A mathematical model can become easier to observe without making a machine easier to control. In experiments on a nonlinear oscillator, promoting the model’s observability score did not improve either state estimation or tracking. A more practical intervention—refreshing the controller’s hidden-state estimate from new measurements at every sampling instant—had a larger effect. That gap between a metric and the performance it is meant to predict is the paper’s central lesson.

Maria Bartzioka and Mohammad Khosravi study how to learn finite-dimensional models of nonlinear systems from input-output records: the measured outputs and the control inputs that produced them. Their approach uses a Koopman representation, a way of describing nonlinear motion as linear evolution after transforming the system into a larger space of features. The transformation can make familiar tools such as linear observers and model predictive control usable for nonlinear systems. But the paper argues that a model that predicts well is not necessarily one a controller can use well. A controller must recover the model’s hidden state from measurements and correctly represent how its own actions change that state.

The distinction matters anywhere a machine must act without seeing its full physical state: a robot estimating its motion from cameras, an industrial process controlled from sparse sensors, or a vehicle steering through changing conditions. The paper does not demonstrate a new deployed controller in those settings. Instead, it gives engineers a more careful way to ask what a learned model is actually good for—and shows that one apparently attractive score may not answer the question they care about.

The Science

The physical systems in the paper are described by a state that changes over time under two influences: the system’s own dynamics and an applied input. Sensors report an output generated from that state. In general, the output is incomplete: two different physical states can produce the same instantaneous reading. A controller must therefore infer what is happening from a history of measurements and the known inputs applied along the way.

Koopman theory tackles nonlinear dynamics by transforming the physical state into a set of observables—features calculated from that state. In the transformed, or “lifted,” space, the researchers seek dynamics that behave approximately like a linear system. The learned model has a matrix governing its free evolution, a matrix for the effect of control inputs, and a linear map from the hidden lifted state back to measured outputs. If the approximation is useful, linear estimation and optimization tools can operate on this representation even when the original system is nonlinear.

The key design choice is to construct the lifted state from a finite, causal window of past outputs and known inputs. “Causal” means the encoder uses information available now or earlier, not future measurements. Including past inputs matters because the measurements reflect both the system and the actions already taken. The encoder produces a candidate hidden state, which the model then propagates forward; the predicted output can be compared with what the sensors report.

The authors separate several requirements that can otherwise be conflated. Observability asks whether different hidden states lead to distinguishable output histories. Input-map compatibility asks whether the chosen representation can describe the effect of inputs with a constant matrix. Controllability asks whether inputs can move the model through its state space. Data informativity asks whether the collected experiments contain enough independent variation to identify the model in the first place. A system can satisfy one of these conditions and fail another.

For observability, they use a finite-horizon observability Gramian. This matrix collects how strongly perturbations in each direction of the lifted state affect the output over a chosen time interval. Directions with small Gramian eigenvalues barely change the predicted measurements, making them difficult to distinguish, especially when sensors are noisy. The learning objective promotes observability while keeping output-prediction error and temporal-consistency error within separate bounds. Temporal consistency means that propagating the encoded state through the model should agree with encoding the later input-output window.

The work also gives structural results. Under assumptions of exactness and observability, an encoder applied to different histories that end at the same physical state must return the same lifted state. In approximate settings, the paper bounds how much those encodings can differ using prediction errors and the weakest observable direction. It also relates the nonlinear system’s iterated Lie brackets—a way of capturing directions reachable through combinations of its dynamics and inputs—to controllability of the lifted model. In particular, under the stated exact-lifting assumptions, an additive exact lift with more coordinates than the original state cannot be fully controllable.

What They Found

The experiments are deliberately split between a system with an exact analytical lift and a controlled Duffing oscillator, a standard nonlinear oscillator model. The exact benchmark gives the researchers ground truth: they know the representation the learning procedure ought to recover. There, encoder initialization determines whether learning recovers that analytical lift. This is a concrete reminder that a training objective does not by itself guarantee the intended representation; where optimization starts can matter.

The Duffing oscillator probes the approximate case. The authors show that an informative, exact additive lifting on an open set is impossible for this system. That result makes the example more than a test of whether a neural model can fit data. It identifies a structural limit: the chosen model form, with a constant additive input matrix, cannot exactly represent the system’s input action across an open region. A successful learned model must therefore be understood as an approximation, not a hidden exact linearization waiting to be discovered.

The most consequential practical result is negative. In the reported Duffing experiments, promoting the nominal observability Gramian improved its measures of observability, but did not improve observer-based state estimation or tracking. The paper does not report a universal failure of observability-aware learning; it reports a mismatch between gains in the model’s nominal metric and gains in these downstream tests. That distinction is crucial. The Gramian is calculated for the learned model, while estimation and control depend on noisy measurements, model mismatch, the observer, the controller and how information is refreshed in operation.

