À la recherche de petites sommes d'exponentielles. (In search of small exponential sums) (Q5956590)

Let \[ S(m,n)= \sum_{t\pmod n} e(mf(t)/n)) \] be the exponential sum associated with a fixed rational function \(f(X)\). If \(f\) is nonconstant, Weil's estimates show that \(S(m,n)\ll n^{1/2+\varepsilon}\) as \(n\to\infty\). Although this bound is essentially best possible, it is conjectured that \(S(1,n)\), for example, can be considerably smaller than \(n^{1/2}\) for suitable \(n\). The present paper establishes this for a wide variety of functions \(f\). This is done by taking \(n\) as a product of two primes. In the case of the classical Kloosterman sum, for example, one has \[ |K(1,1;pq)|\leq (\sqrt{pq})^{5/6+ \varepsilon} \] for infinitely many pairs \(p,q\), for any \(\varepsilon> 0\). The proof begins with the observation that \[ S(1,pq)= S(\overline{p},q) S(\overline{q},p). \] Hence, it suffices for \(S(\overline{p},q)\) to be small. Taking \(q\) to be fixed and allowing \(p\) to vary, this becomes a problem of ``horizontal'' distribution -- the more tractable type.