Suppose a bounded-degree polynomial P vanishes on every combination αa+βb of two distinct points of a set A. Croot, Lev and Pach's trick is to build an ∣A∣×∣A∣ matrix out of the values Bab:=P(αa+βb): because P vanishes off the diagonal, this matrix is essentially diagonal, and diagonal matrices with many nonzero entries need high rank -- but the specific algebraic structure of P forces its rank to stay small no matter how large A is.
∣{a∈A:P(−γa)=0}∣≤2md/2
Detailed analysis
Ellenberg and Gijswijt (2017, Proposition 2, generalising Lemma 1 of Croot-Lev-Pach 2016) fix Fq elements α,β,γ with α+β+γ=0. If P∈Sn≤d satisfies P(αa+βb)=0 for every a=b∈A, then ∣{a∈A:P(−γa)=0}∣≤2md/2. The proof expands P(αx+βy): every monomial of total degree at most d has one of its two halves (x-part or y-part) of degree at most d/2, so P(αx+βy)=∑m∈Mnd/2m(x)Fm(y)+∑m∈Mnd/2m(y)Gm(x) for some polynomials Fm,Gm indexed by monomials m of degree at most d/2. Evaluating at x=a,y=b∈A writes the matrix B with entries Bab:=P(αa+βb) as a sum of 2md/2 rank-one matrices, so rank(B)≤2md/2.
Terms in this step
Rank of a matrix
The minimum number of rank-one matrices (outer products of a column and a row vector) needed to write a matrix as their sum; a diagonal matrix with k nonzero entries has rank exactly k.