Arxiv:0804.2233V4 [Math.NT]

arXiv:0804.2233v4 [math.NT] 3 Nov 2010 aytasomtoso siae aeqaiyrltdt h nor the to related quality have estimates or transformations many m nlgo ´laVngao asta if that Pólya-Vinogradov says of analog epciey gi with again respectively, hoe 1.1. Theorem ω fodrtre n lolreseetp nqaiisfrteecha these for inequalities sieve-type large also and three, order of u r iia xettesmover sum the except similar are Luo ornso litccre,casnmes t. e h ui chara of values cubic mean du in the interested attention yet are etc., of we article numbers, lot this class In a curves, neglected. seen elliptic have of ranks characters to quadratic The theory. (1) osdrdteaaoosma au u o h aeo ui Heck cubic of case the for but value mean analogous the considered mohdvrinof version smoothed a where nain ne ope conjugation. complex under invariant where eesn h reso umto ed otepolmo estimatin of problem the f to approximate leads an summation Applying of statement). orders precise the a reversing for 2.1 Lemma (see opiiiecaatr,and characters, primitive to eo) and below), un u httesmoe ui hrcescnb parameteriz be can characters cubic over form sum the the of that out turns u rtrsl nsuch on result first Our iihe hrceso ie re perntrlyi ayapplica many in naturally appear order given a of characters Dirichlet hsrsl sms iia i em fmto fpof otemi res main the to proof) of method of terms (in similar most is result This sa sd,i hudntb upiigthat surprising be not should it aside, an As coe 6 2018 26, October = ≤ A e > c τ 2 h ro em eas banalreseetp eutfrodrt order for result type large-sieve a obtain also We characters. term. error the eta auso h Dirichlet formu asymptotic the an of obtain values we central particular, In characters. Dirichlet Abstract. πi/ M ( sati usto h lmnsof elements the of subset thin a as χ 3 1 n 0 u rbe a e nltcdffiute hc ebifl kthhere. sketch briefly we which difficulties analytic new has problem Our . stecbcGussmand sum Gauss cubic the is ) = w χ b sacntn htcnb ie xlctyi em fa Euler an of terms in explicitly given be can that constant a is n m X steFuirtasomof transform Fourier the is ( ≤ m Let ENVLE IHCBCCHARACTERS CUBIC WITH VALUES MEAN A = ) eivsiaevrosma au rbesivligodrtreprim three order involving problems value mean various investigate We √ ( q, X 1 w m 3)=1 (0 : S N m n TPA AE N ATE .YOUNG P. MATTHEW AND BAIER STEPHAN is , AB ( X χ n 1. , χ (mod ) 3 L X ∞ ≤ 3 h ui eiu ybl where symbol, residue cubic the , = ≪ fntosi h following the is -functions Q nrdcinadMi results Main and Introduction χ = χ ) 0 0 ∗ q → N ) Q m eoe h rnia character. principal the denotes n L L ( ti ehp otntrlt iwtertoa integers rational the view to natural most perhaps is It . R m ( fntosascae oti aiy ihapwrsvn in saving power a with family, this to associated -functions 1 2 3 ) m χ , easot,cmatyspotdfnto.Then function. supported compactly smooth, a be , 1 2 + nta usoe lmnsof elements over runs instead ) ε w and m = AB w sntacb then cube a not is m eethe Here . Q q Z 1+2 1 [ M = ω > c ihnorm with ] = 2 ε Q . = h nlgu uniiscniee by considered quantities analogous The . cQ n()snetesto eta ausis values central of set the since (1) in 0 m X w ≤ b ∗ B 0 + (0) ntesmover sum the on √ 1 m L fntostitdb characters by twisted -functions O ≤ S N n ( X ( afrtefis oetof moment first the for la A Q = n re(n i)Dirichlet six) (and hree ) usoe lmnsof elements over runs 2 racters. ≤ 37 P of m hsi aua because natural is This . tr aebe relatively been have cters Q druhya characters as roughly ed oatatv questions attractive to e / 38+ Z uso h form the of sums g N p e ntoa qainand equation unctional τ [ ( ω L n ε ( N m ) ) χ with ] fntoson -functions ≤ , ( χ n o xml,the example, For . Q n ) rdc (see product ) etit h sum the restricts in nnumber in tions m n l f[] who [L], of ult N m n 3 ( ra least at or , itive m 3 ) , ≤ Q ,B A, Z (23) ( ω [ ω It ), ] 2 STEPHAN BAIER AND MATTHEW P. YOUNG

Other authors ([FaHL], [FrHL], [Di], [BFH], . . . ) have considered cubic and higher order twists using multiple Dirichlet series. However, the method using the metaplectic Eisenstein series currently requires the ground field to contain the l-th roots of unity (supposing one is twisting by order l Hecke characters). Diaconu and Tian [DT] have developed analytic properties of a multiple Dirichlet series that potentially has applications to the first moment considered in our Theorem 1.1. By taking r = 3, F = Q, and L = Q(ω) (in their Section 3) they obtain a double Dirichlet series roughly of the form m n 3 , Re(s), Re(w) > 1. msN(n)w m N n Z[ω] X∈ X∈ The meromorphic continuation, location of poles, and order of growth of this double Dirich- let series allows one to consider moments similar to that of Theorem 1.1. We thank an anonymous referee for pointing this out to us. However, it is not clear if our Theorem 1.1 can be obtained from [DT]. A consequence of Theorem 1.1 is Corollary 1.2. There exist infinitely many primitive Dirichlet characters χ of order 3 such that L(1/2, χ) = 0. More precisely, the number of such characters with conductor Q is 6/7 ε 6 ≤ Q − . ≫ Proof. Let N3(Q) be the number of primitive Dirichlet characters of order 3 with conductor Q such that L(1/2, χ) does not vanish. Then using Theorem 1.1, Hölder’s inequality and ≤the familiar eight moment bound (2) ∗ L(1/2, χ) 8 Q2+ε | | ≪ q Q χ (mod q) X≤ X for the family of all primitive Dirichlet characters with conductor Q (see Theorem 7.34 of [IK]), we obtain ≤ Q ∗ L(1/2, χ) ∗ L(1/2, χ) 8 1/8N (Q)7/8 Q1/4+εN (Q)7/8 ≪ | |≪ | | 3 ≪ 3 q Q χ (mod q) q Q χ (mod q) ≤ 3 ≤ X χX=χ0 X X 6/7 ε which gives Q − N (Q). ≪ 3 The work of [DT] gives a non-quantitative version of Corollary 1.2. Studies of the moments and nonvanishing of cubic twists of elliptic curves using random matrix theory have been carried out in [DFK]. Nonvanishing of cubic twists of elliptic curves using algebraic methods has been undertaken by [FKK]. As for the second moment, we show Theorem 1.3. Let Q 1. Then we have ≥ (3) ∗ L(1/2+ it, χ) 2 Q6/5+ε(1 + t )6/5+ε. | | ≪ | | q Q χ (mod q) ≤ 3 X χX=χ0 For a rational integer m, let

m s (4) L(s, ψm)= N(n)− n 3 n Z[ω] n 1X∈ (mod 3) ≡ MEAN VALUES WITH CUBIC CHARACTERS 3

denote the Hecke L-function. Then (5) ∗ L(1/2+ it, ψ ) 2 M 3/2+ε(1 + t )4/3, | m | ≪ | | m M X≤ where the star indicates the sum is over squarefree integers. Our proof of Theorem 1.1 uses (5) as a key ingredient; actually, we require a minor variant given by (39) below. We shall establish Theorem 1.3 by using the following large sieve-type result with cubic Dirichlet characters.

