PaperlayerView exact revision on arXiv

Source-linked AI summary

Contraction Theory for Nonlinear Stability Analysis and Learning-based Control: A Tutorial Overview

Hiroyasu Tsukamoto, Soon-Jo Chung, Jean-Jacques E. Slotine

arXiv:2110.00675v8cs.LGcs.ROeess.SYmath.OC

TL;DR

The paper addresses how nonlinear, time-varying systems can obtain incremental exponential stability and robustness guarantees, especially when control and models involve learning errors or unknown dynamics. It provides a tutorial synthesis of contraction theory, convex metric-construction methods, and neural or data-driven control frameworks. The resulting guarantees bound target-to-solution trajectory distances exponentially, including under disturbances and learning errors.

  • Problem

    The paper addresses stability and robustness analysis for nonlinear, non-autonomous systems and learning-based control when dynamics may be uncertain, poorly modeled, or unknown.

  • Method

    The paper reviews contraction theory and develops convex optimization, neural contraction metric, adaptive, and data-driven methods for constructing metrics and associated control or estimation laws.

  • Results

    Contraction-theory learning-based controllers formally keep system trajectories in an exponentially convergent bounded tube around the target despite external disturbances and learning errors.

  • Takeaways & Limitations

    The tutorial shows how contraction theory supplies formal robustness and stability guarantees for learning-based and data-driven automatic control.

  • Takeaways & Limitations

    CV-STEM requires solving an optimization problem or nonlinear system at each time instant, limiting suitability for systems with constrained online computation.

Abstract

from arXiv · show

Contraction theory is an analytical tool to study differential dynamics of a non-autonomous (i.e., time-varying) nonlinear system under a contraction metric defined with a uniformly positive definite matrix, the existence of which results in a necessary and sufficient characterization of incremental exponential stability of multiple solution trajectories with respect to each other. By using a squared differential length as a Lyapunov-like function, its nonlinear stability analysis boils down to finding a suitable contraction metric that satisfies a stability condition expressed as a linear matrix inequality, indicating that many parallels can be drawn between well-known linear systems theory and contraction theory for nonlinear systems. Furthermore, contraction theory takes advantage of a superior robustness property of exponential stability used in conjunction with the comparison lemma. This yields much-needed safety and stability guarantees for neural network-based control and estimation schemes, without resorting to a more involved method of using uniform asymptotic stability for input-to-state stability. Such distinctive features permit the systematic construction of a contraction metric via convex optimization, thereby obtaining an explicit exponential bound on the distance between a time-varying target trajectory and solution trajectories perturbed externally due to disturbances and learning errors. The objective of this paper is, therefore, to present a tutorial overview of contraction theory and its advantages in nonlinear stability analysis of deterministic and stochastic systems, with an emphasis on deriving formal robustness and stability guarantees for various learning-based and data-driven automatic control methods. In particular, we provide a detailed review of techniques for finding contraction metrics and associated control and estimation laws using deep neural networks.

1. Introduction

The paper presents contraction theory as a differential, metric-based framework for incremental exponential stability in nonlinear systems. It emphasizes parallels with linear time-varying analysis and develops robustness guarantees for learning-based control.

  • Contraction theory uses a quadratic Lyapunov function of differential states defined by a uniformly positive definite Riemannian contraction metric.This characterizes incremental exponential convergence among multiple nonlinear system trajectories.
  • Differential dynamics enable Linear Time-Varying systems-type techniques for nonlinear stability analysis and control or estimation synthesis.Exponential stability and the comparison lemma simplify analysis relative to approaches relying on ISS or uniform asymptotic stability.
  • The tutorial develops provable incremental exponential robustness and stability guarantees for learning-based and data-driven automatic control.It also presents systematic convex optimization for constructing contraction metrics and differential Lyapunov functions in deterministic and stochastic systems.
  • The paper is organized around contraction-theory foundations, nonlinear control and estimation, and learning-based control applications.The supplied overview identifies the paper's organizational purpose without detailing the individual sections.

Part I: Nonlinear Stability Analysis (Sec. 2–4)

Part I develops contraction-theory foundations for nonlinear robustness and stability analysis, then formulates control and estimation design through convex optimization.

