A Note on the Zeros of Zeta and L-Functions 1

Home , Riemann hypothesis, Riemann zeta function, Von Mangoldt function

A NOTE ON THE ZEROS OF ZETA AND L-FUNCTIONS

EMANUEL CARNEIRO, VORRAPAN CHANDEE AND MICAH B. MILINOVICH

1 Abstract. Let πS(t) denote the argument of the Riemann zeta-function at the point s = 2 + it. Assuming the Riemann hypothesis, we give a new and simple proof of the sharpest known bound for S(t). We discuss a generalization of this bound (and prove some related results) for a large class of L-functions including those which arise from cuspidal automorphic representations of GL(m) over a number ﬁeld.

1. Introduction

Let ζ(s) denote the Riemann zeta-function and, if t does not correspond to an ordinate of a zero of ζ(s), let 1 S(t) = arg ζ( 1 +it), π 2 where the argument is obtained by continuous variation along straight line segments joining the points 1 2, 2 + it, and 2 + it, starting with the value zero. If t does correspond to an ordinate of a zero of ζ(s), set 1 S(t) = lim S(t+ε) + S(t−ε) . 2 ε→0 The function S(t) arises naturally when studying the distribution of the nontrivial zeros of the Riemann zeta-function. For t > 0, let N(t) denote the number of zeros ρ = β + iγ of ζ(s) with ordinates satisfying 1 0 < γ ≤ t where any zeros with γ = t are counted with weight 2 . Then, for t ≥ 1, it is known that t t t 7 1 N(t) = log − + + S(t) + O . 2π 2π 2π 8 t In 1924, assuming the Riemann hypothesis, Littlewood [12] proved that log t S(t) = O log log t as t → ∞. The order of magnitude of this estimate has never been improved. Using the Guinand-Weil explicit formula for the zeros of ζ(s) and the classical Beurling-Selberg majorants and minorants for characteristic functions of intervals, Goldston and Gonek [9] established Littlewood’s estimate in the form 1 log t |S(t)| ≤ +o(1) (1.1) 2 log log t as t → ∞. The authors [4] recently sharpened this inequality and proved the following theorem. Theorem 1. Assume the Riemann hypothesis. Then 1 log t |S(t)| ≤ +o(1) 4 log log t for suﬃciently large t, where the term of o(1) is O(log log log t/ log log t).

2000 Mathematics Subject Classiﬁcation. 11M06, 11M26, 11M36, 11M41, 41A30. Key words and phrases. Riemann zeta-function, automorphic L-functions, Beurling-Selberg extremal problem, extremal functions, exponential type. 1 In this note, we give a new and much simpler proof of Theorem 1. As in the work of Goldston and Gonek mentioned above, we use the explicit formula in conjunction with the classical majorants and minorants of exponential type for characteristic functions of intervals that were constructed by Beurling and Selberg. In [4], two proofs of Theorem 1 are given. The more direct of these proofs proceeds as follows. Assuming the Riemann hypothesis, for t not corresponding to an ordinate of a zero of ζ(s), it is shown in [4] that 1 X S(t) = F (t−γ) + O(1), π γ where 1 x F (x) = arctan − , (1.2) x 1 + x2 1 and the sum runs over the nontrivial zeros ρ = 2 + iγ of ζ(s) counted with multiplicity. Drawing upon the work of Carneiro and Littmann [5], upper and lower bounds for S(t) are established by replacing F (x) in the sum over zeros by (optimally chosen) real entire majorants and minorants of F (x) and then applying the explicit formula. Using similar ideas, Chandee and Soundararajan [6] proved that log 2 log t log |ζ( 1 +it)| ≤ +o(1) 2 2 log log t as t → ∞, also under the assumption of the Riemann hypothesis. The theory of extremal functions of exponential type has also been recently used in [3] to improve the upper and lower bounds for the pair correlation of zeros of ζ(s) under the Riemann hypothesis. In Section 4, we discuss how to extend Theorem 1 to a large class of L-functions and prove a number of other related results. If an L-function is self-dual, our new proof of Theorem 1 generalizes in a natural way. However, if an L-function is not self-dual, then our new proof does not generalize but we show that an analogue of Theorem 1 is still possible using a modiﬁcation of our previous approach in [4].

2. Beurling-Selberg majorants and minorants

For functions R ∈ L1(R), let Z ∞ Rb(ξ) = R(x) e−2πixξ dx −∞ denote the Fourier transform of R. In the 1930s, Beurling observed that the real entire functions

2 ( ∞ ) 2 sin πz X sgn(m) 2 sin πz H±(z) = + ± π (z−m)2 z πz m=−∞

satisfy the inequalities H−(x) ≤ sgn(x) ≤ H+(x) for all real x, have Fourier transforms Hb ±(ξ) supported in [−1, 1], and satisfy Z ∞ ± H (x)−sgn(x) dx = 1. −∞ Moreover, Beurling showed that among all real entire majorants and minorants for sgn(x) with Fourier transforms supported in [−1, 1], the functions H±(x) minimize the L1(R)-distance to sgn(x). The details of this construction can be found in [20].

