New estimates of short Kloosterman sums (Q650291)

This posthumous paper is based on notes made in 2006 and 2007. The principal result shows that if \(p\) is prime and \(\varepsilon>0\) then \[ \sum_{n\leq X}e_p(an+b\overline{n})\ll_{\varepsilon} X(\log X)^{-1+\varepsilon} \] for any integers \(a,b\) not divisible by \(p\), and \(p^{\varepsilon}\leq X\leq p\). While it is disappointing that the saving is only about \(\log X\), the remarkable feature of this estimate is that it is non-trivial right down to \(p^{\varepsilon}\). The proof is based on bounds for bilinear Kloosterman sums of the shape \[ \sum_{U<u\leq U'}\sum_{V<v\leq V'}a_u b_v e_p(auv+b\overline{uv}), \] which are non-trivial unless both \(U\) and \(V\) are very small. Almost all \(n\leq X\) can be written as \(n=uv\) with \(u,v\) in ranges such as those above (though primes \(n\) clearly cannot). Thus one can imagine that a suitable combinatorial argument will handle those \(n\) with a suitable factorization, leaving only primes and ``near-primes'' to be estimated trivially. This rough argument shows why one cannot expect to do better than an estimate \(O(X(\log X)^{-1})\) by this method. The proof presented in the paper uses the above principle, and develops the necessary combinatorial procedure in full. The paper includes similar results for short Kloosterman sums involving an arithmetic weight. Thus for example there is a constant \(c>0\) such that one has \[ \sum_{n\leq X}\mu(n)e_p(an+b\overline{n})\ll_{\varepsilon} X(\log X)^c \] for \(p^{\varepsilon}\leq X\leq p^{4/7}\).