Primes in tuples. I (Q731202): Difference between revisions

These papers (Part II has been published in Acta Math. 204, No. 1, 1--47 (2010)) present a breakthrough in our understanding of differences between consecutive primes. If \(d_n=p_{n+1}-p_n\) is the difference between consecutive primes then it follows from the Prime Number Theorem that \(d_n/\log p_n\) has average 1. It has been known since the work of Westzynthius in 1931 that \(d_n/\log p_n\) can be arbitrarily large. In contrast, until the present papers the problem of small values had been a well-known open question, being a weaker version of the problem about bounded gaps \(d_n\) or even twin primes, for which \(p_{n+1}-p_n=2\). The main results of the first paper are on the one hand that \[ \liminf_{n\rightarrow\infty}\frac{p_{n+1}-p_n}{\log p_n}=0 \tag{\(*\)} \] and on the other that, if a certain improved version ``\(\text{BV}(\theta)\)'' of the Bombieri--Vinogradov Theorem holds, then \[ \liminf_{n\rightarrow\infty}p_{n+1}-p_n<\infty.\tag{1} \] The hypothesis \(\text{BV}(\theta)\) is that \[ \sum_{q\leq x^{\theta}}\max_{(a,q)=1} \left|\psi(x;q,a)-\frac{x}{\phi(q)}\right|\ll_A x(\log x)^{-A}\tag{2} \] for any constant \(A\), and what is required is that this should hold for some \(\theta>1/2\). The Bombieri--Vinogradov Theorem itself allows any \(\theta<1/2\), so that the smallest improvement would suffice to deduce (1). The Elliott--Halberstam Conjecture would similarly permit us to take any \(\theta<1\), which would be more than sufficient for (1). The paper makes a precise connection between the admissible value of \(\theta\) and the size of \(\liminf_{n\rightarrow\infty}p_{n+1}-p_n\), and shows for example that under the Elliott--Halberstam Conjecture one has \[ \liminf_{n\rightarrow\infty}p_{n+1}-p_n\leq 16.\tag{3} \] The second paper proves a stronger version of (\(*\)), namely that \[ \liminf_{n\rightarrow\infty}\frac{p_{n+1}-p_n}{(\log p_n)^{1/2}(\log\log p_n)^2}<\infty.\tag{4} \] The methods of these papers are also partially successful with \(p_{n+v}-p_n\) for values of \(v\geq 2\). It is shown that under the hypothesis \(\text{BV}(\theta)\) one has \[ \liminf_{n\rightarrow\infty}\frac{p_{n+v}-p_n}{\log p_n}\leq (\sqrt{v}-\sqrt{2\theta})^2\tag{5} \] so that the Elliott--Halberstam conjecture would imply that \[ \liminf_{n\rightarrow\infty}\frac{p_{n+2}-p_n}{\log p_n}=0. \] Unconditionally the \(\liminf\) in (5) is \(\leq(\sqrt{v}-1)^2\), but it is stated that the techniques of the paper may be combined with \textit{H. Maier}'s matrix method [Mich. Math. J. 35, No. 3, 323--344 (1988; Zbl 0671.10037)] to show that \[ \liminf_{n\rightarrow\infty}\frac{p_{n+v}-p_n}{\log p_n}\leq e^{-\gamma}(\sqrt{v}-1)^2, \] where \(\gamma\) is Euler's constant. In particular one would have \[ \liminf_{n\rightarrow\infty}\frac{p_{n+2}-p_n}{\log p_n}\leq 0.096\ldots \] The most obvious outstanding question is whether one might use an improved version of the Bombieri--Vinogradov Theorem (along the lines of the results in [\textit{E. Bombieri, J. B. Friedlander} and \textit{H. Iwaniec}, Acta Math. 156, 203--251 (1986; Zbl 0588.10042)] for example) to establish (1). Unfortunately it seems that the method of the present papers needs a maximum over many values of \(a\) in (2), while known sharpenings of the Bombieri--Vinogradov Theorem require inconvenient restrictions on \(a\). The basic approach in the papers is motivated by Selberg's attack [Collected Papers. Vol. II, Springer-Verlag, New York (1991; Zbl 0729.11001)] on almost-primes \(n(n+2)\), and by the reviewer's extension [Mathematika 44, No. 2, 245--266 (1997; Zbl 0886.11052)] to general \(k\)-tuples. In the latter, one examines \(k\)-tuples \((n+h_1,\ldots,n+h_k)\) and compares the sums \[ S_1=\sum_{N<n\leq 2N}\left\{\sum_{j\leq k}d(n+h_j)\right\} \left(\sum_{d|\prod_{j\leq k}(n+h_j)}\lambda_d\right)^2 \] and \[ S_2=\sum_{N<n\leq 2N}\left(\sum_{d|\prod_{j\leq k}(n+h_j)}\lambda_d\right)^2 \] for suitable sieve weights \(\lambda_d\). If one can show that \(S_1<HS_2\) for some integer \(H\) then there must be an \(n\) with \[ \sum_{j\leq k}d(n+h_j)\leq H-1. \] In the first paper