Estimating Random-Walk Probabilities in Directed Graphs

Abstract: We study discounted random walks in a directed graph. In each vertex, the walk will either terminate with some probability $\alpha$, or continue to a random out-neighbor. We are interested in the probability $\pi(s,t)$ that such a random walk starting in $s$ ends in $t$. We wish to, with constant probability, estimate $\pi(s, t)$ within a constant relative error, unless $\pi(s, t) < \delta$ for some given threshold $\delta$. The current status is as follows. Algorithms with worst-case running time $\tilde O(m)$ and $O(1/\delta)$ are known. A more complicated algorithm is known, which does not perform better in the worst case, but for the average running time over all $n$ possible targets $t$, it achieves an alternative bound of $O(\sqrt{d/\delta})$. All the above algorithms assume query access to the adjacency list of a node. On the lower bound side, the best-known lower bound for the worst case is $\Omega(n^{1/2}m^{1/4})$ with $\delta \leq 1/(n^{1/2}m^{1/4})$, and for the average case it is $\Omega(\sqrt{n})$ with $\delta \leq 1/n$. This leaves substantial polynomial gaps in both cases. In this paper, we show that the above upper bounds are tight across all parameters $n$, $m$ and $\delta$. We show that the right bound is $\tilde\Theta(\min{m, 1/\delta})$ for the worst case, and $\tilde\Theta(\min{m, \sqrt{d/\delta}, 1/\delta})$ for the average case. We also consider some additional graph queries from the literature. One allows checking whether there is an edge from $u$ to $v$ in constant time. Another allows access to the adjacency list of $u$ sorted by out-degree. We prove that none of these access queries help in the worst case, but if we have both of them, we get an average-case bound of $\tilde \Theta(\min{m,\sqrt{d/\delta}, (1/\delta)^{2/3}})$.