 ( p ) p − s ) − 1 ⋯ ( 1 − α d ( p ) p − s ) − 1 L(s,f)=\sum_{n\ge1}\lambda_f(n)n^{-s}=\prod_p(1-\alpha_1(p)p^{-s})^{-1}\cdots(1-\alpha_d(p)p^{-s})^{-1} L ( s , f ) = n ≥ 1 ∑ λ f ( n ) n − s = p ∏ ( 1 − α 1 ( p ) p − s ) − 1 ⋯ ( 1 − α d ( p ) p − s ) − 1
for Re ( s ) > 1 \operatorname{Re}(s)>1 Re ( s ) > 1 , where λ f ( 1 ) = 1 \lambda_f(1)=1 λ f ( 1 ) = 1 , λ f ( n ) ∈ C \lambda_f(n)\in\mathbb{C} λ f ( n ) ∈ C , and α i ( p ) ∈ C \alpha_i(p)\in\mathbb{C} α i ( p ) ∈ C . We assume that L ( s , f ) L(s,f) L ( s , f ) extends to a meromorphic function on C \mathbb{C} C with at most a pole at s = 1 s=1 s = 1 . We further assume that L ( s , f ) L(s,f) L ( s , f ) satisfies a functional equation, in asymmetric form, given by
L ( s , f ) = G f ( s ) L ( 1 − s , f ‾ ) and G f ( s ) = ε f N 1 / 2 − s ∏ j = 1 d Γ R ( 1 − s + κ j ) Γ R ( s + κ j ) , L(s,f)=G_f(s)L(1-s,\overline{f})\quad\text{and}\quad G_f(s)=\varepsilon_fN^{1/2-s}\prod_{j=1}^{d}\frac{\Gamma_{\mathbb{R}}(1-s+\kappa_j)}{\Gamma_{\mathbb{R}}(s+\kappa_j)}, L ( s , f ) = G f ( s ) L ( 1 − s , f ) and G f ( s ) = ε f N 1/2 − s j = 1 ∏ d Γ R ( s + κ j ) Γ R ( 1 − s + κ j ) ,
where L ( s ‾ , f ‾ ) = L ( s , f ) ‾ L(\overline{s},\overline{f})=\overline{L(s,f)} L ( s , f ) = L ( s , f ) . Here N N N is a positive integer (depending on the L L L -function), ∣ ε f ∣ = 1 |\varepsilon_f|=1 ∣ ε f ∣ = 1 , and κ j ∈ C \kappa_j\in\mathbb{C} κ j ∈ C . We assume that the κ j \kappa_j κ j are either real or come in conjugate pairs2 and that Re ( κ j ) > − 1 2 \operatorname{Re}(\kappa_j)>-\frac{1}{2} Re ( κ j ) > − 2 1 . With these assumptions, we define the Z Z Z -function, Z f ( t ) Z_f(t) Z f ( t ) , associated to L ( s , f ) L(s,f) L ( s , f ) by
Z f ( t ) = G f ( 1 2 + i t ) − 1 / 2 L ( 1 2 + i t , f ) . Z_f(t)=G_f\left(\frac{1}{2}+it\right)^{-1/2}L\left(\frac{1}{2}+it,f\right). Z f ( t ) = G f ( 2 1 + i t ) − 1/2 L ( 2 1 + i t , f ) .
Then Z f ( t ) Z_f(t) Z f ( t ) is real-valued for t ∈ R t\in\mathbb{R} t ∈ R , infinitely differentiable, and ∣ Z f ( t ) ∣ = ∣ L ( 1 2 + i t , f ) ∣ |Z_f(t)|=|L(\frac{1}{2}+it,f)| ∣ Z f ( t ) ∣ = ∣ L ( 2 1 + i t , f ) ∣ .
Stationary phase
Throughout the section, we assume N N N is fixed and we allow the implied constant to depend on N N N (and therefore on L ( s , f ) L(s,f) L ( s , f ) ).
Lemma 4.1. Let ℓ > 0 \ell>0 ℓ > 0 be fixed, and let a a a lie in a fixed bounded interval. For large x x x with x < r ≤ 2 x x<r\le2x x < r ≤ 2 x , we have, uniformly in a a a ,
∫ x 2 x t a ( t r e ) i t ℓ d t = 2 π ℓ r a + 1 2 e − i r ℓ + π i 4 + E ( a , ℓ , r , x ) , \int_x^{2x}t^a\left(\frac{t}{r e}\right)^{i t\ell}\,dt=\sqrt{\frac{2\pi}{\ell}}r^{a+\frac{1}{2}}e^{-ir\ell+\frac{\pi i}{4}}+\mathcal{E}(a,\ell,r,x), ∫ x 2 x t a ( r e t ) i t ℓ d t = ℓ 2 π r a + 2 1 e − i r ℓ + 4 π i + E ( a , ℓ , r , x ) ,
where
E ( a , ℓ , r , x ) = O ( ( ℓ x ) a ) + O ( ( ℓ x ) a + 1 ℓ ∣ x − r ∣ + ℓ x ) + O ( ( ℓ x ) a + 1 ℓ ∣ 2 x − r ∣ + ℓ x ) . \mathcal{E}(a,\ell,r,x)=O\left((\ell x)^a\right)+O\left(\frac{(\ell x)^{a+1}}{\ell|x-r|+\sqrt{\ell x}}\right)+O\left(\frac{(\ell x)^{a+1}}{\ell|2x-r|+\sqrt{\ell x}}\right). E ( a , ℓ , r , x ) = O ( ( ℓ x ) a ) + O ( ℓ ∣ x − r ∣ + ℓ x ( ℓ x ) a + 1 ) + O ( ℓ ∣2 x − r ∣ + ℓ x ( ℓ x ) a + 1 ) .
When r ≤ x r\le x r ≤ x or r > 2 x r>2x r > 2 x , there is no main term.
Proof. This is a straightforward consequence of [12 Lemma 2] , after a suitable variable change. □
Proposition 4.2. Let T T T be large, T f = N 1 / d T / ( 2 π ) T_f=N^{1/d}T/(2\pi) T f = N 1/ d T / ( 2 π ) , and c = 1 + 1 / log T f c=1+1/\log T_f c = 1 + 1/ log T f . Assume that
∣ L ( 1 2 + i t , f ) ∣ ≪ ( ∣ t ∣ + 1 ) d / 4 and ∑ n ≥ 1 ∣ λ f ( n ) ∣ n c ≪ ( log T ) β (21) \left|L\left(\frac{1}{2}+it,f\right)\right|\ll(|t|+1)^{d/4}
\quad\text{and}\quad
\sum_{n\ge1}\frac{|\lambda_f(n)|}{n^c}\ll(\log T)^\beta \tag*{(21)} L ( 2 1 + i t , f ) ≪ ( ∣ t ∣ + 1 ) d /4 and n ≥ 1 ∑ n c ∣ λ f ( n ) ∣ ≪ ( log T ) β (21) for some β ≥ 0 \beta\ge0 β ≥ 0 . Then
∫ T 2 T Z f ( t ) d t = A f ∑ T f < n 2 / d ≤ 2 T f λ f ( n ) n 1 / d − 1 / 2 e ( − d 2 ( n 2 N ) 1 / d ) + O ( T d / 4 ( log T ) β ) \int_T^{2T} Z_f(t)\,\mathrm{d}t
=A_f\sum_{T_f<n^{2/d}\le2T_f}\lambda_f(n)n^{1/d-1/2}
e\left(-\frac{d}{2}\left(\frac{n^2}{N}\right)^{1/d}\right)
+O\left(T^{d/4}(\log T)^\beta\right) ∫ T 2 T Z f ( t ) d t = A f T f < n 2/ d ≤ 2 T f ∑ λ f ( n ) n 1/ d − 1/2 e ( − 2 d ( N n 2 ) 1/ d ) + O ( T d /4 ( log T ) β ) + O ( T 1 − d / 4 max α ∈ { 1 , 2 } { ∑ 3 α 4 T f ≤ n 2 / d ≤ 5 α 4 T f ∣ λ f ( n ) ∣ ∣ n 2 / d − α T f ∣ + T } ) , \qquad
+O\left(T^{1-d/4}\max_{\alpha\in\{1,2\}}
\left\{
\sum_{\frac{3\alpha}{4}T_f\le n^{2/d}\le\frac{5\alpha}{4}T_f}
\frac{|\lambda_f(n)|}{|n^{2/d}-\alpha T_f|+\sqrt{T}}
\right\}\right), + O T 1 − d /4 α ∈ { 1 , 2 } max ⎩ ⎨ ⎧ 4 3 α T f ≤ n 2/ d ≤ 4 5 α T f ∑ ∣ n 2/ d − α T f ∣ + T ∣ λ f ( n ) ∣ ⎭ ⎬ ⎫ , where
A f = 2 3 / 2 π ( ε f d N 1 / d ) − 1 / 2 e ( 1 8 ∑ 1 ≤ j ≤ d Re ( κ j ) + 1 8 − d 16 ) , A_f=2^{3/2}\pi\left(\varepsilon_f dN^{1/d}\right)^{-1/2}
e\left(\frac{1}{8}\sum_{1\le j\le d}\operatorname{Re}(\kappa_j)+\frac{1}{8}-\frac{d}{16}\right), A f = 2 3/2 π ( ε f d N 1/ d ) − 1/2 e 8 1 1 ≤ j ≤ d ∑ Re ( κ j ) + 8 1 − 16 d , and the implied constants depend on the L L L -function L ( s , f ) L(s,f) L ( s , f ) .
Remark. The L L L -functions considered in Theorem 1.1, Theorem 1.9, and Theorem 1.10 all satisfy the bounds in (4.1).
Proof. Making the variable change s = 1 2 + i t s=\frac{1}{2}+it s = 2 1 + i t , we have
∫ T 2 T Z f ( t ) d t = 1 i ∫ 1 2 + i T 1 2 + 2 i T G f ( s ) − 1 / 2 L ( s , f ) d s . \int_T^{2T}Z_f(t)\,\mathrm{d}t
=\frac{1}{i}\int_{\frac{1}{2}+iT}^{\frac{1}{2}+2iT}G_f(s)^{-1/2}L(s,f)\,\mathrm{d}s. ∫ T 2 T Z f ( t ) d t = i 1 ∫ 2 1 + i T 2 1 + 2 i T G f ( s ) − 1/2 L ( s , f ) d s .
