The Furstenberg-Sárközy theorem for polynomials in one or more prime variables (Q6970000)

For \(X \subseteq \mathbb Z\) and \(N \in \mathbb N\), let \N\[ D(X,N) = \max \{ |A|: A \subseteq \{ 1, 2, \ldots, N \}, \quad (A - A) \cap X \subseteq \{ 0 \} \}. \] \NIn the case of \(S\) the set of perfect squares, Furstenberg and Sárközy showed independently -- and famously -- that \[D(S,N) = o(N).\] A major objective has been to strengthen this bound. A secondary objective has been to prove generalisations and analogues of such results.\N\NWhen the article under review was written, the strongest upper bound on \(D(S,N)\) was demonstrated by \N\textit{T. F. Bloom} and \textit{J. Maynard} [Compos. Math. 158, No. 8, 1777--1798 (2022; Zbl 1514.11019)], \Nwho showed that \N\[ D(S,N) \ll N(\log N)^{-c \log \log \log N}, \] \Nfor some absolute constant \(c > 0\). This has very recently been sharpened to \N\[ D(S,N) \ll N \exp(-c \sqrt{\log N}) \] \Nby \textit{B. J. Green} and \textit{M. Sawhney} [``New bounds for the Furstenberg-Sárközy theorem'', Preprint, \url{arXiv:2411.17448}]. \NA natural generalisation is to \(h(\mathbb Z)\), where \(h\) is an intersective polynomial, i.e., a polynomial with a root to every modulus. A variant of this is the case of \(h(\mathcal P)\), where \(\mathcal P\) is the set of primes, and \(h\) is \(\mathcal P\)-intersective (for every modulus \(d\), the polynomial has a root that is coprime to \(d\)). In these two settings, the best results to date are by \N\textit{N. Arala} [Funct. Approximatio, Comment. Math. 71, No. 2, 271--296 (2024; Zbl 1562.11019)] and \N\textit{A. Rice} [Acta Arith. 157, No. 1, 69--89 (2013; Zbl 1327.11011)], respectively. In a case of the latter setting, \N\textit{B. Green} [J. Am. Math. Soc. 37, No. 4, 1121--1201 (2024; Zbl 1564.11116)] recently showed that \N\[ D(\mathcal P - 1, N) \ll N^{1-c}. \]\N\NNow Doyle and Rice produce similar results for multivariable polynomials. The image enlarges when more variables are introduced, so the point is to obtain stronger bounds. One then needs to avoid pathologies, e.g., \((x+y)^2\) would be univariate after a change of variables. These situations can be detected geometrically, and the authors make such an assumption at co-finitely many prime reductions. The authors then establish analogues of the aforementioned results. When \(\ell \ge 3\), they get \N\[ D(h(\mathbb Z^\ell), N) \ll_h N \exp(-c \sqrt{\log N}) \] and \N\[ D(h(\mathcal P^\ell), N) \ll_h N \exp(-c (\log N)^{1/4}). \] \NThe case \(\ell = 2\) is also treated, but the exponents are poorer and messier. In the case \(\ell = 1\), the authors show that \N\[ D(h(\mathcal P), N) \ll N (\log N)^{-c \log \log \log N}, \] \Nwhich sharpens the previous record by the second author.