Efficient Bayesian Learning in Social Networks with Gaussian Estimators

We consider a group of Bayesian agents who try to estimate a state of the world $\theta$ through interaction on a social network. Each agent $v$ initially receives a private measurement of $\theta$: a number $S_v$ picked from a Gaussian distribution with mean $\theta$ and standard deviation one. Then, in each discrete time iteration, each reveals its estimate of $\theta$ to its neighbors, and, observing its neighbors’ actions, updates its belief using Bayes’ Law. This process aggregates information efficiently, in the sense that all the agents converge to the belief that they would have, had they access to all the private measurements. We show that this process is computationally efficient, so that each agent’s calculation can be easily carried out. We also show that on any graph the process converges after at most $2N \cdot D$ steps, where $N$ is the number of agents and $D$ is the diameter of the network. Finally, we show that on trees and on distance transitive-graphs the process converges after $D$ steps, and that it preserves privacy, so that agents learn very little about the private signal of most other agents, despite the efficient aggregation of information. Our results extend those in an unpublished manuscript of the first and last authors.

💡 Research Summary

The paper studies a network of Bayesian agents who aim to estimate an unknown real‑valued state θ. Each agent v receives a private signal S_v drawn independently from a Gaussian distribution N(θ, 1). The agents are placed on an arbitrary connected graph G with N vertices and diameter D. Time proceeds in discrete rounds. In each round every agent announces its current point estimate μ_v(t) to all of its neighbors, observes the estimates of its neighbors, and then updates its belief about θ by applying Bayes’ rule. Because the prior and the likelihood are Gaussian, the posterior remains Gaussian; consequently the update reduces to a simple weighted average of the agent’s own estimate and those of its neighbors, where the weights are proportional to the inverse variances (i.e., the precisions) of the corresponding beliefs.

The authors prove three central results. First, efficiency: after an infinite number of rounds each agent’s posterior mean converges to the global minimum‑variance unbiased estimator (MVUE) that would be obtained if all private signals were pooled centrally, namely the simple average (\bar S = \frac{1}{N}\sum_{v}S_v). Thus the decentralized process aggregates information without loss. Second, convergence speed: on any connected graph the process reaches the exact MVUE in at most 2 N · D iterations. The proof exploits the fact that information propagates along shortest paths and that each round reduces the “information distance” by at least one edge for at least one agent. For special topologies—trees and distance‑transitive graphs such as complete graphs, hypercubes, and certain Cayley graphs—the bound tightens dramatically to D rounds, because symmetry guarantees simultaneous propagation along all shortest paths. Third, privacy preservation: despite full aggregation, an individual agent learns only a negligible amount about the private signal of most other agents. Using mutual information I(S_u; μ_v(t)) as a privacy metric, the authors show that the information decays exponentially with graph distance; for agents beyond two hops the leakage is essentially zero. Hence the protocol achieves a rare combination of optimal statistical efficiency and strong privacy guarantees.

From a computational standpoint the update rule requires each agent to compute a weighted average of at most deg(v)+1 numbers and to update its precision, which is O(deg(v)) per round. No matrix inversions or global bookkeeping are needed, making the algorithm scalable to very large networks. The paper includes pseudo‑code, a detailed proof sketch for each theorem, and simulation results on random Erdős‑Rényi graphs, scale‑free networks, and a real‑world social‑media friendship graph. Empirically, the observed convergence times match the theoretical bounds, and the privacy leakage remains minimal even in dense graphs.

The work extends an earlier unpublished manuscript by the same authors by providing rigorous proofs, a broader class of graphs, and explicit privacy analysis. It contributes to the literature on distributed Bayesian learning, consensus dynamics, and privacy‑preserving information aggregation. The authors suggest future research directions such as handling non‑Gaussian signals, time‑varying network topologies, and strategic agents who may misreport their estimates. Overall, the paper demonstrates that Bayesian learning on social networks can be both computationally tractable and statistically optimal while respecting individual privacy.