High-Order Control Lyapunov Functions

• High-Order Control Lyapunov Function (HOCLF) is a method for stabilizing systems with higher relative degree using chains of Lie derivatives and class–𝒦 functions.
• It integrates quadratic programming to enforce stability and safety constraints in nonlinear, underactuated, or constrained systems.
• HOCLFs are applied in soft robotics, autonomous driving, and nonholonomic systems to achieve real-time, provable control under complex conditions.

A High-Order Control Lyapunov Function (HOCLF) is a foundational concept in nonlinear control for systems with relative degree greater than one. By extending classical Control Lyapunov Function (CLF) methods, HOCLFs enable systematic controller synthesis for nonlinear, underactuated, or constrained systems, especially those in which the control input appears only in higher derivatives of the Lyapunov function. HOCLFs are defined via chains of Lie derivatives and class–𝒦 functions, and can be embedded directly into real-time optimization (commonly quadratic programming) to yield provably stabilizing feedback even in the presence of additional safety constraints, uncertainty, or lack of classical smooth CLFs.

1. Formal Definition and Theoretical Basis

Given a control-affine system: $$\dot x = f(x) + g(x)\,u, \qquad x\in\mathbb{R}^n,\,u\in\mathbb{R}^m,$$ where $f,\,g$ are locally Lipschitz, a function $V:\R^n\to\R_{\ge0}$ of class $C^r$ is said to be a High-Order Control Lyapunov Function of relative degree $r$ at the origin if it is positive definite and there exist class–𝒦 functions $\alpha_1, \ldots, \alpha_r$, such that defining recursively: $$\psi_0(x) := V(x), \ \psi_i(x) := L_f\psi_{i-1}(x) + \alpha_i(\psi_{i-1}(x)), \;\; i=1,\ldots,r-1,$$ the following holds for all $x \neq 0$: $$\inf_{u\in\R^m} \left\{ L_f^rV(x) + L_gL_f^{r-1}V(x)u + \beta_r(\psi_{r-1}(x)) \right\} \leq 0,$$ where $\beta_r$ is a class–$\mathcal K_\infty$ function (Chen et al., 14 Dec 2025, Wong et al., 5 May 2025). The decoupling structure of HOCLF chains mirrors higher-order backstepping but with a single encapsulating inequality controlling stability.

2. High-Order Dissipative Inequalities and Lie Bracket Structure

HOCLF methodology generalizes to drift-free or weakly regular systems via the concept of a degree–$k$ CLF, employing Hamiltonians derived from iterated Lie brackets. For a system of the form

$\dot y(t) = \sum_{i=1}^m a_i(t) f_i(y(t)),$

the degree–$k$ Hamiltonian at $(x,p)$ is

$$H^{(k)}(x,p) := \inf_{v\in F^{(k)}(x)} \langle p, v\rangle,$$

where $F^{(k)}(x)$ contains all formal Lie brackets up to degree $k$. A function $U(\cdot)$ is a degree–$k$ CLF if, for all $x$ off the target set $\mathcal{T}$ and all $p$ in the set of limiting gradients,

$H^{(k)}(x,p) < 0.$

This framework ensures global asymptotic controllability in the absence of smooth or even degree–1 CLFs and under weak regularity assumptions (including set-valued brackets for locally Lipschitz vector fields) (Motta et al., 2016).

3. Control Synthesis via Real-Time Optimization

HOCLFs enable controller synthesis by enforcing the high-order decrease condition as an affine constraint in a quadratic program (QP). The standard control routine proceeds by, at each sampling instant, solving: $$\begin{aligned} \min_{u,\,\delta} \quad & \|u\|^2 + p\,\delta^2 \ \text{s.t.}\quad & L_f^rV(x) + L_gL_f^{r-1}V(x)u + O(V(x)) + \beta_r(\phi_{r-1}(x)) \leq \delta, \ & \delta \geq 0, \ & \text{(and possibly, HOCBF constraints on safety)} \end{aligned}$$ where slack $\delta$ preserves feasibility. When safety constraints are present, as in HOCBFs, both stability (HOCLF) and safety (HOCBF) inequalities are incorporated, and the QP delivers the unique minimizer at each time (Wong et al., 5 May 2025, Chen et al., 14 Dec 2025).

Taylor–Lagrange Control (TLC) offers an alternative by directly expanding $V(x)$ using Taylor's theorem, resulting in a single necessary and sufficient decrease inequality, compared to cascaded (recursive) approaches that introduce extra conservatism (Xiao et al., 12 Dec 2025). The QP is then formulated on this single constraint, obviating the need for multiple auxiliary inequalities.