For t > 0, let χ[−t,t](x) denote the characteristic function of the interval [−t, t] (normalized at the end- points). From the above properties of Beurling’s functions H±(z), Selberg observed that the real entire functions 1 n o R±(z) = H±∆(t+z) + H±∆(t−z) (2.1) 2 2 − + satisfy the inequalities R (x) ≤ χ[−t,t](x) ≤ R (x) for all real x and have Fourier transforms supported in [−∆, ∆]. We summarize these (and some other relevant) properties of the functions R±(z) in the following lemma.

Lemma 2. Let t and ∆ be positive real numbers. The even real entire functions R±(z) deﬁned in (2.1) satisfy the following properties: − + (i) R (x) ≤ χ[−t,t](x) ≤ R (x) for all real x ; Z ∞ ± 1 (ii) R (x) − χ[−t,t](x) dx = ; −∞ ∆ (iii) R±(z) e2π∆|Im z| ; (iv) R±(x) min(1, ∆−2(|x| − t)−2) for |x| > t ; (v) Rb±(ξ) = 0 for |ξ| ≥ ∆ ; sin 2πtξ 1 (vi) Rb±(ξ) = + O for |ξ| ≤ ∆. πξ ∆ The implied constants in (iii), (iv) and (vi) are absolute.

Proof. This is a specialization of [9, Lemma 2]. For a further discussion on these Beurling-Selberg extremal functions, as well as proofs of the above assertions, see the survey articles of Montgomery [14], Selberg [19] and Vaaler [20].

3. Proof of Theorem 1

The functional equation for the Riemann zeta-function, in symmetric form, can be written as

ΓR(s)ζ(s) = ΓR(1−s)ζ(1−s)

s − 2 s where ΓR(s) = π Γ( 2 ). Lemma 3. For all t > 0, we have 1 Z t Γ0 R 1 N(t) = ( 2 +iu) du + S(t) + 1. 2π −t ΓR Proof. Using the functional equation for ζ(s), the argument principle, and the Schwarz reﬂection principle, it can be shown that 1 N(t) = arg Γ ( 1 +it) + S(t) + 1 π R 2 for any t > 0. See, for instance, [15, Theorem 14.1]. Since Z t Γ0 R 1 1 ( 2 +iu) du = 2 arg ΓR( 2 +it), −t ΓR the lemma follows.

1 Lemma 4. Let h(s) be analytic in the strip |Im s| ≤ 2 + ε for some ε > 0, and assume that |h(s)| (1 + |s|)−(1+δ) for some δ > 0 when |Re s| → ∞. Then

X ρ− 1 1 1 1 Z ∞ Γ0 h 2 = h + h − + h(u) Re R ( 1 +iu) du i 2i 2i π Γ 2 ρ −∞ R (3.1) ∞ 1 X Λ(n) log n log n − √ bh + bh − , 2π n 2π 2π n=2 3 where the sum on the left-hand side runs over the nontrivial zeros ρ of ζ(s) and Λ(n) is the von Mangoldt function deﬁned to be log p if n = pk, p a prime and k ≥ 1, and zero otherwise.

Proof. This is [9, Lemma 1]. For a similar formula, see [10, Equation (25.10)].

We now deduce Theorem 1 from Lemmas 2, 3 and 4.

Proof of Theorem 1. Assume the Riemann hypothesis and that ∆ ≥ 1 and t ≥ 10. By Lemma 2 (i), we have

