A controller can begin with a narrow range of safe actions, then earn room to maneuver as its observations rule out possible models. That is the central promise of this work on positive systems: processes whose state variables are nonnegative, such as quantities of material, energy, or population. Rather than probe an unfamiliar system with potentially dangerous experiments, the controller learns from the transitions that happen during operation—and can relax its caution only when the remaining possibilities justify it.
The result is a mathematical framework, not a tested industrial product. But it addresses a stubborn challenge in adaptive control: how to make decisions when the system’s dynamics are unknown, observations arrive over time, and safety cannot wait for the model to be learned. The authors show that, under their assumptions, the controller need not carry the entire observation history into every decision. It can instead keep a compact record of accumulated data and a set of models that remain compatible with what has been observed (Bencherki et al., 2026).
The Science
A positive system evolves through a simple accounting rule: the next state equals the effect of the current state, the chosen input, and a disturbance. In the paper’s notation, this is a discrete-time linear system, with matrices describing how the state persists and how inputs act. The state-transition matrix is nonnegative, and the disturbance is also constrained to be nonnegative. In practical terms, the model cannot make a nonnegative quantity spontaneously subtract itself through the state dynamics, and the disturbance can add to the next state but not push it below the model’s prediction.
That last condition is important. It makes observations informative. If a measured next state is too small to be explained by some candidate model plus a nonnegative disturbance, that model can be ruled out. The set of models that can still explain all observed transitions is called the consistency set. It starts as the full set of possibilities and shrinks—or stays the same—as evidence accumulates.
The controller faces an adversarial version of uncertainty. It chooses an input; an unknown but fixed system model and a nonnegative disturbance determine the next state. The design objective is minimax: choose a policy that performs as well as possible against the worst admissible model and disturbance, according to a specified stage cost. That cost includes the state, the input, and a penalty for disturbances. This is not a probability-based forecast. The method does not need to assign likelihoods to models; it asks what can be guaranteed against all possibilities left open by the data.
The safety rule is likewise set-based. At each step, the input must preserve nonnegativity for every model still in the consistency set. If two models remain plausible, the action has to be safe for both. If observations later eliminate one of them, the controller may no longer need to respect that model’s constraint. Its admissible input set can therefore grow as uncertainty falls.
To make this adaptive decision process tractable, the authors build an information state from three pieces: the current physical state, the consistency set, and a historical-data vector. The vector accumulates the observed next states, current states, and inputs. The set remembers which models remain feasible; the vector lets the controller recover the accumulated disturbance penalty relevant to the minimax objective. Together, these pieces are sufficient for the dynamic-programming formulation: the complete raw history is not needed to update the decision rule.
The figure at the heart of this idea is not a plot of a measured real-world outcome; it is the evolving information state itself.
The model set acts like a map of remaining explanations, while the accumulated vector stores what the history contributes to the objective. The paper’s central theoretical result is that this representation yields an exact reformulation as a standard two-player dynamic game. “Exact” matters: the simplification does not enlarge the adversary’s possible trajectories or discard a feasible model.
For a finite collection of candidate systems, the authors go further. They derive a certainty-equivalence controller: at each step, it selects a model from the surviving set according to the accumulated-data criterion and uses an associated feedback gain. Certainty equivalence usually means acting as if a particular model were the right one. Here, that chosen model does not get to override safety: the actual input is constrained so it remains safe for every surviving candidate. Under the paper’s assumptions, the resulting controller is safe and stabilizing and has a robustness guarantee.
What They Found
The first result is conceptual but consequential: the unknown fixed model and disturbance history can be summarized without carrying a separate historical score for every candidate. Earlier approaches for positive systems with positivity-preserving disturbances tracked model-specific histories, whose representation grew with the model family. Here, because disturbances are nonnegative, observations can eliminate inconsistent models. A shared historical-data vector plus the consistency set captures what the controller needs.
The second result is a direct connection between learning and safety. The controller’s permissible inputs are built from the intersection of safety constraints across all currently plausible models. When the set contracts, that intersection can relax. The controller can begin conservatively, then use evidence to justify actions that would have been unsafe under a model now ruled out. This is not a promise that the controller will always learn quickly: an observation may leave the set unchanged. Nor does it mean the system can be deliberately driven into unsafe territory to generate evidence. The safety condition applies throughout.
For finite model sets, the paper supplies a concrete construction. Each candidate model has an input bound chosen to preserve positivity and satisfy additional conditions linked to the cost and stabilization. The shared safe bound is formed from the most restrictive surviving bounds. The controller’s chosen gain is limited by that shared bound, so the action remains admissible for every model in the set. If only one candidate remains, the bound can reduce to that model’s own allowance.
The authors also establish a Bellman-inequality certificate. In plain terms, a Bellman inequality is a mathematical test showing that a proposed controller’s worst-case future cost stays under a chosen upper-bound function. This gives a way to certify performance without having to solve the full optimal minimax problem exactly. The paper’s theorem says that if a suitable upper-bound function and policy satisfy the stated conditions, the policy’s cost is bounded by that function from the starting state.