the sum \(S_1\) is replaced by \[ S_1=\sum_{N<n\leq 2N}\left\{\sum_{j\leq k}\theta(n+h_j)\right\} \left(\sum_{d|\prod_{j\leq k}(n+h_j)}\lambda_d\right)^2, \] where \(\theta(m)=\log m\) for \(m\) prime, and \(=0\) otherwise. Then if \(S_1>(\log 2N)S_2\) there must be an integer \(n\in (N,2N]\) with \[ \sum_{j\leq k}\theta(n+h_j)> \log 2N, \] whence there are at least two primes in the \(k\)-tuple \((n+h_1,\ldots,n+h_k)\). Naturally, the method is doomed to failure unless for every prime \(p\) there is an \(n\) such that none of \(n+h_1,\ldots,n+h_k\) are divisible by \(p\). One says that a \(k\)-tuple which is not excluded in this way is ``admissible''. One therefore seeks to choose coefficients \(\lambda_d\) for \(d\leq R\) say, so as to maximize \(S_1/S_2\). At present we do not know how to do this. However to minimize \(S_2\) one could choose weights close to \(\lambda_d=\mu(d)(\log^+ R/d)^k\), and the authors work with the related choice \[ \lambda_d=\lambda_d^{(k,l)}=\mu(d)(\log^+ R/d)^{k+l}, \] where \(l\) is an integer parameter to be chosen later. Indeed one might use \(\mu(d)(\log^+ R/d)^kW(R/d)\) for a general weight function \(W\), but it appears that \(W(x)=x^l\), for a suitable \(l\), is not far from optimal. One then has the task of evaluating the sums \(S_1\) and \(S_2\). Taking \(k\) and \(l\) to be fixed, and \(0\leq h_i\leq N^{1/5}\) say, one can achieve this providing that \(\text{BV}(\theta)\) holds, and \(R= N^{\theta/2-\delta}\) for some fixed \(\delta>0\). The techniques used here are related to those in earlier work by the first and third authors, see [Integers 3, Paper A05, 66 p., electronic only (2003; Zbl 1118.11039)] and [Proc. Lond. Math. Soc. (3) 95, No. 3, 653--686 (2007; Zbl 1134.11034)]. The outcome is that, for an admissible \(k\)-tuple one has \[ \frac{S_1}{S_2\log 2N}\rightarrow \frac{2\theta k}{k+2l+1}\frac{2l+1}{2l+2}, \tag{6} \] so that a suitable choice for \(k\) and \(l\) will yield (1) providing that \(\theta>1/2\). However when \(\theta=1/2\) one produces bounds approaching 1 from below. To achieve (3) one takes \(k=6\) and works with the admissible 6-tuple \((n,n+4,n+6,n+10,n+12,n+16)\). This however requires a slight modification in the choice of sieve weights, and the authors use a linear combination \(\lambda_d^{(6,0)}+c\lambda_d^{(6,1)}\). To prove (4) one may modify \(S_1\), replacing it by the larger sum \[ S_1'=\sum_{N<n\leq 2N}\left\{\sum_{h=1}^L\theta(n+h)\right\} \left(\sum_{d|\prod_{j\leq k}(n+h_j)}\lambda_d\right)^2, \] where \(L\) tends to infinity with \(N\). Now, if \(S_1'/S_2>\log 2N\) there must be an interval \((n,n+L]\) containing two or more primes. The first paper uses a slightly different approach, but the second paper pursues the above line. Roughly speaking the terms in \(S_1'-S_1\) (that is to say, terms \(n+h\) where \(h\) is not one of the \(h_j\)) make a contribution of order \(L/(\log 2N)\) times \(S_1\). If we use only the Bombieri--Vinogradov value \(\theta=1/2\) then by taking \(k\) large enough we can make the limit in (6) as close to 1 as we like. Thus, if \(L=c\log 2N\) with any small \(c>0\), we may make \(S_1'/S_2>\log 2N\). The result (\(*\)) then follows. The second paper quantifies this approach. There are two major technical issues to be overcome. Firstly one must handle \(k\)-tuples uniformly for values \(k\) tending to infinity. Secondly, the error term \(x(\log x)^{-A}\) in the Bombieri--Vinogradov Theorem (2) needs to be improved. This is achieved by paying special attention to the possible effect of exceptional real zeros of Dirichlet \(L\)-functions. It is clear that one needs to work with \(k\)-tuples for which \(k\leq L\). A second fundamental constraint arises, namely that \(L\gg (\log 3N)/k\), and these limit the method to intervals of length \(\gg (\log 3N)^{1/2}\). The methods in the second paper may be applied to primes from sets other than intervals. Thus it is shown that if \(\mathcal{A}\) is any set of natural numbers for which \[ \#\{a\in\mathcal{A}: a\leq N\}\gg (\log N)^{1/2}(\log\log N)^2 \] then infinitely many elements of \(\mathcal{A}-\mathcal{A}\) are differences of two primes.