Access the full text.

Sign up today, get DeepDyve free for 14 days.

Mathematics
, Volume 2020 (1410) – Oct 29, 2014

/lp/arxiv-cornell-university/the-frequency-and-the-structure-of-large-character-sums-GZjRI80xKP

- ISSN
- 1435-9855
- eISSN
- ARCH-3343
- DOI
- 10.4171/JEMS/799
- Publisher site
- See Article on Publisher Site

THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS JONATHAN BOBER, LEO GOLDMAKHER, ANDREW GRANVILLE, AND DIMITRIS KOUKOULOPOULOS Abstract. Let M () denote the maximum of j (n)j for a given non-principal nN Dirichlet character (mod q), and let N denote a point at which the maximum is at- tained. In this article we study the distribution of M ()= q as one varies over characters (mod q), where q is prime, and investigate the location of N . We show that the dis- tribution of M ()= q converges weakly to a universal distribution , uniformly through- out most of the possible range, and get (doubly exponential decay) estimates for 's tail. Almost all for which M () is large are odd characters that are 1-pretentious. Now, j2(2)j M () j (n)j = qjL(1; )j, and one knows how often the latter expres- nq=2 sion is large, which has been how earlier lower bounds on were mostly proved. We show, though, that for most with M () large, N is bounded away from q=2, and the value of M () is little bit larger than jL(1; )j. Contents 1. Introduction 1 2. The structure of characters with large M () 7 3. Auxiliary results about smooth numbers 9 4. Outline of the proofs of Theorems 1.1 and 1.3 and proof of Theorem 1.2 18 5. Truncating P olya's Fourier expansion 22 6. The distribution function 32 7. Some pretentious results 34 8. Structure of even characters with large M (): proof of Theorem 2.3 42 9. The structure of characters with large M (): proof of Theorems 2.1 and 2.2 46 10. Additional tables 56 References 56 1. Introduction For a given non-principal Dirichlet character (mod q), where q is an odd prime, let M () := max (n) : 1xq nx Date : March 27, 2018. 2010 Mathematics Subject Classi cation. Primary: 11N60. Secondary: 11K41, 11L40. Key words and phrases. Distribution of character sums, distribution of Dirichlet L-functions, pretentious multiplicative functions, random multiplicative functions. arXiv:1410.8189v2 [math.NT] 20 Jun 2017 2 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS This quantity plays a fundamental role in many areas of number theory, from modular arithmetic to L-functions. Our goal in this paper is to understand how often M () is large, and to gain insight into the structure of those characters (mod q) for which M () is large. It makes sense to renormalize M () by de ning M () m() = ; e q= and we believe that m() (1 + o (1)) log log q; (1.1) q!1 where is the Euler{Mascheroni constant. R enyi [R en47] observed that m() c +o (1) q!1 with c = e = 12 = 0:509 : : : Upper bounds on M () (and hence on m()) have a rich history. The 1919 P olya{ Vinogradov Theorem states that m() log q for all non-principal characters (mod q). Apart from some improvements on the implicit constant [Hil88, GS07], this remains the state-of-the-art for the general non-principal char- acter, and any improvement of this bound would have immediate consequences for other number theoretic questions (see e.g. [BG16]). Montgomery and Vaughan [MV77] (as im- proved in [GS07]) have shown that the Generalized Riemann Hypothesis implies m() (2 + o (1)) log log q; q!1 whereas, for every prime q, there are characters (mod q) for which m() (1 + o (1)) log log q; q!1 (see [BC50, GS07], which improve on Paley [Pal]), so the conjectured upper bound (1.1) is the best one could hope for. However, for the vast majority of the characters (mod q), M () is somewhat smaller, and so we study the distribution function ( ) := #f (mod q) : m() > g : '(q) Montgomery and Vaughan [MV79] showed that ( ) for all 1 for any xed q C C 1 (they equivalently phrase this in terms of the moments of M ()). This was recently improved by Bober and Goldmakher [BG13], who proved that, for xed and as q ! 1 through the primes, +A n o e 1=4 =(log ) (1.2) exp 1 + O(1= ) ( ) exp Be ; where B is some positive constant and Z Z 2 1 log I (t) dt (1.3) A = log 2 1 dt (log I (t) t) = 0:088546 : : : ; 2 2 t t 0 2 and I (t) is the modi ed Bessel function of the rst kind given by 2 n (t =4) I (t) = : (1.4) n! n0 We improve upon these results and demonstrate that ( ) decays in a double exponential fashion: THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 3 Theorem 1.1. Let = e log 2. If q is a prime number, and 1 log log q M for some M 4, then +A 2 e e exp (1 + O( )) ( ) exp (1 + O( )) ; 1 q 2 where 2 M=2 = (log ) = + e and = (log )= 1 2 The proof implies that 1 + o (1) of the characters for which m() > are odd, so it !1 makes sense to consider odd and even characters separately. Hence we de ne ( ) = #f (mod q) : (1) = 1 and m() > g ; (1.5) (q)=2 We outline the proof of Theorem 1.1 in Section 4, and ll in the details in subsequent sections. The range of uniformity barely misses implying the upper bound in (1.1); even so, it does show that m() is rarely large. Calculations reveal that ; ; each tend to a universal q q distribution function ; ; , and we will show this for later on. In Figure 1 we graph ( ) for a typical q: Figure 1. , , and for q = 12000017 q q Notice that ( ) = 1 for all :7227, and then it decays quickly: (1) :697; (2) q q q 0:0474 and (3) :000538. For more computational data, the reader is invited to consult Table 1 in Section 10. The distribution of jL(1; )j, for q prime, decays similarly [GS06]: +A 1 e (1.6) ( ) := #f (mod q) : jL(1; )j > e g = exp (1 + O( )) (q) 1=2 M=2 with = + e , uniformly for 1 log log q M , M 1. This similarity is no accident, since for an odd, non-primitive character (mod q), the average of the character sum is G() E (n) = L(1; ); 1Nq nN 4 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS where G() is the Gauss sum, so that jG()j = q. In particular, if L(1; ) is large, then so is m(). Moreover, for as above, we have the pointwise formula G() (n) = (2 (2)) L(1; ); nq=2 which implies that (p) m() 2e 1 ; p>2 a little larger than jL(1; )j if j1(2)j 1. The distribution of 2j (1(p)=p) j can p>2 be analyzed in the same way as the distribution of jL(1; )j. However, even if (2) = 1, then we can show that the average of our character sum up to N is slightly larger than jL(1; )j when N is close to q=2. This builds on ideas in [Bob14]. Theorem 1.2. Let q be an odd prime, 1 log log q C for a suciently large absolute constant C . There exists a subset C ( ) f (mod q) : (1) = 1; jL(1; )j > e g of cardinality L e = (1.7) #C ( ) = 1 + O e #f (mod q) : (1) = 1; jL(1; )j > e g such that if 2 C ( ) and c > 0 is suciently large, then " # G() p (log ) E c c (n) (L(1; ) + log 2) q p : (1.8) q=2q= Nq=2+q= nN We can deduce that m() + e log 2 + O((log ) = ): This result, together with (1.6), suggests that the lower bound in Theorem 1.1 should be sharp, and that m() e (jL(1; )j + log 2) when these quantities are \large". However the numerical evidence indicates that m() is perhaps typically a little bigger: 4.4 4.3 4.2 4.1 4.0 4.0 4.1 4.2 4.3 4.4 Figure 2. Scatter plot of m() (vertical axis) and e (jL(1; )j + log 2) for 9 9 the 13617 odd characters with jL(1; )j > 6:4, and q prime, 10 q 10 + 75543. THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 5 The numerical comparison between and (the jL(1; )j-distribution for odd char- acters ) is given in Figure 3: 1 2 3 Figure 3. log( log ( )) log( log ( )), q = 12000017, with a dashed q q line indicating e log 2. Even characters. The focus above has been on odd characters since almost all characters with m() > are odd. However, we can obtain analogous results for even characters. Theorem 1.3. There exists an absolute constant c 1 such that if q is an odd prime and 1 (log log q M )= 3 for some M 1, then ( ) ( ) p p 3 +A 3 e e 1=2 M=2 + exp p 1 + O + e ( ) exp : The lower bound in the above theorem has the same shape as (1.6) and it might be close to the truth, unlike the corresponding result for (as explained by Theorem 1.2). This is supported by computations, as it can be seen in Figure 4. 1 1.5 2 2.5 L + Figure 4. log( log ( )) log( log ( )), q = 12000017. 3q q 6 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS The distribution function. Classically, the statistical behavior of (n) as varies over all non-principal characters mod q is modelled by X , where the fX : p primeg are indepen- n p dent random variables, each uniformly distributed on the unit circle U := fz 2 C : jzj = 1g, e e e e 1 k 1 k and X = X X for n = p : : : p . This suggests modelling L(1; ), (mod q), by p p 1 1 k L(1; X ) = X =n. Indeed, in [GS03] this was shown to be a very successful model. n1 Similarly, one might guess that the distribution of f (n)g could be accu- (mod q); 6= nN 0 rately modelled by X (n) (and thus M () by M (X )), but this seems unlikely since, nN for xed 2 (0; 1), we have that 2 3 2 2 X X X 4 5 E X q whereas (n) (1 )q (q) nq nq (mod q) 6= as q ! 1. Moreover Harper [Har13] recently showed that X (n) is not normally nN distributed, and subsequently we have no idea what its distribution should be (though [CS12] shows that X (n) is normally distributed provided y is not too large). N<nN +y So, what is the right way to model M ()? One reason the above model failed is that it does not take account of characters' periodicity. The periodicity of , via the formula (n)e(an=q) = (a)G(); n=1 2ix where e(x) = e , leads to P olya's formula [MV07, eqn. (9.19), p. 311]: X X (n)(1 e n) G() q log q (n) = + O 1 + (1.9) 2i n z nq 1jnjz p p (which is periodic for (mod 1) ). SincejG()j = q, this suggests modelling M ()=( q=2) = 2e m() by the random variable X (1 e(n)) S = max ; 01 n2Z n6=0 where X is a random variable independent of the X 's, with probability 1/2 of being 1 1 p or 1, and we have set X = X X . The in nite sum here converges with probability 1, n 1 n so that S is well de ned, and we can consider its distribution function, ( ) := Prob(S > 2e ): The following result con rms our intuition that S serves as a good model for 2e m(). Theorem 1.4. If b a > 0 and is continuous at every point of [a; b], then the sequence of functions f g converges to uniformly on [a; b]. In particular, the sequence of q q prime distributions f g converges weakly to . q q prime Theorem 1.4 will be shown in Section 6. Since is a decreasing function, it has at most countably many discontinuities, and so Theorem 1.4 implies that lim ( ) = q!1; q prime q ( ) for almost all > 0. We conjecture that is a continuous function. This follows by the methods of Section 5 applied to X in place of (n). n THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 7 We can prove that, for any 1, +A 2 2 e (log ) e log exp 1 + O ( ) exp 1 + O by considering two points and where is continuous, with < < so that 1 2 1 2 ( ) ( ) ( ). We can bound ( ) and ( ) by Theorems 1.4 and 1.1, and then 2 1 1 2 the result follows by letting ! and ! . 