{"id":"arxiv/2604.22138","url":"https://arxiv.org/abs/2604.22138","tags":["math.OC","eess.SY"],"score":0,"title":"Training Dynamics and Induced Preconditioning of ReLU Policy Gradient for Scalar LQR","domain":"math.OC","source":"arxiv","authors":["Jhojan A. Rodriguez-Gil","César A. Uribe"],"reviews":0,"subject":"math.OC","summary":"Policy gradient is not invariant to reparameterization; the training dynamics of the weights of a neural controller differ from those of the induced gains. We make that difference exact for deterministic scalar discounted linear-quadratic regulation with bias-free one-hidden-layer ReLU policies. We show that policy gradient on the weights induces preconditioning on the gains' dynamics, alongside an explicit quadratic finite-step correction and an activation-group error. Moreover, we identify explicit conditions on the step sizes and initialization parameters that guarantee that the sequence…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-04-24","reactions":{}}