Swiss Federal Institute of Technology Zurich Department of Mathematics Bachelor Thesis Fall 2010 Troy Koltes Soundararajan’s Upper Bound on Moments of the Riemann Zeta Function Submission Date: Advisor: December 2010 Dr. Paul-Olivier Dehaye CONTENTS iii Contents 1 Introduction 1 2 The 2.1 2.2 2.3 2.4 3 3 4 5 6 Main Theorem and its Corollary Preliminaries . . . . . . . . . . . . . . A conjectured bound for µ(S(T, V )) . Soundararajan’s main theorem . . . . An upper bound on Mk (T ) . . . . . . 3 Proof of the Main Theorem 3.1 Outline . . . . . . . . . . . 3.2 An upper bound on log(|ζ( 21 3.3 Two lemmas . . . . . . . . 3.4 Proof of Theorem 2.3 . . . . . . . . + it)|) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 . 9 . 9 . 19 . 23 4 Extension to the Family of Quadratic Dirichlet L-functions 29 5 Summary 31 Bibliography 32 iv CONTENTS Chapter 1 Introduction The Riemann zeta function is defined as ζ(s) = X 1 n≥1 ns , Re s > 1. By analytic continuation, it becomes a meromorphic function on the entire complex plane, with a simple pole at s = 1. It vanishes at even negative integers, the trivial zeros. The famous Riemann Hypothesis states that all non-trivial zeros of ζ(s) lie on the critical line Re s = 12 . We inspect moments of the Riemann zeta function. For T > 0 and k ∈ C, these are defined as Z T |ζ( 12 + it)|2k dt. Mk (T ) = (1.1) 0 Additionally assuming the Riemann Hypothesis, the integrand in Mk (T ) is thus integrated along the critical line of zeros of the Riemann zeta function. We will present a theorem by Soundararajan [8] which, eventually, implies the following asymptotic upper bound for Mk (T ) for real k > 0, conditional on the Riemann Hypothesis: Mk (T ) k,ε T (log T )k 2 +ε , ∀ε > 0. (1.2) Rather than deriving any new results, we try to make Soundararajan’s proof more explicit and include as many details as possible. In particular, an attempt is made at directly verifying all computations that are only sketched in the original paper. Finding bounds of Mk (T ) is a hard problem, even more so when one removes the assumption of the Riemann Hypothesis. For real k > 0, it has been speculated that, in fact, a stronger relation than the one stated above holds, without assuming the Riemann Hypothesis: 2 Mk (T ) ∼ Ck T (log T )k . (1.3) Keating and Snaith [4] have conjectured a precise value for Ck as a consequence of results from random matrix theory. Up to now, this asymptotic equivalence has only been proved for k = 1, 2 by Hardy/Littlewood and Ingham, respectively. For k = 1/m (with m a positive integer), a result by Heath-Brown [2] states that 2 2 T (log T )k k Mk (T ) k T (log T )k . 1 2 Introduction Jutila [3] improved on this result by showing that the dependence on k can be removed. Assuming that the Riemann Hypothesis is true, the lower bound Mk (T ) k T (log T )k has been proved by Ramachandra [6] for real k > 0. 2 In the following chapter, we present the main theorem, which gives us bounds on the measure of large values of |ζ( 12 + it)|. We also derive the corollary yielding the upper bound on Mk (T ) stated in Equation (1.2). The third chapter will discuss the proof of the main theorem in full detail. This requires a proposition which elegantly bounds the logarithm of the absolute value of ζ(s) along a certain part of the critical line. Finally, a mean value estimate for certain Dirichlet series helps us complete the proof of the main theorem. In Chapter 4, we describe an extension of the previous results to families of L-functions in the particular case of quadratic Dirichlet L-functions, under the assumption of the Generalized Riemann Hypothesis. Chapter 2 The Main Theorem and its Corollary 2.1 Preliminaries Throughout this presentation, we write logk (x) for log log . . . log(x) if k ≥ 3. | {z k times } Also, for two functions f, g : C → C we will use the standard notation f g ⇐⇒ f = O(g). We denote by µ the Lebesgue measure on (R, B(R)). The characterization of the moments of the Riemann zeta function will rely on the main theorem. This theorem bounds, for large T , the measure of the set S(T, V ) := {t ∈ [T, 2T ] | log |ζ( 12 + it)| ≥ V }. (2.1) The following restatement of Mk (T ) will be essential in proving the corollary that leads to the desired upper bound on the moments of ζ(s). For T > 0, Z 2T T |ζ( 12 + it)| 2k Z 2T dt = 1 e2k log |ζ( 2 +it)| dt (2.2) T = Z 2T Z log |ζ( 1 +it)| 2 T Z 2ke2kV dV dt −∞ 2T Z ∞ = 2k 1l T Z ∞ = 2k −∞ (t) 1 {t | log |ζ( 2 +it)|≥V } e2kV µ(S(T, V )) dV. · e2kV dV dt (2.3) (2.4) −∞ We can exchange integrals in (2.3) since (2.2) exists; this is the case because ζ( 12 + it) is bounded in the compact interval [T, 2T ]. 3 4 2.2 The Main Theorem and its Corollary A conjectured bound for µ(S(T, V )) The main Theorem 2.3, due to Soundararajan [8], will prove a weaker version of a conjecture further below. This conjecture is derived from a result by Selberg [7],[5]; to be more precise, Selberg proved a limit theorem (in the probabilistic sense) for the modulus of the Riemann zeta function on the critical line. It states that, unconditionally, for any fixed a ∈ R, q µ(S(T, a lim 1 2 log log T )) T T →∞ where = 1 − Φ(a), (2.5) x t2 1 e− 2 dt Φ(x) = √ 2π −∞ is the standard normal distribution function. One of the central steps of Selberg’s proof is to find a suitable approximation of ζ(s) on the critical line. In fact, the proof of the main theorem in the following chapter also strongly relies on a proposition with a similar bound. Z Conjecture 2.1. For V ≥ 0 and T large, ! √ log log T V2 µ(S(T, V )) T · exp − , V log log T (2.6) where the implicit constant is absolute. .q This conjecture is natural because a “substitution” of a in Equation (2.5) by a 1 2 log log T Φ0 (x) x , and the approximation 1 − Φ(x) ∼ for x → ∞, would lead to the speculated upper bound. Unfortunately, this substitution is not possible as Selberg’s theorem is only valid √ for V of size log log T . If one could prove Conjecture 2.1, this would have strong consequences for the upper bound on Mk (T ). The following corollary displays that we would obtain the upper bound 2 T (log T )k , with the implicit constant solely depending on k. 2 Corollary 2.2. If Conjecture 2.1 is true, then Mk (T ) k T (log T )k for real k > 0. Proof. Let Z be a random variable (on some probability space) such that Z ∼ N (0, 1). Then, Mk (2T ) − Mk (T ) = Z 2T T |ζ( 21 + it)|2k dt Z ∞ = 2k = 2k e2kV µ(S(T, V )) dV −∞ Z k log log T 2 e 2kV Z ∞ µ(S(T, V )) dV + 2k −∞ ≤ 2kT Z k log log T 2 e 2kV Z ∞ dV + 2k −∞ 2kT e k2 log log T Z ∞ + 2k k 2 log log T k 2 log log T √ 2kV e T k 2 log log T e2kV µ(S(T, V )) dV e2kV µ(S(T, V )) dV log log T V2 exp − V log log T ! dV 2.3 Soundararajan’s main theorem 5 √ ≤ 2kT e k2 log log T ≤ 2kT ek 2 log log T = 2kT ek 2 log log T ! log log T ∞ 2kV V2 e exp − + 2kT k dV log log T 2 log log T −∞ √ q 2 2 √ ( log log T ) + 4k T π k · E[exp (2k 12 log log T · Z)] 2 log log T √ 2 + 8kT π · ek log log T Z 2 k T (log T )k . (2.7) By telescoping and using Mk (0) = 0 it holds by (2.7) that Mk (T ) = ∞ X Mk (2−j T ) − Mk (2−j−1 T ) j=0 ∞ X k 2−j−1 T (log(2−j−1 T ))k 2 j=0 k2 ≤ T (log T ) ∞ X 1 j+1 2 j=0 2 = T (log T )k , which proves the corollary. 