Abstract
The second moment of the Riemann zeta-function twisted by a normalized Dirichlet polynomial with coefficients of the form \((\mu \star \Lambda _1^{\star k_1} \star \Lambda _2^{\star k_2} \star \cdots \star \Lambda _d^{\star k_d})\) is computed unconditionally by means of the autocorrelation of ratios of \(\zeta \) techniques from Conrey et al. (Proc Lond Math Soc (3) 91:33–104, 2005), Conrey et al. (Commun Number Theory Phys 2:593–636, 2008) as well as Conrey and Snaith (Proc Lond Math Soc 3(94):594–646, 2007). This in turn allows us to describe the combinatorial process behind the mollification of
where \(\zeta ^{(k)}\) stands for the kth derivative of the Riemann zeta-function and \(\{\lambda _k\}_{k=1}^d\) are real numbers. Improving on recent results on long mollifiers and sums of Kloosterman sums due to Pratt and Robles (Res Number Theory 4:9, 2018), as an application, we increase the current lower bound of critical zeros of the Riemann zeta-function to slightly over five-twelfths.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
1.1 Integral moments and autocorrelation ratios of L-functions
Although applications of random matrix theory in number theory started with Montgomery’s pair correlation conjecture [54] in the mid 1970s, it is during the last two decades that the use of random matrices as a tool in the study of L-functions has become indispensable.
A local statistic is a figure that involves exclusively correlations between zeros separated on a scale of a few mean spacings. In [54], Montgomery conjectured that in the limit of a large height T on the critical line, any local statistic is supplied by the associated statistic for eigenvalues from the Gaussian unitary ensemble (GUE). These conjectures were numerically tested by Odlyzko and found to have incredible agreement [57]. The leading order of statistics involving zeros of \(\zeta \) and statistics involving eigenvalues is identical and asymptotically no factors of arithmetical nature appear [25, p. 594]. On the other hand, from the work of Bogomolny and Keating [6], it is expected that arithmetical contributions will be significant in the lower order terms.
The zeros of \(\zeta \) are not the only quantities of interest, naturally one is interested in the zeros of other L-functions and their associations to other types of matrices. Katz and Sarnak [43, 44] proposed that local statistics of zeros of families of L-functions could be understood by the eigenvalues of matrices coming from classical compact groups, see also the work of Rudnick and Sarnak in [62]. It is thus believed that families of L-functions can be modeled by the characteristic polynomials from such groups. These could be unitary, symplectic or orthogonal. The calculation performed by Iwaniec, Luo and Sarnark [40] for the one-level densities of families of L-functions with each symmetry type was in agreement with random matrix theory and further showed that there is no arithmetic component in the leading terms.
It is only recently that global, rather local, statistics were contrasted with classical compact groups. A distinct feature of global statistics is that an arithmetical factor does appear in the leading order terms.
Keating and Snaith argued in groundbreaking papers [41, 42] (first conjectured by Conrey and Ghosh [23]) that the leading terms of the moments of an L-function are the product of a characteristic polynomial from a random matrix and an Euler product. Namely, for the general 2k moment we expect that
where \(a_k\) is the Euler product
The value of \(g_k\) is the quantity associated with the eigenvalues of random matrices. Prior to [41, 42], Conrey and Ghosh [23] had conjectured that \(g_3=42\) and that in general \(g_k\) is an integer. Conrey and Gonek [24] later conjectured that \(g_4 = 24,024\). Using the above-mentioned techniques from random matrix theory, Keating and Snaith conjectured a closed formula for \(g_k\) which is given by the product
Proving that the above candidate formula for \(g_k\) is an integer is not a trivial matter [32, p. 196].
This type of result can be generalized by considering averages of ratios of products of L-functions (on the number theoretical side) or of characteristic polynomials (on the random matrix side). These results were mostly established and illustrated, sometimes conjecturally when it comes to the number theoretical aspect, by Conrey et al. [18], Conrey et al. [22], and Conrey and Snaith [25] among others.
These conjectures on the averages of ratios of products of \(\zeta \)-functions (autocorrelation of ratios) are useful not only for global statistics of zeros, as one would naturally expect, but also for local statistics. It has been suggested in fact that autocorrelation ratios of characteristic polynomials are more fundamental to random matrices than correlation functions ([7] and [25, p. 595]). Thus, the same thing could be argued for autocorrelation ratios in the L-function universe. The reason why autocorrelation ratios are useful is because they provide many local or global statistic (n-level correlations, discrete moments, etc...). Moreover, the ratios conjectures imply Montgomery’s pair correlation conjecture [22, p. 594], and they contain additional information that can be utilized to make precise conjectures about the distributions of zeros of L-functions.
Autocorrelation ratios with usually one or two L-functions in the numerator and in the denominator are enough to cover a very wide spectrum of applications, but there is no limit to the size they can accommodate. A salient feature of this paper is that later we will need many zeta-functions in both the numerator and in the denominator and this carries a heavy combinatorial price. On the other hand, Bogomolny and Keating [6] needed a heavy duty analysis of the Hardy–Littlewood prime pair conjectures to obtain the lower order terms of local statistics where arithmetical components appear. A nice feature of autocorrelations of ratios of L-functions is that they bypass those issues [25, p. 595].
Mollifiers are fundamentally important objects in the study of the moments of zeta and their arithmetic consequences. There will be plenty to say about this in a moment, but, roughly speaking, they are used to mine information about small values of L-functions, specially zeros, as well as to bound the number of zeros either in a vertical strip to the right of \({\text {Re}}(s) = \frac{1}{2}\) or at \({\text {Re}}(s)=\frac{1}{2}\). Mollifiers have also been employed to extract nonvanishing results at the central point for families of L-functions (see, among very many examples, [46, 47, 53, 65]).
Unfortunately, none of these results ever comes cheap. Even the simplest examples require sophisticated and very long analysis, and regrettably this paper is no exception. Improvements on the underlying technology have somewhat decreased the length and complexity of the calculations. For example, using the autocorrelation ratios technique, Young [71] was able to shorten Levinson’s original proof that more than one-third of the zeros of zeta are on the critical line from fifty pages to eight.Footnote 1 Further refinements on mollifiers still require lengthy calculations, however.
What is surprising is that unlike other averages of families considered in [25, p. 596], ‘there does not seem to be a random matrix analogue of mollifying as there is nothing that naturally corresponds to a partial Dirichlet series.’
Before we move on to describe the mechanism of autocorrelation ratios, we mention that the ratios can also be used to study moments of \(|\zeta '(\rho +a)|\) and allied quantities. Specifically, Conrey and Snaith showed how to obtain all the other lower order terms for these averages in [25, §7].
While difficult computations can be simplified with the autocorrelation ratios, it must be stressed that one needs to assume the Riemann hypothesis (RH), or generalized Riemann hypothesis depending on the L-function, and therefore there is a limit to how useful they are. In this paper, we provide unconditional results by using the underlying techniques and ideas behind the autocorrelation of ratios of \(\zeta \) but without using the conjectures themselves. This is of particular importance because the application we provide is an improvement on the proportion of zeros on the Riemann zeta-function on the critical line, see Theorem 1.2. Naturally, RH cannot be assumed for this type of application. Having said this, it is also important to add that, as an illustrative check, Conrey and Snaith obtained the leading terms of the simplest mollified moment in [25, §5.1] as well as higher mollified moments [25, §6] under the ratios conjecture (and hence under RH).
1.2 The ratios conjecture
Since we are only concerned with the Riemann zeta-function, we need not step outside the unitary family. Assume the Riemann hypothesis, set \(s=\frac{1}{2}+it\) and let us follow §2.1 of [25]. Farmer [30, 31] was the first to put forward the asymptotic conjecture
as \(T \rightarrow \infty \), provided that \({\text {Re}}(\gamma ), {\text {Re}}(\delta )>0\).
The approximate functional equation states that
where R is a remainder and \(XY=t/(2\pi )\). Now, use (1.1) for the zeta-functions that appear in the numerator of the integrand and use ordinary Dirichlet series for the zeta-functions in the denominator (\(\frac{1}{\zeta (s)} = \sum \mu (n)n^{-s}\)). A rule of thumb (the so-called ‘recipe’ [18, p. 52]) tells us we only need to be concerned with the pieces for which there is the same number of \(\chi (s)\) and \(\chi (1-s)\) due to oscillations. The next step is to integrate term by term and keep only the diagonal pieces and complete all the sums that we arrive at. For the first term of (1.1), this procedure boils down to
This is the perennial expression that appears, in some way or another, in all calculations involving autocorrelations of ratios of L-functions and it is what allowed Young to simplify Levinson’s proof from fifty to eight pages. Now, we only have 0 and 1 as possibilities for h and k. Thus, a simple analysis (it will not be this easy again later) shows that the sum on the right-hand side is equal to
This means that the Euler product on the right-hand side is given by the following ratio of products of \(\zeta \)
where A is the ‘arithmetical factor’
As will become clearer later in our exposition, the piece from the other term coming from (1.1) is essentially the same except that \(\alpha \) is replaced by \(-\beta \), \(\beta \) is replaced by \(-\alpha \) and it is affected by a multiplication by
These manipulations allowed Conrey et al. [22] to obtain a more precise ratios conjecture.
Conjecture 1.1
(Conrey et al. [22]). If \(-\frac{1}{4}< {\text {Re}}(\alpha ) < \frac{1}{4}\), \(\frac{1}{\log T} \ll {\text {Re}}(\delta ) < \frac{1}{4}\) and \({\text {Im}}(\alpha ),{\text {Im}}(\delta ) \ll _\varepsilon T^{1-\varepsilon }\) for every \(\varepsilon >0\), then
The key to obtaining lower order term in the pair correlations is embedded in the above conjecture. One needs to differentiate with respect to \(\alpha \) and \(\beta \) and then set \(\gamma = \alpha \) and \(\delta = \beta \). It is important (and substantially more so later on) to note that \(A(\alpha ,\beta ,\alpha ,\beta )=1\). It is also useful to see that
Unfortunately, we will not have recourse to such neat formulas in our analysis. This differentiation process turns the above into the following.
Theorem 1.1
(Conrey et al. [25]). If Conjecture 1.1 is true, then
provided that \(1/\log T \ll {\text {Re}}(\alpha ),{\text {Re}}(\beta )<\frac{1}{4}\).
The last sum over p will appear frequently in the latter sections.
One example of moments of logarithmic derivatives is taken from [22, p. 628]. It illustrates the presence of the arithmetical factor. Assuming a variant of the ratios conjecture one has
where c(r) is a function of r which is uniformly bounded for \(|r| < 1/4-\varepsilon \) and is given by
Oddly enough, although we are working in the context of global statistics and we thus expect arithmetical terms to be present, it so happens that only the simplest ones survive after undergoing a certain combinatorial process. Indeed, the arithmetical factors A and their derivatives, which are sums over primes like the one above, conspire to either become zero or to get absorbed in an error term, thus (luckily) leaving us only with terms for which the arithmetical factor is equal to one.
It is important to mention the celebrated ‘Five authors’ (Conrey, Farmer, Keating, Rubinstein and Snaith) conjecture [18, p. 44] regarding the 2k moments of zeta. First, we recall the Vandermonde
and the notation \({e}(z) = \mathrm{e}^{2\pi i z}\).
Conjecture 1.2
(CFKRS 2005). Suppose g(t) is a suitable weight function. Then,
where \(P_k\) is a polynomial of degree \(k^2\) given by the 2k-fold residue
where one integrates over small circles about \(z_i=0\), with
and \(A_k\) is the Euler product
More generally,
where
with the path of integration being small circles surrounding the poles \(\alpha _i\).
This conjecture displays the rich structure behind the moments of zeta. We shall be needing a special type of moment related to the case \(k=1\) for our purposes, and we have chosen to write our result (Theorem 7.1) in a way that parallels the structure of Conjecture 1.2.
Moreover in [18, 22, 25], a lot of effort is invested in explicating the combinatorial structure of the permutations sums, arithmetical factors as well as double products that appear in certain formulae (notable results in this direction are given by [18, Lemma 2.5.1, §2.7], [25, §5] and [22, §6.4, §7.2]). In our findings, we also come across formulae and concepts that require a similar effort but for which the existing ideas that have appeared in the literature do not seem to apply directly as far as the enumeration and the combinatorics are concerned.
Lastly, in the words of Conrey and Snaith [25, p. 596] ‘before embarking on such a [moment] calculation, it would be useful to know ahead of time what the answer is.’ Calling it a ‘painful calculation,’ as they do, is nothing short of accurate. The presentation we have decided to adopt follows this philosophy closely. We have started with simple examples where the combinatorics are undemanding so that objects can be counted ‘by hand’ before moving on to the general principles. Even when considering the general principles, we have paused at critical steps to fall back to special cases (which have not have appeared in the literature before) to better illustrate the underlying blueprint of our results.
1.3 Motivation and choice of the mollifiers
We set \(s = \sigma +it\) with \(\sigma , t \in \mathbb {R}\) and denote by \(\zeta (s) = \sum _{n\ge 1}n^{-s}\) the Riemann zeta-function for \(\sigma >1\), and otherwise by analytic continuation. Now, let
N(T) denote the number of zeros \(\rho = \beta + i \gamma \), counted with multiplicity, of \(\zeta (s)\) inside the rectangle \(0< \beta < 1\) and \(0< \gamma < T\),
\(N_0(T)\) denote the number of zeros, counted with multiplicity, of \(\zeta (s)\) such that \(\beta = \tfrac{1}{2}\) and \(0< \gamma < T\).
