Pairs of diagonal quadratic forms and linear correlations among sums of two squares (Q2353615)

Let \[ q_1(X_1,\ldots,X_6)=a(X_1^2+X_2^2)+a'(X_3^2+X_4^2) \] and \[ q_2(X_1,\ldots,X_6)=b(X_1^2+X_2^2)+b'(X_3^2+X_4^2)+b''(X_5^2+X_6^2) \] be quadratic forms with rational integer coefficients satisfying \(ab'-a'b\not=0\). The goal of this paper is to investigate the sum \[ S(B)=\sum_{\substack{ \mathbf{x}\in\mathbb{Z}^6\\ 2\nmid Q_1(\mathbf{x})\\ Q_2(\mathbf{x})=0}} r(Q_1(\mathbf{x})) W(B^{-1}\mathbf{x}) \] as \(B\rightarrow\infty\). Here \(W\) is a fixed smooth weight function. In effect this sum counts rational points in \(\mathbb{P}^7\) lying on a certain singular intersection of two quadrics. An asymptotic formula for a somewhat more general singular intersection of quadrics in \(\mathbb{P}^n\), for any \(n\geq 8\), was achieved in the authors' recent work [Compos. Math. 149, No. 9, 1457--1494 (2013; Zbl 1350.11047)]. This is the motivation for the present paper, the challenge being to handle quadratics in as few variables as possible. The sum \(S(B)\) can also be viewed as counting quadruples of positive integers \((n_1,n_2,n_3,n_4)\), lying on a line in \(\mathbb{P}^3\), with weight \(r(n_1)r(n_2)r(n_3)r(n_4)\). In this formulation the problem was considered by the reviewer [Lond. Math. Soc. Lect. Note Ser. 303, 133--176 (2003; Zbl 1161.11387)], who established an asymptotic formula saving a small power of \(\log B\) (with a slightly restricted type of weight function). A more general approach by \textit{L. Matthiesen} [Acta Arith. 154, No. 3, 235--306 (2012; Zbl 1294.11169)] handles arbitrary \(k\)-tuples \((n_1,\ldots,n_k)\), but gives no explicit error term. In contrast to these works, the present paper gives a small power saving, showing that \[ S(B)=cB^4+O(B^{4-1/160+\varepsilon}) \] for any fixed \(\varepsilon>0\). As in the authors' earlier work one starts with the relation \(\tfrac14 r(n)=(1*\chi_4)(n)\) and uses Dirichlet's hyperbola trick, producing sums \[ \sum_{\substack{ \mathbf{x}\in\mathbb{Z}^n\\ d\mid Q_1(\mathbf{x})\\ Q_2(\mathbf{x})=0}} W(B^{-1}\mathbf{x}), \] for various \(d\ll B\). As before one then imposes the condition \(d\mid Q_2(\mathbf{x})\), and picks out terms with \(Q_2(\mathbf{x})/d=0\) via the \(\delta\)-method of Duke, Friedlander and Iwaniec [\textit{W. Duke} et al., Invent. Math. 112, No. 1, 1--8 (1993; Zbl 0765.11038)], as developed by the reviewer [J. Reine Angew. Math. 481, 149--206 (1996; Zbl 0857.11049)]. This procedure leads to sums \[ S_{d,q}(\mathbf{m})=\sum_{\substack{ 1\leq a\leq q\\ (a,q)=1}} \; \sum_{\substack{ \mathbf{k}\pmod{dq}\\ d\mid Q_i(\mathbf{k}), (i=1,2)}} e_{dq}(aQ_2(\mathbf{k})+\mathbf{m}.\mathbf{k}), \] averaged over \(d\) and \(q\). The strategy in the earlier paper was to sum non-trivially over \(q\), but trivially over \(d\). The behaviour of \(S_{d,q}(\mathbf{m})\) with respect to \(q\) only involves the quadratic form \(Q_2\), so this is relatively straightforward. In the present paper one reverses the procedure and averages non-trivially over \(d\), but trivially with respect to \(q\). This is advantageous because \(d\) typically ranges over a larger set than \(q\) does. However one now has a problem in which both forms \(Q_1\) and \(Q_2\) appear, and it is only in the special case described that the authors are able to control the sums sufficiently for a non-trivial estimation to be obtained.