Formal Uncertainty Propagation for Stochastic Dynamical Systems with Additive Noise

Steven Adams, Eduardo Figueiredo, Luca Laurenti · 2025

Read Formal Uncertainty Propagation for Stochastic Dynamical Systems with Additive Noise by Steven Adams, Eduardo Figueiredo, Luca Laurenti (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-11219.txt.

Introduction

Steven Adams, Eduardo Figueiredo, and Luca Laurenti tackle formal uncertainty tracking for noisy moving systems in a 2025 paper. Engineers often know a system's starting state and its random kicks only roughly. The question is how that roughness grows as the system runs, possibly forever.

The authors wrap the starting and noise distributions in sets of nearby candidates, with closeness measured by a transport distance. Drawing on quantization, optimal transport, and stochastic methods, they build sets centered on blends of bell curves that provably contain the true sets. The guarantee holds for both short and endless horizons. Benchmark tests from control and machine learning show the method works in practice, giving safety checks for systems where bounds must be more than a guess.

Catalog record

CREATORSteven Adams, Eduardo Figueiredo, Luca Laurenti
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-11219

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Formal Uncertainty Propagation for Stochastic Dynamical Systems with Additive Noise Steven Adams, Eduardo Figueiredo, Luca Laurenti Abstract In this paper, we consider discrete-time non-linear stochastic dynamical systems with additive process noise in which both the initial state and noise distributions are uncertain. Our goal is to quantify how the uncertainty in these distributions is propagated by the system dynamics for possibly infinite time steps. In particular, we model the uncertainty over input and noise as ambiguity sets of probability distributions close in the ρ \rho -Wasserstein distance and aim to quantify how these sets evolve over time. Our approach relies on results from quantization theory, optimal transport, and stochastic optimization to construct ambiguity sets of distributions centered at mixture of Gaussian distributions that are guaranteed to contain the true sets for both finite and infinite prediction time horizons. We empirically evaluate the effectiveness of our framework in various benchmarks from the control and machine learning literature, showing how our approach can efficiently and formally quantify the uncertainty in linear and non-linear stochastic dynamical systems. I Introduction Modern control systems are commonly uncertain. This uncertainty is due not only to the inherently stochastic dynamics [ 1 ] , but also to the use of data-driven methods and statistical estimators [ 2 , 3 ] . The models of such systems are consequently not only stochastic, but the probability distributions of the various random variables are themselves uncertain [ 4 ] . In many applications, where failures may have catastrophic consequences, these uncertainties cannot be ignored but must be formally accounted for during the temporal evolution of the system. Unfortunately, propagating a set of distributions through non-linear stochastic dynamics is computationally intractable [ 5 ] . This leads to the main question in this paper: Can we efficiently estimate the state of a non-linear stochastic dynamical system with uncertain initial and noise distributions over a possibly infinite time horizon while providing formal guarantees of correctness? Various recent works have considered formal uncertainty propagation for dynamical systems. The more restrictive, but already intractable, problem of propagating known distributions over deterministic non-linear dynamics [ 5 ] has been well studied, using