2.3 Soundararajan’s main theorem Theorem 2.3. Assume the Riemann Hypothesis. Let V ≥ 3 and T be large. √ If 10 log log T ≤ V ≤ log log T , then V V2 µ(S(T, V )) T √ exp − log log T log log T 4 1− log3 T ! . (2.8) If log log T < V ≤ 21 (log log T ) log3 T , then V V2 µ(S(T, V )) T √ exp − log log T log log T 7V 1− 4(log log T ) log3 T 2 ! . (2.9) If 21 (log log T ) log3 T < V , then µ(S(T, V )) T exp − 1 V log V 129 . (2.10) All implicit constants above are absolute. Proof. See p. 23. In the range 0 ≤ V ≤ log log T , Jutila [3] showed that a similar bound holds: V2 µ(S(T, V )) T exp − log log T where the implicit constants are absolute. V 1+O log log T ! , (2.11) 6 The Main Theorem and its Corollary Interestingly, to prove this approximation, Jutila showed that 2 T (log T )k Mk (T ) T (log T )k 2 1 is true in the case k = m , m ∈ N, T ≥ 2, with absolute implicit constants. This was an improvement on Heath-Brown’s [2] result (for the same k): 2 2 T (log T )k k Mk (T ) k T (log T )k . However, Soundararajan’s bound on Mk (T ) will be a consequence of the bounds on µ(S(T, V )) and not vice versa. It is clear that Soundararajan’s and Jutila’s bounds are closely related to Conjecture 2.1. Consequently, Soundararajan’s proof of the bound on Mk (T ) as a corollary of Theorem 2.3 will use an approach similar to the proof of Corollary 2.2. 2.4 An upper bound on Mk (T ) The following corollary, assuming the Riemann Hypothesis, gives the bound on Mk (T ) mentioned in the introduction; it is nearly of the conjectured magnitude described in Equation (1.3). Corollary 2.4. Assume the Riemann Hypothesis. Then for every real k > 0 and every ε > 0, it holds that Z T Mk (T ) = 0 |ζ( 12 + it)|2k dt k,ε T (log T )k 2 +ε . (2.12) Proof. See p. 8. To prove Corollary 2.4, we will use the following crude combination of Theorem 2.3 and Jutila’s bound in Equation (2.11). Proposition 2.5. Assume the Riemann Hypothesis. For every real k > 0, we have ( µ(S(T, V )) k T (log T )o(1) exp(−V 2 / log log T ), T (log T )o(1) exp(−4kV ), if 3 ≤ V ≤ 4k log log T if V > 4k log log T. √ Proof of Proposition 2.5. Note that in the range 3 ≤ V ≤ 10 log log T , Jutila’s theorem (see Equation (2.11)) yields V2 µ(S(T, V )) T exp − log log T ! V2 exp(o(1)) T exp − log log T ! . By the √ conditions imposed on V , we can rewrite the r.h.s. of (2.8) in the following way: If 10 log log T < V ≤ log log T , then V V2 µ(S(T, V )) T √ exp − log log T log log T 4 1− log3 T ! 2.4 An upper bound on Mk (T ) ≤T p 7 log log T (log T ) o(1) = T (log T ) 4 log3 T V2 exp − log log T V2 exp − log log T ! ! . The last step follows from the fact that 1 2 log3 T = o(1) · log log T and from taking exponentials, which yields (2.13) √ log log T = (log T )o(1) . If 4k ≤ 1, then the case 3 ≤ V < 4k log log T is taken care of by the two bounds above. If 4k > 1, we pick T large enough to assure that 4k log log T ≤ 21 (log log T ) log3 T . Then, in the range log log T < V ≤ 4k log log T (≤ 21 (log log T ) log3 T ), the r.h.s. of (2.3) can be bounded as follows: V V2 µ(S(T, V )) T √ exp − log log T log log T 7V 1− 4(log log T ) log3 T 2 ! V2 224k 3 log log T 49 log log T log log T exp − + − log log T log3 T 16(log3 T )2 ≤ 4kT p = 4kT p log log T · (log T ) 224k3 (log3 T )−49/16 (log3 T )2 V2 = T (log T )o(1) exp − log log T V2 · exp − log log T ! ! ! . The last step above is proved by means of a similar argument as in (2.13): log(4k) + 1 2 · log3 T + 224k 3 (log3 T ) − 49/16 (log log T ) = o(1) log log T, (log3 T )2 (2.14) By taking exponentials, the desired equality follows. For the range V > 4k log log T , we consider all three cases of Theorem 2.3 to find a simpler bound. If 4k log log T < V ≤ log log T (i.e., if k < 14 ), then µ(S(T, V )) T 4kV (log log T ) 4 log log T log log T exp − + log log T log3 T p = T (log T )o(1) exp(−4kV ). If log log T < 4k log log T < V ≤ 12 (log log T ) log3 T , let T be large enough to assure that 256k log log T < 21 (log log T ) log3 T . We consider two subcases: If log log T < 4k log log T < V ≤ 256k log log T , µ(S(T, V )) 256kT = 256kT p p 7 · 256k log log T 1− 4(log log T ) log3 T log log T exp −4kV 2 ! C(k) log log T D(k) log log T log log T exp −4kV + − log3 T (log3 T )2 8 The Main Theorem and its Corollary = T (log T )o(1) exp(−4kV ), where C(k), D(k) ∈ R are constants dependent on k. Considering the second subcase, if 256k log log T < V ≤ 12 (log log T ) log3 T , then we have Tp V2 µ(S(T, V )) log log T (log3 T ) exp − 2 log log T 256kV ≤ T (log T )o(1) exp − 64 o(1) = T (log T ) exp(−4kV ). 7 1− 8 2 ! We now turn to the last case. If 12 (log log T ) log3 T < 4k log log T < V , let T be large enough to ensure that V > e129·4k . Then, 1 µ(S(T, V )) T exp − V log e129·4k 129 = T exp(−4kV ) T (log T )o(1) exp(−4kV ). Hence, we have proved for V > 4k log log T that µ(S(T, V )) T (log T )o(1) exp(−4kV ). (2.15) Combining the above bounds yields the proposition. We can now use the weaker approximations from Proposition 2.5 to find the wanted upper bound on Mk (2T ) − Mk (T ) and, consequently, also on Mk (T ). Proof of Corollary 2.4, p.6. Let Z be a random variable (on some probability space) such that Z ∼ N (0, 1). Then, Mk (2T ) − Mk (T ) = Z 2T T |ζ( 12 + it)|2k dt Z ∞ = 2k ≤ 2kT e2kV µ(S(T, V )) dV −∞ Z 3 e2kV dV −∞ + 2kT (log T ) o(1) Z 4k log log T 2kV e 3 + 2kT (log T ) o(1) Z ∞ V2 exp − log log T ! dV e2kV −4kV dV 4k log log T k T e6k + T (log T )o(1) log log T · E[exp(2k p o(1) + T (log T ) q 1 2 log log T · Z)] 2 · exp(−8k log log T ) k T + T (log T )o(1) · ek 2 log log T + T (log T )o(1) 2 = T (log T )o(1)+k . Therefore, for any k > 0, ε > 0, it holds that Mk (2T ) − Mk (T ) k,ε T (log T )k 2 +ε . By telescoping in the same way as in Corollary 2.2, we can bound Mk (T ) by T (log T )k as well. The claim follows. 