Theorem 1.4. Let (am)m N be an arbitrary sequence of complex numbers. Then ∈ 2 (6) ∗ ∗ a χ(m) ∆(Q, M) ∗ a 2 , m ≪ | m| Q

where the star at the sum over m indicates that it is taken over squarefree integers and (7) ∆(Q, M)=(QM)ε min Q5/3 +M, Q4/3 +Q1/2M, Q11/9 +Q2/3M, Q+Q1/3M 5/3 +M 12/5 . { } For comparison, note that the ordinary large sieve gives the following weaker results

1 8 2+ε 1 2 5 +ε ∗ L( , χ) Q , ∗ L( , χ) Q 4 , | 2 | ≪ | 2 | ≪ q Q χ (mod q) q Q χ (mod q) ≤ 3 ≤ 3 X χX=χ0 X χX=χ0 1 9 +ε ∗ L( , χ) Q 8 . | 2 |≪ q Q χ (mod q) ≤ 3 X χX=χ0 The 8th moment is deduced simply by embedding the family of cubic characters into the family of all Dirichlet characters of conductor Q, and using the known bound (2). The estimates for the ﬁrst and second moments follow≤ by Cauchy’s inequality. We point to [E], Section 7, for some early large sieve-type results on general r-th order characters. Related results to Theorem 1.4 are Heath-Brown’s quadratic and cubic large sieves [Hea1] [Hea2]. The quadratic large sieve states ♭ (8) ∗ a χ (m) 2 (Q + M)(QM)ε ∗ a 2, m d ≪ | m| d Q m M m M |X|≤ X≤ X≤ where the sum over d runs over fundamental discriminants, and χd is the associated primitive quadratic character. The cubic large sieve states

∗ ∗ m 2 2/3 ε ∗ 2 (9) am (M + N +(MN) )(MN) am , n 3 ≪ | | n Z[ω] m Z[ω] m M N(X∈n) N N(Xm∈) M X≤ ≤ ≤ where the stars indicate that m, n run over squarefree elements of Z[ω] that are congruent to 1 (mod 3). Both of these results are proved with a recursive use of Poisson summation. Our method of proof of Theorem 1.4 uses (9) (after some transformations), and avoids recursion. One of the new diﬃculties with treating cubic Dirichlet characters is the asymmetry between χ and m. It takes some calculation to see that a direct application of (9), choosing am to have support on rational integers, is not better than (7). Note that m M means N(m) M 2 so that (9) implies ∆(Q, M) (QM)ε(Q + M 2 +(QM 2)2/3). For≤ example, M = √Q≤gives here ∆(Q, √Q) Q4/3+ε while≪ the fourth bound of (7) gives ∆(Q, √Q) Q6/5+ε, which in fact is the key to≪ proving (3). ≪ 4 STEPHAN BAIER AND MATTHEW P. YOUNG

It is of great interest to extend these results to higher-order characters and to different number fields. In general we wish to understand these families of twists in the Katz-Sarnak sense [KS1] [KS2]. One obvious analytic issue is the degree of the field extension Q(e2πi/l)/Q, as discussed following equation (1) above. It is plausible that our methods could generalize to l = 4, 6 since for these cases this degree is also 2. However, for the application of the central values of L-functions, we also require estimates for the sum of l-th order Gauss sums, which become somewhat worse as l increases (e.g. see Proposition 1 of [P2]). We can generalize some of our results to sextic characters. 3 Theorem 1.5. Theorems 1.3 and 1.4 remain valid when the condition χ = χ0 is replaced 6 by the weaker condition χ = χ0. The proof is nearly identical to those of Theorems 1.3 and 1.4 so we omit the details. A motivating example for this generalization to sextic twists is to understand the behavior of the family of elliptic curves y2 = x3 + b, where b Z. These curves have complex multiplication by Q(ω), and have L-functions ∈ that can be expressed using the sextic residue symbol (4b/n)6, where n runs over elements of Z[ω]. The study of this family of L-functions clearly leads to double sums of the form addressed in Theorem 1.5.

Acknowledgements. The second-named author would like to thank B. Brubaker and S.J. Patterson for useful comments. Parts of this work were done when the first-named author visited Texas A&M University in November 2007 and September 2009. He wishes to thank this institution for the invitation, its warm hospitality during his pleasant stays, and for financial support. Both authors would like to thank A. Diaconu for helping us to understand his work [DT] with Y. Tian, and to an anonymous referee for pointing out that [DT] may be relevant to our investigations. 2. Preliminaries In this section we provide various tools used throughout the paper. 2.1. Properties of the cubic characters. The cubic characters are related to the arithmetic of the quadratic number field Q(ω), ω = e2πi/3, with ring of integers Z[ω] and discrim- inant 3. This field has class number one and has six units, 1,ω,ω2 , and one ramified prime− 1 ω dividing 3. Each principal ideal 0 = (n) Z[ω]±{ with (n, 3)} = 1 has a unique generator−n 1 (mod 3); sometimes we may implicitly6 ⊂ choose such a generator. ≡ Lemma 2.1. The primitive cubic Dirichlet characters of conductor q coprime to 3 are of m the form χn : m ( n )3 for some n Z[ω], n 1 (mod 3), n squarefree and not divisible by any rational primes,→ with norm N(∈n)= q. ≡ This analysis can also be found in [DFK] e.g. but we shall present this here for completeness. We refer to Chapter 9 of [IR] for basic properties of cubic residues. Proof. To classify the characters we use an approach similar to that of [Da], Chapter 5. By multiplicativity, it suffices to consider the case that q = pa is a prime power. It is not hard to show that there is a primitive character of conductor p if and only if p 1 (mod 3), in which case there are exactly two such characters (mod p), each being the≡ square of the other. If a 2 and p = 3 then there is no primitive character of order 3, as any character of order 3 must≥ be induced6 from one (mod p). MEAN VALUES WITH CUBIC CHARACTERS 5

These cubic characters are intimately connected with the cubic residue symbol in the ring Z[ω]. If p 1 (mod 3) then p = ππ with N(π) = p, and then there is an associated cubic − ≡ l l N(π) 1 character l ( ) . This cubic character is defined by the conditions l 3 (mod π), → π 3 π 3 ≡ with l 1,ω,ω2 . It follows directly from the definition that for l Z, ( l ) = ( l ) . π 3 π 3 π 3 Thus for prime∈ { conductor} p there is a one-to-one correspondence between primitive∈ Dirichlet characters of order 3 and conductor p, and cubic residue symbols χπ with N(π) = p. By multiplicativity we extend this one-to-one correspondence to squarefree q and elements n of Z[ω] such that N(n) = q. It is easy to see that N(n) is squarefree (as an element of Z) if and only if n is squarefree (as an element of Z[ω]) and n has no rational prime divisor. Recall that the cubic reciprocity law states that for m, n Z[ω], m, n 1 (mod 3), ∈ ≡ ± m n = . n 3 m 3 Some sources state this for m n 1 (mod 3), but the fact that ( 1)3 = 1 easily allows for this slight generalization. The≡ supplement≡ − states that if π = 1+3−a+3bω,− where a, b Z, then ∈ 1 ω − = ωa. π 3 Note 3 = ω2(1 ω)2, and that N(π)=(1+3a)2 (1+3a)3b +9b2 1+3a +3b (mod 9). Therefore,− a simple− calculation shows − ≡ ω = ω2a+2b, π 3 and hence 3 = ωb. π 3 It follows easily that for any n 1 (mod 3), n Z[ω], written in the form n =1+3c +3dω, d ≡ ∈ then (3/n)3 = ω . In particular, for n 1 (mod 3), we have (3/n)3 = 1 if and only if n 1, 4, 7 (mod 9), and in general (3/n≡) only depends on n (mod 9). The functions ≡ 3 n ((1 ω)/n)3 and n (ω/n)3 are ray class characters (mod 9). →Suppose− that m Z[ω→], m not a cube nor a unit. Then the function ψ : (n) m ∈ m → n 3 defined on ideals (n) Z[ω] coprime to 3, where n 1 (mod 3), gives a class group character of modulus 9m. Hence⊂ (see Theorem 12.5 of [I1])≡ the Hecke L-function