The function G f ( s ) G_f(s) G f ( s ) is analytic and nonzero for Re ( s ) ≥ 1 / 2 \operatorname{Re}(s)\ge1/2 Re ( s ) ≥ 1/2 and Im ( s ) ≥ T \operatorname{Im}(s)\ge T Im ( s ) ≥ T , when T T T is sufficiently large. Since this region is convex, we can define an analytic branch of G f ( s ) − 1 / 2 G_f(s)^{-1/2} G f ( s ) − 1/2 in it, and we choose the branch that agrees with the one defining Z f ( t ) Z_f(t) Z f ( t ) on the segment from 1 2 + i T \frac{1}{2}+iT 2 1 + i T to 1 2 + 2 i T \frac{1}{2}+2iT 2 1 + 2 i T . Cauchy’s Theorem, applied to the positively oriented rectangle with vertices
[ 1 2 + i T , c + i T , c + 2 i T , 1 2 + 2 i T ] , \left[\frac{1}{2}+iT,c+iT,c+2iT,\frac{1}{2}+2iT\right], [ 2 1 + i T , c + i T , c + 2 i T , 2 1 + 2 i T ] ,
implies that
∫ T 2 T Z f ( t ) d t = 1 i ∫ c + i T c + 2 i T G f ( s ) − 1 / 2 L ( s , f ) d s + H 1 − H 2 , \int_T^{2T}Z_f(t)\,\mathrm{d}t
=\frac{1}{i}\int_{c+iT}^{c+2iT}G_f(s)^{-1/2}L(s,f)\,\mathrm{d}s+H_1-H_2, ∫ T 2 T Z f ( t ) d t = i 1 ∫ c + i T c + 2 i T G f ( s ) − 1/2 L ( s , f ) d s + H 1 − H 2 ,
where c = 1 + 1 / log T f c=1+1/\log T_f c = 1 + 1/ log T f and
H k = 1 i ∫ 1 2 + k i T c + k i T G f ( s ) − 1 / 2 L ( s , f ) d s H_k=\frac{1}{i}\int_{\frac{1}{2}+kiT}^{c+kiT}G_f(s)^{-1/2}L(s,f)\,\mathrm{d}s H k = i 1 ∫ 2 1 + k i T c + k i T G f ( s ) − 1/2 L ( s , f ) d s
for k = 1 , 2 k=1,2 k = 1 , 2 . Note that trivially
∣ H k ∣ ≤ ( c − 1 2 ) max 1 2 ≤ σ ≤ c ∣ G f ( σ + k i T ) − 1 / 2 L ( σ + k i T , f ) ∣ . (22) |H_k|\le\left(c-\frac{1}{2}\right)\max_{\frac{1}{2}\le\sigma\le c}
\left|G_f(\sigma+kiT)^{-1/2}L(\sigma+kiT,f)\right|. \tag*{(22)} ∣ H k ∣ ≤ ( c − 2 1 ) 2 1 ≤ σ ≤ c max G f ( σ + k i T ) − 1/2 L ( σ + k i T , f ) . (22)
Since ∣ G f ( 1 2 + i t ) ∣ = 1 |G_f(\frac{1}{2}+it)|=1 ∣ G f ( 2 1 + i t ) ∣ = 1 for t ∈ R t\in\mathbb{R} t ∈ R , by (4.1) we see that
∣ G f ( 1 2 + i t ) − 1 / 2 L ( 1 2 + i t , f ) ∣ ≪ t d / 4 (23) \left|G_f\left(\frac{1}{2}+it\right)^{-1/2}L\left(\frac{1}{2}+it,f\right)\right|\ll t^{d/4} \tag*{(23)} G f ( 2 1 + i t ) − 1/2 L ( 2 1 + i t , f ) ≪ t d /4 (23)
as t → ∞ t \to\infty t → ∞ . Also, by Stirling’s formula for the gamma function, we have
G f ( σ + i t ) − 1 / 2 = ε f − 1 / 2 e ( 1 8 ∑ j = 1 d Re ( κ j ) − d 16 ) ( N 1 / d t 2 π ) d ( σ − 1 2 ) / 2 ( N 1 / d t 2 π e ) i t d / 2 { 1 + O ( 1 ∣ t ∣ ) } , (24) \mathcal{G}_f(\sigma+it)^{-1/2}
= \varepsilon_f^{-1/2}e\left(\frac{1}{8}\sum_{j=1}^{d}\operatorname{Re}(\kappa_j)-\frac{d}{16}\right)\left(N^{1/d}\frac{t}{2\pi}\right)^{d(\sigma-\frac{1}{2})/2}\left(N^{1/d}\frac{t}{2\pi e}\right)^{itd/2}\left\{1+O\left(\frac{1}{|t|}\right)\right\}, \tag*{(24)} G f ( σ + i t ) − 1/2 = ε f − 1/2 e ( 8 1 j = 1 ∑ d Re ( κ j ) − 16 d ) ( N 1/ d 2 π t ) d ( σ − 2 1 ) /2 ( N 1/ d 2 π e t ) i t d /2 { 1 + O ( ∣ t ∣ 1 ) } , (24)
uniformly for σ \sigma σ in a compact interval and t t t sufficiently large. So, again by (21) , we deduce that
∣ G f ( c + i t ) − 1 / 2 L ( c + i t , f ) ∣ ≪ t d / 4 ∑ n ≥ 1 ∣ λ f ( n ) ∣ n c ≪ t d / 4 ( log T ) β (25) \left|\mathcal{G}_f(c+it)^{-1/2}L(c+it,f)\right|\ll t^{d/4}\sum_{n\geq1}\frac{|\lambda_f(n)|}{n^c}\ll t^{d/4}(\log T)^\beta \tag*{(25)} G f ( c + i t ) − 1/2 L ( c + i t , f ) ≪ t d /4 n ≥ 1 ∑ n c ∣ λ f ( n ) ∣ ≪ t d /4 ( log T ) β (25)
as t → ∞ t \to\infty t → ∞ . Using (4.3) and (4.5), an application of the Phragmén–Lindelöf principle now implies that
∣ H 1 ∣ + ∣ H 2 ∣ ≪ T d / 4 ( log T ) β . |H_1|+|H_2|\ll T^{d/4}(\log T)^\beta. ∣ H 1 ∣ + ∣ H 2 ∣ ≪ T d /4 ( log T ) β .
Next we expand L ( s , f ) L(s,f) L ( s , f ) as an absolutely convergent Dirichlet series, interchange the sum and the integral, and then make the variable change s = c + i t s=c+it s = c + i t , to see that
1 i ∫ c + i T c + 2 i T G f ( s ) − 1 / 2 L ( s , f ) d s = ∑ n ≥ 1 λ f ( n ) n c ∫ T 2 T G f ( c + i t ) − 1 / 2 n − i t d t = ε f − 1 / 2 e ( 1 8 ∑ j = 1 d Re ( κ j ) − d 16 ) ∑ n ≥ 1 λ f ( n ) n c ∫ T 2 T ( N 1 / d t 2 π ) d ( c − 1 2 ) / 2 ( N 1 / d t 2 π n 2 / d e ) i t d / 2 d t + O ( T d / 4 ( log T ) β ) . \begin{aligned}
\frac{1}{i}\int_{c+iT}^{c+2iT}\mathcal{G}_f(s)^{-1/2}L(s,f)\,ds
=\sum_{n\geq1}\frac{\lambda_f(n)}{n^c}\int_T^{2T}\mathcal{G}_f(c+it)^{-1/2}n^{-it}\,dt \\
&=\varepsilon_f^{-1/2}e\left(\frac{1}{8}\sum_{j=1}^{d}\operatorname{Re}(\kappa_j)-\frac{d}{16}\right)\sum_{n\geq1}\frac{\lambda_f(n)}{n^c}\int_T^{2T}\left(N^{1/d}\frac{t}{2\pi}\right)^{d(c-\frac{1}{2})/2}\left(N^{1/d}\frac{t}{2\pi n^{2/d}e}\right)^{itd/2}\,dt
+O\left(T^{d/4}(\log T)^\beta\right).
\end{aligned} i 1 ∫ c + i T c + 2 i T G f ( s ) − 1/2 L ( s , f ) d s = n ≥ 1 ∑ n c λ f ( n ) ∫ T 2 T G f ( c + i t ) − 1/2 n − i t d t = ε f − 1/2 e ( 8 1 j = 1 ∑ d Re ( κ j ) − 16 d ) n ≥ 1 ∑ n c λ f ( n ) ∫ T 2 T ( N 1/ d 2 π t ) d ( c − 2 1 ) /2 ( N 1/ d 2 π n 2/ d e t ) i t d /2 d t + O ( T d /4 ( log T ) β ) .
Here, the second equality comes after applying Stirling’s formula for G f ( c + i t ) − 1 / 2 \mathcal{G}_f(c+it)^{-1/2} G f ( c + i t ) − 1/2 in (4.4) and then using (4.1) to obtain the big-O O O term. Making the variable change N 1 / d t / ( 2 π ) ↦ t N^{1/d}t/(2\pi)\mapsto t N 1/ d t / ( 2 π ) ↦ t and then combining estimates, we have deduced that
∫ T f 2 T f Z f ( t ) d t = ε f − 1 / 2 e ( 1 8 ∑ j = 1 d Re ( κ j ) − d 16 ) 2 π N 1 / d ∑ n ≥ 1 λ f ( n ) n c ∫ T f 2 T f t d ( c − 1 2 ) / 2 ( t n 2 / d e ) i t d N − 1 / d d t + O ( T d / 4 ( log T ) β ) . \int_{T_f}^{2T_f}Z_f(t)\,dt
=\varepsilon_f^{-1/2}e\left(\frac{1}{8}\sum_{j=1}^{d}\operatorname{Re}(\kappa_j)-\frac{d}{16}\right)\frac{2\pi}{N^{1/d}}\sum_{n\geq1}\frac{\lambda_f(n)}{n^c}\int_{T_f}^{2T_f}t^{d(c-\frac{1}{2})/2}\left(\frac{t}{n^{2/d}e}\right)^{itdN^{-1/d}}\,dt
+O\left(T^{d/4}(\log T)^\beta\right). ∫ T f 2 T f Z f ( t ) d t = ε f − 1/2 e ( 8 1 j = 1 ∑ d Re ( κ j ) − 16 d ) N 1/ d 2 π n ≥ 1 ∑ n c λ f ( n ) ∫ T f 2 T f t d ( c − 2 1 ) /2 ( n 2/ d e t ) i t d N − 1/ d d t + O ( T d /4 ( log T ) β ) .
We now use Lemma 4.1 to estimate the integral on the right-hand side. The main term is
A f ∑ τ f < n 2 / d ≤ 2 T f λ f ( n ) n 1 / d − 1 / 2 e ( − d 2 ( n 2 N ) 1 / d ) , A_f\sum_{\tau_f<n^{2/d}\leq2T_f}\lambda_f(n)n^{1/d-1/2}e\left(-\frac{d}{2}\left(\frac{n^2}{N}\right)^{1/d}\right), A f τ f < n 2/ d ≤ 2 T f ∑ λ f ( n ) n 1/ d − 1/2 e ( − 2 d ( N n 2 ) 1/ d ) ,
while the error term (using (4.1) and the notation in Lemma 4.1) is
≪ ∑ n ≥ 1 ∣ λ f ( n ) ∣ n c { T d / 4 + ε ( d 2 ( c − 1 2 ) , π d N − 1 / d , n 2 / d , T f ) } ≪ T d / 4 ( log T ) β + T 1 + d / 4 max α ∈ { 1 , 2 } { ∑ n ≥ 1 ∣ λ f ( n ) ∣ n c 1 ∣ n 2 / d − α T f ∣ + T } ≪ T d / 4 ( log T ) β + T 1 − d / 4 max α ∈ { 1 , 2 } { ∑ 3 α 4 T f ≤ n 2 / d ≤ 5 α 4 T f ∣ λ f ( n ) ∣ ∣ n 2 / d − α T f ∣ + T } . \begin{aligned}
\ll\sum_{n\geq1}\frac{|\lambda_f(n)|}{n^c}\left\{T^{d/4}+\varepsilon\left(\frac{d}{2}\left(c-\frac{1}{2}\right),\pi dN^{-1/d},n^{2/d},T_f\right)\right\} \\
&\ll T^{d/4}(\log T)^\beta+T^{1+d/4}\max_{\alpha\in\{1,2\}}\left\{\sum_{n\geq1}\frac{|\lambda_f(n)|}{n^c}\frac{1}{|n^{2/d}-\alpha T_f|+\sqrt{T}}\right\} \\
&\ll T^{d/4}(\log T)^\beta+T^{1-d/4}\max_{\alpha\in\{1,2\}}\left\{\sum_{\frac{3\alpha}{4}T_f\leq n^{2/d}\leq\frac{5\alpha}{4}T_f}\frac{|\lambda_f(n)|}{|n^{2/d}-\alpha T_f|+\sqrt{T}}\right\}.
\end{aligned} ≪ n ≥ 1 ∑ n c ∣ λ f ( n ) ∣ { T d /4 + ε ( 2 d ( c − 2 1 ) , π d N − 1/ d , n 2/ d , T f ) } ≪ T d /4 ( log T ) β + T 1 + d /4 α ∈ { 1 , 2 } max { n ≥ 1 ∑ n c ∣ λ f ( n ) ∣ ∣ n 2/ d − α T f ∣ + T 1 } ≪ T d /4 ( log T ) β + T 1 − d /4 α ∈ { 1 , 2 } max ⎩ ⎨ ⎧ 4 3 α T f ≤ n 2/ d ≤ 4 5 α T f ∑ ∣ n 2/ d − α T f ∣ + T ∣ λ f ( n ) ∣ ⎭ ⎬ ⎫ .
