A quantitative replica-symmetric bound of Sherrington--Kirkpatrick model in the entire de Almeida--Thouless region

What the math gives to ML

The paper gives a quantitative stability criterion for a high-dimensional random binary system: the replica-symmetric fixed point is stable when the squared interaction strength times the average squared local response, \(\beta^2\mathbb{E}\,\operatorname{sech}^4(\cdot)<1\). Inside this region, two replicas driven by the same random couplings have overlap fluctuations of order \(N^{-1}\), so the system avoids macroscopic sensitivity to initialization or sampling noise. A plausible neural transfer is a stochastic binary or recurrent layer whose gain is explicitly constrained by an empirical de Almeida–Thouless susceptibility estimate. This is not a theorem for arbitrary neural networks, but it provides a concrete stability-controlled architecture and an inexpensive scalar diagnostic for detecting critical or unstable mean-field dynamics.