s m s L(s, ψm)= ψm((n))N(n)− = N(n)− n 3 (n) n Z[ω] X n 1X∈ (mod 3) ≡ is associated to a weight one cusp form of level 3N(9m) and nebentypus character χ(a) = ( 3/a)ψ ((a)), where ( 3/a) is the Kronecker symbol. − m − 2.2. On the Gauss sums. It turns out that the Gauss sum associated to the Dirichlet character χn (on Z) deﬁned in Lemma 2.1 is the same one as the corresponding Hecke character (on Z[ω]). We now prove this important fact. Recall the deﬁnition of the standard Gauss sum for n Z[ω], n 1 (mod 3) (as in [H-BP] for instance), ∈ ≡ d g(n)= eˇ(d/n), wheree ˇ(z) = exp(2πi(z + z)). n 3 d (modX n) 6 STEPHAN BAIER AND MATTHEW P. YOUNG

Our notation differs from [H-BP] as we reserve e(z) for the more standard exp(2πiz). Recall that n has no rational prime divisor, so (n, n¯) = 1. By definition, x 2πix N(n) τ(χn)= e . n 3 1 x N(n) ≤ X≤ Now write x yn¯ +yn ¯ (mod nn), where y varies over a set of representatives in Z[ω] (mod n), and≡ heren ¯ is the complex conjugate of n. It is easy to see that as y varies (mod n), x varies (mod N(n)), using thatx ¯ = x and the Chinese Remainder Theorem. We then see yn¯ 2πi( y + y ) τ(χn)= e n n . n 3 y (modX n) It is a consequence of cubic reciprocity that ( n¯ ) = 1 for n 1 (mod 3), so we have n 3 ≡ (10) τ(χn)= g(n). Now we collect some basic facts on the Gauss sums. It is well-known that (11) g(n)3 = µ(n)N(n)n, whence one derives the pleasant fact that g(n) vanishes unless n is squarefree. Generalize the definition of the Gauss sum by setting x rx g(r, n)= eˇ . n 3 n x (modX n) See [H-BP], pp.123-124 for the following formulas. First, s (12) g(rs,n)= g(r, n), if (s, n)=1. n 3 Furthermore, if (n1, n2) = 1 then n (13) g(r, n n )= 1 g(r, n )g(r, n )= g(n r, n )g(r, n ). 1 2 n 1 2 2 1 2 2 3 We also compute, for π prime in Z[ω], k 1, ≥ N(π2), k =3, (14) g(π2, πk)= − (0, otherwise. In addition, we shall require the fact that (15) g(r, n)=0, if π2 n, π ∤ r, | which follows immediately from (11) and (12).

Lemma 2.2. Suppose n1, n2, δ Z[ω] are squarefree, 1 (mod 3), with norms that are pairwise relatively prime. Then ∈ ≡ n δ δ (16) τ(χ χ )= 2 τ(χ )τ(χ )τ(χ ). n1 n2δ n n n1 n2 δ 1 3 2 3

Proof. The conditions ensure that χn1 χn2δ is a primitive character. It follows from the deﬁnition of the cubic residue symbol that m = m , so for m Z, χ (m) = χ (m). n 3 n 3 ∈ n n Thus τ(χ χ ) = τ(χ ). Repeatedly using (10), (13), and cubic reciprocity, we get n1 n2δ n1n2δ (16). MEAN VALUES WITH CUBIC CHARACTERS 7

2.3. The approximate functional equation. Using an approximate functional equation (see Theorem 5.3 of [IK]), we have Proposition 2.3. Let χ be an odd primitive Dirichlet character χ of conductor q, and make the following deﬁnitions: Let

3 +α+s Γ 2 1 G(s) s s/2 2 V (x)= g (s)x− ds, where g (s)= π− . α α α 3 +α 2πi (1) s 2 Z Γ 2

1 1/2 Furthermore, let ǫ(χ)= i− q− τ(χ) be essentially the (normalized) Gauss sum and set

3 α 2 − q α Γ 2 X = − . α 3 +α π Γ 2 2 1 Finally let A and B be positive real numbers such that AB = q. Then for any Re(α) < 2 we have | |

∞ ∞ 1 χ(m) m χ(m) m (17) L( + α, χ)= Vα + ǫ(χ)Xα V α . 2 1 +α 1 α 2 A 2 − B m=1 m m=1 m − X X For α = 0 we set V0 = V .

2.4. Poisson summation. We shall require two versions of the Poisson summation formula. Suppose that w is a smooth, compactly-supported function on the positive reals. Let χ be a primitive Dirichlet character of conductor q. Then m M hM (18) w χ(m)= τ(χ) χ(h)w . M q q m Z h Z X∈ X∈ This is well-known. For the latter version, we directly quoteb Lemma 10 of [Hea2]. Let m m χ(m) = n n where n1 and n2 are elements of Z[ω] that are coprime to each other, 1 3 2 3 and to 3, and are squarefree. Then χ is a primitive character on Z[ω] of modulus n1n2. Lemma 2.4. We have N(m) χ(√ 3)g(n )g(n )M N(k) w χ(m)= − 1 2 χ(k)ˇw M , M N(n1n2) N(n1n2) m Z[ω] k Z[ω] s ! X∈ X∈ where ∞ ∞ wˇ(t)= w(N(x + yω))ˇe(t(x + yω)/√ 3)dxdy. − Z−∞ Z−∞ 3. Proof of Theorem 1.1 Here we give the proof of Theorem 1.1 though a variant on the bound (5) used in a critical way is proved in Section 4.3. We ﬁrst note that by Lemma 2.1, q N(n) := ∗ L( 1 , χ)w = ′ L( 1 , χ )w M 2 Q 2 n Q (q,3)=1 χ (mod q) n 1 (mod 3) 3 ≡ X χX=χ0 X 8 STEPHAN BAIER AND MATTHEW P. YOUNG

where the prime indicates the sum runs over squarefree elements n of Z[ω] that have no rational prime divisor. Applying the approximate functional equation, Proposition 2.3, with A B = N(n) gives = + , where n M M1 M2 ∞ ′ χn(m) m N(n) 1 = V w M √m An Q n 1 (mod 3) m=1 ≡ X X and ∞ χ (m) m N(n) = ′ ǫ(χ ) n V w . M2 n √m B Q n 1 (mod 3) m=1 ≡ X X Deﬁne A by AB = Q so that A = A N(n) A for all n under consideration, in view of the n Q ≍ support of w. We shall treat and with diﬀerent methods. The results are summarized with M1 M2 Lemma 3.1. We have

1/2+ε 3/4 1/6+ε (19) = cQw(1) + O(Q A + QA− ), M1 and e (20) Q5/6B1/6 + Q2/3B5/6. M2 ≪ Choosing B = Q7/19, whence A = Q12/19 gives Theorem 1.1. The constant c is given more explicitly in Section 3.2 below. Our approach for 1 employs the summation over n to transform the expression into one involving Hecke LM-functions. Then we bound this new expression with (39) which is a consequence of Theorem 1.4. In a previous version of this paper (available on the arxiv), we set up a complicated recursive technique that was later used (with other ingredients) by the second author in a simpler setting [Y]. The cancellation in 2 comes from the sum of cubic Gauss sums, which follows from the work of Patterson showingM that these cubic Gauss sums appear as Fourier coeﬃcients of metaplectic Eisenstein series [P1]. See Lemma 3.2 below for the estimate on the sum of cubic Gauss sums.

