GCD sums and sum-product estimates (Q2303676)

Let \(f\colon \mathbb{N} \to \mathbb{C}\) be finitely-supported, and let \(\alpha \in (0,1]\). This article concerns \textit{GCD sums} \[ \sum_{a,b} f(a) \overline{f(b)} \frac{ (a,b)^{2 \alpha} } { (ab)^\alpha } . \] These have applications to finding large values of the Riemann zeta function on the line \(\Re(s) = \alpha\), and to distribution modulo 1. Though the authors have something to say about all values of \(\alpha \in (0,1]\), this review will focus on the interesting special case \(\alpha = 1/2\), for which the sum is \[ S(f) := \sum_{a,b} f(a) \overline{f(b)} \frac{ (a,b) } {\sqrt{ab}}. \] Sharpening an estimate of \textit{A. Bondarenko} and \textit{K. Seip} [Bull. Lond. Math. Soc. 47, No. 1, 29--41 (2015; Zbl 1375.11055)], \textit{R. de la Bretèche} and \textit{G. Tenenbaum} [Proc. Lond. Math. Soc. (3) 119, No. 1, 104--134 (2019; Zbl 1458.11141)] showed that \[ |S(f)| \le \exp \left( (2\sqrt 2 + o(1)) \sqrt{ \frac{ \log N \cdot \log \log \log N}{\log \log N}} \: \: \right) \| f \|_2^2, \] and that \(2\sqrt 2\) cannot be replaced by a smaller constant. The main idea of the article is that stronger estimates can be obtained if \(f\) is assumed not to be multiplicatively-structured. Let \(K \ge 3\), and suppose the multiplicative energy of \(f\) admits the bound \[ \sum_{ab = cd} f(a) f(b) \overline{ f(c) f(d)} \le K \| f \|_2^2. \] The authors, Bloom and Walker, show that if \(\| f \|_1 \ge 3\) and \(K \ge \log \| f\|_1\) then \[ S(f) \ll \frac{ \exp \left( O \left( \sqrt{\log K \cdot \log \log \|f\|_1} \: \right) \right ) }{ O(1) + \log \|f \|_1} \| f \|_2^2, \] the proof of which adapts a probabilistic framework of \textit{M. Lewko} and \textit{M. Radziwiłł} [Adv. Math. 305, 280--297 (2017; Zbl 1375.11056)]. They then apply this to the additively-structured function \(n \mapsto r_A(n)1_\mathbb{N},\) where \(A \subset \mathbb{N}\) is finite and \(r_A(n) = \# \{ (a,b) \in A^2: a-b = n \}\). In this case \(\| f \|_2^2\) is essentially the additive energy \[ E(A) := \sum_{n \in \mathbb{Z}} r(n)^2. \] Using a sum-product-type estimate of \textit{O. Roche-Newton} and \textit{M. Rudnev} [Isr. J. Math. 209, Part 2, 507--526 (2015; Zbl 1378.52019)], they obtain \[ S(r_A(n)1_\mathbb{N}) \ll (\log N)^{O(1)} \exp \left( O \left( \sqrt{ \log (N^3 / E(A)) \cdot \log \log N} \right) \right) E(A), \] where \(N = \# A\). Finally, an application to distribution modulo 1 is presented. Let \(\mathcal{A}\) be a strictly increasing sequence of positive integers, and for \(N \in \mathbb{N}\) denote by \(A_N\) the truncation of \(\mathcal{A}\) to its first \(N\) terms. The sequence \(\mathcal{A}\) is \textit{metric poissonian} if, for almost all \(\beta \in [0,1]\), for all \(s > 0\) we have \[ N^{-1} \sum_{\substack{ a,b \in A_N \\ a \ne b \\ \| \beta (a-b) \| \le s/N}} 1 \to 2s \qquad (N \to \infty). \] Quantum mechanics propaganda aside, the characterisation of the metric poissonian property has proved to be a rich and difficult task on which we appear to be stuck. Sharpening an estimate of \textit{C. Aistleitner} et al. [Isr. J. Math. 222, No. 1, 463--485 (2017; Zbl 1388.11043)], the authors conclude that if \(C\) is a large absolute constant and \[ E(A_N) \ll N^3 / (\log N)^C \] then \(\mathcal{A}\) is metric Poissonian. The next question is whether any \(C > 1\) suffices here; we know from \textit{A. Walker}'s paper [Mathematika 64, No. 1, 230--236 (2018; Zbl 1429.11136)] that it fails for \(C=1\). The result applies in particular when all of the terms of \(\mathcal{A}\) are squares, by an estimate of \textit{T. Sanders} [Anal. PDE 5, No. 3, 627--655 (2012; Zbl 1320.11009)] .