Space-efficient population protocols for exact majority on general graphs

We study exact majority consensus in the population protocol model. In this model, the system is described by a graph $G = (V,E)$ with $n$ nodes, and in each time step, a scheduler samples uniformly at random a pair of adjacent nodes to interact. In the exact majority consensus task, each node is given a binary input, and the goal is to design a protocol that almost surely reaches a stable configuration, where all nodes output the majority input value. We give improved upper and lower bounds for exact majority in general graphs. First, we give asymptotically tight time lower bounds for general (unbounded space) protocols. Second, we obtain new upper bounds parameterized by the relaxation time $\tau_{\mathsf{rel}}$ of the random walk on $G$ induced by the scheduler and the degree imbalance $\Delta/\delta$ of $G$. Specifically, we give a protocol that stabilizes in $O\left( \tfrac{\Delta}{\delta} \tau_{\mathsf{rel}} \log^2 n \right)$ steps in expectation and with high probability and uses $O\left( \log n \cdot \left( \log\left(\tfrac{\Delta}{\delta}\right) + \log \left(\tfrac{\tau_{\mathsf{rel}}}{n}\right) \right) \right)$ states in any graph with minimum degree at least $\delta$ and maximum degree at most $\Delta$. For regular expander graphs, this matches the optimal space complexity of $\Theta(\log n)$ for fast protocols in complete graphs [Alistarh et al., SODA 2016 and Doty et al., FOCS 2022] with a nearly optimal stabilization time of $O(n \log^2 n)$ steps. Finally, we give a new upper bound of $O(\tau_{\mathsf{rel}} \cdot n \log n)$ for the stabilization time of a constant-state protocol.