3.1. Evaluating 1. First we work on 1. We shall detect the condition that n 1 (mod 3) has no rationalM prime divisor usingM the formula ≡

1, n has no rational prime divisor, (21) µ (d)= Z 0, otherwise. d n,d Z ( d 1|X (mod∈ 3) ≡

Here we deﬁne µZ(d)= µ( d ), the usual M¨obius function. The choice of d up to unit, namely d 1 (mod 3) is natural| for| the arithmetic of the ring Z[ω]. We apply this formula and change≡ variables n dn to the sum over n. Since d is squarefree as an element of Z[ω], the condition that dn is→ squarefree then simply means that n is squarefree and (d, n) = 1. Thus

m ∞ d 3 ∗ m m Q N(nd) 1 = µZ(d) V w . M √m n 3 A N(nd) Q d Z m=1 n 1 (mod 3) d 1X (mod∈ 3) X ≡ X ≡ (n,d)=1 MEAN VALUES WITH CUBIC CHARACTERS 9

Now we use Möbius inversion again (writing µω(l) for the Möbius function on Z[ω]) to detect the condition that n is squarefree, getting m ∞ 2 = µ (d) µ (l) dl 3 (d,l,m), M1 Z ω √m M1 d Z l 1 (mod 3) m=1 d 1X (mod∈ 3) ≡ X X ≡ where m m Q N(ndl2) 1(d,l,m)= V w . M n 3 A N(ndl2) Q n 1 (mod 3) ≡(n,dX)=1 Next we use the Mellin transform of the weight function to express the sum over n as a contour integral involving the Hecke L-function. By Mellin inversion, m Q N(ndl2) 1 Q s V w = f(s)ds, A N(ndl2) Q 2πi N(ndl2) Z(2) where e ∞ m s 1 f(s)= V x w(x)x − dx. 0 A Z Integration by parts shows f(es) is a function satisfying the bound for all Re(s) 1 ≥ 4 100 100 f(s) (1 + s )− (1 + m/A)− . e ≪ | | With this notation, and with the definition (4), then e 1 Q s (d,l,m)= L(s, ψ )f(s)ds. M1 2πi N(dl2) m Z(2) We estimate 1 by moving the contour to the half line. Whene m is a cube the Hecke L-function hasM a pole at s = 1. We set to be the contribution to of these residues, M0 M1 and 1′ to be the remainder. We shall defer the analysis of 0 to Section 3.2. ByM bounding everything with absolute values, we see that M

1 √Q 100 ∞ 1 100 ′ (1 + m/A)− L( + it, ψm) (1 + t )− dt. |M1|≪ N(dl2) √m | 2 | | | d √Q N(l) √Q m Z−∞ ≪X X≪ X Since d Z, then N(d)=pd2 so that the sums over d and l are easily computed. Finally we use the∈ estimate (39), which is a close relative to (5), to bound the sum over m. Putting everything together, we obtain 1/2+ε 3/4 (22) ′ Q A , |M1|≪ 1/6+ε In Section 3.2 we show = cQw(1) + O(QA− ) which combined with (22) gives (19). M0 3.2. Computing 0. Recall that M e m ∞ 2 Q = µ (d) µ (l) dl 3 f(1)Res L(s, ψ ), M0 Z ω √m N(dl2) s=1 m d Z l 1 (mod 3) m=1 d 1X (mod∈ 3) ≡ X X ≡ e where using the Mellin convolution formula shows s ∞ m 1 A G(s) f(1) = V x w(x)dx = w(1 s) g(s)ds. 0 A 2πi (1) m − s Z Z e e 10 STEPHAN BAIER AND MATTHEW P. YOUNG

From the discussion in Section 2.1, it is not diﬃcult to see that ψm is the principal character only if m is a cube, in which case

s L(s, ψ )= ζ (s) (1 N(π)− ), m Q(ω) − π 3m Y| and ζ (s) is the Dedekind zeta function for the ﬁeld K. Let c = 2π be the residue of K ω 6√3 ζQ(ω)(s) at s = 1, evaluated using the Kronecker limit formula. Then

∞ f(1) 1 µZ(d) µω(l) = c Q (1 N(π)− ) . M0 ω m3/2 − d2 N(l2) m=1 π 3m d Z,(d,m)=1 (l,m)=1 X e Y| d∈ 1X (mod 3) l 1X (mod 3) ≡ ≡ Computing the sums over d and l explicitly, we obtain

∞ f(1) 1 2 2 = c Q (1 N(π)− ) (1 p− ) (1 N(π)− ). M0 ω m3/2 − − − m=1 π 3m p∤3m π∤3m X e Y| Y Y The two products over π combine rather nicely to give

∞ 1 1 f(1) 1 1 2 1 = c ζ− (2)ζ− (2)Q (1 + N(π)− )− (1 p− )− . M0 ω Q(ω) m3/2 − m=1 π 3m p 3m X e Y| Y| Let ∞ u 1 1 2 1 Z(u)= m− (1 + N(π)− )− (1 p− )− , − m=1 π 3m p 3m X Y| Y| which is holomorphic and bounded for Re(u) 1+ δ > 1. Then ≥ 1 1 1 s 3 G(s) = c ζ− (2)ζ− (2)Q A Z( +3s)w(1 s) g(s)ds. M0 ω Q(ω) 2πi 2 − s Z(1) We move the contour of integration to 1/6+ ε, crossinge a pole at s = 0 only. The new 1/6+ε − contour contributes O(A− Q), while the pole at s = 0 gives 1 1 (23) cQw(1), where c = cωζQ−(ω)(2)ζ− (2)Z(3/2). Note that Z(u) converges absolutely at u =3/2 so it is easy to express Z(3/2) explicitly as an Euler product, if desired.e

1 1 3.3. Estimating . Using the calculation ǫ(χ )= i− g(n)N(n)− 2 , we have M2 n ∞ 1 1 m ′ χn(m)g(n) N(n) 2 = i− V w . M 1 B N(n) Q m=1 m 2 n 1 (mod 3) X ≡ X In this section we show p

Lemma 3.2. For any m Z[ω], write m = m0m1 where m0 is a unit times a power of 1 ω and m 1 (mod 3). Then∈ we have − 1 ≡ ′ χn(m)g(n) N(n) 2/3+ε 1/6 5/6 1/6+ε H′(n, Q) := w Q N(m) + Q N(m1)− . N(n) Q ≪ n Z[ω] n 1X∈ (mod 3) ≡ p MEAN VALUES WITH CUBIC CHARACTERS 11

