Neural Contraction Metrics with Formal Guarantees for Discrete-Time Nonlinear Dynamical Systems
Haoyu Li, Xiangru Zhong, Bin Hu, Huan Zhang · 2025
Read Neural Contraction Metrics with Formal Guarantees for Discrete-Time Nonlinear Dynamical Systems by Haoyu Li, Xiangru Zhong, Bin Hu, Huan Zhang (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-2504-17102.txt.
Introduction
Haoyu Li, Xiangru Zhong, Bin Hu, and Huan Zhang tackle a hard control problem, in a paper published in two thousand twenty-five. The problem is finding a proven measure of how fast a nonlinear system settles. Such measures, called contraction metrics, certify stability, robustness, and convergence.
The hard part is that complex systems rarely yield these metrics through older convex methods. Those methods also assume smooth dynamics, while modern controllers built from neural networks switch in steps and are not smooth. The authors instead train a neural network to propose a contraction metric for discrete-time systems. They then add formal checks so the learned metric carries a guarantee, not just a good fit. The work bridges learning and proof for systems steered by network-based controllers. It aims to make such controllers safer to trust.
Catalog record
| CREATOR | Haoyu Li, Xiangru Zhong, Bin Hu, Huan Zhang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-17102
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Neural Contraction Metrics with Formal Guarantees for Discrete-Time Nonlinear Dynamical Systems Haoyu Li, Xiangru Zhong, Bin Hu, Huan Zhang Abstract Contraction metrics are crucial in control theory because they provide a powerful framework for analyzing stability, robustness, and convergence of various dynamical systems. However, identifying these metrics for complex nonlinear systems remains an open challenge due to the lack of scalable and effective tools. This paper explores the approach of learning verifiable contraction metrics parametrized as neural networks (NNs) for discrete-time nonlinear dynamical systems. While prior works on formal verification of contraction metrics for general nonlinear systems have focused on convex optimization methods (e.g. linear matrix inequalities, etc) under the assumption of continuously differentiable dynamics, the growing prevalence of NN-based controllers, often utilizing ReLU activations, introduces challenges due to the non-smooth nature of the resulting closed-loop dynamics. To bridge this gap, we establish a new sufficient condition for establishing formal neural contraction metrics for general discrete-time nonlinear systems assuming only the continuity of the dynamics. We show that from a computational perspective, our sufficient condition can be efficiently verified using the state-of-the-art neural network verifier α , β \alpha,\!\beta -CROWN, which scales up non-convex neural network verification via novel integration of symbolic linear bound propagation and branch-and-bound. Built upon our analysis tool, we further develop a learning method for synthesizing neural contraction metrics from sampled data. Finally, our approach is validated through the successful synthesis and verification of NN contraction metrics for various nonlinear examples. \coltauthor 1 1 1 Equal Contribution. 1 1 footnotemark: 1 University of Illinois Urbana-Champaign (UIUC), Department of Electrical and Computer Engineering (ECE) and Department of Computer Science (CS) keywords Contraction Metrics, Formal Verification, Neural Networks, Nonlinear Systems 1 Introduction Contraction theory is a prominent framework for analyzing the stability and safety of control systems. It examines how trajectories of a system remain close and converge to each other ( Lohmiller and Slotine, 1998 ; Tran et al., 2018 ; Giesl et al., 2022 ; Manchester and Slotine, 2017 ; Tsukamoto et al., 2021 ; Bullo, 2022 ; Dawson et al., 2023 ) . In contrast to Lyapunov theory, which assesses stability with respect to a specific equilibrium point ( Lyapunov, 1992 ) , contraction analysis focuses on the existence of a Riemannian-type metric d in which the system is contracting ( Simpson-Porco and Bullo, 2014 ) . This approach certifies that the distance between any pair of trajectories diminishes over time, ensuring that all solutions converge toward each other regardless of their initial conditions. Contraction theory has been applied to analyze and synthesize controllers for both continuous-time ( Tsukamoto and Chung, 2020 ; Manchester and Slotine, 2017 ; Sun et al., 2021 ; Tsukamoto et al., 2021 ) and discrete-time systems ( Wei et al., 2021 ; Wei et al., 2022 ) . However, many of these approaches are often heuristic-based methods and lack formal verification. In this paper, we focus on synthesizing and formally verifying a neural network (NN) parametrized contraction metric that is valid over the