Self-organized robustness in mean-field interacting systems

Abstract

Self-organization is a defining feature of living systems, with order often maintained through interactions between constituent units rather than centralized feedback. We introduce a tractable mean-field model of self-organized robustness, formulated as meta-optimization over the system's response to perturbations. The resulting interaction structure has an intuitive picture as a dynamically modulated landscape (``seascape'') whose shape is determined self-consistently to accelerate relaxation back to equilibrium. The collective dynamics follows an optimized Wasserstein gradient flow toward an attractor in the space of collective states. When communication is limited, interactions preferentially encode slowly relaxing modes and modes that are frequently perturbed. The model further shows that robust collective states are associated with flatter equilibrium landscapes and predicts a continuum of intermediate ``reservoir states'' in such systems. The model offers a perspective of self-organization as a hierarchical associative memory that operates on the scale of a collective of interacting computational units.