Two parties want to simulate a joint probability distribution. They need shared randomness — correlated coin flips that both can see. In a point-to-point channel, the minimum rate of shared randomness needed equals the mutual information of the target distribution. This is a classical result.
In a network — multiple senders, multiple receivers, intermediate nodes — the problem becomes harder. Each pair of communicating parties may need different amounts of shared randomness, and the randomness available at intermediate nodes can be split, combined, or relayed. The question: what's the minimum total randomness the network needs to simulate all the target distributions simultaneously?
Jain & Nair (arXiv:2603.13197) show that in communication networks, shared randomness can be compressed beyond what point-to-point theory predicts. The key insight: randomness that serves one pair of users can sometimes be reused to serve another pair, if the network topology allows it. A random seed at an intermediate node can be partially observed by multiple receivers, each extracting the portion relevant to their own simulation task.
The compression is not unlimited — the paper characterizes the achievable rate region, showing exactly how much savings the network topology permits. The bounds are tight for several network families, including broadcast channels and multiple-access channels.
The conceptual contribution: randomness in networks is a shared resource, not a dedicated one. In point-to-point communication, each link needs its own randomness. In a network, the same randomness can do double duty because the correlations it creates are partially redundant across links. The minimum total randomness is less than the sum of the per-link requirements — the network structure itself provides a kind of randomness compression.