By summing trivially over m one easily deduces (20). Recall that the prime on the sum over n indicates that the sum is restricted to squarefree numbers having no rational prime divisor. This feature causes some difficulties. Our first move in the proof of Lemma 3.2 is to use Möbius inversion, i.e., (21), to remove the condition that n has no rational prime divisor. We simplify the resulting expression using the the identity g(dn) = g(d)g(n)χn(d) following from (13), the fact that g(n)=0 1 unless n is squarefree, and using the notation g(c)= g(c)N(c)− 2 . This gives

2 H′(n, Q)= µZ(d)˜g(d)H(dm, Q/d ), d Z e d 1X (mod∈ 3) ≡ where χn(dm)g(n) N(n) H(dm,X)= 1 w . N(n) 2 X n Z[ω] n 1X∈ (mod 3) ≡ We estimate H with the following

Lemma 3.3. For any l Z[ω], write l = l0l1 where l0 is a unit times a power of 1 w, and l 1 (mod 3). Then we∈ have − 1 ≡ 1/2+ε 1/4 5/6 1/6+ε H(l,X) X N(l ) + X N(l )− . ≪ 1 1 Before proving Lemma 3.3, we show how Lemma 3.2 follows from it. We treat d Y and d > Y separately, where Y is a parameter to be chosen. For d Y we use| Lemma| ≤ 3.3, while| | for d > Y we use the trivial bound H(l,X) X. Thus,| | ≤ writing m = m m | | ≪ 0 1 where m0 is a power of 3 and m1 is coprime to 3, we have 1/2+ε 5/6 Q 1/4 Q 1/6+ε Q H′(n, Q) N(dm) + N(m )− + , ≪ d2 d2 1 d2 d Y d Y d >Y |X|≤ |X|≤ |X| which simplifies as 1/2+ε 1/4 1 5/6 1/6+ε H′(n, Q) Q √Y N(m) + QY − + Q N(m )− . ≪ 1 1/3 1/6 Optimally choosing Y = Q N(m)− gives Lemma 3.2. Proof of Lemma 3.3. The difficulty in estimating H(l,X) is apparently a technicality: the sum is not quite in a form that allows us to quote from the literature, in particular Section 4 of [H-BP]. Our goal is to manipulate H(l,X) until it meets these conditions. Before elaborating on this discussion we first do some minor simplifications that ease the comparison to the literature. In this section we use the convention that all sums over elements of Z[ω] are restricted to elements 1 (mod 3). Writing l = l l as above, and using cubic reciprocity, we see that ≡ 0 1 l n l χ (l) := = 0 . n n l n 3 1 3 3 l0 From the discussion in Section 2.1, the function λ(n) = n 3 is a ray class character (mod 9). Thus n λ(n) l g(n) 1 3 N(n) H(l,X)= 1 w . N(n) 2 X n Z[ω] X∈ 12 STEPHAN BAIER AND MATTHEW P. YOUNG

Note that the identity (12) implies n g(n)= g(l , n) for (n, l ) = 1. Letting l1 1 1 λ(n)g(r, n) h(r, s)= , N(n)s (n,rX)=1 and introducing the Mellin transform of w, we get 1 (24) H(l,X)= w(s)Xsh( 1 + s,l )ds. 2πi 2 1 Z(2) Clearly the series defining h(r, s) converges absolutelye and uniformly on any region Re(s) 3 3 ≥ 2 + δ > 2 . We need to know the analytic behavior of h(r, s), i.e., meromorphic continuation, location of poles, and order of growth. In Section 4 of [H-BP] these properties are explicitly given but for slightly different functions, such as g(r, n) g(r, n) g(r, n) ψ(r, s)= , ψ (r, s)= , ψ (r, s)= . N(n)s α N(n)s α N(n)s n n 0 (mod α) (n,α)=1 X ≡ X X Precisely, with the following we summarize results from Lemmae 4 of [H-BP], and Theorems 9.1 and 8.1 of [P2]. Lemma 3.4. The function ψ(r, s) has meromorphic continutation to the complex plane. It is holomorphic in the region Re(s) > 1 except possibly for a pole at s = 4/3. Furthermore, letting σ = 3 + ε, and σ σ σ 1 , s 4 > 1 , we have 1 2 1 ≥ ≥ 1 − 2 | − 3 | 12 1 (σ1 σ) 2 σ1 σ ψ(r, s) N(r) 2 − (1 + t ) − . ≪ 2 If r = r1r2 is cubefree, then the residue satisfies 1/6+ε res ψ(r, s) N(r )− . s=4/3 ≪ 1 Note h(r, s) differs from ψ(r, s) only in the additional presence of the ray class character λ, and the coprimality condition (n, r) = 1. The presence of λ is unimportant, but removing the condition (n, r) = 1 unfortunately seems to require some elaborate gyrations. Lemma 3.5. Lemma 3.4 holds with ψ(r, s) replaced by h(r, s). Before proving Lemma 3.5 we show how it implies Lemma 3.3. We move the line of 1 integration in (24) to Re(s)= 2 + ε, crossing a pole at s =5/6, which contributes 5/6 1/6+ε X N(l )− . ≪ 1 The main contribution comes from the new line of integration, which gives X1/2+εN(l )1/4. ≪ 1 This completes the proof of Lemma 3.3. Proof of Lemma 3.5. We shall use inclusion-exclusion type arguments to reduce the estimation of h(r, s) to that of ψ(r, s). To this end, we collect some of these results with the following Lemma 3.6. Suppose f, α are squarefree and (r, f)=1, and set λ(n)g(r, n) λ(n)g(r, n) h(r,f,s)= , h (r, s)= . N(n)s α N(n)s (n,rfX)=1 (n,αX)=1 MEAN VALUES WITH CUBIC CHARACTERS 13

2 3 Furthermore suppose r = r1r2r3 where r1r2 is squarefree, and let r3∗ be the product of primes dividing r3. Then µ (a)λ(a)g(r, a) (25) h(r,f,s)= ω h(ar, s), N(a)s a f X| 2 3 2 (26) h(r1r2r3,s)= h(r1r2,r3∗,s), 2 3 2 3s 1 2 (27) h(r r ,s)= (1 λ(π) N(π) − )− h (r r ,s), 1 2 − r1 1 2 π r2 Y| (28) 2 3 2 3s 1 1 2s 2 2 2 h (r r ,s)= (1 λ(π) N(π) − )− µ (a)N(a) − λ(a) g(r r /a, a)h (r r /a,s). r1 1 2 − ω 1 2 1 1 2 π r1 a r1 Y| X| Before embarking on the technical details of this proof, we show how it proves Lemma 3.5. The function h1(r, s) is identical to ψ(r, s) except it is twisted by λ(n), the ray class character of modulus 9. Then h1 satisﬁes the properties of Lemma 3.4, the necessary generalizations having been carried out in [P2] for example. By working backwards and using (25-28), we see that h(r, s) has meromorphic continuation and potential pole at s =4/3 only, and