  • Nonlinear Stability Analysis: Sections 2–4 establish fundamental contraction-theory results and derive convex optimization methods for constructing contraction metrics.The framework addresses nonlinear robustness, stability, optimal feedback control, and estimation.

Part II: Learning-based Control (Sec. 5–9)

Part II extends contraction theory to learning-based and data-driven control, including neural contraction metrics, adaptive methods, learned models, estimation, and motion planning. The framework provides formal robustness guarantees while addressing online computational limits.

  • Learning-based Control: Learning-based control, estimation, and motion planning use deep neural networks to design contraction metrics, while adaptive control addresses parametric uncertainty.The paper also proposes model-free contraction-theory versions for learning-based and data-driven control.
  • Contraction Theory: Contraction theory defines incremental exponential stability as convergence of all solution trajectories to one trajectory regardless of initial conditions.The paper presents this as a generalized stability notion rather than a replacement for traditional Lyapunov or incremental asymptotic stability.
  • Contraction Theory: Exponential stability and the comparison lemma support intuitive ISS and finite-gain Lp stability proofs for autonomous and non-autonomous nonlinear systems.This avoids relying on uniform asymptotic stability in the described analyses.
  • Contraction Theory: Contraction theory simplifies nonlinear input-output and H∞ stability analysis by using a quadratic differential-state Lyapunov function.The paper contrasts this with finding a Lyapunov function through a Hamilton-Jacobi inequality.
  • Construction of Contraction Metrics: Analytical and optimization-based methods construct contraction metrics, including SDRE, CV-STEM, and differential feedback through Control Contraction Metrics.CV-STEM minimizes an upper bound on steady-state distance between perturbed and unperturbed trajectories.
  • Neural Contraction Metrics: Numerical metric-design schemes require solving optimization problems or nonlinear equations online, motivating approximations such as SOS functions and neural networks.This limitation is central to the transition toward neural contraction metrics.
  • Learning-based Guarantees: Treating learning error as an external disturbance yields exponentially bounded target-to-learned trajectory distance with steady-state bounds proportional to learning error.NCM and NSCM are introduced for real-time computation of contraction metrics under deterministic and stochastic disturbances.

2. Contraction Theory

This section reviews prior contraction-theory results that simplify and generalize Lyapunov theory for subsequent robustness and stability analyses.

  • The paper reviews foundational contraction-theory results for use in formal robustness and stability guarantees across later sections.The review is explicitly presented as a simplification and generalization of Lyapunov theory.

2.1. Fundamentals

Contraction theory studies differential dynamics through a uniformly positive definite metric, converting nonlinear stability analysis into conditions on a quadratic differential Lyapunov function. A contraction metric is equivalent to incremental exponential convergence of all solution trajectories, including systems without stable fixed points.

  • Differential dynamics: Contraction theory analyzes smooth time-varying nonlinear systems through differential displacements and their associated differential dynamics.The system is assumed smooth, ensuring local existence and uniqueness of solutions; differential displacements are infinitesimal variations at fixed time.
  • Contraction metric: The differential Lyapunov function V = δx⊤M(x, t)δx is quadratic in δx, reducing the search for a Lyapunov function to finding a positive-definite matrix M.This contrasts with general nonlinear Lyapunov theory, where the candidate V(x, t) may be any scalar function and can require solving a PDE.
  • Deterministic contraction: A uniformly positive definite M(x, t) satisfying the contraction condition makes all solution trajectories converge exponentially to one trajectory, with rate α, regardless of initial conditions.The converse also holds, yielding a necessary and sufficient characterization of incremental exponential convergence.
  • Deterministic contraction: Differential lengths decay as ∥δz(t)∥≤∥δz0∥e^−αt, and path integration extends this exponential convergence from infinitesimal displacements to finite distances between trajectories.Thus, the length of any finite path converges exponentially to zero from arbitrary initial conditions.
  • Incremental exponential stability: Contraction implies incremental exponential stability between any pair of trajectories, rather than stability only with respect to an equilibrium point.The framework also permits contracting systems that have no stable fixed point.
  • Partial contraction: Partial contraction analyzes selected system components while treating the remaining state as a time-varying parameter, providing stability insight when full contraction is difficult to prove.Its convex contraction condition can support numerical construction of contraction metrics even when analytic construction is challenging.