Combining estimates, the proposition now follows. □
Dirichlet L L L -functions: Proof of Theorem 1.1
For a Dirichlet L L L -function L ( s , χ ) L(s,\chi) L ( s , χ ) associated to a primitive character χ \chi χ (mod q q q ), Proposition 4.2 implies that
∫ T 2 T Z χ ( t ) d t = A χ ∑ T χ < n ≤ 2 T χ χ ( n ) n e ( − n 2 2 q ) + O ( T 1 / 4 log T ) + O ( T 3 / 4 max α ∈ { 1 , 2 } { ∑ 3 α 4 T χ ≤ n 2 ≤ 5 α 4 T χ 1 ∣ n 2 − α T χ ∣ + T } ) , (26) \begin{aligned}
\int_T^{2T} Z_\chi(t)\,dt
= A_\chi\sum_{\sqrt{\mathcal{T}_\chi}<n\le\sqrt{2\mathcal{T}_\chi}} \chi(n)\sqrt{n}e\left(\frac{-n^2}{2q}\right)
+ O\left(T^{1/4}\log T\right) \\
&+ O\left(T^{3/4}\max_{\alpha\in\{1,2\}}\left\{\sum_{\frac{3\alpha}{4}\mathcal{T}_\chi\le n^2\le\frac{5\alpha}{4}\mathcal{T}_\chi}\frac{1}{|n^2-\alpha\mathcal{T}_\chi|+\sqrt{T}}\right\}\right), \tag*{(26)}
\end{aligned} ∫ T 2 T Z χ ( t ) d t = A χ T χ < n ≤ 2 T χ ∑ χ ( n ) n e ( 2 q − n 2 ) + O ( T 1/4 log T ) + O T 3/4 α ∈ { 1 , 2 } max ⎩ ⎨ ⎧ 4 3 α T χ ≤ n 2 ≤ 4 5 α T χ ∑ ∣ n 2 − α T χ ∣ + T 1 ⎭ ⎬ ⎫ , (26)
where
A χ = 2 3 / 2 i α / 2 π q 1 / 4 τ ( χ ) e ( α 8 + 1 16 ) and T χ = q T 2 π . A_\chi=\frac{2^{3/2}i^{\alpha/2}\pi}{q^{1/4}\sqrt{\tau(\chi)}}e\left(\frac{\alpha}{8}+\frac{1}{16}\right)
\quad\text{and}\quad
\mathcal{T}_\chi=\frac{qT}{2\pi}. A χ = q 1/4 τ ( χ ) 2 3/2 i α /2 π e ( 8 α + 16 1 ) and T χ = 2 π q T .
Here and throughout this proof, all implied constants are allowed to depend on q q q . It makes the calculations cleaner if we first estimate the error terms and then sum over dyadic intervals (to recover the integral from 0 0 0 to T T T ) before evaluating the main term. Since the second big-O O O term can be bounded in a similar manner when α = 1 \alpha=1 α = 1 or α = 2 \alpha=2 α = 2 , we only give the details of how to estimate this term when α = 1 \alpha=1 α = 1 . In this case, we partition the interval [ 3 T χ / 4 , 5 T χ / 4 ] [\sqrt{3\mathcal{T}_\chi}/4,\sqrt{5\mathcal{T}_\chi}/4] [ 3 T χ /4 , 5 T χ /4 ] into three subintervals:
V 1 = [ T χ − log T , T χ + log T ] , V_1=\left[\sqrt{\mathcal{T}_\chi}-\log T,\sqrt{\mathcal{T}_\chi}+\log T\right], V 1 = [ T χ − log T , T χ + log T ] ,
V 2 = [ 3 T χ / 4 , T χ − log T ] , V_2=\left[\sqrt{3\mathcal{T}_\chi}/4,\sqrt{\mathcal{T}_\chi}-\log T\right], V 2 = [ 3 T χ /4 , T χ − log T ] ,
and
V 3 = ( T χ + log T , 5 T χ / 4 ] . V_3=\left(\sqrt{\mathcal{T}_\chi}+\log T,\sqrt{5\mathcal{T}_\chi}/4\right]. V 3 = ( T χ + log T , 5 T χ /4 ] .
Accordingly, we define
Σ j = T 3 / 4 ∑ n ∈ V j 1 ∣ T χ − n 2 ∣ + T , \Sigma_j=T^{3/4}\sum_{n\in V_j}\frac{1}{|\mathcal{T}_\chi-n^2|+\sqrt{T}}, Σ j = T 3/4 n ∈ V j ∑ ∣ T χ − n 2 ∣ + T 1 ,
for 1 ≤ j ≤ 3 1\le j\le3 1 ≤ j ≤ 3 , so that this error term is O ( Σ 1 + Σ 2 + Σ 3 ) O(\Sigma_1+\Sigma_2+\Sigma_3) O ( Σ 1 + Σ 2 + Σ 3 ) . For n ∈ V 1 n\in V_1 n ∈ V 1 , we have
Σ 1 ≪ T 3 / 4 ∑ n ∈ V 1 1 T ≪ T 1 / 4 ∣ V 1 ∣ ≪ T 1 / 4 log T . \Sigma_1\ll T^{3/4}\sum_{n\in V_1}\frac{1}{\sqrt{T}}\ll T^{1/4}|V_1|\ll T^{1/4}\log T. Σ 1 ≪ T 3/4 n ∈ V 1 ∑ T 1 ≪ T 1/4 ∣ V 1 ∣ ≪ T 1/4 log T .
For n ∈ V 2 n\in V_2 n ∈ V 2 , by the integral test, we observe that
Σ 2 ≪ T 1 / 4 + T 3 / 4 ∫ V 2 d u T χ + T χ − u 2 . \Sigma_2\ll T^{1/4}+T^{3/4}\int_{V_2}\frac{du}{\sqrt{\mathcal{T}_\chi}+\sqrt{\mathcal{T}_\chi}-u^2}. Σ 2 ≪ T 1/4 + T 3/4 ∫ V 2 T χ + T χ − u 2 d u .
By partial fractions, the integral on the right-hand side is equal to
1 2 T χ + T χ ∫ V 2 ( 1 T χ + T χ − u + 1 T χ + T χ + u ) d u ≪ log T T . \frac{1}{2\sqrt{\mathcal{T}_\chi}+\sqrt{\mathcal{T}_\chi}}\int_{V_2}\left(\frac{1}{\sqrt{\mathcal{T}_\chi}+\sqrt{\mathcal{T}_\chi}-u}+\frac{1}{\sqrt{\mathcal{T}_\chi}+\sqrt{\mathcal{T}_\chi}+u}\right)\,du\ll\frac{\log T}{\sqrt{T}}. 2 T χ + T χ 1 ∫ V 2 ( T χ + T χ − u 1 + T χ + T χ + u 1 ) d u ≪ T log T .
Therefore, Σ 2 ≪ T 1 / 4 log T \Sigma_2\ll T^{1/4}\log T Σ 2 ≪ T 1/4 log T . We can bound the contribution from the n ∈ V 3 n\in V_3 n ∈ V 3 similarly to n ∈ V 2 n\in V_2 n ∈ V 2 and conclude that Σ 3 ≪ T 1 / 4 log T \Sigma_3\ll T^{1/4}\log T Σ 3 ≪ T 1/4 log T , as well. Estimating similarly when α = 2 \alpha=2 α = 2 , we conclude that
∫ T 2 T Z χ ( t ) d t = A χ ∑ T χ < n ≤ 2 T χ χ ( n ) n e ( − n 2 2 q ) + O ( T 1 / 4 log T ) . \int_T^{2T} Z_\chi(t)\,dt
= A_\chi\sum_{\sqrt{\mathcal{T}_\chi}<n\le\sqrt{2\mathcal{T}_\chi}}\chi(n)\sqrt{n}e\left(\frac{-n^2}{2q}\right)+O(T^{1/4}\log T). ∫ T 2 T Z χ ( t ) d t = A χ T χ < n ≤ 2 T χ ∑ χ ( n ) n e ( 2 q − n 2 ) + O ( T 1/4 log T ) .
Summing this estimate dyadically over the intervals [ T / 2 , T ] [T/2,T] [ T /2 , T ] , [ T / 4 , T / 2 ] [T/4,T/2] [ T /4 , T /2 ] , [ T / 8 , T / 4 ] , … [T/8,T/4],\ldots [ T /8 , T /4 ] , … and then using the definition of T χ T_\chi T χ and A χ A_\chi A χ , we derive that
∫ 0 T Z χ ( t ) d t = 2 3 / 2 q 1 / 4 i a / 2 π τ ( χ ) ( a 8 + 1 16 ) ∑ n ≤ q T 2 π χ ( n ) n e ( − n 2 2 q ) + O ( T 1 / 4 log T ) . \int_0^T Z_\chi(t)\,\mathrm{d}t=\frac{2^{3/2}}{q^{1/4}}\frac{\mathrm{i}^{a/2}\pi}{\sqrt{\tau(\chi)}}\left(\frac{a}{8}+\frac{1}{16}\right)\sum_{n\leq\sqrt{\frac{qT}{2\pi}}}\chi(n)\sqrt{n}e\left(\frac{-n^2}{2q}\right)+O\left(T^{1/4}\log T\right). ∫ 0 T Z χ ( t ) d t = q 1/4 2 3/2 τ ( χ ) i a /2 π ( 8 a + 16 1 ) n ≤ 2 π q T ∑ χ ( n ) n e ( 2 q − n 2 ) + O ( T 1/4 log T ) .
Applying partial summation to the sum over n n n on the right-hand side, we derive that
∑ n ≤ x χ ( n ) n e ( − n 2 2 q ) = S ( χ ) 3 q x 3 / 2 + O ( x 1 / 2 ) . \sum_{n\leq x}\chi(n)\sqrt{n}e\left(\frac{-n^2}{2q}\right)=\frac{S(\chi)}{3q}x^{3/2}+O(x^{1/2}). n ≤ x ∑ χ ( n ) n e ( 2 q − n 2 ) = 3 q S ( χ ) x 3/2 + O ( x 1/2 ) .
The desired formula follows by putting x = q T / ( 2 π ) x=\sqrt{qT/(2\pi)} x = q T / ( 2 π ) . Note that we may assume that a = 0 a=0 a = 0 as otherwise, if a = 1 a=1 a = 1 , then S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 . The final assertion of Theorem 1.1 follows from Theorem 1.5, since N ( 8 p ) = p − 3 2 > 0 N(8p)=\frac{p-3}{2}>0 N ( 8 p ) = 2 p − 3 > 0 for every prime p ≥ 5 p\geq5 p ≥ 5 . This completes the proof of Theorem 1.1.
Holomorphic cusp forms and Hecke–Maass cusp forms
Averages of coefficients of L L L -functions in short intervals
In order to use Proposition 4.2 to prove Theorem 1.9 and Theorem 1.10, we need to understand how the normalized coefficients of primitive holomorphic cusp forms and primitive Hecke–Maass cusp forms behave in short intervals.
