Euclidean distance degree and mixed volume (Q2098233)

This pretty paper studies a special case of the algebraic Euclidean distance problem: given a polynomial \(f\in \mathbb{C} [x_1, \ldots , x_n]\) and a generic point \((u_1, \ldots , u_n) \in \mathbb{C}^n,\) determine the number of critical points of the function that computes squared Euclidean distance from \(u\) to the hypersurface \(V(f).\) Here, ``Euclidean distance function'' refers to the complexified distance on \(\mathbb{R}^n,\) rather than the Hermitian distance, so this function is given by \begin{align*} d_u : V(f) &\to \mathbb{C} \\ (x_1, \ldots , x_n) &\mapsto (x_1-u_1)^2 + \cdots + (x_n - u_n)^2. \end{align*} The number of such critical points is known as the Euclidean distance degree of \(V(f),\) and denoted \(\operatorname{EDD} (f)\) by the authors. When \(f\) is generic in the space of all degree-\(d\) polynomials, it is well-known from prior work that \[ \operatorname{EDD} (f) = d \, \displaystyle\sum_{i=0}^{n-1} (d-1)^i. \] The authors' main result is a refinement of this formula to the case where \(f\) is generic in the space of all polynomials with a fixed monomial support set \(\mathcal{A} \subset \mathbb{Z}_{\ge 0}^{n}\) that contains the origin (ie, \(f\) has a nonzero constant term.) They prove that \(\operatorname{EDD} (V(f)) = \operatorname{MV} (P, P_1, \ldots , P_n),\) where \begin{itemize} \item[(1)] \(P = \operatorname{conv} (\mathcal{A})\) is the Newton polytope of \(f\), \item[(2)] \(P_1, \ldots , P_n\) are the Newton polytopes of certain Lagrange multiplier equations whose supports are governed by \(\mathcal{A},\) and \item[(3)] \(\operatorname{MV}\) denotes the mixed volume. \end{itemize} They conclude by evaluating their formula when \(P\) is a rectangular parallelepiped. The result, in some sense, is ``expected'' due to a celebrated theorem of Bernstein. However, the conclusion is non-trivial: for \(f\) that \textit{do not} satisfy certain genericity conditions, we may have a strict inequality \(\operatorname{EDD} (V(f)) < \operatorname{MV} (P, P_1, \ldots , P_n)\). These genericity conditions turn out to be the main tool used in the proof. If the inequality is strict, then Bernstein's ``other'' theorem states that a solution must exist to some \textit{facial system} associated to the Lagrangian equations. When \(f\) is sufficiently generic, the authors prove these facial systems have no solutions.