Control Lyapunov Functional Overview

• Control Lyapunov functionals are an extension of classical Lyapunov functions that embed control parameters to guarantee system stability across various dynamic models.
• They enable the synthesis of stabilizing feedback laws, such as Sontag’s universal formula, which ensures asymptotic stability and inverse optimality for nonlinear and uncertain systems.
• Modern computational techniques, including convex relaxation, SOS programming, and neural network-based methods, facilitate efficient CLF construction for high-dimensional and safety-critical control applications.

A control Lyapunov functional (CLF, often called a control Lyapunov function in finite dimensions) is a mathematical construct central to modern nonlinear feedback stabilization, robust control design, performance analysis, and formal specification of optimality in control systems. The CLF generalizes classical Lyapunov functions by embedding the control-parameterization explicitly within the Lyapunov decrease condition, thus providing a unified criterion for the existence of stabilizing feedbacks across a wide class of deterministic, stochastic, and infinite-dimensional systems.

1. Formal Definition and General Structure

For a control system described by

$$\dot x = f(x) + G(x)u, \quad x\in\mathbb{R}^n,~u\in\mathbb{R}^m,$$

a function $V:\mathbb{R}^n\to\mathbb{R}$ is a control Lyapunov function if:

• $V\in C^1$, $V(0)=0$, and $V(x)>0$ for all $x\neq0$,
• $V$ is radially unbounded: $V(x)\to\infty$ as $|x|\to\infty$,
• For all $x\neq0$,

$$\inf_u \Big( \nabla V(x)\cdot f(x) + \nabla V(x)\cdot G(x)u \Big) < 0.$$

This property implies that at each state (excluding the equilibrium), one can find a control input rendering $V$ strictly decreasing. The condition can be localized (“local CLF”) or global, depending on the domain over which the inequality holds (Bongard et al., 4 Feb 2026). In broader contexts (e.g. systems with disturbances, constraints, or infinite dimensions), the CLF is extended to a control Lyapunov functional, which acts on function spaces or includes generalized derivatives.

2. Sontag-Type Feedback and Inverse Optimality

Given a CLF, Sontag's universal formula constructs a continuous, often smooth, state-feedback law ensuring asymptotic stabilization: $u_S(x)= -R^{-1} b(x)^\top \lambda(x),$ where

$$a(x) = \nabla V(x)\cdot f(x),~b(x) = \nabla V(x)\cdot G(x),$$

and

$$\lambda(x) = \frac{a(x) + \sqrt{a(x)^2 + x^\top Q x \, b(x) R^{-1} b(x)^\top}}{b(x) R^{-1} b(x)^\top},$$

with positive-definite weighting matrices $Q,R$. This controller not only guarantees strict decrease of $V$ but is inverse-optimal: it minimizes a trajectory cost functional with weighting tied to the CLF structure,

$$J(u) = \frac{1}{2} \int_{t_0}^\infty \frac{1}{\lambda(x)} \Big(x^\top Qx + u^\top R u \Big)dt,$$

which recovers the standard quadratic cost in linear regimes or for appropriate choice of $V$ (Bongard et al., 4 Feb 2026).

3. Global and Local CLF Construction in Nonlinear Systems

For linearizable or feedback-linearizable systems (admitting a global diffeomorphism $z=T(x)$), a quadratic form

$$V(x) = \frac{1}{2} z(x)^\top \tilde P z(x),~\tilde P = (T'(0))^{-\top} P (T'(0))^{-1},$$

with $P$ from the algebraic Riccati equation,

$A^\top P + P A - P B R^{-1} B^\top P + Q = 0,$

yields a global CLF and ensures global asymptotic stability under Sontag’s feedback. Locally, the LQR value function in the original coordinates can serve as a CLF, ensuring that the region of attraction under Sontag's controller is at least as large as that of the linear design (Bongard et al., 4 Feb 2026). In both cases, the CLF construction is constructive and computationally tractable.