X n + o X + 0 ≤ R (γ) − χ[−t,t](γ) = R (γ) − 2 N(t) (3.2) γ γ

1 + where the sum runs over the ordinates of the nontrivial zeros ρ = 2 +iγ of ζ(s). Note that R (u) is even and satisfies the conditions of Lemma 4. In fact, the growth condition required in Lemma 4 follows by Lemma 2 (iii)-(iv) via the Phragmén-Lindelöfprinciple. Using inequality (3.2) in conjunction with Lemma 3 and (3.1), we derive that 1 Z ∞ n o Γ0 1 ∞ Λ(n) log n 1 + R 1 X + + S(t) ≤ R (u)−χ[−t,t](u) ( +iu) du − √ Rb + R − 1. (3.3) 2π Γ 2 2π n 2π 2i −∞ R n=2 + 1 π∆ Lemma 2 (iii) implies that R 2i e , while Lemma 2 (v)-(vi) imply that the sum over prime powers in (3.3) is X Λ(n) 1 1 1 √ + eπ∆ n log n ∆ ∆ n≤e2π∆ by the Prime Number Theorem and partial summation. Therefore 1 Z ∞ n o Γ0 + R 1 π∆ S(t) ≤ R (u)−χ[−t,t](u) ( 2 +iu) du + O e . (3.4) 2π −∞ ΓR It remains to estimate the integral in (3.4). Note that, apart from the gamma factor, this is precisely the integral that Beurling and Selberg set out to minimize. Stirling’s formula for the gamma function implies that Γ0 1 u 1 R 1 2 ± iu = log + O ΓR 2 2π u as u → ∞. Hence, by Lemma 2 (iv), since ∆ ≥ 1, we see that Z ∞ n o Γ0 Z ∞ log u log t + R 1 R (u)−χ[−t,t](u) ( 2 +iu) du 2 2 du 2 1. 2t ΓR 2t ∆ (u−t) ∆ t Similarly, we have Z −2t n o Γ0 log t + R 1 R (u)−χ[−t,t](u) ( 2 +iu) du 2 1. −∞ ΓR ∆ t Thus, by Lemma 2 (ii), it follows that Z ∞ n o Γ0 Z 2t n o Γ0 + R 1 + R 1 R (u)−χ[−t,t](u) ( 2 +iu) du = R (u)−χ[−t,t](u) ( 2 +iu) du + O(1) −∞ ΓR −2t ΓR Z 2t log t n + o ≤ + O(1) R (u)−χ[−t,t](u) du + O(1) 2 −2t Z ∞ log t n + o ≤ + O(1) R (u)−χ[−t,t](u) du + O(1) 2 −∞ log t 1 = + O + 1 . 2∆ ∆ 4 Combining the above estimates, we find that log t S(t) ≤ + Oeπ∆ . 4π∆ Choosing π∆ = log log t − 2 log log log t, we deduce that 1 log t log t log log log t S(t) ≤ + O . 4 log log t (log log t)2 This is the upper bound implicit in Theorem 1. Similarly, starting with the observation that X R−(γ) − 2 N(t) ≤ 0, γ we can use Lemmas 2, 3 and 4 to derive the inequality log t S(t) ≥ − + Oeπ∆ . 4π∆ Again, choosing π∆ = log log t − 2 log log log t, it follows that 1 log t log t log log log t S(t) ≥ − + O . 4 log log t (log log t)2 This completes the proof of Theorem 1. A remarkable property of the Beurling-Selberg majorants and minorants for the characteristic function of an interval is that the L1(R)-distance of the approximations, as described in Lemma 2 (ii), is independent of the length of the interval. For this reason, the above proof essentially works equally well in bounding the number of zeros of ζ(s) with ordinates in the intervals [−t, t] as it does for bounding the number of zeros with ordinates in [0, t]. Therefore, since ζ(s) has exactly the same number of zeros with imaginary parts in the intervals [−t, 0] and [0, t], we are able to use the Beurling-Selberg majorants and minorants to bound S(t) − S(−t) = 2S(t) just as sharply as we are able to bound S(t) − S(0) = S(t). This is, apart from a few technical issues, the key point which enables us to improve Goldston and Gonek’s bound for S(t) in (1.1) by a factor of 2. These observations also allow us to generalize the proof of Theorem 1 to self-dual L-functions. However, in order to generalize Theorem 1 to L-functions which are not self-dual, it seems that the techniques in our previous paper [4] are necessary. We have chosen to give the proof for the zeros of ζ(s) separately, and in full detail, in order to illustrate the simplicity of the method.

4. Extension to general L-functions

In this section, we discuss how to prove an analogue of Theorem 1 for a large class of L-functions. Although our results are stated more generally, the L-functions we have in mind arise from automorphic representations of GL(m) over a number ﬁeld. Let m be a natural number. For Re(s) > 1, we suppose that an L-function, L(s, π), of degree m is given by the absolutely convergent Dirichlet series and Euler product of the form ∞ m −1 X λπ(n) Y Y αj,π(p) L(s, π) = = 1 − ns ps n=1 p j=1 where λπ(1) = 1, λπ(n) ∈ C, and the local parameters αj,π(p) are complex numbers which satisfy

ϑm |αj,π(p)| ≤ p 5 for a constant 0 ≤ ϑm ≤ 1 depending only on m. We deﬁne the completed L-function, Λ(s, π), by

Λ(s, π) = L(s, π∞) L(s, π)

where m s/2 Y L(s, π∞) = N ΓR(s+µj). j=1 s − 2 s Here, as above, ΓR(s) = π Γ( 2 ), the natural number N denotes the conductor of L(s, π), and the spectral parameters µj are complex numbers which are either real or come in conjugate pairs and are assumed 1 1 to satisfy the inequality Re(µj) > − 2 . We assume that the completed L-function Λ(s, π) extends to a meromorphic function of order 1 in C with at most poles at s = 0 and s = 1, each of order less than or equal to m. We further assume that Λ(s, π) satisﬁes the functional equation

Λ(s, π) = κ Λ(1−s, π˜)

where the root number κ is a complex number of modulus 1 and Λ(s, π˜) = Λ(¯s, π). Hence, by our assumptions

on the complex numbers µj, we have L(s, π˜) = L(¯s, π) and L(s, π˜∞) = L(s, π∞). If L(s, π˜) = L(s, π), then we say that L(s, π) is self-dual. Note that, by the functional equation, the (possible) poles of Λ(s, π) at s = 0 and s = 1 have the same order which we denote by r(π). As stated above, we have 0 ≤ r(π) ≤ m.

