Multidimensional policy-parameter resonance analysis
Establish the resonance behavior of the primal-dual learning recursion for a multidimensional policy parameter, where the constraint gradient is not an eigenvector of the Lagrangian curvature, and quantify the error introduced by reducing the coupled system of dimension d+1 to a scalar curvature–sensitivity pair.
References
Two extensions remain open. The first is a multidimensional policy parameter: the scalar pair (a, c) is exact only when d = 1 or when the constraint gradient is an eigenvector of the curvature, and otherwise the tongue comes from a coupled system of dimension d + 1. We expect the mechanism to persist on the mode most strongly coupled to the constraint, but reducing a d-dimensional curvature to the single number a carries an uncontrolled error; closing that gap is a spectral question about the pencil, not a new dynamical mechanism.
A separate, unresolved puzzle: Stooke et al. (2020, Figure 1) sweep the integral gain KI , our α, and observe the cost oscillations shrink and vanish as the gain grows, which read naively runs against Proposition 1 and Theorem 1. We have not run their sweep ourselves. A conjectured reconciliation: their perturbation has a spectrum fixed by the rate of policy improvement rather than swept, so as α grows ˜ω0 ∝ √α leaves that excitation band, predicting an amplitude non-monotone in KI , maximal where ˜ω0 crosses it. That prediction is testable on the same sweep, which we leave to future work.