Power-law and log-periodic degree tails for a family of probability generating function equations arising in evolving networks

What the math gives to ML

The paper gives an explicit, implementable branching-process construction whose stopped population size has a heavy tail with a non-constant multiplicatively periodic correction, rather than an ordinary pure power law. This provides a principled distribution for dynamically varying computational width or sparse expert fanout, with a tunable finite mean but occasional very large computations. The transferable asset is the exact geometric stopping rule and the resulting discrete-scale-invariant tail, which can replace ad hoc fixed-capacity or Gaussian width noise in sparse neural architectures. The safest first use is a capped stochastic MoE or tree-like MLP whose per-token fanout is sampled from this process and trained with an unbiased compute/load penalty.