Non-Minimal State Representations from Input-Output Data

This direction is merged into another direction: generalization to nonlinear systems. Its last version stays readable and citable here.

Version 1 of 113 papers17 conceptsfrozen 2026-10-04

In classical control theory, the "state" of a system is a compact vector that captures everything about the past necessary to predict the future. If you know the state, you don't need the history. However, in many real-world applications, this internal state is hidden, and we only have access to a stream of inputs ($u$) and outputs ($y$).

The traditional solution is a two-step process: first, identify a mathematical model of the system; second, design an observer (like a Kalman filter) to estimate the hidden state. A growing body of research suggests we can bypass this complexity by using a non-minimal state representation—a surrogate state constructed directly from a finite window of historical input and output data.

The History as a State Proxy

The core intuition is that for an observable system, the recent history of what we did (inputs) and what happened (outputs) contains exactly the same information as the hidden internal state. By stacking these observations into a large vector, we create an "extended" or "non-minimal" state.

Mohamed Naveed Gul Mohamed, Raman Goyal, and Suman Chakravorty formalize this as the information-state. In An Information-State Based Approach to Linear Time Varying System Identification and Control, they demonstrate that if the history window is long enough to satisfy observability requirements, this representation is mathematically sufficient for optimal feedback control. This shifts the definition of a state from a physical variable to an abstract information-theoretic construct, a concept Paweł Wachel traces back to Nerode equivalence in On State-Space Representations of General Discrete-Time Dynamical Systems.

Data-Enabled Predictive Control

One of the most influential applications of this idea is Data-Enabled Predictive Control (DeePC). Introduced by Jeremy Coulson, John Lygeros, and Florian Dorfler in Data-Enabled Predictive Control: In the Shallows of the DeePC, DeePC uses raw data trajectories to perform optimal tracking.

Instead of solving a model-based optimization, DeePC uses a window of recent measurements as an implicit state estimate. This allows the controller to account for the system's current condition directly through its behavioral history, eliminating the need for separate observers. Similarly, Johannes Köhler, Julian Berberich, and Matthias A. Müller show in On the design of terminal ingredients for data-driven MPC how to design stability guarantees (terminal costs and sets) directly from this lag-based representation.

Bridging to Reinforcement Learning

This non-minimal framing also solves a major headache in Reinforcement Learning (RL): dealing with partially observable environments. Standard Q-learning assumes you can see the full state. When you can't, the problem becomes much harder.

In An efficient data-based off-policy Q-learning algorithm for optimal output feedback control of linear systems, Mohammad Alsalti, Matthias A. Müller, and Victor G. Lopez use the input-output history window to transform an output-feedback problem into a state-feedback problem. By treating the history as the state, they can apply efficient RL algorithms that converge quadratically to optimal solutions without ever knowing the underlying system's internal physics.

Stability and Robustness of Data Models

If we use a large window of data to represent a small system, we introduce "extra" dynamics—essentially, the model has more degrees of freedom than the physical system. A natural concern is whether these extra components might become unstable.

Joo-Won Lee and colleagues address this in Input-Output Data-Driven Representation: Non-Minimality and Stability. They prove that these "latent poles" are inherently stable when the model is constructed using standard methods like the Moore-Penrose pseudoinverse.

Furthermore, this framework allows for formal safety proofs. Anne Koch, Frank Allgöwer, and Julian Berberich demonstrate in Provably Robust Verification of Dissipativity Properties from Data that we can verify "dissipativity"—a property critical for stability—using only noisy measurement data. By mapping the history window to Linear Matrix Inequalities (LMIs), they provide tight guarantees that hold even when the system model is completely unknown.

Why It Works

The success of non-minimal representations relies on a fundamental trade-off: we trade compactness for directness.

  1. Observability: A sufficient window of past I/O is a one-to-one mapping to the internal state.
  2. Linearity: In linear systems, this mapping is a simple linear transformation, making the "extended state" just as easy to work with as the original state.
  3. Implicit Filtering: By using a window of data, the controller implicitly performs the work of an observer, smoothing out noise and capturing the system's momentum without needing a separate Kalman filter.

From stabilizing power converters (Feiran Zhao et al., Direct Adaptive Control of Grid-Connected Power Converters) to optimizing complex trajectories, the shift toward history-based states is enabling a more robust, model-free approach to control.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_non_minimal_state_representations_from_i_8bfaa401,
  title = {Non-Minimal State Representations from Input-Output Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-minimal-state-representations-from-input-output-data/rdr_8bfaa401eb7c49119931a2a372df0f58},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_non_minimal_state_representations_from_i_8bfaa401_v1,
  title = {Non-Minimal State Representations from Input-Output Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-minimal-state-representations-from-input-output-data/rdr_8bfaa401eb7c49119931a2a372df0f58/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}