It is well known [66, Ch. IX] that the asymptotic formula for N(T) is given
where the term S(T) is
as \(T \rightarrow \infty \). Let \(\kappa \) be the proportion of zeros on the critical line, i.e.,
In 1942, Selberg [63] showed that \(1 \ge \kappa > 0\). This means that a positive proportion of non-trivial zeros of the Riemann zeta-function lies on the critical line.
Let Q(x) be a real polynomial satisfying \(Q(0)=1\) and \(Q'(x)=Q'(1-x)\). Set d to be the degree of Q, so that \(d = \deg (Q) \ge 1\). We then define the differential operator V by
where, for large T, we set
Using the functional equation of \(\zeta (s)\), Littlewood’s lemma and the arithmetic and geometric mean inequalities, Conrey [16] (see also [50]) showed that
Here, \(\sigma _0 = 1/2 - R/L\) where R is a bounded positive real of our choice and \(\psi \) is a mollifier.
A mollifier is a regular function designed to dampen the large values of \(\zeta (s)\) so the product \(V\psi \) is expected to be smaller than V. To mollify \(\zeta \), one uses a Dirichlet polynomial
with suitable coefficients b(n, y) and an acceptable length \(y=T^{\theta }\), where \(0< \theta < 1\).
A wide range of coefficients b(n, y) have been studied in the literature. Levinson [50] first used
with \(\theta = \frac{1}{2}-\varepsilon \). Along with the choice \(Q(x)=1-x\), Levinson was able to prove that \(\kappa > \frac{1}{3}\) in 1974.
In [19, p. 7], a comparison between Selberg’s and Levinson’s methods is made. Essentially these two methods are ‘diametrically opposed.’ Indeed, Selberg’s method is based on counting sign changes of the suitably normalized and mollified Riemann zeta-function and this is a very safe, if not entirely effective, procedure. One cannot get a negative (worse than trivial) bound for the counting number. However, due, among other things, to the fact that the zeros are not evenly spaced, Selberg’s method fails to produce significant values of \(\kappa \). On the other hand, Levinson’s method is a gamble as it could produce negative bounds for the counting number of critical zeros if the pertinent estimates are wasteful. If the mollification is ‘nearly perfect,’ then it opens the possibility for \(100\%\), or at least substantially higher values of \(\kappa \). Therefore, it behooves us to perfect the technique of Levinson’s method as much as possible and present it in its greatest flexibility and generality. This is indeed one of the goals of this article.
The next refinement is due to Conrey [15] who further generalized the mollifier to
where P(x) is a real polynomial such that \(P(0)=0\) and \(P(1)=1\). Combined with further refinements on the polynomial Q, such as taking \(d= 5\) and keeping \(\theta = \frac{1}{2}-\varepsilon \), Conrey showed that \(\kappa > 0.36581\) along with other results on the proportion of zeros of derivatives of the Riemann zeta-function on \({\text {Re}}(s) = \frac{1}{2}\).
The next improvement would be arithmetical in nature. In [2], Balasubramanian, Conrey and Heath-Brown examined the error terms from the mean value integral in (1.2) and the resulting exponential sums. In particular, they showed using Vaughan’s identity [67] on \(1/\zeta (s)\) and Weil’s bound for Kloosterman sums that one could push the size of the length \(\theta \) past the \(\frac{1}{2}\) barrier to \(\frac{9}{17}\) when the coefficients are given by (1.3). The consequences of Hooley’s conjecture \(R^*\), see [38], are also discussed.
Further improvements of this result using (1.3) are worked out by Conrey in [16], where he uses results from Deshouillers and Iwaniec [26, 27] on exponential sums to unconditionally prove that \(\theta =\tfrac{4}{7}-\varepsilon \) which results in \(\kappa > 0.4088\).
Further choices of b(n, y) have been proposed. Following the work of Luo and Yao [52], Feng [33] proposed
where \(K=2,3,\ldots \) is an integer of our choice and \(y = T^\theta \) for some \(0<\theta <1\). The set \(\{p_i\}_{i=1}^k\) is composed of distinct primes, and \(P_k\) are certain polynomials not unlike P above. Working with the two-piece mollifier
where \(y_C = T^{\theta _C}\) and \(y_F = T^{\theta _F}\) with \(\theta _C = \frac{4}{7}-\varepsilon \) and \(\theta _F = \frac{3}{7}-\varepsilon \), Feng proved that \(\kappa > 0.4107\).
The size of \(\theta _F\) was initially taken to be \(\frac{4}{7}\) and Feng later reduced it to \(\frac{1}{2}\). The proportion of zeros associated to \(\theta _F = \frac{1}{2}-\varepsilon \) is \(\kappa > 0.4128\). However, in [8, 48, 58, 60], a gap was found in Feng’s argument which reduces the size to \(\frac{3}{7}\), unless some work is done at the exponential sum level of the error terms. In [58], it is shown, by decomposing the error terms associated to \(b_F(n,y)\) into Type I and Type II sums and handling the resulting incomplete Kloosterman sums, that one can take \(\theta _F = \frac{6}{11} - \varepsilon \), thereby validating Feng’s claim that \(\theta _F = \frac{1}{2}-\varepsilon \) and \(\kappa > 0.4128\).
The best bound for an arbitrary coefficient \(a_n\) of a generic Dirichlet series \(\sum _{n \le T^{\theta }} a_n n^{-s}\) is \(\theta = \frac{17}{33} - \varepsilon \). This is due to Bettin et al. [4]. Its key ingredient is an improvement of a result of Duke, Friedlander and Iwaniec [29] on trilinear Kloosterman sums due to Bettin and Chandee [3]. We also remark that in [61], Robles and Zaharescu along with Roy proved that the bilinear Kloosterman sums of [29] lead to \(\theta = \frac{48}{95} - \varepsilon \), but this result was obtained shortly after the publication of [4].
Somewhat inspired by [52] and certainly drawing from the autocorrelation of ratios, Bui et al. [10] introduced a second piece to Conrey’s mollifier, namely they worked with
Here, \(\mu _2\) is given by the Dirichlet convolution \(\mu _2(h) = (\mu \star \mu )(h)\) and \(\chi (s)\) is such that \(\zeta (s) = \chi (s)\zeta (1-s)\), i.e., \(\chi (s) = 2^s \pi ^{s-1}\sin (\frac{1}{2}\pi s)\Gamma (1-s)\). In this case, \(y_2 = T^{1/2-\varepsilon }\) and \(P_2\) is a polynomial with similar properties to those of \(P_1\). They obtained \(\kappa > 0.4105\).
In [33, p. 515] and [10, p. 36], the idea of crossing all mollifiers
was remarked and it was hinted at that this would be a ‘technically difficult’ thing to do. This was accomplished in [60].
Lastly, a family of mollifiers that generalizes (1.4) was studied in [49] and independently and almost simultaneously by Sono in [64].
It was suspected from an argument of Farmer (‘\(\theta =\infty \) conjecture’), see [30] and [5, p. 1], that mollifiers might be optimal when their size is \(1-\varepsilon \) in that they produce \(100\%\) of zeros on the critical line. However, another intriguing recent result in this direction is due to Bettin and Gonek [5]. They prove with a very short and elegant argument involving Mellin transforms and Parseval’s formula that if one takes \(\theta = \infty \), then RH would follow (not just \(100\%\)). Of course, we are very far away from such lengths of mollifiers. Nevertheless, theoretically this approach opens the door to a direction toward RH via the moments.
In this paper, we propose to mollify the whole perturbed Riemann zeta-function. In other words, we mollify V(s) for a general d. As pointed out in the literature, see, e.g., [16, §3] and [33, p. 515], the idea behind Selberg’s method is to mollify \(\zeta (s)\) directly. However, in Levinson’s framework, what one needs to mollify is the whole perturbed function V(s). This is not an easy task and one runs into serious combinatorial difficulties. Indeed, as remarked in [33, Remark (c)], ‘it is too complicated to optimize exactly the coefficients of the mollifier.’ It is in fact too complicated to even display the terms of the mollified moment, let alone optimize them.
To accomplish this task, we examine the behavior of the inverse 1 / V(s) as the degree d increases. Namely, we will be studying the expression
as a function of d and the complex variable s. We shall be able to compute the mean value integral appearing in (1.2) with any desired degree of accuracy in terms of d.
In other words, instead of examining
we examine
and provide the clarity needed to extract the rich features that this eventual autocorrelation of ratios of products of \(\zeta \)-functions has to offer. More explicitly, as mentioned in the discussion after Conjecture 1.2 above, the object of study is the \(k=1\) case of the autocorrelation functions (sometimes called shifted moments, see [17, p. 1]). In our case, the shifts have their origins in (1.2). It will become clearer as we proceed that (1.6) will lead to integrals of a ratios of several products of shifted zeta functions [see (6.3) and (7.4) below] such as \(G(z_1, \ldots , z_{2k})\) in Conjecture 1.2 and [22, equation (5.12)].
Although computing the moment integrals of the zeta-function twisted by a general Dirichlet series associated to the mollification is the main and most difficult target of our research (see Theorem 7.1), we can give an immediate application. An interim optimization of the parameters at our disposal yields \(\kappa > 0.417293\) and \(\kappa ^* \ge 0.407511\) where \(\kappa ^*\) denotes the proportion of simple zeros on the critical line.
Theorem 1.2
More than five-twelfths of the non-trivial zeros of the Riemann zeta-function are on the critical line.
One of the most satisfactory features of the result we present for the twisted second moment is the amount of adjustability that it exhibits. Therefore, we expect computational enthusiasts to push the boundaries of the terms we present. Moreover, if or when the length of present or future mollifiers is increased, number theorists should be able to use the theoretical and numerical procedures we present in this article to refine \(\kappa \) or other arithmetical quantities of interest (see, e.g., [25] for applications of autocorrelation ratios in pair correlations, distributions, discrete moments, connections to random matrix theory, etc). We end our discussion with some problems for future work in Sect. 9.
Remark 1.1
Shortly before we presented this paper, Wu [70] uploaded his result on the twisted mean square and the critical zeros of Dirichlet L-functions. Our conclusions partially overlap, and the methodologies are independent of each other. Our bounds for the critical zeros (of Riemann or Dirichlet) are in agreement.
2 Preliminary tools
We shall devote this section to presenting the tools we will need throughout the paper. Let \(\nu (n)\) denote the number of distinct primes of n. We use \(n=p_1p_2 \ldots p_r\) to denote the prime factorization of a general square-free number. If n is a square-free number, then \(\nu (n)=r\) and \(\mu (n)=(-1)^r\).
The generalized von Mangoldt function \(\Lambda _k(n)\) is defined as (see [13, 39, 59] for applications to zeros)
for \(k \in \mathbb {N}\). If \(k=1\), then we have \(\Lambda _1(n) = \Lambda (n)\), the usual von Mangoldt function. For \({\text {Re}}(s)>1\), its Dirichlet series is given by
where \(\zeta ^{(k)}\) stands for the k the derivative of \(\zeta (s)\) with respect to s. We also note the following identity
Arithmetically, this means that
Moreover, for \({\text {Re}}(s)>1\), we can write
where
The advantage of working with \(\Lambda _{\mathcal {L},k}\) instead of \(\Lambda _k(n)\) is that \(\Lambda _{\mathcal {L},k}\) will be zero when n is not a power of a prime, whereas this is certainly not the case for the much more combinatorially complicated arithmetical function \(\Lambda _k(n)\).
We denote by \(\mathcal {P}(k)\) the representation of unordered partitions of k into positive parts [45, p. 14], i.e.,
and \(\mathcal {C}(k,n)\) stands for the ordered partitions of the integer k into n nonnegative parts, i.e.,
Moreover, we also define
The multinomial coefficients are given by
An identity involving \(\Lambda _k\) that can occasionally become useful is
for square-free n as described earlier.
The polynomials \(B_{n,k}(x_1, x_2, \ldots , x_{n-k+1})\) denote the partial or incomplete exponential Bell polynomials, whereas \(B_n(x_1, x_2, \ldots , x_n)\) will denote the nth complete exponential Bell polynomials. A good introduction to these combinatorial objects can be found for instance in [14, § 3.3]. For the sake of completeness, and given the role they will play shortly, we shall define and illustrate their main properties.
The Bell polynomials are defined by
where \(\divideontimes (k,n)\) indicates that the sum is over \(\{j_1, \ldots , j_{n-k+1}\} \in \mathcal {C}(k,n) \cap \mathcal {P}(k)\). The complete exponential Bell polynomials are given by the sum
The partial Bell polynomials can be computed efficiently by a recursion relation
where \(B_{0,0}=1\), \(B_{n,0}=0\) for all \(n \ge 1\) and \(B_{0,k}=0\) for \(k \ge 1\). When it comes to the complete Bell polynomials, we have the recursion
with \(B_0=1\). Lastly, the generating function is given by
The exponential Bell polynomial encodes the information related to the ways a set can be partitioned, and given a Bell polynomial \(B_n\) we can separate the partial Bell polynomial \(B_{n,k}\) by collecting all those monomials with degree k. We shall now illustrate some examples.
Let us for instance take \(B_{3,k}(x_1, x_2, x_3)\). We immediately find
which we can represent pictorially as in [28, 69] (Fig. 1).
