New bounds for $\mathrm{GL}(3)\times\mathrm{GL}(2)$ shifted convolution sums

Abstract

Shifted convolution sums of $\mathrm{GL}(3)\times\mathrm{GL}(2)$ type have been extensively studied in several important settings, while a number of natural cases remain open. In this paper, we first revisit the shifted convolution sums involving $λ_π(1,n)$ and the divisor function $d(n)$, improving the recent power saving of $1/42$ obtained by Blomer and Li to $1/22$. We also treat the corresponding sums involving $λ_π(1,n)$ and $r(n,\mathcal{Q})$, settling a previously open case and obtaining a power saving of $1/22$. Here $λ_π(1,n)$ denotes the normalized Fourier coefficients of a Hecke--Maass cusp form $π$ on $\mathrm{SL}(3,\mathbb{Z})$, and $r(n,\mathcal{Q})$ denotes the number of representations of a positive integer $n$ by a positive definite integral binary quadratic form $\mathcal{Q}$. The proof combines two circle methods with a different organization of the major-arc analysis. A key point is that the nonzero Poisson frequencies are kept together in a single coefficient vector, rather than estimated separately, which allows them to be controlled by a finite matrix norm. The zero Poisson frequency is treated separately: its contribution is independent of the reduced residue class, and its arithmetic mean cancels the constant term in the complementary Jutila weight before absolute values are taken. We make this cancellation explicit, including the relevant normalization and local factors. The same mechanism extends to the binary quadratic-form setting, while the remaining Jutila contribution is handled by the available correlation estimates.