Bounded Independence for $k$-Min-Wise Hashing: Tight Bounds and Limitations of Structured Hashing

Abstract

Min-wise hashing and its $k$-min-wise extension are fundamental tools in sampling, sketching, similarity estimation, etc. A standard approach to constructing such families is bounded independence. For ordinary min-wise hashing, the required degree of independence is fully understood: $Θ(\log 1/δ)$-wise independence is both sufficient and necessary. For $k$-min-wise hashing, however, the best previous result only showed that $O(k\log\log1/δ+\log1/δ)$-wise independence suffices, with no matching lower bound. We give a tight characterization of the amount of bounded independence required for $k$-min-wise hashing, proving that $Θ(k+\log 1/δ)$-wise independence is both sufficient and necessary. This improves the previous upper bound and provides a matching lower bound. Consequently, the standard construction of bounded-independent hash families has seed length $O\bigl((k+\log 1/δ)\cdot\log(N/δ)\bigr)$. In particular, for polynomially small $δ$ and any $Ω(\log N) \le k \le N^{1 - c}$, it achieves the optimal seed length $O(k\log N)$. We further investigate two standard low-independence hash families. For random affine functions over $\mathbb{F}_2$, which form a pairwise independent family, we show that the multiplicative error is $Ω(\log N)$ even for ordinary min-wise hashing. For simple tabulation hashing, which is $3$-wise independent and performs well for ordinary min-wise hashing, we show that it incurs a multiplicative error $Ω(N)$ for $k$-min-wise hashing whenever $k\ge 4$.