\(B_{3,3}(x_1,x_2,x_3)\) (extreme left), \(B_{3,2}(x_1,x_2,x_3)\) (3 middle diagrams) and \(B_{3,1}(x_1,x_2,x_3)\) (extreme right)
In this case, \(x_1\) indicates the presence of a block with a single element, \(x_2\) the presence of a block with two elements and \(x_3\) a block with three elements. Since the coefficient of \(B_{3,2}\) is 3, we obtain three different ways of partitioning a block of 3 elements into 2 blocks, one block of 1 element and one block of 2 elements. We also note that \(B_3(1,1,1)=B_3=5\), which is the Bell number associated to 3. This represents the total number of diagrams.
Similarly, if we now consider \(B_{4,k}(x_1,x_2,x_3,x_4)\) for \(k=1,2,3,4\) then we obtain
\(B_{4,k}(x_1,x_2,x_3,x_4)\) for \(k=1,2,3,4\)
We remark again that the total number of diagrams, or Bell number, is \(B_4=15\) (Fig. 2). With these tools in mind, let us now proceed.
3 Constructing a mollifier
In this section, we will present the ideas behind the construction of a mollifier and we will capitalize on how they have been constructed up until now before we explain the approach we have taken.
3.1 The zeroth-order case \(d=0\)
In this case, going back to (1.5) with \(d=0\), one simply has
from which we get a mollifier of the form
When we impose that the cutoffs on the polynomial \(P_0\), i.e., \(P_0\) be such that \(P_0(0)=0\) and \(P_0(1)=1\), we then see that this is in agreement with (1.3).
3.2 The linear case \(d=1\)
Now we have to deal with the first derivative. This is the case contemplated by Feng [33]. Going back to (1.5), we formally get
by the use of the binomial theorem for fractional powers
Here, \(\Lambda ^{\star k}\) stands for convolving \(\Lambda \) with itself exactly k times. Suppose that n is square-free, then
According to Feng’s conjecture ([33, p. 516]), if n had had a square divisor, then the coefficients coming from (3.2) would contribute a lower order term to the mean value integrals \(\int |V\psi (\sigma _0+it)|^2 \mathrm{d}t\). This means that we could simply ignore the n’s for which \(\mu ^2(n)=0\). However, since Feng’s claim is not substantiated, we must operate by supposing that n is square-free. Otherwise, the computation of the convolution (3.3) becomes very difficult. Therefore, by keeping n square-free, we get
Here, the \(p_i\) denote distinct primes. This suggests a mollifier of the form
The conditions on \(P_{1,k}\) are that \(P_{1,k}(0)=0\) for all k. We note the following remarks.
- (1)
Feng has set the convention of starting at \(K=2\).
- (2)
Here, \(K=2,3,\ldots \) is an integer of our choice coming from the truncation of the infinite sum over k. The higher K is, the more precise the mollification. However, this is achieved at the cost of adding extra terms that require taxing computational resources.
- (3)
The sign alternator \((-1)^k\) has been absorbed into the polynomials \(P_{1,k}\), i.e.,
$$\begin{aligned} (-1)^k P_{1,k}(x) = (-1)^k \sum _{i=0}^{\deg P} a_{1,k,i} x^i = \sum _{i=0}^{\deg \tilde{P}} \tilde{a}_{1,k,i} x^i = \tilde{P}_{1,k}(x) \end{aligned}$$where
$$\begin{aligned} {\left\{ \begin{array}{ll} \tilde{a}_{1,k,i} = (-1)^k a_{1,k,i}, \\ \deg P = \deg \tilde{P}. \end{array}\right. } \end{aligned}$$ - (4)
Alternatively, we could have written
$$\begin{aligned} \psi _{d=1}(s) = \sum _{n \le y_1} \frac{\mu ^2(n)n^{\sigma _0-1/2}}{n^s} \sum _{k=2}^K \frac{1}{\log ^k y_1} (\mu \star \Lambda ^{\star k})(n) P_{1,k} \bigg (\frac{\log (y_1/n)}{\log y_1}\bigg ), \end{aligned}$$since \(\mu ^2(n)\) will discriminate square-free numbers.
3.3 The quadratic case \(d=2\)
Before proceeding with the general case, it will be instructive to see how adding the second derivative increases substantially the complexity of the combinatorics associated to this problem. The degree d is small enough that a trick that changes \(\zeta ''/\zeta \) into derivatives of \(\zeta '/\zeta \) is sufficient to obtain a useful mollifier. The effect of working with \(d=2\) is that the expression \(\mathcal {M}\) in (1.5) becomes
Now we use \(k=1\) in (2.1) to replace \(\zeta ''/\zeta \) by an expression involving only \(\zeta '/\zeta \) and hence
Let us then look at the Dirichlet convolution a little bit more closely. For n square-free, in a general power setting, we have
Using \(a=k-j+2i\) as well as \(b=j-i\) and inserting this into the above expression for \(\mathcal {M}(s,2)\) while keeping n square-free yields
Hence, the mollifier should be of the form
Again \(K=2,3,\ldots \) is a positive integer of our choice. We see that there are as many polynomials as there are primes (in this case \(k+i\) polynomials). We could also have written
since the term \(\mu ^2(n)\) will discriminate square-free numbers.
3.4 The general \(d \ge 0\) case
We now relax the condition on n and forgo the computation of the Dirichlet convolution. The advantage of operating this way will become clearer in the proof of our main result, Sects. 5.1 and 6. The general \(d \ge 0\) mollifier we want to use is given by
We now use the convolution
so that we end up with
This suggests a mollifier of the form
Here, the polynomials P are such that if \(d=0\), then \(P_{0,\ell } \equiv P_0\) with \(P_0(0)=0\) as well as \(P_0(1)=1\); and if \(d > 0\), then \(P_{d,\ell }(0)=0\) for all \(\ell \ge 0\).
As a check, we observe that \(\mu \star \Lambda ^{\star 0} = \mu \) and also when \(d=0\) the K-truncation disappears (i.e., \(K=\ell =0\)) so we are left with the mollifier
which is the Conrey–Levinson mollifier with \(y_0 = T^{\theta _0}\) where \(\theta _0 = \frac{4}{7}-\varepsilon \). We shall take \(y_d = N = T^{\theta _d}\) with \(\theta _d = \frac{4}{7}-\varepsilon \) for all \(d \ge 0\). As we shall explain in Sect. 6, there is no need to be concerned with \(\mu ^2(n)\).
3.5 Combinatorial interpretation of the mollifier
Before proceeding with the mean value integral, it is worth pausing to see from a different angle the structure of these mollifiers. Another way to interpret the combinatorial meaning behind this mollification is to apply Faà di Bruno’s formula [14]
to \(\exp (\log \zeta (s))\) so that we can write the following representation
Therefore, picking up from \(\mathcal {M}(d,s)\) we arrive at
Let us set
We place ourselves in a general setting of two Bell polynomials and three powers
This allows us to now plot the exact representations of the Bell diagrams. We first note that
which means that we will have 15 diagrams (Fig. 3).
Pictorial representation of (3.8)
As additional examples, we will now increase the precision of the truncation by keeping two Bell polynomials but taking more powers. That means
This produces
diagrams (Fig. 4).
Pictorial representation of (3.9)
Furthermore, if we instead increase the number of polynomials and keep three powers, then
Setting \(x_1=x_2=x_3=1\) in (3.10) yields 250 diagrams (Fig. 5).
Pictorial representation of (3.10)
4 Main result for the moment integral
Recall that \(L := \log T\) and let
Moreover, we shall denote the twisted second moment by
where \(\varPhi \) is a smooth function supported on [1, 2] and satisfying \(\varPhi ^{(j)}(x) \ll _j \log ^j T\). The starting point is the following improvement of [58, Theorem 1.2].
Theorem 4.1
Let \(\alpha , \beta \ll L^{-1}\). Then, one has
with \(\mathcal {E}\) given by the following choices
We note that the second case was only proved for \(D=1\) in [58], but the proof below shows that it can be adapted to \(D \ge 0\). This second case will no longer be needed as we can ‘improve’ it to the third case by relaxing the condition that discriminates square-free numbers.Footnote 2 However, we leave it in the theorem for chronological accuracy or in case it becomes useful in another moment integral problem of this type. The first case (when \(\alpha =\beta =0\)) is due to Bettin, Chandee and Radziwiłł [4]. As mentioned earlier, the key to that result is the recent improvement of bilinear Kloosterman sums of Duke, Friedlander and Iwaniec [29] due to Bettin and Chandee [3].
Effectively, this means that if one considers a Dirichlet polynomial whose coefficients are given by the third case, then one can ‘push’ the size of \(\theta \) from \(\frac{6}{11}\) to \(\frac{4}{7}\). Also note that if \(D=0\) in the third case, then one recovers \(\mu \star \Lambda ^{\star 0} = \mu \), that is the Conrey–Levinson mollifier. This means that the Feng mollifier [33, 48, 60] and all its generalizations can be taken to have size \(\theta _d = \frac{4}{7} - \varepsilon \) for \(D \ge 0\) just as in Conrey’s mollifier.
Proof of Theorem 4.1
We need to adapt the proof appearing in [58, §3.2.5] since only the error terms are affected and the main terms remain exactly the same.
We assume familiarity with [58] and its notation. Following through the proof of [58, §3.2.5], we must bound the quantity
where \(|r(n)| \ll n^\epsilon \) is some function. Recall that \(n_i \asymp N_i/d\) and \(n_1 \le N\). The coefficients \(a_m\) are a finite linear combination of functions, so using linearity and the definition of the \(a_m\) we see it suffices to bound
Using the binomial theorem and the additivity of the logarithm, we may separate d and \(n_1\) from one another in the polynomial P. We see it suffices to bound
where j is some fixed, nonnegative integer.
We are now faced with the task of separating d and \(n_1\) in the arithmetic factor
In the second case of the theorem, we had a \(\mu ^2\) factor in our coefficients which meant we could automatically take d and \(n_1\) to be coprime to one another, and this simplified things somewhat.
We factor \(n_1 \rightarrow h n_1\), where \(h \mid d^\infty \) and \(n_1\) is coprime to d. The quantity to bound therefore becomes
By the additivity of the logarithm and the binomial theorem, we may separate h and \(n_1\) in \((\log hn_1)^j\), so that we must bound
for some nonnegative integers j and k (j is not necessarily the same as before).
We claim that for all \(k \ge 1\) we may write \(\Lambda _k(n)\) as a finite linear combination of functions of the form
where the \(j_i\) are nonnegative integers. We proceed by induction. The base case \(k = 1\) is trivial. Now assume it is true for k. The recurrence formula gives
By the induction hypothesis, \(\Lambda _k\) is a linear combination of functions of the desired form, and therefore so is \(\Lambda \star \Lambda _k\). To see that \(\log (n) \Lambda _k(n)\) is also of the desired form, it suffices to apply the induction hypothesis and note that for any arithmetic functions \(f_1, \ldots , f_J\) we have
where \(g_{j,i} = f_j\) if \(i \ne j\), and \(g_{i,i} = f_i \log \).
The quantity to bound is therefore a finite linear combination of quantities of the form
We now give the argument that shows how to separate \(n_1\) from dh in expressions of this form. We have
Since \(\log (\ell _i)^{j_i}\Lambda (\ell _i)\) is supported on prime powers and \(n_1\) is coprime to dh, we see that \((\mu \star \log ^{j_1}\Lambda \star \cdots \star \log ^{j_R}\Lambda )(dhn_1)\) is the sum of a bounded number of functions of the form
We write \(m = m'm''\), where \(m' \mid dh\) and \(m'' \mid n_1\) to see that this last quantity is
which gives the desired separation. It follows that the quantity in question is a finite linear combination of sums of the form
where G is some function satisfying \(|G(h)| \ll \tau (dh)^{O(1)} (\log N)^{O(1)}\), and the \(j_i\) are nonnegative integers that are not necessarily the same as before.
Now that we have separated the variables, we perform a combinatorial decomposition on the function \((\mu \star \log ^{j_1}\Lambda \star \cdots \star \log ^{j_R}\Lambda )(n_1)\) to reduce to Type I and Type II exponential sums. For \(n \le x\) and \(K \ge 1\), a fixed integer, the multinomial theorem and Heath-Brown’s identity [36] imply
Performing the usual dyadic decompositions and combinatorial maneuvers (see [58] for more details), we see that the sum in question is a linear combination of \(O((N_1/dh)^\epsilon )\) sums of the form
where \(\alpha \) and \(\beta \) are arithmetic functions supported on dyadic intervals, \(U \asymp N_1/dh, V \asymp N_2/d\). We have dropped the h summation, temporarily, but we shall return to it later. In the Type I case, we have that \(\alpha \) is supported on integers \(n \le W\) and \(\beta = \log ^S\) for some nonnegative integer S. Here, \(W \ll (N_1/dh)^{1/3}\) is a parameter at our disposal. In the Type II case, we have that \(\alpha \) and \(\beta \) are supported on integers in \([W,(N_1/dh)W^{-1}]\).
