Analyzing AdamW Optimizer State Effects on Minibatch Perturbation Dynamics

AdamW is modeled as a finite-horizon input-state-output system to track how optimizer memory influences future loss effects. The study uses a signed response operator to map localized gradient perturbations to their delayed magnitude and timing in subsequent training updates.