{"id":"arxiv/2608.21304","url":"https://arxiv.org/abs/2608.21304","tags":["math.NA","cs.NE"],"score":0,"title":"Basin-Preserving Discretizations of Modern Hopfield Retrieval Dynamics: Energy Cells, Dissipation, and the Attention Limit","domain":"math.NA","source":"arxiv","authors":["Francisco R. Villatoro"],"reviews":0,"subject":"math.NA","summary":"The retrieval dynamics of a modern Hopfield network is the gradient flow of a log-sum-exp energy, while the attention update is its exact difference-of-convex minimization step. We study which time discretizations preserve not only energy decay and equilibria but also basins of attraction. We introduce energy cells, connected components of sublevel sets containing one attractor and no other critical point. Our main theorem shows that every finite energy cell below the escape energy is contained simultaneously in the basin of the continuous flow, every relaxed attention map…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-08-21","reactions":{}}