2 +ε Chapter 3 Proof of the Main Theorem 3.1 Outline The diagram in Figure 3.1 describes the relations between different lemmas, propositions and corollaries used in Soundararajan’s proof of Theorem 2.3 and of the bound on the moments Mk (T ) (Corollary 2.4). We have already seen that finding approximations for the measures of S(T, V ) is the crucial step in finding this bound on Mk (T ). The proof of Theorem 2.3 itself is the consequence of a bound on the function log |ζ( 12 + it)| by a sum involving the von Mangoldt function Λ(n) (and another term of similar magnitude), which will be described in Proposition 3.1. Interestingly, the contribution of zeros of ζ(s) near 12 + it to this bound is minor. To raise the effectiveness of the proposition, we show in Lemma 3.5 that a restriction of the sum to primes is unproblematic, i.e., the resulting error is, in an appropriate sense, negligible. Then, a second lemma (Lemma 3.6) gives us a general bound for mean values, or “moments”, of Dirichlet sums over primes. Applying this mean value estimate to the sum from the proposition and considering different ranges for the concerning variables eventually gives us the wanted upper bounds on µ(S(T, V )). The proof of the proposition bounding log |ζ( 12 + it)| will involve manipulating the product 0 expansion of ζζ (s) and Stirling’s approximation formula for Γ(s). During this computation, a need for a bound on log |ζ(σ0 + it)|, for σ0 > 1/2, arises (Corollary 3.4). This will be 0 taken care of by bounding ζζ (s) (using residue calculus) and integrating. 3.2 An upper bound on log(|ζ( 21 + it)|) As mentioned before, Soundararajan’s proof of Theorem 2.3 relies on the subsequent “tailor-made” bound on the logarithm of |ζ(s)| along the critical line, conditional on the Riemann Hypothesis. Denote by Λ the (arithmetical) von Mangoldt function: ( Λ(n) := log p, if n = pk for some prime p and some k ≥ 1 0, otherwise. 9 (3.1) 10 Proof of the Main Theorem Bound on the moments 2 Mk (T )k,ε T (log T )k +ε Corollary 2.4 Simplified bound on µ(S(T, V )) (2 cases) Proposition 2.5 General bound on µ(S(T, V )) (3 cases) Theorem 2.3 Reduction of the sum in the bound on log |ζ( 12 + it)| to a sum over primes Lemma 3.5 Bound on log |ζ( 12 + it)| by a sum involving Λ(n) Proposition 3.1 Mean value approximation of a short Dirichlet sum Lemma 3.6 Bound on log |ζ(σ0 + it)| for σ0 > 1/2 Corollary 3.4 Reformulation 0 of − ζζ (s) by residue calculus Lemma 3.3 Figure 3.1: Relations between different results in view of Theorem 2.3. The red (thick) boxes contain statements relying on the Riemann Hypothesis. 3.2 An upper bound on log(|ζ( 12 + it)|) 11 Proposition 3.1. Assume the truth of the Riemann Hypothesis. Let T be large, t ∈ [T, 2T ], and 2 ≤ x ≤ T 2 . Also, let λ0 be the unique positive real number such that e−λ0 = λ0 + 12 λ20 . Then the following upper bound holds for any λ ≥ λ0 : log |ζ( 21 + it)| ≤ Re log( nx ) (1 + λ) log T 1 . (3.2) + +O 1/2+λ/log x+it log n log x 2 log x log x n n≤x Λ(n) X Proof. See p. 16. Remark. The existence and uniqueness of a λ0 ∈ (0, ∞) as defined in Proposition 3.1 follows because f : [0, ∞) → R x 7→ −e−x + x + x2 2 (3.3) is strictly increasing, with f (0) = −1 and f (1) > 1. In fact, it holds that 0.491 < λ0 < 0.492. The following corollary of Proposition 3.1 gives a nice bound on the absolute value of ζ( 21 + it). Corollary 3.2. Assume the truth of the Riemann Hypothesis. For all large t, it holds that |ζ( 12 3 log t + it)| ≤ exp . 8 log log t Proof. We bound the logarithm in Equation (3.1) directly for x = (log T )2−ε , where ε > 0. Let t ∈ [T, 2T ]: log |ζ( 21 log n log( nx ) (1 + λ0 ) log T 1 + +O + it)| ≤ 1/2 2 log x log log T n log n log x n≤x X ≤ 1 X n≤(log T )2−ε n1/2 + (1 + λ0 ) log T + O(1) 4 − 2ε log log T (1 + λ0 ) log T + O(1) 4 − 2ε log log T log T (1 + λ0 ) = o(1) + . log log T 4 − 2ε ≤ (log T ) 2−ε 2 + For T large and after taking exponentials, choosing t = T yields |ζ( 21 + iT )| ≤ exp log T log log T The claim then follows by choosing ε < o(1) + (1 + λ0 ) 4 − 2ε . 2−4λ0 3 . Before proving Proposition 3.1, we state a lemma which allows us to deduce a corollary bounding log |ζ(σ0 + it)| for σ0 > 12 . This will eventually help us bound log |ζ(s)| on the critical line. Define R := {ρ 6= −2, −4, −6, . . . | ζ(ρ) = 0}, the nontrivial zeros of ζ(s). If the Riemann Hypothesis is assumed, write R = {ρ = 21 + iγ | ζ(ρ) = 0}, for γ real. 12 Proof of the Main Theorem Lemma 3.3. For any s 6= 1 with s ∈ / R, i.e. not coinciding with a zero of ζ(s) on the critical line, and for any x ≥ 2, it holds that X Λ(n) log(x/n) ζ0 1 − (s) = + s ζ n log x log x n≤x − 0 ζ ζ 0 (s) + 1 X xρ−s log x ρ (ρ − s)2 ∞ 1 X x−2k−s x1−s + , (1 − s)2 log x log x k=1 (2k + s)2 where the two sums converge absolutely. The sum over ρ is understood as X lim T →∞ |Im ρ|<T xρ−s . (ρ − s)2 Note that this lemma is unconditional on the truth of the Riemann Hypothesis. Proof. Recall the series expansion of the logarithmic derivative of ζ on Re z > 1, ∞ X Λ(n) ζ0 . − (z) = ζ nz n=1 (3.4) Write s = σ + it (with σ, t ∈ R) and set c := max(1, 2 − σ). If w ∈ [c − i∞, c + i∞], we have ( Re(s + w) = σ + c = σ + 1, 2, σ>1 σ ≤ 1. Concluding that Re(s + w) > 1, we can apply the series expansion (3.4) to the following integral, which yield: 1 2πi Z c+i∞ c−i∞ ζ0 xw 1 − (s + w) 2 dw = ζ w 2πi = Z c+i∞ X ∞ Λ(n) xw ns+w w2 c−i∞ n=1 dw Z c+i∞ w log(x/n) ∞ X Λ(n) 1 e n=1 ns 2πi c−i∞ w2 (3.5) dw. (3.6) The exchange of the sum and the integral above is permissible since the sum converges uniformly absolutely on Re w = c. To evaluate the integral in (3.6), we use residue calculus: First, let x < n; then log( nx ) < 0. If a, b ∈ C, denote the oriented line segment from a to b by [a, b]. For any B > 0, let C(B) be the right half-circleof radius B with center c, oriented clockwise, i.e. C(B) := B · e−iaπ | a ∈ [−1/2, 1/2] . Also define M (B) := C(B) ∪ [c − iB, c + iB]. Then, Z c+iB w log(x/n) e c−iB w2 Z dw = M (B) ew log(x/n) dw − w2 Z C(B) ew log(x/n) dw. w2 (3.7) Since the integrand is holomorphic in a neighborhood of the simply connected set with boundary M (B), an application of Cauchy’s integral theorem yields that the integral along 3.2 An upper bound on log(|ζ( 12 + it)|) 13 M (B) vanishes. Additionally, since log( nx ) < 0, we have Z πB ew log(x/n) dw sup exp B log( nx ) cos(aπ) ≤ 2 2 C(B) w B a∈[−1/2,1/2] π π x π ≤ B exp B log( n ) cos( 2 ) = B → 0, if B → ∞. Letting B → ∞ on both sides of Equation (3.7), it follows that Z c+i∞ w log(x/n) e w2 c−i∞ dw = 0, if x < n. Now, let x ≥ n; then log( nx ) ≥ 0. Similarly to the first case, for any B > c, let D(B) denote the left half-circleof radius B with center c, oriented counterclockwise, i.e. D(B) := B · eiaπ | a ∈ [1/2, 3/2] . Also set N (B) := D(B) ∪ [c − iB, c + iB]. Therefore, Z c+iB w log(x/n) e w2 c−iB Z dw = N (B) ew log(x/n) dw − w2 Z D(B) ew log(x/n) dw. w2 (3.8) The integrand has a pole at w = 0. Residue calculus yields Z N (B) ew log(x/n) ew log(x/n) dw = 2πi Res w=0 w2 w2 = 2πi Resw=0 1 + w log( nx ) + O(|w|2 ) w2 = 2πi log( nx ). As in the first case, Z ew log(x/n) πB π dw → 0, ≤ 2 exp B log( nx ) cos( π2 ) = 2 D(B) w B B if B → ∞. Thus, letting B → ∞ in Equation (3.8), we have Z c+i∞ w log(x/n) e c−i∞ w2 dw = 2πi log( nx ), if x ≥ n. We insert this result into (3.6) and find 1 2πi Z c+i∞ c−i∞ X Λ(n) ζ0 xw − (s + w) 2 dw = log(x/n). ζ w ns n≤x (3.9) For k, j ∈ Z>0 , set Bk = −Re(s) − 2k − 1. Also, select Tj in such a way that Tj → ∞ as j → ∞ and for all ordinates γ = Im ρ, where ρ ∈ R, we have ||γ|−Tj −Im s| (log(Tj ))−1 , for j large (see Davenport and Montgomery [1], Section 17). By our choice of Bk , we can apply some estimates by Davenport and Montgomery [1], Section 17. For fixed j and w ∈ [Bk + iTj , Bk − iTj ] we have the bound 0 ζ (s + w) log(2|s + w|) log(2|w|), ζ (3.10) 14 Proof of the Main Theorem keeping in mind that s is fixed. This estimate can be applied since Re(s + w) ≤ −1 for all k. For fixed k and w ∈ [Bk + iTj , c + iTj ] or w ∈ [Bk − iTj , c − iTj ], whenever Re(s + w) ≤ −1, a similar estimate is valid: 0 ζ (s + w) log(2|Tj |). ζ For w ∈ [Bk + iTj , c + iTj ] with −1 ≤ Re(s + w) ≤ 2, we have, by our choice of Tj , 0 ζ (s + w) log2 (|Tj |). ζ Finally, if Re(s + w) ≥ 2, ζ0 ζ (s + w) is bounded on these two segments. We can now integrate along the rectangle with vertices c − iTj , c + iTj , Bk + iTj , Bk − iTj and apply Cauchy’s theorem. Denote, in the same order, the four sides by γi , i = 1, 2, 3, 4. Then we have for fixed k: Z ζ0 xw Bk x−1 (c + 1)xc − (s + w) 2 dw ≤ 2 O(log(2Tj )) + O(log2 (Tj )), γ2,4 ζ w Tj + 1 Tj2 (3.11) where we split up the integration paths γ2 and γ4 at [1 + iTj ] and [1 − iTj ], respectively, estimating the two resulting integrals separately with the bounds of Davenport and Montgomery [1]. Thus, this integral vanishes for j → ∞. On γ3 for Tj → ∞, we rewrite Z ζ0 xw − (s + w) 2 dw = ζ w γ3 Z ∞ −∞ ζ0 xBk +it (s + Bk + it) dt . ζ (Bk + it)2 (3.12) Letting k → ∞, the integrand vanishes. To be able to exchange the limit and the integral, we need to show that the integrand is absolutely integrable, uniformly for large k. Choose a positive integer k0 ≥ −Re s/2. This implies that Bk ≤ −1 when k ≥ k0 . For these k, the integrand above is bounded in t ∈ [−1, 1]. Hence, using the fact that log(1 + y) ≤ y for y ≥ 0, we find Z ∞ 0 xBk +it ζ (s + Bk + it) dt ≤ O (Bk + it)2 −∞ ζ Z ∞ log(2|Bk + it|) 2 1 (Bk2 + t2 ) ! dt =O ! Z ∞ log(2) + 21 log(Bk2 + t2 ) dt 2 2 ≤O Z ∞ log t2 + (1 + Bk2 /t2 ) 1 1 Z ∞ ≤O 1 Bk + t t2 (1 + Bk2 /t2 ) 2 log t 1 + dt t2 t2 ! dt = O(1), uniformly in k ≥ k0 . Hence, we may apply the Dominated Convergence Theorem, which yields that the integral along [Bk + i∞, Bk − i∞] vanishes for k → ∞. Consequently, the two horizontal and the vertical integrals tend to 0 if we first let j → ∞, then k → ∞. 3.2 An upper bound on log(|ζ( 12 + it)|) 15 We can now use residue calculus to rewrite the l.h.s. of (3.5). 0 w Define f (w) := − ζζ (s + w) xw2 . Since we count residues in the region left of the line 0 Re w = c, the contributing residues of − ζζ (s+w) come from the region left of Re(w+s) = c + σ > 1 by definition of c. Therefore, all (trivial and non-trivial) zeros of ζ(s) contribute. 1 2πi Z c+i∞ f (w) dw = Resw=0 f (w) + Resw=1−s f (w) + X c−i∞ Resw=ρ−s f (w) ρ + ∞ X Resw=−2k−s f (w) k=1 ζ0 = − (s) log x − ζ 0 ζ ζ 0 (s) + ∞ X xρ−s X x1−s x−2k−s − − , 2 (1 − s)2 (−2k − s)2 ρ (ρ − s) k=1 where the sum over ρ is understood in the symmetric sense as X lim T →∞ |Im ρ|<T xρ−s . (ρ − s)2 Then this sum converges absolutely (see Davenport and Montgomery [1], Section 17), and the absolute convergence of the sum on the right follows by geometric summation and |x| > 1. The claimed equality follows after combining this with (3.9) and dividing by log x. 1 2 Corollary 3.4. Let T be large, σ0 > holds that and 2 ≤ x ≤ T 2 . For any s 6= 1 s.t. s ∈ / R, it Λ(n) log(x/n) 1 ζ0 − (σ0 + it) nσ0 +it log n log x log x ζ n≤x log |ζ(σ0 + it)| = Re X 1 X + log x ρ Z ∞ σ0 ! xρ−s dσ + O((log x)−1 ) , (ρ − s)2 (3.13) where the sum over ρ is understood as X Z ∞ lim T →∞ |γ|<T σ0 xρ−s dσ. (ρ − s)2 Proof. We integrate the real parts of both sides of the equality in Lemma 3.3 for fixed t ∈ [T, 2T ] from σ0 > 12 to ∞. For 2 ≤ x ≤ T 2 , we have log |ζ(σ0 + it)| = Z ∞ −Re σ0 ζ0 (σ + it) dσ ζ Λ(n) log(x/n) 1 ζ0 − (σ0 + it) nσ0 +it log n log x log x ζ n≤x = Re + + X 1 X log x ρ 1 log x Z ∞ σ0 Z ∞X ∞ σ0 xρ−s 1 dσ − 2 (ρ − s) log x ! x2k−s dσ . (2k + s)2 k=1 Z ∞ σ0 x1−s dσ (1 − s)2 16 Proof of the Main Theorem The sum over ρ can be exchanged with the integral since it converges absolutely (see Davenport and Montgomery [1]), uniformly in 12 ≤ σ < ∞. In fact, the integrals of the last two terms in the r.h.s. of Lemma 3.3 are O((log x)−1 ) for 2 ≤ x ≤ T 2 , t ∈ [T, 2T ]. This is true because looking at the integral of the former term for large x (and therefore also T ) yields: Z Z ∞ 1−σ ∞ x1−s 1 x dσ dσ ≤ σ0 (1 − s)2 log x log x σ0 t2 Z ∞ ≤ 1 T 2 log x e(1−σ) log x dσ σ0 x1−σ0 x(log x)2 1 . log x ≤ Also, bounding the integral of the last term of the r.h.s. of Lemma 3.3, we get ∞ Z ∞ ∞ 1 Z ∞X 1 X x−2k−σ x2k−s dσ ≤ dσ log x σ0 (2k + s)2 log x k=1 σ0 |2k + σ0 + it|2 k=1 ≤ 1 ∞ X x−2k σ02 xσ0 log x k=1 log x 1 , log x where the exchange of the sum and the integral is possible due to the uniform absolute convergence of the sum on 12 ≤ σ < ∞. The claim of the corollary follows. We now present the proof of Proposition 3.1. Using Stirling’s formula, we will bound the difference of log |ζ( 12 + it)| and log |ζ(σ0 + it)| for σ0 > 21 . Combining this with a version of Corollary 3.4, we obtain the desired estimate of log |ζ( 12 + it)|. Proof of Proposition 3.1, p.11. Since we are assuming the truth of the Riemann Hypothesis, we have R = {ρ = 21 + iγ | ζ(ρ) = 0}, with γ real. When considering s = σ + it ∈ C, we pick t in such a way that t 6= γ for all ρ ∈ R. Define F (s) = F (σ + it) := X 1 1 2σ − 1 + = s−ρ s−ρ (σ − 1/2)2 + (t − γ)2 Im(ρ)>0 Im(ρ)>0 X = σ − 1/2 . (σ − 1/2)2 + (t − γ)2 ρ∈R X Since X 1 1 1 + ≤ < ∞, 2 ρ ρ |ρ| Im(ρ)>0 Im(ρ)>0 X the series converges absolutely (see Davenport and Montgomery [1]). Clearly, F (s) ≥ 0 for σ ≥ 1/2. 