For t > 0, let N(t, π) denote the number of zeros ρπ = βπ + iγπ of Λ(s, π) which satisfy 0 ≤ βπ ≤ 1 and 1 2 −t ≤ γπ ≤ t where any zeros with γπ = ±t are counted with weight 2 . When t is not an ordinate of a zero of Λ(s, π), a standard application of the argument principle gives Z t 0 1 L 1 N(t, π) = 2 +iu, π∞ du + S(t, π) + S(t, π˜) + 2 r(π), (4.1) π −t L where Z ∞ 0 1 1 1 L S(t, π) = arg L 2 +it, π = − Im σ+it, π dσ. (4.2) π π 1/2 L If t does correspond to an ordinate of a zero of Λ(s, π), we deﬁne 1 S(t, π) = lim S(t+ε, π) + S(t−ε, π) . (4.3) 2 ε→0 Then the formula in (4.1) holds for all t > 0. We now prove an analogue of Theorem 1 for the functions S(t, π) in terms of the quantity m C(t, π) = C(π) |t| + 1 , where C(π) = C(0, π) is often called the analytic conductor of Λ(s, π) and is deﬁned by m Y C(π) = N |µj| + 3 . j=1

Theorem 5. Assume the generalized Riemann hypothesis for Λ(s, π). Then, for all t > 0, we have

3/m ! 1 ϑm log C(t, π) log C(t, π) log log log[C(t, π) ] |S(t, π)| ≤ + + O 2 (4.4) 4 2 log log[C(t, π)3/m] log log[C(t, π)3/m]

1If L(s, π) arises from an irreducible cuspidal representation of GL(m), then this assumption is known to hold (see [13]). 2Note the difference between the definition of N(t, π) and the definition for N(t) given in the introduction. If L(s, π) = ζ(s), then N(t, π) = 2N(t). 6 and, consequently, we have ! 1 log C(π) log C(π) log log log[C(π)3/m] ord Λ(s, π) ≤ + ϑm + O . (4.5) 1 3/m 3/m 2 s= 2 2 log log[C(π) ] log log[C(π) ] The implied constants above are universal.

Remarks. (i) If π is an irreducible cuspidal representation of GL(m) over a number ﬁeld, then Luo, Rudnick and 1 1 Sarnak [13] have shown that ϑm ≤ 2 − m2+1 . This bound can be improved when m ≤ 4 (see [1]). For instance, 7 it is known that ϑ2 = 64 from the work of Kim and Sarnak [11] and the corresponding generalization of their bound to number ﬁelds due to Nakasuji [16] and to Blomer and Brumley [1]. The generalized Ramanujan-

Petersson conjecture asserts that ϑm = 0 for all m ≥ 1. This conjecture is trivially true for the Riemann zeta- function and primitive Dirichlet L-functions, and is known for L-functions attached to classical holomorphic modular forms due the work of Deligne [7] and of Deligne and Serre [8]. (ii) If L(s, π) is the Riemann zeta-function, then we recover Theorem 1 from (4.4). Further examples are discussed in Section 6.

(iii) If L(s, π) is an L-function attached to an elliptic curve over Q, then the estimate in (4.5) recovers the well-known result of Brumer [2] on the analytic rank of an elliptic curve in terms of its conductor. In general, the estimate in (4.5) gives a sharpened form of [10, Proposition 5.21].

Proof of Theorem 5. By (4.2) and (4.3), it suffices to prove the inequality (4.4) in the case where t does not correspond to an ordinate of a zero of Λ(s, π). We start by recording the explicit formula for Λ(s, π), which can be established by modifying the proof of [10, Theorem 5.12]. For functions h satisfying the conditions in Lemma 4, we have 1 Z ∞ 0 X ρπ − 1 1 1 L h 2 = r(π) h + h − + h(u) Re 1 +iu, π du i 2i 2i π L 2 ∞ ρ −∞ π (4.6) ∞ 1 X 1 log n − log n − √ Λπ(n) bh + Λπ˜ (n) bh , 2π n 2π 2π n=2 where the sum runs over the zeros ρπ = βπ + iγπ of Λ(s, π) and the coefficients Λπ(n), which are supported on primes powers, are defined via the Dirichlet series

0 ∞ d L X Λπ(n) − log L(s, π) = − (s, π) = ds L ns n=1 which converges absolutely for Re(s) > 1. Logarithmically diﬀerentiating the Euler product, it can be shown that

ϑm |Λπ(n)| ≤ m Λ(n) n (4.7) where Λ(n) is the von Mangoldt function deﬁned in Lemma 4.

The self-dual case. We ﬁrst handle the case where L(s, π) is self-dual to show how our proof for the zeros of ζ(s) in the previous section can be extended. By Lemma 2 (i), assuming the generalized Riemann hypothesis for Λ(s, π), we have X n + o X + 0 ≤ R (γπ) − χ[−t,t](γπ) = R (γπ) − N(t, π).

