Optimal Document Exchange and New Codes for Insertions and Deletions

Abstract

We give the first communication-optimal document exchange protocol. For any $n$ and $k < n$ our randomized scheme takes any $n$-bit file $F$ and computes a $\Theta(k \log \frac{n}{k})$-bit summary from which one can reconstruct $F$, with high probability, given a related file $F'$ with edit distance $ED(F,F') \leq k$. The size of our summary is information-theoretically order optimal for all values of $k$, giving a randomized solution to a longstanding open question of [Orlitsky; FOCS'91]. It also is the first non-trivial solution for the interesting setting where a small constant fraction of symbols have been edited, producing an optimal summary of size $O(H(\delta)n)$ for $k=\delta n$. This concludes a long series of better-and-better protocols which produce larger summaries for sub-linear values of $k$ and sub-polynomial failure probabilities. In particular, the recent break-through of [Belazzougui, Zhang; FOCS'16] assumes that $k < n^\epsilon$, produces a summary of size $O(k\log^2 k + k\log n)$, and succeeds with probability $1-(k \log n)^{-O(1)}$. We also give an efficient derandomized document exchange protocol with summary size $O(k \log^2 \frac{n}{k})$. This improves, for any $k$, over a deterministic document exchange protocol by Belazzougui with summary size $O(k^2 + k \log^2 n)$. Our deterministic document exchange directly provides new efficient systematic error correcting codes for insertions and deletions. These (binary) codes correct any $\delta$ fraction of adversarial insertions/deletions while having a rate of $1 - O(\delta \log^2 \frac{1}{\delta})$ and improve over the codes of Guruswami and Li and Haeupler, Shahrasbi and Vitercik which have rate $1 - \Theta\left(\sqrt{\delta} \log^{O(1)} \frac{1}{\epsilon}\right)$.