Mathematical principle describing how random matrix eigenvalues distribute, predicting optimal width expansion in neural networks.