On Hilbert's 17th problem in low degree (Q2360118)

Let \(\mathbb R\) be a real closed field and let \(n\) be a positive integer. By a famous theorem of \textit{A. Pfister} [Invent. Math. 4, 229--237 (1967; Zbl 0222.10022)], it is known that every positive semidefinite polynomial in \(\mathbb R [x_1,\ldots,x_n]\) can be expressed as a sum of at most \(2^n\) squares of rational functions in \(\mathbb R (x_1,\ldots,x_n)\). Cassels, Ellison and Pfister [\textit{J. W. S. Cassels} et al., J. Number Theory 3, 125--149 (1971; Zbl 0217.04302)] gave an example of a positive semidefinite real polynomial in \(n=2\) variables of degree \(3n=6\) which cannot be written as a sum of \(2^n-1=3 \) squares of rational functions over \(\mathbb R\). Thus, the bound \(2^n\) for the number of squares required is in general the best possible. On the other hand, \textit{D. Hilbert} [Math. Ann. 32, 342--350 (1888; JFM 20.0198.02)] had already proven that for positive semidefinite real polynomials in \(n=2\) variables of degree \(d\leq 2n = 4\), \(2^n-1=3\) squares (even of polynomials) suffices. It is this theme that is taken up in the paper under review; that is, to show that when the degree \(d\) is not too large in relation to \(n\), every positive semidefinite polynomial in \(\mathbb R [x_1\ldots,x_n]\) of degree \(d\) is a sum of \(2^n-1\) squares of rational functions in \(\mathbb R(x_1,\ldots, x_n)\). In particular, this is shown to be the case whenever \(d\leq 2n-2\), or when \(d=2n\) and either \(n\) is even, \(n=3\), or \(n=5\). The inequality \(d\leq 2n\) has a natural geometric interpretation, reflecting the rational connectedness of an associated algebraic variety. The strategy used to prove these results originates in work of \textit{J. L. Colliot-Thélène} [Arch. Math. 58, No. 4, 392--396 (1992; Zbl 0738.14023); Compos. Math. 86, No. 2, 235--243 (1993; Zbl 0774.12002)]. The idea is to consider the homogenization \(F\) of the polynomial \(f\), introduce the algebraic surface \(Y:=\{Z^2+F=0\}\), and then translate the question of whether or not \(f\) can be written as a sum of \(2^n-1\) squares in \(\mathbb R (x_1,\ldots,x_n)\) into a cohomological property of the variety \(Y\).