First Steps in Algorithmic Fewnomial Theory

TL;DR

This paper explores the computational complexity of deciding real roots of polynomial systems, identifying phase transitions in NP-hardness, and introduces new polynomial-time decidable classes related to A-discriminants and bounds on zero sets.

Contribution

It establishes the complexity thresholds for the feasibility problem of polynomial systems with a fixed number of monomials and introduces new polynomial-time algorithms for certain A-discriminant sign decisions.

Findings

NP-hardness for m > n+2 monomials

Polynomial-time algorithms for m <= n+2 monomials

New bounds on real zero sets of (n+2)-nomial systems

Abstract

Fewnomial theory began with explicit bounds -- solely in terms of the number of variables and monomial terms -- on the number of real roots of systems of polynomial equations. Here we take the next logical step of investigating the corresponding existence problem: Let FEAS_R denote the problem of deciding whether a given system of multivariate polynomial equations with integer coefficients has a real root or not. We describe a phase-transition for when m is large enough to make FEAS_R be NP-hard, when restricted to inputs consisting of a single n-variate polynomial with exactly m monomial terms: polynomial-time for m<=n+2 (for any fixed n) and NP-hardness for m<=n+n^{epsilon} (for n varying and any fixed epsilon>0). Because of important connections between FEAS_R and A-discriminants, we then study some new families of A-discriminants whose signs can be decided within polynomial-time. (A-discriminants contain all known resultants as special cases, and the latter objects are central in algorithmic algebraic geometry.) Baker's Theorem from diophantine approximation arises as a key tool. Along the way, we also derive new quantitative bounds on the real zero sets of n-variate (n+2)-nomials.