Set-based State Estimation for Unknown Dynamical Systems from Noisy Output Data with Probabilistic Guarantees

Georgios Pantazis, Michelle S. ChongarXiv 2026 · arXiv

TL;DR

This paper introduces a data-driven framework for estimating the hidden states of a dynamical system without knowing its underlying physics or having access to "ground truth" state measurements for training. By leveraging scenario optimization and a novel "diameter surrogate" based on sensor geometry, the authors provide rigorous probabilistic guarantees that the true state remains within a calculated set of a specific maximum size.

What Problem Does This Paper Address?

In control theory and robotics, "state estimation" is the art of guessing what is happening inside a system (the state) based on what we can see (the outputs). Traditionally, this requires a mathematical model of the system's physics. While recent data-driven methods have tried to bypass the need for a model, they usually require "training data" where the internal states are known—a luxury often unavailable in real-world scenarios.

The fundamental challenge the authors tackle is: How do you build a "tight" bounding box around a system's state when you don't know the system's equations and you've never actually seen the state directly?

Without a model, it is difficult to ensure that your estimated set is both accurate (contains the true state) and useful (isn't so large that it's meaningless). This paper bridges that gap by providing a way to learn these sets from noisy input-output trajectories alone.

Methods

The authors propose a methodology that combines statistical learning with geometric constraints:

  1. Output Set Learning: They parameterize rectangular "output sets" (boxes) as functions of the system's input history. These boxes are designed to capture where the system's output is likely to land.
  2. The Diameter Surrogate: Since the internal state is unknown, its set diameter cannot be measured directly. The authors derive a mathematical bound showing that the state set's diameter is constrained by the width of the output sets and the geometry of the sensor map (the matrix $C$ that relates state to output). They use this bound as a "surrogate" in their optimization.
  3. Scenario Optimization with Relaxation: They formulate a Linear Program (LP) that seeks to minimize the state set diameter while maximizing the "coverage" of training trajectories. To handle noise and outliers, they use a "relaxation" approach, allowing some trajectories to fall outside the sets during training in exchange for tighter bounds overall.
  4. A Posteriori Guarantees: Using the framework of scenario optimization (specifically recent results by Garatti and Campi), they provide a distribution-free certificate. This means they can calculate exactly how likely it is that a brand-new, unseen trajectory will violate either the state inclusion or the diameter bounds.

Datasets

The paper does not use standard public datasets. Instead, it evaluates the methodology using a synthetic 2D nonlinear dynamical system.

  • System Dynamics: A complex nonlinear model involving tanh functions and Gaussian process noise.
  • Training/Evaluation: The model was trained on varying numbers of trajectories ($N = 500, 1000, 2000$) and evaluated against 2500 independent test trajectories.
  • Noise: The simulation included both measurement noise (uniform distribution) and process noise (Gaussian), testing the algorithm's robustness to stochastic disturbances.

Key Insights and Ideas

  • Geometry as a Proxy for Physics: The most striking insight is that you don't need to know the dynamics (how the state changes over time) to bound the state set diameter, provided you know the sensing map (how the state turns into an output). The sensor matrix $C$ provides enough geometric structure to "back out" a diameter bound.
  • The Coverage-Tightness Trade-off: There is a quantifiable tension between how many data points you want to "cover" and how small you want your estimation sets to be. The "relaxation penalty" ($\rho$) in their optimization acts as a dial to tune this trade-off.
  • No State Measurements Required: Unlike many "data-driven" reachability methods that require full state logging during a training phase, this method works entirely in the input-output space. This makes it significantly more applicable to systems where internal states are physically impossible to measure directly.

Key Results

  • Theoretical Validation: Numerical simulations confirmed that the empirical violation probability (how often the state actually left the predicted box) stayed strictly below the theoretical probabilistic bounds derived in the paper.
  • Generalization: As the number of training trajectories ($N$) increased, the empirical violations decreased, and the probabilistic bounds became less conservative (tighter).
  • Robustness to Noise: The method successfully narrowed the state sets over time even under process noise, suggesting that the attenuation of initial uncertainty can dominate the accumulation of noise in certain regimes.

Limitations and Future Work

  • Linear Sensing Map: A core assumption is that the sensor map ($h(x) = Cx$) is linear and known. While common, many systems have nonlinear sensors (e.g., cameras or rangefinders), which this current framework does not yet support.
  • Computational Complexity: While the formulation is a tractable Linear Program, the number of constraints grows with the number of training samples and the time horizon, which might pose challenges for very high-dimensional systems or extremely long trajectories.
  • Future Directions: The authors suggest extending this work to nonlinear sensing maps and exploring its applications in secure state estimation, where an attacker might be trying to inject false data into the system.

Summary updated

Abstract

We develop a methodology for set-based state estimation from noisy input-output trajectories, without a predefined model structure for the unknown state dynamics or access to training state measurements. We derive a bound on the state set diameter that depends on the sensor map and leverage it as a surrogate to formulate a tractable linear program balancing state set tightness against output trajectory coverage. Combining this certificate with deterministic or probabilistically certified measurement noise sets and leveraging the results from scenario optimization with relaxation, we provide joint a posteriori probabilistic guarantees for state inclusion and state set diameter. Numerical simulations corroborate our developed theoretical guarantees.
BibTeX
@misc{pantazis2026set,
  author = {Georgios Pantazis and Michelle S. Chong},
  title = {{Set-based State Estimation for Unknown Dynamical Systems from Noisy Output Data with Probabilistic Guarantees}},
  year = {2026},
  eprint = {2611.19716},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19716},
}
Figure