Lemma 4.3. Let f ∈ S k ( Γ 0 ( N ) , χ ) f\in S_k(\Gamma_0(N),\chi) f ∈ S k ( Γ 0 ( N ) , χ ) be a primitive holomorphic cusp form. Then
∑ x < n ≤ x + x ∣ λ f ( n ) ∣ ≪ f x , as x → ∞ . (27) \sum_{x<n\leq x+\sqrt{x}}|\lambda_f(n)|\ll_f \sqrt{x},\qquad\text{as }x\to\infty. \tag*{(27)} x < n ≤ x + x ∑ ∣ λ f ( n ) ∣ ≪ f x , as x → ∞. (27) Let φ \varphi φ be a primitive Hecke–Maass cusp form of level N N N with nebentypus χ \chi χ . Then, for any ε > 0 \varepsilon>0 ε > 0 , we have
∑ x < n ≤ x + x ∣ λ φ ( n ) ∣ ≪ φ , ε x 11 / 20 + ε , as x → ∞ . (28) \sum_{x<n\leq x+\sqrt{x}}|\lambda_\varphi(n)|\ll_{\varphi,\varepsilon}x^{11/20+\varepsilon},\qquad\text{as }x\to\infty. \tag*{(28)} x < n ≤ x + x ∑ ∣ λ φ ( n ) ∣ ≪ φ , ε x 11/20 + ε , as x → ∞. (28)
Proof. If f ∈ S k ( Γ 0 ( N ) , χ ) f\in S_k(\Gamma_0(N),\chi) f ∈ S k ( Γ 0 ( N ) , χ ) is a primitive holomorphic cusp form, then it is known that
∑ p ≤ x ∣ λ f ( p ) ∣ 2 p = log log x + O ( 1 ) . \sum_{p\leq x}\frac{|\lambda_f(p)|^2}{p}=\log\log x+O(1). p ≤ x ∑ p ∣ λ f ( p ) ∣ 2 = log log x + O ( 1 ) .
For instance, this follows from [30 Proposition 2.3] and partial summation. By Cauchy’s inequality,
∑ p ≤ x ∣ λ f ( p ) ∣ p ≤ ∑ p ≤ x 1 p ∑ p ≤ x ∣ λ f ( p ) ∣ 2 p ≤ log log x + O ( 1 ) . (29) \sum_{p\leq x}\frac{|\lambda_f(p)|}{p}\leq\sqrt{\sum_{p\leq x}\frac{1}{p}}\sqrt{\sum_{p\leq x}\frac{|\lambda_f(p)|^2}{p}}\leq\log\log x+O(1). \tag*{(29)} p ≤ x ∑ p ∣ λ f ( p ) ∣ ≤ p ≤ x ∑ p 1 p ≤ x ∑ p ∣ λ f ( p ) ∣ 2 ≤ log log x + O ( 1 ) . (29)
The multiplicative function ∣ λ f ( n ) ∣ |\lambda_f(n)| ∣ λ f ( n ) ∣ belongs to the class M M M defined by Shiu in [31 ] . Therefore, using (4.9) in [31 Theorem 1] , we derive the bound in (4.7).
To derive (4.8), we cannot use Shiu’s result since it is not known if the coefficients λ φ ( n ) \lambda_\varphi(n) λ φ ( n ) satisfy the Ramanujan–Petersson conjecture. Instead, we use the Rankin–Selberg estimate
∑ n ≤ x ∣ λ φ ( n ) ∣ 2 = c φ x + O ( x 3 / 5 + ε ) , (30) \sum_{n\leq x}|\lambda_\varphi(n)|^2=c_\varphi x+O(x^{3/5+\varepsilon}), \tag*{(30)} n ≤ x ∑ ∣ λ φ ( n ) ∣ 2 = c φ x + O ( x 3/5 + ε ) , (30)
which holds for some constant c φ > 0 c_{\varphi} > 0 c φ > 0 . An estimate of this form is classical for primitive holomorphic cusp forms. This estimate can be proved in the same way using the Rankin–Selberg L L L -function L ( s , φ ⊗ φ ‾ ) L(s,\varphi\otimes\overline{\varphi}) L ( s , φ ⊗ φ ) and applying [3 Theorem 4.1] . Using (30) and Cauchy’s inequality, we see that
∑ x < n ≤ x + x ∣ λ φ ( n ) ∣ ≤ x 1 4 ∑ x < n ≤ x + x ∣ λ φ ( n ) ∣ 2 ≪ x 11 20 + ε . \sum_{x<n\le x+\sqrt{x}} |\lambda_{\varphi}(n)| \le x^{\frac{1}{4}}\sqrt{\sum_{x<n\le x+\sqrt{x}} |\lambda_{\varphi}(n)|^2} \ll x^{\frac{11}{20}+\varepsilon}. x < n ≤ x + x ∑ ∣ λ φ ( n ) ∣ ≤ x 4 1 x < n ≤ x + x ∑ ∣ λ φ ( n ) ∣ 2 ≪ x 20 11 + ε .
This completes the proof of the lemma. □
Proof of Theorem 1.9
Let f ∈ S k ( Γ 0 ( N ) , χ ) f \in S_k(\Gamma_0(N),\chi) f ∈ S k ( Γ 0 ( N ) , χ ) be a primitive holomorphic cusp form. Then Proposition 4.2 implies that
∫ T 2 T Z f ( t ) d t = 2 π ε f 1 / 2 N 1 / 4 e ( k 8 ) ∑ T f < n ≤ 2 T f λ f ( n ) e ( − n N ) + O ( T 1 / 2 log T ) (31) \int_T^{2T} Z_f(t)\,\mathrm{d}t = \frac{2\pi}{\varepsilon_f^{1/2}N^{1/4}}e\left(\frac{k}{8}\right)\sum_{T_f<n\le2T_f}\lambda_f(n)e\left(-\frac{n}{\sqrt{N}}\right)+O(T^{1/2}\log T) \tag*{(31)} ∫ T 2 T Z f ( t ) d t = ε f 1/2 N 1/4 2 π e ( 8 k ) T f < n ≤ 2 T f ∑ λ f ( n ) e ( − N n ) + O ( T 1/2 log T ) (31)
+ O ( T 1 / 2 max α ∈ { 1 , 2 } { ∑ 3 α 4 T f ≤ n ≤ 5 α 4 T f ∣ λ f ( n ) ∣ ∣ n − α T f ∣ + T } ) . \qquad+O\left(T^{1/2}\max_{\alpha\in\{1,2\}}\left\{\sum_{\frac{3\alpha}{4}T_f\le n\le\frac{5\alpha}{4}T_f}\frac{|\lambda_f(n)|}{|n-\alpha T_f|+\sqrt{T}}\right\}\right). + O T 1/2 α ∈ { 1 , 2 } max ⎩ ⎨ ⎧ 4 3 α T f ≤ n ≤ 4 5 α T f ∑ ∣ n − α T f ∣ + T ∣ λ f ( n ) ∣ ⎭ ⎬ ⎫ .
As in the proof of Theorem 1.1, before estimating the main term, we first estimate the second big-O O O term in (31) and then we sum over dyadic intervals to recover the integral from 0 0 0 to T T T .
When α = 1 \alpha=1 α = 1 , we partition the interval [ 3 T f / 4 , 5 T f / 4 ] [3T_f/4,5T_f/4] [ 3 T f /4 , 5 T f /4 ] into
V 1 = [ T f − T , T f + T ] , V 2 = [ 3 T f / 4 , T f − T ) , and V 3 = ( T f + T , 5 T f / 4 ] . V_1=[T_f-\sqrt{T},T_f+\sqrt{T}],\qquad V_2=[3T_f/4,T_f-\sqrt{T}),\qquad\text{and}\qquad V_3=(T_f+\sqrt{T},5T_f/4]. V 1 = [ T f − T , T f + T ] , V 2 = [ 3 T f /4 , T f − T ) , and V 3 = ( T f + T , 5 T f /4 ] .
Accordingly, we let
Σ j = T 1 / 2 ∑ n ∈ V j ∣ λ f ( n ) ∣ ∣ n − T f ∣ + T , \Sigma_j=T^{1/2}\sum_{n\in V_j}\frac{|\lambda_f(n)|}{|n-T_f|+\sqrt{T}}, Σ j = T 1/2 n ∈ V j ∑ ∣ n − T f ∣ + T ∣ λ f ( n ) ∣ ,
for j = 1 , 2 , 3 j=1,2,3 j = 1 , 2 , 3 , and the second error term of (31) is O ( Σ 1 + Σ 2 + Σ 3 ) O(\Sigma_1+\Sigma_2+\Sigma_3) O ( Σ 1 + Σ 2 + Σ 3 ) . When n ∈ V 1 n\in V_1 n ∈ V 1 , we simply note that
1 ∣ n − T f ∣ + T ≤ 1 T , \frac{1}{|n-T_f|+\sqrt{T}}\le\frac{1}{\sqrt{T}}, ∣ n − T f ∣ + T 1 ≤ T 1 ,
and then the first estimate in Lemma 4.3 implies that
Σ 1 ≤ ∑ n ∈ V 1 ∣ λ f ( n ) ∣ ≪ T 1 / 2 . \Sigma_1\le\sum_{n\in V_1}|\lambda_f(n)|\ll T^{1/2}. Σ 1 ≤ n ∈ V 1 ∑ ∣ λ f ( n ) ∣ ≪ T 1/2 .
For n ∈ V 2 n\in V_2 n ∈ V 2 , we further partition V 2 V_2 V 2 into
V 2 = ⋃ j = 1 J U j : = ⋃ j = 1 J { n ∈ V 2 : T f − 2 j T ≤ n < T f − 2 j − 1 T } , V_2=\bigcup_{j=1}^{J}U_j:=\bigcup_{j=1}^{J}\left\{n\in V_2:T_f-2^j\sqrt{T}\le n<T_f-2^{j-1}\sqrt{T}\right\}, V 2 = j = 1 ⋃ J U j := j = 1 ⋃ J { n ∈ V 2 : T f − 2 j T ≤ n < T f − 2 j − 1 T } ,
where J = ⌊ log ( T f 4 T ) / log ( 2 ) ⌋ J=\left\lfloor\log\left(\frac{T_f}{4\sqrt{T}}\right)/\log(2)\right\rfloor J = ⌊ log ( 4 T T f ) / log ( 2 ) ⌋ . On each interval U j U_j U j , we have ∣ n − T f ∣ ≥ 2 j − 1 T |n-T_f|\ge2^{j-1}\sqrt{T} ∣ n − T f ∣ ≥ 2 j − 1 T . We partition U j U_j U j into 2 j − 1 2^{j-1} 2 j − 1 intervals of length T \sqrt{T} T and then apply the first estimate in Lemma 4.3 to each of these 2 j − 1 2^{j-1} 2 j − 1 intervals. It follows that
T 1 / 2 ∑ n ∈ U j ∣ λ f ( n ) ∣ ∣ n − T f ∣ + T ≪ 2 − j ∑ n ∈ U j ∣ λ f ( n ) ∣ ≪ T 1 / 2 . T^{1/2}\sum_{n\in U_j}\frac{|\lambda_f(n)|}{|n-T_f|+\sqrt{T}}\ll2^{-j}\sum_{n\in U_j}|\lambda_f(n)|\ll T^{1/2}. T 1/2 n ∈ U j ∑ ∣ n − T f ∣ + T ∣ λ f ( n ) ∣ ≪ 2 − j n ∈ U j ∑ ∣ λ f ( n ) ∣ ≪ T 1/2 .
Since J = O ( log T ) J=O(\log T) J = O ( log T ) , we deduce that Σ 2 ≪ T 1 / 2 log T \Sigma_2\ll T^{1/2}\log T Σ 2 ≪ T 1/2 log T . A similar calculation for n ∈ V 3 n\in V_3 n ∈ V 3 implies that Σ 3 ≪ T 1 / 2 log T \Sigma_3\ll T^{1/2}\log T Σ 3 ≪ T 1/2 log T . Thus, when α = 1 \alpha=1 α = 1 , the second big-O O O term in (31) is O ( T 1 / 2 log T ) O(T^{1/2}\log T) O ( T 1/2 log T ) . Similarly, when α = 2 \alpha=2 α = 2 , a straightforward adaptation of our calculation shows that this error term is again O ( T 1 / 2 log T ) O(T^{1/2}\log T) O ( T 1/2 log T ) , so
∫ T 2 T Z f ( t ) d t = 2 π ε f 1 / 2 N 1 / 4 e ( k 8 ) ∑ T f < n ≤ 2 T f λ f ( n ) e ( − n N ) + O ( T 1 / 2 log T ) . \int_T^{2T} Z_f(t)\,\mathrm{d}t=\frac{2\pi}{\varepsilon_f^{1/2}N^{1/4}}e\left(\frac{k}{8}\right)\sum_{\mathcal{T}_f<n\leq2\mathcal{T}_f}\lambda_f(n)e\left(-\frac{n}{\sqrt{N}}\right)+O(T^{1/2}\log T). ∫ T 2 T Z f ( t ) d t = ε f 1/2 N 1/4 2 π e ( 8 k ) T f < n ≤ 2 T f ∑ λ f ( n ) e ( − N n ) + O ( T 1/2 log T ) .
