Absence of critical scaling in the Schelling segregation model

What the math gives to ML

The paper identifies a quantitative mechanism in which transitions are caused by finite, subcritical relocation cascades rather than critical long-range correlations. A branching ratio R below one predicts a finite mean cascade size of 1/(1-R), while susceptibility grows sharply as R approaches one. This can transfer to neural networks as a cascade-stability controller: perturb activations, gradients, or parameter blocks, estimate how many new computational events are triggered, and regulate learning rate or residual gain before the effective branching ratio reaches one.