Vanishing Spaces of Random Sets and Applications to Reed-Muller Codes
We study the following natural question on random sets of points in $\mathbb{F}_2^m$: Given a random set of $k$ points $Z=\{z_1, z_2, \dots, z_k\} \subseteq \mathbb{F}_2^m$, what is the dimension of the space of degree at most $r$ multilinear polynomials that vanish on all points in $Z$? We show that, for $r \leq γm$ (where $γ> 0$ is a small, absolute constant) and $k = (1-ε) \cdot \binom{m}{\leq r}$ for any constant $ε> 0$, the space of degree at most $r$ multilinear polynomials vanishing on a random set $Z = \{z_1,\ldots, z_k\}$ has dimension exactly $\binom{m}{\leq r} - k$ with probability $1 - o(1)$. This bound shows that random sets have a much smaller space of degree at most $r$ multilinear polynomials vanishing on them, compared to the worst-case bound (due to Wei (IEEE Trans. Inform. Theory, 1991)) of $\binom{m}{\leq r} - \binom{\log_2 k}{\leq r} \gg \binom{m}{\leq r} - k$. Using this bound, we show that high-degree Reed-Muller codes ($\text{RM}(m,d)$ with $d > (1-γ) m$) "achieve capacity" under the Binary Erasure Channel in the sense that, for any $ε> 0$, we can recover from $(1 - ε) \cdot \binom{m}{\leq m-d-1}$ random erasures with probability $1 - o(1)$. This also implies that $\text{RM}(m,d)$ is also efficiently decodable from $\approx \binom{m}{\leq m-(d/2)}$ random errors for the same range of parameters.