Summing these estimates over the dyadic intervals { [ 2 − j − 1 T , 2 − j T ] } j ≥ 0 \{[2^{-j-1}T,2^{-j}T]\}_{j\geq0} {[ 2 − j − 1 T , 2 − j T ] } j ≥ 0 , we obtain that
∫ 0 T Z f ( t ) d t = 2 π ε f 1 / 2 N 1 / 4 e ( k 8 ) ∑ n ≤ T f λ f ( n ) e ( − n N ) + O ( T 1 / 2 log T ) . (32) \int_0^T Z_f(t)\,\mathrm{d}t=\frac{2\pi}{\varepsilon_f^{1/2}N^{1/4}}e\left(\frac{k}{8}\right)\sum_{n\leq\mathcal{T}_f}\lambda_f(n)e\left(-\frac{n}{\sqrt{N}}\right)+O(T^{1/2}\log T). \tag*{(32)} ∫ 0 T Z f ( t ) d t = ε f 1/2 N 1/4 2 π e ( 8 k ) n ≤ T f ∑ λ f ( n ) e ( − N n ) + O ( T 1/2 log T ) . (32)
A classical result of Wilton [34 ] (see [17 Theorem 5.3] for full generality) states that
∑ n ≤ x λ f ( n ) e ( n α ) ≪ f x 1 / 2 log x \sum_{n\leq x}\lambda_f(n)e(n\alpha)\ll_f x^{1/2}\log x n ≤ x ∑ λ f ( n ) e ( n α ) ≪ f x 1/2 log x
uniformly for α ∈ R \alpha\in\mathbb{R} α ∈ R . Therefore, the sum over n n n in (32) is O ( T 1 / 2 log T ) O(T^{1/2}\log T) O ( T 1/2 log T ) . This completes the proof of Theorem 1.9.
4.4.3. Proof of Theorem 1.10. Since the proof of this theorem is very similar to the proof of Theorem 1.9, we simply indicate the necessary changes. If φ \varphi φ is a primitive Hecke–Maass cusp form of level N N N with nebentypus χ \chi χ , then Proposition 4.2 implies that
∫ T 2 T Z φ ( t ) d t = 2 π ε φ 1 / 2 N 1 / 4 e ( δ 4 ) ∑ T φ < n ≤ 2 T φ λ φ ( n ) e ( − n N ) + O ( T 1 / 2 log T ) + O ( T 1 / 2 max α ∈ { 1 , 2 } { ∑ 3 α 4 T φ ≤ n ≤ 5 α 4 T φ ∣ λ φ ( n ) ∣ ∣ n − α T φ ∣ + T } ) , \begin{aligned}
\int_T^{2T} Z_\varphi(t)\,\mathrm{d}t={}&\frac{2\pi}{\varepsilon_\varphi^{1/2}N^{1/4}}e\left(\frac{\delta}{4}\right)\sum_{\mathcal{T}_\varphi<n\leq2\mathcal{T}_\varphi}\lambda_\varphi(n)e\left(-\frac{n}{\sqrt{N}}\right)+O(T^{1/2}\log T)\\
&+O\left(T^{1/2}\max_{\alpha\in\{1,2\}}\left\{\sum_{\frac{3\alpha}{4}\mathcal{T}_\varphi\leq n\leq\frac{5\alpha}{4}\mathcal{T}_\varphi}\frac{|\lambda_\varphi(n)|}{|n-\alpha\mathcal{T}_\varphi|+\sqrt{T}}\right\}\right),
\end{aligned} ∫ T 2 T Z φ ( t ) d t = ε φ 1/2 N 1/4 2 π e ( 4 δ ) T φ < n ≤ 2 T φ ∑ λ φ ( n ) e ( − N n ) + O ( T 1/2 log T ) + O T 1/2 α ∈ { 1 , 2 } max ⎩ ⎨ ⎧ 4 3 α T φ ≤ n ≤ 4 5 α T φ ∑ ∣ n − α T φ ∣ + T ∣ λ φ ( n ) ∣ ⎭ ⎬ ⎫ ,
where T φ = N T / ( 2 π ) \mathcal{T}_\varphi=\sqrt{NT}/(2\pi) T φ = N T / ( 2 π ) . Estimating the second big-O O O term as we did in Theorem 1.9 but instead using the second estimate in Lemma 4.3, we deduce that
∫ T 2 T Z φ ( t ) d t = 2 π ε φ 1 / 2 N 1 / 4 e ( δ 4 ) ∑ T φ < n ≤ 2 T φ λ φ ( n ) e ( − n N ) + O ( T 11 / 20 + ε ) . \int_T^{2T} Z_\varphi(t)\,\mathrm{d}t=\frac{2\pi}{\varepsilon_\varphi^{1/2}N^{1/4}}e\left(\frac{\delta}{4}\right)\sum_{\mathcal{T}_\varphi<n\leq2\mathcal{T}_\varphi}\lambda_\varphi(n)e\left(-\frac{n}{\sqrt{N}}\right)+O(T^{11/20+\varepsilon}). ∫ T 2 T Z φ ( t ) d t = ε φ 1/2 N 1/4 2 π e ( 4 δ ) T φ < n ≤ 2 T φ ∑ λ φ ( n ) e ( − N n ) + O ( T 11/20 + ε ) .
Summing these estimates over the dyadic intervals { [ 2 − j − 1 T , 2 − j T ] } j ≥ 0 \{[2^{-j-1}T,2^{-j}T]\}_{j\geq0} {[ 2 − j − 1 T , 2 − j T ] } j ≥ 0 , we obtain that
∫ 0 T Z φ ( t ) d t = 2 π ε φ 1 / 2 N 1 / 4 e ( δ 4 ) ∑ n ≤ T φ λ φ ( n ) e ( − n N ) + O ( T 11 / 20 + ε ) . (33) \int_0^T Z_\varphi(t)\,\mathrm{d}t=\frac{2\pi}{\varepsilon_\varphi^{1/2}N^{1/4}}e\left(\frac{\delta}{4}\right)\sum_{n\leq\mathcal{T}_\varphi}\lambda_\varphi(n)e\left(-\frac{n}{\sqrt{N}}\right)+O(T^{11/20+\varepsilon}). \tag*{(33)} ∫ 0 T Z φ ( t ) d t = ε φ 1/2 N 1/4 2 π e ( 4 δ ) n ≤ T φ ∑ λ φ ( n ) e ( − N n ) + O ( T 11/20 + ε ) . (33)
By [14 Theorem 3] (see also [8 ] or [18 Theorem 8.1] ), we have
∑ n ≤ x λ φ ( n ) e ( n α ) ≪ φ x 1 / 2 log x \sum_{n\leq x}\lambda_\varphi(n)e(n\alpha)\ll_\varphi x^{1/2}\log x n ≤ x ∑ λ φ ( n ) e ( n α ) ≪ φ x 1/2 log x
uniformly for α ∈ R \alpha\in\mathbb{R} α ∈ R . Therefore, the sum over n n n in (33) is O ( T 1 / 2 log T ) O(T^{1/2}\log T) O ( T 1/2 log T ) . This completes the proof of Theorem 1.10.
Some observations on the distribution of C χ C_\chi C χ
Recall that
C χ = 2 3 / 4 π 1 / 4 e ( 1 16 ) 3 q 1 / 2 τ ( χ ) 1 / 2 S ( χ ) . C_\chi=\frac{2^{3/4}\pi^{1/4}e\left(\frac{1}{16}\right)}{3q^{1/2}\tau(\chi)^{1/2}}S(\chi). C χ = 3 q 1/2 τ ( χ ) 1/2 2 3/4 π 1/4 e ( 16 1 ) S ( χ ) .
From our evaluation of S ( χ ) S(\chi) S ( χ ) , we see that
∣ C χ ∣ ≤ A q , |C_\chi|\le A_q, ∣ C χ ∣ ≤ A q ,
where
A q = 2 3 / 4 π 1 / 4 2 a 3 q 1 / 4 A_q=\frac{2^{3/4}\pi^{1/4}2^a}{3q^{1/4}} A q = 3 q 1/4 2 3/4 π 1/4 2 a
with
a = { ω ( q ) + 1 / 2 , if 8 ∥ q , ω ( q ) + 1 , if 16 ∥ q , ω ( q ) + 3 / 2 , if 32 ∣ q . a=\begin{cases}
\omega(q)+1/2, & \text{if }8\mathrel{\|}q,\\
\omega(q)+1, & \text{if }16\mathrel{\|}q,\\
\omega(q)+3/2, & \text{if }32\mid q.
\end{cases} a = ⎩ ⎨ ⎧ ω ( q ) + 1/2 , ω ( q ) + 1 , ω ( q ) + 3/2 , if 8 ∥ q , if 16 ∥ q , if 32 ∣ q .
and equality is achieved when q q q is square-full (and C χ ≠ 0 C_\chi\ne0 C χ = 0 ). For square-full q q q , then, the only question about the distribution of C χ C_\chi C χ is its sign.
The proof of Theorem 1.1 involves a choice of square root of G χ ( s ) G_\chi(s) G χ ( s ) in the region { Re ( s ) ≥ 1 / 2 } ∖ [ 1 , ∞ ) \{\operatorname{Re}(s)\ge1/2\}\setminus[1,\infty) { Re ( s ) ≥ 1/2 } ∖ [ 1 , ∞ ) . The argument of G χ G_\chi G χ is the same as the argument of τ ( χ ) \tau(\chi) τ ( χ ) on the segment [ 1 / 2 , 1 ) [1/2,1) [ 1/2 , 1 ) , so this is equivalent to a choice of square root of the Gauss sum, and we do not see a way of making this choice that is not somewhat arbitrary. As stated after Theorem 1.1 , we use the square root for which Z χ ( 0 ) > 0 Z_\chi(0)>0 Z χ ( 0 ) > 0 when L ( 1 / 2 , χ ) ≠ 0 L(1/2,\chi)\ne0 L ( 1/2 , χ ) = 0 , which is also the choice made by the LMFDB. As the argument of L ( 1 / 2 , χ ) 2 L(1/2,\chi)^2 L ( 1/2 , χ ) 2 is the argument of τ ( χ ) \tau(\chi) τ ( χ ) , this is the same as setting the argument of τ ( χ ) 1 / 2 \tau(\chi)^{1/2} τ ( χ ) 1/2 to the argument of L ( 1 / 2 , χ ) L(1/2,\chi) L ( 1/2 , χ ) .
Somewhat curiously, it seems that with this choice, C χ C_\chi C χ has a slight (presumably asymptotically negligible) positivity bias. For example, up to modulus 500,000, there are 226,470,976 characters for which C χ C_\chi C χ is nonzero. We find that C χ > 0 C_\chi>0 C χ > 0 for 117,548,676 characters, which is about 51.9% of the time. Up to modulus one million there are 880,161,637 characters for which C χ ≠ 0 C_\chi\ne0 C χ = 0 , and 51.53% of them are positive. Table 1 lists the number of positive and negative values of C χ C_\chi C χ for some square-full numbers with small prime factors.
Table 1. Some statistics on the positivity and negativity of C χ C_{\chi} C χ , where the square root of τ ( χ ) \tau(\chi) τ ( χ ) is chosen so that Z χ ( 0 ) > 0 Z_{\chi}(0) > 0 Z χ ( 0 ) > 0 .
We now consider more generally the distribution of C χ C_\chi C χ when q q q is not square-full. First we observe that from the fact that C χ C_\chi C χ must be a real number, one can make some deductions about signs of ratios of Gauss sums.
