On the Riesz means of \(\frac{n}{\varphi(n)}\). II (Q477312)

In this sequence of three closely linked papers, the authors prove bounds for the error term \(E_{k}(x)\) in the asymptotic \[ \frac{1}{k!} \sum_{n \leq x} \frac{n}{\varphi(n)} (1 - \frac{n}{x})^{k} = c_{1}(k)x + c_{2}(k)\log x + c_{3}(k) + E_{k}(x) , \] where \(k\) is a natural number and \(\varphi(n)\) is the Euler totient function. \textit{R. Sitaramachandrarao} [Rocky Mt. J. Math. 15, 579-588 (1985; Zbl 0584.10027)] conjectured that \(E_{1}(x) \ll_{\delta} x^{-3/4 + \delta}\) (and proved the estimate \(\ll x^{-1/5}\)), and in Part I [the authors, Hardy-Ramanujan J. 36, 8--20 (2013; Zbl 1358.11108)], the authors conjecture this should hold uniformly for \(E_{k}(x)\) for all natural numbers \(k\). The authors also prove in Part I that \[ E_{k}(x) \ll_{\delta} x^{-1/2 + \delta}, \] and in this paper they refine this to the estimates \[ E_{1}(x) \ll \frac{\log^{5/4}x \log\log x}{x^{1/2}}, \qquad E_{k}(x) \ll_{k} \frac{\log^{1/4}x \log\log x}{x^{1/2}} \quad \text{ for all } k \geq 2 . \] (The result is actually more precise, as discussed below.) Finally, in Part III [the authors, Acta Arith. 170, No. 3, 275--286 (2015; Zbl 1358.11110)] they show that \[ E_{k}(x) \ll \frac{1}{k x^{1/2}} \exp(-c \frac{\log^{1/3}x}{(\log\log x)^{1/3}}) \quad \text{ for all } k \geq 2 , \] and they prove the conjecture for \(k \geq 2\) assuming the Riemann Hypothesis. The proofs follow a classical contour integration approach. In Part I [Zbl 1358.11108], the line of integration is moved to \(\mathrm{Re}(s) = -1/2 + \delta\), and the argument can be completed using straightforward bounds together with the functional equation and the second moment of the zeta function. In this paper, the line of integration is moved to \(\mathrm{Re}(s)=-1/2\), and the argument completed using Vinogradov-Korobov pointwise bounds for \(\zeta(1+it)\) and \(1/\zeta(1+it)\), together with an estimate for the number of large values of \(1/\zeta(1+it)\). In Part III [Zbl 1358.11110], the line of integration is moved even further, to exploit the Vinogradov-Korobov zerofree region or (for the conditional result) the Riemann Hypothesis. Reviewer's remarks: The form of some of the bounds in these papers seems, to the reviewer, perhaps a little strange. Thus, the authors actually assert and show in this paper that, for any real \(c^{*} \geq 10\) and any integer \(k \geq 2\), one has \[ E_{k}(x) \ll_{\varepsilon} \frac{\max(4^{k},c^{*(2/3 + \varepsilon)}) \log x}{x^{c^* k -1}} + c^{* 1/2} \frac{\log^{1/4}x \log\log x}{e^{k} x^{1/2}} . \] But this bound seems, to the reviewer, to be optimised for all \(k\) (and dominated by the second term) by taking \(c^* = 10\), provided \(x\) is larger than an absolute constant. The factor \(1/k\) in the bound from Part III [loc. cit.] also seems as though it could be replaced by \(e^{-k}\), as in this paper, since in the proof the authors replace an integral roughly like \(\int_{T_0}^{T} \frac{dt}{t^{k - 1/6}}\) by an integral \(\int_{1}^{T} \frac{dt}{t^{k - 1/6}}\), whereas keeping the original provides an extra saving of the shape \(T_{0}^{-k}\) (where \(T_0\) here is a large constant).