γπ γπ 7 Since S(t, π) = S(t, π˜) and R+(z) is an even function, the formula for N(t, π) in (4.1) and the explicit formula imply that Z ∞ 0 1 + L 1 S(t, π) ≤ R (u) − χ[−t,t](u) Re 2 +iu, π∞ du 2π −∞ L 1 X Λπ(n) log n 1 − √ Rc+ + r(π) R+ − 1 . 2π n 2π 2i n≤e2π∆ Using parts (ii), (iii) and (vi) of Lemma 2, along with the estimate in (4.7), it follows that 1 Z ∞ n o L0 + 1 π(1+2 ϑm)∆ S(t, π) ≤ R (u) − χ[−t,t](u) Re 2 +iu, π∞ du + O m e 2π −∞ L m Z ∞ 0 1 1 N 1 X n o Γ +iu+µj = log + R+(u) − χ (u) Re 2 du + O m eπ(1+2 ϑm)∆ . 4π∆ πm 4π [−t,t] Γ 2 j=1 −∞ To estimate each integral term in the sum above, it is convenient to use Stirling’s formula in the form Γ0 1 (s) = log(s+1) − + O(1) (4.8) Γ s 1 1 which is valid for Re (s) > 0, say. Since Re (µj) > − 2 for each j, we note that Re ( 2 + iu + µj) > 0. Also, for each j, we break the integral into three ranges by writing Z ∞ 0 1 n + o Γ 2 +iu+µj R (u) − χ[−t,t](u) Re du = I1 + I2 + I3 −∞ Γ 2

where the range of integration in I1 is over the interval (−∞, −(|t| + 1)(|µj| + 3)), the range of integration

in I2 is over the interval [−(|t| + 1)(|µj| + 3), (|t| + 1)(|µj| + 3)], and the range of integration in I3 is over the

interval ((|t| + 1)(|µj| + 3), ∞).

To estimate I3, we use (4.8) and Lemma 2 (iv) to derive that Z ∞ 1 1 log[(|t| + 1)(|µj| + 3)] I3 2 2 log u + O(1) du 2 . (|t|+1)(|µj |+3) ∆ (u − t) ∆ (|t| + 1)(|µj| + 3)

The same bound holds for I1. It remains to estimate I2. If we write µj = aj + ibj, then some care is needed to handle the range of integration over the interval J = [−bj − 1, −bj + 1] since it may be the case that aj is 1 close to − 2 . We can estimate the contribution from the interval J to I2 by using (4.8) and either observing that the quantity 1 ( 1 + a ) Re − = − 2 j (4.9) 1 1 2 2 2 + iu + µj ( 2 + aj) + (u + bj) is negative (so that this portion of the integral may be omitted) or by observing that the function on the right-hand side of (4.9) is actually integrable in u (since it is a Poisson kernel) and using the fact that + R (u) − χ[−t,t](u) is uniformly bounded. From either observation, we see that the contribution to I2 from

the interval J is O(1). Therefore, letting I = [−(|t| + 1)(|µj| + 3), (|t| + 1)(|µj| + 3)], we have Z n + o 1 I2 ≤ R (u) − χ[−t,t](u) log (|t| + 1)(|µj| + 3) + 2 + |µj| du + O(1) I−J Z ∞ n + o ≤ R (u) − χ[−t,t](u) log (|t| + 1)(|µj| + 3) du + O(1) −∞ log[(|t| + 1)(|µ | + 3)] = j + O(1). ∆

8 Combining these estimates, we arrive at the inequality log C(t, π) log C(t, π) S(t, π) ≤ + O + O m eπ(1+2 ϑm)∆ . 4π∆ ∆2

3/m 3/m Choosing π (1+2 ϑm) ∆ = log log C(t, π) − 2 log log log C(t, π) , we derive the upper bound implicit in (4.4). Similarly, starting with the observation that

X − R (γπ) − 2 N(t, π) ≤ 0,

γπ we can derive the inequality log C(t, π) log C(t, π) S(t, π) ≥ − + O + O m eπ(1+2 ϑm)∆ . 4π∆ ∆2

3 3/m 3/m Again choosing π (1 + 2 ϑm) ∆ = log log C(t, π) − 2 log log log C(t, π) , we can establish the lower bound implicit in (4.4). This completes the proof of the bound for S(t, π) in Theorem 5 in the case where L(s, π) is self-dual. We note that if L(s, π) is not self-dual, then the above proof leads to the inequality

3/m ! 1 log C(t, π) log C(t, π) log log log[C(t, π) ] S(t, π) + S(t, π˜) ≤ + ϑm + O 2 , 2 log log[C(t, π)3/m] log log[C(t, π)3/m] and it does not seem possible to derive the inequality in (4.4) from this estimate.