Proposition 5.1. Suppose that ψ 2 \psi_2 ψ 2 is a primitive Dirichlet character (mod 2 a 2^a 2 a ) with a ≥ 4 a\ge4 a ≥ 4 and that χ 2 \chi_2 χ 2 is a primitive character (mod 2 a − 1 2^{a-1} 2 a − 1 ) such that ψ 2 2 = χ 2 \psi_2^2=\chi_2 ψ 2 2 = χ 2 . Let α \alpha α be an integer such that ψ 2 ( 5 ) = e ( α / 2 a − 2 ) \psi_2(5)=e(\alpha/2^{a-2}) ψ 2 ( 5 ) = e ( α / 2 a − 2 ) (equivalently, χ 2 ( 5 ) = e ( α / 2 a − 3 ) \chi_2(5)=e(\alpha/2^{a-3}) χ 2 ( 5 ) = e ( α / 2 a − 3 ) ). If a > 5 a>5 a > 5 , and a a a is even or α ≡ ± 1 ( m o d 8 ) \alpha\equiv\pm1\pmod{8} α ≡ ± 1 ( mod 8 ) ), then
τ ( ψ 2 ) 2 τ ( χ 2 ) = 2 a / 2 + 1 / 2 e ( ( − 4 α ) 1 8 ) . \frac{\tau(\psi_2)^2}{\tau(\chi_2)}=2^{a/2+1/2}e\left(\left(\frac{-4}{\alpha}\right)\frac{1}{8}\right). τ ( χ 2 ) τ ( ψ 2 ) 2 = 2 a /2 + 1/2 e ( ( α − 4 ) 8 1 ) . If a > 5 a>5 a > 5 is odd and α ≡ ± 3 ( m o d 8 ) \alpha\equiv\pm3\pmod{8} α ≡ ± 3 ( mod 8 ) , then
τ ( ψ 2 ) 2 τ ( χ 2 ) = 2 a / 2 + 1 / 2 e ( α 8 ) . \frac{\tau(\psi_2)^2}{\tau(\chi_2)}=2^{a/2+1/2}e\left(\frac{\alpha}{8}\right). τ ( χ 2 ) τ ( ψ 2 ) 2 = 2 a /2 + 1/2 e ( 8 α ) . If a = 5 a=5 a = 5 , then
τ ( ψ 2 ) 2 τ ( χ 2 ) = 8 e ( 1 2 − α 8 ) . \frac{\tau(\psi_2)^2}{\tau(\chi_2)}=8e\left(\frac{1}{2}-\frac{\alpha}{8}\right). τ ( χ 2 ) τ ( ψ 2 ) 2 = 8 e ( 2 1 − 8 α ) .
If a = 4 a=4 a = 4 , then
τ ( ψ 2 ) 2 τ ( χ 2 ) = 4 2 e ( − ( − 4 α ) 1 8 ) . \frac{\tau(\psi_2)^2}{\tau(\chi_2)}=4\sqrt{2}\,e\left(-\left(\frac{-4}{\alpha}\right)\frac{1}{8}\right). τ ( χ 2 ) τ ( ψ 2 ) 2 = 4 2 e ( − ( α − 4 ) 8 1 ) .
Proof. The small cases can be computed directly, so suppose that a > 6 a>6 a > 6 . Choose a prime p > 5 p>5 p > 5 such that − p ≡ α ( m o d 8 ) -p\equiv\alpha\pmod{8} − p ≡ α ( mod 8 ) . Let ψ p \psi_p ψ p be a primitive even non-quadratic character (mod p p p ) and χ p = ψ p 2 \chi_p=\psi_p^2 χ p = ψ p 2 , and put χ = χ 2 χ p \chi=\chi_2\chi_p χ = χ 2 χ p and ψ = ψ 2 ψ p \psi=\psi_2\psi_p ψ = ψ 2 ψ p , hence χ \chi χ is a primitive character modulo q = 2 a − 1 p q=2^{a-1}p q = 2 a − 1 p . Then S ( χ ) = S 2 ( ψ , − 1 ) ≠ 0 S(\chi)=S_2(\psi,-1)\ne0 S ( χ ) = S 2 ( ψ , − 1 ) = 0 and we have the factorization
S 2 ( ψ , − 1 ) = ψ 2 ( p ) 2 S 2 ( ψ 2 , − p ) ψ p ( 2 a ) 2 S 2 ( ψ p , − 2 a ) . S_2(\psi,-1)=\psi_2(p)^2S_2(\psi_2,-p)\psi_p(2^a)^2S_2(\psi_p,-2^a). S 2 ( ψ , − 1 ) = ψ 2 ( p ) 2 S 2 ( ψ 2 , − p ) ψ p ( 2 a ) 2 S 2 ( ψ p , − 2 a ) .
If a a a is even or if 2 2 2 is a square (mod p p p ), then
S 2 ( ψ , − 1 ) = ψ 2 ( p ) 2 S 2 ( ψ 2 , − p ) ψ p ( 2 a ) S 2 ( ψ p , − 1 ) S_2(\psi,-1)=\psi_2(p)^2S_2(\psi_2,-p)\psi_p(2^a)S_2(\psi_p,-1) S 2 ( ψ , − 1 ) = ψ 2 ( p ) 2 S 2 ( ψ 2 , − p ) ψ p ( 2 a ) S 2 ( ψ p , − 1 )
and from our evaluation of S 2 S_2 S 2 we find that
S 2 ( ψ , − 1 ) 2 = ψ 2 ( p ) 4 S 2 ( ψ 2 , − p ) 2 ψ p ( 2 a ) 2 S 2 ( ψ p , − 1 ) 2 = ψ 2 ( p ) 4 ( 4 ψ 2 ‾ ( − p ) τ ( ψ 2 ) ) 2 ψ p ( 2 a ) 2 ( τ ( ψ p ) + ϵ p ( − 1 ) τ ( ϵ p ψ p ) ) 2 = 16 ψ 2 ( p ) 2 τ ( ψ 2 ) 2 ψ p ( 2 a ) 2 ( τ ( ψ p ) + ϵ p ( − 1 ) τ ( ϵ p ψ p ) ) 2 . \begin{aligned}
S_2(\psi,-1)^2
&=\psi_2(p)^4S_2(\psi_2,-p)^2\psi_p(2^a)^2S_2(\psi_p,-1)^2\\
&=\psi_2(p)^4\left(4\overline{\psi_2}(-p)\tau(\psi_2)\right)^2\psi_p(2^a)^2\left(\tau(\psi_p)+\epsilon_p(-1)\tau(\epsilon_p\psi_p)\right)^2\\
&=16\psi_2(p)^2\tau(\psi_2)^2\psi_p(2^a)^2\left(\tau(\psi_p)+\epsilon_p(-1)\tau(\epsilon_p\psi_p)\right)^2.
\end{aligned} S 2 ( ψ , − 1 ) 2 = ψ 2 ( p ) 4 S 2 ( ψ 2 , − p ) 2 ψ p ( 2 a ) 2 S 2 ( ψ p , − 1 ) 2 = ψ 2 ( p ) 4 ( 4 ψ 2 ( − p ) τ ( ψ 2 ) ) 2 ψ p ( 2 a ) 2 ( τ ( ψ p ) + ϵ p ( − 1 ) τ ( ϵ p ψ p ) ) 2 = 16 ψ 2 ( p ) 2 τ ( ψ 2 ) 2 ψ p ( 2 a ) 2 ( τ ( ψ p ) + ϵ p ( − 1 ) τ ( ϵ p ψ p ) ) 2 .
Now,
C χ 2 = 2 3 / 2 π 1 / 2 e ( 1 / 8 ) 9 q τ ( χ ) S 2 ( ψ , − 1 ) 2 . C_\chi^2=\frac{2^{3/2}\pi^{1/2}e(1/8)}{9q\tau(\chi)}S_2(\psi,-1)^2. C χ 2 = 9 q τ ( χ ) 2 3/2 π 1/2 e ( 1/8 ) S 2 ( ψ , − 1 ) 2 .
Factoring the Gauss sum as τ ( χ ) = χ 2 ( p ) χ p ( 2 a − 1 ) τ ( χ 2 ) τ ( χ p ) = ψ 2 ( p ) 2 ψ p ( 2 a − 1 ) 2 τ ( χ 2 ) τ ( χ p ) \tau(\chi)=\chi_2(p)\chi_p(2^{a-1})\tau(\chi_2)\tau(\chi_p)=\psi_2(p)^2\psi_p(2^{a-1})^2\tau(\chi_2)\tau(\chi_p) τ ( χ ) = χ 2 ( p ) χ p ( 2 a − 1 ) τ ( χ 2 ) τ ( χ p ) = ψ 2 ( p ) 2 ψ p ( 2 a − 1 ) 2 τ ( χ 2 ) τ ( χ p ) and putting in our evaluation of S 2 S_2 S 2 , we have
C χ 2 = 16 2 3 / 2 π 1 / 2 e ( 1 / 8 ) 9 q τ ( ψ 2 ) 2 τ ( χ 2 ) ψ p ( 4 ) ( τ ( ψ p ) + ϵ p ( − 1 ) τ ( ϵ p ψ p ) ) 2 τ ( χ p ) . C_\chi^2=16\frac{2^{3/2}\pi^{1/2}e(1/8)}{9q}\frac{\tau(\psi_2)^2}{\tau(\chi_2)}\psi_p(4)\frac{\left(\tau(\psi_p)+\epsilon_p(-1)\tau(\epsilon_p\psi_p)\right)^2}{\tau(\chi_p)}. C χ 2 = 16 9 q 2 3/2 π 1/2 e ( 1/8 ) τ ( χ 2 ) τ ( ψ 2 ) 2 ψ p ( 4 ) τ ( χ p ) ( τ ( ψ p ) + ϵ p ( − 1 ) τ ( ϵ p ψ p ) ) 2 .
Now, upon inserting q = 2 a − 1 p q=2^{a-1}p q = 2 a − 1 p , multiplying and dividing by a quadratic Gauss sum, and recalling (17) we see that
C χ 2 = 16 e ( 1 − ϵ p ( − 1 ) 8 ) 2 3 / 2 π 1 / 2 e ( 1 / 8 ) 9 ⋅ 2 a − 1 p τ ( ψ 2 ) 2 τ ( χ 2 ) ( ψ p ( 4 ) τ ( ϵ p ) ( τ ( ψ p ) + ϵ p ( − 1 ) τ ( ϵ p ψ p ) ) 2 τ ( χ p ) ) = 32 e ( 2 − ϵ p ( − 1 ) 8 ) 2 3 / 2 π 1 / 2 9 ⋅ 2 a − 1 p τ ( ψ 2 ) 2 τ ( χ 2 ) ( Re ( J ( ψ p , ϵ p ) τ ( ϵ p ) ) + ϵ p ( − 1 ) ) . \begin{aligned}
C_\chi^2
&=16e\left(\frac{1-\epsilon_p(-1)}{8}\right)\frac{2^{3/2}\pi^{1/2}e(1/8)}{9\cdot2^{a-1}\sqrt{p}}\frac{\tau(\psi_2)^2}{\tau(\chi_2)}
\left(\frac{\psi_p(4)}{\tau(\epsilon_p)}\frac{\left(\tau(\psi_p)+\epsilon_p(-1)\tau(\epsilon_p\psi_p)\right)^2}{\tau(\chi_p)}\right)\\
&=32e\left(\frac{2-\epsilon_p(-1)}{8}\right)\frac{2^{3/2}\pi^{1/2}}{9\cdot2^{a-1}\sqrt{p}}\frac{\tau(\psi_2)^2}{\tau(\chi_2)}
\left(\operatorname{Re}\left(\frac{J(\psi_p,\epsilon_p)}{\tau(\epsilon_p)}\right)+\epsilon_p(-1)\right).