techniques such as moment matching [ 6 ] , discretization-based methods [ 7 , 8 ] or approximate numerical integration [ 9 , 10 ] . However, these techniques commonly lack formal guarantees or are computationally demanding due to the need to discretize the full state space, and, critically, do not account for the fact that the actual underlying (state or noise) distributions are rarely known exactly. The problem becomes even more challenging for uncertain distributions, as it requires propagating a set of distributions through the system dynamics. While this setting is receiving increasing attention, most existing approaches are either restricted to linear dynamics, lack formal guarantees, or do not scale to large prediction horizons [ 11 , 12 , 13 ] . For example, [ 11 ] proposes a framework for propagating sets of distributions close in Wasserstein distance through deterministic dynamics over multiple time steps, but the method is effectively limited to linear systems due to numerical tractability issues. In the stochastic setting, [ 12 ] approximate the evolving state distributions with mixture distributions, while providing guarantees of correctness in the Total Variation metric. However, the resulting bounds generally become uninformative for small time horizons. Recently, [ 13 ] proposed propagating an uncertain distribution through non-linear dynamics by approximating it with discrete distributions with quantification of the resulting uncertainty in Wasserstein distance. While the method supports general stochastic dynamics, it requires constructing Cartesian products of state- and noise discretizations, limiting scalability. In this paper, we present an algorithmic framework for formally propagating an uncertain input distribution through a non-linear dynamical system with possibly uncertain additive noise over a possibly infinite time horizon. Our approach relies on selecting a reference distribution from the input set, which is then propagated through the stochastic dynamics by approximating it with a mixture of Gaussian distributions. The resulting distribution is then used to construct a tractable convex set that is guaranteed to contain all possible distributions after propagation. To quantify uncertainty, we rely on the Wasserstein distance [ 14 ] , which not only guarantees that distributions close in this distance have similar moments, but also enables the use of results from optimal transport [ 14 , 15 ] . These results allow us to build a framework that not only efficiently propagates formal uncertainty sets but, for contracting systems, is even suitable for infinite-horizon propagation, as the radii of the resulting sets converge to a fixed point. We empirically illustrate the efficacy of our framework in various benchmarks. The experiments illustrate how our approach is able to successfully propagate uncertainty in various non-linear dynamical systems. In summary, the main contributions of this work are: • an algorithmic framework for obtaining tractable relaxations of sets of distributions propagated through non-linear dynamics with uncertain additive noise, with guarantees in Wasserstein distance, • a proof that the Wasserstein bounds converge to a fixed point for contracting dynamics, • extensive empirical validation of our approach on several benchmarks, including neural network dynamics with multiple hidden layers. II Notation For a vector x\in{\mathbb{R}}^{n} we use {{{x}}^{(i)}} to denote the i -th element of {{x}} . Furthermore, for region {\mathcal{X}}\subset{\mathbb{R}}^{n} , the indicator function for {\mathcal{X}} is denoted as {\mathbf{1}}_{{\mathcal{X}}}(x)=\begin{cases}1&\text{if }x\in{\mathcal{X}}\\ 0,&\text{otherwise}\end{cases}. Given a measurable space ({\mathcal{X}},{\mathcal{B}}({\mathcal{X}})) with {\mathcal{B}}({\mathcal{X}}) being the \sigma -algebra, we denote by {\mathcal{P}}({\mathcal{X}}) the set of probability distributions on ({\mathcal{X}},{\mathcal{B}}({\mathcal{X}})) . The set of probability distributions with finite moments up to order \rho\in\mathbb{N}_{>0} , {\mathcal{P}}_{\rho}({\mathcal{X}}) , is defined as the set of all \mathbb{P}\in{\mathcal{P}}({\mathcal{X}}) such that \int_{{\mathcal{X}}}\|{{x}}\|^{\rho}\mathbb{P}(d{{x}})<\infty , where \|\cdot\| is the L_{\rho} -norm. In this paper, for a metric space {\mathcal{X}} , {\mathcal{B}}({\mathcal{X}}) is assumed to be the Borel \sigma -algebra of {\mathcal{X}} . For a random variable {{x}} taking values in {\mathcal{X}} , \mathbb{P}_{{{x}}}\in{\mathcal{P}}({\mathcal{X}}) represents the probability measure associated to {{x}} . For \rho\in\mathbb{N}_{>0} and \mathbb{P},\mathbb{Q}\in{\mathcal{P}}_{\rho}({\mathcal{X}}) , the \rho -Wasserstein distance between \mathbb{P} and \mathbb{Q} is defined as: {\mathbb{W}}_{\rho}(\mathbb{P},\mathbb{Q})=\left(\inf_{\gamma\in\Gamma(\mathbb{P},\mathbb{Q})}\int_{{\mathcal{X}}\times{\mathcal{X}}}\|{{x}}-{{x}}^{\prime}\|^{\rho}\gamma(d{{x}},d{{x}}^{\prime})\right)^{\frac{1}{\rho}} (1) where \Gamma(\mathbb{P},\mathbb{Q})\subset{\mathcal{P}}({\mathcal{X}}\times{\mathcal{X}}) represents the set of probability distributions with marginal distributions \mathbb{P} and \mathbb{Q} . For \mathbb{P}\in{\mathcal{P}}_{\rho}({\mathcal{X}}) and \theta\geq 0 the set of distributions closer than \theta to \mathbb{P} in the \rho -Wasserstein distance, also called \rho -Wasserstein ambiguity set , is denoted by {\mathbb{B}}_{\theta}(\mathbb{P})=\left\{\mathbb{Q}\in{\mathcal{P}}_{\rho}({\mathcal{X}})\mid{\mathbb{W}}_{\rho}(\mathbb{P},\mathbb{Q})\leq\theta\right\}\subset{\mathcal{P}}_{\rho}({\mathcal{X}}). (2) For N\in\mathbb{N} , \Pi^{N}=\{\pi\in{\mathbb{R}}^{N}_{\geq 0}\;:\;\sum_{i=1}^{N}{{{\pi}}^{(i)}}=1\} is the N -simplex. A discrete probability distribution \mathbb{D}\in{\mathcal{P}}({\mathcal{X}}) is defined as \mathbb{D}=\sum_{i=1}^{N}{{{\pi}}^{(i)}}\delta_{{{c}}_{i}} , where \delta_{{c}} is the Dirac delta function centered at location {{c}}\in{\mathcal{X}} and \pi\in\Pi^{N} . The set of discrete probability distributions on {\mathcal{X}} with at most N locations is denoted as {\mathcal{D}}_{N}({\mathcal{X}})\subset{\mathcal{P}}({\mathcal{X}}) . For measurable spaces ({\mathcal{X}},{\mathcal{B}}({\mathcal{X}})) and ({\mathcal{Y}},{\mathcal{B}}({\mathcal{Y}})) , a probability distribution {\mathbb{P}}\in{\mathcal{P}}({\mathcal{X}}) and a measurable function g:{\mathcal{X}}\rightarrow{\mathcal{Y}}\subseteq{\mathbb{R}}^{q} , we use g\#\mathbb{P} to denote the push-forward measure of \mathbb{P} by g , i.e., the measure on {\mathcal{Y}} such that for all {\mathcal{A}}\in{\mathcal{B}}({\mathcal{Y}}) , (g\#\mathbb{P})({\mathcal{A}})=\mathbb{P}(g^{-1}({\mathcal{A}})) . For probability distributions \mathbb{P},\mathbb{Q}\in{\mathcal{P}}({\mathcal{X}}) , we use \mathbb{P}\ast\mathbb{Q}\in{\mathcal{P}}({\mathcal{X}}) to denote the convolution of \mathbb{P} and \mathbb{Q} , i.e., (\mathbb{P}\ast\mathbb{Q})({\mathcal{A}})=\int_{{\mathcal{X}}\times{\mathcal{X}}}{\mathbf{1}}_{\mathcal{A}}({{x}}+{{x}}^{\prime})\mathbb{P}(d{{x}})\mathbb{Q}(d{{x}}^{\prime}) for all {\mathcal{A}}\in{\mathcal{B}}({\mathcal{X}}) . III Problem Formulation We consider a non-linear stochastic process with additive noise described as: {{x}}_{k+1}=f({{x}}_{k})+{{\omega}}_{k},\quad{{x}}_{0}\sim\mathbb{P}_{{{x}}_{0}},{{\omega}}_{k}\sim\mathbb{P}_{{{\omega}}} (3) where, for state space {\mathcal{X}}\subseteq{\mathbb{R}}^{n} , f:{\mathcal{X}}\mapsto{\mathcal{X}} is a possibly non-linear measurable piecewise Lipschitz continuous function

Read the complete text