Stabilization Regimes in Misspecification

Background: In a misspecified setting, irreducible prediction bias can lead to controllers that fail to stabilize the true, partially observed system, resulting in infinite LQR cost. The analysis in the paper provides numerical examples of stabilization but leaves the theoretical characterization incomplete.

Question / Future Work: Theoretically characterize the conditions under which multi-step predictors yield a stabilizing controller in the misspecified setting, while single-step or intermediate predictors fail to stabilize the system, which is currently left as a theoretical characterization to future work based on numerical observations.