1 ε (σ1 σ) 2 σ1 σ h(r,f,s) N(f) N(r) 2 − (1 + t ) − . ≪ The analogous bound on h(r, s) follows. The estimate on the residue follows by a similar method. Proof of Lemma 3.6. Using M¨obius to remove the condition (n, f) = 1 gives µ (a)λ(a) λ(n)g(r, an) h(r,f,s)= ω . N(a)s N(n)s a f (n,r)=1 X| X Notice that if π a then π ∤ r so if in addition π n then by (15), g(r, an) = 0. Thus we may assume (n, a) =| 1, in which case g(r, an)= g(ar,| n)g(r, a) by (13), and hence (25) holds. 2 3 2 From (12) it follows that g(r1r2r3, n)= g(r1r2, n) provided (n, r3) = 1, whence (26) holds. Now we prove (27). For this we introduce some new notation as follows. Let ab2 Z[ω] and let π be prime such that (ab, π) = 1. Then ∈ λ(n)g(ab2π2, n) λ(n)g(ab2π2, n) λ(n)g(ab2π2, n) h (aπ2, b2,s) := = . 2 N(n)s N(n)s − N(n)s (n,aπ)=1 (n,a)=1 (n,a)=1,π n X X X | j 2 2 j Writing in the latter sum n = π n′, where (n′, π) = 1, then we have g(ab π , π n′) = j+2 2 2 2 j 2 2 j 2 j 2 j g(π ab , n′)g(ab π , π ) by (13). Using (12) we get g(ab π , π )= (ab /π )3g(π , π ), which 2 2 3 2 2 2 is nonzero if and only if j = 3, from (14). Thus we get g(ab π , π n′)= N(π )g(ab π , n′). In summary, we have shown − 2 2 2 2 λ(n)g(ab π , n) 3 2 3s 2 2 h (s, aπ , b )= + λ(π) N(π) − h (s, aπ , b ), 2 N(n)s 2 (n,aX)=1 which when rearranged states 2 2 2 2 λ(n)g(ab π , n) 3 2 3s 1 λ(n)g(ab π , n) = (1 λ(π )N(π) − )− . N(n)s − N(n)s (n,aπX)=1 (n,aX)=1 14 STEPHAN BAIER AND MATTHEW P. YOUNG

An induction argument on the number of prime divisors of b gives

2 2 3 2 3s 1 λ(n)g(r1r2, n) h(r r ,s)= (1 λ(π) N(π) − )− , 1 2 − N(n)s π r2 (n,r1)=1 Y| X which is the same as (27). Finally, the relation (28) is a slight generalization of Lemma 3(i) of [H-BP], the only diﬀerence being that the sums in (28) are twisted by λ(n). Since λ(n) is completely multiplicative, an inspection of the argument of [H-BP] easily shows that the proof generalizes to give (28).

4. The cubic large sieve In this section we establish our cubic large sieve, Theorem 1.4. It is easy to reduce the expression in question, namely the left hand side of (6), to a sum of similar expressions with the additional summation conditions (q, 3)=1 and (m, 3) = 1 included. Thus it suﬃces to estimate 2 2 ∗ ∗ ′ ∗ amχ(m) = amχn(m) =: T (Q, M), Q

2 2 ′ ∗ 2 2 (29) B1(Q, M) := sup am − amχm(n) , where am = am , (a ) || || || || | | m Q

Let W : R R be a ﬁxed smooth, nonnegative, compactly-supported function such that W (x) 1 for→ 1 x 2. It follows that B (Q, N) is bounded by ≥ ≤ ≤ 3 2 2 N(n) ∗ B3(Q, N) sup am − W amχm(n) . ≤ (a ) || || Q m n m X X Expanding the square and rearranging the summation, the right-hand side takes the form

2 ∗ N(n) sup a − a a W χ (n)χ (n). || m|| m1 m2 Q m1 m2 (am) m ,m n X1 2 X As in [Hea2], it will turn out that we may restrict attention to the case in which m1 and m2 are coprime. We deﬁne another norm B4 corresponding to the above sum with the restriction (m1, m2) = 1 included by

2 ∗ N(n) B (Q, M) := sup a − a a W χ (n)χ (n). 4 || m|| m1 m2 Q m1 m2 (am) n (m1X,m2)=1 X We further deﬁne a norm C1(M, Q) dual to B1(Q, M) by 2 2 ∗ ′ C1(M, Q) := sup bn − bnχn(m) , (b ) || || n MMX/Q 16 STEPHAN BAIER AND MATTHEW P. YOUNG

where the sum over K in (36) runs over powers of 2. We postpone the proof of Lemma 4.1 to the following sections and now deduce Theorem 1.4. Recall that we need to prove that (37) B (Q, M) (QM)ε min Q5/3 + M, Q4/3 + Q1/2M, Q11/9 + Q2/3M, Q + Q1/3M 5/3 + M 12/5 . 1 ≪ { } The first estimate in the minimum follows from (30) and the trivial bound B1(Q, M) = C (M, Q) C (M, Q). The second and the third estimate are obtained by combining (30) 1 ≤ 2 and (31), with v =2, 3, and then using B1(Q, M) C2(M, Q). All that remains is to show the last inequality≤ in (37). This bound is most relevant for the second moment of cubic Dirichlet L-functions, i.e., (3). For this, we use the relations between the various norms. Specifically, we shall start with the already-established bound (38) B (Q, M) (QM)ε Q11/9 + Q2/3M 1 ≪ (third term in the minimum in (37)) as an initial estimate, deduce bounds for B3 and B4 from it and then work backwards, obtaining new bounds for B3 and finally B1. In details, we begin by combining (33) with (38) to get ε 1/2 1/2 11/9 2/3 B (Q, M) (QM) Q X− (X + X M). 2 ≪ The worst case is X = Q which shows B2(Q, M) also satisfies (38). Repeating the argument, we have B (Q, M) (QM)ε(Q11/9 + Q2/3M). 3 ≪ Combining this with (36), we obtain ε 2 11/9 2/3 4 1 B (Q, M) Q +(QM) QM − max K + K M : K M Q− 4 ≪ ≤ ε 6 1 2 ε 11/9 2/3 +(QM) M Q− K− − (K + K M) K M 4/Q ≥X ε 2/9 26/9 1/3 5/3 Q +(QM) (Q− M + Q M ). ≪ From this and (35), we deduce that 2/9 26/9 1/3 5/3 Q Q − M Q M B (Q, M) +(QM)ε + 3 ≪ ∆ ∆ ∆ ∆ ∆ 1 1 2 1 2 ! 2 for some positive integers ∆1, ∆2 with ∆2 ∆1. The worst case is ∆2 = ∆1 = 1. Using this together with the trivial bound B (Q, M) ≥ B (Q, M) gives 1 ≤ 3 ε 2/9 26/9 1/3 5/3 B (Q, M) Q +(QM) Q− M + Q M . 1 ≪ This bound can in general be improved by taking Q larger, so we use the increasing property (32) to replace Q by Q1+ε + M 11/5 which gives the desired bound B (Q, M) (QM)ε Q + Q1/3M 5/3 + M 12/5 . 1 ≪ ✷ This completes the proof of Theorem 1.4.

