Polynomials vanishing on grids: The Elekes-R\'onyai problem revisited

In this paper we characterize real bivariate polynomials which have a small range over large Cartesian products. We show that for every constant-degree bivariate real polynomial $f$, either $|f(A,B)|=\Omega(n^{4/3})$, for every pair of finite sets $A,B\subset{\mathbb R}$, with $|A|=|B|=n$ (where the constant of proportionality depends on ${\rm deg} f$), or else $f$ must be of one of the special forms $f(u,v)=h(\varphi(u)+\psi(v))$, or $f(u,v)=h(\varphi(u)\cdot\psi(v))$, for some univariate polynomials $\varphi,\psi,h$ over ${\mathbb R}$. This significantly improves a result of Elekes and R\'onyai (2000). Our results are cast in a more general form, in which we give an upper bound for the number of zeros of $z=f(x,y)$ on a triple Cartesian product $A\times B\times C$, when the sizes $|A|$, $|B|$, $|C|$ need not be the same; the upper bound is $O(n^{11/6})$ when $|A|=|B|=|C|=n$, where the constant of proportionality depends on ${\rm deg} f$, unless $f$ has one of the aforementioned special forms. This result provides a unified tool for improving bounds in various Erd\H os-type problems in geometry and additive combinatorics. Several applications of our results to problems of these kinds are presented. For example, we show that the number of distinct distances between $n$ points lying on a constant-degree parametric algebraic curve which does not contain a line, in any dimension, is $\Omega(n^{4/3})$, extending the result of Pach and de Zeeuw (2013) and improving the bound of Charalambides (2012), for the special case where the curve under consideration has a polynomial parameterization. We also derive improved lower bounds for several variants of the sum-product problem in additive combinatorics.