Epistemic Cascade

An epistemic cascade is a sequential process in which agents adopt a belief not because they have evaluated the evidence independently, but because they observe prior adoption by others and rationally infer that those others possess private information supporting the belief. The result is a herding dynamics in which a community converges on a belief that may be false, even though every individual agent acted rationally given their limited information. Epistemic cascades are a canonical pathology of epistemic networks with sequential information flow, and they demonstrate that rational individual updating can produce collectively irrational outcomes when network structure is ignored. The classic model (Bikhchandani, Hirshleifer, and Welch 1992) shows that once a cascade begins, public information overwhelms private signals, and further evidence is ignored.

The Cascade as a System Attractor

An epistemic cascade is not merely a statistical phenomenon. It is a system attractor — a self-reinforcing configuration that the network settles into and that resists displacement. Once a cascade begins, the public signal (the observed consensus) becomes so strong that private signals are swamped regardless of their quality. The system is not broken; it is doing exactly what rational Bayesian updating should do given the information structure. The problem is that the information structure is endogenous: the consensus is itself produced by the updating process, and the updating process treats the consensus as exogenous evidence.

This is the hallmark of a feedback loop with positive gain and insufficient damping. The public signal amplifies itself through the agents' responses, and the amplification continues until the private signal is effectively zero. The cascade is not a failure of rationality. It is a failure of network design: the topology of the information flow creates a positive feedback loop that the agents' rationality is powerless to resist.

Network Topology and the Cascade

The canonical BHW model assumes a sequential line: agent 1 acts, agent 2 observes agent 1, agent 3 observes agents 1 and 2, and so on. This is a path graph — the simplest connected network. Real epistemic communities are not paths. They are complex networks with clustering, community structure, and heterogeneous connectivity. The topology of these networks determines whether rational updating produces convergence, polarization, or persistent disagreement.

In a complete network (everyone sees everyone), agents converge rapidly because every agent has access to the full history of actions. If early signals are misleading, the entire network converges on the wrong belief quickly. In a clustered network with limited inter-cluster connectivity, subgroups can maintain divergent beliefs because their information neighborhoods are effectively isolated. The network does not cascade to a single belief; it fragments into epistemic echo chambers. In a small-world network, cascades can propagate rapidly across the entire network through long-range bridges, but the same bridges can also transmit corrective signals if the bridge agents have access to diverse information sources.

The network epistemology literature, drawing on Kevin Zollman's work, has established that network topology is not merely a moderator of cascade dynamics. It is the primary determinant. The BHW model is not a general theory of epistemic cascades; it is a special case that applies only to path graphs. The general theory must be formulated in terms of graph structure, spectral properties, and phase transitions in belief dynamics.

Phase Transitions in Epistemic Networks

An epistemic network can be modeled as a dynamical system on a graph, where each node's state is its belief and the edges represent information flow. The dynamics are governed by the agents' updating rules (Bayesian, bounded-rational, heuristic) and the network's adjacency structure. The key question is: at what network density does the system shift from persistent disagreement to rapid convergence? And at what clustering coefficient do subgroups become stable epistemic echo chambers?

These are phase-transition questions, and they have phase-transition answers. There exists a critical connectivity threshold — analogous to the percolation threshold in random graphs — below which the network cannot sustain a global cascade and above which a single signal can propagate to the entire network. There exists a critical clustering coefficient — analogous to the modularity threshold in community detection — above which the network fragments into disconnected epistemic basins. The precise values of these thresholds depend on the updating rule, the signal-to-noise ratio of private information, and the heterogeneity of node influence. But the existence of the thresholds is robust.

The practical implication is that epistemic cascades cannot be prevented by exhorting individuals to think