We remark that a further cycle in the above process does not lead to an improvement of our result. For convenience, we enclose a table displaying the estimates for B1(Q, M) that we get for various ranges. This table should be read as follows. If the fractions α and β are the (n 1)-th and n-th entries, respectively, in the ﬁrst row, and the term T is the n-th entry in the− second row, then the estimate B (Q, M) (QM)εT holds in the range M α < Q M β. 1 ≪ ≤ MEAN VALUES WITH CUBIC CHARACTERS 17

Range 3 6 6 3 9 108 11 5 5 7 5 2 5 55 5 2 ∞ Bound M Q5/3 Q1/2M Q4/3 Q2/3M Q11/9 M 12/5 Q1/3M 5/3 Q 4.3. Proof of Theorem 1.3. Before we turn to the proof of Lemma 4.1, we establish Theorem 1.3 which is in fact an easy consequence of Theorem 1.4. Since all steps are standard, we will only sketch the arguments. We ﬁrst establish (3). Using (17), the approximate functional equation, with A = B = √q and α = it, and Cauchy’s inequality, we estimate the second moment in question by 2 ∞ ∗ 2 ∗ χ(m) m L(1/2+ it, χ) 2 1 Vit . | | ≤ +it √q q Q χ (mod q) q Q χ (mod q) m=1 m 2 X≤ X3 X≤ X3 X χ =χ0 χ =χ0

The analytic conductor of L(1/2+ it, χ) is q(1 + t ) so that by Proposition 5.4 of [IK], 1/2 R ≍ | | Vit(x) R (1+ x(1+ t )− )− for any R> 0. Thus we may truncate m so that m M := (Q(1 +≪t ))1/2+ε with| a| negligibly small error. ≤ Then| we| break the summations over q and m into dyadic intervals and remove the weight 2 Vit using the Mellin transform. We further write m = d n, where n is squarefree, and use the Cauchy-Schwarz inequality again. Eventually, we arrive at sums of the form 2

1 ∗ ∗ χ(m) 1 d +it d √2M Q

4.4. Proof of Lemma 4.1, estimate (30). The key point in establishing our cubic large sieve is the estimation of the norm C2(M, Q), which we do in this subsection. We point out that the ordinary large sieve inequality gives only the weaker bound C (M, Q) M + Q2. 2 ≪ Recall that C2(M, Q) is the norm associated to the sum 2 ′ S(M, Q) := bnχn(m) , M

Next, we extract the greatest common divisor δ of n1 and n2, changing variables via n1 δn1, n δn . The coprimality conditions become (n , n )=1and (δn , δn ) = 1. From→ these 2 → 2 1 2 1 2 conditions, combined with the facts that n1δ and n2δ are squarefree and have no rational prime divisor, we see that (N(n1), N(n2δ)) = 1. Thus

′ m SW (M, Q)= bn ∆δb W χn χn δ(m), 1 n2∆δ M 1 2 ∆,δ,n1,n2 (m,N(∆))=1 (N(n1)X,N(n2δ))=1 X 2 where we use that χδχδ = χδ = χδ. We still need to remove the coprimality condition in the sum over m before we can apply Poisson summation. Doing this by using the M¨obius function, we get m (40) S (M, Q)= ′ b b µ(l)χ χ (l) W χ χ (m). W n1∆δ n2∆δ n1 n2δ M/l n1 n2δ ∆,δ,n1,n2 l N(∆) m | (N(n1)X,N(n2δ)=1 X X

Now the characters χn1 χn2δ(m) in the above expression are primitive, and we have a smooth sum over m. Applying the Poisson summation formula in the form given in (18), we have

m Mτ(χn1 χn2δ) hM (41) W χn1 χn2δ(m)= χn1 χn2δ(h)W . M/l lN(n1n2δ) lN(n1n2δ) m Z h Z X∈ X∈ c MEAN VALUES WITH CUBIC CHARACTERS 19

When h = 0, then the summand above is zero unless n1 = n2 = δ = 1. Hence, the contribution of h =0 to SW (M, Q), say S0(M, Q) satisﬁes S (M, Q) M 1+ε ′ b 2 M 1+ε b 2. 0 ≪ | ∆| ≪ || || X∆ Let SW′ (M, Q) be the contribution to SW (M, Q) from h = 0. We analyze S′ now, where we need to show 6 5/3 ε 2 (42) S′ (M, Q) Q (QM) b . W ≪ | n| n X Our strategy is to apply Heath-Brown’s cubic large sieve estimate (9). We need the following easy consequence of (9): Suppose that dn,dn′ are arbitrary complex numbers supported on squarefree n Z[ω], n 1 (mod 3), with N(n) X. Then by Cauchy’s inequality, ∈ ≡ ≤ m 2 1/2 n 2 1/2 2/3+ε (43) dmdn′ ( dm ) ( dn′ ) X dm dn′ . | n 3 | ≤ | | | m 3 | ≪ || ||·|| || m,n m m n X X X X Proof of (42). First observe that we may freely truncate the sum over h for Q2l h (QM)ε =: H, | | ≤ N(δ)N(∆)2M

since W has rapid decay. Precisely, if we let SW′ (M, Q) = SW′′ (M, Q)+ E where SW′′ (M, Q) 100 2 is the contribution to S′ (M, Q) from 0 < h H, then E (MQ)− b . Further, we W | | ≤ ≪ || || note thatc the relevant range for n1, n2 is N(n1), N(n2) Q/N(∆δ) since the coeﬃcients bn are supported at n Z[ω] with N(n) Q. Combining≍ (40), (41), using Lemma 2.2, and changing variables n∈ n , we arrive at≍ the following bound 1 → 1 ′ ′ 1 1 (44) S′′ (M, Q) M U(∆, δ, l, h) , W ≪ N(δ)1/2 l | | ∆ δ l ∆ 0< h H X X X| X| |≤ where

′ hM n1 U(∆, δ, l, h) := W c∆,δ,l,h(n1)c∆′ ,δ,l,h(n2) , lN(n1n2δ) n2 3 N(n1),N(n2) Q/N(∆δ) X≍ (N(n1),N(n2))=1 c and the coeﬃcients c,c′ satisfy the bounds N(δ∆) 1/2 N(δ∆) 1/2 c (n) b , c′ (n) b . ∆,δ,l,h ≪ Q | n∆δ| ∆,δ,l,h ≪ Q | n∆δ| Now we are almost ready to use Heath-Brown’s cubic large sieve inequality in the form (43) to bound the sum U(∆, δ, l, h). The only obstacle is that the variables n1 and n2 are not separated due to the coprimality condition (N(n1), N(n2)) = 1 and the weight function W . This is only a technical obstacle since one can use M¨obius inversion to remove the coprimality condition, and the Mellin inversion formula to remove the weight function, both atc essentially no cost. Hence N(δ∆) 1/3 U(∆, δ, l, h) (QM)ε ′ b 2. ≪ Q | n| n X Inserting this into (44) and summing trivially over all the other variables gives (42). 20 STEPHAN BAIER AND MATTHEW P. YOUNG

4.5. Proof of Lemma 4.1, estimate (31). To prove the self-referential estimate (31) for C2(M, Q), we introduce a dual norm 2 2 ′ (45) C2′ (Q, M) := sup am − amχn(m) . (a ) || || m Q

2 1/v 2 1 1/v ′ C′ (Q, M) a − Q − c χ (m) , 2 ≪ || m|| m n Q

v 1 − 1/v 1 1/v 2 2 j v 1/v (46) C′ (Q, M) Q − a − c C′ (Q, 2 M ) . 2 ≪ || m|| | m| 2 j=0 2j M v