\end{aligned} C χ 2 = 16 e ( 8 1 − ϵ p ( − 1 ) ) 9 ⋅ 2 a − 1 p 2 3/2 π 1/2 e ( 1/8 ) τ ( χ 2 ) τ ( ψ 2 ) 2 ( τ ( ϵ p ) ψ p ( 4 ) τ ( χ p ) ( τ ( ψ p ) + ϵ p ( − 1 ) τ ( ϵ p ψ p ) ) 2 ) = 32 e ( 8 2 − ϵ p ( − 1 ) ) 9 ⋅ 2 a − 1 p 2 3/2 π 1/2 τ ( χ 2 ) τ ( ψ 2 ) 2 ( Re ( τ ( ϵ p ) J ( ψ p , ϵ p ) ) + ϵ p ( − 1 ) ) .
We know that this quantity is a positive real number. Since ∣ J ( ψ p , ϵ p ) / τ ( ϵ p ) ∣ = 1 \left|J(\psi_p,\epsilon_p)/\tau(\epsilon_p)\right|=1 ∣ J ( ψ p , ϵ p ) / τ ( ϵ p ) ∣ = 1 and C χ ≠ 0 C_\chi\ne0 C χ = 0 , the last factor is nonzero and has the sign of ϵ p ( − 1 ) \epsilon_p(-1) ϵ p ( − 1 ) . So if ϵ p ( − 1 ) = 1 \epsilon_p(-1)=1 ϵ p ( − 1 ) = 1 , then
e ( 1 / 8 ) τ ( ψ 2 ) 2 τ ( χ 2 ) e(1/8)\frac{\tau(\psi_2)^2}{\tau(\chi_2)} e ( 1/8 ) τ ( χ 2 ) τ ( ψ 2 ) 2
is positive. If ϵ p ( − 1 ) = − 1 \epsilon_p(-1)=-1 ϵ p ( − 1 ) = − 1 , then
e ( − 1 / 8 ) τ ( ψ 2 ) 2 τ ( χ 2 ) e(-1/8)\frac{\tau(\psi_2)^2}{\tau(\chi_2)} e ( − 1/8 ) τ ( χ 2 ) τ ( ψ 2 ) 2
is positive.
This covers the case when a a a is even or 2 2 2 is a square (mod p p p ). When a a a is odd and 2 2 2 is not a square, we also need to consider ϵ p ( − 2 ) \epsilon_p(-2) ϵ p ( − 2 ) , which is determined by p p p (mod 8 8 8 ), which is determined by α \alpha α (mod 8 8 8 ), and similar calculations lead to the second case of the proposition. □
Having evaluated the sign of τ ( ψ 2 ) 2 / τ ( χ 2 ) \tau(\psi_2)^2/\tau(\chi_2) τ ( ψ 2 ) 2 / τ ( χ 2 ) , we can examine C χ C_\chi C χ for characters modulo a power of 2 2 2 times an odd prime squared to do the same for odd primes. We omit the details of the following, which can also be deduced from work of Pigno and Pinner [28 Corollary 1.2] and the observation that τ ( ψ p ) 2 τ ( ψ p 2 ) = J ( ψ p , ψ p ) \frac{\tau(\psi_p)^2}{\tau(\psi_p^2)}=J(\psi_p,\psi_p) τ ( ψ p 2 ) τ ( ψ p ) 2 = J ( ψ p , ψ p ) .
Proposition 5.2. Suppose that ψ p \psi_p ψ p is a primitive Dirichlet character (mod p a p^a p a ), where p p p is an odd prime and a > 1 a>1 a > 1 . Let α \alpha α be an integer such that ψ p ( p + 1 ) = e ( α / ( p a − 1 ) ) . \psi_p(p+1)=e(\alpha/(p^{a-1})). ψ p ( p + 1 ) = e ( α / ( p a − 1 )) . Then
ψ p ( 4 ) τ ( ψ p ) 2 τ ( ψ p 2 ) = { p a / 2 , if a is even , ϵ p ( α ) p a / 2 , if a is odd and p ≡ 1 ( m o d 4 ) , − i ϵ p ( α ) p a / 2 , if a is odd and p ≡ 3 ( m o d 4 ) . \frac{\psi_p(4)\tau(\psi_p)^2}{\tau(\psi_p^2)}=
\begin{cases}
p^{a/2}, & \text{if }a\text{ is even},\\
\epsilon_p(\alpha)p^{a/2}, & \text{if }a\text{ is odd and }p\equiv1\pmod{4},\\
-i\epsilon_p(\alpha)p^{a/2}, & \text{if }a\text{ is odd and }p\equiv3\pmod{4}.
\end{cases} τ ( ψ p 2 ) ψ p ( 4 ) τ ( ψ p ) 2 = ⎩ ⎨ ⎧ p a /2 , ϵ p ( α ) p a /2 , − i ϵ p ( α ) p a /2 , if a is even , if a is odd and p ≡ 1 ( mod 4 ) , if a is odd and p ≡ 3 ( mod 4 ) .
Proof. If α \alpha α is a square (mod p p p ) or 2 2 2 is not a square (mod p p p ), then we can choose q = 2 a 2 p a q=2^{a_2}p^a q = 2 a 2 p a for some value of a 2 ≥ 6 a_2\ge6 a 2 ≥ 6 such that 2 a 2 + 1 α 2^{a_2+1}\alpha 2 a 2 + 1 α is a square (mod p p p ). If α \alpha α is not a square and 2 2 2 is, then we can choose some q = 2 a 2 p a p ′ a q=2^{a_2}p^ap'^{a} q = 2 a 2 p a p ′ a for some other prime p ′ ≥ 5 p'\ge5 p ′ ≥ 5 which is not a square (mod p p p ), so that 2 a 2 + 1 α p ′ 2^{a_2+1}\alpha p' 2 a 2 + 1 α p ′ is a square (mod p p p ).
Then for a primitive even χ 2 \chi_2 χ 2 (mod 2 a 2 2^{a_2} 2 a 2 ) with χ 2 ( 5 ) = e ( α 2 / 2 a 2 − 2 ) \chi_2(5)=e(\alpha_2/2^{a_2-2}) χ 2 ( 5 ) = e ( α 2 / 2 a 2 − 2 ) , α 2 ≡ − p a p ′ \alpha_2\equiv-p^ap' α 2 ≡ − p a p ′ (mod 8 8 8 ) (with p ′ p' p ′ possibly equal to 1 1 1 ) and any even primitive character χ p ′ \chi_{p'} χ p ′ (mod p ′ p' p ′ ), the product χ = χ 2 ψ p 2 χ p ′ \chi=\chi_2\psi_p^2\chi_{p'} χ = χ 2 ψ p 2 χ p ′ will satisfy the conditions of Proposition 1.7 and C χ ≠ 0 C_\chi\ne0 C χ = 0 . Then we can proceed as in the proof of the previous proposition and use the fact that C χ 2 C_\chi^2 C χ 2 is positive to get the result. □
Finally, putting these propositions together, we give an alternate expression for C χ C_\chi C χ .
Proposition 5.3. Suppose that C χ ≠ 0 C_\chi\ne0 C χ = 0 for a primitive character χ \chi χ (mod q q q ) with q = ∏ p ∣ q p e p q=\prod_{p\mid q}p^{e_p} q = ∏ p ∣ q p e p . Let ψ = ∏ p ∣ q ψ p \psi=\prod_{p\mid q}\psi_p ψ = ∏ p ∣ q ψ p be a primitive character (mod 2 q 2q 2 q ) such that ψ 2 = χ \psi^2=\chi ψ 2 = χ . Write q 2 = ∏ p ∥ q p q_2=\prod_{p\parallel q}p q 2 = ∏ p ∥ q p . Then
C χ 2 = 2 3 2 + 2 a π 1 2 9 q 1 2 ∏ p ∣ q 2 ϵ p ( − 2 q / p ) ∏ p ∣ q 2 ( 2 Re ( J ( ψ p , ϵ p ) τ ( ϵ p ) ) + 2 ϵ p ( − 2 q / p ) ) . C_\chi^2=\frac{2^{\frac32+2a}\pi^{\frac12}}{9q^{\frac12}}\prod_{p\mid q_2}\epsilon_p(-2q/p)\prod_{p\mid q_2}\left(2\operatorname{Re}\left(\frac{J(\psi_p,\epsilon_p)}{\tau(\epsilon_p)}\right)+2\epsilon_p(-2q/p)\right). C χ 2 = 9 q 2 1 2 2 3 + 2 a π 2 1 p ∣ q 2 ∏ ϵ p ( − 2 q / p ) p ∣ q 2 ∏ ( 2 Re ( τ ( ϵ p ) J ( ψ p , ϵ p ) ) + 2 ϵ p ( − 2 q / p ) ) . where
a = { ω ( q / q 2 ) + 1 / 2 , if 8 ∥ q , ω ( q / q 2 ) + 1 , if 16 ∥ q , ω ( q / q 2 ) + 3 / 2 , if 32 ∣ q . a=
\begin{cases}
\omega(q/q_2)+1/2, & \text{if }8\parallel q,\\
\omega(q/q_2)+1, & \text{if }16\parallel q,\\
\omega(q/q_2)+3/2, & \text{if }32\mid q.
\end{cases} a = ⎩ ⎨ ⎧ ω ( q / q 2 ) + 1/2 , ω ( q / q 2 ) + 1 , ω ( q / q 2 ) + 3/2 , if 8 ∥ q , if 16 ∥ q , if 32 ∣ q .
Proof. We work as in the proofs of the previous propositions, and as in Equation (18) , factoring S 2 ( ψ , − 1 ) S_2(\psi,-1) S 2 ( ψ , − 1 ) and the Gauss sum τ ( χ ) \tau(\chi) τ ( χ ) . We will find that
C χ 2 = B 2 3 2 + 2 a π 1 2 9 q 1 2 ∏ p ∣ q 2 ( 2 Re ( J ( ψ p , ϵ p ) τ ( ϵ p ) ) + 2 ϵ p ( − 2 q / p ) ) C_\chi^2=B\frac{2^{\frac32+2a}\pi^{\frac12}}{9q^{\frac12}}\prod_{p\mid q_2}\left(2\operatorname{Re}\left(\frac{J(\psi_p,\epsilon_p)}{\tau(\epsilon_p)}\right)+2\epsilon_p(-2q/p)\right) C χ 2 = B 9 q 2 1 2 2 3 + 2 a π 2 1 p ∣ q 2 ∏ ( 2 Re ( τ ( ϵ p ) J ( ψ p , ϵ p ) ) + 2 ϵ p ( − 2 q / p ) )
for some root of unity B B B . Since C χ 2 > 0 C_\chi^2>0 C χ 2 > 0 and each factor in the product is nonzero and has the sign of ϵ p ( − 2 q / p ) \epsilon_p(-2q/p) ϵ p ( − 2 q / p ) , we know that B = ∏ p ∣ q 2 ϵ p ( − 2 q / p ) B=\prod_{p\mid q_2}\epsilon_p(-2q/p) B = ∏ p ∣ q 2 ϵ p ( − 2 q / p ) . □
Now suppose that q = 2 m p q=2^mp q = 2 m p for an odd prime p p p . As ψ p \psi_p ψ p varies, the ratio J ( ψ p , ϵ p ) / τ ( ϵ p ) J(\psi_p,\epsilon_p)/\tau(\epsilon_p) J ( ψ p , ϵ p ) / τ ( ϵ p ) becomes equidistributed on the unit circle as p → ∞ p\to\infty p → ∞ (see the remark following Proposition 2.5 ), and we do not see any reason why the vanishing or nonvanishing of S ( χ ) S(\chi) S ( χ ) should interact with this equidistribution. Thus it seems that, after shifting and scaling, C χ 2 C_\chi^2 C χ 2 should be distributed like the real part of a random variable uniformly distributed on the unit circle. Provided that the choice of square root behaves as if it is independent of everything else, we should then also expect that C χ / A q C_\chi/A_q C χ / A q behaves like the real part of a uniformly distributed random variable on the unit circle (using the identity cos 2 x = 2 ( cos x ) 2 − 1 \cos2x = 2(\cos x)^2 - 1 cos 2 x = 2 ( cos x ) 2 − 1 ). Specifically, this suggests the following conjecture.