3.2 An upper bound on log(|ζ( 12 + it)|) 17 By an application of the Hadamard product theorem, ζ(s) can be written as a product over its nontrivial zeros (see Titchmarsh [9]). For all s ∈ C, it holds that s exp(s(log(2π) − 1 − γ0 /2)) Y s eρ , 1− 1 ρ 2(s − 1)Γ(1 + 2 s) ρ ζ(s) = where γ0 = limn→∞ − log n + nk=1 n1 is the Euler-Mascheroni constant. Taking the logarithmic derivative, this yields P X ζ0 γ0 1 1 Γ0 1 1 , (s) = log(2π) − 1 − − − (1 + 12 s) + + ζ 2 s−1 2 Γ s − ρ ρ ρ (3.14) where the sum over ρ converges absolutely using X ρ and the convergence of we find 1 ρ |ρ|2 . P Considering the real parts of both sides of Equation (3.14), ζ0 γ0 (s) = log(2π) − 1 − ζ 2 γ0 = log(2π) − 1 − 2 Re 1 1 X s + = s−ρ ρ ρ(s − ρ) ρ − Re X 1 1 Γ0 Re − Re (1 + 21 s) + s−1 2 Γ ρ − Re X 1 1 1 Γ0 1 − Re (1 + 21 s) + F (s) + + . s−1 2 Γ ρ ρ Im(ρ)>0 1 1 + s−ρ ρ (3.15) √ The sum on the right of (3.15) evaluates to 1 − γ20 − log(2 π) (see Davenport and Montgomery [1], Section 12). Hence we can simplify (3.15): Re ζ0 1 1 1 Γ0 (s) = log(π) − Re − Re (1 + 12 s) + F (s). ζ 2 s−1 2 Γ (3.16) Stirling’s formula for Γ says that, for σ > 0, Γ(s) = 2π s 1 s 2 s e 1 + O(|s|−1 ) . (3.17) Again, we take the logarithmic derivative and get Γ0 1 (s) = log s − + O(|s|−1 ) = log s + O(|s|−1 ). Γ 2s (3.18) Inserting this approximation into (3.16), we find Re ζ0 1 1 1 (s) = Re + log π − Re log( 12 s + 1) + O(|s|−1 ) + F (s) ζ s−1 2 2 1 = − log( 12 s + 1) + F (s) + O(1). 2 (3.19) If we fix σ and take t ∈ [T, 2T ], then, for T large, −Re ζ0 1 (s) = log T − F (s) + O(1). ζ 2 (3.20) 18 Proof of the Main Theorem We integrate the l.h.s. of (3.20) for fixed t along the real part from Z σ0 1 2 −Re 1 2 to σ0 > 1 2 and get ζ0 (σ + it) dσ = Re log ζ( 12 + it) − log ζ(σ0 + it) ζ = log |ζ( 12 + it)| − log |ζ(σ0 + it)|. Doing the same for the r.h.s. of (3.20), we end up with Z σ0 log T 1 2 2 + O(1) − F (s) dσ = log T + O(1) (σ0 − 12 ) − 2 = (σ0 − Let h(x) = log(1 + x2 ) − x2 , 1+x2 1 2) Z σ0 F (σ + it) dσ 1/2 (σ0 − 12 )2 + (t − γ)2 log T 1X . + O(1) − log 2 2 ρ (t − γ)2 x ∈ R. Then h : R → R is smooth with h(0) = 0. Since h0 (x) = 2x3 , (1 + x2 )2 we find that 0 is a critical point and also a maximum of h. It follows that log(1 + x2 ) ≥ x2 . 1 + x2 We use this to find the bound log |ζ( 12 + it)| − log |ζ(σ0 + it)| = (σ0 − 12 ) ≤ (σ0 − 1 2) (σ0 − 1/2)2 log T 1X + O(1) − 2 2 ρ (σ0 − 1/2)2 + (t − γ)2 log T F (σ0 + it) − + O(1) . 2 2 (3.21) We now use Corollary 3.4 to remove the dependence on log |ζ(σ0 + it)| and conclude the claim of the proposition. Observe that the contribution to log |ζ(σ0 + it)| from the nontrivial zeros ρ ∈ R can be bounded as follows: XZ ∞ X Z ∞ xρ−s X x1/2−σ x1/2−σ0 dσ ≤ dσ = 2 2 2 σ0 (ρ − s)2 σ0 |σ0 + it − ρ| ρ ρ ρ ((σ0 − 1/2) + (t − γ) ) · log x = F (σ0 + it)x1/2−σ0 . (σ0 − 1/2) log x Working this bound and (3.20) into the r.h.s. of Corollary 3.4, we obtain log |ζ(σ0 + it)| ≤ Re X n≤x Λ(n) nσ0 +it log n + O((log x)−1 ) + log(x/n) log T F (σ0 + it) + − log x 2 log x log x F (σ0 + it)x1/2−σ0 . (σ0 − 1/2)(log x)2 We combine this inequality with (3.21) by addition, which yields a bound for the logarithm of |ζ(s)| on the critical line: 3.3 Two lemmas log |ζ( 21 19 log( nx ) log T Λ(n) + it)| ≤ Re + nσ0 +it log n log x 2 n≤x X 1 1 σ0 − + 2 log x x1/2−σ0 1 σ0 − 1/2 + F (σ0 + it) − − 2 (σ0 − 1/2)(log x) log x 2 ! + O((log x)−1 ). (3.22) Set σ0 = 12 + logλ x for any λ ≥ λ0 . Remember that for these λ, f as defined in (3.3) is nonnegative, hence e−λ − λ − λ2 /2 ≥ 0. Observe that x1/2−σ0 1 σ0 − 1/2 e−λ − λ − λ2 /2 − − = ≤ 0, (σ0 − 1/2)(log x)2 log x 2 λ log x since log x > 0. Therefore, the term including F (σ0 + it) in (3.22) is nonpositive and can be disregarded in the computation of an upper bound. This is a crucial step since this term includes the contribution from zeros ρ close to 12 + it. These zeros therefore have a minor effect on the bound for log |ζ( 12 + it)| and we obtain the claimed approximation log |ζ( 12 + it)| ≤ Re 3.3 log( nx ) (1 + λ) log T 1 + +O . (3.23) 1/2+λ/log x+it 2 log x log x n log n log x n≤x X Λ(n) Two lemmas Before we continue with the proof of the main theorem, we will need two lemmas. The first lemma shows that restricting the sum in Proposition 3.1 to primes gives only a small error term of order log3 T . The second lemma is a mean value estimate for Dirichlet series, which will be applied to Proposition 3.1 in the proof of the main theorem. Lemma 3.5. Assume the truth of the Riemann Hypothesis. Let T ≤ t ≤ 2T , 2 ≤ x ≤ T 2 and σ ≥ 12 . Then, X Λ(n) log(x/n) log3 T + O(1). nσ+it log n log x n≤x (3.24) n6=p Proof. First, consider the sum over n = pk with k ≥ 3. Then, X X 1 log(p) X 1 Λ(n) log(x/n) ≤ pkσ log(pk ) nσ+it log n log x p3/2 k k n≤x n=pk k≥3 p ≤x k≥3 (3.25) p ≤x k≥3 ≤ X 1 n≤x n3/2 1, (3.26) for x (and therefore T ) large. Now, we look at the contribution from n = p2 . Section 18 in Davenport and Montgomery [1] gives, under assumption of the Riemann Hypothesis, the following bound on P ψ(z) := n≤z Λ(n) for z ≥ 2 large: √ ψ(z) = z + O( z log2 z). 