The general case. In order to prove (4.4) for general L-functions (not necessarily self-dual), we follow the alternative approach described in [4]. Note that 1 Z 5/2 L0 S(t, π) = − Im σ+it, π dσ + O(m) π 1/2 L (4.10) Z 5/2 0 0 1 L L 5 = − Im (σ+it, π) − ( 2 +it, π) dσ + O(m). π 1/2 L L This estimate follows from (4.2) and the coeﬃcient bound in (4.7) since

Z ∞ 0 Z ∞ Z ∞ L X |Λπ(n)| X |Λπ(n)| X |Λπ(n)| Im σ+it, π dσ ≤ dσ = dσ = = O(m). L nσ nσ n5/2 log n 5/2 5/2 n≥1 n≥1 5/2 n≥1 To estimate the second integral on the right-hand side of (4.10), we use the partial fraction decomposition [10, Equation (5.24)]

0 m 0 L 1 N 1 X Γ s + µj r(π) r(π) X 1 1 − (s, π) = log m + − Bπ + + − + (4.11) L 2 π 2 Γ 2 s s − 1 s − ρπ ρπ j=1 ρπ where Bπ is a complex number. Now insert (4.11) into (4.10). Evidently, the constant terms cancel. Moreover, it is not diﬃcult to estimate the other terms which do not involve the zeros of Λ(s, π). Writing 1 5 s = σ + it with 2 ≤ σ ≤ 2 , we have r(π) Im = O(m) s and r(π) t Im = −r(π) . s − 1 (σ−1)2 + t2

3The power of 3/m can be replaced by a/m for any ﬁxed constant a > e (so that log log a > 0). 9 The function on the right-hand side of the second expression is integrable in σ as long as t 6= 0 (in fact, it is integrable from σ = −∞ to σ = ∞ since it is a Poisson kernel). Thus, the contributions from these terms 1 5 to (4.10) is O(m). We use (4.8) to handle the terms involving the gamma factors. For 2 ≤ σ ≤ 2 , we have Re (s + µj) > 0. Thus, for |s + µj| > 2, it follows that Γ0 s + µ s + µ Im j = Im log j + O(1) = O(1). Γ 2 2

If |s + µj| ≤ 2, we write µj = aj + ibj and observe that 0 Γ s + µj 2 2 (t + bj) Im = −Im + O(1) = 2 2 . Γ 2 s + µj (σ + aj) + (t + bj)

Again, this is integrable in σ since it is a Poisson kernel (and if t = −bj it is simply equal to 0). Summing each of these factors from j = 1 to j = m we get another term of size O(m). It therefore follows that

Z 5/2 0 0 1 L L 5 S(t, π) = − Im (σ + it, π) − 2 + it, π dσ + O(m) π 1/2 L L Z 5/2 1 X (t − γπ) (t − γπ) = − dσ + O(m) 1 2 2 2 π 1/2 (σ − ) + (t − γπ) 4 + (t − γπ) γπ 2 Z 5/2 1 X (t − γπ) (t − γπ) = − dσ + O(m) (4.12) 1 2 2 2 π 1/2 (σ − ) + (t − γπ) 4 + (t − γπ) γπ 2 1 X 2 2 (t − γπ) = arctan − 2 + O(m) π (t − γ) 4 + (t − γπ) γπ 1 X = G(t − γ ) + O(m) π π γπ

2 ± where G(x) = arctan(2/x) − 2x/(4 + x ). We now use the extremal functions m∆(x) of exponential type 2π∆ constructed in [4, Lemma 6].4 These are real entire functions satisfying the following properties:

(P1) For all real x we have C C − ≤ m−(x) ≤ G(x) ≤ m+ (x) ≤ 1 + x2 ∆ ∆ 1 + x2 where C is a universal constant and, for all complex z, we have 2 ± ∆ 2π∆|Im (z)| m (z) e . ∆ 1 + ∆|z| ± ± (P2) We have supp mb ∆ ⊂ [−∆, ∆] and mb ∆(ξ) 1 for all ξ ∈ [−∆, ∆].

(P3) The L1(R)-distances to G(x) are minimized, with values given by Z ∞ Z ∞ n + o n − o π m∆(x) − G(x) dx = G(x) − m∆(x) dx = . −∞ −∞ 2∆ By (4.12) and (P1), we have 1 X S(t, π) ≤ m+ (t − γ ) + O(m). π ∆ π γπ

4Note that G(x) = F (x/2) for F deﬁned in (1.2). The extremal functions in [4, Lemma 6] must be dilated accordingly. 10 + Using the explicit formula (4.6) with h(z) = m∆(t − z), the properties (P1) - (P3) above imply that m Z ∞ 0 1 1 N 1 X Γ + it − iu + µj S(t, π) ≤ log + m+ (u) Re 2 du + O m ∆ eπ(1+2 ϑm)∆ . 4π∆ πm 2π2 ∆ Γ 2 j=1 −∞

+ Integrating this against the majorant m∆(u) gives a term of size π π log (|it + µ | + 3) + O(1) ≤ log [(|t| + 1)(|µ | + 3)] + O(1). 2∆ j 2∆ j Summing over j and combining estimates, we arrive at the inequality log C(t, π) S(t, π) ≤ + O m ∆ eπ(1+2 ϑm)∆ . 4π∆ 3/m 3/m Choosing π (1+2 ϑm) ∆ = log log C(t, π) − 3 log log log C(t, π) , we derive the upper bound implicit in (4.4). The proof of the lower bound for S(t, π) implicit in (4.4) is analogous.

