Nonnegative functions as squares or sums of squares (Q2368787)

In 1978 Fefferman and Phong stated (and sketchily proved) a lemma that any nonnegative \({\mathcal C}^{\infty}\) function in \({\mathbb R}^n\) is a sum of squares \({\mathcal C}^{1,1}\). In the reviewed paper it is proved for \(n\geq 4\) that such a regularity condition is sharp, namely, there exist nonnegative \({\mathcal C}^{\infty}\) functions \(f : {\mathbb R}^n \rightarrow {\mathbb R}\) that are not sums of squares of \({\mathcal C}^{2}\) functions. It is also proved that a general nonnegative \({\mathcal C}^{2}\) function of one variable has a \({\mathcal C}^{1}\) admissible square root, but no better regularity can be assured. If the function is \({\mathcal C}^{4}\) and its values at all its local minima are controlled it has a \({\mathcal C}^{2}\) admissible square root, but no better regularity can be assured. In both cases, increasing the regularity of the nonnegative function up to \({\mathcal C}^{\infty}\) does not provide a better result.