4.6. Proof of Lemma 4.1, estimates (32)-(36). In this section, we establish the remaining estimates (32)-(36) in Lemma 4.1 in which the norms Bi(Q, M) are compared. The estimate (32) says that the norm B1(Q, M) is essentially increasing in Q. It is easy to describe the idea behind the proof: simply take coefficients am supported on multiples of a fixed prime p. This extends the size of Q by a factor N(p) without essentially changing the size of the norm B1. There is a slight technical issue regarding coprimality with p that can be circumvented by averaging over p. The details are essentially the same as in Lemma 9 of [Hea1] and we therefore omit this proof. Next, we compare B1 and B2. We recall that in the definition (29) of B1, the outer sum ranges over squarefree n Z[ω] that are not divisible by any rational prime. We further recall ∈ that B2 is defined in the same way as B1 with the condition that n is not divisible by any rational prime being removed. Hence, we have the trivial inequality B (Q, M) B (Q, M). 1 ≤ 2 Conversely, we want to prove the estimate (33) of B2 in terms of B1. To reduce the sum over squarefree n Z[ω] to sums over squarefree n Z[ω] that are not divisible by any rational prime, we extract∈ rational divisors, getting ∈ 2 2 ∗ ∗ ′ ∗ T ∗(Q, M) := a χ (n) a χ (n) , m m ≤ m m Q

Breaking the outer sum over k on the right-hand side into O(log 2Q) dyadic intervals, we ﬁnd that 2 ∗ ′ ∗ T ∗(Q, M) log(2Q) sup amχm(n) ≪ 1 X Q ≤ ≤ (Q/X)1/2 k 2(Q/X)1/2 X/4

∗ N(n) n n am1 am2 W Q m1 3 m2 3 (m1,m2)=1 n Z[ω] X X∈ N(k)Q k k ∗ ˇ = Q bm1 bm2 W 2 , s(m1m2) m1 3 m2 3 k Z[ω] (m1,m2)=1 ! X∈ X with Wˇ being a certain weight function of rapid decay and √ 3 g(m) b := a − . m m m m2 3 Now, similarly as in [Hea2], we separate the variables m1 and m2 using the Mellin transform of the weight function Wˇ , and using M¨obius inversion on the coprimality condition (m1, m2)= 1. We then use the Cauchy-Schwarz inequality, and after a short calculation, arrive at the estimate for M 1 (47) ≥ ε 2 4 1 6+ε 2 ε B (Q, M) QM − max B (K, M): K M Q− + M /Q K− − B (K, M) 4 ≪ 3 ≤ 3 K>M 4/Q X where K runs over powers of 2. This corresponds to Lemma 8 in [Hea2]. We also have the trivial bound B (Q, M) Q if M < 1. Combining this with (47), we get (36). ✷ 4 ≪ References [BL] Baier, S.; Zhao, L. On the low-lying zeros of Hasse-Weil L-functions for elliptic curves, arXiv:0708.2987v3 [math.NT]. [BFH] Brubaker, B.; Friedberg, S.; Hoﬀstein, J. Cubic twists of GL(2) automorphic L-functions. Invent. Math. 160 (2005), no. 1, 31–58. [CFKRS] Conrey, J.; Farmer, D.; Keating, J.; Rubinstein, M.; Snaith, N. Integral moments of L-functions, to appear in Proc. London Math. Soc. [Da] Davenport, H. Multiplicative number theory. Third edition. Revised and with a preface by Hugh L. Montgomery. Graduate Texts in Mathematics, 74. Springer-Verlag, New York, 2000. [DFK] David, C.; Fearnley, J.; Kisilevsky, H. On the vanishing of twisted L-functions of elliptic curves. Experiment. Math. 13 (2004), no. 2, 185–198. [Di] Diaconu, A. Mean square values of Hecke L-series formed with r-th order characters. Invent. Math. 157 (2004), no. 3, 635–684. 22 STEPHAN BAIER AND MATTHEW P. YOUNG

[DT] Diaconu, A.; Tian, Y. Twisted Fermat curves over totally real fields. Ann. of Math. (2) 162 (2005), no. 3, 1353–1376. [E] Elliott, P. D. T. A. On the mean value of f(p). Proc. London Math. Soc. (3) 21 1970 28–96. [FaHL] Farmer, D.; Hoffstein, J.; Lieman, D. Average values of cubic L-series. Automorphic forms, automorphic representations, and arithmetic (Fort Worth, TX, 1996), 27–34, Proc. Sympos. Pure Math., 66, Part 2, Amer. Math. Soc., Providence, RI, 1999. [FKK] Fearnley, J.; Kisileveky, H.; Kuwata, M. Vanishing and Non-Vanishing Dirichlet Twists of L- Functions of Elliptic Curves. arXiv:0711.1771v1 [math.NT]. [FrHL] Friedberg, S.; Hoffstein, J.; Lieman, D. Double Dirichlet series and the n-th order twists of Hecke L-series. Math. Ann. 327 (2003), no. 2, 315–338. [Hea1] Heath-Brown, D.R. A mean value estimate for real character sums. Acta Arith. 72 (1995), 235–275. [Hea2] Heath-Brown, D.R. Kummer’s conjecture for cubic Gauss sums. Isr. J. Math. 120 (2000), 97–124. [H-BP] Heath-Brown, D. R.; Patterson, S. J. The distribution of Kummer sums at prime arguments. J. Reine Angew. Math. 310 (1979), 111–130. [IR] Ireland, K.; Rosen, M. A classical introduction to modern number theory. Second edition. Graduate Texts in Mathematics, 84. Springer-Verlag, New York, 1990. [I1] Iwaniec, H. Topics in Classical Automorphic Forms, Graduate Studies in Mathematics, 17. American Mathematical Society, Providence, RI, 1997. [I2] Iwaniec, H. On the order of vanishing of modular L-functions at the critical point. Sém. Théor. Nombres Bordeaux (2) 2 (1990), no. 2, 365–376. [IK] Iwaniec, H.; Kowalski, E. Analytic Number Theory. American Mathematical Society Colloquium Pub- lications, 53. American Mathematical Society, Providence, RI, 2004. [KS1] Katz, N.; Sarnak, P. Random Matrices, Frobenius Eigenvalues, and Monodromy. American Mathe- matical Society Colloquium Publications, 45. American Mathematical Society, Providence, RI, 1999. [KS2] Katz, N.; Sarnak, P. Zeroes of zeta functions and symmetry, Bull. Amer. Math. Soc., 36, 1-26 (1999). [Lem] Lemmermeyer, F. Reciprocity laws. From Euler to Eisenstein. Springer Monographs in Mathematics, Springer, Berlin (2000). [LP] Livné, R.; Patterson, S. J. The first moment of cubic exponential sums. Invent. Math. 148 (2002), no. 1, 79–116. [L] Luo, W. On Hecke L-series associated with cubic characters. Compos. Math. 140 (2004), no. 5, 1191– 1196. [P1] Patterson, S. J. A cubic analogue of the theta series. J. Reine Angew. Math. 296 (1977), 125–161. [P2] Patterson, S. J. The distribution of general Gauss sums and similar arithmetic functions at prime arguments. Proc. London Math. Soc. (3) 54 (1987), no. 2, 193–215. [Y] Young, M. P. The first moment of quadratic Dirichlet L-functions, Acta Arith. 138 (2009) no.1, 73–99.

Jacobs University, School of Engineering and Science, P.O. Box 750561, 28725 Bremen, Germany E-mail address: [email protected] Department of Mathematics, Texas A&M University, College Station, TX 77843-3368, U.S.A. E-mail address: [email protected]