Let us consider first the case of a Type I sum. Using the binomial theorem to separate variables, we have
where \(\alpha \) is not necessarily the same as before, and \(E \ll W\). We use summation by parts and the usual Weil bound for incomplete Kloosterman sums [68] to obtain that the sum on f is
Summing over all our variables and recalling \(E \ll W\), we obtain a Type I bound of
We now turn to a Type II sum. We use the binomial theorem to separate variables in \(\log ^j\) and the Mellin transform to separate variables in \(F_{N_1}\). We obtain
where \(EF \asymp U\) and \(W \ll E\ll F \ll UW^{-1}\). By a result of Deshouillers and Iwaniec ([27, Lemma 1], the variable h corresponds to \(\rho \) and bounds on h arise from that lemma) we obtain
We now set \(W = U^{1/4}\) to balance the Type I and Type II bounds. We use the bound \(E^{-a} \ll W^{-a} = U^{-a/4}\) for \(a < 0\), and \(E^b \ll U^{b/2}\) for \(b > 0\), and note that in each term the power of h is smaller than the power of U. The variable U is of size \(N_1/dh\); thus, when U appears there is a hidden h. The significance of the power of h being less than the power of U is that we need some positive power of h in the denominator at the end of the day so we can sum over h and not have it blow up. Since
we find that, in the notation of [58], we have
Thus, taking \(N = T^{\frac{4}{7} - \varepsilon }\) is permissible. \(\square \)
5 The square-free terms and Feng’s conjecture
We pause at this point to explain what happened to Feng’s conjecture (unproved claim, rather) that only square-free terms contribute to the integral appearing in Theorem 4.1. It is not entirely clear how Feng guessed that numbers containing a square would ‘contribute a lower order term for the mean value integral’ [33, p. 516].
It is plausible that the approach Feng followed, which is a mixture of elementary inductions juxtaposed with Mertens’s formulas \(\sum _{p \le y} \frac{\log p}{p} = \log y + O(1)\) and \(\sum _{p|n} \frac{\log p}{p} \ll \log \log n\), only works reasonably well if n is square-free. Feng rewrites the mollifier by removing the Dirichlet convolutions and explicitly writing their result. He uses
which is only valid when n is square-free. Attempting to obtain a general formula for \((\mu \star \Lambda ^{\star k})(n)\) when n contains a square runs into combinatorial difficulties (see also [51, p. 309] for a similar commentary on mollifying \(\frac{1}{\zeta '}\) instead of \(\frac{\zeta '}{\zeta }\)). In fact, a key result of this is Lemma 9. This lemma explicitly demands that we deal only with square-free numbers when multiplying mollifiers. Again, adapting such a result for arbitrary n would not be easy.
For \(\ell = 1\), \((\mu \, \star \, \Lambda ^{\star \ell }) = \mu ^2(n)(\mu \, \star \,\Lambda ^{\star \ell })\), the plots below illustrate the differences between the unrestricted convolution \((\mu \star \Lambda ^{\star \ell })\) and the ‘square-free convolution’ \(\mu ^2(n)(\mu \star \Lambda ^{\star \ell })\) for \(\ell =2\) and \(\ell =3\), respectively (Figs. 5, 6).
Left: Plots of \(\mu ^2(n)(\mu ^{\star 2} \star \log n)(n)\) in yellow, \((\mu ^{\star 2} \star \log n)(n)\) in blue. Right: Plot of \((\mu ^{\star 2} \star \log n)(n)-\mu ^2(n)(\mu ^{\star 2} \star \log n)(n)\), for \(1 \le n \le 1000\)
Next (Fig. 7).
Left: Plots of \(\mu ^2(n)(\mu ^{\star 3} \star \log ^{\star 2} n)(n)\) in yellow, \((\mu ^{\star 3} \star \log ^{\star 2} n)(n)\) in blue. Right: Plot of \((\mu ^{\star 3} \star \log ^{\star 2} n)(n)-\mu ^2(n)(\mu ^{\star 3} \star \log ^{\star 2} n)(n)\), for \(1 \le n \le 1000\)
Moreover (Fig. 8).
Left: Plots of \(\mu ^2(n)(\mu ^{\star 4} \star \log ^{\star 3} n)(n)\) in yellow, \((\mu ^{\star 4} \star \log ^{\star 3} n)(n)\) in blue. Right: Plot of \((\mu ^{\star 4} \star \log ^{\star 3} n)(n)-\mu ^2(n)(\mu ^{\star 4} \star \log ^{\star 3} n)(n)\), for \(1 \le n \le 1000\)
The plots below illustrate the difference between the partial sums of the restricted and unrestricted convolutions (Figs. 9, 10).
Left: Plots of \(\sum _{n \le x} (\mu \star \Lambda ^{\star 2})(n)\) in blue and \(\sum _{n \le x} \mu ^2(n)(\mu \star \Lambda ^{\star 2})(n)\) in orange for \(1 \le x \le 1000\). Right: Plots of \(\sum _{n \le x} (\mu \star \Lambda ^{\star 2})(n)-\sum _{n \le x} \mu ^2(n)(\mu \star \Lambda ^{\star 2})(n)\) for \(1 \le x \le 1000\)
Left: Plots of \(\sum _{n \le x} (\mu \star \Lambda ^{\star 3})(n)\) in blue and \(\sum _{n \le x} \mu ^2(n)(\mu \star \Lambda ^{\star 3})(n)\) in orange for \(1 \le x \le 1000\). Right: Plots of \(\sum _{n \le x} (\mu \star \Lambda ^{\star 3})(n)-\sum _{n \le x} \mu ^2(n)(\mu \star \Lambda ^{\star 3})(n)\) for \(1 \le x \le 1000\)
5.1 Proof of Feng’s conjecture
We examine here only the case \(K = 2\); the case of higher values of K works out similarly. We examine the main term
where \(a_d\) is our mollifier coefficient for \(K = 2\), i.e.,
where P is a polynomial satisfying \(P(0) = 0\). We write \(a_d = b_d + c_d\), where \(b_d\) is the Feng coefficient
and \(c_d\) is the difference of \(a_d\) and \(b_d\). The reader may like to prove, as an exercise, that \(c_d\) vanishes unless d is of the form \(d = p^k m\), where p is a prime, \(k \ge 2\), and m is a square-free integer coprime to p, in which case one has
Putting \(a_d = b_d + c_d\) into (5.2), we get the ‘Feng’ part, which has \(b_d\) and \(b_e\), and then three other terms all of which involve at least one of \(c_d\) or \(c_e\). We expect that these latter three terms are negligible.
We investigate the crossed case with \(b_d\) and \(c_e\). Write h(d) for the sum over \(q_1\) and \(q_2\). Using what we know about \(c_e\), and ignoring the factor of \(T \widehat{\varPhi }(0)\), we see we may write this expression as
Now we put the sum on \(p^k\) as the outermost sum. There are two cases to consider: \(p \mid d\) and \(p \not \mid d\). We consider here only the ‘generic’ case \(p \not \mid d\). We therefore have
We now change variables \(m \rightarrow em\), where \(e \mid d\) and \((m,d)=1\), and arrive at
Next, we interchange the order of summation so the e-sum is the next sum after the \(p^k\)-sum. We change variables \(d \rightarrow ed\) to get
To make the d- and m-sums independent from one another, we apply Möbius inversion to remove the condition that m and d are coprime. Interchanging orders of summation and changing variables yet again, we obtain
Observe that for coprime square-free integers m, n we have
By two applications of this rule and exploiting the additivity of the logarithm, we find
Thus, we have six different cases to consider.
5.1.1 The m-sum
Before we consider these cases, let us estimate the sum over m, which does not depend on which of the six cases we are in. For technical reasons, we do not proceed immediately to writing the sum as an integral. Instead, it is advantageous to use Möbius inversion to deal with the coprimality condition \((m,er)= 1\). Our sum becomes
It turns out to be helpful to have this averaging over f at our disposal.
We write
so our sum becomes
Writing the logarithm as an integral and interchanging the order of summation and integration, we have
Here, \(c = \frac{1}{\log N}\), say. By an Euler product computation, we see that the integrand (apart from the \(s^{-j-1}\)) is equal to
We use the classical zero-free region for \(\zeta \) and some bounds for \(\zeta ^{-1}\) close to the 1-line, see [66, Ch. III]. We can show that the m-sum is equal to
plus
We calculate this residue by taking j derivatives in s and then letting \(s \rightarrow 0\). One can bound this term by
here we have used the fact that
It is also important that \(|\beta | \asymp \frac{1}{\log N}\). We therefore obtain a total bound of
and upon summing over f we obtain
5.1.2 The d-sums
Let us now turn to the various cases for the sum over d. Due to symmetry, there are actually only two d-sums we need to consider:
as well as
We start with the sum involving h(d). Opening up h(d) and interchanging the order of summation, we have
This inner sum on d is almost identical to the m-sum we estimated earlier. We similarly use Möbius inversion to handle the condition \((d,er) = 1\), and then write the sum as an integral and use the zero-free region for \(\zeta \). We find the sum on d is
and we obtain the sum bound upon summing over \(q_1\) and \(q_2\).
Let us now turn to the other type of d-sum. We wish to estimate
Since d is square-free, we may write
where \(\mathfrak {p}\) is a prime (we have used all the Roman letters traditionally associated with primes, so we have to branch out a bit). We interchange the order of the \(\mathfrak {p}\)- and d-summations, and then argue as before. The bounds we obtain are of a similar shape.
5.1.3 Clean-up
It is now a routine matter to sum up all of the bounds, obtaining a final bound of
Observe that we could have afforded to lose one more logarithm, and we still would have had an acceptable bound. The reason for this is that since we are looking at Feng’s \(K = 2\), we get to divide by two logarithms in order to make the coefficients bounded. However, the difference between Feng’s \(K=2\) and our \(K=2\) is like a Conrey mollifier on average, which already has bounded coefficients. This accounts for our bound being one logarithm smaller than we actually need it to be.
6 Specializing the coefficients
Let us now take
where \(P_{(\cdot )}\) is a polynomial associated to the coefficients a or b satisfying some conditions such as \(P_{(\cdot )}(0)=0\) and \(P_{(\cdot )}(1)=1\). For the moment, \(\mathfrak {a}_n\) and \(\mathfrak {b}_n\) will be arbitrary coefficients that will later have certain parameters that will also affect \(P_{(\cdot )}\). By the Mellin representation of P, we may write
as well as
We go back to the right-hand side of Theorem 4.1 and write
The double sum over d and e, which we call \(\mathcal {S}\), requires a closer look. We will handle the first term of the t-integral, that is the term involving \(\zeta (1+\alpha +\beta )\), as the second term will have the symmetries \(\alpha \rightarrow -\beta \) and \(\beta \rightarrow -\alpha \) as well as a premultiplication by \(T^{-\alpha -\beta }\).
We have reached the point where we need some structure on \(\mathfrak {a}\) and \(\mathfrak {b}\) that we may exploit to our advantage.
6.1 The linear case \(d=1\)
To get a taste of things to come, let us first assume that
as well as
This means that (leaving out the denominator \((\log N)^{\ell _1+\ell _2}\) for \(I_{1}\))
Write \(d=d_0 d_1 \ldots d_{\ell _1}\) and \(e=e_0 e_1 \ldots e_{\ell _2}\) so that
We now employ the incredibly useful formula
where the contour of integration is a small circle around the origin. This leads us to
The advantage of this formula is that in the inner sum of \(\mathcal {S}_1\) we now have products of multiplicative functions (the \(\mu \)’s) and completely multiplicative functions (the \(d^x\)’s and the \(e^y\)’s) instead of \(\log \)’s and \(\Lambda \)’s. The next step is to write this as an Euler product so that
To recover the constant (\(p^0\)) and linear terms (\(p^1\)), we restrict the choices of the d’s and the e’s to the following nine possibilities. To enumerate them in a simple and fast manner, we shall denote the choices by the labels \(\{d_0,d_i,e_0,e_j\}\) where \(i=1,2,\ldots ,\ell _1\) and \(j=1,2,\ldots ,\ell _2\).
- (1)
For \(\{0,0,0,0\}\) we get 1 in the Euler product. This is our constant term.
- (2)
The case \(\{1,0,0,0\}\) yields
$$\begin{aligned} \frac{\mu ^{\star \ell _1+1}(p)}{[p, 1]} \frac{(p, 1)^{\alpha +\beta }}{p^{\alpha +s}} = -\frac{\ell _1+1}{p^{1+\alpha +s}}. \end{aligned}$$ - (3)
The case \(\{1,0,1,0\}\) yields
$$\begin{aligned} \frac{\mu ^{\star \ell _1+1}(p)\mu ^{\star \ell _2+1}(p)}{[p, p]} \frac{(p, p)^{\alpha +\beta }}{p^{\alpha +s}p^{\beta +u}} = \frac{(\ell _1+1)(\ell _2+1)}{p^{1+s+u}}. \end{aligned}$$ - (4)
The case \(\{1,0,0,1\}\) yields
$$\begin{aligned} \frac{\mu ^{\star \ell _1+1}(p)}{[p, p]} \frac{(p, p)^{\alpha +\beta }}{p^{\alpha +s} p^{\beta +u+w_j}} = -\frac{\ell _1+1}{p^{1+s+u+w_j}}. \end{aligned}$$ - (5)
The case \(\{0,1,0,0\}\) yields
$$\begin{aligned} \frac{1}{[p, 1]} \frac{(p, 1)^{\alpha +\beta }}{p^{\alpha +s+z_i}} = \frac{1}{p^{1+\alpha +s+z_i}}. \end{aligned}$$ - (6)
By symmetry with \(\{1,0,0,1\}\), the case \(\{0,1,1,0\}\) yields
$$\begin{aligned} -\frac{\ell _2+1}{p^{1+s+u+z_i}}. \end{aligned}$$ - (7)
The case \(\{0,1,0,1\}\) is the most difficult as it mixes the variables z and w. We have
$$\begin{aligned} \frac{1}{[p, p]} \frac{(p, p)^{\alpha +\beta }}{p^{\alpha +s+z_i} p^{\beta +u+w_j}} = \frac{1}{p^{1+s+u+z_i+w_j}}. \end{aligned}$$ - (8)
By symmetry with \(\{1,0,0,0\}\), the case \(\{0,0,1,0\}\) yields
$$\begin{aligned} -\frac{\ell _2+1}{p^{1+\beta +u}}. \end{aligned}$$ - (9)
Lastly, the case \(\{0,0,0,1\}\) is symmetric with respect to \(\{0,1,0,0\}\), and hence we get
$$\begin{aligned} \frac{1}{p^{1+\beta +u+w_j}}. \end{aligned}$$
If we now insert these terms into \(\mathcal {S}_1\), then we arrive at
We can write this more concisely with an autocorrelation-type ratio of zeta-functions [18, 22, 25] as
Here \(A = A_{\alpha ,\beta }(z_1,\ldots ,z_{\ell _1},w_1,\ldots ,w_{\ell _2};s,u)\) is an arithmetical factor that is absolutely convergent in some half-plane containing the origin. Examination of the nine cases above indicates that it is given by
We may now go back to \(I_1\), the first half of I, and tidy up a bit so that
where \(\mathcal {E}_3\) is the third case of Theorem 4.1 and where
Let us define the integrand to be
Our first observation is that
and hence at \(\mathbf {z}=\mathbf {w}=\mathbf {0}\) we end up with
On the other hand, for the arithmetical factor A we find that
and hence at \(\mathbf {z}=\mathbf {w}=\mathbf {0}\) we arrive at
The property that \(\mathfrak {M}_{\alpha ,\beta }(\mathbf {0},\mathbf {0};\beta ,\alpha ) = A_{\alpha ,\beta }(\mathbf {0},\mathbf {0};\beta ,\alpha ) = 1\) will become useful shortly.