Proof of (4.5). Finally, the estimate in (4.5) follows from (4.1) and (4.4) by noting that

lim N(t, π) = ord Λ(s, π). t→0 1 s= 2 This completes the proof of Theorem 5. The bound for S(t, π) in Theorem 5 can improved in t-aspect for automorphic L-functions. In particular, let π be an irreducible unitary cuspidal representation of GL(m) over Q and let L(s, π) be the corresponding L-function. Setting Λπ(n) = Λ(n)aπ(n) and using the fact that the Rankin-Selberg L-function L(s, π × π˜) has a simple pole at s = 1, Rudnick and Sarnak [18, Section 2.4] observed that

2 2 X Λ(n)|aπ(n)| log n log x ∼ n 2 n≤x as x → ∞. Since 2 2 X |Λπ(n)| X Λ(n)|aπ(n)| log n ≤ n n n≤x n≤x

and the coeﬃcients Λπ(n) are supported on prime powers, an application of Cauchy’s inequality shows that

X |Λπ(n)| √ = O px log x n n≤x where the implied constant depends on π (and thus on C(π) and m). Using this estimate in the above proof, we can derive the following result.

Theorem 6. Let π be an irreducible unitary cuspidal representation of GL(m) over Q and let L(s, π) be the corresponding L-function. Then, assuming the generalized Riemann hypothesis for L(s, π), we have m log t log t log log log t |S(t, π)| ≤ + O 4 log log t (log log t)2 11 when t is suﬃciently large where the implied constant depends on m and the representation π.

5. The lowest zero of an L-function

In addition to bounding the order of vanishing at the central point, Theorem 5 also allows us to bound the height of the lowest zero of a completed L-function satisfying the hypotheses of the previous section in terms of its analytic conductor. In particular, we prove the following theorem (which sharpens a recent result of Omar [17] for automorphic L-functions). Theorem 7. Assume the generalized Riemann hypothesis for Λ(s, π). Then there is at least one zero 1 ρπ = 2 + iγπ of Λ(s, π) satisfying ! 1 π log log log C(π)3/m |γπ| ≤ + ϑm + O 2 2 log log[C(π)3/m] log log[C(π)3/m] when C(π)3/m is suﬃciently large where the implied constant is universal.

1 1 Proof. Observe that if Λ( 2 , π) = 0, then the theorem holds. Thus, we may assume that Λ( 2 , π) 6= 0. If t is small, say 0 < t ≤ 1, then by applying (4.8) and (4.9) we can show that

Z t 0 m Z t 0 1 1 L t 1 X Γ + iu + µj 1 + iu, π ) du = log N/πm + Re 2 du π L 2 ∞ π 2π Γ 2 −t j=1 −t t = log C(π) + O(m). π Using this estimate in (4.1) and observing that r(π) ≥ 0, it follows that t N(t, π) ≥ log C(π) + S(t, π) + S(t, π˜) + O(m) (5.1) π for 0 < t ≤ 1. When 0 ≤ t ≤ 1 we have

log C(π) ≤ log C(t, π) ≤ m log 2 + log C(π) (5.2)

which implies that

3 3 3 log m log C(π) ≤ log m log C(t, π) ≤ log 3 log 2 + m log C(π) . (5.3) Using (5.2) and (5.3) in Theorem 5 we get 3 ! 1 ϑm (m log 2 + log C(π)) (m log 2 + log C(π)) log log 3 log 2 + m log C(π) |S(t, π)| ≤ + + O 2 4 2 log log[C(π)3/m] log log[C(π)3/m] 3/m ! 1 ϑm log C(π) log C(π) log log log C(π) = + + O 2 + O(m). 4 2 log log[C(π)3/m] log log[C(π)3/m] From (5.1) we arrive at

3/m 3/m 3/m ! t m 3/m 1 m log C(π) m log C(π) log log log C(π) N(t, π) ≥ log C(π) − + ϑm + O 2 . π 3 2 3 log log[C(π)3/m] log log[C(π)3/m] The right-hand side of this inequality is positive when ! 1 π log log log C(π)3/m t ≥ + ϑm + O 2 , 2 log log[C(π)3/m] log log[C(π)3/m]

and this concludes the proof. 12 6. Examples

In this section, we state a few consequences of Theorem 5. For the deﬁnitions and relevant properties of the zeta and L-functions described in the following examples, see [10, Chapter 5].

Example 8. Let q > 2 be an integer, let χ be a primitive Dirichlet character modulo q, and let L(s, χ) be the corresponding Dirichlet L-function. Then Theorem 5 implies that 1 log q(|t|+1) 1 log q |S(t, χ)| ≤ + o(1) and ord L(s, χ) ≤ + o(1) 1 4 log log q(|t|+1) s= 2 2 log log q 1 1 assuming the generalized Riemann hypothesis for L(s, χ) where S(t, χ) = π arg L( 2 + it, χ).

Example 9. Let f be a holomorphic newform of weight k ≥ 1 and level q, let

X (k−1)/2 2πinz f(z) = λf (n) n e n≥1 P −s be its normalized Fourier expansion at the cusp ∞, and let L(s, f) = n≥1 λf (n)n for Re(s) > 1 be the corresponding L-function. Then √ √ 1 log q(k+3)(|t|+1) log q(k+3) |S(t, f)| ≤ + o(1) √ and ord L(s, f) ≤ 1 + o(1) √ 1 2 log log q(k+3)(|t|+1) s= 2 log log q(k+3) 1 1 assuming the generalized Riemann hypothesis for L(s, f) where S(t, f) = π arg L( 2 + it, f).