For continuum uncertainty—an uncountable range of possible models—the consistency set is generally a difficult object to store. The paper identifies structured classes for which it can nevertheless be represented recursively with a fixed, finite-dimensional description. The point is scalability: the size of the information state need not grow with the number of model pairs, even when that number is infinite. This result is conditional on the uncertainty class having the needed representation; it does not make arbitrary continuous uncertainty easy.
The paper’s finite-model construction can be pictured as a funnel of allowable actions. At first, the funnel is the overlap of the safety regions for all candidates. Removing a candidate can widen it, while the true system remains inside the surviving set. This is a useful way to understand the contribution, but it is a schematic interpretation rather than a reported experimental curve.
The provided paper material does not include numerical experiment values, plotted data arrays, or quantitative comparisons. So there is no honest basis for numerical charts showing performance, learning speed, or robustness. The equations and assumptions provide theoretical results, not measured gains that can be turned into a chart. A useful visual summary is therefore qualitative rather than quantitative:
These chart references cannot responsibly be populated with fabricated measurements. In a data visualization, empty data would be misleading; the paper’s central results are proofs and a framework, not a set of numerical outcomes supplied in the material here. The figures referenced above should likewise be understood as explanatory placements, not claims that the provided text contains specific numbered plots.
Why This Changes Things
The broader problem is familiar well beyond control theory. A water network, a chemical process, a storage system, or a biological production line may have dynamics that are only partly known. A controller must keep quantities within safe ranges while acting on the system, and the act of acting changes what can be learned. In this setting, a policy that waits for perfect identification may be useless; a policy that assumes the best-fitting model is certainly correct may be unsafe.
This paper offers a middle path. It treats uncertainty not as a probability distribution that must be guessed, but as a collection of models that remain possible. Safety is enforced against the whole collection. Learning does not require a special experiment: every observed transition may remove explanations that cannot account for it. As the collection narrows, the controller’s options can widen.
That structure is particularly relevant for positive systems because their constraints are not incidental. A negative state may be physically meaningless: negative inventory, negative concentration, or negative population is not merely an undesirable operating point. Positivity also creates an order that allows comparisons between models and supports the authors’ construction of robust input bounds. The system’s restriction becomes part of the learning mechanism.
The framework also speaks to a practical design tension. Conservative robust control often protects against all modeled possibilities, but can sacrifice performance by acting as though every possibility will remain plausible forever. Adaptive control can improve performance, yet learning and safety can conflict. By making the safe action set depend explicitly on evidence, this method gives a formal route to reduce conservatism without treating safety as an afterthought.
Still, the results should not be mistaken for a general recipe for safe AI or autonomous infrastructure. The paper assumes a fixed unknown model, a particular linear positive-system structure, nonnegative disturbances, and a known compact uncertainty set. Real systems can have model drift, sensor error, delays, unmodeled interactions, or disturbances that do not fit the assumed sign restriction. The admissible input set must also be nonempty, and the finite-model controller relies on conditions that may be demanding in a specific application.
The distinction between “safe under the model family” and “safe in the world” is essential. The proof covers all candidate models in the specified set; it cannot protect against a true system omitted from that set. Nor does a theoretical stability guarantee substitute for validation against physical hardware, operational limits, and failures not represented in the equations.
What’s Next
The next step is to learn whether these compact information states can be computed quickly enough for demanding applications. Even a finite candidate set can be large, and updating a consistency set may require solving many feasibility tests. For continuum uncertainty, the paper’s fixed-dimensional representation applies only to structured classes. Identifying useful classes and efficient algorithms will determine how widely the approach travels.
A second test is empirical. The supplied material describes numerical experiments in the paper’s organization, but contains no experiment details or results from which to report numbers. Future work should compare the controller with fixed robust policies and other adaptive approaches on realistic systems, measuring not only cost but also safety margins, how quickly the model set contracts, and how often the action set actually expands.
The assumptions around disturbances deserve close attention too. Nonnegative disturbances make each observed transition a one-sided test of the model. In settings where disturbances can subtract as well as add, or where measurements are noisy, a candidate may be hard to eliminate reliably. Extending the method to those cases without weakening the safety claim is a substantial open problem.
There is also a deeper question about exploration. This controller learns from the data generated by safe operation, but the paper’s central promise is not that it actively chooses actions to maximize information. Its minimax formulation accounts for the consequences of uncertainty, yet the practical balance between informative actions and immediate performance depends on the chosen costs and model constraints. Understanding when safe operation alone yields enough information—and when it does not—will matter for deployment.
The paper’s most useful idea is that uncertainty need not be treated as a permanent tax on action. In the right class of systems, evidence can change not only what a controller believes, but what it is permitted to do. Turning that theorem into trustworthy practice will require stronger tests against imperfect models and real disturbances. But it opens a promising design principle: make caution responsive to what the system has actually shown, while keeping every action safe for the possibilities that remain.