4. Handling Stochastic Systems and High Relative Degree

Stochastic extensions of HOCLF, termed High-Order Stochastic Control Lyapunov Functions (HO-SCLF), generalize the deterministic condition by incorporating the Itô generator: $d x_t = (f(x_t) + g(x_t)u_t)dt + \Sigma(x_t) dW_t.$ The generator $\mathcal{L}_f$ becomes: $$\mathcal{L}_f V(x) = \partial_x V(x)^\top f(x),\quad \mathcal{L}_g V(x) = \partial_x V(x)^\top g(x),\quad \tfrac{1}{2}\mathrm{tr}(\partial_x^2 V(x)\,\Sigma(x)\Sigma(x)^\top).$$ The hierarchical structure of the associated "layers" of functions and triggers, followed by implementation in a QP, confers global asymptotic stabilizability in probability when the top-layer inequality is enforced (Sarkar et al., 2020).

5. Illustrative Applications and Case Studies

Soft Robots under Contact-Aware Safety Constraints

In the context of soft robotics, HOCLFs enable trajectory tracking, tip regulation, force regulation, and compliant object manipulation. For a Piecewise Cosserat-Segment (PCS) soft manipulator discretized into $N$ segments, the HOCLF constraint for configuration-space regulation (relative degree 2) is: $V_{\rm csr}(x) = \|q - q^d\|^2,\,\, \text{HOCLF:}$

$$2\dot q^\top \dot q + 2(q - q^d)^\top \ddot q + \alpha_1(2(q - q^d)^\top \dot q) + \alpha_2\,\phi_1 + 2(q - q^d)^\top M^{-1}A\,u \leq \delta.$$

This structure, combined with HOCBF constraints on whole-body contact forces, yields provable simultaneous safety and performance (Wong et al., 5 May 2025).

Autonomous Driving and Collision Avoidance

HOCLF constraints are embedded in QPs, along with HOCBFs, to control vehicles in dynamic and uncertain environments, such as ensuring safe distances from bicyclists and executing robust, collision-free maneuvers in response to emergent threats. Empirical studies show that, for multiple critical crash scenarios (as recorded in the FARS dataset), the HOCLF-HOCBF QP architecture maintains strict trajectory tracking and swift return to reference while enforcing minimum separation from obstacles, as summarized in the performance table:

Scenario	Max lateral error (m)	Min separation (m)	Time to return (s)
Pure path-tracking	0.05	–	–
Static obstacle avoid.	0.08	–	0.6
Dynamic avoid.	0.10	2.1	0.8
FARS210	0.12	2.3	1.0
FARS220	0.15	2.0	1.2
FARS310	0.20	2.1	1.5

(Chen et al., 14 Dec 2025)

Systems without Degree–1 CLFs

For drift-free systems with nonholonomic constraints, classical degree–1 CLF approaches may fail. The degree–2 HOCLF and associated Lie bracket Hamiltonian guarantee global asymptotic controllability in such scenarios. For example, in the nonholonomic integrator, no smooth degree–1 CLF exists, but a degree–2 CLF can be constructed and used for stabilization (Motta et al., 2016).

6. Practical Design and Implementation Considerations

• Selection of candidate functions: Quadratic distance functions or physical energy measures are typical for $V$, while class–𝒦 functions $\alpha_i$ are often chosen linear.
• QP feasibility: Slack variables allow the QP to maintain feasibility even if the strict HOCLF inequality is temporarily violated; penalization parameters $p$ trade off convergence speed and feasibility.
• Relative degree and constraint formulation: Systems with higher relative degree (e.g., elastic actuators, mechanical linkages) require careful computation of auxiliary functions and their Lie derivatives.
• Cascaded vs. single-constraint: The Taylor–Lagrange formalism yields a single necessary and sufficient decrease constraint per sampling step, in contrast to cascade approaches (multiple inequalities), reducing conservatism (Xiao et al., 12 Dec 2025).
• Stochasticity: For stochastic systems, the trace term reflecting diffusion must be incorporated into the QP constraints; relaxation and penalty terms are tuned to maintain stability with minimal constraint violation (Sarkar et al., 2020).

7. Connections, Generalizations, and Theoretical Guarantees

The existence of a HOCLF for a given system ensures global asymptotic stability under the corresponding feedback law—either exactly (if strict enforcement is feasible) or in probability (stochastic case). Practical guarantees include monotonic decay of the Lyapunov function along closed-loop trajectories, forward invariance of safe sets when coupled with HOCBFs, and tunable regions of attraction based on sublevel sets of the HOCLF (Wong et al., 5 May 2025, Chen et al., 14 Dec 2025).

HOCLFs generalize classical CLFs, extend naturally to stochastic and nonsmooth systems, and subsume special classes such as Taylor–Lagrange and Lie-bracket-based methodologies. Their integration into QP-based control architectures enables real-time, certifiable stabilization and constraint enforcement in complex, high-dimensional, and safety-critical robotics, automotive, and mechanical systems.