The Euclidean distance degree of Fermat hypersurfaces (Q346561)

Finding the point on a surface that is closest to a given point \(u\) is an important problem with numerous applications. Often, it can be solved by numerical methods. However, these may not be sufficiently reliable in certain situations. If the surface is algebraic, an alternative is to compute the critical points of the squared distance function on the surface. This is an algebraic problem with finitely many solutions outside the surface's singular locus. The number of solutions for algebraic hypersurfaces of degree \(d\) in the complex extension of \(n\)-dimensional projective space is called the \textit{(Euclidean) distance degree}. For generic hypersurfaces, it has recently been computed as \(d\sum_{i=0}^{n-1}(d-1)^i\) [\textit{J. Draisma} et al., Found. Comput. Math. 16, No. 1, 99--149 (2016; Zbl 1370.51020)]. In non-generic cases it may be less. A natural next step is the investigation of special classes of surfaces. In this article, the author at first considers Fermat hypersurfaces in complex projective space, which are defined by the polynomial \(x_0^d+\cdots+x_n^d\). Here, the distance degree equals the generic hypersurface bound minus a sum over multiplies of \(\delta(m,p)\) -- the number of integer \(m\)-tupels \((t_1,\ldots,t_m)\) with \(1 \leq t_i \leq p\) such that \(1+\sum x^{2t_i}\) divides the \(p\)-th cyclotomic polynomial. A closed formula for \(\delta(m,p)\) seems to be unknown but it is a polynomial periodic function in \(p\) that can be efficiently computed. Interestingly, the difference between the respective distance degrees of general hypersurfaces and Fermat hypersurfaces can attain arbitrarily large values. For affine Fermat hypersurfaces, given by \(x_1^d + \cdots x_n^d = 1\), a similar formula for the affine distance degree is obtained. As usual, it counts the critical points over the complex numbers. The author provides the upper bound \(\sqrt{2}^{25n^2-3n+2}(n+2)^{5n}\) for the number of real solutions but conjectures that the actual upper bound may be just \(2^n-1\). Finally he shows that for a generic choice of complex numbers \(a_0,\ldots,a_n\), the distance degree of the scaled Fermat hypersurface \(x_0^d/a_0 + \cdots + x_n^d/a_n\) equals the distance degree for generic hypersurfaces.