1 2 Acknowledgements. The authors would like to thank Sandro Bettin, Adam Harper, K. Soundararajan, Akshay Venkatesh, John Voight, and Trevor Wooley for some helpful discus- sions. AG and DK are partially supported by Discovery Grants from the Natural Sciences and Engineering Research Council of Canada. Notation. Given a positive integer n, we let P (n) and P (n) be the largest and small- est prime divisors of n, with the notational convention that P (1) = 1 and P (1) = 1. Furthermore, we write P (y; z) for the set of integers all of whose prime factors lie in (y; z]. The symbol d (n) denotes the number of ways n can be written as a product of k positive integers. We denote with !(n) the number of distinct primes that divide n, and with (n) the total number of prime factors of n, counting multiplicity. Finally, we use the symbol 1 to indicate the characteristic function of the set A. For example, 1 equals 1 if A (n;a)=1 (n; a) = 1 and 0 otherwise. 2. The structure of characters with large M () We now determine more precise information about the structure of characters with large M (). Theorem 2.1. Let q be an odd prime and 1 log log q C for a suciently large absolute constant C . There exists a set C ( ) f (mod q) : m() > g of cardinality e = (2.1) #C ( ) = 1 + O e #f (mod q) : m() > g for which the following holds: If 2 C ( ), then (1) is an odd character. (2) Let = N =q 2 [0; 1], and a=b be the reduced fraction for which j a=bj 1=(b ) with b . Let b = b if b is prime, and b = 1 otherwise. Then 0 0 3=4 j1 (p)j (log ) ; and 1=4 (2.2) pe p6=b b (p) m() = e 1 + O log : (2.3) (b ) p p6=b Granville and Soundararajan showed in [GS07] that if M () is large then must pretend to be a character of small conductor and of opposite parity (that is, is \ -pretentious"). Building on the techniques of [GS07], Theorem 2.2 shows that, for the vast majority of characters with large M (), the character is the trivial character, and must be odd (and so 1-pretentious as in (2.2)). The discrepancy between Theorem 1.1 and (1.6) means that the error term in (2.3) cannot be made o(1). 8 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Inequality (2.2) allows us to establish accurate estimates for (n) for all N . Here nN and for the rest of the paper, for u 0 we set P (u) = e (t)dt; (2.4) where is the Dickman{de Bruijn function, de ned by (u) = 1 for u 2 [0; 1] and via the dierential-delay equation u (u) = (u 1) for u > 1. Then we have the following estimate: Theorem 2.2. Let q; ; ; a; b; b and 2 C ( ) be as in Theorem 2.1. Given 2 [0; 1], let 0 q 10 10 k=` be the reduced fraction for which j k=`j 1=(` ) with ` . De ne u > 0 by j k=`j = 1=(`e ). (a) If b = 1, then 3=4 e i 1 P (u) + O(( log ) ) if ` = 1; (n) = 3=4 G() + O(( log ) ) if ` > 1: n q (b) If b = b, then e i 1 1=b 3=4 (n) = + O(( log ) ); G() 1 (b)=b n q where 1 P (u) if ` = 1; < v1 (b) 1 (b) 1 + P (u) if ` = b , v 1; b b 1 1 otherwise: Moreover, if we take = and write j a=bj = 1=(be ), then we have that 2 3=4 j1 (b)j (log ) (1 P (u )) : 2 1=4 Notice that if = k=` then P (u) = 1, so 1 P (u) is a measure of the distance between and k=`. Theorems 2.1 and 2.2 will be proved in Section 9. Analogous results can also be proven for even characters: Theorem 2.3. Let q be an odd prime and 1 log log q C for a suciently large constant C . There exists a set C ( ) f (mod q) : (1) = 1; m() > g of cardinality + e = (2.5) #C ( ) = 1 + O e #f (mod q) : (1) = 1; m() > g for which the following statements hold. If 2 C ( ) then (1) If = N =q 2 [0; 1], and a=b is the reduced fraction for which j a=bj 1=(b ) with b , then b = 3. (2) We have that (p) log ; and (2.6) pe p6=3 THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 9 e 3 (2.7) m() = L 1; + O(log ): 2 3 (3) Given 2 [0; 1], let k=` be the reduced fraction for which j k=`j 1=(` ) with 10 3u ` . De ne u > 0 by j k=`j = 1=(`e ). Then v1 > k (3 ) > v P (u) + O( log ) if ` = 3 for some v 1; X v1 3 3 (n) = G() > nq p O( log ) otherwise: The proof of Theorem 2.3 will be given in Section 8. 3. Auxiliary results about smooth numbers Before we get started with the proofs of our main results, we state and prove some facts about smooth numbers. As usual, we set (x; y) = #fn x : P (n) yg: We begin with the following simple estimates. Lemma 3.1. For u 1, we have that O(1) (u);j (u)j : u log u Proof. The claimed estimates follow by Corollary 2.3 in [HT93] and Corollary 8.3 in Section III.5 of [Ten95]. Lemma 3.2. For y 2 and u 1, we have that X p 1 log y log y + e : n u n>y P (n)y Proof. Without loss of generality, we may assume that y 100. When v 2 [1; log y], de Bruijn [dB51] showed that log(v + 1) v v (3.1) (y ; y) = y (v) 1 + O : log y Therefore, Lemma 3.1 yields that v v (3.2) (y ; y) (y=v) 1 v log y : Together with partial summation, this implies that 1 log y n u 3=2 u (log y) y <ne P (n)y Finally, if = 2= log y, then we note that X X Y p p 1 1 1 2 log y 2 log y e e 1 : 1 1 n n p 3=2 P (n)y py (log y) n>e P (n)y 10 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Since p = 1 + O(log p= log y) for p y, the product over the primes is log y, which completes the proof of the lemma. We also need the following result, where P is de ned by (2.4). Lemma 3.3. If y 2 and u > 0, then = P (u) e log y + O(1): ny P (n)y In particular, lim P (u) = 1. u!1 Proof. If u 1, then the lemma follows from the estimate 1=n = log x +O(1). Assume nx now that u 1, and set v = minfu; log yg. Lemma 3.2 implies that X X p 1 1 log y = + O(e ): n n u v ny ny + + P (n)y P (n)y We use de Bruijn's estimate (3.1) and partial summation to deduce that Z v X X 1 1 d (y ; y) = log y + O(1) + = log y + O(1) + n n y v u y ny y<ny + + P (n)y P (n)y = log y + O(1) + (log y) (t)dt = (log y) (t)dt + O(1) = (log y) (t)dt + O(1); thus completing the proof of the lemma. Remark 3.1. Using the slightly more accurate approximation 0 u ( 1)x (u) xe (x; y) = x(u) + + O log y (log y) for u 2 [; 1][ [1 + ; log y] (see [Sai89]), one can similarly deduce the stronger estimate = P (u) e log y + (u) + o (1) y!1 ny P (n)y for all u > 0. Finally, we have the following key estimate. Its second part is a strengthening of Theorem 11 in [FT91]. Lemma 3.4. Let y 10 and 2 [1=2; 1=2). (a) We have that X X e(n) 1 = + O(1): n n P (n)y n1=jj P (n)y THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 11 (b) There is an absolute constant c 2 such that if c 2, a=b is a reduced fraction 0 1 c c c 1 1 0 with b (log y) and j a=bj minf1=(b(log y) );jj=(log y) g, then e(n) 1 (b) log log y = log + (1 (u)) + O ; n 1 e(a=b) (b) log y P (n)y where u is de ned by y = j a=bj and is von Mangoldt's function. Proof. (a) If n 1=jj, then e(n) = 1 + O(njj), and so X X X e(n) 1 jj 1: n n n1=jj n1=jj n1=jj + + P (n)y P (n)y Hence it suces to show that e(n) 1: (3.3) n n>1=jj P (n)y Note that if y > 1=jj, then e(n) 1=jj<ny using partial summation on the estimate e(n) 1=jj. Therefore, in order to com- nx plete the proof of (3.3), it suces to show that e(n) 1; n>w P (n)y where w := maxfy; 1=jjg. Moreover, Lemma 3.2 reduces the above relation to showing that e(n) 1; (3.4) n w<nw P (n)y 0 2 log log y= log log log y 0 u where w = y . In particular, we may assume that jj 1=w . If x = y 2 [w; w ], then Theorem 10 in [FT91] and relation (3.2) imply that log(u + 1) x e(n) (x; y) ; log y e log y nx P (n)y provided that jj > (log y) =x, for C suciently large. (This lower bound on jj guarantees that the parameter q in Theorem 10 of [FT91] is 2, as required.) Consequently, if we set 00 C w = maxfy; (log y) =jjg, then partial summation implies that e(n) 1: 00 0 w <nw P (n)y 12 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Then (3.4) follows from e(n) 1; (3.5) w<nw P (n)y C 00 C which is trivial unless jj (log y) =y, so that w = (log y) =jj. C+1 m Let y = w (1 + 1=(log y) ) for m 1. Let M be the largest integer for which 00 C+1 y w . Then M (log y) log log y. If x 2 (y ; y ] for some m 2 f1; : : : ; Mg and M m1 m log x u = , then Theorem 3 in [Hil86] implies that log y log(u + 1) (x; y) = (y ; y) + (x y )(u) 1 + O m2 m2 log y = (y ; y) + (x y )(u) + O : m2 m2 C+2 (log y) Therefore, partial summation and Lemma 3.1 imply that e(n) e(x) x y 1 +jj(y y ) m1 m m1 = (u) + (u) dx + O C+2 n x x log y (log y) m1 y <ny m1 m P (n)y e(x) 1 +jj(y y ) m m1 = (u)dx + O : C+2 x (log y) m1 Summing the above relation over m 2 f1; : : : ; Mg and bounding trivially the contribution of n 2 (y ; w ] implies that e(n) e(x) log log y = (u)dx + O : n x log y w<nw P (n)y Finally, integration by parts gives us that Z Z 00 00 w w e(x) d e(x) (u) (u)dx = dx x dx 2i x w w 00 Z w 0 e(x) e(x) (u) (u) = (u) + dx 2 2 2ix 2i x x log y x=w 1; jjw by the de nition of w and Lemma 3.1. This completes the proof of (3.5) and thus the proof of part (a) of the lemma. (b) Write = a=b + and set N = 1=(jj log y). We note that, since jj jj=(log y) < jj, by assumption, we must have that b 2. We start by estimating the part of the sum with n > N . We claim that e(n) log log y (3.6) n log y P (n)y n>N THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 13 Note that if y > N , then e(n) 1 jj log y 1 n jjN jj log y N<ny by partial summation and the estimate e(n) 1=jj. Therefore, it suces to prove nx that e(n) log log y n log y P (n)y n>N where N := maxfN; yg. Furthermore, Lemma 3.2 reduces the above estimate to showing that e(n) log log y (3.7) n log y P (n)y 0 00 N <nN 00 2 log log y= log log log y 00 where N = y . For this sum to have any summands we need that N N . 0 00 Next, we x X 2 [N ; N =2] and estimate e(n) P (n)y nx for x 2 [X; 2X ]. Let B be the constant in Theorem 10 of [FT91] when the parameters A and there equal 100 and 1=10, respectively, and set c = B + 1 and Q = 2X=(log y) . There is some reduced fraction r=s = r(X )=s(X ) with j r=sj 1=(sQ) and s Q. Note that B B (log y) (log y) 2 jj (log y) jj = : N X Q In particular, we must have that s 2. If, in addition, s (log y) =2, then Theorem 11 in [FT91] and relation (3.2) imply that 1 log(u + 1) 1 x e(n) (x; y) + ; 1:9 100 2:9 (log y) log y (log y) (log y) P (n)y nx log x where u = . Therefore log y e(n) 1 2:9 n (log y) (3.8) P (n)y X<nminf2X;N g if s (log y) =2. Consider now X such that 2 s < (log y) =2 and set = r=s. We claim that B+1 B+1 X N (log y) . We argue by contradiction: assume, instead, that X > N (log y) . Then Q > 2N log y = 2=jj and thus jj > 1=(sQ) and Q > 2b. In particular, r=s 6= a=b and thus 1=(bs) ja=b r=sj jj + 1=(sQ) jj + 1=(2bs), which implies that s 1=(2bjj) c 2 (log y) =2 (log y) =2, a contradiction. 