On rounding error sums related to the circle problem (Q1603249)

The classical lattice point problems of Gauss and Dirichlet lead in a natural way to the estimation of sums \[ S_1:= \sum_{\sqrt{t/2}<n\leq\sqrt{t}} \psi\left(\sqrt{t-n^2}\right) \qquad \text{and} \qquad S_2:= \sum_{1\leq n\leq\sqrt{t}} \psi\left({t\over n}\right) \] where \(\psi\) is the rounding error function (fractional part minus \({1\over 2}\)). For both sums, the upper bound \(O(t^{23/73+\varepsilon})\) has been established by M. Huxley. \medskip The main question of the present paper is how far the range of summation in these sums can be extended with this estimate still valid. It turns out (Theorem 1) that in \(S_1\) one may sum over any subinterval of \([-\sqrt{t},\sqrt{t}]\). Further, the efficiency of the old Van der Corput's method is compared with Huxley's. Concerning \(S_2\) the author shows that one can even go a bit further (any subinterval of \([1,t^{2/3}]\) leads to a bound \(O(t^{1/3})\)), but that for an interval of order \(t\) the sum is definitively \(\gg t\). \smallskip The method of proof is based on a tricky application of the idea that lattice points in a planar domain can be counted in two different ways.