Control Contraction Metric: Theory & Applications

• Control Contraction Metric (CCM) is a smooth, uniformly bounded Riemannian metric that certifies exponential contraction of infinitesimal state displacements in control-affine nonlinear systems.
• The framework integrates differential feedback along minimizing geodesics to achieve trajectory tracking and stabilization with universal exponential convergence guarantees.
• Its dual convex formulation via LMIs facilitates robust synthesis and supports extensions to output-feedback, distributed, and data-driven control designs.

A Control Contraction Metric (CCM) is a smooth, uniformly bounded Riemannian metric for a control-affine nonlinear system under which infinitesimal state displacements can be made to contract exponentially by an appropriate differential feedback. In the CCM framework, trajectory tracking and stabilization are formulated on the tangent bundle rather than directly on the state space: one certifies contraction of differential dynamics and then integrates the resulting differential control along a path, typically a minimizing geodesic, between the current and target states. This yields constructive nonlinear feedback laws with universal exponential stabilizability guarantees, convex dual formulations, and coordinate-invariant criteria for nonlinear control design (Manchester et al., 2013, Manchester et al., 2015).

1. Definition and differential formulation

For the control-affine system

$$\dot x \;=\; f(x,t)\;+\;B(x,t)\,u,\qquad x\in\mathbb R^n,\;u\in\mathbb R^m,$$

the differential dynamics along a trajectory $(x(t),u(t))$ are

$$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$

A Riemannian metric is a smooth, uniformly bounded family of inner products

$$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$

A metric $M(x,t)$ is a CCM with rate $\lambda>0$ if, for all $(x,u,t)$,

$$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$

This kernel-form condition states that contraction must hold in differential directions orthogonal, with respect to the metric, to the control distribution (Manchester et al., 2015).

The dual metric $W(x,t)=M(x,t)^{-1}$ is central because it yields convex formulations. A standard sufficient dual condition introduces a differential feedback multiplier $Y(x,t)$ and requires

$(x(t),u(t))$0

from which one recovers the differential gain $(x(t),u(t))$1 (Manchester et al., 2015). In the earlier universal-stabilizability formulation, an equivalent parameterization uses a nonnegative scalar $(x(t),u(t))$2 through

$(x(t),u(t))$3

or, in dual form,

$(x(t),u(t))$4

(Manchester et al., 2013).

A strong CCM supplements contraction with Killing-field conditions on the input vector fields,

$(x(t),u(t))$5

so that the effect of the control channels on the differential energy is affine and the contraction test can be imposed in the null-space of $(x(t),u(t))$6 (Shiromoto et al., 2018).

2. Universal exponential stabilizability and controller realization

The central theorem of the convex CCM construction states that if there exists a smooth dual metric $(x(t),u(t))$7 and matrix function $(x(t),u(t))$8 such that $(x(t),u(t))$9 is uniformly bounded and the pointwise LMI

$$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$0

holds for all $$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$1, then every trajectory is open-loop controllable to any other trajectory with exponential rate $$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$2, there exists a sampled-data feedback controller attaining global exponential tracking with rate $$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$3, and there exists a continuous feedback law, defined almost everywhere, with the same guarantee. In all cases the overshoot is $$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$4 (Manchester et al., 2015).

The finite controller is obtained by integrating a differential controller along a path connecting the target trajectory $$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$5 and the current state $$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$6. With differential gain $$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$7, the exact realization is

$$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$8

where $$\delta\dot x \;=\; A(x,u,t)\,\delta x \;+\; B(x,t)\,\delta u, \quad A(x,u,t)\;=\;\frac{\partial f}{\partial x}(x,t)\;+\sum_{i=1}^m u_i\,\frac{\partial b_i}{\partial x}(x,t).$$9 is the minimizing geodesic under the metric $$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$0 joining $$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$1 to $$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$2 (Shiromoto et al., 2018). In the scalar-multiplier formulation, the integrated controller can be written as

$$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$3

and the associated Riemannian distance $$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$4 satisfies $$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$5 (Manchester et al., 2013).

These results are universal in the sense used by the original formulation: every nominal trajectory of the open-loop system can serve as the reference trajectory. The emphasis is therefore not on stabilization of a single equilibrium alone, but on reference-independent incremental exponential stability of all trajectories (Manchester et al., 2013).

