Faster Parallel Algorithm for Approximate Shortest Path

We present the first $m\,\text{polylog}(n)$ work, $\text{polylog}(n)$ time algorithm in the PRAM model that computes $(1+\epsilon)$-approximate single-source shortest paths on weighted, undirected graphs. This improves upon the breakthrough result of Cohen~[JACM'00] that achieves $O(m^{1+\epsilon_0})$ work and $\text{polylog}(n)$ time. While most previous approaches, including Cohen's, leveraged the power of hopsets, our algorithm builds upon the recent developments in \emph{continuous optimization}, studying the shortest path problem from the lens of the closely-related \emph{minimum transshipment} problem. To obtain our algorithm, we demonstrate a series of near-linear work, polylogarithmic-time reductions between the problems of approximate shortest path, approximate transshipment, and $\ell_1$-embeddings, and establish a recursive algorithm that cycles through the three problems and reduces the graph size on each cycle. As a consequence, we also obtain faster parallel algorithms for approximate transshipment and $\ell_1$-embeddings with polylogarithmic distortion. The minimum transshipment algorithm in particular improves upon the previous best $m^{1+o(1)}$ work sequential algorithm of Sherman~[SODA'17]. To improve readability, the paper is almost entirely self-contained, save for several staple theorems in algorithms and combinatorics.