Schwartz-Zippel Lemma

August 08, 2024

Terminology

  • A polynomial is identically zero if it evaluates to zero for all elements of its domain.
  • The degree of a term of a multivariate polynomial is the sum of the degrees of each variable in the term.
  • A polynomial is multilinear if it is linear in each of its terms when each of the other terms is held fixed, in other words, each variable has degree 1 or 0.
  • The total degree of the polynomial is the degree of the highest degree term.
    • For an mm-variate multilinear polynomial, the total degree must be m\leq m because no variable can have degree >1\gt 1.

Schwartz-Zippel Lemma

  • Let F\mathbb{F} be any field and let g:FmFg: \mathbb{F}^m \rightarrow \mathbb{F} be a nonzero mm-variate polynomial of total degree d\leq d. Then on any finite set SFS \subseteq \mathbb{F}
    PrxSm[g(x)=0]dS\Pr_{x \leftarrow S^m}[g(x) = 0]\leq \frac{d}{|S|}

For a subset SFS \subseteq \mathbb{F}

  • The product set Sm=S×S×SS^m = S \times S \times S \dots from which xx is drawn contains Sm|S|^m values.
  • The probability that a g(x)=0g(x) = 0 for a random vector xSmx \in S^m is dS\leq \frac{d}{|S|}, so there are at most dSm1d \cdot |S|^{m-1} values of xSmx \in S^m where g(x)=0g(x) = 0.
  • It follows that any two polynomials pap_a and pbp_b of total degree d\leq d can agree on at most dS\frac{d}{|S|} fraction of the points in SmS^m, or dSm1d \cdot |S|^{m-1} values of xSmx \in S^m
    • papbp_a - p_b cannot be a polynomial with more than dSm1d \cdot S^{m-1} roots

For a Finite Field Fp\mathbb{F}_p

  • For an mm-variate polynomial g:FpmFpg: \mathbb{F}_p^m \rightarrow \mathbb{F}_p of total degree d\leq d, the probability that an mm-dimensional vector xFpx \in \mathbb{F}_p will be a root of gg is d/p\leq d/p.
    PrxFpm[g(x)=0]dFpPrxFpm[g(x)=0]dp\Pr_{x \in \mathbb{F}^m_p}[g(x) = 0]\leq \frac{d}{|\mathbb{F}_p|} \Rightarrow \Pr_{x \in \mathbb{F}^m_p}[g(x) = 0]\leq \frac{d}{p}
  • There are pmp^m possible mm-dimensional vectors in Fpm\mathbb{F}^m_p, so there are dpm1\leq d\cdot p^{m-1} roots of gg.
    pmdp=dpm1p^m \cdot \frac{d}{p} = d\cdot p^{m-1}

Univariate Case

  • Reed-Solomon and univariate Lagrange interpolation for an nn-dimensional vector both resulted in unique univariate polynomials of degree n1\leq n-1. Applying Schwartz-Zippel in the univariate case, we can verify their distance amplifying properties.
  • Let pap_a be the unique univariate extension of degree n1\leq n-1 for a vector aFpna \in \mathbb{F}_p^n and pbp_b be the the univariate extension of degree n1\leq n-1 for a different vector bFpnb \in \mathbb{F}_p^n. Then let q=pa=pbq = p_a=p_b be a degree n1\leq n-1 polynomial. By Schwartz-Zippel:
    PrxFp1[q(x)=0]n1p\Pr_{x \in \mathbb{F}^1_p}[q(x) = 0] \leq \frac{n-1}{p}
  • Thus, pap_a and pbp_b have an n1p\frac{n-1}{p} probability of agreeing at any point xFpx \in \mathbb{F}_p.

Profile picture

Written by Breads