3. Convexity, intrinsic structure, and relation to adjacent frameworks

A defining feature of CCMs is that the search problem becomes convex in dual variables. The metric $$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$6 is a $$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$7-tensor, and under a smooth coordinate change $$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$8 with Jacobian $$\langle \delta_1,\delta_2\rangle_{x,t} \;=\;\delta_1^T\,M(x,t)\,\delta_2, \quad \alpha_1 I \preceq M(x,t)\preceq \alpha_2 I.$$9, it transforms as

$M(x,t)$0

while the dual metric transforms as $M(x,t)$1. By direct substitution, the CCM LMI is preserved under smooth diffeomorphisms $M(x,t)$2 and affine feedback transformations $M(x,t)$3, so the conditions are intrinsic to the geometry of the system rather than artifacts of a particular parametrization (Manchester et al., 2015).

For feedback-linearizable systems,

$M(x,t)$4

with $M(x,t)$5 controllable, one may choose a constant LTI Lyapunov metric $M(x,t)$6 satisfying

$M(x,t)$7

and pull it back to a CCM for the original nonlinear plant. Accordingly, CCM existence is necessary and sufficient, up to technical regularity, for exact feedback linearizability plus stabilizability (Manchester et al., 2015).

CCMs also connect directly to control Lyapunov functions. For any fixed target trajectory $M(x,t)$8, the Riemannian distance

$M(x,t)$9

is a control Lyapunov function, while the differential Lyapunov condition is imposed on variational dynamics rather than on the state directly (Manchester et al., 2013). In comparison with LPV gain scheduling, the CCM-based approach has been described as an extended gain-scheduled control scheme that achieves global reference-independent stability and performance through an exact control realization integrating a series of local LPV controllers on a path between the current and reference states (Wang et al., 2019).

4. Computational synthesis and online implementation

The offline CCM synthesis problem is infinite-dimensional because the contraction inequality is pointwise in state and time. A standard workflow is to parameterize the dual metric $\lambda>0$0 and the auxiliary decision variables $\lambda>0$1 or $\lambda>0$2 in a finite basis, impose the pointwise LMI together with $\lambda>0$3, and then enforce feasibility by gridding or by sum-of-squares relaxation when the dynamics are polynomial (Manchester et al., 2013, Manchester et al., 2015). In the implementation steps summarized for the 2015 formulation, one parameterizes $\lambda>0$4 as a linear combination of chosen basis functions, solves the resulting finite-dimensional SDP, reconstructs $\lambda>0$5, computes at each time a minimal-energy path $\lambda>0$6 joining $\lambda>0$7 and $\lambda>0$8, and integrates the path-integrable differential controller

$\lambda>0$9

along $(x,u,t)$0 (Manchester et al., 2015).

The original paper illustrates the computational role of convexity with the unstable polynomial system

$(x,u,t)$1

After solving an LQR problem at the origin, one searches for polynomial $(x,u,t)$2 and $(x,u,t)$3 satisfying an SOS relaxation together with $(x,u,t)$4 and $(x,u,t)$5. Using YALMIP + Mosek, the resulting semidefinite program is solved in a few tenths of a second, and the resulting CCM controller coincides with LQR near the origin yet remains globally stabilizing (Manchester et al., 2015).

The main online burden is geodesic computation. A pseudospectral method based on Chebyshev polynomials parameterizes the geodesic, enforces endpoint constraints linearly, approximates the energy by Clenshaw-Curtis quadrature, and solves the resulting nonlinear program by quasi-Newton iterations (Leung et al., 2016). For real-time settings where repeatedly solving a geodesic optimization is undesirable, a continuous-time dynamic realization introduces an internal path state $(x,u,t)$6 with boundary conditions $(x,u,t)$7, $(x,u,t)$8, and evolves it by a forward-plus-gradient flow so that the internal path converges to a geodesic while the plant tracks the reference. In that construction, the online computation is shifted from repeated nonlinear optimization to ODE integration on the path state (Wang et al., 2019).

5. Structural extensions and generalized settings

Several extensions preserve the differential-geometric core of CCMs while altering the admissible geometry, information structure, or target set.

Extension	Main addition	Paper
Output-feedback	Dual observer and controller problems are solved by pointwise LMIs, and a separation principle holds	(Manchester et al., 2014)
Distributed nonlinear control	Separable block-diagonal metrics yield distributed controllers with neighbor-based communication and chordal SDP decompositions	(Shiromoto et al., 2018)
Finsler manifolds	CCMs are generalized to non-Riemannian metrics; sampled-data control need not compute globally shortest paths in real time	(Chaffey et al., 2018)
Lie groups	The group is embedded as a constraint manifold in Euclidean space and the CCM search is reformulated as convex conditions	(Wu et al., 2024)
Submanifold stabilization	A convex dual criterion guarantees exponential convergence to $(x,u,t)$9	(Manchester et al., 2015)

In the output-feedback formulation, the observer design problem is dual to the controller design problem: one searches for an observer-contraction metric $$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$0 satisfying a pointwise LMI obtained by the transpose substitution $$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$1, $$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$2. When both controller and observer metrics exist, the combined output-feedback law is exponentially stable, yielding a nonlinear separation principle (Manchester et al., 2014).