We will find it expedient to examine the logarithm of A in order to turn the products into sums. One has
For all cosmetic purposes, (6.4) combined with (6.5) and (6.6) and supplemented with (6.7) is as far as we can go before things take an ugly turn due to the combinatorics involved in this situation. Indeed, this is the price to pay for having used (6.2). However, it is customary [10, 49, 60, 71] to decouple the complex variables s and u and then perform the sums over the indices i and j to give cleaner terms.
We will start with \(\ell _1 = \ell _2 =1\), and \(\ell _1 = \ell _2 =2\) in \(d=1\), then move on to \(d=2\).
6.2 Choosing the truncation \(\ell \)
6.2.1 The case \(\ell _1=\ell _2=1\)
In order to keep things readable, we shall specialize to \(\ell _1 = \ell _2 = 1\) and then delineate the path to general \(\ell _1\) and \(\ell _2\). In this simpler case, a direct residue calculus computation shows that
Here, the notation \(f^{(a,b)}(x,y)\) indicates the a-th derivative with respect to x and the b-th derivative with respect to y. Therefore,
where
by deforming the path of integration of the s, u-integrals to \({\text {Re}}(s)={\text {Re}}(u)=\delta \) with \(\delta >0\). Next use Dirichlet series to see that
We have used the Dirichlet convolution
Now we take \(\delta \asymp L^{-1}\) and bound integrals trivially to get \(J_{1} \ll L^{i+j+2}\). We then use a Taylor expansion so that \(A_{\alpha ,\beta }^{(m,n)}(0,0;s,u) = A_{\alpha ,\beta }^{(m,n)}(0,0;\beta ,\alpha ) + O(|s-\beta |+|u-\alpha |)\) for \(m,n \in \{0,1\}\). Each factor of s and u saves a factor of \(\log T\) and \(\alpha ,\beta \ll 1/\log T\). We know that \( A_{\alpha ,\beta }^{(0,0)}(0,0;\beta ,\alpha ) = 1\) and we use (6.7) to find the other three values that we are missing in the expression for \(J_{1}\).
By logarithmic differentiation, we have that
We concentrate on the left-hand side and observe from (6.7) that
Consequently,
Therefore,
By symmetry, one also gets that \(A_{\alpha ,\beta }^{(0,1)}(0,0;\beta ,\alpha ) = 0\). We next move on to the derivative with respect to \(z_1\) and with respect to \(w_1\). Unfortunately, the trick with the logarithmic derivative will not work as cleanly as it has hitherto. To find \(A_{\alpha ,\beta }^{(1,1)}\), we make use of Faà di Bruno’s formula
where we have found it useful to temporarily consolidate the notation from \(\mathbf {z}=(z_1, \ldots , z_{\ell _1})\) and \(\mathbf {w} = (w_1, \ldots , w_{\ell _2})\) to simply \(\mathbf {z}=(z_1, \ldots , z_{\ell _1}, z_{\ell _1+1}, \ldots , z_{\ell _2})\). Here, \(\pi \) runs through the set \(\Pi (\ell _1+\ell _2)\) of all partitions of \(\{1,2,\ldots ,\ell _1+\ell _2\}\) and \(B \in \pi \) indicates that the variable B runs through the list of all blocks of the partition \(\pi \). For \(\ell _1 = \ell _2 = 1\), the left-hand side of (6.8) is given by the expression
Consequently,
For the right-hand side of (6.8), we get
Since \(A_{\alpha ,\beta }(0,0;\beta ,\alpha )=1\) and \(A_{\alpha ,\beta }^{(1,0)}(0,0;\beta ,\alpha ) = A_{\alpha ,\beta }^{(0,1)}(0,0;\beta ,\alpha ) = 0\), we finally see that at \(w_1 = z_1 = 0\) and \(s=\beta \) and \(u=\alpha \) the above expression reduces to
Hence comparing the left- and right-hand sides of (6.8), we obtain
Note that this is the term appearing in [25, Theorem 2.5]. We remark that for \(|\alpha + \beta | \le \varepsilon < \tfrac{1}{4}\) we get
We can assemble our findings back in \(J_{1}\) to obtain
We quickly explain how to get the main terms in rough strokes before moving on the general idea. For \(J_{1,1}\), we may now write
where
The critical thing to notice is that the variables s and u have been separated. These integrals are special cases of Lemma 7.1 that will proved later in the wider narrative of the general case. This will require a careful use of the standard zero-free region of the Riemann zeta-function and the Laurent series around \(s=0\)
where \(C_0\) is the Euler constant. At any rate, these integrals are seen to be equal to [71, p. 546]
where the contours are small circles of radii one centered around the origin. Some re-arrangements lead to
by the change of variables \(x \rightarrow x \log N\) and \(y \rightarrow y \log N\). Going back to \(I_1\), we can perform the sums over i and j so that
Using Euler–Maclaurin summation (see [10, Corollary 4.5], [48, Lemma 2.7] and [60, Lemma 3.6]), the above can be turned into
Next, we look at \(J_{1,2}\). The procedure parallels the one above. We see that
where
as well as
Now use the standard zero-free region and bounds of \(\zeta \), as well as the Laurent series around \(s=0\)
(more on this later in Lemma 7.1) so that
Again we sum over i and j and obtain
Euler–Maclaurin summation turns this into
The cases \(J_{1,3}\) and \(J_{1,4}\) are identical, up to the obvious symmetries. We do \(J_{1,3}\). The same strategy as before leads us to
Summing over i and j yields
Lastly, Euler–Maclaurin transforms this into
By symmetry
Finally, the last term is very similar to the first one since
so that the sum over i and j takes us to
We remark that with respect to \(\log N\) this last term is of smaller order of magnitude smaller than the other terms \(I_{1,i}\) for \(i \in \{1,2,3,4\}\).
The final step is to assemble back \(I_1(\alpha ,\beta )\) with \(I_{2}(\alpha ,\beta )\), where \(I_2\) was given by the second part of Theorem 4.1 or of Eq. (6.1). This is an easy step that produces a lot of attrition. We start with \(I_{1,1}\), \(I_{1,2}\) and move on chronologically. For the very first term, we see that
However, if we focus on the part involving \(\alpha \) and \(\beta \) only, then
Next, we bring in the differential operators Q to get
At \(\alpha = \beta = -R/L\), the above reduces to
Finally, the main term coming from this subdivision of I is given by
and putting all these things back together produces
Next, for \(I_{2}\) we had
Moreover,
and therefore
We move on to \(I_{3}\) following the same strategy as before. For this term,
Also
Therefore,
By symmetry,
Finally, we come to the last term that was of smaller order of magnitude in \(\log N\). This term is similar to \(I_{1}\) except for the presence of \((1-u)^2\) in the integrand and the power of \(\log N\) in the denominator. The sum over primes is less than \(\zeta (2)\) and
as \(T \rightarrow \infty \), and since \(N = T^{\theta }\) and \(\alpha , \beta \ll L^{-1}\)
Thus, for the integral in the term
we will get a term of size \(\tfrac{T}{\log ^3 N}\log T \ll \tfrac{T}{\log ^2 T}\), and hence \(I_{5,1}(\alpha ,\beta ) + T^{-\alpha -\beta }I_{5,1}(-\beta ,-\alpha )\) is an error term of size \(T/L^2 + T/L \ll T/L\).
Alternatively, to deal with \(I_{5}\) we set \(n=2\) in [56, Entry 2.5] so that
Using \(c=\frac{3}{2}\), \(a=-1\) and \(x=p^{1+\alpha +\beta }\), we can turn the denominator in the sum into an integral
For this last part, we have
We examine the term containing \(\alpha \) and \(\beta \) and observe that
Bring in the differential operators Q to get
Thus, we are left with
Hence, the term associated to \(A_{\alpha ,\beta }^{(1,1)}(0,0;\beta ,\alpha )\) is an error term.
We also remark that
could prove useful in other scenarios to turn the sum into zeta functions.
6.2.2 The case \(\ell _1=\ell _2=2\)
Suppose we now move on to \(\ell _1 = \ell _2 = 2\). We know that \(A_{\alpha ,\beta }(0,0,0,0;\beta ,\alpha ) = 1\) and a direct calculation from (6.7) shows that
Next, for the second derivatives
For the triple and fourth derivatives,
The first four equations imply that
We now use (6.8) to see that
By symmetry,
Likewise
By symmetry,
For the triple derivatives, we get
By symmetry
Finally, (presenting only the surviving terms to save space)
In other words,
Computing the contour integrals directly leads us to
Here, \(\widetilde{\mathfrak {M}}_{\alpha ,\beta }(z_1, z_2, w_1, w_2; s,u)\) is the same as \(\mathfrak {M}_{\alpha ,\beta }(z_1, z_2, w_1, w_2; s,u)\) but with \(A_{\alpha ,\beta }(z_1, z_2, w_1, w_2; s,u)\) replaced by 1. The reason behind this change is that most of the terms of
involve partial derivatives of A that are zero and the ones that are not zero will eventually become error terms of size O(T / L). Thus, for practical purposes we only really need to retain the terms that are multiplied by \(A_{\alpha ,\beta }(0,0,0,0;s,u)\). It is now clear that the process to achieve the main terms from here is the same as in the case \(\ell _1 = \ell _2 = 1\), only that there are lot more of them.
7 The general case \(d \ge 0\)
We now specialize the coefficients to
as well as
The task ahead is to compute the arithmetical sum \(\mathcal {S}_d = \sum _{hm=kn}\) appearing in (6.1). This is precisely where the autocorrelation of ratios of \(\zeta \) technique plays again a fundamental role and where we appreciate why it is unnecessary to discriminate square-free numbers n as discussed in Sect. 3. To this end, define
by writing the generalized von Mangoldt function as \(\Lambda _q = \mu \star \log ^q\) and setting \(L_d := \ell _1 + \ell _2 + \cdots + \ell _d\).
We have the rather unfortunate situation where it becomes difficult to recycle the alphabet, but it should be clear that only the d’s in (d, e), [d, e] and \(d^{\alpha +s}\) are being summed. The d’s involved in the logs are a parameter of our choice.
We now write
and
so that \(\mathcal {S}_d\) becomes
To remove the various powers of \(\log \), we now employ Cauchy’s integral formula
where the contour of integration is a small circle around the origin. This further transforms \(\mathcal {S}_d\) into
The multiplicativity of the above expression is now conducive for a representation as an Euler product format. One has that
To recover the main terms, we find the constant (\(p^0\)) and linear (\(p^1\)) terms of the above Euler product. To do so, we must constrain the choices of the d’s and the e’s to nine possibilities, much like in the \(d=1\) case. While the computation is not complicated, it is so cumbersome that we provide it below for the benefit of the reader. The enumeration will be as follows. Label the choices by \(\{d_0, d_{q,i}, e_0, e_{\bar{q},j} \}\) for some \(q,\bar{q}\in \{1,2,\ldots ,d\}\), some \(i \in \{1,2,\ldots ,\ell _1\}\) and some \(j \in \{1,2,\ldots ,\ell _2\}\).
- (1)
For \(\{0,0,0,0\}\) we naturally get 1 in the Euler product. This will be the constant term.
