The Brun-Titchmarsh theorem (Q1912277)

Let \(a\) and \(q\) be relatively prime integers, and let \(\pi(x;q,a)\) denote the number of primes \(p\leq x\) with \(p\equiv a \pmod q\). Let \(\beta\) denote \(\log x/\log q\). Montgomery and Vaughan used their version of the large sieve to prove that \((1)\;\pi(x;q,a) \leq (2/(1-\beta)) x/(\phi(q) \log x)\). Motohashi, Goldfeld, and Iwaniec have given sharpenings of this for various values of \(\beta\). In particular \textit{H. Iwaniec} [J. Math. Soc. Japan 34, 95-123 (1982; Zbl 0486.10033)] proved that the bound in (1) holds with \(2/(1-\beta)\) replaced by \(5/(5-6\beta) + \epsilon\) when \(5/12 < \beta \leq 9/20 - \epsilon\) and \(8/(6-7\beta) + \epsilon\) when \(9/20 < \beta \leq 2/3-\epsilon\). In this article, the author gives a further improvement; he shows that (1) holds with \(2/(1-\beta)\) replaced by \(4/(2-\beta)+ \epsilon\) for \(9/20 - \epsilon < \beta < 1/2- \epsilon\). This result was stated without proof in another article by \textit{H. Iwaniec} [Acta Arith. 37, 307-320 (1980; Zbl 0444.10038)]. The proof uses the bilinear remainder form of the linear sieve and a theorem of Iwaniec on large values of Dirichlet polynomials. The author also gives a further conditional improvement of the Brun-Titchmarsh theorem. The condition is in terms of character sums. Let \(\theta\) be a non-negative constant with the property that for any \(\epsilon >0\), there exists \(\eta=\eta(\epsilon)\) such that \(\sum_{l\leq L} \chi(l) \ll_\epsilon L q^{-\eta}\) for all non-principal characters \(\chi \pmod q\) and all \(L\geq q^\theta +\epsilon\). Results of Burgess show that one may take \(\theta=3/8\), and for cube-free \(q\), one may take \(\theta = 1/4\). It is conjectured that one may take \(\theta=0\), and this is known if \(q\) is a power of a fixed prime. The author shows that if \(3/7 < \beta \leq 9/20 - \epsilon\), one may obtain an improved version of the Brun-Titchmarsh estimate, provided that one also knows that \(\theta \leq 7\beta /3-1\). The improvement is stated in terms of integrals of the upper and lower bound sifting functions for the linear sieve, and it is too involved to give here. The proof uses Buchstab iteration to raise the level of distribution ``most of the time''.