The speed of distributed computer programs is limited by the physical location of data rather than the spectral mixing mathematicians previously blamed.
April 23, 2026
Original Paper
Locality, Not Spectral Mixing, Governs Direct Propagation in Distributed Offline Dynamic Programming
arXiv · 2604.18615
The Takeaway
Physical locality governs how fast information can spread through a distributed network during dynamic programming. This discovery overturns years of technical assumptions that focused on the spectral properties of the network graph. Gossip algorithms are mathematically proven to be slower than direct propagation methods in these scenarios. Engineers have been optimizing the wrong variables when trying to speed up large-scale distributed systems. It means we can build much faster cloud infrastructure by focusing on data proximity rather than network topology.
From the abstract
We study the communication complexity of distributed offline dynamic programming, where a fixed batch dataset is partitioned across (M) machines connected by the data-induced dependency graph. We compare two paradigms: direct boundary-value propagation, which follows Bellman dependencies, and gossip averaging, which mixes local estimates. Our results show that **locality** is the fundamental driver of round complexity. In particular, we prove that no method can achieve (\varepsilon)-accuracy in