- (2)
The case \(\{1,0,0,0\}\) yields
$$\begin{aligned} \frac{\mu ^{\star L_d+1}(p)}{[p,1]} \frac{(p,1)^{\alpha +\beta }}{p^{\alpha +s}} = - \frac{L_d+1}{p^{1+\alpha +s}}. \end{aligned}$$ - (3)
By symmetry, the case \(\{0,0,1,0\}\) yields
$$\begin{aligned} - \frac{\bar{L}_d+1}{p^{1+\beta +u}}. \end{aligned}$$ - (4)
From the case \(\{1,0,1,0\}\), we obtain
$$\begin{aligned} \frac{\mu ^{\star L_d+1}(p)\mu ^{\star \bar{L}_d+1}(p)}{[p,p]} \frac{(p,p)^{\alpha +\beta }}{p^{\alpha +s}p^{\beta +u}} = \frac{(L_d+1)(\bar{L}_d+1)}{p^{1+s+u}}. \end{aligned}$$ - (5)
The case \(\{1,0,0,1\}\) is
$$\begin{aligned} \frac{\mu ^{\star L_d+1}(p)}{[p,p]} \frac{(p,p)^{\alpha +\beta }}{p^{\alpha +s}p^{\beta +u+w_{\bar{q},j}}} = -\frac{L_d+1}{p^{1+s+u+w_{\bar{q},j}}}. \end{aligned}$$ - (6)
The symmetric case to (5), which is \(\{0,1,1,0\}\), is given by
$$\begin{aligned} - \frac{\bar{L}_d+1}{p^{1+s+u+z_{q,i}}}. \end{aligned}$$ - (7)
The case \(\{0,1,0,0\}\) yields
$$\begin{aligned} \frac{1}{[p,1]} \frac{(p,1)^{\alpha +\beta }}{p^{\alpha +s+z_{q,i}}} = \frac{1}{p^{1+\alpha +s+z_{q,i}}}. \end{aligned}$$ - (8)
Immediately, we see that for \(\{0,0,0,1\}\) we get
$$\begin{aligned} \frac{1}{p^{1+\beta +u+w_{\bar{q},j}}}. \end{aligned}$$ - (9)
The last case is the most unpleasant one because it mixes z’s and w’s. For \(\{0,1,0,1\}\), we arrive at
$$\begin{aligned} \frac{1}{[p,p]} \frac{(p,p)^{\alpha +\beta }}{p^{\beta +u+w_{\bar{q},j}} p^{\alpha +s+z_{q,i}}} = \frac{1}{p^{1+s+u+z_{q,i}+w_{\bar{q},j}}}. \end{aligned}$$
We now insert these terms back into \(\mathcal {S}_d\) and write
The last step is to write this in as an autocorrelation ratio of zeta-functions. This naturally leads us to
Here, \(A=A_{\alpha ,\beta }(\{z_{q,i}\}, \{w_{\bar{q},j}\}; s, u) = A_{\alpha ,\beta }(\mathbf {z},\mathbf {w};s,u)\) is again an arithmetical factor that is absolutely convergent in some half-plane containing the origin. We can re-construct this arithmetical term by assembling back the nine cases above. The term \(A_{\alpha , \beta }\) is given by
We remark that
Hence at \(\mathbf {z}=\mathbf {w}=\mathbf {0}\) we obtain
Finally, we may define \(\zeta (1+\alpha +\beta ) \times ({\text {integrand}})\) to be
and observe that
so that \(\mathfrak {M}_{\alpha ,\beta }(\mathbf {0},\mathbf {0}; \beta , \alpha )=1\). Indeed, the formulas
are key properties that will be used frequently. We summarize this result below.
Theorem 7.1
Let \(\mathfrak {a}_d\) and \(\mathfrak {b}_e\) be given by (7.1) and (7.2), respectively, and set
One has that
where \(\mathfrak {M}_{\alpha ,\beta }\) is given by (7.6).
Again, this is as far as one can go before things get out of hand and in fact it is bad enough as it is already. The procedure to obtain the main terms is exactly the same as it was in the case \(d=1\). We give the general recipe below before we proceed.
- (1)
Choose the degree d of the polynomial Q. The higher the degree, the more taxing the calculation.
- (2)
Choose a truncation \(\ell , {\bar{\ell }}\). Once more, the higher the truncation, the more taxing the calculation.
- (3)
Use the logarithmic derivative of (7.5) in conjunction with Faà di Bruno’s formula (6.8) and with (7.7) to isolate the derivatives of \(A_{\alpha ,\beta }(\{z_{q,i}\}, \{w_{\bar{q},j}\}; s, u)\) at \(s=\alpha \), \(u=\beta \) and \(\{z_{q,i}\}=\{w_{\bar{q},j}\}=0\). Most of the derivatives will be 0 and the remaining ones will go into an error term of size \(\ll T/L\). In other words, only the terms that have a coefficient \(A_{\alpha ,\beta }(\mathbf {0},\mathbf {0};s,u)\) will survive.
- (4)
Insert the results back into \(I_1\) (recall that \(I_2\) is basically the same as \(I_1\) up to symmetries) and deform the paths of integration to \({\text {Re}}(s)={\text {Re}}(u)=\delta \) with \(\delta >0\) then take \(\delta \asymp L^{-1}\) and bound trivially. This allows us to change \(A_{\alpha ,\beta }(\mathbf {0},\mathbf {0};s,u)\) into \(A_{\alpha ,\beta }(\mathbf {0},\mathbf {0};\beta ,\alpha )\) with an acceptable error term.
- (5)
Use Dirichlet convolutions of appropriate \(\mathbf {1} \star \Lambda _1^{k_1} \star \Lambda _2^{k_2} \star \cdots \star \Lambda _d^{k_d}\) to separate the complex variables s and u along with knowledge of (7.7). The truncation of the n-sum will be at N.
- (6)
Divide and conquer strategy: separate the resulting terms and apply Lemma 7.1 below to each one of them.
- (7)
Sum over i and j to recover the polynomials \(P_{d,\ell }\) and \(P_{d,{\bar{\ell }}}\).
- (7)
Use Euler–Maclaurin result (Lemma 7.2) below to turn the remaining sums involving arithmetical functions into integrals.
Remark 7.1
If we look at the nine cases that we considered earlier in this section to get to the autocorrelation ratio of \(\mathcal {S}_d\), it is clear that each of these cases would remain unaffected by the presence of \(\mu ^2(d)\) and \(\mu ^2(e)\). Indeed, each case would be multiplied by either \(\mu ^2(p^0)=\mu ^2(1)=1\) or \(\mu ^2(p^1)=\mu ^2(p)=(-1)^2=1\), where p is always a prime. We will be explicit about this important point. If we had taken
instead of (7.1), then the affected possibilities (once symmetries are accounted for) in the Euler product are
- (1)
The case \(\{0,0,0,0\}\) yields
$$\begin{aligned} \frac{\mu ^2(1)\mu ^2(1)}{1 \times 1} = 1. \end{aligned}$$ - (2)
The case \(\{1,0,0,0\}\) yields
$$\begin{aligned} \frac{\mu ^2(p) \mu ^{\star L_d+1}(p) \mu ^2(1)}{[p,1]} \frac{(p,1)^{\alpha +\beta }}{p^{\alpha +s}} = - \frac{L_d+1}{p^{1+\alpha +s}}. \end{aligned}$$ - (4)
The case \(\{1,0,1,0\}\) yields
$$\begin{aligned} \frac{\mu ^2(p)\mu ^{\star L_d+1}(p)\mu ^2(p)\mu ^{\star \bar{L}_d+1}(p)}{[p,p]} \frac{(p,p)^{\alpha +\beta }}{p^{\alpha +s}p^{\beta +u}} = \frac{(L_d+1)(\bar{L}_d+1)}{p^{1+s+u}}. \end{aligned}$$ - (5)
The case \(\{1,0,0,1\}\) yields
$$\begin{aligned} \frac{\mu ^2(p)\mu ^{\star L_d+1}(p)\mu ^2(1)}{[p,p]} \frac{(p,p)^{\alpha +\beta }}{p^{\alpha +s}p^{\beta +u+w_{\bar{q},j}}} = -\frac{L_d+1}{p^{1+s+u+w_{\bar{q},j}}}. \end{aligned}$$ - (7)
The case \(\{0,1,0,0\}\) yields
$$\begin{aligned} \frac{\mu ^2(1)\mu ^2(1)}{[p,1]} \frac{(p,1)^{\alpha +\beta }}{p^{\alpha +s+z_{q,i}}} = \frac{1}{p^{1+\alpha +s+z_{q,i}}}. \end{aligned}$$ - (9)
The case \(\{0,1,0,1\}\) yields
$$\begin{aligned} \frac{\mu ^2(1)\mu ^2(1)}{[p,p]} \frac{(p,p)^{\alpha +\beta }}{p^{\beta +u+w_{\bar{q},j}} p^{\alpha +s+z_{q,i}}} = \frac{1}{p^{1+s+u+z_{q,i}+w_{\bar{q},j}}}. \end{aligned}$$
Moreover, the arithmetical factor that would arise from (7.8) also behaves nicely.
It is likely that the present authors would have not noticed this had they not used the autocorrelation of ratios approach. Indeed, this shows that the powerful autocorrelation of ratios technique is really the ‘way to do things.’
In any case, Feng’s approach is certainly commendable for his main terms are accurate and his outstanding intuition about the size of the mollifier was spot on.
7.1 Example: the quadratic case \(d=2\)
We have illustrated the mechanism of this twisted moment in its easiest manifestation, that is when \(d=1\) and when the truncations are small, i.e., \(\ell _1, \ell _2 \le 2\). Now that we have given the general procedure, it is advisable to see what happens when the degree of Q is increased from 1 to 2. Indeed, this will make the above presentation more agreeable and in fact these terms have never before appeared in the literature.
With \(d=2\), we have \(L_d = \ell _1 + \ell _2\) and \(\bar{L}_d = {\bar{\ell }}_1 + {\bar{\ell }}_2\). Let us set \(\ell _1 = \ell _2 = {\bar{\ell }}_1 = {\bar{\ell }}_2 = 1\) for simplicity and use \(\widetilde{\mathfrak {M}}_{\alpha ,\beta }(z_{1,1},z_{2,1},w_{1,1},w_{2,1};s,u)\) to denote \(\mathfrak {M}_{\alpha ,\beta }(z_{1,1},z_{2,1},w_{1,1},w_{2,1};s,u)\) with the arithmetical factor \(A_{\alpha ,\beta }\) replaced by 1.
We start with the arithmetical product. We can compute the logarithm from (7.5) of \(A_{\alpha ,\beta }\) so that
Using (7.9), it is straightforward to show that
etc. The pattern as to what terms will survive becomes clearer (some of the ‘balanced’ mixed derivatives). However, it does not really matter because these surviving terms will get absorbed in an error term of size O(T / L) much like in the case \(d=1\) that we handled previously.
A lengthy residue calculus computation, again we are paying the toll for having used (7.3), shows that (Mathematica 11.2 was used to produce this)
Each of these cases is now treated as in Sect. 5.1 with the assistance of Lemmas 7.1 and 7.2, see below. A way to automate this process would be most welcome.
7.2 Resuming the general case \(d \ge 0\)
We shall finish our delineation of the main terms. An inspection shows that a general term of (using the compactified notation of Theorem 7.1)
is of the form
where \(\Psi (k_1, k_2, \ldots , k_{d}; l_1, l_2, \ldots , l_d; m_1, m_2, \ldots , m_{d})\) is a constant independent of the complex variables \(\alpha , \beta , s\) and u. The combinatorial meaning of
follows from the residue calculus process and it involves partitioning the structure of d and \(L_d, \bar{L}_d\). What is critically important about this is not so much the powers and coefficients, which are achievable in an elementary if ugly way as we just showed, but its structure in terms of the variables \(\alpha , \beta , s\) and u. In Sect. 9, we raise the question of whether some of the combinatorial tools such as [18, Lemma 2.5.1] could be helpful in elucidating the meaning of these contour integrals. In any case, in what follows, we will need to decouple s and u and to do this, we shall use the Dirichlet convolution
This should not be surprising as it somewhat tallies up with the structure of the coefficients in Theorem 4.1.
Going back to (6.1) with these new coefficients and considering only the first part of the integral (i.e., the one involving \(\zeta (1+\alpha +\beta )\)) yields
where
Deforming the path of integration of the s, u-integrals to \({\text {Re}}(s)={\text {Re}}(u)=\delta \) with \(\delta >0\), we come to see that
Following the recipe, take \(\delta \asymp L^{-1}\) and bound trivially to get \(J_{1,d} \ll L^{i+j+\sum r \ell _r +\sum {\bar{r}} \ell _{\bar{r}}}\). Next, use a Taylor expansion \(A_{\alpha ,\beta }^{(m,n)}(\mathbf {0},\mathbf {0};s,u)=A_{\alpha , \beta }^{(m,n)}(\mathbf {0},\mathbf {0};\beta , \alpha )+ O(|s-\beta |+|u-\alpha |)\). This puts us in the following situation
by the aid of (7.11) and where
as well as
We shall need one further tool from complex analysis before proceeding.
Lemma 7.1
Suppose \(0< \delta \asymp L^{-1}\), \(i\ge 1\) and \(\alpha \ll L^{-1}\). Let \(N \ge n \ge 0\) and \(l_r=0,1,2,\ldots \) for \(r=1,\ldots ,d\). One then has for some \(\nu \asymp \log \log N\) that
where
with the contour of integration being a circle of radius one centered at the origin and enclosing \(-\alpha \).
Proof
The idea behind this proof is to use the standard zero-free region of \(\zeta \), see for instance [10, Lemma 6.1]. Let U be a large parameter with \(U\rightarrow \infty \) and \(U = o(T)\) as \(T\rightarrow \infty \), which will be chosen at the end of the proof. The above integral is, by Cauchy’s theorem, equal to the sum of the residues at \(s=0\) and at \(s=-\alpha \) plus the sum over line integrals over the segments \(\gamma _1 = \{s=it \,:\, t \in \mathbb {R}, \; |t| \ge U\}\), \(\gamma _2 = \{ s = \sigma \pm iU \, : \, -c/\log U \le \sigma \le 0\}\), and \(\gamma _3 = \{ s = -c/\log U + it \, : \, |t| \le U\}\), where c is some fixed positive constant such that \(\zeta (1+\alpha +s)\) has no zeros in the region on the right-hand side of the contour determined by the \(\gamma _i\)’s. Another two requirements on c are that the estimate
holds in this region, and that
see [55, Theorem 6.7]. Figure 11 illustrates the contour of integration.
