Learning-based Homothetic Tube MPC with Non-Asymptotic Guarantees

What the math gives to ML

The paper provides a constructive mechanism for combining non-asymptotic regularized least-squares confidence sets with robust homothetic-tube propagation. Its transferable asset is a computable uncertainty-to-stability pipeline: maintain a high-probability parameter polytope, propagate its effect through a feedback or model rollout, and tighten admissible states and inputs by the resulting tube. A strong neural-network transfer is uncertainty-aware rollout control for learned world models or neural state-space models, where local Jacobian or last-layer uncertainty is converted into a certified tube around the nominal trajectory. The method makes falsifiable predictions: confidence radii should shrink with accumulated excitation, tube radii should obey a geometric recursion, and constraint violations should remain below the declared confidence failure probability.