14 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Using the above information, we are going to show that e(n) 1 (3.9) n log y P (n)y X<nminf2X;N g when 2 s < (log y) =2. If this relation holds, then (3.7) follows by a straightforward dyadic decomposition argument. We use the stronger relation (3.8) for the O(log N ) = 1:1 B+1=2 00 O((log y) ) dyadic intervals corresponding to X 2 [N (log y) ; N ], and we use (3.9) for 0 B+1=2 00 the O(log log y) dyadic intervals with X 2 [N ; minfN (log y) ; N g]. It remains to prove (3.9) when 2 s < (log y) =2, in which case X lies in the inter- 0 B+1 00 B val [N ; minfN (log y) ; N g]. In addition, note that jj 1=(sQ) (log y) =X . Let B+1 m X = X (1 + 1=(log y) ) for m 1. Let M be the largest integer for which X m M 00 B+1 minf2X; N g. Then M (log y) . Consider x 2 (X ; X ] for some m 2 f1; : : : ; Mg m1 m log x and set u = . Since s (log y) =2, Theorems 2 and 5 in [FT91] imply that log y X X X e(nr=s) = e(mr=c) + + s=cd P (n)y P (n)y nx mx=d (m;c)=1 X X X e(rj=c) x = 1 + O B+100 (c) (log y) 1jc s=cd P (n)y (j;c)=1 mx=d (m;c)=1 X X (c) x = (g) (x=(gd); y) + O ; B+100 (c) (log y) s=cd gjc since e(rj=c) = (c) (see, for example, Lemma 7.6 below). Note that 1jc; (j;c)=1 X X (c) 1 (g) = 0 (c) gd s=cd gjc for s > 1. Therefore X X X (c) x (x; y) x e(nr=s) = (g) ; y + O : B+100 (3.10) (c) gd gd (log y) s=cd P (n)y gjc nx We apply Theorem 3 in [Hil86] to nd that x (x; y) X (X ; y) x X log k m2 m2 m2 ; y = ; y + u (u) k k k k k log y + O : B+2 k(log y) THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 15 for x 2 [X ; X ] and kjs. Consequently, m1 m X X X (c)(g)(x X ) log k m2 e(nr=s) = u (u) dg(c) log y s=cd gjc P (n)y nx + c + O ; B+2 (log y) for some constant c = c(X; y; m; r; s) 2 R. Since e(n) = e(n)e(nr=s), applying partial summation as in the proof of part (a), we deduce that X X X e(n) (c) (g) e(x) log(dg) = u (u) dx n d(c) g x log y m1 X <nX s=cd gjc m1 m P (n)y 1 +jj(X X ) m m1 + O : B+2 (log y) 0 00 Setting X = minf2X; N g, summing the above relation over m 2 f1; : : : ; Mg, and bounding trivially the contribution of n 2 (X ; X ] implies that Z 0 X X X e(n) (c) (g) e(x) log(dg) 1 = u (u) dx + O : n d(c) g x log y log y s=cd X<nX gjc P (n)y Finally, we have that Z 0 Z X 2X e(x) log(dg) log(dg) dx (log 2) log(dg) u (u) dx = ; x log y log y x log y X X whence 0 1 X X X e(n) 1 c log(dg) 1 @ A 1 + ; n log y d(c) log y X<nX s=cd gjc P (n)y which completes the proof of (3.9) and thus of (3.6). To conclude, we have shown that X X e(n) e(n) log log y = + O : n n log y + + P (n)y P (n)y nN Since e(n) = e(n) + O(njj), we further deduce that X X e(n) e(an=b) log log y = + O : n n log y + + P (n)y P (n)y nN Observe that X X e(na=b) 1 e(an=b) = log n 1 e(a=b) n nminfN;yg n>minfN;yg (3.11) 1 1 = log + O ; 1 e(a=b) minfN; ygka=bk 16 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS where kxk denotes the distance of x from its nearest integer. Note that ka=bk jjjj c c 0 0 jj=2 jj(log y) =2 by our assumption that jj jj=(log y) . Since we also have that jj 1=(2(log y) ), a consequence of the fact that b 2, the error term in (3.11) is 1 1 1 2(log y) 1 + + c 1 Njj yjj (log y) y log y by our assumption that jj jj=(log y) . Hence, the lemma is reduced to showing that e(an=b) (b) log log y = (1 (u)) + O : n (b) log y P (n)y y<nN By Lemma 3.2, it suces to show that e(an=b) (b) log log y = (1 (u)) + O ; (3.12) n (b) log y P (n)y y<ny v 2 log log y= log log log y c where y := minfN; y g. Since 2 b (log y) , the argument leading to (3.10) implies that X X X (c)(g) x (x; y) x e(an=b) = ; y + O (c) gd gd (log y) b=cd gjc P (n)y nx for x 2 [y; y ]. Therefore partial summation yields 0 1 X X X X X B C e(an=b) (c)(g) 1 1 1 B C = + O @ A 98 n dg(c) n n (log y) v v y<ny b=cd gjc y=(dg)<ny =(dg) P (n)y v + y<ny P (n)y P (n)y 0 1 X X X B C (c)(g) 1 1 B C = log(dg) + O : @ A 98 dg(c) n (log y) v v b=cd gjc y =(dg)<ny P (n)y Arguing as in Lemma 3.3, we apply partial summation on relation (3.1) to deduce that 1 (1 + log(dg)) = (v) log(dg) + O : n log y v v y =(dg)<ny P (n)y Consequently, X X X e(an=b) (c)(g) log(dg) 1 = (1 (v)) + O n dg(c) log y b=cd gjc P (n)y y<ny (b) 1 = (1 (v)) + O (b) log y (b) log log y = (1 (u)) + O ; (b) log y THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 17 since v = minfu log log y= log y; 2 log log y= log log log yg and (u) u by Lemma 3.1. This completes the proof of (3.12) and thus the proof of the lemma. Corollary 3.5. Let be a Dirichlet character modulo q and 2 R. For all z; y 1, we have (n)(1 e(n)) log log y 2e log y + 2 log 2 + O : n log y 1jnjz P (n)y Proof. If is an even character, then (n)=n + (n)=(n) = 0, so that X X X (n)(1 e(n)) (n)e(n) 1 1 = 2 2 = 2e log y + O ; n n n log y nz 1jnjz P (n)y P (n)y P (n)y by the Prime Number Theorem, as desired. We may therefore assume that is an odd character, so that X X X (n)(1 e(n)) (n)(1 cos(2n)) 1 cos(2n) = 2 2 : n n n nz n1 1jnjz + + P (n)y P (n)y P (n)y If jj 1=(log y) , where c is the constant from Lemma 3.4(b), then Lemma 3.4(a) implies that X X X X 1 cos(2n) 1 1 1 = + O(1) + O(1) n n n n + c n1 n>1=jj n(log y) 0 P (n)y P (n)y P (n)y e log y log log y + O(1); which is admissible. Finally, assume that jj > 1=(log y) , and consider a reduced fraction 2c 2c c 0 0 0 a=b with b (log y) and j a=bj 1=(b(log y ) ) jj=(log y) . Then Lemma 3.4(b) and the Prime Number Theorem imply that 1 cos(2n) (b) log log y = e log y + logj1 e(a=b)j (1 (u)) + O ; n (b) log y n1 P (n)y where j a=bj = y . Since (u) 1 and j1 e(a=b)j 2, we deduce that 1 cos(2n) log log y e log y + log 2 + O ; n log y n1 P (n)y which concludes the proof. 18 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS 4. Outline of the proofs of Theorems 1.1 and 1.3 and proof of Theorem 1.2 We rst deal with Theorem 1.3. For the lower bound, note that if is even and q > 3, then [MV07, eqn. (9.18), p. 310] yields the inequality X X q (n)(e(n=3) e(n=3)) M () (n) = : 2 n n=1 nq=3 Since we also have that (4.1) e(n=3) e(n=3) = i 3 ; we deduce that p p (n) 3q 3q M () = L 1; 2 n 2 3 n=1 for all even characters . The lower bound in Theorem 1.3 is a direct consequence of the above inequality and of the following result, whose proof is a straightforward application of the methods in [GS07]: Theorem 4.1 (Granville - Soundararajan). If is a character modulo some b 2 f1; 2; 3g, X is the set of odd or of even characters mod q, and 1 log log qM for some M 0, then +A 1 (b) e 1=2 M=2 # 2 X : jL(1; )j > e = exp 1 + O + e : jXj b Finally, the upper bound in Theorem 1.3 follows from Theorems 4.1 and 2.3. (The proof of the latter theorem is independent of the proof of the upper bound of Theorem 1.3, as we will see.) Next, we turn to Theorem 1.1. Its lower bound is a direct consequence of Theorems 1.2 and 4.1. So it remains to outline the proof of the upper bound in Theorem 1.1, as well as to prove Theorem 1.2. In the heart of these two proofs lies a moment estimate which implies that the bulk of the contribution in P olya's Fourier expansion (1.9) comes from smooth inputs: X X (n)(1 e(n)) (n)(1 e(n)) n n 1jnjz 1jnjz P (n)y for most and any . To state this more precisely, we need some notation. Given a set of positive integers A, set (n)e(n) S () = max ; 2[0;1] n n2A THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 19 in the special case when A = fn 2 N : n z; P (n) > yg, write S () in place of S (). y;z A Observe that X X (n)(1 e(n)) (n)(1 e(n)) max 2[0;1] n n 1jnjz 1jnjz P (n)y X X X (n)(1 e(n)) (n) (n)e(n) = max + max (4.2) 2[0;1] 2[0;1] n n n 1jnjz 1jnjz 1jnjz + + + P (n)>y P (n)>y P (n)>y X X (n) (n)e(n) 2 + 2 max 4S (): y;z n 2[0;1] n nz nz + + P (n)>y P (n)>y Our next goal is to show that S () is small for most . To do this, we will prove in Section y;z 5 that high moments of S () are small. As a straightforward application of our moment y;z bounds we will get the following theorem. 11=21 Theorem 4.2. If q 2 N, 3 y q and 2 [1= log y; 1], then # (mod q) : S () > e 11=21 y log log y y;q 1=(500 log log q) exp 1 + O + q : (q) log y log y We now show how to complete the proof of Theorems 1.1 and 1.2. Proof of the upper bound in Theorem 1.1. Let 2 [0; 1] be such that M () = (n) . nq By P olya's expansion (1.9), (4.2) and Lemma 3.5, we nd that e (n)(1 e(n)) 1=43 m() = + O(q ) 2 n 11=21 1jnjq log log y log y + + 2e S 11=21 () + O y;q log y for all y 10, where = e log 2. We set y = e for some parameter 2 [1= log y; 1] to be chosen shortly. Theorem 4.2 then implies that 2 2 e log 1=(500 log log q) ( ) exp 1 + O + q : Taking = 1 completes the proof of the upper bound in Theorem 1.1. We conclude this section with the proof of Theorem 1.2. Proof. The set C ( ) in which we work is de ned by C ( ) = f (mod q) : (1) = 1; jL(1; )j > e ; jS 11=21 ()j 1g; q y;q +c where we have set y = e for some constant c > 0. If the constant C in the statement of Theorem 1.2 is large enough, then Theorems 4.2 and 4.1 imply that (1.7) does hold. 20 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Assume now that 2 C ( ). Using partial summation on the P olya{Vinogradov inequal- ity, then that jS 11=21 ()j 1, and nally Lemma 3.2, we obtain y;q X X X (n) (n) (n) 1=43 L(1; ) = + O(q ) = + O(1) = + O(1): n n n (4.3) + + 11=21 P (n)y P (n)y nq 11=21 nq Given that jL(1; )j > e , we deduce that X X (n) 1 (4.4) e < jL(1; )j = + O(1) + O(1) e + O(1); n n + + P (n)y P (n)y by Mertens's estimate. Therefore 1 (p) 1 1 1 = 1 + O : p p py Taking logarithms, we nd that XX 1 Re( (p)) 1 jp py j=1 that is, is \1-pretentious". Since j1 uj 2Re(1 u) for u 2 U, the above inequality and the Cauchy{Schwarz inequality imply that XX j1 (p)j log log z (4.5) pz j=1 for all z 2 [10; y]. Moreover, since j1 (n)j j1 (p )j; p kn we nd that X X X j1 (n)j 1 j1 (p )j n n + + j P (n)z P (n)z p kn (4.6) XX X j1 (p )j 1 (log z) log log z p : p m pz j1 P (m)z Now, since is odd for 2 C ( ), P olya's Fourier expansion (1.9) implies that X X G() (n) cos(2n) (n) = L(1; ) + O(log q): (4.7) i n nq n=1 Set G() R () = (n) (L(1; ) + log 2): nq THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 21 2c If 1=y j 1=2j 1= and c is suciently large, then Lemma 3.4 and relation (4.7) imply that 0 1 X X G() cos(2n) log @ A R () = (n) L(1; ) + O i n nq P (n)y 0 1 X X G() (n) cos(2n) cos(2n) log @ A = + O : i n n n=1 P (n)y Now, Lemma 3.2 and relation (4.6) yield that X X (n) cos(2n) cos(2n) n n n=1 P (n)y X X (n) cos(2n) cos(2n) 1 = + O n n n=1 P (n)y nq X X (n) cos(2n) ((n) 1) cos(2n) (log ) = + + O p n n 2c nq <nq P (n)>y P (n)y c cos(2n) (log ) =: + O ; 2c <nq for some complex numbers c of modulus 2. Finally, if j 1=2j 1=y, then we simply note the trivial bound R () q , which follows by our assumption that S () 1 11=21 y;q for the we are working with. So (1.8) will certainly follow if we show that 1=2+1= c cos(2n) log d : 2 c n 1=21= 2c <nq By Cauchy{Schwarz, it suces to show that Z c c 1=2+1= 2 c cos(2n) (log ) (4.8) := d : 2 c n 1=21= 2c <nq For convenience, set B = , and note that 1=2+1=B m+n B (1) m + n m n cos(2m) cos(2n)d = f + f ; 2 2 B B 1=21=B where sin(2u) if u 6= 0; 2u f (u) := 1 otherwise: 22 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Therefore 0 1 X X X B C 1 1 B C jf (k=B)j + : @ A mn mn 2 2 k0 B <m;nq B <m;nq m+n=k mn=k Note that X X 1 1 2 m + n 1 2 log k k>2B k>2B mn k mn k m+n=k m+n=k 1m;nk1 B <m;nq and X X X X 1 1 1 1 1 mn n(n + k) k n n 2 2 2 mn=k n>B B <nk n>maxfk;B g B <m;nq 1 2 log k 1 k>2B + : k maxfk; B g Using the bound f (k) minf1; B=kg, we conclude that log B c 2 Since B = for c 2, (4.8) follows. This completes the proof of (1.8). Finally, note that relation (1.8) clearly implies that m() jL(1; ) + log 2j + O((log ) = ): Relations (4.3) and (4.6) with z = y imply that L(1; ) = e + O( log ) 2 2 1=2 2 2 If L(1; ) = a + ib then a=ja + ibj = (1 + b =a ) = 1 + O(b =a ) = 1 + O((log )= ). 