20 Proof of the Main Theorem Using Abel summation, we find X Λ(n) n≤z ns = ψ(z) +s zs Z z ψ(u)u−s−1 du. 1 Set s = 2it for T ≤ t ≤ 2T . Then, for z large, X log p p≤z p2it √ √ O( u log2 u) du + O( z log2 z) 2it 2it−1 2it 2it+1 n z u 1 u n≤z ! Z z√ √ 1 u log2 u 2it = 2it−1 + 2it−1 z log2 z) + 2itO du + O( 2it+1 z z (1 − 2it) u 1 z √ 2 + z log z T ≤ X Λ(n) = 1 + 2it Z z 1 + where the implicit constant is absolute. From this, we deduce a (conditional) upper bound on the sum over inverse primes to the 2it-th power, using Abel summation. Assume z ≤ T : X 1 p≤z p2it = X log p p≤z 1 p2it log p ! ! √ √ Z z u/T + u log2 (u) z z log2 z + +O =O du T log z log z u log2 u 2 √ √ z =O + z log z + O(1 + x) T log z √ = O( z log z), To we will split up the sum (over primes). Let √ prove our upper bound for prime squares, 2 ≤ y ≤ T and first assume that y > log4 T . Then, for T large, X X X 1 log(y/p) 1 log(y/p) 1 log(y/p) ≤ + 2σ+2it 2σ+2it 2σ+2it log y p log y p log y p≤y p p≤log4 T log4 T <p≤y X 1 log(y/p) X 1 log(y/p) + ≤ p log y 4 p2σ+2it log y p≤log4 T log T <p≤y X 1 X 1 log(y/p) ≤ + p 4 p2σ+2it log y p≤log4 T log T <p≤y X 1 log(y/p) ≤ O(log3 T ) + (3.27) 2σ+2it log y log4 T <p≤y p Clearly, from these bounds we can deduce that for y ≤ log4 T , we have X 1 log(y/p) = O(log T ). 3 2σ+2it log y p≤y p 3.3 Two lemmas 21 We turn back to the estimation of the right sum in Equation (3.27) with y > log4 T . Abel 1 summation with the arithmetical function a(n) := 1l{n prime} nlog(y/n) 2it log y and the C -function √ φ(y) := 1/y 2σ yields, for 2 ≤ y ≤ T , T ≤ t ≤ 2T and T large: X X log(y/p) X log(y/p) 1 log(y/p) 1 1 = − 4 2σ 2it 2it 2σ 2it p log y (log T )2σ log4 T <p≤y p p log y p≤y p log y y p≤log4 T Z y X log(y/p) 2σ du + 2σ+1 p2it log y log4 T u p≤u X X Z y 1 log p 2σ y ≤ 2σ + O(1) + du + 2σ+1 y log4 T u p≤u p2it p≤u p2it log y √ √ Z y u log u u log(u)2 1 u = O(1) + O + + du 2σ+1 2σ+1 log4 T u2σ+1 u log y |{z} T u log y ≤1 log T 1 1 + + 2 4 log T log (T ) log y log y ≤O ! = O(1), where the implicit constant is absolute since we assumed that σ is fixed. The bound on R y log2 u 2 log y √ . The claim of the lemma follows the rightmost integral follows by 1 u3/2 du = O y √ by replacing y with x. We now state and prove the second lemma, which will give us an upper bound on means of short Dirichlet series. Lemma 3.6. Let T be large and 2 ≤ x ≤ T . Choose k ∈ Z>0 in such a way that xk ≤ For any complex sequence (ap ), indexed by primes, it holds that k Z 2T X 2 X 2k ap |ap | . dt T k! 1/2+it p p T p≤x p≤x Proof. We rewrite the k-th power of the sum in the integrand as X ap p≤x p1/2+it k = X bk,x (n) The coefficient bk,x is defined as follows: Let n = ( bk,x (n) = n1/2+it n≤xk αi i=1 pi . Qr . Then, it follows that if ri=1 αi = 6 k Pr Q k r α i i=1 api , if i=1 αi = k. α1 ,...,αr P 0, This representation is a consequence of the multinomial theorem. Rewriting the integral in the claim yields Z 2T X ap 2k dt = p1/2+it T p≤x X m,n≤xk bk,x (m)bk,x (n) √ mn Z 2T it n T m dt T log T . (3.28) 22 Proof of the Main Theorem X |bk,x (n)|2 =T n n≤xk ! X +O m,n≤xk |b (m)bk,x (n)| √ k,x , mn| log(m/n)| (3.29) m6=n by separating the terms with m = n and m 6= n. By the binomial formula, it holds that √ 2|bk,x (m)bk,x (n)/ mn| ≤ |bk,x (m)|2 /m + |bk,x (n)|2 /n. Applying this to the rightmost term in (3.29), we find the bound X m,n≤x m6=n |b (m)bk,x (n)| √ k,x ≤ mn| log(m/n)| k X m,n≤xk m6=n X |bk,x (n)|2 X =2 n≤xk n m≤xk m6=n X |bk,x (n)|2 =4 n≤xk n n≤xk n≤xk X X |bk,x (n)|2 ≤ n X |bk,x (n)|2 n n−m n )| 1 ≤ 4xk n − m 1≤m<n n≤xk T n≤x 1 | log(1 − 1≤m<n n xk log(xk ) X |bk,x (n)|2 1 =4 | log(m/n)| n k X X |bk,x (n)|2 xk ≤4 |bk,x (m)|2 /m + |bk,x (n)|2 /n | log(m/n)| ≤4 1 | log(m/n)| 1≤m<n X X |bk,x (n)|2 n≤xk n n n−m 1≤m<n X 1 X |bk,x (n)|2 m n k X 1≤m<xk n≤x T log(T / log T ) X |bk,x (n)|2 log T n≤xk n , since xk ≤ T / log T . Hence, 2k X |bk,x (n)|2 . dt T Z 2T X ap 1/2+it p T p≤x n≤xk n The sum on the r.h.s. can be bounded by X |bk,x (n)|2 n≤xk n X = X p1 <···<pr ≤x αP 1 ,...,αr ≥1 αi =k k α1 , . . . , αr !2 |ap1 |2α1 · · · |apr |2αr pα1 1 · · · pαr r ! X ≤ k! X p1 <···<pr ≤x αP 1 ,...,αr ≥1 αi =k = k! X |ap |2 p≤x p k |ap1 |2α1 · · · |apr |2αr α1 , . . . , αr pα1 1 · · · pαr r k , where the inequality holds since k α1 , . . . , α r ! = k! ≤ k!. α1 ! · · · αr ! 3.4 Proof of Theorem 2.3 3.4 23 Proof of Theorem 2.3 Using the previous lemmas, we are now able to prove Theorem 2.3, using a few more bounds and suitable choices of parameters in Proposition 3.1. √ Proof of Theorem 2.3, p.5. We assume 10 log log T ≤ V ≤ large and V > 38 log T / log log T , we have 3 8 log T / log log T since for T µ(S(T, V )) = µ({t ∈ [T, 2T ] | log |ζ( 12 + it)| ≥ V }) = 0 by Corollary 3.2, so the range V > Define A := 3 8 log T / log log T does not contribute to the measure. 1 2 log3 T, log log T 2V if V ≤ log log T if log log T < V ≤ 21 (log log T ) log3 T if V > 21 (log log T ) log3 T. log3 T, 1, (3.30) Also set x = T A/V (≤ T ) and z = x1/ log log T . Inserting Lemma 3.5 into Proposition 3.1 for λ = λ0 and x, z as just defined, yields X x log( ) Λ(n) 1 + λ0 log T 1 n + + O log |ζ( 12 + it)| ≤ 2 log x log x n≤x n1/2+λ0 /log x+it log n log x log( xp ) 1 + λ0 X log p + V + O(log3 T ) ≤ 2A p≤x p1/2+λ0 /log x+it log p log x ≤ S1 (t) + S2 (t) + with 1 + λ0 V + O(log3 T ), 2A X log( xp ) 1 S1 (t) = p≤z p1/2+λ0 /log x+it log x and X x log( ) 1 p S2 (t) = . z<p≤x p1/2+λ0 /log x+it log x For any t ∈ S(T, V ), we have either S1 (t) ≥ V 1− 7 8A or S2 (t) ≥ V . 8A This statement is true since assuming the contrary, i.e., reversing the inequalities and making them strict would yield, using λ0 < 299/400: V ≤ log |ζ( 12 + it)| ≤ S1 (t) + S2 (t) + But V 400A 1 + λ0 V V + O(log3 T ) < V − + O(log3 T ). 2A 400A = O(log3 T ) does not hold for any choice of V and A (compare (3.30)). Let k ≤ V /A − 1 be a positive integer. By our choice of A and since T is large, we have A T V xk ≤ T V ( A −1) ≤ T 2 log log T log3 T ≤ T . log T 24 Proof of the Main Theorem This means that we can apply Lemma 3.6 with ap = 0, 1 pλ0 / log x p≤z p > z. log(x/p) log x , This yields Z 2T 2k |S2 (t)| T Z 2T X ap 2k dt = dt p1/2+it T (3.31) p≤x k log2 (x/p) T k! p1+2λ0 / log x log2 x z<p≤x 1 X ≤ T k! X 1 p z<p≤x k T (k(log log x − log log z + O(1)))k T k log log T A/V log T A/(V log log T ) !!k ! + O(1) k = T k(log3 T + O(1)) . (3.32) Fix k as the largest integer satisfying k ≤ V /A − 1. Observe that Z 2T 2k |S2 (t)| T V dt ≥ µ({t ∈ [T, 2T ] | S2 (t) ≥ V /(8A)}) 8A 2k . Using this together with (3.32) and absorbing the O(1) into log3 T , we deduce the bound µ ({t ∈ [T, 2T ] | S2 (t) ≥ V /8A}) T 8A V 2k (2k log3 T )k . (3.33) 2 Since T is large, observe √ that for any choice of V , we have A ≤ log3 T ≤ log(V ) = 2 log(V ) by our assumption 10 log log T ≤ V . Also note that 2k > V /A > 1. This gives a more concise bound in (3.33): T 8A V 2k 2V log(V 2 ) A (2k log3 T )k T V −2k (8A)2V /A !V /A = T V −2V /A (256AV log V )V /A T V −V /A (512 log2 V )V /A V T V − 2A V log V = T exp − 2A . The last bound follows from the fact that (512 log2 V )V /A = O(V V /(2A) ) because lim T →∞ 512 log2 V √ V !V /A = 0, (3.34) 3.4 Proof of Theorem 2.3 25 since T → ∞ implies that V → ∞ and A ≥ 1. In short, we find V µ ({t ∈ [T, 2T ] | S2 (t) ≥ V /8A}) T exp − log V 2A . (3.35) We proceed in a similar way to estimate µ ({t ∈ [T, 2T ] | S1 (t) ≥ V1 := V (1 − 7/(8A))}). Assume 1 ≤ k ≤ log(T / log T )/ log z, so z k√≤ T / log T is satisfied. Applying Lemma 3.6 and the simplified Stirling formula k! = O( k(k/e)k ) with the complex sequence ap = log(x/p) 1 , pλ0 / log x log x p≤z 0, p > z, we find Z 2T T |S1 (t)|2k dt T k! X1 p≤z p k √ k log log z k T k e √ k log log T k T k , e using z = T A/(V log log T ) ≤ T . By the same method as for S2 (t), the measure can be bounded in the following way: √ k log log T k µ ({t ∈ [T, 2T ] | S1 (t) ≥ V1 }) T k . eV12 (3.36) Choose ( k= bV12 / log log T c, V ≤ (log log T )2 b10V c, V > (log log T )2 . Selected in this way, for V ≤ log log T , we have A = k≤ 1 2 log3 T and V12 V2 log(T / log T ) 2 log log T log(T / log T ) ≤ ≤V ≤ V = , log log T log log T log T log3 T log z log T since T is large. For log log T < V ≤ (log log T )2 , we have A = log2V log3 T and we find that V2 2 log(T / log T ) 2 log(T / log T ) V12 k≤ ≤ ≤ V = , log log T log log T log T (log3 T ) log z since T is large. Further, in the second range for k, if V > (log log T )2 , we use A = 1. It holds that log(T / log T ) log log T log(T / log T ) k ≤ 10V ≤ V = log T log z since T is large. We conclude that k ≤ log(T / log T )/ log z is satisfied. Hence, with this choice of k we can further bound Equation (3.36): 26 Proof of the Main Theorem µ ({t ∈ [T, 2T ] | S1 (t) ≥ V1 }) V1 T√ log log T V 2 / log log T 1 1 e √ +T V V12 V exp − T√ log log T log log T ! V V12 T√ exp − log log T log log T ! V12 V exp − T√ log log T log log T ! 10V log log T eV 2 (1 − 7/(8A))2 10V eV (1 − 7/(8A))2 √ 10 V log V + T exp − 10V log 2 + T exp log V + 10V log(256) − 5V log V 2 !! + T exp (−4V log V ) . (3.37) Summarizing the above, we have V12 V exp − µ ({t ∈ [T, 2T ] | S1 (t) ≥ V1 }) T √ log log T log log T ! + T exp (−4V log V ) . (3.38) Combining (3.35) and (3.38) will yield the claim of the theorem. Since µ(S(T, V )) ≤ µ ({t ∈ [T, 2T ] | S1 (t) ≥ V1 or S2 (t) ≥ V /(8A)}) ≤ µ ({t ∈ [T, 2T ] | S1 (t) ≥ V1 }) + µ ({t ∈ [T, 2T ] | S2 (t) ≥ V /(8A)}) , we estimate these two measures for each of the three ranges in Theorem 2.3 separately. Set B1 (T, V ) := {t ∈ [T, 2T ] | S1 (t) ≥ V1 } and B2 (T, V ) := {t ∈ [T, 2T ] | S2 (t) ≥ V /(8A)}. √ √ If 10 log log T < V ≤ log log T , then A = 12 log3 T . Bounding (3.35) for 10 log log T < V ≤ 12 log log T and using 2V / log log T ≤ 1 yields V µ(B2 (T, V )) T exp − log V log3 T V log3 T T exp − (log 10 + ) log3 T 2 2V 2 T exp − log log T V2 T exp − log log T ! 2 1 + log3 T 2 V V2 T√ exp − log log T log log T In the second subcase, 1 2 ! 4 1− log3 T ! . log log T < V ≤ log log T , we find, using V / log log T ≤ 1, V log V log3 T V T exp − (− log 2 + log3 T ) log3 T µ(B2 (T, V )) T exp − V2 T exp − log log T 4 1− log3 T ! 3.4 Proof of Theorem 2.3 27 V2 V exp − T√ log log T log log T 4 1− log3 T ! . For√ the bound on (3.38), it is not necessary to consider these subcases separately. For 10 log log T < V ≤ log log T , we directly find V V2 µ(B1 (T, V )) T √ exp − log log T log log T 7 1− 4 log3 T 2 ! + T exp(−4V log V ) ! 14 V V2 1− T√ exp − log log T 4 log3 T log log T log3 T + T exp −4V log 10 + 2 ! 2 V V 4 T√ exp − 1− + T exp (−8V ) log log T log3 T log log T V V2 T√ exp − log log T log log T 4 1− log3 T ! V2 V exp − T√ log log T log log T 4 1− log3 T ! 8V 2 + T exp − log log T ! . √ This proves Theorem 2.3 for the range 10 log log T < V ≤ log log T . If log log T < V ≤ 12 (log log T ) log3 T , then A = log log T 2V V 2 log V µ(B2 (T, V )) T exp − (log log T ) log3 T V2 T exp − log log T log3 T . Bounding (3.35) yields ! ! V V2 T√ exp − log log T log log T 7V 1− 4(log log T ) log3 T 2 ! . The bound on (3.38) follows directly from this choice of range for V : V V2 µ(B1 (T, V )) T √ exp − log log T log log T 7V 1− 4(log log T ) log3 T 2 ! 7V 1− 4(log log T ) log3 T 2 ! 7V 1− 4(log log T ) log3 T 2 ! + T exp(−4V log V ) V2 V exp − T√ log log T log log T V2 + T exp − log log T ! V V2 T√ exp − log log T log log T This proves Theorem 2.3 for the range log log T < V ≤ 21 (log log T ) log3 T . . 28 Proof of the Main Theorem If 12 (log log T ) log3 T < V , then A = 1. Again, we first bound (3.35): µ(B2 (T, V )) T exp − V log V 2 1 V log V 129 T exp − . Finally, we want to prove that µ(B1 (T, V )) T exp − 1 V log V 129 . (3.39) The right term in the r.h.s. of Equation (3.38) clearly fulfills this bound, so we only need to consider the left term. We want to show that V V2 1 √ exp − + V log V 64 log log T 129 log log T ! = o(1). (3.40) Since V > 12 (log log T ) log3 T , write V = α(log log T ) log3 T with some α > 1/2 which may depend on T . Keeping in mind that T is large, the l.h.s. of (3.40) is reformulated as: exp log α + log3 T α2 (log log T )(log3 T )2 + log4 T − 2 64 1 α log log T (log3 T )(log α + log3 T + log4 T ) 129 !! 1 1 log α 1 2 2 + + = exp α log log T (log3 T ) o(1) − 64 129α 129 α log3 T + | 2 2 = exp α log log T (log3 T ) 1 1 + o(1) − | 64 {z 129α} {z = o(1) } !! (3.41) <0 = o(1), since α > 1/2 (uniformly) implies that the exponent in (3.41) tends to −∞ for T → ∞. Hence, we have shown that the bound in (3.39) holds, which completes the proof of Soundararajan’s Theorem 2.3. Chapter 4 Extension to the Family of Quadratic Dirichlet L-functions Under the assumption of the Generalized Riemann Hypothesis, Soundararajan has extended Proposition 3.1, which is the crucial element of the proof of Theorem 2.3, to work for certain families of L-functions as well (see [8]). In particular, he gives a modified version of the bound in Proposition 3.1 for quadratic Dirichlet L-functions. Recall that an integer d is a fundamental discriminant if and only if one of the following conditions holds: d ≡ 1 (mod 4) and d is square-free d = 4m for some square-free integer m ≡ 2 or 3 (mod 4). Proposition 4.1. Assume the