Example 10. Let K be a ﬁnite extension of Q with discriminant dK and degree [K : Q], and let ζK (s) 1/[K: ] denote the associated Dedekind zeta-function. Set DK = |dK | Q . Then [K : Q] log DK (|t|+3) 1 log(|dK |+3) |SK (t)| ≤ + o(1) and ord ζK (s) ≤ + o(1) 1 4 log log DK (|t|+3) s= 2 2 log log(|dK |+3) 1 1 assuming the generalized Riemann hypothesis for ζK (s) where SK (t) = π arg ζK ( 2 + it).

Acknowledgements.

EC acknowledges support from CNPq-Brazil grants 302809/2011 − 2 and 477218/2013 − 0, and FAPERJ grant E −26/103.010/2012. MBM is supported in part by an AMS-Simons Travel Grant and the NSA Young Investigator Grant H98230-13-1-0217.

References

[1] V. Blomer and F. Brumley, On the Ramanujan conjecture over number ﬁelds, Ann. of Math. 174 (2011), no. 1, 581–605. [2] A. Brumer, The average rank of elliptic curves. I, Invent. Math. 109 (1992) 445–472. [3] E. Carneiro, V. Chandee, F. Littmann and M. B. Milinovich, Hilbert spaces and the pair correlation of zeros of the Rie- mann zeta-function, to appear in J. Reine Angew. Math. (Crelle’s Journal). Preprint at http://arxiv.org/abs/1406.5462.

[4] E. Carneiro, V. Chandee and M. B. Milinovich, Bounding S(t) and S1(t) on the Riemann hypothesis, Math. Ann. 356 (2013), no. 3, 939–968. [5] E. Carneiro and F. Littmann, Bandlimited approximation to the truncated Gaussian and applications, Constr. Approx. 38, no. 1 (2013), 19–57. 1 [6] V. Chandee and K. Soundararajan, Bounding |ζ( 2 +it)| on the Riemann Hypothesis, Bull. London Math. Soc. 43 (2011), no. 2, 243–250. [7] P. Deligne, La conjecture de Weil. I, Inst. Hautes Etudes´ Sci. Publ. Math. No. 43 (1974), 273–307. [8] P. Deligne and J.-P. Serre, Formes modulaires de poids 1, Ann. Sci. Ecole´ Norm. Sup. (4) 7 (1974), 507–530 (1975). 13 [9] D. A. Goldston and S. M. Gonek, A note on S(t) and the zeros of the Riemann zeta-function, Bull. London Math. Soc. 39 (2007), 482–486. [10] H. Iwaniec and E. Kowalski, Analytic Number Theory, AMS Colloquium Publications, vol. 53 (2004). [11] H. Kim and P. Sarnak, Reﬁned estimates towards the Ramanujan and Selberg conjectures, J. Amer. Math. Soc. 16

(2003), 139–183, appendix to H. Kim, Functoriality for the exterior square of GL4 and the symmetric fourth of GL2. [12] J. E. Littlewood, On the zeros of the Riemann zeta-function, Proc. Camb. Philos. Soc. 22 (1924), 295–318. [13] W. Luo, Z. Rudnick and P. Sarnak, On Selberg’s eigenvalue conjecture, Geom. Funct. Anal. 5 (1995), no. 2, 387–401. [14] H. L. Montgomery, The analytic principle of the large sieve, Bull. Amer. Math. Soc. 84 (1978), 547–567. [15] H. L. Montgomery and R. C. Vaughan, Multiplicative Number Theory I. Classical Theory, Cambridge Studies in Ad- vanced Mathematics 97, Cambridge University Press (2007). [16] M. Nakasuji, Generalized Ramanujan conjecture over general imaginary quadratic ﬁelds, Forum Math. 24 (2012), no. 1, 85–98. [17] S. Omar, On small zeros of automorphic L-functions, C. R. Math. Acad. Sci. Paris 352 (2014), nos. 7-8, 551–556. [18] Z. Rudnick and P. Sarnak, Zeros of principal L-functions and random matrix theory. A celebration of John F. Nash, Jr. Duke Math. J. 81 (1996), no. 2, 269–322. [19] A. Selberg, Lectures on sieves, in Collected papers, Vol. II, Springer, Berlin (1991), 65–247. [20] J. D. Vaaler, Some extremal functions in Fourier analysis, Bull. Amer. Math. Soc. 12 (1985), 183–215.

IMPA - Instituto de Matematica Pura e Aplicada - Estrada Dona Castorina, 110, Rio de Janeiro, RJ, Brazil 22460-320 E-mail address: [email protected]

Department of Mathematics, Burapha University, 169 Long-Hard Bangsaen Road, Saen Sook Sub-district, Mueang District, Chonburi 20131, Thailand E-mail address: [email protected]

Department of Mathematics, University of Mississippi, University, MS 38677 USA E-mail address: [email protected]