2a log 2+(log 2) 1=2 a 2 Therefore jL(1; ) + log 2j=jL(1; )j = (1 + ) = 1 + log 2 + O( ), 2 2 2 a +b ja+ibj and so jL(1; ) + log 2j = jL(1; )j + log 2 + O(log )= ), which completes the proof of the theorem. 5. Truncating Polya's Fourier expansion In this section we show that for most we can limit the Fourier expansion to a sum over very smooth numbers without much loss, which is the content of Theorem 4.2. We prove this theorem by showing that high moments of S are small: y;z Theorem 5.1. Let q and k be integers with 3 k (log q)=(400 log log q). For k log k 11=21 y z q , we have that 2 1 O(k) 1 e k log y e 2k O(k log log y= log y) S () e + : y;z 19k (q) y (log y) (mod q) One consequence of Theorem 5.1 is the desired conclusion that S () is usually small: y;z THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 23 Deduction of Theorem 4.2 from Theorem 5.1. We may assume that y and q are large. Let be the proportion of characters (mod q) such that S () > e . Moreover, set 11=21 y;q y log q k = min ; ; log y 400 log log q where c is a constant to be determined. Then Theorem 5.1 implies that 2k 2 2k O(k) (e ) k log y e 2k O(k log log y= log y) S 11=21 () e + y;q 19k (q) ey (log y) (mod q) k+O(k log log y= log y) e ; which completes the proof. We prove Theorem 5.1 as an application of the following technical estimates. Proposition 5.2. Let q and k be two integers with 3 k (log q)=(400 log log q), and A fn 2 N : y < n z; P (n) > yg, where y and z are two positive real numbers such 3 11=21 that k y z q . Then 1 1 2k k=21 1=10 S () y + q : 40k (q) (log y) (mod q) Proposition 5.3. Let q and k be two integers with 3 k (log q)=(400 log log q), and A fn 2 N : y < n z; P (n) > yg, where y and z are two positive real numbers such log log k that k log k y z k . Then O(k log log y= log y) k O(k) 1 e k e 2k S () + : k 50k (q) (ey log y) (log y) (mod q) Before we proceed to the proof of these propositions, let us see how we can apply them to deduce Theorem 5.1. Deduction of Theorem 5.1 from Propositions 5.2 and 5.3. Set Y = maxfy; k g. Then X X X (n)e(n) (n)e(n) (n)e(n) = + n n n nz nz nz + + + P (n)>y y<P (n)Y P (n)>Y X X X X (a) (b)e(ab) (a) (b)e(ab) = + : a b a b az az 1<bz=a 1<bz=a + + P (a)y b2P (y;Y ) P (a)Y P (b)>Y where P (y; Y ) is the set of integers all of whose prime factors lie in (y; Y ]. We let X X (n)e(n) (n)e(n) (1) (2) S () = max and S () = max ; w w 2[0;1] n 2[0;1] n y<nw Y <nw n2P (y;Y ) P (n)>Y so that (1) (2) X X S () S () z=a z=a (5.1) S () + : y;z a a + + P (a)y P (a)Y 24 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS We shall bound the moments of each summand appearing above, individually. (1) We start with the summand involving S (). Here we may assume that y k (and (1) 3 0 log log k thus Y = k ), else P (y; Y ) = f1g and so S () = 0 for all w. Set w = minfw; k g and note that 0 1 B C 1 1 (1) (1) (1) B C S () = S () + O = S () + O 0 0 w w w @ A 100 n (log y) + 3 P (n)k log log k nk by Lemma 3.2. So Minkowski's inequality and Proposition 5.3 imply that 1 1 0 1 0 1 2k 2k X X 1 1 (1) (1) 2k 2k 100 @ A @ A S () S () + O(1=(log y) ) w w (q) (q) (mod q) (mod q) O(log log y= log y) 25 e + O(1=(log y) ) ey log y Hence, applying H older's inequality and Mertens' estimate 1=a = e log y + O(1), P (a)y we arrive at the estimate (5.2) 0 1 2k (1) (1) 2k 2k1 X X X X S () S () 1 (e log y + O(1)) z=a z=a @ A (q) a (q) a + + (mod q) P (a)y (mod q) P (a)y 2k e k log y O(k log log y= log y) 24 e + O(1=(log y) ) : ey (2) Next, in order to bound the summand in (5.1) involving S (), we observe that 1 1 (2) 2k S () z=a 40k (q) (log Y ) (mod q) for all m 1 by Proposition 5.2. Therefore H older's inequality implies that 0 1 2k (2) (2) 2k O(k) 2k1 X X X X S () S () 1 e (log Y ) z=a z=a @ A (q) a (q) a + + (mod q) P (a)Y P (a)Y (mod q) (5.3) O(k) O(k) e e 38k 38k (log Y ) (log y) Finally, relations (5.1), (5.2) and (5.3), together with an application of Minkowski's inequal- ity, imply that 2k X 2 1 e k log y 2k O(k log log y= log y) 19 (5.4) S () e + O(1=(log y) ) : y;z (q) ey (mod q) THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 25 We note that, for all ; > 0, we have that 2k 2k k 2k 2k k =(2k) (5.5) ( + ) ( + )(1 + ) ( + )e : p p p p p Indeed, if , then + (1 + ), whereas if > , then + (1 + ). Combining (5.4) and (5.5) completes the proof of Theorem 5.1. Our next task is to show Proposition 5.2. First, we demonstrate the following auxiliary lemma. Lemma 5.4. Let 2 (0; 1] and k 2 be an integer. Uniformly for (2 + )=(2 + 2) and 1+ for y k , we have that d (n) O(k= log k) e : P (n)>y Proof. Lemma 3.1 in [BG13], which is a generalization of Lemma 4 in [GS06], implies that r 2 d (p ) log = log I (2k=p ) + O(k=p ); 2r r=0 1+ where I is de ned by (1.4). Note that if p y k , then k=p < 1, which implies that 2 2 k 1 k 1 I (2k=p ) 1 + 1 + O : 2 2 2 p m! p m1 So we arrive at the estimate r 2 2 d (p ) k log ; 2r 2 p p r=0 which in turn yields that 0 1 2 2 2 X X d (n) k k k @ A log ; 2 2 21 n p y log y log k p>y P (n)>y 1+ by our assumptions that y k and (2 + )=(2 + 2) 3=4. Proof of Proposition 5.2. Without loss of generality, we may assume that q is large, else the result is trivially true. Set A(N ) = A\ (N=e; N ], so that S () S j (): A(e ) log y<jlog z+1 H older's inequality with p = 2k=(2k 1) and q = 2k implies that ! ! 2k1 X X 2k 4k 2k S () j S j () A A(e ) 4k 2k1 log y<jlog z+1 log y<jlog z+1 (5.6) 4k 2k j S () ; A(e ) 2k+1 (log y 1) log y<jlog z+1 26 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS which reduces the problem to bounding 2k S () A(N ) (q) (mod q) for N 2 [y; ez]. In order to do this, we rst decouple the , the point where the maximum S () occurs, from the character . We accomplish this by noticing that for every R 2 N A(N ) and for every 2 (0; 1], there is some r 2 f1; 2; : : : ; Rg such that j r=Rj 1=R. Then X X (n)e(n) (n)e(nr=R) N S () = = + O : A(N ) n n R n2A(N ) n2A(N ) 21=20 We choose R = N . Then Minkowski's inequality implies that 2k O(k) (n)e(nr=R) e 2k 2k1 S () 2 max + O A(N ) k=10 1rR n N n2A(N ) (5.7) 2k O(k) X X (n)e(nr=R) e 2k1 2 + O ; k=10 n N r=1 n2A(N ) which reduces Proposition 5.2 to bounding 2k X X 1 (n)e(nr=R) S := : N;r (q) n (mod q) n2A(N ) Notice that X X 1 d (n; N )(n) (5.8) S = ; N;r (q) n k k (mod q) (N=e) <nN P (n)>y where X Y d (n; N ) := e(n r=R): k j n n =n 1 k j=1 n ;:::;n 2A(N ) 1 k Clearly, jd (n; N )j d (n; N ) := 1: k k n n =n n ;:::;n 2A(N ) 1 k So opening the square in (5.8), and summing over (mod q), we nd that X X d (m; N ) d (n; N ) k k S : N;r m n (5.9) k k k k k k N =e <mN N =e <nN (m;q)=1; P (m)>y nm (mod q); P (n)>y THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 27 (1) (2) The right hand side of (5.9) is at most S + 2S , where N;r N;r X X X d (m; N ) d (m; N ) d (n; N ) k k k (1) (2) S := and S := : N;r N;r m m n k k k k k k k N =e <mN N =e <mN m<nN (m;q)=1 nm (mod q) P (m)>y P (n)>y P (m)>y We shall bound each of these sums in a dierent way. Firstly, note that k=2 2 O(k) e d (m) e (1) S ; (5.10) N;r k=2 3=2 k=2 N m N P (m)>y by Lemma 5.4 with = 1, which is admissible. (2) (2) Next, we bound S . Note that n m + q > q for n and m in the support of S . In N;r N;r (2) particular, for S to have any summands, we need that N > q. We choose j 2 f2; : : : ; kg N;r such that j1 j (5.11) N q < N : Then 2k X X (2) S d (m; N ) d (n; N ) k k N;r 2k k k k k k k N =e <mN N =e <nN (m;q)=1 nm (mod q) (5.12) 2k X X X = d (m; N ) d (g; N ) d (h; N ): k kj j 2k k kj j mN gN hN (m;q)=1 (g;q)=1 hgm (mod q) Our goal is to bound D(a) = d (h; N ); hN ha (mod q) for every a 2 f1; : : : ; qg that is coprime to q. First, assume that j > 1000. Note that D(a) is supported on integers h which can be written as a product h = n n with each of the 1 j factors n lying in the interval (N=e; N ] and having all their prime factors > y. In particular, (n ) log N= log y for all ` 2 f1; : : : ; jg and consequently j log N j log q log q (h) ; log y j 1 log y 0:999 log y by (5.11). In particular, log q log k (h) 0:334 0:999 log y 0:999 log y d (h; N ) j k = q q ; by our assumptions that y k and that j k. This inequality also holds when j 1000, j 1000 1000 since in this case d (h; N ) d (h) h for all > 0 and h N N q for the j j numbers h in the range of D(a). So, no matter what j is, we conclude that j j N N 0:334 D(a) q 1 ; 0:666 0:116 (5.13) q Rq hN ha (mod q) 28 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS 21=20 11=20 since R N q . Inserting (5.13) into (5.12), we arrive at the estimate 2k j 2k e N e (2) k kj (5.14) S N N = : N;r 2k 0:116 0:116 N Rq Rq Combining relations (5.10) and (5.14) with (5.9), we deduce that O(k) 2k e e S + : N;r k=2 0:116 N Rq Together with (5.7), the above estimate implies that 21=20 O(k) O(k) O(k) O(k) O(k) 1 N e e e e e 2k S () + + + ; A(N ) k=2 0:116 k=20 0:116 k=20 (q) N q N q N (mod q) since we have assumed that k 3. Together with (5.6), this implies that O(k) 4k 4k X X 1 e j j 2k S () + y;z 2k+1 jk=20 0:116 (q) (log y) e q log y<jlog z+1 (mod q) O(k) 2k1 O(k) 4k+1 e (log y) e (log z) + : k=20 0:116 2k+1 y q (log y) 11=21 Since z q and k (log q)=(400 log log q), Proposition 5.2 follows. Proof of Proposition 5.3. Without loss of generality, we may assume that y is large enough. We start in a similar way as in the proof of Proposition 5.2: we set R = bz c and note that (n)e(nr=R) S () = max + O(1=z ): 1rR n2A Therefore, Minkowski's inequality implies that 1 1 0 1 0 1 2k 2k 2k X X X 1 1 (n)e(nr=R) 2k 2 @ A @ A S () max + O(1=z ) 1rR (q) (q) n (mod q) (mod q) n2A 0 1 1 2k 2k X X X 1 (n)e(nr=R) @ A + O(1=z ): (q) n r=1 n2A (mod q) We claim that 2k O(k log log y= log y) k O(k) X X 1 (n)e(nr=R) e k e (5.15) S := + k 100k (q) n (ey log y) (log y) (mod q) n2A for all r 2 f1; : : : ; Rg. Proposition 5.3 follows immediately if we show this relation, since we would then have that 2k 1 k 2k 3 O(k log log y= log y) 50 S () z e + O(1=(log y) ) (q) ey log y (mod q) O(k log log y= log y) k O(k) e k e k 50k (ey log y) (log y) THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 29 log log k by our assumption that z k and (5.5), which establishes Proposition 5.3. Arguing as in the proof of Proposition 5.2 and setting 0 k d (n) = #f(n n ) 2 N : n = n n ; y < n z (1 j k)g; 1 k 1 k j we nd that 0 0 X X d (m) d (n) k k (1) (2) S S + 2 S ; r r (5.16) m n k k m>y n>y (m;q)=1; P (m)>y nm (mod q); P (n)>y where 0 0 0 X X X d (m) d (m) d (n) (1) k (2) k k S := and S := : r r m m n n>m P (m)>y (m;q)=1 nm (mod q) m2P (y;z) P (n)>y We shall bound each of these sums in a dierent way. (2) We start with bounding S . Set 1 log(y=k) log log k = min ; 2 log k log k 1+ for large enough k, so that y k . Also, let = (2 + )=(2 + 2). Fix m 2 N with (m; q) = 1, and note that if n m (mod q) with n > m, then n m + q > q. Therefore X X d (n) 1 (2) k S (m) := n n n 1 k n>m n ;:::;n 2(y;z] 1 k nm (mod q) n n >q P (n)>y P (n n )>y n n m (mod q) 1 k X X n n (5.17) 1 k r 1 r 1r ;:::;r log(z=y)+1 ` ` ye <n ye 1 k r ++r >log(q=y ) P (n )>y (1`k) 1 k ` n n m (mod q) 1 k X X 1: k r ++r 1 k y e r 1 r 1r ;:::;r log(z=y)+1 ` ` 1 k ye <n ye r ++r >log(q=y ) P (n )>y (1`k) 1 k ` n n n (mod q) 1 k We x r ; : : : ; r as above, set y = ye for all ` 2 f1; : : : ; kg, and choose j 2 f2; : : : ; kg such 1 k ` that y y q < y y : (5.18) 1 j1 1 j 30 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Then, for any a 2 N that is coprime to q, the Cauchy{Schwarz inequality implies that X X 1 d (n) y =e<n y ny y 1 j ` ` ` P (n )>y (1`j) P (n)>y n n a (mod q) na (mod q) 1 j 0 1 0 1 1=2 1=2 X X B C B C B C B C 1 d (n) @ A @ A ny y ny y 1 j 1 j na (mod q) P (n)>y 0 1 1=2 B C y y d (n) 1 j j B C (y y ) : 1 j @ 2 A q n ny y 1 j P (n)>y So, applying Lemma 5.4, we deduce that O(k) X 1+2 (y y ) e y y 1 j 1 j O(k) 1 e ; q q y =e<n y ` ` ` P (n )>y (1`j) n n a (mod q) 1 j O(k) since y ez e and y y q, by (5.18). Inserting this estimate into (5.17), we j 1 j1 deduce that O(k) k O(k) k e (log z) e (log z) (2) (5.19) S (m) : 1 =3 q q Consequently, we immediately deduce that (5.20) O(k) k O(k) 2k k O(k) 2k k e (log z) log z e (log log k) (log k) e (log log k) (log k) (2) S = =3 =3 (log log k)=(3 log k) q log y q q O(k) 100k (log y) which is admissible. (1) It remains to bound S . This will be done in a very dierent way. We observe that if (X ) are the random variables de ned in the introduction, then n n1 2 3 2k 6 n 7 (1) S E : 4 5 n2P (y;z) n>1 We have that ( ) X Y X X X X 1 n p p = 1 + 1 = 1 + exp + O n p p y log y y<pz y<pz n2P (y;z) n>1 = 1 + e + O(1=y); THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 31 where T := : y<pz Therefore h i 1 2k 2k (1) T 2k (S ) E e 1 + O(1=y); by Minkowski's inequality. Fix 2 [1= log y; 1]. When jTj , then je 1j e jTj, whereas T jTj jTj ` if jTj , then we use the trivial bound je 1j 2e 2e (jTj=) , for any ` 2 N. Therefore 1 1 (1) 2k 2kjTj ` 2k 2k 2k (S ) e E jTj + 2 E e (jTj=) + O(1=y): We have that O(k= log y) k X X 1 1 e k 2k E jTj = k! ; 2 k p p (p p ) (ey log y) 1 2k 1 k y<p ;:::;p z y<p ;:::;p z 1 2k 1 k p p =p p 1 k k+1 2k k 2 since k! (k=e) k and 1=p = 1=(y log y) + O(1=(y log y)) by the Prime Number p>y Theorem. Moreover, 1 1 m m X X (2k) (2k) 1=2 2kjTj ` ` m+` ` 2m+2` E e (jTj=) = E jTj E jTj m! m! m=0 m=0 1 m=2+`=2 m O(1) (2k) (m + `)! e m! y log y m=0 `=2 m=2 `=2 O(1) O(1) 2 O(1) e ` 1 e k e ` o(k) 2 2 y log y (m=2)! y log y y log y m=0 m+` O(m) 2 as (m + `)! m!`!2 and m! = (m=2)!e , with y k. Choosing ` = bc y log yc 0 2 O(k)4c y log y for an appropriate small constant c > 0 makes the left hand side e for some c > 0. We take = log log y= log y to conclude that 0 2 0 2 2kjTj ` O(k)4c y(log log y) = log y 2kc (log log y) E e (jTj=) e e ; where we used our assumption that k is large and y k log k. We thus conclude that 2k k 0 2 (1) O(k log log y= log y) c (log log y) S e + 2e ey log y O(k log log y= log y) k e k 0 2 O(k)c k(log log y) + e ; (ey log y) by (5.5). Together with (5.16) and (5.20), this completes the proof of relation (5.15) and, thus, of Proposition 5.3. 32 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS 6. The distribution function In this section, we prove Theorem 1.4. Throughout this section we x > 0 and a large 5=9 5=9 odd prime number q such that (log log q) 2 log log log q. Set y = expf(log log q) g and 1 (n)(1 e(n)) m () = max : 2e 2[0;1] n n2Z; n6=0 P (n)y Relation (4.2) and Lemma 3.2 imply that 1= log y jm() m ()j 2e S () + O q : 11=21 y;q Therefore, if q is large enough, then Theorem 4.2 and Lemma 3.2 imply that 1 y #f (mod q) : jm () m()j > 2= log yg exp = o (1) ( ); y q!1 q (q) 2(log y) 5=9 where the last relation follows by Theorem 1.1 and our assumption that (log log q) 2 log log log q. Therefore we conclude that (1 + o (1)) ( + 2= log y; y) ( ) (1 + o (1)) ( 2= log y; y); q!1 q q q!1 q where (t; y) := #f (mod q) : m () > tg: (6.1) q y (q) We perform the same analysis on ( ). With a slight abuse of notation, we set 1 X (1 e(n)) m (X ) = max 2[0;1] 2e n n2Z; n6=0 P (n)y The analogue of Theorem 4.2 for the random variables X is (more easily) proven by the same method with a few simple changes: we proceed exactly as in the proof of Theorem 5.1 but wherever we evaluated a sum (1=(q)) (h=k) there (as 0, unless h k (mod q) (mod q) when it equals 1), we now evaluate an expectation E[X X ] which equals 0 unless h = k, h k when it equals 1. (In fact this only happens in the proofs of (5.9) and of (5.16).) We then conclude that ( + 2= log y; y) + o ( ( )) ( ) ( 2= log y; y) + o ( ( )); (6.2) q!1 q q!1 q where (t; y) := Prob(m (X ) > t): (We could have written o (( )) in place of o ( ( )) in (6.2) but this would have q!1 q!1 q required proving a lower bound for ( ). In order to avoid this technical issue, we use for comparison ( ) whose size we already know by Theorem 1.1.) Next, for each p y we x a parameter of the form = 2=k with k 2 N to be chosen, and a partition of the unit p p p p circle into arcs fI ; : : : ; I g of length . Moreover, we let w be the point in the middle p;1 p;k p p;j of the arc I and we de ne Z to be the set of ((y) + 1)-tuples z := (z ; z ; z ; z ; : : : ) with p;j 1 2 3 5 THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 33 z 2 f1; 1g and z 2 fw : 1 j k g for all primes p y. Given such a choice of z 1 p p;j p and n 2 N, we set z = z and z = z z . Moreover, similarly to before, we let n e n 1 n p kn 1 z (1 e(n)) m (z) = max : 01 2e n n2Z; n6=0 P (n)y We will show that there exist choices of and a constant C > 0 such that if X = z and p 1 1 X belongs to the arc I centred at z for all p y, then jm (X )m (z)j C= log y. This p p;j p y y immediately implies that ( + C= log y; y) (y; ) ( C= log y; y); (6.3) where (t; y) := #fz 2 Z : m (z) > tg: jZj It remains to show thatjm (X )m (z)j C= log y if is chosen appropriately. We choose y y p 3 3 these parameters so that maxf(log p)=2; 1g=(log y) maxflog p; 2g=(log y) . Then the condition that jz X j for each prime p y implies that jz X j (logjnj)=(log y) p p p n n for all n 2 Znf0g with P (n) y. Hence X X X 1 (log n) (log p) 1 1 jm (X ) m (z)j = ; y y 3 e 3 n (log y) p (log y) m log y n1 py m1 + + e1 P (n)y P (m)y which proves our claim. We now prove similar inequalities for , namely that (1 + o (1)) ( + C= log y; y) ( ; y) (1 + o (1)) ( C= log y; y): (6.4) q!1 q q!1 5=3 Theorem 9.3 of [Lam08] states that if I is an arc of length 1=(log log q) for each p p p y, then #f (mod q) : (p) 2 I for each p yg (q ! 1): p p (q) py The same methods can easily be adapted to also show that 1 1 #f (mod q) : (1) = ; (p) 2 I for each p yg (q ! 1) p p (q) 2 py 5=9 5=3 for 2 f1; 1g. Since y = expf(log log q) g, the inequality 1=(log log q) is indeed satis ed for our choice of , thus completing the proof of (6.4). Finally, combining relations (6.1), (6.2), (6.3) and (6.4), we obtain ( + (2C + 2)= log y) (1 + o (1)) ( ) ( (2C + 2)= log y): q!1 q If is continuous in [a; b], then it is also uniformly continuous. It then follows immediately from the above estimate that ! as q ! 1 over primes, uniformly on [a; b]. In order to see that converges also weakly to , consider a continuous function f : R ! R of bounded support. Fix > 0. Since has at most countable many discontinuity points, we deduce that there is an open set E of Lebesgue measure < that contains all discontinuities of . If I is a bounded closed interval containing the support of f , then the set I n E is 34 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS compact and thus it can be written as a nite union of closed intervals. Since ( ) q q prime converges uniformly to on each such interval, it does so on E n I as well. Therefore Z Z lim sup f ( ) ( )d f ( )( )d q!1 R R q prime 0 1 @ A kfk 2 meas(E) + meas(I n E) lim sup sup j ( ) ( )j 1 q q!1 2InE q prime < 2kfk : Since was arbitrary, we conclude that Z Z lim f ( ) ( )d = f ( )( )d; q!1 R R q prime thus completing the proof of Theorem 1.4. 7. Some pretentious results In this section, we develop some general tools we will use to prove Theorems 2.1 and 2.3. We begin by stating a result that allows us to concentrate on the case when is a rational number with a relatively small denominator. 5 5 Lemma 7.1. Let y 2, z (log y) , be a Dirichlet character, 2 R and B 2 [(log y) ; z]. Let a=b be a reduced fraction with b B and j a=bj 1=(bB). Then X X (n)e(n) (n)e(na=b) = + O(log B); n n 1jnjz 1jnjN + + P (n)y P (n)y where N = minfz;jb aj g. Proof. This follows immediately by the second part of Lemma 4.1 in [Gol12] (see also Lemma 6.2 in [GS07]). When b is large, we have the following result. Lemma 7.2. Let j a=bj 1=b , where (a; b) = 1. For all z; y 3, we have that 5=2 (n)e(n) (log b) log b + log log y + log y: nz P (n)y Proof. This is Corollary 2.2 in [Gol12], which is based on a result due to Montgomery and Vaughan [MV77]. For smaller b, we shall use the following formula. Lemma 7.3. Let be a Dirichlet character and (a; b) = 1. For z; y 1 we have that X X X X (n)e(an=b) 2 (b=d)d (n) (n) = (a)G( ) : n b (d) n 1jnjz djb (mod d) nzd=b + + P (n)y P (n)y odd THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 35 Proof. If (c; d) = 1, then e(c=d) = (c)G( ): (d) (mod d) So, writing b=d for the greatest common divisor of n and b, we nd that X X X (n)e(na=b) (b=d)d (m)e(am=d) n b m 1jnjz djb 1jmjzd=b + + P (n)y P (n)y; (m;d)=1 X X X (b=d)d (m) (m) = (a)G( ) : b(d) m djb (mod d) 1jmjzd=b P (n)y Finally, observe that the innermost sum vanishes if is an even character, whereas if is odd it equals (m) (m) 2 : mz=d P (m)y This concludes the proof of the lemma. Following Granville and Soundararajan [GS07], we are going to show that the all the terms in the right hand side in Lemma 7.3 are small unless is induced by some xed character which depends at most on and y. As in [GS07], in order to accomplish this, we de ne a certain kind of \distance" between two multiplicative functions f and g of modulus 1: 1 Re(f (p)g(p)) D(f; g; y) := : py Then we let = (; y) be a primitive character of conductor D = D(; y) log y such that D(; ; y) = min D(; ; y): dlog y (7.1) (mod d) primitive We need a preliminary result on certain sums of multiplicative functions, which is the second part of Lemma 4.3 in [GS07]. Lemma 7.4. Let f : N ! U be a multiplicative function. For z; y 1 we have that f (n) (log y) expfD (f; 1; y)=2g: nz P (n)y Then we have the following \repulsion" result. Lemma 7.5. Let , y and be as above. If is a Dirichlet character modulo d y, of conductor log y, that is not induced by , then X p (n) (n) 1=2+ 2=4+o(1) (log y) (y ! 1): nz P (n)y 36 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Proof. Let (mod d ) be the primitive character inducing . Since 1 1 2 2 2 D (; ; y) D (; ; y) D (; ; y) O(log log log d); 1 1 pjd Lemma 3.4 in [GS07] and the de nition of imply that D (; ; y) 1 + o(1) log log y (y ! 