Generalized Riemann Hypothesis. Let x ≥ 2, λ0 as defined in Proposition 3.1 and d be a fundamental discriminant. Denote by χd the corresponding d-th primitive quadratic character, defined as the Kronecker symbol: χd (n) = nd . Then, log L( 12 , χd ) Λ(n)χd (n) log(x/n) 1 + λ0 log |d| 1 + +O . ≤ 1/2+λ / log x 0 2 log x log x n log n log x 2≤n≤x X Note that, due to the Generalized Riemann Hypothesis, we have L( 12 , χd ) ≥ 0: Assuming L( 21 , χd ) < 0 would imply the existence of a real zero ρ with 12 < ρ < 1 since L(2, χd ) > 0 and L(σ, χd ) 6= 0 for real σ ≥ 1. If L( 12 , χd ) = 0, we interpret log L( 21 , χd ) as −∞ to keep the inequality in the proposition valid. Comparing this with the original proposition for | log(ζ(s))|, we see that the real variable t has been replaced by the fundamental discriminant d, for example in the numerator (log |d|)(1 + λ0 ) of the second term. We also remark that the sum in the proposition restricted to prime squares now is of order 21 log log x instead of log3 x due to the fact that χd (p2 ) = dp dp = 1 when p - d and that the “dampening” effect of 1/pit in the sum is not present any more. After some more work which consists in finding a statement similar to Lemma 3.6, Soundararajan showed that it is possible to use the modified proposition to find an analogue of Theorem 2.3. Since d is discrete, the measure applied in the 29 30 Extension to the Family of Quadratic Dirichlet L-functions modified case is the counting measure, i.e., the modified theorem bounds the number of d with absolute value ≤ X such that log L( 12 , χd ) ≥ V + 12 log log X. Notice the additional 1 2 log log X compared to the definition of the set S(T, V ) for the Riemann zeta function due to the larger effect of the sum in Proposition 4.1 restricted to prime squares. More precisely, Soundararajan stated the modified theorem for 2 different ranges of X: Theorem 4.2. Assume the Generalized Riemann Hypothesis. For X ≥ 2 large and √ log log X ≤ V = o((log log X) log3 X), it holds for fundamental discriminants d that ! #{d | X ≥ |d| and log L( 21 , χd ) V2 ≥ V + log log X} X exp − (1 + o(1)) . 2 log log X 1 2 If V ≥ (log log X) log3 X, we have #{d | X ≥ |d| and log L( 21 , χd ) ≥ V + 21 log log X} X exp (−cV log V ) for a constant c > 0. We see that the modified theorem states quite similar upper bounds to the ones for the Riemann zeta function. Finally, this theorem is used to find a bound on the moments of L( 12 , χ): Corollary 4.3. Assume the Generalized Riemann Hypothesis. Let X ≥ 2 be large. Then for every real k > 0 and every ε > 0, it holds that X 1 L( 21 , χd )k k,ε X(log X) 2 k(k+1)+ε , |d|≤X where the sum is over fundamental discriminants d. Again, we see the similarity to the original bound apart from the additional term k/2 in the exponent, which arises because we are bounding the number of d with absolute value ≤ X such that log L( 12 , χd ) ≥ V + 21 log log X instead of just log L( 12 , χd ) ≥ V as in the original Corollary 2.4. To prove Corollary 4.3, the sum is interpreted as an integral with respect to an appropriate counting measure, and a similar reformulation as seen in Equation (2.4) is used. Implementing the bounds on the (counting) measure from the modified Theorem 4.2 yields the stated result. Chapter 5 Summary Approximating moments of the Riemann zeta function is a difficult problem. In this presentation we followed Soundararajan’s proof of the bound Mk (T ) k,ε T (log T )k 2 +ε for real k > 0 and ε > 0, under the assumption of the Riemann Hypothesis (Corollary 2.4). An attempt was made to give as many details as possible to make the proof more explicit than in the original paper. To prove the approximation of the moments, a theorem bounding the measure of the set S(T, V ) as defined in Chapter 1 was used (Theorem 2.3). Proving this theorem required several auxiliary statements, as depicted in Figure 3.1. The crucial point was Proposition 3.1, which allowed us to bound |ζ( 12 + it)| by a certain sum involving the von Mangoldt function. In addition to being helpful in proving Theorem 2.3, this proposition yielded Corollary 3.2, an upper bound on |ζ( 21 + it)|. The actual proof of Theorem 2.3 relied on the proposition, where the sum was restricted to primes (Lemma 3.5). A mean value estimate (Lemma 3.6) applied to this sum over primes yielded, with some more work, the upper bounds from Theorem 2.3. Lastly, a slight modification of the proposition eventually was an important step in finding an upper bound on moments for the family of quadratic Dirichlet L-functions, as mentioned in Chapter 4. In fact, a version of Soundararajan’s theorem and bound on moments can also be applied to other families of L-functions, for example the family of quadratic twists of an elliptic curve (see [8]). Obviously, a future proof of the bound on Mk (T ) without using the Riemann Hypothesis, confirming the conjecture mentioned in the introduction (see Equation (1.3)), would be a vast improvement of the result covered in this paper. Still, Soundararajan’s bound might give valuable hints to the “right direction” of a possible proof of the unconditional result. 31 32 Summary Bibliography [1] H. Davenport and H. Montgomery, Multiplicative number theory, Springer Verlag, 2000. [2] D. Heath-Brown, Fractional moments of the Riemann zeta-function, Journal of the London Mathematical Society, 2 (1981), p. 65. [3] M. Jutila, On the value distribution of the zeta-function on the critical line, Bulletin of the London Mathematical Society, 15 (1983), p. 513. [4] J. Keating and N. Snaith, Random matrix theory and ζ(1/2+ it), Communications in Mathematical Physics, 214 (2000), pp. 57–89. [5] A. Laurincikas, Limit theorems for the Riemann zeta-function, Kluwer Academic Publishers, 1996. [6] K. Ramachandra, Some remarks on the mean value of the Riemann zeta-function and other Dirichlet series, I, Hardy-Ramanujan Journal, 1 (1978), pp. 1–15. [7] A. Selberg, Old and new conjectures and results about a class of Dirichlet series, Proceedings of the Amalfi Conference on Number Theory, Collected Papers (Vol. II), 1989. [8] K. Soundararajan, Moments of the Riemann zeta function, Annals of Mathematics, 170 (2009), pp. 981–993. [9] E. Titchmarsh and D. Heath-Brown, The theory of the Riemann zeta-function, Oxford University Press, USA, 1986. 33