A Tight Lower Bound on Distributed Random Walk Computation

We consider the problem of performing a random walk in a distributed network. Given bandwidth constraints, the goal of the problem is to minimize the number of rounds required to obtain a random walk sample. Das Sarma et al. [PODC'10] show that a random walk of length $\ell$ on a network of diameter $D$ can be performed in $\tilde O(\sqrt{\ell D}+D)$ time. A major question left open is whether there exists a faster algorithm, especially whether the multiplication of $\sqrt{\ell}$ and $\sqrt{D}$ is necessary. In this paper, we show a tight unconditional lower bound on the time complexity of distributed random walk computation. Specifically, we show that for any $n$, $D$, and $D\leq \ell \leq (n/(D^3\log n))^{1/4}$, performing a random walk of length $\Theta(\ell)$ on an $n$-node network of diameter $D$ requires $\Omega(\sqrt{\ell D}+D)$ time. This bound is {\em unconditional}, i.e., it holds for any (possibly randomized) algorithm. To the best of our knowledge, this is the first lower bound that the diameter plays a role of multiplicative factor. Our bound shows that the algorithm of Das Sarma et al. is time optimal. Our proof technique introduces a new connection between {\em bounded-round} communication complexity and distributed algorithm lower bounds with $D$ as a trade-off parameter, strengthening the previous study by Das Sarma et al. [STOC'11]. In particular, we make use of the bounded-round communication complexity of the pointer chasing problem. Our technique can be of independent interest and may be useful in showing non-trivial lower bounds on the complexity of other fundamental distributed computing problems.

💡 Research Summary

The paper addresses a fundamental question in distributed computing: how many communication rounds are required to generate a true random‑walk sample of length ℓ on an n‑node network of diameter D when each edge can carry only O(log n) bits per round (the CONGEST model). Prior work by Das Sarma et al. (PODC 2010, STOC 2011) gave an algorithm that accomplishes this in ˜O(√ℓ D + D) rounds, and left open whether the multiplicative √ℓ·√D term is inherent.

The authors prove an unconditional lower bound that matches the upper bound up to polylogarithmic factors. Specifically, for any parameters satisfying D ≤ ℓ ≤ ( n/(D³ log n) )¹⁄⁴, there exists a family of n‑node graphs of diameter D such that any (possibly randomized) distributed algorithm must spend Ω(√ℓ D + D) rounds to produce a correct random‑walk sample, regardless of which of the three natural problem variants is considered (destination reports source, source reports destination, or every node learns its position in the walk). The bound also extends to the more general CONGEST(B) model, where each edge can transmit B bits per round, as long as ℓ ≤ ( n/(D³ B) )¹⁄⁴.

The technical contribution is a novel reduction from bounded‑round two‑party communication complexity to distributed time lower bounds that explicitly incorporates the network diameter as a trade‑off parameter. The authors construct a family of graphs G(Γ, κ, Λ) that contain Γ parallel paths of length Θ(κ^Λ) and two distinguished terminals s and t. The graph has Θ(Γ·κ^Λ) nodes and diameter Θ(κ^Λ). By embedding an instance of the r‑round pointer‑chasing problem (for which Nisan and Wigderson proved an Ω(Γ·κ) message lower bound) into the structure of G(Γ, κ, Λ), they show that any distributed algorithm solving the random‑walk problem on G must simulate the pointer‑chasing protocol, thereby inheriting its communication cost. Choosing parameters Γ≈ℓ/D and κ≈√D yields the desired Ω(√ℓ D) term, while the additive D term comes from the inherent need to propagate information across the network’s diameter.

The proof also handles randomness by assuming all parties share an infinite public random string, ensuring that the lower bound holds for Monte‑Carlo algorithms with any constant error probability. To extend the result to simple graphs, the authors replace high‑capacity edges with multiple parallel edges and double ℓ, preserving the lower bound.

Overall, the paper settles the optimality of the Das Sarma et al. algorithm, demonstrates that the product of walk length and network diameter cannot be avoided, and introduces a versatile technique that may be applied to other global distributed problems where the network’s geometry plays a crucial role.