Curve \(\gamma \) in the proof of Lemma 7.1
Then, one has
since \(i \ge 1\). Moreover, since \(n\le N\)
and finally
Appropriately choosing \(U \asymp \log N\) gives an error of size \(O((\log \log N)^{i+\sum _{r=1}^{d}r l_r}) = O(\log N)\). The last thing we need to do is collect the residues, which we write as the contour integral
where the contour is now a small circle \(\Omega \) of radius \(\asymp L^{-1}\) centered around the origin such that \(-\alpha \in \Omega \). Since the radius of the circle tends to zero as \(T \rightarrow \infty \), we may use the Laurent expansions around \(s = -\alpha \)
where, we recall, \(C_0\) is the Euler constant, as well as
to finally arrive at
Using a direct estimation yields that the right-hand side above is equal to
which is precisely the definition of \(\Upsilon (d,\mathbf {l})\) given in the statement of the lemma. \(\square \)
The resulting contour integral leads to three very different cases which we now need to distinguish. Set
7.2.1 Case A: Terms for which \(\omega (d,\mathbf {l}) = -1\)
In this case, we have
by the change of variable \(x \rightarrow x \log N\) and where we have adopted the shorter notation
with
We take the chance now to sum over i (the index coming from the polynomial decomposition of \(P_{d,\ell }\)), so that
7.2.2 Case B: Terms for which \(\omega (d,\mathbf {l}) = 0\)
We now deal with
by the use of Cauchy’s integral theorem and where
We perform the sum over i so that
7.2.3 Case C: Terms for which \(\omega (d,\mathbf {l}) > 0\)
This is the most complicated case. We have
where
For the third case, we employ
for \(\tau =0,1,2,\ldots \). Here, P is a polynomial in \(\log q\) of degree \(\tau \), see [10, p. 56] and [60, p. 285]. Only the first term of the right-hand side above contributes when we insert this expression into \(\Upsilon (d,\mathbf {l})\). This is because the contribution from the second term vanishes by taking the contour to be arbitrary large. The explicit calculation is as follows (temporarily set \(\tau = \omega (d,\mathbf {l})-1\) and \(q=\frac{N}{n}\)):
by the use of the change of variables \(t=q^{-a}\). Lastly, we do the sum over i and obtain
Let us now recap. We had \(I_d(\alpha ,\beta ) = I_{1,d}(\alpha ,\beta ) + I_{2,d}(\alpha ,\beta ) = I_{1,d}(\alpha ,\beta ) + T^{-\alpha -\beta }I_{1,d}(-\beta ,-\alpha )+O(T/L)\), where
The terms \(F_d\) are given by the three different cases
depending on whether \(\omega (d,\mathbf {l}) = -1\), \(\omega (d,\mathbf {l}) = 0\) and \(\omega (d,\mathbf {l}) > 0\), respectively. Similarly
depending on whether \(\omega (d,\mathbf {k}) = -1\), \(\omega (d,\mathbf {k}) = 0\) and \(\omega (d,\mathbf {k}) > 0\), respectively. Here, \(\mathcal {U}, \mathcal {V}\) and \(\mathcal {W}\) are given by (7.12), (7.13), (7.14), (7.15).
So we must now cross each of all possible \(3 \times 3 = 9\) terms. In fact, due to symmetries, there will only be six different cases, much like in the Euler product.
To deal with the sum over n, we will need the following Euler–Maclaurin lemma.
Lemma 7.2
We have
for smooth functions F and H in the interval [0, 1].
Proof
Let g(n) be an arithmetic function with
for some \(c_g>0\) and \(k_g\ge 1\). Then,
for F, H and s as in the lemma. The proof of this statement is almost the same as the proof of [48, Lemma 2.4], and we thus omit it. It is thus enough to show that
To prove (7.18), we need
with g(n) as in (7.17). The proof of these three equations uses partial summation, and we omit the details. Finally, let f(n) be an arithmetic function with
then
We have
We now have
The last equality can be proved for instance by induction over \(k_f\), which completes the proof of (7.21). We now can show that we have for each \(k\ge 1\)
We prove this equation by induction. We have \(\Lambda _1= \Lambda \), and thus the case \(k=1\) is trivial. We thus assume the statement holds for a given k. We use (2.1) together with (7.19) and (7.20) and get
In a similar way, one can show that
We now can prove (7.18) and thus complete the proof of the lemma. We argue by induction. The case \(k_1=k_2=\cdots =k_m=0\) follows immediately from (7.21). Suppose now that (7.18) for given \(k_1\), \(k_2\), ...\(k_m\). We then get with (7.21)
This ends the proof. \(\square \)
All but one of the \(\log N\)’s in \(I_{1,d}\) will cancel with the \(\log N\)’s coming from the Euler-Maclaurin expression leaving us only with \(\frac{1}{\log N}\) in \(I_{1,d}\).
The last step is to apply the differential operators Q on \(I_{1,d}(\alpha ,\beta )+T^{-\alpha -\beta }I_{1,d}(-\beta ,-\alpha )\)
From the above procedure, a term of the form \(\log N\) (such as \(\log (N^{x+y}T)\) for example) will come out. This term will be combined with the term \(\frac{1}{\log N}\) above to produce an expression like \(\frac{\log (N^{x+y}T)}{\log N} = \frac{\theta (x+y)+1}{\theta }\), thereby removing the \(\log T\) dependence. For the terms involving derivatives of the arithmetical factor \(A_{\alpha ,\beta }\), it can be seen that the resulting expression after the application of the two Q’s will be of the form \(\frac{\log N}{\log ^2 N} \asymp \frac{1}{\log T}\), thereby producing error terms of size \(\frac{T\widehat{\varPhi }(0)}{\log T} \ll T/L\).
This ends our delineation of the final main terms.
8 Numerical aspects
The below numerical calculations are similar to the ones that appeared in the older version of Feng’s paper on the arXiv. Feng obtains \(\kappa \ge 0.417288\) which was rounded up to 0.4173. Since we were never able to achieve a \(\frac{4}{7}\) length in the past, we had to rely on our code of the main terms (which matched Feng’s). However, we are now able to reach \(\frac{4}{7}\); we are only too happy to give credit to Feng and recover part of his numerical setup. Let us set \(d=1\) and \(K=3\) in (3.4) as well as
Moreover, we slightly increase the number of terms of the polynomials P of the mollifiers (Feng only used \(\deg P_{1}=4\), \(\deg P_{2}=2\) and \(\deg P_{3}=1\)) and take
as well as
Letting \(\varepsilon \rightarrow 0\) yields \(\kappa \ge 0.417293962\). Using only a linear polynomial in Q yields the proportion of simple zeros on the critical line, see [1, 37, 51]. Taking
as well as
along with \(R=1.1167\), \(\theta =\frac{4}{7}\), \(d=1\) and \(K=3\) and letting \(\varepsilon \rightarrow 0\) yields \(\kappa ^* \ge 0.407511457\).
9 Further remarks and future work
We end our paper with some questions and discussions on how the methodology we have presented could be used in future work.
- (1)
Naturally, the most pressing question is whether these ideas can be applied to other L-functions in the Selberg class. Along with this comes their associations with compact groups.
- (2)
Equally important would be the effect of these mollification refinements on other arithmetical objects such as discrete moments, logarithmic moments, k-th moments and pair correlations as illustrated in, e.g., [25].
- (3)
In the proof of Theorem 4.1, it was important to exploit the additivity of the logarithm in \(a_n \sim (\mu \star \log )(n)\). This allowed us to use Vaughan and Heath-Brown identities [36, 67] and get \(\theta = \frac{4}{7}-\varepsilon \). Moreover, in [4] it is shown that the best size one can take for coefficients \(a_n\) about which we know nothing (other than \(a_n \ll _\varepsilon n^{\varepsilon }\)) is \(\theta = \frac{17}{33}-\varepsilon \). It is an interesting question to ask how one can increase the length of the Dirichlet polynomials without specializing the coefficients \(a_n\) too much.
- (4)
As mentioned in Sect. 1.3, the mixing of (1.4) and the Feng mollifier (\(d=1\)) is technically difficult. It is possible to obtain a better result for \(\kappa \) by working with a mollifier of the form
$$\begin{aligned} \psi (s)&= \sum _{\ell =0}^K (-1)^\ell \sum _{\ell _1 + \ell _2 + \cdots + \ell _d = \ell } \left( {\begin{array}{c}\ell \\ \ell _1, \ell _2, \ldots , \ell _d\end{array}}\right) \\&\quad \times \sum _{n \le y_d} \frac{n^{\sigma _0-1/2}}{n^s} \frac{(\mu \star \Lambda ^{\star \ell _1} \star \Lambda _2^{\star \ell _2} \star \cdots \star \Lambda _d^{\star \ell _d}) (n)}{(\log y_d)^{\sum _{r=1}^d r \ell _r}} P_{d,\ell } \bigg ( \frac{\log (y_d/n)}{\log y_d} \bigg ) \\&\quad \quad + \chi (s + \tfrac{1}{s}-\sigma _0) \sum _{hk \le y_2} \frac{\mu _2(h)h^{\sigma _0-1/2}k^{1/2-\sigma _0}}{h^sk^{1-s}}P_2 \bigg (\frac{\log (y_2/hk)}{\log y_2}\bigg ). \end{aligned}$$If necessary, then further pieces from [49, 64] may be brought in. This would qualify as a painful calculation. Moreover, it would also be interesting to set \(d=1\) and analyze how a large truncation of \(\ell \) could yield terms from a higher degree \(d \ge 2\). What would be a good balance between the degree d and the truncation \(\ell \)?
- (5)
Is there a way to find a better expression (if possible a comfortable one such as [18, Lemma 2.5.1]) for the contour integrals that yield the logarithmic derivatives of \(\zeta \)? See for instance the end of Sects. 6.2 and 7.1. Another way to rewrite \(\mathfrak {M}_{\alpha ,\beta }\) in (7.6) is
$$\begin{aligned} \mathfrak {M}_{\alpha ,\beta }(\mathbf {z},\mathbf {w};s,u)&= (-1)^{1 \times \ell _1 + 2 \times \ell _2 + \cdots d \times \ell _d}(-1)^{1 \times {\bar{\ell }}_1 + 2 \times {\bar{\ell }}_2 + \cdots d \times {\bar{\ell }}_d} \\&\quad \times \frac{\zeta (1+s+u)^{(L_d+1)(\bar{L}_d+1)}\zeta (1+\alpha +\beta )}{\zeta (1+\beta +u)^{\bar{L}_d+1} \zeta (1+\alpha +s)^{L_d+1}} \\&\quad \times \bigg (\prod _{q=1}^d \prod _{i=1}^{\ell _q} \frac{\partial ^q}{\partial z_{q,i}^q}\bigg ) \bigg (\prod _{\bar{q}=1}^d \prod _{j=1}^{{\bar{\ell }}_{\bar{q}}} \frac{\partial ^{\bar{q}}}{\partial w_{\bar{q},j}^{\bar{q}}}\bigg ) \\&\quad \times \bigg \{ \bigg (\prod _{q=1}^d \prod _{i=1}^{\ell _q} \prod _{\bar{q}= 1}^{d} \prod _{j=1}^{{\bar{\ell }}_{\bar{q}}} \zeta (1+s+u+w_{q,i}+z_{\bar{q},j})\bigg ) \\&\quad \times \bigg (\prod _{q=1}^d \prod _{i=1}^{\ell _q} \frac{\zeta (1+\alpha +s+z_{q,i})}{\zeta ^{\bar{L}_d + 1}(1+s+u+z_{q,i})}\bigg )\\&\quad \times \bigg (\prod _{\bar{q}=1}^d \prod _{j=1}^{{\bar{\ell }}_{\bar{q}}} \frac{\zeta (1+\beta +u+w_{\bar{q},j})}{\zeta ^{L_d+1}(1+s+u+w_{\bar{q},j})}\bigg )\\&\quad \times A_{\alpha ,\beta }(\varvec{z}, \varvec{w}, s, u) \bigg \}\bigg |_{\varvec{z}= \varvec{w} = 0}. \end{aligned}$$Some attempts to obtain a closed formula for this problem indicate that it is much too cumbersome and that the easiest way is to proceed with mathematical software.
- (6)
In [34], Feng and Wu used the \(d=1\) version of a close variant of \(\psi _d\) to show that infinitely often consecutive non-trivial zeros of the Riemann zeta-function differ by at least 2.7327 times the average spacing, and infinitely often they differ by at most 0.5154 times the average spacing, under RH. In other words, if
$$\begin{aligned} \lambda = \limsup (\gamma -\gamma ')\frac{\log \gamma }{2\pi } \quad \hbox {and}\quad \mu = \liminf (\gamma -\gamma ')\frac{\log \gamma }{2\pi }, \end{aligned}$$then, under RH, they prove that \(\lambda > 2.7327\) and \(\mu < 0.5154\). See also [9, 12] among others. Since \(\psi _{d=1}\) is a special case of a wider general family, can one improve the results of Feng and Wu with a higher degree d? In a recent paper of Bui and Milinovich [11], these bounds are improved. The question remains whether the technique could be useful. Very recently, Goldston and Turnage-Butterbaugh [35] have assumed RH and improved results in this direction. In particular using the weights developed by Wu (which have their root in Feng’s mollifier), they have shown that there are infinitely many zeros of the zeta-function whose differences are smaller than 0.50412 times the average space.