By contrast, updating the lifted-state estimate through the encoder at every sampling instant had a larger effect on tracking accuracy. Rather than relying only on a hidden state propagated forward by the model, this procedure repeatedly uses the newest input-output window to re-encode the state. In intuitive terms, it gives the controller regular chances to correct its account of where the system is, rather than trusting a forecast to drift forward unchecked.

The paper also formalizes why a good fit alone is insufficient. The encoder’s output should not depend arbitrarily on which past trajectory happened to reach the current state. If it does, the controller may be handed different latent descriptions of the same physical condition. Under exact assumptions, the authors establish independence from that history; for approximate models they provide a bound on the discrepancy, tied to prediction quality and observability. This connects a machine-learning design question—what does the encoder remember?—to a systems question—can the state used for control be consistently inferred?

The methodology combines neural encoding with integral regression. Instead of estimating time derivatives from noisy measurements, integral regression fits the continuous-time model by relating changes over intervals to accumulated model dynamics. The learned representation is then paired with a lifted-state estimator and model predictive control, which chooses actions by optimizing predicted future behavior subject to constraints. The paper’s experiments assess estimation and tracking separately from the nominal observability objective; that separation makes the reported caveat visible.

The available paper text reports these results qualitatively, without numerical experiment values in the supplied material. For that reason, the findings here are described without invented performance percentages or plotted quantities.

Why This Changes Things

The broader contribution is a more disciplined vocabulary for data-driven control. “The model predicts well” is not the same as “the hidden state can be recovered from sensors.” Neither statement means that a constant input matrix captures how control acts, or that the experimental data were rich enough to identify the model. By distinguishing these issues, the framework helps avoid a common failure mode: optimizing a single score and treating it as a proxy for the whole control problem.

This has practical consequences for engineering. If estimation is poor, adding more training data may not be the first answer; the sensor history may not distinguish the relevant states. If a model responds incorrectly to proposed inputs, a better encoder may not fix a structural mismatch in the additive model. If identification is underdetermined because the inputs did not sufficiently excite the system, apparent fit on familiar trajectories may not support reliable action elsewhere. The paper’s categories point toward different remedies for different failures.

Its result about lifted controllability also challenges the intuition that more features are always better. A high-dimensional representation may capture useful nonlinear structure, but exact additive lifting under the paper’s assumptions cannot make every lifted direction reachable when the lifted dimension exceeds the physical state dimension. The learned model can still be useful as an approximate predictor; the point is that dimension alone does not create genuine control authority. Engineers should distinguish adding coordinates that help represent outputs from adding independent directions the actuator can actually reach.

The Duffing result sharpens another lesson: some limitations belong to the model class, not merely to the training process. A constant input matrix is convenient because it preserves linear structure for estimation and model predictive control. But a nonlinear system’s input effect can depend on its current state. If those effects cannot be represented by a constant matrix, training harder within that form cannot yield an informative exact representation on the region in question. A state-dependent or bilinear model may be a better fit, though it brings its own complexity and must be assessed on the same output-feedback task.

Most important, the experiment separates a pleasing internal score from a useful external outcome. An observability Gramian is meaningful: it quantifies output sensitivity under a specified model and horizon. But it does not automatically account for all the ways real estimation can fail. The result is not that observability should be ignored. It is that improvements in a model-based observability measure should be verified against actual estimation and control performance rather than assumed to transfer.

For systems that affect people or infrastructure, that distinction is not academic. A robot that tracks a model’s internal state poorly may move unpredictably even if its training loss is low. A process controller may look accurate on recorded trajectories but respond badly when asked to move toward a new target. The study offers a framework for exposing such gaps earlier, before a model’s mathematical elegance is mistaken for operational reliability.

What’s Next

The findings are bounded by the systems and experiments studied. The Duffing oscillator is a controlled, simulated example, and the paper’s structural claims rely on explicit assumptions about exact lifting, observability and admissible trajectories. These results do not establish that encoder refresh will outperform observability promotion in every system, nor that a particular learned representation is safe or robust under real-world noise, unmodeled disturbances or actuator limits.

A natural next step is to test the framework on physical systems and compare more model classes. In particular, the Duffing analysis suggests a direct question: when a constant input map is structurally inadequate, do state-dependent or bilinear lifted models improve output-feedback control enough to justify their added complexity? Another is whether observability objectives can be redesigned to better predict the performance of a particular observer under actual measurement noise, rather than only improve a nominal Gramian criterion.

The paper also leaves room to study the interaction between data collection and control. A model may be controllable in principle but poorly identified because the training inputs failed to excite the relevant dynamics. Designing experiments that make the data informative—without unsafe or impractical excitation—could be as important as designing the encoder itself. Testing representation quality across changes in operating conditions would also reveal when a learned lift remains useful and when it must be updated.

The paper’s grounded message is that a controller needs a representation that is not merely linear-looking or predictive, but recoverable from what sensors actually reveal and faithful to the effects of actions. Its most useful finding may be the gap it exposes: improving a mathematical proxy did not improve the task, while a simple change in how the hidden state was refreshed mattered more. Closing that gap is the work ahead for learning-based control.