Recursive Sketching For Frequency Moments

TL;DR

This paper introduces a recursive sketching technique that improves the space complexity for computing frequency moments in data streams, reducing poly-logarithmic factors and requiring only limited independence.

Contribution

The authors present a simple recursive sketching method that achieves near-optimal space complexity for frequency moments, surpassing previous poly-logarithmic factor bounds.

Findings

Achieves a space complexity of $O( ext{log}(m) ext{log}(nm)( ext{log} ext{log} n)^4 n^{1-2/k})$

Requires only 4-wise independence, simplifying implementation

Approaches the iterated logarithm $ ext{log}^* n$ in complexity

Abstract

In a ground-breaking paper, Indyk and Woodruff (STOC 05) showed how to compute $F_k$ (for $k>2$) in space complexity $O(\mbox{\em poly-log}(n,m)\cdot n^{1-\frac2k})$, which is optimal up to (large) poly-logarithmic factors in $n$ and $m$, where $m$ is the length of the stream and $n$ is the upper bound on the number of distinct elements in a stream. The best known lower bound for large moments is $\Omega(\log(n)n^{1-\frac2k})$. A follow-up work of Bhuvanagiri, Ganguly, Kesh and Saha (SODA 2006) reduced the poly-logarithmic factors of Indyk and Woodruff to $O(\log^2(m)\cdot (\log n+ \log m)\cdot n^{1-{2\over k}})$. Further reduction of poly-log factors has been an elusive goal since 2006, when Indyk and Woodruff method seemed to hit a natural "barrier." Using our simple recursive sketch, we provide a different yet simple approach to obtain a $O(\log(m)\log(nm)\cdot (\log\log n)^4\cdot n^{1-{2\over k}})$ algorithm for constant $\epsilon$ (our bound is, in fact, somewhat stronger, where the $(\log\log n)$ term can be replaced by any constant number of $\log $ iterations instead of just two or three, thus approaching $log^*n$. Our bound also works for non-constant $\epsilon$ (for details see the body of the paper). Further, our algorithm requires only $4$-wise independence, in contrast to existing methods that use pseudo-random generators for computing large frequency moments.