- (7)
In [20, 21], the asymptotic large sieve is used along with Levinson’s method to obtain lower bounds for the proportion of simple zeros on the critical line of the twists by primitive Dirichlet characters of a fixed L-function of degree 1, 2 or 3. For a certain family of Dirichlet L-functions, Conrey, Iwaniec and Soundararajan prove that at least \(56\%\) of the zeros are on the critical line and are simple. For aesthetic reasons, it would be desirable to increase this proportion to more than \(\frac{3}{5}\). This would necessitate an analysis of more complicated mollifiers. The asymptotic large sieve technology would not apply directly, and additional difficulties would arise.
Notes
At one point in Levinson’s original paper, there are twenty-four cancellations going on simultaneously! See [51, p. 308] for further details.
It is not exactly an improvement but rather a different problem altogether.
References
Anderson, R.J.: Simple zeros of the Riemann zeta-function. J. Number Theory 17, 176–182 (1983)
Balasubramanian, R., Conrey, B., Heath-Brown, D.R.: Asymptotic mean square of the product of the Riemann zeta-function and a Dirichlet polynomial. J. Reine Angew. Math. 357, 161–181 (1985)
Bettin, S., Chandee, V.: Trilinear forms with Kloosterman fractions. Adv. Math. 328, 1234–1262 (2018)
Bettin, S., Chandee, V., Radziwiłł, M.: The mean square of the product of \(\zeta (s)\) with Dirichlet polynomials. J. Reine Angew. Math. 729, 51–79 (2017)
Bettin, S., Gonek, S.: The \(\theta = \infty \) conjecture implies the Riemann hypothesis. Mathematika 63(1), 29–33 (2017)
Bogomolny, E.B., Keating, J.P.: Gutzwiller’s trace formula and spectral statistics: beyond the diagonal approximation. Phys. Rev. Lett. 77, 1472–1475 (1996)
Borodin, A., Strahov, E.: Averages of characteristic polynomials in random matrix theory. Commun. Pure Appl. Math. 59, 161–253 (2006)
Bui, H.M.: Critical Zeros of the Riemann Zeta-Function. Preprint. arXiv:1410.2433
Bui, H.M.: Large gaps between consecutive zeros of the Riemann zeta-function. J. Number Theory 131, 67–95 (2011)
Bui, H.M., Conrey, B., Young, M.P.: More than 41% of the zeros of the zeta function are on the critical line. Acta Arith. 150.1, 35–64 (2011)
Bui, H.M., Milinovich, M.B.: Gaps between zeros of the Riemann zeta-function. Q. J. Math. 69, 403–423 (2018)
Bui, H.M., Milinovich, M.B., Ng, N.: A note on the gaps between consecutive zeros of the Riemann zeta-function. Proc. Am. Math. Soc. 138(12), 4147–4175 (2010)
Chaubey, S., Malik, A., Robles, N., Zaharescu, A.: Zeros of normalized combinations of \(\xi ^{(k)}(s)\) on \(\operatorname{Re}(s)=1/2\). J. Math. Anal. Appl. 461, 1771–1785 (2018)
Comtet, L.: Advanced Combinatorics. Reidel, Kufstein (1974)
Conrey, J.B.: Zeros of derivatives of the Riemann’s \(\xi \)-function on the critical line. J. Number Theory 16, 49–74 (1983)
Conrey, J.B.: More than two fifths of the zeros of the Riemann zeta function are on the critical line. J. Reine Angew. Math. 399, 1–26 (1989)
Conrey, J.B., Farmer, D.W., Keating, J.P., Rubinstein, M.O., Snaith, N.C.: Autocorrelation of random matrix polynomials. Commun. Math. Phys. 237, 365–395 (2003)
Conrey, J.B., Farmer, D.W., Keating, J.P., Rubinstein, M.O., Snaith, N.C.: Integral moments of \(L\)-functions. Proc. Lond. Math. Soc. (3) 91, 33–104 (2005)
Conrey, J.B., Iwaniec, H.: Critical Zeros of Lacunary \(L\)-Functions. Preprint. arXiv:1607.03288
Conrey, J.B., Iwaniec, H., Soundararajan, K.: Critical zeros of Dirichlet L-functions. J. Reine Angew. Math. 681, 175–198 (2013)
Conrey, J.B., Iwaniec, H., Soundararajan, K.: Critical Zeros of Dirichlet L-Functions. Preprint. arXiv:1105.1176
Conrey, J.B., Farmer, D.W., Zirnbaeur, M.R.: Autocorrelation of ratios of \(L\)-functions. Commun. Number Theory Phys. 2, 593–636 (2008)
Conrey, J.B., Ghosh, A.: A conjecture for the sixth power moment of the Riemann zeta-function. Int. Math. Res. Not. 15, 775–780 (1998)
Conrey, J.B., Gonek, S.M.: High moments of the Riemann zeta function. Duke Math. J. 107, 577–604 (2001)
Conrey, J.B., Snaith, N.C.: Applications of the \(L\)-functions ratios conjectures. Proc. Lond. Math. Soc. 3(94), 594–646 (2007)
Deshouillers, J.M., Iwaniec, H.: Kloosterman sums and Fourier coefficients of cusp forms. Invent. Math. 70, 219–288 (1982)
Deshouillers, J.M., Iwaniec, H.: Power mean values of the Riemann zeta function II. Acta Arith. 48, 305–312 (1984)
Dickau, R.: Bell Number Diagrams. Wolfram Demonstrations Project (2011). http://demonstrations.wolfram.com/BellNumberDiagrams/
Duke, W., Friendlander, J., Iwaniec, H.: Bilinear forms with Kloosterman fractions. Invent. Math. 128, 23–43 (1997)
Farmer, D.W.: Long mollifiers of the Riemann zeta-function. Mathematika 40, 71–87 (1993)
Farmer, D.W.: Mean value of Dirichlet series associated with holomorphic cusp forms. J. Number Theory 49, 209–245 (1994)
Farmer, D.W.: Basic analytic number theory. In: Mezzadri, F., Snaith, N.C. (eds.) Recent Perspectives in Random Matrix Theory and Number Theory. London Mathematical Society Lecture Note Series, vol. 322. Cambridge University Press, Cambridge (2007)
Feng, S.: Zeros of the Riemann zeta function on the critical line. J. Number Theory 132, 511–542 (2012)
Feng, S., Wu, X.: On gaps between zeros of the Riemann zeta-function. J. Number Theory 132, 1385–1397 (2012)
Goldston, D.A., Turnage-Butterbaugh, C.L.: A Note on Small Gaps Between Zeros of the Riemann Zeta-Function. Preprint. arXiv:1904.06001
Heath-Brown, D.R.: Prime numbers in short intervals and a generalized Vaughan identity. Can. J. Math. 34(6), 1365–1377 (1982)
Heath-Brown, D.R.: Simple zeros of the Riemann zeta-function on the critical line. Bull. Lond. Math. Soc. 11, 17–18 (1979)
Hooley, C.: On the greatest prime factor of a cubic polynomial. J. Reine Angew. Math. 303(304), 21–50 (1978)
Ivić, A.: On the asymptotic formulas for a generalization of von Mangoldt’s function. Rend. Mat., (6) 10, 51–59 (1977)
Iwaniec, H., Luo, W., Sarnak, P.: Low lying zeros of families of L-functions. Inst. Hautes Études Sci. Publ. Math. 91, 55–131 (2000)
Keating, J.P., Snaith, N.C.: Random matrix theory and \(\zeta (1/2 + it)\). Commun. Math. Phys. 214, 57–89 (2000)
Keating, J.P., Snaith, N.C.: Random matrix theory and \(L\)-functions at \(s = 1/2\). Commun. Math. Phys. 214, 91–110 (2000)
Katz, N.M., Sarnak, P.: Random Matrices, Frobenius Eigenvalues, and Monodromy, vol. 45. American Mathematical Society Colloquium Publications, Providence (1999)
Katz, N.M., Sarnak, P.: Zeroes of zeta functions and symmetry. Bull. Am. Math. Soc. (N.S.) 36, 1–26 (1999)
Knafo, E.R.: Variance of Distribution of Almost Primes in Arithmetic Progressions. Ph.D. Thesis, University of Toronto (2006)
Kowalski, E., Michel, P., VanderKam, J.: Non-vanishing of high derivatives of automorphic \(L\)-functions at the center of the critical strip. J. Reine Angew. Math. 526, 1–34 (2000)
Kowalski, E., Michel, P., VanderKam, J.: Mollification of the fourth moment of automorphic \(L\)-functions and arithmetic applications. Invent. Math. 142, 95–151 (2000)
Kühn, P., Robles, N., Zeindler, D.: On a mollifier of the perturbed Riemann zeta-function. J. Number Theory 174, 274–321 (2017)
Kühn, P., Robles, N., Zeindler, D.: On mean values of mollifiers and \(L\)-functions associated to primitive cusp forms. Math. Z. 291, 661–709 (2019)
Levinson, N.: More than one third of zeros of Riemann’s zeta-function are on \(\sigma = \tfrac{1}{2}\). Adv. Math. 13, 383–436 (1974)
Levinson, N.: In: Nohel, J.A., Sattinger, D.H. (eds.) Selected Papers of Norman Levinson, vol. 2. Birkhauser, Boston (1996)
Luo, S., Yao, Q.: A lower bound for zeros of Riemann’s zeta function on the line \(\sigma = \tfrac{1}{2}\). Acta Math. Sin. (Chin. Ser.) 24, 390–400 (1981)
Milinovich, M.B.: Mean-Value Estimates for the Derivative of the Riemann Zeta-Function. Ph.D. thesis, University of Rochester (2008)
Montgomery, H.L.: The pair correlation of zeros of the zeta function. Analytic number theory, St. Louis, MO. In: Diamond, H.G. (ed.) Proceedings of Symposia in Pure Mathematics, vol. 23, pp. 181–193. American Mathematical Society, Providence, RI, 1972 (1973)
Montgomery, H.L., Vaughan, R.C.: Multiplicative Number Theory I: Classical Theory. Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge (2007)
Oberhettinger, F.: Tables of Mellin Transforms. Springer, Berlin (1974)
Odlyzko, M.: On the distribution of spacings between zeros of the zeta function. Math. Comput. 48, 273–308 (1987)
Pratt, K., Robles, N.: Perturbed moments and a longer mollifier for critical zeros of \(\zeta \). Res. Number Theory 4, 9 (2018)
Robles, N., Roy, A.: Unexpected average values of generalized von Mangoldt functions in residue classes. J. Aust. Math. Soc. (to appear)
Robles, N., Roy, A., Zaharescu, A.: Twisted second moments of the Riemann zeta-function and applications. J. Math. Anal. Appl. 434, 271–314 (2016)
Robles, N., Roy, A., Zaharescu, A.: Moments of the Riemann zeta-function twisted by a Dirichlet series. Unpublished private manuscript (2016)
Rudnick, Z., Sarnak, P.: Zeros of principal \(L\)-functions and random matrix theory. A celebration of John F. Nash Jr. Duke Math. J. 81, 269–322 (1996)
Selberg, A.: On the zeros of Riemann’s zeta-function. Skr. Norske Vid. Akad. Oslo I(10), 1–59 (1942)
Sono, K.: An application of generalized mollifiers to the Riemann zeta-function. Kyushu J. Math. 72, 35–69 (2018)
Soundararajan, K.: Nonvanishing of quadratic Dirichlet \(L\)-functions at \(s=\tfrac{1}{2}\). Ann. Math. 152, 447–488 (2000)
Titchmarsh, E.C.: The Theory of the Riemann Zeta-Function, 2nd edn. Oxford University Press, Oxford (1986)
Vaughan, R.C.: Sommes trigonométriques sur les nombres premiers. Proc. C. R. Acad. Sci. Paris Sér. A 285, 981–983 (1977)
Weil, A.: On some exponential sums. Proc. Nat. Acad. Sci. USA 34, 204–207 (1948)
Weisstein, E.W.: Bell numbers. From MathWorld—A Wolfram Web Resource (1999). http://mathworld.wolfram.com/BellNumber.html
Wu, X.: The twisted mean square and critical zeros of Dirichlet \(L\)-functions. Mathe. Z. (2018). https://doi.org/10.1007/s00209-018-2209-8
Young, M.P.: A short proof of Levinson’s theorem. Arch. Math. 95, 539–548 (2010)
Acknowlegements
The authors would like to acknowledge Matthew Faust of the Illinois Geometry Lab project for helping them write the code that produced their results as well as Siegfried Baluyot for proofreading an earlier version of this manuscript. Moreover, the authors would like to thank Hung Bui and Arindam Roy for useful discussions. For part of this work, the first author was supported by NSF Grant DMS-1501982. The authors are extremely grateful to the anonymous referees for their meticulous checking, for thoroughly reporting countless typos and inaccuracies as well as for their valuable comments. These corrections and additions have made the manuscript clearer and more readable.
Author information
Authors and Affiliations
Corresponding author
Additional information
Dedicated to Brian Conrey on the occasion of the 30th anniversary of his ‘Two-fifths’ paper.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Pratt, K., Robles, N., Zaharescu, A. et al. More than five-twelfths of the zeros of \(\zeta \) are on the critical line. Res Math Sci 7, 2 (2020). https://doi.org/10.1007/s40687-019-0199-8
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s40687-019-0199-8
Keywords
- Riemann zeta-function
- Critical line
- Zeros
- Mollifier
- Incomplete Kloosterman sums
- Autocorrelation ratios
- Generalized von Mangoldt functions
- Convolution structure
- Bell diagrams