Well-Specified Predictor Efficiency Comparison

Background: The statistical efficiency of different predictor architectures (single-step rollout, direct multi-step, and single-step trained with multi-step loss) is compared for linear dynamical systems in a well-specified setting where the system is fully observed. All methods exhibit an asymptotic error decay rate of $1/N$.

Question / Future Work: Compare the asymptotic prediction error decay rates of single-step, intermediate (single-step with multi-step loss), and direct multi-step predictors in the well-specified (fully observed) linear system setting, focusing on how the horizon $H$ influences the efficiency gap between them.