2.2. Path-Length Integral and Robust Incremental Stability Analysis

Path-integral formulations convert differential contraction into Lyapunov-like measures of trajectory separation, supporting exponential convergence and explicit robustness bounds under deterministic and stochastic perturbations.

  • Path-integral formulation: The transformed squared length and path integral connect contraction metrics to incremental exponential stability between arbitrary solution trajectories.The shortest path integral defines the Riemannian distance along a minimizing geodesic.
  • Path-integral formulation: The comparison lemma converts differential decay inequalities into incremental exponential stability for every pair of trajectories.The contraction conditions yield decay rates for both transformed squared length and path-integral Lyapunov-like functions.
  • Deterministic perturbation: Bounded transformed disturbances make the perturbed trajectory converge exponentially to a bounded error ball.The result requires Θd ∈ L∞ and uses uniform bounds on the contraction metric.
  • Stochastic perturbation: The stochastic analysis uses virtual systems and infinitesimal-generator inequalities to extend deterministic incremental stability arguments to independent Wiener-driven trajectories.Comparison and Markov inequalities produce the stated moment and probability bounds.
  • Scope of stochastic analysis: The stochastic framework can be generalized beyond second moments and Gaussian white noise to other moment orders and selected noise processes.The paper notes extensions to p-th moments, compound Poisson shot noise, and bounded-measure Lévy noise.
  • Stochastic perturbation: Stochastic contraction provides expected-error and probability bounds under uniformly positive definite, Lipschitz contraction metrics and bounded diffusion terms.The stochastic theorem imposes metric regularity and a strengthened contraction condition accounting for diffusion and metric-derivative terms.

2.3. Finite-Gain Lp Stability and Hierarchical Contraction

Contraction theory extends robustness analysis to finite-gain Lp and ISS properties, while hierarchical contraction propagates disturbance bounds through interconnected contracting subsystems.

  • Finite-gain Lp stability: Contraction theory simplifies ISS and finite-gain Lp analysis by using exponential stability and the comparison lemma instead of uniform asymptotic stability.The resulting analysis also supports nonlinear H∞ control and disturbance attenuation.
  • Finite-gain Lp stability: A contracting system with bounded perturbations is finite-gain Lp stable for p ∈ [1, ∞] and also exhibits input-to-state stability.The gain bound is independent of the time horizon under the stated metric bounds.
  • Hierarchical contraction: Hierarchical contraction bounds the path-length errors of two interconnected contracting dynamics when the interconnection and disturbances satisfy the stated boundedness conditions.Each subsystem receives a bound determined by its contraction rate and transformed disturbance magnitude.
  • Hierarchical contraction: The hierarchical finite-gain Lp result extends recursively to an arbitrary number of hierarchically combined groups.The bound uses subsystem-specific factors ζ_i determined by p and each contraction rate α_i.

2.4. Contraction Theory for Discrete-time Systems

Discrete-time contraction mirrors the continuous-time framework by using uniformly positive definite discrete metrics and yields robust trajectory-separation bounds under bounded perturbations.

  • Robustness bounds: A virtual path between unperturbed and perturbed discrete trajectories enables a discrete robust-contraction bound analogous to the continuous-time result.The path is parameterized by μ ∈ [0,1] with the two trajectories as particular solutions.
  • Discrete-time contraction: The discrete-time formulation defines contraction through a smooth coordinate transformation and a uniformly positive definite metric M_k(x_k,k).The contraction factor satisfies α ∈ (0,1), with metric bounds imposed uniformly over states and time indices.
  • Scope: The discrete-time techniques also support hybrid systems and stochastic nonlinear systems when combined with the corresponding continuous-time robustness theorems.For sufficiently small discretization intervals, stochastic discrete-time contraction reduces to the continuous-time case.
  • Design implication: The steady-state robust bounds in deterministic, stochastic, and discrete-time contraction depend on the ratio of the metric bounds.This dependence is used later in convex optimization-based control and estimation synthesis.

3. Robust Nonlinear Control and Estimation

The paper develops contraction-based nonlinear control and estimation methods using SDC and CCM formulations, convex metric synthesis, and extensions to non-affine, stochastic, and learning-based settings.