Conjecture 5.4. Fix m ≥ 3 m \ge3 m ≥ 3 , and for a prime p p p let q = 2 m p q = 2^m p q = 2 m p . For any subinterval [ α , β ] ⊆ [ − 1 , 1 ] [\alpha,\beta] \subseteq[-1,1] [ α , β ] ⊆ [ − 1 , 1 ] , as p → ∞ p \to\infty p → ∞ over the primes, we have
# { primitive χ ( m o d q ) such that C χ A q ∈ [ α , β ] ∖ { 0 } } # { primitive χ ( m o d q ) such that C χ ≠ 0 } ⟶ 1 π ∫ α β 1 1 − x 2 d x . \frac{\#\left\{ \text{primitive }\chi\pmod q\text{ such that }\frac{C_\chi}{A_q}\in[\alpha,\beta]\setminus\{0\}\right\}}{\#\left\{ \text{primitive }\chi\pmod q\text{ such that }C_\chi\ne0\right\}}
\longrightarrow\frac{1}{\pi}\int_{\alpha}^{\beta}\frac{1}{\sqrt{1-x^2}}\,\mathrm{d}x. # { primitive χ ( mod q ) such that C χ = 0 } # { primitive χ ( mod q ) such that A q C χ ∈ [ α , β ] ∖ { 0 } } ⟶ π 1 ∫ α β 1 − x 2 1 d x . More generally, if q = q 1 q 2 q=q_1q_2 q = q 1 q 2 , where q 1 q_1 q 1 is square-full (and divisible by 8) and q 2 q_2 q 2 is square-free, then for ω ( q 2 ) \omega(q_2) ω ( q 2 ) fixed, we should expect that as long as the smallest prime factor of q 2 q_2 q 2 tends to infinity, C χ / A q C_\chi/A_q C χ / A q should be asymptotically distributed like the product of ω ( q 2 ) \omega(q_2) ω ( q 2 ) independent copies of the real part of a uniform random variable on the unit circle. That is, we guess
lim P − ( q 2 ) → ∞ ω ( q 2 ) = k # { primitive χ ( m o d q ) such that C χ / A q ∈ [ α , β ] ∖ { 0 } } # { primitive χ ( m o d q ) such that C χ ≠ 0 } = ∫ α β ϕ k ( x ) d x , \lim_{\substack{P^-(q_2)\to\infty\\ \omega(q_2)=k}}
\frac{\#\left\{ \text{primitive }\chi\pmod q\text{ such that }C_\chi/A_q\in[\alpha,\beta]\setminus\{0\}\right\}}{\#\left\{ \text{primitive }\chi\pmod q\text{ such that }C_\chi\ne0\right\}}
=
\int_{\alpha}^{\beta}\phi_k(x)\,\mathrm{d}x, P − ( q 2 ) → ∞ ω ( q 2 ) = k lim # { primitive χ ( mod q ) such that C χ = 0 } # { primitive χ ( mod q ) such that C χ / A q ∈ [ α , β ] ∖ { 0 } } = ∫ α β ϕ k ( x ) d x ,
where P − ( q 2 ) P^{-}(q_2) P − ( q 2 ) denotes the smallest prime factor of q 2 q_2 q 2 and ϕ k ( x ) \phi_k(x) ϕ k ( x ) is the density function of the random variable ∏ j = 1 k cos ( 2 π X j ) \prod_{j=1}^{k}\cos(2\pi X_j) ∏ j = 1 k cos ( 2 π X j ) , where the X j X_j X j are independent and uniformly distributed on [ 0 , 1 ] [0,1] [ 0 , 1 ] .
If q 2 q_2 q 2 has a fixed small prime factor, then the distribution will be distorted by the small number of values taken by S 2 ( ψ p , r ) S_2(\psi_p,r) S 2 ( ψ p , r ) for this small prime factor, and the distribution may tend to something else. Figure 11 plots the distribution of C χ C_\chi C χ for a few q q q around one hundred million with differing factorization patterns.
Figure 11. Normalized histograms of C χ / A q C_\chi/A_q C χ / A q for the nonzero values of C χ C_\chi C χ for primitive characters, mod q q q for various q q q .
Finally, we briefly remark on the apparent relationship between C χ C_\chi C χ and low order biases in the spacing between zeros of L ( s , χ ) L(s,\chi) L ( s , χ ) . Recall from (1.7) that we define M χ + ( T ) M_\chi^{+}(T) M χ + ( T ) and M χ − ( T ) M_\chi^{-}(T) M χ − ( T ) respectively to be the measures of the sets of t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] for which Z χ ( t ) Z_\chi(t) Z χ ( t ) is positive and Z χ ( t ) Z_\chi(t) Z χ ( t ) is negative. Equivalently, M χ ( T ) : = M χ + ( T ) − M χ − ( T ) M_\chi(T) := M_\chi^{+}(T)-M_\chi^{-}(T) M χ ( T ) := M χ + ( T ) − M χ − ( T ) measures the difference between the sum of the even-indexed gaps g 2 n g_{2n} g 2 n and the sum of the odd-indexed gaps g 2 n − 1 g_{2n-1} g 2 n − 1 between zeros of L ( s , χ ) L(s,\chi) L ( s , χ ) on the critical line up to height T T T .
Figure 12 presents plots of M χ ( T ) M_\chi(T) M χ ( T ) for characters modulo 1616 = 2 4 ⋅ 101 1616=2^4\cdot101 1616 = 2 4 ⋅ 101 and 2312 = 2 3 ⋅ 17 2 2312=2^3\cdot17^2 2312 = 2 3 ⋅ 1 7 2 , where C χ C_\chi C χ does not vanish. In the latter case, ∣ C χ ∣ \lvert C_\chi\rvert ∣ C χ ∣ is constant when it does not vanish, as 2312 2312 2312 is square-full. For 1616 1616 1616 we see a range of nonzero values of C χ C_\chi C χ , and in Figure 13 we see what seems to be a clear linear relationship between C χ C_\chi C χ and M χ ( T ) M_\chi(T) M χ ( T ) . A natural guess is that there is some constant A A A such that when C χ ≠ 0 C_\chi\ne0 C χ = 0 we have
Figure 12. Plots of M χ + ( T ) − M χ − ( T ) M_\chi^{+}(T)-M_\chi^{-}(T) M χ + ( T ) − M χ − ( T ) , the difference between the amount of time that Z χ ( t ) Z_\chi(t) Z χ ( t ) spends positive and negative, for the 49 characters (mod 1616) with C χ ≠ 0 C_\chi\ne0 C χ = 0 and the 64 characters (mod 2312) with C χ ≠ 0 C_\chi\ne0 C χ = 0 .
Figure 13. C χ C_{\chi} C χ (horizontal axis) compared to M χ = M χ + ( 20000 ) − M χ − ( 20000 ) M_{\chi}=M_{\chi}^{+}(20000)-M_{\chi}^{-}(20000) M χ = M χ + ( 20000 ) − M χ − ( 20000 ) (vertical axis) for the 49 characters mod 1616 for which C χ ≠ 0 C_{\chi}\ne0 C χ = 0 .
M χ ( T ) ∼ A ⋅ C χ T 3 / 4 , M_\chi(T)\sim A\cdot C_\chi T^{3/4}, M χ ( T ) ∼ A ⋅ C χ T 3/4 ,
though some logarithmic factors could appear as well.
Appendix A. A direct proof that C is real
In this appendix, we give a direct proof that the constant C χ C_{\chi} C χ defined in Theorem 1.1 is real.
Lemma A.1. For every primitive character χ \chi χ modulo q q q , we have
S ( χ ) = τ ( χ ) q e ( − 1 / 8 ) S ( χ ) ‾ . S(\chi)=\frac{\tau(\chi)}{\sqrt{q}}e(-1/8)\overline{S(\chi)}. S ( χ ) = q τ ( χ ) e ( − 1/8 ) S ( χ ) . In particular, C χ C_{\chi} C χ is real.
Proof. When q q q is odd, the formula holds because S ( χ ) = 0 S(\chi)=0 S ( χ ) = 0 . Hence, we may assume that q q q is even. For primitive characters χ \chi χ modulo q q q , we have
χ ( n ) = τ ( χ ) q ∑ a = 1 q χ ( a ) ‾ e ( − a n / q ) , \chi(n)=\frac{\tau(\chi)}{q}\sum_{a=1}^{q}\overline{\chi(a)}e(-an/q), χ ( n ) = q τ ( χ ) a = 1 ∑ q χ ( a ) e ( − an / q ) ,
and thus
S ( χ ) = τ ( χ ) q ∑ a = 1 q χ ( a ) ‾ ∑ n = 1 2 q e ( − n 2 + 2 a n 2 q ) = τ ( χ ) q ∑ a = 1 q χ ( a ) ‾ e ( a 2 2 q ) ∑ n = 1 2 q e ( − ( n + a ) 2 2 q ) . (34) S(\chi)=\frac{\tau(\chi)}{q}\sum_{a=1}^{q}\overline{\chi(a)}\sum_{n=1}^{2q}e\left(-\frac{n^{2}+2an}{2q}\right)=\frac{\tau(\chi)}{q}\sum_{a=1}^{q}\overline{\chi(a)}e\left(\frac{a^{2}}{2q}\right)\sum_{n=1}^{2q}e\left(-\frac{(n+a)^{2}}{2q}\right). \tag*{(34)} S ( χ ) = q τ ( χ ) a = 1 ∑ q χ ( a ) n = 1 ∑ 2 q e ( − 2 q n 2 + 2 an ) = q τ ( χ ) a = 1 ∑ q χ ( a ) e ( 2 q a 2 ) n = 1 ∑ 2 q e ( − 2 q ( n + a ) 2 ) . (34)
upon completing the square. The quadratic Gauss sum can be evaluated as
∑ n = 1 2 q e ( − ( n + a ) 2 2 q ) = ∑ n = 1 2 q e ( − n 2 2 q ) = { 0 , q odd , 2 q e ( − 1 / 8 ) , q even . \sum_{n=1}^{2q}e\left(-\frac{(n+a)^{2}}{2q}\right)=\sum_{n=1}^{2q}e\left(-\frac{n^{2}}{2q}\right)=
\begin{cases}
0, & q\text{ odd},\\
2\sqrt{q}e(-1/8), & q\text{ even}.
\end{cases} n = 1 ∑ 2 q e ( − 2 q ( n + a ) 2 ) = n = 1 ∑ 2 q e ( − 2 q n 2 ) = { 0 , 2 q e ( − 1/8 ) , q odd , q even .
Since q q q is even, we have
∑ a = 1 q χ ( a ) ‾ e ( a 2 2 q ) = 1 2 S ( χ ) , \sum_{a=1}^{q}\overline{\chi(a)}e\left(\frac{a^{2}}{2q}\right)=\frac{1}{2}S(\chi), a = 1 ∑ q χ ( a ) e ( 2 q a 2 ) = 2 1 S ( χ ) ,
e n d s m i d = 1 {ends_mid=1} e n d s m i d = 1
and therefore, inserting the previous two identities into (A.1), we conclude that
S ( χ ) = τ ( χ ) q e ( − 1 / 8 ) S ( χ ) ‾ . S(\chi)=\frac{\tau(\chi)}{\sqrt{q}}e(-1/8)\overline{S(\chi)}. S ( χ ) = q τ ( χ ) e ( − 1/8 ) S ( χ ) .
Consequently, for either choice of square root,
e ( 1 / 16 ) S ( χ ) τ ( χ ) ∈ R , \frac{e(1/16)S(\chi)}{\sqrt{\tau(\chi)}}\in\mathbb{R}, τ ( χ ) e ( 1/16 ) S ( χ ) ∈ R ,
and hence C χ C_{\chi} C χ is real. This completes the proof of the lemma. □
Acknowledgments
JB is supported by the Heilbronn Institute for Mathematical Research. ZG was supported in part by a fellowship from the Fields Institute. MBM was supported in part by the NSF grant DMS 2401461. The authors would like to thank Valentin Blomer, Andy Booker, Brian Conrey, Jerzy Kaczorowski, and the referees for a number of useful comments that helped to improve the paper.
References