Don't truncate, decompose: mean-field dynamics of long-range quantum systems from strongly correlated states

What the math gives to ML

The paper provides a nonstandard mechanism for handling strongly correlated states in strongly long-range systems: instead of truncating a cumulant hierarchy, it decomposes the many-body state into independently evolving mean-field components and reconstructs correlations from the decomposition. The transferable asset is an ensemble representation in which low-order statistics and even non-Gaussian multimodality are generated by the distribution of mean-field trajectories, avoiding a fragile Gaussian or finite-cumulant closure. A promising neural-network analogue is an ensemble-valued recurrent or state-space layer whose particles evolve by a shared mean-field map while outputs are reconstructed from empirical moments, preserving multiple attractors and symmetry-breaking branches.