Bilinear forms with Kloosterman sums

Philippe Michel, Will Sawin and I have just finished the first draft of a paper where we prove estimates for general bilinear forms of the type
\sum_{m\leq M}\sum_{n\leq N}\alpha_m\beta_n K(mn)
where (1) K(x)=\mathrm{Kl}_k(mn;p) is a (normalized) hyper-Kloosterman sum (for k=2, this is a classical Kloosterman sum) modulo a prime p; and (2) the ranges M and N are such that we have non-trivial bounds even if M=N is a bit smaller than \sqrt{p} in logarithmic scale. In other words, we obtain non-trivial results below the “Pólya-Vinogradov range”.

The basic strategy to get this result is not new: it was devised by Fouvry and Michel a number of years ago (inspired at least in part by earlier work of Friedlander-Iwaniec and by the Vinogradov-Karatsuba-style “shift” method to estimate certain short exponential sums). What was missing (despite the strong motivation provided by applications that were known to follow from such a result, one of which is described in a recent preprint of Blomer, Fouvry, Milicevic, Michel and myself) was a way to prove certain estimates for (complete) sums over finite fields, of the type

\sum_{r\in \mathbf{F}_p} R(r,\lambda_1,b_1,b_2,b_3,b_4)\overline{R(r,\lambda_2,c_1,c_2,c_3,c_4)}\ll p^{3/2}


R(r,\lambda,b_1,b_2,b_3,b_4)= \sum_{s\in\mathbf{F}_p}e\Bigl(\frac{\lambda s}{p}\Bigr)K(s(r+b_1))K(s(r+b_2))\overline{K(s(r+b_3))K(s(r+b_4))}

unless the parameters (\lambda_1,\lambda_2, b_i, c_i) are in some “diagonal” positions. And we cannot afford too many diagonal cases…

The main contribution of our paper (much of which comes from the ideas of Will!) is to find a relatively robust approach to such estimates.

This relies, as one can expect, from extensive algebraic-geometric arguments to apply the Riemann Hypothesis over finite fields. In fact, from this point of view, this paper is by far the most complicated I’ve ever been involved in. We use, among other things:

  • The Riemann Hypothesis over finite fields, in its most general version of Deligne — indeed, we use it multiple times;
  • The interpretation of the sum over s (in the sum R above) as itself a trace function of sbome sheaf on the space of parameters (\lambda,r,b_1,b_2,b_3,b_4); this follows from the formalism of étale cohomology, which is also used in many other ways (e.g., to detect irreducibility of sheaves by properties of the top-degree cohomology);
  • A very general version of the Euler-Poincaré characteristic formula in étale cohomology – this comes from SGA5;
  • The formalism and properties of vanishing and nearby cycles in étale cohomology, and in particular their relations with local monodromy representations of sheaves on curves;
  • The global \ell-adic Fourier transform of Deligne as well as the local Fourier transform of Laumon;
  • A special case of the homogeneous Fourier transform of Laumon (which we might be able to avoid, although with an argument involving perverse sheaves);
  • Katz’s theory of Kloosterman and hypergeometric sheaves, in particular with respect to the computation of their geometric monodromy groups (and its implication through the Goursat-Kolchin-Ribet Criterion), but also (and equally importantly) with respect to their local monodromy properties;
  • The diophantine criterion for geometric irreducibility (which is again a case of the Riemann Hypothesis)…

Many of these are results and ideas that I was aware of but had never actually used before, and I learnt a lot by seeing how Will exploited and combined them.  I will try to write a few more posts later to (attempt to) explain and motivate them (and how we use them) from an analytic nunber theorist’s viewpoint.  The theory of vanishing cycles, in particular, should have many more applications in extending the range of applicability of Deligne’s Riemann Hypothesis to problems in analytic number theory.

The paper is dedicated to Henryk Iwaniec, who has been over the years the most eloquent and powerful advocate for a deeper use of the work of Deligne (and Katz and others) in applications to analytic number theory.

Published by


I am a professor of mathematics at ETH Zürich since 2008.

5 thoughts on “Bilinear forms with Kloosterman sums”

  1. Nice! What do you think is the prognosis for using these sorts of methods to prove Zhang-type equidistribution results? As you know, in Polymath8a we identified a potential way to improve the level of distribution if one could control a certain complicated multiple sum of Kloosterman sums, but that looks to be a slightly different expression to what you have here.

    1. Hi Terry,

      Yes in fact together with Will we have the required bounds at for the sums that were left to do in Poymath8a from the ideas we had with Paul once at least the sums you get when you reduce everything to prime moduli. In fact this is a bit easier than the present paper. As you know there is a painful (but doable) process in passing from the pure prime to the highly composite case. Now that the bilinear sum paper is behind us one can turn to that line of applications and even investigate wilder transformations


  2. A slightly different question (which I may have asked you before at some point): regarding completed bilinear sums $\sum_{m \in {\bf F}_p} \sum_{n \in {\bf F}_p} \alpha_m \beta_n K(mn)$, is there hope of getting square root cancellation, i.e. a gain of $p^{-1/2+o(1)}$ as is the case with the Fourier sums $\sum_{m \in {\bf F}_p} \sum_{n \in {\bf F}_p} \alpha_m \beta_n e(mn/p)$? Taking the usual Cauchy-Schwarz one can presumably get $p^{-1/4+o(1)}$ cancellation, but in principle one can keep taking Cauchy-Schwarz again and again and push the gain all the way to the optimal square root level. Of course, the algebraic geometry becomes much worse when doing so…

    1. Yes, there is a lot of hope.

      The bound you mention for K(mn) with K a general trace function is due to Ping Xi in his thesis using exactly the method you describe and using Katz’s theory of multiplicative convolution to handle the algebraic geometry.

      Inspired by this, I worked out a generalization to K(f(m,n)) for an arbitrary polynomial f, and even to arbitrary sheaves and higher dimensions. It should essentially replace 1/4 with 1/2-epsilon in your algebraic regularity lemma while also generalizing from definable functions to trace functions (and precisely 1/2 in some cases). I have not finished writing it up, though.

  3. Off-topic: You mentioned a 2nd edition for your book with Iwaniec. Is there any news? An estimate for release date? Thanks. Paul.

Comments are closed.