Algebraic trace functions over the primes (Q2510816)

In this paper the authors establish power-saving bounds for sums of the form \[ \sum_{q \; \text{prime}} K(q) V(q/X) , \qquad \sum_{n} \mu(n) K(n) V(n/X) , \] where \(K(\cdot)\) is an \textit{isotypic trace weight} mod \(p\) that is not \textit{exceptional}, and where (roughly speaking) \(V(\cdot)\) is a smooth function approximating the indicator function \(\mathbf{1}_{[0,1]}\), and \(X\) is not too much smaller than \(p\). The authors explain that an isotypic trace weight is a complex-valued version of the trace on an isotypic trace sheaf modulo \(p\), which they further explain is ``\dots a geometrically isotypic \(\ell\)-adic Fourier sheaf\dots [in the sense of Katz] \dots which is pointwise pure of weight 0''. It is beyond the competence of this reviewer to comment further on this, but it perhaps suffices here to note that trace weights include additive characters or Dirichlet characters evaluated at polynomial arguments, products of such functions, Kloosterman sums, hyper-Kloosterman sums, and many other interesting special cases. A trace weight is said to be exceptional if it is proportional to \(\chi(n) \Psi(n)\) for some Dirichlet character \(\chi\) and some additive character \(\Psi\) (mod \(p\)). As the authors note, the restriction of their results to non-exceptional \(K\) is essentially saying that they cannot prove anything that would correspond to a zero-free strip for Dirichlet \(L\)-functions (which is unsurprising since their proofs use a Type I/Type II sums method, and in general such methods show that the primes do not destroy cancellation rather than showing that they produce it). Nevertheless the authors describe several nice applications of their bounds, for example to the average size of the error term in the prime number theorem in progressions when averaging over classes \(a\) which are values of polynomials. As mentioned, the proofs use combinatorial decompositions of the von Mangoldt function and of \(\mu(n)\) (in this case Heath-Brown's identity) to reduce matters to bounding sums of the form \[ \sum_{m_1 , \ldots, m_k} \alpha_1(m_1) \ldots \alpha_k(m_k) \sum_{n_1 , \ldots, n_k} V_1(n_1) \ldots V_k(n_k) V(\frac{m_1 \ldots m_k n_1 \ldots n_k}{X}) K(m_1 \ldots m_k n_1 \ldots n_k) , \] where the \(\alpha_i\) are arithmetic functions that should be considered fairly arbitrary (apart from having restricted support), and the \(V_i\) are smooth functions. It is at this point (possibly after an application of Cauchy-Schwarz to remove some of the weights \(\alpha_i\)) that deep ingredients from algebraic geometry, as well as some spectral theory and the amplification method, are required. Several of the relevant estimates are imported from another paper of the authors [\textit{É. Fouvry} et al., Geom. Funct. Anal. 25, No. 2, 580--657 (2015; Zbl 1344.11036)], although additional work is required here and occupies the final section of the paper. The authors note that one important aspect is that a twisted divisor function which arises is treated ``\dots in a fully automorphic manner'' as the Fourier coefficient of a certain Eisenstein series, rather than by more elementary means. The paper is certainly not a self-contained account of trace weights, but is probably a good place for a reader with a more analytic background to start to become familiar with them. It includes a very nicely written 15 page introduction which explains the results, the ideas of their proofs, and many connections to other results and questions.