no. 16 · Tuesday 17 November 2026
Boundary Collisions and Weak Excitation in LQR Policy-Gradient Flows
Abstract
Closed-loop stability concerns all initial states, whereas an LQR cost evaluates only the directions excited by a prescribed initial-state distribution. We study how this distinction affects Euclidean policy-gradient flows. For every nonzero deficient covariance rank, we construct relatively open families of controllable data and open sets of stabilizing initial gains whose exact gradient trajectories reach the stability boundary in finite optimization time. The state Gramian can remain positive definite at every time before collision. By separating the cost contribution of a critical mode, we obtain a local collision criterion based on direction attraction and motion toward the stability boundary. Adding excitation of strength $\varepsilon$ yields matching $Θ(\sqrt\varepsilon)$ bounds on the local minimum stability margin for trajectories that would otherwise collide. For a class of single-input, corank-one systems, these bounds hold over the entire trajectory without a common compact-set assumption. Numerical examples further illustrate the dynamics and scaling laws.
- fields
- Optimization and Control, Dynamical Systems