  • Robust control and estimation: Contraction theory provides explicit trajectory-separation bounds useful for robust feedback control, including H∞ control and disturbance attenuation.The framework targets nonlinear systems subject to deterministic or stochastic perturbations.
  • Control formulations: The tutorial focuses on control-affine systems because their contraction-based controller designs are less complicated, while non-affine systems require implicit or augmented-input treatments.Non-affine control can use ˙u as an input or solve an implicit controller equation iteratively.
  • Control formulations: For non-affine systems, iterative learning can solve an implicit stabilizing controller equation for unknown residual dynamics without deteriorating the stated stability performance.The paper presents this as a learning-based approach for real-time implementation concerns.
  • SDC formulation: The generalized SDC formulation expresses nonlinear differences through state-dependent matrices and supports optimal control and estimation laws with global stability.Its extended linear form keeps the analysis close to LTV systems theory.
  • Metric synthesis: The tutorial derives convex optimization-based contraction conditions and relates them to bounded-real and KYP lemmas for control and estimation synthesis.It also reviews applications to deep-neural-network metrics, uncertain systems, and model-free learning-based control.

4. Convex Optimality in Contraction Theory

The paper formulates contraction-metric design and robust control or estimation synthesis as convex optimization problems. CV-STEM minimizes steady-state tracking-error bounds while supporting deterministic and stochastic systems, with analogous control-effort trade-offs and practical computational considerations.

  • CV-STEM formulation: CV-STEM optimally designs contraction metrics by minimizing steady-state tracking-error bounds subject to convex contraction constraints.The framework covers SDC- and CCM-based feedback control and state estimation for deterministic and stochastic disturbances.
  • Optimization structure: The optimization uses decision variables ν, χ, and W̄, with deterministic and stochastic systems imposing different contraction constraints.Here W̄=νW, ν=m, χ=m/m, and P is a performance-based convex cost function.
  • Design trade-offs: Increasing c1 relative to c0 reduces control effort but increases steady-state tracking error, whereas decreasing c1 has the opposite trade-off.The weights play an analogous role to state and control costs in LQR.
  • State estimation: The same contraction-theoretic approach designs nonlinear state estimators and yields convex steady-state upper bounds on estimation error.The estimator uses a positive definite matrix M satisfying a contraction-metric constraint and exploits control-estimation duality in differential dynamics.
  • Implementation: The CV-STEM requires solving an optimization problem at each time instant, while neural contraction metrics approximate this scheme for real-time implementation.The approximation is intended to retain formal stability guarantees associated with the CV-STEM framework.

5. Contraction Theory for Learning-based Control

The paper applies contraction theory to learning-based control, estimation, and system identification by representing learning mismatch as a perturbation of contracting dynamics. This yields formal tracking-error guarantees under disturbances and learning errors, while neural contraction metrics address the online computational burden.

  • Common formulation: The framework covers learning-based tracking control, state estimation, and system identification through a common virtual-system formulation.The virtual path connects particular solutions such as the target and actual states, the true and estimated states, or learned and underlying dynamics.
  • Learning-based control: Learning-based control is analyzed by modeling the difference between learned and ideal feedback laws as a learning-error perturbation.The ideal closed loop is assumed contracting, and the perturbation is ΔL=B(x,t)(uL−u∗).
  • Error bounding: Learning errors can be bounded on compact sets when the learned and target functions are bounded or Lipschitz under the stated assumptions.The perturbation typically consists of the learning error multiplied by a bounded or Lipschitz continuous function.
  • Comparison and limitation: A noncontracting framework can provide mathematical guarantees, but its bound may diverge through the exponential term e^(Lg+ϵℓ1)t and become conservative.
  • Robust guarantees: Contraction-based tracking-error bounds are exponentially bounded linearly in learning error even when external disturbances are present.
  • Robust guarantees: Contraction theory guarantees trajectories remain in a tube with an exponentially convergent bounded radius around the target despite external disturbances and learning errors.The bounds become tighter as the learning errors ϵℓ0 and ϵℓ1 decrease with more training data used to verify the learning condition.
  • Neural contraction metrics: Neural Contraction Metrics model CV-STEM optimization with deep neural networks to support real-time computation while retaining formal robustness and optimality guarantees despite modeling errors.

