Towards the exact nature of a certain error term (Q2277512)

For a square-free natural number \(k\), let \(\delta_ k\) be the multiplicative arithmetic function defined by \(\delta_ k(n)=\max \{d\in\mathbb N: d\mid n\), \((d,k)=1\}\). For the error term \(E_ k(x)=\sum_{n\leq x}\delta_ k(n)-kx^ 2/2\sigma (k),\) it is proved here that if \(k=p_ 1\cdots p_ r\) where \(p_ 1<\ldots<p_ r\) are primes, then \[ \limsup_{x\to \infty}\frac{| E_ k(x)|}{x}\leq \frac{p_ 1}{p_ 1+1}2^{r-1}-\sum^{r}_{j=2}\frac{p_ 1p_ 2\cdots p_{j-1}}{(p_ 1+1)(p_ 2+1)\cdots (p_ j+1)}2^{r-j}.\tag{1} \] This improves a result of the author and \textit{R. Balasubramanian} [Arch. Math. 56, 37--40 (1991; Zbl 0686.10033)]. It is also proved that for general square-free \(k\), no significant improvement of (1) is possible.