Notes on data-driven output-feedback control of linear MIMO systems

Mohammad Alsalti, Victor G. Lopez, Matthias A. Müller

TL;DR

Data-driven control often struggles with multi-output (MIMO) systems because standard methods of building "surrogate" states from past data cause mathematical rank deficiencies. This paper solves the problem by introducing a novel, data-derived non-minimal state vector that guarantees full rank. This unlocks direct, robust output-feedback control for general MIMO systems without requiring an explicit system model.

What Problem Does This Paper Address?

Direct data-driven control has gained massive popularity because it allows engineers to design controllers directly from historical input-output data, completely bypassing the tedious step of system identification (building a mathematical model). While this works beautifully when you can measure the full internal state of a system, practical applications usually only give us access to a few outputs.

To get around this, researchers often construct a "non-minimal state" by stacking a window of past inputs and outputs. For Single-Input Single-Output (SISO) systems, this trick works perfectly. However, for Multi-Input Multi-Output (MIMO) systems, it breaks down.

The authors prove that for any system with more than one output ($p > 1$), the standard non-minimal state representation is inherently uncontrollable, meaning the resulting data matrices can never achieve full row rank. To make their math work, previous literature frequently restricted the class of systems to those where the number of outputs multiplied by the system lag equals the system order ($p\ell = n$). Because this condition rarely holds in the real world (generally $p\ell \ge n$), a generalized data-driven output-feedback method for MIMO systems has remained out of reach.

Methods

The authors propose a clever data-based workaround. Instead of blindly stacking all past outputs into the surrogate state vector, they use a data-driven permutation approach to isolate and retain only the linearly independent components of the output history.

  1. Row Selection via Permutation: By analyzing the data matrix of past inputs and outputs, they use standard Gaussian elimination to find an invertible permutation matrix ($\Pi_2$). This matrix separates the $n$ linearly independent rows from the redundant, linearly dependent ones, yielding a selection matrix $\bar{\Theta}$.
  2. The New Non-Minimal State: They define an alternative non-minimal state vector: $$z_t = \begin{bmatrix} u_{[t-\ell,t-1]} \\ \bar{\Theta}y_{[t-\ell,t-1]} \end{bmatrix}$$ This vector acts as a valid state for an alternative representation that is controllable and guarantees full row rank when the system is persistently excited.

  3. Handling Noise: When data is corrupted by noise, the strict rank conditions fail. The authors extend their method by applying a truncated Singular Value Decomposition (SVD) to approximate the clean data matrix before performing the row selection.

This architecture allows designers to directly apply control techniques originally developed for the simpler input-state setting to output-feedback problems, as illustrated below:

Fig. 1. An illustration of the closed-loop system, where a feedback controller based on a non-minimal state zt is used.

Datasets

No standard external datasets were used. The authors validated their approach using a synthetic numerical simulation of a third-order, open-loop unstable LTI MIMO system.

Key Insights and Ideas

  • The Uncontrollability Proof: A major "aha" moment is the formal proof that the traditional past-input/past-output state vector makes multi-output systems structurally uncontrollable. This explains why previous optimization attempts in the literature frequently encountered infeasible solutions.
  • Model-Free Coordinate Selection: You do not need to know the underlying system matrices ($A, B, C, D$) to find the right state coordinates. The compression matrix $\bar{\Theta}$ can be computed entirely from raw data using basic linear algebra.
  • Restoring Matrix Conditioning: By stripping away redundant output data, the method drastically improves the numerical conditioning of the data matrices, making optimization algorithms stable even in the presence of measurement noise.

Key Results

The authors compared their proposed state vector ($z_t$) against the traditional extended state vector ($\xi_t$) on a simulated noisy, unstable 3rd-order MIMO system where $p\ell = 4 > 3 = n$.

  • Matrix Conditioning: The traditional data matrix was severely ill-conditioned, yielding a condition number of $7.4017 \times 10^3$. In contrast, the authors' proposed data matrix achieved a clean condition number of $179.9305$.
  • Controller Feasibility: Due to the ill-conditioning and uncontrollability of the traditional state, the control design Linear Matrix Inequalities (LMIs) were completely infeasible. The proposed method successfully solved the LMIs, returning a robustly stabilizing output-feedback controller gain matrix $K$.

Limitations and Future Work

The current framework is strictly limited to linear time-invariant (LTI) systems. Additionally, while the truncated SVD approach handles small amounts of noise well, a comprehensive framework for highly noisy data remains an open challenge.

The authors note that the primary direction for future work is extending this non-minimal state definition to nonlinear systems. This will involve exploring how to select user-defined basis functions, potentially leveraging Koopman operator theory or autoencoder-based lifting techniques, to ensure similar rank conditions are satisfied by design.

Summary updated

Abstract

Recent works have approached the data-driven design of dynamic output-feedback controllers for discrete-time LTI systems by constructing non-minimal state vectors composed of past inputs and outputs. Depending on the system's complexity (order $n$, lag $\ell$ and number of outputs $p$), it was observed in several works that such an approach presents significant limitations. In particular, many works require to restrict the class of LTI systems to those satisfying the relation $p\ell=n$. In this note, we show how to address the general MIMO case (for which $p\ell\geq n$ in general) by constructing an alternative non-minimal state vector from data. Different from the existing literature, our method guarantees the satisfaction of certain rank conditions when the system is persistently excited, thereby facilitating the direct data-driven dynamic output-feedback control of MIMO systems by applying methods that were originally developed for the input-state data setting.
BibTeX
@article{alsalti2023notes,
  author = {Mohammad Alsalti and Victor G. Lopez and Matthias A. Müller},
  title = {{Notes on data-driven output-feedback control of linear MIMO systems}},
  journal = {IEEE Trans. Automatic Control},
  year = {2023},
  doi = {10.1109/tac.2025.3553073},
}
Figure