1): The claimed estimate then follows by Lemma 7.4 above. When applying Lemma 7.3, we will need to evaluate the Gauss sum that arises. In order to do this, we shall use the following classical result (see, for example, Theorem 9.10 in [MV07, p. 289]). Lemma 7.6. Let be a character modulo d induced by the primitive character modulo d . Then G( ) = (d=d ) (d=d )G( ): 1 1 1 1 We also need the following simple estimate, which we state below for easy reference. Lemma 7.7. Let f : N ! U be a completely multiplicative function. For all a 2 N, we have that 0 1 X Y X X f (n) f (p) f (n) a log p @ A = 1 + O n p n (a) p nz nz pja pja (n;a)=1 and 0 1 X Y X X f (n) f (p) f (n) a log p @ A = 1 + O : n p n (a) p nz pja nz pja (n;a)=1 Proof. We write dja if pja for all primes pjd. Then 0 1 X X X X X X f (n) f (d) f (m) f (d) f (m) log d @ A = = + O n d m d m d 1 1 1 nz mz dja mz=d dja dja (m;a)=1 (m;a)=1 0 1 Y X X f (p) f (m) a log p @ A = 1 + O : p m (a) p pja mz pja (m;a)=1 The second part is proved similarly, starting from the identity X X X f (n) (d)f (d) f (m) = : n d m nz dja mz=d (n;a)=1 Combining the above results, we prove the following simpli ed version of Lemma 7.3. THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 37 Lemma 7.8. Let , y, and D be as above, and consider a real number z 1 and a reduced 1=100 fraction a=b with 1 b (log y) and (b; q) = 1. If either D - b or is even, then (n)e(na=b) 0:86 (log y) ; 1jnjz P (n)y whereas if Djb and is odd, then (n)e(na=b) 2e 2 !(b=D) D (;;y)=2+O(1) 0:86 (log y) min p (2=3) ; e + O((log y) ): 1jnjz P (n)y Proof. By Lemma 7.5, we see that if is a character modulo d that is not induced by , then X p (m)(m) 1=2+ 2=4+o(1) 0:854 (log y) (log y) (x 1): (7.2) mx P (n)y So the rst result follows by Lemma 7.3. Finally, if Djb and is odd, then Lemmas 7.3 and 7.6, and (7.2) imply that X X X (n)e(na=b) 2(a)G() (b=d)d (d=D)(d=D) (n)(n) n b (d) n 1jnjz djb ndz=b; (n;d)=1 + + d0 (mod D) P (n)y P (n)y 0:86 + O((log y) ): Writing d = Dc, noticing that (c; D) = 1 if (c) 6= 0, and using the trivial bound 1=n log(b=d) to extend the sum over n dz=b to a sum over n z, we nd dz=b<nz that X X X (n)e(na=b) 2(a)G() D (b=(Dc))c (c)(c) (n)(n) n b (D) (c) n 1jnjz cjb=D nz; P (n)y (c;D)=1 P (n)y (n;cD)=1 0:86 + O((log y) ): Setting m = b=D and applying Lemma 7.7 with f (n) = (n)(n)1 1 + and m in (n;cD)=1 P (n)y place of a we deduce that X X Y (n)e(na=b) 2(a)G() D (m=c)c (c)(c) (p)(p) = 1 n b (D) (c) p 1jnjz cjm pjm; p-cD (c;D)=1 P (n)y (n)(n) 0:86 + O((log y) ): nz P (n)y (n;mD)=1 38 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Since we have assumed that (b; q) = 1, we have that X Y (m=c)c (c)(c) (p)(p) (c) p cjm pjm; p-cD (c;D)=1 (p)(p) (p)(p) (p)(p) (7.3) = (m) 1 1 1 1 1=p p p pjm; p-D m(D) 1 (p)(p) = (m) : (mD) 1 (p)(p)=p pjm; p-D If z 2 C with jzj = 1, then 1 z 2 1 z=p 1 + 1=p Therefore the absolute value of the sum in (7.3) is m(D) 2 (mD) 1 + 1=p pjm Since we also have that mD = b, we deduce that X X Y (n)e(na=b) 2 b (n)(n) 2 0:86 p + O((log y) ) n (b) n 1 + 1=p Dm 1jnjz P (n)y pjm P (n)y (n;b)=1 ( ) 3=2 2 b 2 D (;;y)=2+O(1) p min e log y + O(1); (log y)e (b) Dm 0:86 + O((log y) ) 1 + 1=p pjm 2e 2 !(m) D (;;y)=2+O(1) 0:86 (log y) min (2=3) ; e + O((log y) ); 1 2 where we used Lemma 7.4 and the inequality 2=5 2=3 for p 2. This p 1+1=p completes the proof of the lemma. Finally, if pretends to be 1 in a strong way, then we can get a very precise estimate on the sum of (n)e(an=b)=n over smooth number using estimates for such numbers in arithmetic progressions due to Fouvry and Tenenbaum [FT91]. THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 39 Lemma 7.9. Let z; y 2, be a character mod q, and a=b be a reduced fraction with 1 b (log y) . Then X X X (n)e(na=b) (c)(d) (n) = + O(E) n (c)d n nz b=cd nz; (n;c)=1 P (n)y P (n)y Y X (q ) (b ) (p) (n) 1 1 = (1 (p)) 1 + O(E); (q ) (b ) p n 1 1 pjb nz; (n;b)=1 P (n)y where b is the largest divisor of b with (b ; q) = 1 and b = b q , with 1 1 1 1 b j1 (p)j E = 1 + (e 1) log log y where = D(; 1; y) log log y: (b) p 1 py Proof. First, note that the inequality D(; 1; y) log log y follows by the Cauchy{ Schwarz inequality and the fact that j1 zj 2Re(1 z) for z 2 U. We write = h 1. j j j1 Then we have that jh(p )j = j(p ) (p )j j1 (p)j 2 and thus jh(m)j e (log y) : (7.4) P (m)y log log y Moreover, observe that, using Lemma 3.2, we may assume that z y . We begin by estimating the sum S(t) := (n)e(na=b) nt P (n)y 400 200 for t 2 [(log y) ; z]. Set ` = (log y) and note that X X S(t) = (d) (m)e(ma=c) b=cd mt=d P (m)y; (m;c)=1 X X X = (d) h(k) e(k`a=c) b=cd kt=d `t=(dk) + + P (k)y; (k;c)=1 P (`)y; (`;c)=1 X X X X (7.5) = (d) h(k) e(kja=c) 1 b=cd 1jc kt=(d` ) `t=(dk) + + (j;c)=1 P (k)y; (k;c)=1 P (`)y; `j (mod c) 0 1 B C tb jh(k)j B C + O @ A (b) k t=(log y) <kt P (k)y where we bounded trivially the sum over ` when k > t=(d` ) (note that d` (log y) for 0 0 all djb). 40 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Next, we need an estimate for the sum `t=(dk) P (`)y; `j (mod c) when dk t=` . Note that, since b (log y) , Theorems 2 and 5 in [FT91] imply that X X 1 t 1 = 1 + O (c) dk(c)(log y) `t=(dk) `t=(dk) + + P (`)y; `j (mod c) P (`)y; (`;c)=1 when t=(dk) y; the same result also holds when t=(dk) y by elementary techniques since t=(dkc) ` =c (log y) for cjb and dk t=` . So, using the identity 0 0 e(kja=c) = (c); 1jc (j;c)=1 we deduce that X X X (c)(d) S(t) = h(k) 1 (c) b=cd kt=(d` ) `x=(dk) + + P (k)y; (k;c)=1 P (`)y; (`;c)=1 0 1 B C t tb jh(k)j B C + O + ; @ 2 A (log y) (b) k t=(log y) <kt P (k)y where we used (7.4). We get the same right side no matter what the value of a, as long as (a; b) = 1. Hence X X S(t) = (n)e(nr=b) + R(t) (b) 1rb nt (r;b)=1 P (n)y for some function R(t) satisfying the bound t tb jh(k)j R(t) + : (log y) (b) k t=(log y) <kt P (k)y Letting d = (n; b), and writing n = md and b = cd so that (m; c) = 1, we deduce that X X (c)(d) S(t) = (m) + R(t) (c) b=cd mx=d P (m)y; (m;c)=1 THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 41 by [Dav00, eq (7), p. 149], for all t 2 [(log y) ; z]. Therefore partial summation implies that X X (n)e(na=b) (n)e(na=b) = + O(log log y) n n nz (log y) <nz P (n)y P (n)y X X (c)(d) (m) (c)d m b=cd (log y) =d<mz=d; (m;c)=1 P (m)y dR(t) + + O(log log y): (log y) log log y Integrating by parts and applying relation (7.4) and our assumption that z y we conclude that dR(t) b log log y jh(k)j b log log y 1 + 1 + (e 1): 400 t (b) k (b) (log y) k>1 P (k)y Since we also have that X X X (c)(d) (m) log log y log log y; (c)d m (c)d b=cd b=cd m2[1;(log y) =d][(z=d;z] (m;c)=1; P (m)y we deduce that X X X (n)e(na=b) (c)(d) (m) = + O(E): (7.6) n (c)d m nz b=cd mz; (m;c)=1 P (n)y P (m)y Applying Lemma 7.7 with f (n) = (n)1 1 + and b in place of a implies that (n;c)=1 P (n)y 0 1 X Y X X (m) (p) (n) b log p @ A = 1 + O : m p n (b) p mz; (m;c)=1 pjb; p-c nz; (n;b)=1 pjb + + P (m)y P (n)y Inserting this formula into (7.6) leads to an error term of size X X 1 b log p E + E; (c)d (b) p b=cd pjb and a main term of Y X (p) (n) p n pjb nz; (n;b)=1 P (n)y times X Y Y (c)(d) (p) (q ) (b ) 1 1 1 = (1 (p)); (c)d p (q ) (b ) 1 1 b=cd pjc pjb thus completing the proof of the lemma. 42 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Corollary 7.10. Let q be an integer that either equals 1 or is prime. Let z; y 2, and a=b be a reduced fraction with 1 < b (log y) . Then X X e(na=b) 1 1 1 b b=q q>1 = + O 1 + log log y : n (q) n q (b) nz nz + + P (n)y P (n)y (n;q)=1 (n;q)=1 8. Structure of even characters with large M (): proof of Theorem 2.3 3 +c The goal of this section is to prove Theorem 2.3. Throughout this section, we set y = e for some constant c. We will show this theorem with C ( ) := (mod q) : 6= ; (1) = 1; S () 1; m() > ; 11=21 q y;q where the quantity S () is de ned as in Section 4. Theorem 4.2 and the lower bound in y;z Theorem 1.3, which we already proved in the beginning of Section 4 (independently of the proof of Theorem 2.3), guarantee that the cardinality of C ( ) satis es (2.5), provided that the constant c in the de nition of y and the constant C in the statement of Theorem 2.3 are large enough. We x a large log log q and we consider a character 2 C ( ). Let = N =q. Then X X 1 (n)e(n) 1 (n)e(n) 1=43 m() = + O(q ) = + O(1) 2e n 2e n 11=21 11=21 1jnjq 1jnjq P (n)y 1 (n)e(n) = + O(1); 2e n n2Znf0g P (n)y by (1.9), our assumption that jS 11=21 ()j 1 for 2 C ( ), and Lemma 3.2. As in the y;q statement of Theorem 2.3, we approximate by a reduced fraction a=b with b and 10 10 j a=bj 1=(b ). We let N = 1=jb aj and apply Lemma 7.1 with z = 1 to nd that 1 (n)e(an=b) m() = + O(log ): (8.1) 2e n 1jnjN P (n)y Since 2 C ( ), we must have that m() > , which implies that (n)e(an=b) 2e O(log ): (8.2) 1jnjN P (n)y We now proceed to show that satis es properties (1) and (2) of Theorem 2.3. Proof of Theorem 2.3 - Property (1). We choose (mod D) with D log y to satisfy (7.1). We claim that (8.3) b = D = 3 and = ; the rst claim being equivalent to a=b 2 f1=3; 2=3g. THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 43 Firstly, note that Lemma 7.2 in conjunction with (8.2) implies that b 1. Equation (8.2) also tells us that we must be in the second case of Lemma 7.8 (provided that is large enough), that is to say Djb and is odd. Since is even, we conclude that is odd. Thus 3 D b 1. Moreover, the last inequality in Lemma 7.8 implies that (n)e(an=b) 2e 3 !(b=D) 0:86 sum p (2=3) + O( ): 1jnjN P (n)y Comparing this inequality with (8.2), we deduce that b = D = 3 and thus = (=3), the quadratic character modulo 3, which completes the proof of our claim and hence of the fact that Property (1) holds. Proof of Theorem 2.3 - Property (2). We start with the proof of (2.7). Note that X X (n)e(an=3) (n)(e(an=3) e(an=3)) = ; n n 1jnjN nN P (n)y P (n)y where a 2 f1; 2g. Then using (4.1), we nd that X X (n)e(an=3) a (n) = i 3 ; n 3 n jnjN nN P (n)y P (n)y so that, by (8.1), (n) m() = + O(log ): 2e n nN P (n)y Hence, we conclude that p X p X (n) 2e O(log ) 3 3 n n nN nN; (n;3)=1 P (n)y P (n)y 2 1 = 3 e log y + O(1) 3 n n>N; (n;3)=1 P (n)y = 2e + O(1) 3 : n>N; (n;3)=1 P (n)y Therefore log ; n>N; (n;3)=1 P (n)y which, in turn, implies that n n X X (n) (n) 3 3 = + O(log ); n n nN P (n)y P (n)y 44 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS so that (n) (8.4) m() = + O(log ): 2e n P (n)y Finally, we have that (n) 1=43 L 1; = + O(q ) 3 n 11=21 nq X X (n) 1 (n)(e(n=3) e(n=3)) 3 1=43 = + p + O(q ) n n i 3 11=21 11=21 nq nq + + P (n)y P (n)>y (n) = + O(log ); P (n)y by the P olya{Vinogradov inequality, our assumption that 2 C ( ) and Lemma 3.2. In- serting the above estimates into (8.4) completes the proof of (2.7). Finally, we prove (2.6). For convenience, we set (n) = (n) . Then (8.4) and our e + assumption that M () > q for 2 C ( ) imply that X X (n) 2 1 e log y + O(log ) = + O(log ): n 3 n + + P (n)y P (n)y (n;3)=1 Since this lower bound is also an upper bound, we deduce that (p) 1 log 1 1 = 1 + O : p p py p6=3 Then, the argument leading to (4.5) implies that XX j1 (p)j log p ; (8.5) py j=1 p6=3 thus completing the proof of slightly more than Property (2). Proof of Theorem 2.3 - Property (3). De ne w via the relation j k=`j = 1=(`y ). Note 3 +c . So we may show the theorem with w in place of u. that w = u(1 + O(1= )) as y = e THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 45 Arguing as at the beginning of Section 8, and applying Lemma 7.1 with z = 1, we nd that X X 1 (n)e( n) (n) = + O(1) G() 2i n n q n2Z; n6=0 P (n)y 1 (n)e(kn=`) = + O(log ) 2i n 1jnjy P (n)y (n) sin(2kn=`) = + O(log ): ny P (n)y We note that inequality (8.5) and the argument leading to (4.6) imply that j1 (n)j log ; (8.6) n P (n)y (n;3)=1 n j where we have set (n) = (n) . We write n = 3 m with (m; 3) = 1, so that j j X X X (3 ) (m) sin(23 km=`) (n) = + O(log ) G() 3 m j=0 n q P (m)y w j my =3 (m;3)=1 j j X X sin(23 km=`) p (3 ) = + O( log ); 3 m j=0 P (m)y my by (8.6) and as 1=m j. Using the formula 2 sin(2m=3) = 3 , we w j w y =3 <my deduce that j j X X X 1 (3 ) 2 sin(2m=3) sin(23 km=`) (n) = p + O( log ) G() 3 m n q j=0 P (m)y my j+1 j+1 3 k` 3 k+` X X 1 (3 ) cos(2m ) cos(2m ) 3` 3` = p + O( log ): 3 m j=0 P (m)y my j+1 3 k` If ` is not a power of 3, then 2= Z for each j, so Corollary 7.10 implies that 3` (n) log G() n q 46 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS j+1 3 k` as claimed. Finally, if ` = 3 and 2 f1; 1g, then 2= Z, unless j = v 1 and 3` k (mod 3), so that = . Therefore Corollary 7.10 and Lemma 3.3 imply that v1 X X (3 ) 1 (n) = p + O( log ) v1 G() 3 m n q P (m)y my k v1 (3 ) p = e P (w) + O( log ): v1 This completes the proof of Theorem 2.3. 