largest possible domain for a given discrete-time nonlinear dynamical system. Although previous studies such as Giesl and Wendland (2019) , Giesl et al. (2024) , Pokkakkillath and Giesl (2024) , and Fitzsimmons and Liu (2024) employ numerical quadrature or learning-based approaches to synthesize contraction metrics, they share common limitations. First, they all implicitly assume the smoothness of the system dynamics. However, with the advancements of deep learning, NN-based controllers have been popularized ( Chang et al., 2019 ; Dai et al., 2021 ; Wu et al., 2023 ; Yang et al., 2024 ; Wang and Fazlyab, 2024 ) , and the use of ReLU/LeakyReLU activations ( Dai et al., 2021 ; Wu et al., 2023 ; Yang et al., 2024 ) just naturally introduce non-smoothness into the closed-loop dynamics. Consequently, traditional synthesis and verification techniques that rely on the Jacobian matrix of the dynamics are no longer applicable. Second, many existing works depend on convex optimization techniques such as linear matrix inequalities for the verification and learning of contraction metrics, which unavoidably increases the conservativeness of the resulting metrics. To overcome these challenges, we devise a novel verification condition for discrete-time contraction analysis that depends neither on matrix inequalities nor on the Jacobian of the dynamics. Besides handling nonsmooth systems, this novel formulation also significantly reduces the complexity of verification conditions. It allows the reformulation of contraction certification into a verification problem within a bounded set on a computation graph, making it well-suited for advanced neural network verification tools like \alpha,\!\beta -CROWN ( Zhang et al., 2018 ; Xu et al., 2020 ; Xu et al., 2021 ; Wang et al., 2021 ) , which has demonstrated its exceptional scalability in the verification of neural network controllers ( Everett et al., 2023 ; Yang et al., 2024 ) . In contrast to traditional SMT and MIP solvers, \alpha,\!\beta -CROWN is far more efficient for this task, as it leverages the neural network architecture and benefits from GPU acceleration. By combining our novel verification formulation with the efficient neural network verifier, we achieved the first formally verified contraction metrics for a state feedback system with an NN controller. To summarize, our main contributions include: • We develop a theoretical framework for formal contraction certificates applicable to potentially non-smooth discrete-time systems, proposing a novel verification condition for contraction metrics that eliminates the need for both matrix inequalities and the Jacobian of the dynamics; • Unlike previous work with verified contraction metrics ( Fitzsimmons and Liu, 2024 ) , which uses expensive verifiers like SMT, our new formulation allows the combination of advanced neural network verification tool \alpha,\!\beta -CROWN, enabling us to use effective learning-based NN contraction metrics while ensuring efficient formal verification of contraction ; • Our experiments on several autonomous and NN-controlled systems demonstrate that our learned NN contraction metric can produce a large verified region of contraction, sometimes close to the upper bound. For the first time in literature, we present verified neural contraction metrics for a neural network-controlled state-feedback pendulum system . 2 Preliminaries Notations Our paper primarily focuses on a discrete time-invariant closed-loop control system \displaystyle x_{k+1} \displaystyle=g(x_{k},\pi(x_{k}))=f(x_{k}) (1) where x_{k} is the system state, g denotes the open-loop dynamics, \pi denotes the policy, and f denotes the closed-loop nonlinear dynamics. We are interested in two cases: i) f involves neural networks, or ii) f is just a general nonlinear system. We denote the solution of the system starting with initial condition x=x_{0} to be \phi(k,x_{0}) , i.e. it satisfies \phi(k+1,x_{0})=f(\phi(k,x_{0})) . We denote the set of all positive semi-definite/positive definite matrices by \mathbb{S}_{+}^{n\times n} and \mathbb{S}_{++}^{n\times n} respectively. For two matrices A and B , we write A\prec B or A\preceq B to indicate that B-A is positive definite or positive semi-definite, respectively. A function d:\mathcal{X}\times\mathcal{X}\to\mathbb{R}_{\geq 0} is said to be a metric on \mathcal{X} if the following conditions holds: (1) d(x,x)=0,\,\forall x\in\mathcal{X} ; (2) d(x,y)=d(y,x),\forall x,y\in\mathcal{X} ; (3) d(x,z)\leq d(x,y)+d(y,z),\forall x,y,z\in\mathcal{X} . We denote the norm induced by a matrix P\in\mathbb{S}_{++}^{n\times n} as \|\cdot\|_{P} such that \|x\|_{P}=\sqrt{x^{\top}Px} , and we denote B(x;\epsilon) as the infinity norm ball around x with radius \epsilon , i.e. B(0;\epsilon)=\{y:\|y-x\|_{\infty}\leq\epsilon\} . Contraction Metrics This paper focuses on contraction