In fruit fly egg chambers, border cells follow a chemical gradient toward the oocyte. The mathematics are clean: chemoattractant diffuses from the source, cells read the concentration slope, migration follows. But a phase-field model that accounts for real tissue geometry reveals that intersections and narrow passages flatten local gradients, slowing or stalling cells that should — by the signal alone — keep moving. The geometry of the channel overrides the strength of the signal.
When unrestricted information flows through a network of perfectly rational Bayesian agents, the group's collective beliefs get worse, not better. Each agent processes incoming signals correctly. The update rules are sound. But the network topology introduces correlations between signals that agents treat as independent. At high-connectivity nodes — the intersections of the information network — evidence becomes so redundant that rational updating amplifies noise rather than extracting truth.
In both systems, the signal exists and is real. The degradation happens in the channel, not at the source.
The cell migration case is physical. Tissue geometry creates regions where diffusion equilibrates locally — a narrow passage between nurse cells allows the chemoattractant to reach near-uniform concentration across the passage width. The cell reads this flattened gradient as “no direction,” even though the global gradient clearly points toward the oocyte. The bottleneck doesn't block the signal. It homogenizes it, which is worse. A blocked signal can be routed around. A flat gradient offers no information about which way to go.
The information network case is structural. A highly connected agent receives signals from many neighbors, most of whom have already been influenced by overlapping sources. The agent, treating each signal as fresh evidence, updates confidently in a direction that reflects network topology rather than underlying truth. The hub doesn't block information. It concentrates it, and concentration without independence is amplification of bias.
The common mechanism: both systems have a transport equation (diffusion of molecules, propagation of beliefs) operating on a graph (tissue geometry, social network). Where the graph creates convergence points — bottlenecks, hubs — the transported quantity loses its gradient structure. Molecules equilibrate. Beliefs correlate. In both cases, the agents doing the reading (cells, Bayesians) have no way to distinguish a genuinely flat signal from a topologically flattened one.
This suggests a design principle: the fidelity of gradient-following depends more on channel geometry than on signal strength. Doubling the chemoattractant concentration won't help a cell stuck in a bottleneck. Doubling the information quality won't help an agent drowning in correlated evidence. The fix is architectural — reshape the channel, not amplify the source.