For large-scale networked systems, separable or sum-separable CCMs impose a block-diagonal metric

$$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$3

with matching sparsity in the dual variable $$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$4. If the pointwise LMI holds with this structure, the resulting feedback is $$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$5-admissible and exponentially stabilizes every target trajectory with rate $$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$6; when the relevant graph is chordal, the global LMI decomposes into clique-level LMIs (Shiromoto et al., 2018).

The Finsler generalization replaces the quadratic Riemannian differential energy with a Finsler-Lyapunov function $$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$7, allowing non-Riemannian metrics and asymmetric distances. In that setting, open-loop and sampled-data control constructions are available, and the sampled-data construction does not require real-time computation of globally shortest paths (Chaffey et al., 2018). On Lie groups, CCM design is achieved by viewing the group as a constrained embedded manifold, parameterizing the metric on the tangent bundle through a lower-dimensional dual variable $$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$8, and imposing convex LMI conditions on the induced tangent coordinates (Wu et al., 2024).

6. Amendment, robustness, and data-driven developments

A technically important correction was given in the 2017 amendment to the convex and intrinsic theory. The original proof of Theorem 1 required an additional condition to guarantee integrability of differential control signals. In the amended statement, one introduces

$$\delta_x^T\,MB=0 \quad\implies\quad \delta_x^T\!\bigl(\dot M + A^T M + M A + 2\lambda\,M\bigr)\,\delta_x \;<\;0.$$9

and requires the existence of a continuous scalar function $W(x,t)=M(x,t)^{-1}$0 such that

$W(x,t)=M(x,t)^{-1}$1

This condition is automatically satisfied if $W(x,t)=M(x,t)^{-1}$2 has constant rank, and also under the strong CCM conditions where $W(x,t)=M(x,t)^{-1}$3. A counterexample shows that if the input matrix drops rank, the weaker original conditions may not imply stabilizability of all trajectories (Manchester et al., 2017). A common misconception is therefore that the original weak kernel condition is always sufficient; the amendment shows that integrability is the subtle point.

Robust variants use CCMs to quantify disturbance amplification and to build invariant tubes. Robust Control Contraction Metrics (RCCMs) extend the differential inequalities to include $W(x,t)=M(x,t)^{-1}$4, seek a universal $W(x,t)=M(x,t)^{-1}$5-gain $W(x,t)=M(x,t)^{-1}$6, and compute invariant tubes around nominal trajectories that are valid for any nominal plan. Under mild assumptions, the RCCM approach yields tighter tubes than a CCM-plus-ISS construction (Zhao et al., 2021). In autonomous racing, a CCM computed for a perturbed single-track model is used to parameterize a homothetic tube for nonlinear MPC constraint tightening; the resulting robust nonlinear MPC introduces only a single additional state variable into the prediction model compared to the nominal scheme (Bongard et al., 4 Feb 2026).

Recent work has also replaced analytic parameterizations by learned ones. Neural approaches co-synthesize a dual metric and controller by embedding CCM residuals directly in training losses; examples include neural parameterizations of $W(x,t)=M(x,t)^{-1}$7 with post-training validation by grid-plus-Lipschitz certificates or conformal regression (Sun et al., 2020), offline-data methods that jointly learn an implicit dynamics model, a policy, and a contraction metric using augmented sample pairs (Rezazadeh et al., 2021), and actor-critic formulations in which a contraction metric generator is updated within an RL loop and the reward includes the metric-weighted tracking error (Cho et al., 28 May 2025). Distributional-RL variants estimate uncertainty in the CCM certificate itself and report at least a $W(x,t)=M(x,t)^{-1}$8 improvement in tracking error under large aerodynamic disturbances and at least a 3 times improvement in contraction rate relative to QuaDRED-MPC (Wang et al., 2022).

Application-specific implementations preserve the same core idea: certify incremental exponential stability through a metric and then incorporate additional structure. A robust neural control design for three-drone slung payload manipulation uses a CCM baseline controller with saturation together with an uncertainty and disturbance estimator, and states that the modularized design achieves a zero trajectory tracking error if the disturbances meet certain assumptions (Liang et al., 1 Oct 2025). For cooperative aerial payload transportation with variable-length cables, a neural CCM controller and a neural feedback controller are jointly trained for the payload subsystem, while cable-length control is handled in a decoupled channel for obstacle avoidance (Lo et al., 18 Jun 2026). Across these developments, the underlying object remains the same: a metric on differential dynamics that renders all trajectories exponentially convergent under a path-integrated feedback law.