6. Learning-based Robust Control and Estimation

This section presents neural contraction metrics for learning-based control and estimation, replacing online convex optimization with learned metric evaluations while retaining robustness and stability guarantees under stated learning-error conditions.

  • Neural Contraction Metrics: A Neural Contraction Metric is a deep neural network model of a contraction metric trained from convex-optimization solutions for control or estimation.The framework can use sampled states and metrics obtained from CV-STEM formulations.
  • Neural Contraction Metrics: NCM evaluation requires one function evaluation per time instant to obtain the contraction metric, avoiding online optimization and enabling real-time feedback control and state estimation.This supports applications requiring computationally efficient online operation.
  • Robustness and Stability Guarantees: NCM-based control and estimation inherit incremental robustness and stability from the internal contracting property of the learned metric.The paper derives corresponding learning-based control and estimation theorems for deterministic and stochastic systems.
  • Robustness and Stability Guarantees: Metric construction errors are converted into explicit control and estimation error bounds when the learning error is sufficiently small on compact state, input, and time sets.For control, the bound uses ϵℓ1 = ¯ρ¯b^2ϵℓ; for estimation, it uses ϵℓ1 = ¯ρ¯c^2ϵℓ.
  • Examples: In simulations, NCM and CV-STEM estimation errors remain below the optimal steady-state bound, while NCM keeps disturbed spacecraft trajectories inside the prescribed error tube at lower computational cost.The reported examples cover Lorenz state estimation and spacecraft motion around obstacles.
  • NCMs as Lyapunov Functions: Using an NCM directly as a contraction metric makes the CLF relaxation variable decrease with smaller learning error, reaching zero when NCM modeling error is zero.This provides an alternative stability perspective to results based on the original CV-STEM metric or CCM.

7. Learning-based Robust Motion Planning

This section combines contraction-theoretic robust tracking with learned motion-planning policies. LAG-ROS provides exponentially bounded tracking and state-constraint guarantees without online target-trajectory computation, while retaining lower computational cost than robust tube-based planning.

  • Planner Motivation: Robust tube-based planning keeps disturbed trajectories in an exponentially bounded tube around a target but requires online computation of the target trajectory and control input.This requirement can be unrealistic for systems with limited computational resources.
  • LAG-ROS Design: LAG-ROS bridges learning-based and robust tube-based planning by learning an optimal robust feedback control law from state, local-environment, and time information.Its training data are sampled from contraction-theoretic robust control constructions.
  • Stability Guarantees: LAG-ROS achieves exponentially bounded trajectory-tracking error through its internal feedback structure, unlike feedforward learning-based planners whose bound increases exponentially with time.The paper identifies this property as relevant to safety-critical guidance and control.
  • Cart-Pole Example: In the cart-pole example, LAG-ROS and robust tube-based planning satisfy the exponential bound for all times, whereas the learning-based planner has a diverging bound and increasing deviation.The comparison averages tracking errors over 50 simulations and reports a small standard deviation for LAG-ROS and robust tube-based planning.
  • LAG-ROS Design: LAG-ROS achieves online control without solving a motion-planning problem while retaining robustness and stability properties associated with the robust feedback controller.The paper reports significantly lower computational cost than the robust tube-based planner.
  • State Constraint Satisfaction: LAG-ROS satisfies state constraints under its sampled-target conditions, even with learning error, and can operate across environments using localized observations from a single policy.The localization method extracts local environment information from global information for distributed use.
  • Scope and Limitations: The framework assumes fixed known disturbance-size bounds, which can produce conservative state constraints and motivates learning-based estimation of unknown disturbance components.This scope boundary is discussed for both deterministic and stochastic disturbances.

8. Learning-based Adaptive Control

The section develops contraction-theory-based adaptive control for nonlinear systems with parametric uncertainty, including deep-learning-based adaptive metrics and explicit robustness guarantees. Under stated assumptions, bounded disturbances and learning errors yield exponential tracking bounds, while zero learning error can recover asymptotic stability.