9. The structure of characters with large M (): proof of Theorems 2.1 and 2.2 The goal of this section is to prove Theorems 2.1 and 2.2. Throughout this section, we set +c y = e for some constant c > 0 (note that this a dierent value of y than in the previous section). We will show this theorem with C ( ) := (mod q) : 6= ; S 11=21 () 1; m() > ; q 0 y;q where the quantity S () is de ned as in Section 4. Theorem 4.2 and the lower bound y;z in Theorem 1.1, which we already proved in the beginning of Section 4, guarantee that the cardinality of C ( ) satis es (2.1), provided that the constant c in the de nition of y and the constant C in the statement of Theorem 2.1 are large enough. We x a large log log q and we consider a character 2 C ( ). Let 2 [0; 1) be such that M () = (n) : nq As in the statement of Theorem 2.1, we pick a reduced fraction a=b such that 1 b and ja=bj 1=(b ), and we de ne b to equal b if b is prime, and 1 otherwise. Following the argument leading to (8.1), we deduce that X X 1 (n) (n)e(an=b) m() = + O(log ); (9.1) 2e n n n2Znf0g 1jnjN + + P (n)y P (n)y where N = 1=jb aj . Proof of Theorem 2.1 - Property (1). We choose (mod D) with D log y to satisfy (7.1). Since 2 C ( ), we must have that m() > , which implies that X X (n) (n)e(an=b) 2e O(log ): (9.2) n n n2Znf0g 1jnjN + + P (n)y P (n)y We claim that is odd; D = 1; = 1; and D(; 1; y) 1; (9.3) the rst relation being Property (1). We separate two cases. THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 47 1=100 First, assume that b . Then we apply Lemma 7.2 to nd that (n)e(an=b) 11=300 1jnjN P (n)y which, together with (9.2), implies that (n) 299=300 (9.4) j1 (1)j 2e O( ): P (n)y Then we must have that (1) = 1. Furthermore, the rst part of Lemma 7.8 implies that D = 1 and thus = 1. Finally, Lemma 7.4 and (9.4) imply that D(; 1; y) 1, which completes the proof of (9.3) in this case. 1=100 Finally, assume that b . Suppose that either is even or 6= 1. Then Lemma 7.8 implies that (n) 0:86 n2Znf0g P (n)y So (9.2) becomes (n)e(an=b) 0:86 2e O( ): (9.5) 1jnjN P (n)y Thus we must be in the second case of Lemma 7.8 as far as the above sum is concerned, that is to say that Djb and is odd. Then the second part of Lemma 7.8 implies that (n)e(an=b) 2e !(b=D) 0:86 (2=3) + O( ): 1jnjN P (n)y Comparing this inequality with (9.5), we deduce that b = D = 1 and thus = 1, provided that is large enough. But then has to be odd, which contradicts our initial assumption. So we conclude that our initial assumption must be wrong, that is to say must be odd and = 1, so that D = 1. 1=100 It remains to show that D(; 1; y) 1 in the case when b . We apply Lemma 7.4 and the second part of Lemma 7.8 to deduce that X X (n) (n)e(an=b) 2 0:86 expfD (; 1; y)=2g + : n n 1jnjN 1jnjz + + P (n)y P (n)y Combining the above inequality with (9.2) yields the estimate D(; 1; y) 1, thus complet- ing the proof of our claim (9.3) and, consequently, of Property (1). 48 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS In order to prove Property (2) in Theorem 2.1, we need an intermediate result: we set (n) L () = n1; (n;d)=1 P (n)y for d 2 N, and X p p j1 (p)j = D(; 1; y) log log p 1 py as in Lemma 7.9, where we used (9.3). Additionally, we set o(1) E = 1 + (e 1) log = ( ! 1); (b) (1) (2) and we write L () = L () + L (), where d d X X (n) (n) (1) (2) L () = and L () = : d d n n nN; (n;d)=1 n>N; (n;d)=1 + + P (n)y P (n)y The intermediate result we need to show is that >jL ()j + O(E) if b is not a prime power; jL ()j + O E if b = p ; e 2; (9.6) m() = e > p > b : jL ()j + O E if b is prime: (b) Before proving this, we show how to use it to complete the proof of Theorem 2.1. Proof of Theorem 2.1 - Property (2). We argue as in the proof of Property (2) in Theorem 2.3. For (2.3), note that our assumption that 2 C ( ) implies, as S () 1, that, for 11=21 y;q all d , X X X (n) (g)(g) (m) n g m + + gjd P (n)>y P (mg)>y 11=21 (n;d)=1 mq =g 11=21 nq 0 1 X X X (g)(g) (m) (g) log g @ A = + O g m g gjd gjd P (m)>y 11=21 mq 0 1 X X (g) log(2g) d log p @ A 1 + (log log ) : g (d) p gjd pjd THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 49 Therefore, X X X (n) (n) (n) 1=43 2 = + O(q ) = + O((log log ) ) n n n (n;d)=1 (n;d)=1 P (n)y 11=21 nq (n;d)=1 11=21 nq = L () + O((log log ) ) for all d , by the P olya{Vinogradov inequality, and Lemma 3.2. So (2.3) follows from (9.6) but with the weaker error term O(E ) in place of O( log ), where E = E if 1 1 e 1=2+o(1) b = p , and E = E otherwise, so that E . We argue much like we did getting to 1 1 to (4.5): We have 1 1 Y Y b (p) 1 m() = 1 + O (E ) e 1 ; e (b ) p p p6=b py py so that (p) 1 1 1 = 1 + O(E = ) p p p6=b py and therefore 1 Re((p)) E py p6=b This implies that 1=2 j1 (p)j E log p 1 py p6=b In particular, = j1 (p)j=(p 1) 1, so that we always have that E log py (b) and E log . This completes the proof of Property (2) in Theorem 2.1. Proof of (9.6). We separate two main cases. Case 1. Assume that b = 1. Then (9.1) and the fact that is odd imply that X X (n) (n) m() = e + O(log ) n n P (n)y nN (9.7) P (n)y (2) = e jL ()j + O(log ): Since m() > , we nd that X X 1 1 (2) e + O(log ) jL ()j e + O(log ): n n n>N nN + + P (n)y P (n)y 50 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS (2) Consequently, 1=n log , which in turn gives us that L () = L () + + 1 nN; P (n)y 1 O(log ). Inserting this estimate into (9.7), we deduce that m() = e jL ()j + O(log ); that is to say, (9.6) holds (with a stronger error term). Case 2. Assume that 1 < b . Then (9.1) and Lemma 7.9 implies that X X (n) (n)e(na=b) m() = e + O(log ) n n nz nN + + P (n)y P (n)y (b) (p) (1) = e L () L () (1 (p)) 1 + O(E): (b) p pjb Now Lemma 7.7, applied with f (n) = (n)1 + and b in place of a, implies that P (n)y Y X (p) b log p (1) (1) L () = L () 1 + O : (9.8) 1 b p (b) p pjb pjb So, if we set Y Y (b) (p) C = 1 (1 (p)) 1 ; (9.9) (b) p pjb pjb then we have that (2) (1) m() = e L () + C L () + O(E) (9.10) (b) (1) = e L () L () (1 (p)) + O(E): (b) pjb Our goal is to show, using formula (9.10), that (c) 1 for all cjb. Indeed, if this were true, then C b=(b) (1 (p)=p) . We start by observing that pjb X Y (d) (c) (p) C = 1 1 : (9.11) d (c) p b=cd pjb; p-c Indeed, reversing the last steps of the proof of Lemma 7.9, we nd that 1 1 Y X Y (b) (p) (d) (c) (p) 1 (1 (p)) = 1 : (b) p d (c) p pjb b=cd pjb; p-c Moreover, Y X X X X X (p) (n) (n) (d) (m) 1 = = = p n n d m b=cd pjb pjn ) pjb djb pjn ) pjb pjm ) pjb (n;b)=d (m;c)=1 X Y (d) (p) = 1 : d p b=cd pjb; p-c THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 51 Combining the two above relations, we obtain (9.11). Using (9.11) and the inequality j1 (p)=pj = p=jp (p)j , we deduce that p1 1 (c) b=(b) b jC j 1 = (9.12) b d (c) c=(c) (b) b=cd for b > 1. Inserting this into (9.10), and since m() > = e + O(1); P (n)y we obtain that (2) (1) jL ()j +jC jjL ()j + O(E) 1 b P (n)y (9.13) (2) (1) jL ()j + jL ()j + O(E): 1 b (b) We also have 0 1 X X X 1 (b) 1 b log p (1) @ A jL ()j = + O n b m (b) p nN; (n;b)=1 mN pjb P (m)y P (n)y by Lemma 7.7. Together with (9.13), this implies that (2) (2) jL ()j + O(E) =: S + O(E): N<nz; P (n)y Since this holds as an upper bound too, we deduce that (2) (2) (9.14) jL ()j = S + O(E): Substituting (9.14) into (9.13) we obtain b 1 (1) (1) jL ()j = + O(E) =: S + O(E) (9.15) (b) n nN P (n)y and then comparing (9.13) with the displayed line above, (b) (9.16) jC j = 1 + O( ); b 1 (j) where we have set := E=S for j 2 f1; 2g. Case 2a. Assume that b has at least two distinct prime factors. If b 6= 6, then we can nd e f 0 0 e f 0 !(b ) two distinct primes p and q such that b = p q b with p; q - b , p q 6= 6, and (b ) 2 (which only fails for b = 2 or 6). Therefore, taking absolute values in (9.9), we nd that 1 1 (b) j1 (p)jj1 (q)j 1 (p) 1 (q) jC j 1 + 1 1 1 1 : e f b (p ) (q ) p p q q 52 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Note that 1 (`) Re(1 (` )) Re(1 (`)) (9.17) 1 1 = exp exp : ` ` k` ` Therefore Re(1 (p)) Re(1 (q)) j1 (p)j j1 (q)j + log 1 + : e f p q (p ) (q ) e f e f Now (p q ) (8=7) pq when p q 6= 6 so that, for = Re(1(p)) and = Re(1(q)), j1 (p)j j1 (q)j j1 (p)jj1 (q)j 2 7 log 1 + + e f e f (p ) (q ) (p q ) (8=7) pq 8 p q 2 2 since j1 (p)j = 2 and j1 (q)j = 2 . We therefore deduce that Re(1 (p)) Re(1 (q)) ; : p q Now j1 (p)j = 2Re(1 (p)), and so j1 (p)jj1 (q)j pq b ; 1 1 and therefore, by (9.9), (p) C = (1 + O( )) 1 : b 1 pjb Substituting this into (9.10), and using (9.8), yields (9.6) in this case, except if b = 6. When b = 6, we use relations (9.11) and (9.17) to deduce that (b) 1 1 3 Re(1 (2)) Re(1 (3)) jC j + exp + exp : b 3 2 2 2 3 Since the left hand side is 1 + O( ) and Re(1 (2)); Re(1 (3)) 0, we deduce that Re(1 (2)); Re(1 (3)) , as before. Proceeding now as in the case b 6= 6 completes the proof (9.6) when b = 6 too. Case 2b. Now suppose that b = p is a prime power with e 2. Using (9.11), we see that e2 j e1 (b) 1 (p ) (p ) C = 1 + : j e1 b p p p j=0 De ne so that jj = 1 and C = jC j. Then b b e2 j e1 1 1 (p ) 1 (p ) (b) 1 + = 1 jC j j e1 p p p b j=0 which is 0 and O( ) by (9.12) and (9.16). Taking real parts, and noting that each Re(1 (p )) 0, we deduce that Re(1 ); Re(1 (p))=p = O( ) by considering the j = 0 and 1 terms. Hence j1 (p)j = j(1 (p))j j1 j +j1 (p)j p 1 THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 53 and therefore (p) 1=2 C = (1 + O( )) 1 : Substituting this into (9.10), and using (9.8), yields the result in this case. Case 2c. Now suppose that b = p is a prime, so that C = p=(p 1). Hence we cannot use (9.16) to gain information on (p). Now, using Lemma 7.7, we have p log p p p 1 log p (2) (2) (2) jL ()j = L () + O S + O : 1 p p (p) p jp (p)j p p Combining this with (9.14) yields that (p 1)=jp (p)j 1 + O( ); that is 1 (p) 1 + 1 + O( ); p 1 Taking real parts, we deduce that Re(1 (p)) 1 1 + 1 + O( ); p 1 with the lower bound being trivial. This implies that j1 (p)j p: Using Lemma 7.7, we then conclude that (2) (1) (2) (1) L () + C L () = L () + L () 1 b 1 p p 1 p p log p (2) (1) = L () + L () + O p p p (p) p 1 p p p (2) (1) = L () + L () + O E ; p p p 1 p 1 which, together with (9.10), completes the proof of (9.6) in this last case too. Remark 9.1. Note that can be quite big if N is small and if is not very close to a=b. So we cannot say more than the last formula without more information on the location of . We conclude the paper with the proof of Theorem 2.2. +c Proof of Theorem 2.2. Let y = e for a large enough constant c, as above, and de ne w via the relation j k=`j = 1=(`y ). Note that w = u(1 + O(1= )). So we may show the theorem with w in place of u. Arguing as at the beginning of this section, and applying Lemma 7.1 with z = 1, we nd that X X X i 1 (n) 1 (n)e(kn=`) (n) = + O(log ) G() 2 n 2 n n q n2Z; n6=0 P (n)y P (n)y 1jnjy X X (n) (n) cos(2kn=`) = + O(log ): n n + + P (n)y P (n)y ny 54 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS Note that (2.2) and the argument leading to (4.6) imply that j1 (n)j 3=4 ( log ) (9.18) n P (n)y (n;b )=1 for all > 0. Hence is 1-pretentious. Now suppose that b = 1. Substituting (9.18) in our formula for (n), we obtain n q X X X i 1 cos(2kn=`) 3=4 (n) = + O(( log ) ) G() n n + + n q P (n)y P (n)y ny If ` > 1, then we bound the second sum using Corollary 7.10. If ` = 1, then the result follows from Lemma 3.3. This concludes the proof of part (a). Finally, assume that b is a prime number so that b = b. Writing n = b m with (m; b) = 1, we nd that 0 1 j j X X X X B C i (b ) (m) (m) cos(2b km=`) B C (n) = + O(log ) j @ A G() b m m + + n q j=0 P (m)y P (m)y (m;b)=1 my ; (m;b)=1 0 1 j j X X X B C (b ) 1 cos(2b km=`) 3=4 B C = + O(( log ) ) j @ A b m m + + j=0 P (m)y P (m)y (m;b)=1 my ; (m;b)=1 by (9.18) and the trivial estimate 1=m j log b j log . If ` = 1, then we w j w y =b <my have that X X X i (b ) 1 3=4 (n) = + O( log ) G() b m n q j=0 P (m)y m>y ; (m;b)=1 1 1=b 3=4 = e (1 P (w)) + O(( log ) ) 1 (b)=b v j by Lemmas 7.7 and 3.3. Next, if ` 6= b for all v 0, then b k=` 2= Z for all j 0. So Corollary 7.10 implies that X X X i (b ) 1 1 1=b 3=4 3=4 (n) = + O( log ) = e + O(( log ) ) G() b m 1 (b)=b n q j=0 P (m)y (m;b)=1 as claimed. Finally, assume that ` = b for some v 1. The terms with j v contribute j v X X (b ) 1 1 1=b (b ) = e (1 P (w)) + O(1) j v b m 1 (b)=b b j=v P (m)y m>y ; (m;b)=1 THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 55 by Lemmas 7.7 and 3.3. When j v 1, we apply Corollary 7.10. The total contribution of those terms is v1 j v1 X X X (b ) 1 (b ) 1 1 + + O(log ) j v1 b m b (b) m + + j=0 P (m)y P (m)y (m;b)=1 my ; (m;b)=1 v v v1 1 1 (b )=b (b ) P (w) = e 1 + + O(log ) v1 b 1 (b)=b b b 1 by Lemmas 7.7 and 3.3. Putting the above estimates together, yields the estimate v1 e i 1 (b)=b (b) 1 (b) (n) = 1 + P (u) + O() G() 1 1=b b b 1 n q 3=4 1=4 with = (log ) = . To nish the proof of the theorem, we specialize the above formula when = , in which case k=` = a=b, so that v = 1. Then the modulus of the left hand side equals m(), which is > by assumption. On the other hand, if we set z = (1(b))=(b 1), then 1 (b)=b 1 (b) 1 + P (u )z 1 + P (u ) = : 1 1=b b 1 1 + z Consequently, 1 + P (u )z 1 + O(): 1 + z Since 0 P (u ) 1 and 2Re(z) = jzj 0, we have that 2 2 2 1 + P (u )z jzj (1 P (u ) ) + 2(1 P (u ))Re(z) (1 P (u ))jzj 0 0 0 0 = 1 1 : 2 2 1 + z j1 + zj j1 + zj Putting together the above inequalities proves the last claim of part (b). Hence the proof of Theorem 2.2 is now complete. 56 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS 10. Additional tables even odd all q min mean :9999 max min mean :9999 max mean :9999 10000019 0.728 0.994 1.74 2.11 0.788 1.51 3.35 3.74 1.25 3.25 10000079 0.725 0.994 1.75 2.05 0.795 1.51 3.35 3.81 1.25 3.26 10000103 0.725 0.994 1.75 2 0.793 1.51 3.34 3.83 1.25 3.25 10000121 0.724 0.994 1.75 2.02 0.793 1.51 3.34 3.78 1.25 3.26 10000139 0.726 0.994 1.75 2.01 0.797 1.51 3.35 3.74 1.25 3.25 10000141 0.719 0.994 1.74 2.02 0.79 1.51 3.33 3.82 1.25 3.25 10000169 0.721 0.994 1.75 2.06 0.788 1.51 3.34 3.73 1.25 3.25 10000189 0.709 0.994 1.75 2.01 0.793 1.51 3.35 3.71 1.25 3.25 10000223 0.73 0.994 1.75 2 0.783 1.51 3.34 3.81 1.25 3.25 10000229 0.723 0.994 1.74 2.04 0.784 1.51 3.33 3.79 1.25 3.25 10000247 0.716 0.994 1.74 2.05 0.794 1.51 3.34 3.71 1.25 3.25 10000253 0.724 0.994 1.75 2.05 0.783 1.51 3.34 3.75 1.25 3.25 11000027 0.724 0.994 1.75 2.03 0.797 1.51 3.34 3.72 1.25 3.25 11000053 0.733 0.994 1.75 2.07 0.781 1.51 3.34 3.81 1.25 3.25 11000057 0.707 0.994 1.75 2.03 0.79 1.51 3.34 3.74 1.25 3.25 11000081 0.724 0.994 1.75 2.01 0.789 1.51 3.33 3.77 1.25 3.25 11000083 0.724 0.994 1.75 2.05 0.799 1.51 3.34 3.84 1.25 3.25 11000089 0.728 0.994 1.75 2.05 0.794 1.51 3.33 3.77 1.25 3.26 11000111 0.724 0.994 1.75 2.03 0.796 1.51 3.33 3.72 1.25 3.26 11000113 0.719 0.994 1.75 2.01 0.781 1.51 3.34 3.73 1.25 3.25 11000149 0.731 0.994 1.75 2.03 0.805 1.51 3.34 3.72 1.25 3.25 11000159 0.722 0.994 1.75 2 0.797 1.51 3.33 3.83 1.25 3.25 11000179 0.724 0.994 1.74 2.03 0.796 1.51 3.35 3.86 1.25 3.26 11000189 0.728 0.994 1.75 2.1 0.794 1.51 3.34 3.68 1.25 3.26 12000017 0.723 0.994 1.74 2.08 0.8 1.51 3.34 3.8 1.25 3.26 12000029 0.72 0.994 1.74 2.06 0.791 1.51 3.33 3.84 1.25 3.26 12000073 0.735 0.994 1.75 2.05 0.794 1.51 3.34 3.9 1.25 3.26 12000091 0.728 0.994 1.75 2.02 0.794 1.51 3.35 3.73 1.25 3.26 12000097 0.719 0.994 1.75 2.08 0.788 1.51 3.34 3.75 1.25 3.25 12000127 0.724 0.994 1.75 2.09 0.794 1.51 3.34 3.71 1.25 3.25 12000133 0.727 0.994 1.75 2.11 0.785 1.51 3.34 3.73 1.25 3.25 12000239 0.715 0.994 1.75 2.02 0.797 1.51 3.34 3.8 1.25 3.25 12000253 0.713 0.994 1.75 2.02 0.786 1.51 3.34 3.76 1.25 3.26 Table 1. The minimum, maximum, mean, and :9999-quantile for m() over even, odd, and all nontrivial mod q for some selected values of q. References [BC50] P. T. Bateman and S. Chowla, Averages of character sums. Proc. Amer. Math. Soc. 1 (1950), 781{787. [BG13] J. Bober and L. Goldmakher, The distribution of the maximum of character sums. Mathematika 59 (2013), no. 2, 427{442. THE FREQUENCY AND THE STRUCTURE OF LARGE CHARACTER SUMS 57 [BG16] , P olya{Vinogradov and the least quadratic nonresidue. Math. Ann. 366 (2016), no. 1-2, [Bob14] J. Bober, Averages of character sums. Preprint (2014). arXiv:1409.1840 [CS12] S. Chatterjee and K. Soundararajan, Random multiplicative functions in short intervals. Int. Math. Res. Not. IMRN (2012), no. 3, 479{492. [Dav00] H. Davenport, Multiplicative number theory. Third edition. Revised and with a preface by Hugh L. Montgomery. Graduate Texts in Mathematics, 74. Springer-Verlag, New York, 2000. [dB51] N. G. de Bruijn, On the number of positive integers x and free of prime factors > y. Nederl. Acad. Wetensch. Proc. Ser. A. 54 (1951), 50{60. [DE52] H. Davenport and P. Erd} os, The distribution of quadratic and higher residues. Publ. Math. Debrecen 2 (1952), 252{265. [FT91] E. Fouvry and G. Tenenbaum, Entiers sans grand facteur premier en progressions arithm etiques. Proc. London Math. Soc. (3) 63 (1991), no. 3, 449{494. [Gol12] L. Goldmakher, Multiplicative mimicry and improvements to the P olya{Vinogradov inequality. Al- gebra Number Theory 6 (2012), no. 1, 123{163. [GS03] A. Granville and K. Soundararajan, Distribution of values of L(1; ). Geom. Funct. Anal. 13 (2003), no. 5., 992{1028. [GS06] , Extreme values of j (1 + it)j. The Riemann zeta function and related themes: papers in honour of Professor K. Ramachandra, 65{80, Ramanujan Math. Soc. Lect. Notes Ser., 2, Ramanujan Math. Soc., Mysore, 2006. [GS07] , Large character sums: pretentious characters and the P olya{Vinogradov theorem. J. Amer. Math. Soc. 20 (2007), no. 2, 357{384. [Har13] A. J. Harper, On the limit distributions of some sums of a random multiplicative function. J. Reine Angew. Math. 678 (2013), 95{124. [Hil86] A. Hildebrand, On the number of positive integers x and free of prime factors > y. J. Number Theory 22 (1986), no. 3, 289{307. [Hil88] , Large values of character sums. J. Number Theory 29 (1988), no. 3, 271{296. [HT93] A. Hildebrand and G. Tenenbaum. Integers without large prime factors. J. Th eor. Nombres Bordeaux 5 (1993), no. 2, 411{484. [Lam08] Y. Lamzouri, The two-dimensional distribution of values of (1 + it). Int. Math. Res. Not. IMRN (2008), Vol. 2008, 48 pp. doi:10.1093/imrn/rnn106 [MV77] H. L. Montgomery and R. C. Vaughan, Exponential sums with multiplicative coecients. Invent. Math. 43 (1977), no. 1, 69{82. [MV79] , Mean values of character sums. Canad. J. Math., 31(3):476{487, 1979. [MV07] , Multiplicative number theory. I. Classical theory. Cambridge Studies in Advanced Mathe- matics, 97. Cambridge University Press, Cambridge, 2007. [Pal] R. E. A. C. Paley, A Theorem on Characters. J. London Math. Soc. S1-7 no. 1, 28. [R en47] A. A. R enyi. On some new applications of the method of Academician I. M. Vinogradov. Doklady Akad. Nauk SSSR (N. S.) 56 (1947), 675{678. [Sai89] E. Saias, Sur le nombre des entiers sans grand facteur premier. (French) [On the number of integers without large prime factors] J. Number Theory 32 (1989), no. 1, 78{99. [Ten95] G. Tenenbaum, Introduction to analytic and probabilistic number theory. Translated from the second French edition (1995) by C. B. Thomas. Cambridge Studies in Advanced Mathematics, 46. Cambridge University Press, Cambridge, 1995. 58 J. BOBER, L. GOLDMAKHER, A. GRANVILLE, AND D. KOUKOULOPOULOS JB: Heilbronn Institute for Mathematical Research, School of Mathematics, University of Bristol, Howard House, Queens Avenue, Bristol BS8 1SN, United Kingdom E-mail address : j.bober@bristol.ac.uk LG: Department of Mathematics and Statistics, Bronfman Science Center, Williams Col- lege, 18 Hoxsey St, Williamstown, MA 01267, USA E-mail address : leo.goldmakher@williams.edu AG: Departement de mathematiques et de statistique, Universite de Montreal, CP 6128 succ. Centre-Ville, Montreal, QC H3C 3J7, Canada E-mail address : andrew@dms.umontreal.ca DK: Departement de mathematiques et de statistique, Universite de Montreal, CP 6128 succ. Centre-Ville, Montreal, QC H3C 3J7, Canada E-mail address : koukoulo@dms.umontreal.ca

Mathematics – arXiv (Cornell University)

**Published: ** Oct 29, 2014

Loading...

You can share this free article with as many people as you like with the url below! We hope you enjoy this feature!

Read and print from thousands of top scholarly journals.

System error. Please try again!

Already have an account? Log in

Bookmark this article. You can see your Bookmarks on your DeepDyve Library.

To save an article, **log in** first, or **sign up** for a DeepDyve account if you don’t already have one.

Copy and paste the desired citation format or use the link below to download a file formatted for EndNote

Access the full text.

Sign up today, get DeepDyve free for 14 days.

All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.