Remarks on Bootstrap Percolation in Metric Networks

We examine bootstrap percolation in d-dimensional, directed metric graphs in the context of recent measurements of firing dynamics in 2D neuronal cultures. There are two regimes, depending on the graph size N. Large metric graphs are ignited by the occurrence of critical nuclei, which initially occupy an infinitesimal fraction, f_* -> 0, of the graph and then explode throughout a finite fraction. Smaller metric graphs are effectively random in the sense that their ignition requires the initial ignition of a finite, unlocalized fraction of the graph, f_* >0. The crossover between the two regimes is at a size N_* which scales exponentially with the connectivity range \lambda like_* \sim \exp\lambda^d. The neuronal cultures are finite metric graphs of size N \simeq 10^5-10^6, which, for the parameters of the experiment, is effectively random since N« N_. This explains the seeming contradiction in the observed finite f_ in these cultures. Finally, we discuss the dynamics of the firing front.

💡 Research Summary

The paper investigates bootstrap percolation—a threshold‐driven activation process—in directed metric graphs of dimension d, motivated by recent measurements of firing dynamics in two‑dimensional neuronal cultures. In bootstrap percolation each node becomes active only when it has at least m active incoming neighbors. The central question is how the critical initial active fraction f* (the smallest fraction of nodes that must be initially active to trigger a macroscopic cascade) depends on the size N of the network and on the spatial range λ of connections.

The authors first distinguish two regimes determined by a crossover size N*. For networks much larger than N* (N ≫ N*), the system behaves as a true metric network: rare spatial “critical nuclei” of infinitesimal relative size can appear purely by chance. Once such a nucleus forms, its interior nodes already satisfy the m‑neighbor condition, and the active region expands outward, eventually engulfing a finite fraction of the whole graph. In this regime the critical initial fraction tends to zero, f* → 0, because the presence of even a single nucleus is sufficient to ignite the whole system.

For smaller networks (N ≪ N*), the probability of finding a nucleus becomes negligible. The graph then behaves effectively as a random network, where connections are independent of distance. In this case a finite, non‑localized fraction of nodes must be initially active to achieve percolation, i.e., f* > 0. By estimating the expected number of nuclei as P ≈ N exp(−c λ^d) (with c a constant that depends on m and d) and setting P ≈ 1, the authors obtain the crossover size N* ≈ exp(c λ^d). Thus N* grows exponentially with the d‑th power of the connectivity range λ.

Applying these results to the experimental neuronal cultures, the authors note that the cultures contain N ≈ 10^5–10^6 neurons, while the effective connectivity range (set by axonal/dendritic spread) corresponds to λ of order a few tens of cell diameters. Plugging realistic values into the expression for N* yields N* on the order of 10^11 or larger. Consequently the cultures lie deep in the N ≪ N* regime, meaning they are effectively random graphs despite their underlying metric structure. This explains why experiments observe a finite critical fraction f* rather than the vanishing value predicted for an infinite metric network.

The paper also addresses the dynamics of the firing front once a nucleus has formed. By treating the front as a moving interface, the authors derive a continuum equation for its velocity v(x,t) that incorporates the distance‑dependent connection probability and the threshold m. The analysis shows that the front initially accelerates as it sweeps through the dense core of the nucleus, then decelerates when it reaches regions where the probability of finding enough active neighbors drops sharply with distance. This theoretical front profile matches qualitatively the spatiotemporal patterns reported in the neuronal culture experiments.

Finally, the authors discuss extensions such as incorporating heterogeneous degree distributions (e.g., scale‑free connectivity), adding inhibitory links, or allowing the threshold m to evolve in time. They suggest that combining empirical connectivity maps with the present analytical framework could yield precise quantitative predictions for real neural tissue.

In summary, the study provides a clear theoretical picture of bootstrap percolation on metric networks, identifies an exponential crossover size N* that separates a “critical‑nucleus” regime (f* → 0) from a random‑graph regime (f* > 0), and demonstrates that the finite‑size neuronal cultures studied experimentally belong to the latter. The work also offers a tractable description of the subsequent firing front, thereby linking static percolation thresholds to the dynamic spread of activity in spatially embedded neural systems.