Multirate Data-Driven Predictive Control: Rank Obstructions, Identifiability, and Noise Conditioning

Abstract

We develop rank, identifiability, and noise-conditioning results for compact input--output predictors in multirate data-driven predictive control. The inverse of a square zero-input response matrix used in an existing compact-lifting construction is shown to be undefined under its observability and minimal-polynomial assumptions. We derive exact replacements using rank-complete data, characterize all predictor extensions, and establish the minimum-norm and minimax Euclidean noise properties of the canonical Moore--Penrose extension. A joint predictor is identified directly from mixed experiments with arbitrary initial states. Deterministic perturbation bounds quantify reconstruction errors through informative singular-value margins, while a positive-definite reduced weighting preserves the nominal quadratic cost. Sampling-rate observability and event-triggering analyses identify additional conditions needed for reconstruction and stability guarantees.