Streaming Algorithms with Large Approximation Factors

We initiate a broad study of classical problems in the streaming model with insertions and deletions in the setting where we allow the approximation factor $\alpha$ to be much larger than $1$. Such algorithms can use significantly less memory than the usual setting for which $\alpha = 1+\epsilon$ for an $\epsilon \in (0,1)$. We study large approximations for a number of problems in sketching and streaming and the following are some of our results. For the $\ellp$ norm/quasinorm $|x|p$ of an $n$-dimensional vector $x$, $0 < p \le 2$, we show that obtaining a $\poly(n)$-approximation requires the same amount of memory as obtaining an $O(1)$-approximation for any $M = n^{{\Theta(1)}$.} For estimating the $\ellp$ norm, $p > 2$, we show an upper bound of $O(n^{1-2/p} (\log n \allowbreak \log M)/\alpha^{2})$ bits for an $\alpha$-approximation, and give a matching lower bound, for almost the full range of $\alpha \geq 1$ for linear sketches. For the $\ell2$-heavy hitters problem, we show that the known lower bound of $\Omega(k \log n\log M)$ bits for identifying $(1/k)$-heavy hitters holds even if we are allowed to output items that are $1/(\alpha k)$-heavy, for almost the full range of $\alpha$, provided the algorithm succeeds with probability $1-O(1/n)$. We also obtain a lower bound for linear sketches that is tight even for constant probability algorithms. For estimating the number $\ell_0$ of distinct elements, we give an $n^{{1/t}$-approximation} algorithm using $O(t\log \log M)$ bits of space, as well as a lower bound of $\Omega(t)$ bits, both excluding the storage of random bits.