CUBIC FORMS MODULO ONE (Q4465115)

Let \(C(x_1,\dots, x_s)\in \mathbb{R}[x_1,\dots, x_s]\) be a cubic form. It was shown by \textit{R. J. Cook} [Acta Arith. 34, 91--102 (1978; Zbl 0372.10014)] that for any \(N\geq 1\) one can find a non-zero integer vector \((x_1,\dots, x_s)\) with \(\max| x_i|\leq N\) and such that \[ \| C(x_1,\dots, x_s)\|\ll_{\varepsilon, s} N^{-c(s)+\varepsilon}. \] In Cook's work the exponent \(c(s)\) takes the value \(c(s)= 1/4\) for \(1\leq s\leq 111\), and slowly increases to \(1\) for \(s\geq 112\). The value \(c(1)= 1/4\) is an earlier result of \textit{I. Danićić} [Mathematika, Lond. 3, 91--102 (1958; Zbl 0085.03304) or Proc. Lond. Math. Soc., III. Ser. 8, 161--176 (1958; Zbl 0081.27104)], and the results for \(s\leq 111\) are a trivial deduction from this. The present paper achieves non-trivial results as soon as \(s\geq 3\). Precisely, it shows that one may take \(c(1)= c(2)= 1/4\) and \[ c(s)= {s+ 4sc(t)\over 5s+ 4+ 16c(t)}\qquad (s\geq 3), \] where \(t= 1+ [s/2]\). Thus \(c(3)= 6/23\) for example. One still finds that \(c(s)\) tend slowly to \(1\), though somewhat faster than before. As with the work of Cook one applies a quasi-diagonalization process so that one can bring to bear techniques for diagonal forms. The present paper cleverly adapts a method from \textit{H. Davenport's} work [Philos. Trans. R. Soc. Lond., Ser. A 251, 193--232 (1959; Zbl 0084.27202)] on cubic forms in 32 variables to split off successive cube terms from the form. This process is considerably more efficient than that used by Cook.