  • Adaptive control framework: Deep learning-based adaptive control improves the real-time performance of robust contraction-based control for nonlinear systems with parametric uncertainty.The method targets both model-based and model-free robust-control results and can trade computational cost for generality through differential state feedback.
  • Adaptive control framework: The adaptive-control framework assumes structured parameter uncertainty, including a nominal parameter estimate and, in the basic formulation, matched uncertainty through the input distribution.The nominal model can remain useful when modeling errors are bounded, while the adaptive formulation addresses cases where those assumptions may fail.
  • Robustness and stability guarantees: With nonzero disturbance scaling, the system is robust against bounded deterministic and stochastic disturbances; with zero learning error and σ=0, tracking is asymptotically stable.These guarantees hold in the compact set S under the stated metric, boundedness, and learning-error conditions.
  • Robustness and stability guarantees: The contraction proof uses a virtual system and a differential Lyapunov function to convert adaptation and disturbance terms into an exponential error bound.The comparison argument yields a differential inequality of the form V̇ℓ≤−αaVℓ+σ√γ ϑ̄, which produces bound (8.5).
  • Extensions: The section extends adaptive contraction techniques to multiplicatively separable systems and indicates compatibility with robotics, high-fidelity spacecraft, basis-function, and DNN models.The extension supports robustness to bounded external disturbances and asymptotic stability under adaptive CV-STEM conditions.
  • Adaptive neural contraction metrics: An adaptive neural contraction metric (aNCM) is a deep-neural-network model of an optimal parameter-dependent contraction metric obtained through an adaptive convex optimization formulation.Including the parameter estimate ˆθ among the metric inputs increases representational power relative to a nominal-system NCM.

9. Contraction Theory for Learned Models

This section applies contraction theory when nonlinear dynamics are learned from trajectory data or contain unknown control-non-affine residuals. It shows that contracting learned models and spectrally normalized neural residuals can support explicit robustness and stability guarantees.

  • Learned models: Model-free control is motivated by settings where the true dynamics are poorly modeled or completely unknown and only trajectory data are available.The learned model fL is used in place of the unknown true dynamics ftrue.
  • Learned models: If the learned dynamics render the modeled system contracting and the learning error is bounded, contraction theory provides robustness and stability guarantees in a compact set.Theorem 9.1 uses a bounded contraction metric and an error condition with ϵℓ1=0 to establish the result.
  • Learned models: The resulting tracking bound tightens as the verified learning-error bound ϵℓ0 decreases with more training data.The paper explicitly connects smaller verified modeling error to a tighter bound in Theorem 9.1.
  • Robust control with DNN residuals: For unknown control-non-affine residuals, the true dynamics are decomposed into a known control-affine part and an unknown residual modeled by a spectrally normalized DNN.The controller combines a nominal stabilizing input with the learned residual model, while Lipschitz constraints support the contraction argument.
  • Robust control with DNN residuals: A Lipschitz residual model with the required bound makes the discrete-time controller a contraction mapping and yields explicit robustness to deterministic and stochastic disturbances.The controller is analyzed through the mapping F(u)=u∗−rL and its Lipschitz constant in u.
  • Robust control with DNN residuals: The learned-residual controller extends contraction-based robust control to systems whose DNN components depend nonlinearly on the control input.The Lagrangian example states that the tracking error is exponentially bounded under the learned-residual design.
  • Learning contraction metrics: Learning contraction metrics from sampled trajectories enlarges the region where the learned metric satisfies contraction as the contraction-violation error decreases.For uniform samples, the paper reports ϵℓ≤O(k · polylog(N)/N) decay rates for several function classes.

10. Concluding Remarks

The paper presents contraction theory as a framework that simplifies incremental exponential stability analysis and supports formal guarantees for learning-based and data-driven control, estimation, and motion planning.

  • Contraction theory generalizes and simplifies Lyapunov-based methods for incremental exponential stability analysis of nonlinear non-autonomous systems.Its differential dynamics resemble linear time-varying systems, enabling LMI and convex optimization formulations.
  • LMI and convex optimization formulations support systematic nonlinear control and estimation synthesis through contraction metrics.These formulations provide a constructive route to metric-based stability analysis.
  • Contraction-based machine learning control methods augment existing learning frameworks with formal robustness and stability guarantees.The paper presents multiple methods that use contraction-theoretic results for learning-based control.
  • The surveyed methods provide mathematical tools for formally guaranteeing safety and stability in learning-based and data-driven control, estimation, and motion planning.The stated application scope includes high-performance robotic and autonomous systems.
Loading 2110.00675v8…