4. Extensions: Robustness, Infinite-Dimensionality, and Higher-Order Constructions

CLF methodologies extend to uncertain and infinite-dimensional settings. For linear systems with bounded disturbances,

$\dot x = A x + Bu + B_w w, ~ \|w\|\leq\delta,$

a quadratic CLF yields robust stabilization if the CLF’s derivative admits decrease irrespective of the disturbance, enforced via a robust control Lyapunov inequality and solved via semi-definite programming: $$\inf_{u}\sup_{\|w\|\le\delta}\nabla V(x)^\top [A x + B u + B_w w ] \leq -\alpha(V(x)),$$ with the region of attraction characterized as an ellipsoid whose size depends explicitly on the disturbance bound (Yang et al., 2024). For distributed parameter, time-delay, or hybrid systems, CLF functionals act on trajectory segments or function spaces, with their decrease verified via operator inequalities in Hilbert space, further generalizing finite-dimensional CLF theory.

In nonholonomic or underactuated systems lacking degree-1 (classical) CLFs, higher-degree control Lyapunov functionals are constructed by embedding dissipativity via iterated Lie-bracket Hamiltonians: $$H^{(k)}(x,p) = \inf_{v\in F^{(k)}(x)} \langle p, v \rangle, \quad F^{(k)} = \{\text{vector fields and their Lie brackets up to order } k\}.$$ Degree-$k$ CLFs guarantee global asymptotic controllability even if smooth degree-1 CLFs do not exist (Motta et al., 2016).

5. Computational and Data-Driven Methods for CLF Synthesis

Numerous algorithms operationalize the search for CLFs and their associated stabilizing policies:

• Convex Relaxation and SOS: For polynomial systems, CLF search reduces to sum-of-squares (SOS) or semidefinite programming problems checking positivity and decrease conditions over the state space (Ravanbakhsh et al., 2018).
• Counterexample-Guided Learning: Iterative frameworks alternate between candidate CLF generation, verification on samples or via SMT solvers, and counterexample refinement. This approach encompasses both classical basis-function parameterizations and high-capacity neural network representations (Chang et al., 2020, Wu et al., 2023, Ravanbakhsh et al., 2018).
• Physics-Informed Learning: Neural network CLFs are trained to minimize the residual of the Hamilton-Jacobi-Bellman (HJB) or Zubov-type PDE, augmented with trajectory data generated by Pontryagin's Maximum Principle, and subsequently formally verified via SMT-based logic (Liu et al., 2024).
• Robust and Safe Synthesis: For safety-critical or robust settings, CLFs are patched with control barrier functions (CBFs) using algorithmic smoothing (e.g., softmax barriers, bump functions) and strict Farkas’ lemma to ensure consistency and maximize invariant domains, with formal certificates produced via SMT (Liu et al., 2 Oct 2025).
• Exit-Time and Curse-of-Dimensionality-Free Methods: For high-dimensional nonlinear systems, concatenation of local CLFs and value functions generated from exit-time optimal control problems allows global CLF synthesis with tractable computational burden (Yegorov et al., 2019).

6. Region of Attraction, Safety, and Performance Considerations

The explicit construction and verification of CLF sublevel sets enable formal certification of regions of attraction (RoA). In modern approaches, the goal is to maximize the certifiable RoA, as quantified by the largest invariant sublevel set where the CLF decrease condition holds and where formal verification can be established (via MILP, SOS, or SMT methods) (Wu et al., 2023, Liu et al., 2 Oct 2025, Liu et al., 2024). Combining CLFs with control barrier functions yields certifiably safe sets guaranteeing both asymptotic stabilization and forward-invariance of safety constraints. The interplay between local quadratic (LQR-based) and global CLF structures quantifies performance tradeoffs: near the setpoint, optimality in the classical quadratic sense is recovered; away from the setpoint, inverse-optimality or distorted cost functionals tied to the CLF structure may dominate (Bongard et al., 4 Feb 2026).

A sample table compares CLF constructions and their domains for the reverse Van der Pol system (Liu et al., 2024):

Method	Verified Region {level ≤c}	Verification Time
QCLF (ARE)	c_P ≈ 0.45	<0.1 s
Rational CLF	≈ 0.45	–
SOS degree-6	≈ 0.32	–
Neural CLF	c_2 ≈ 0.83	≲ 5 s (dReal)

Both neural and classical CLFs are now routinely constructed with formal, machine-verifiable stability and safety certificates, surpassing standard polynomial or rational CLFs in both the volume of the verified safe RoA and computation times for typical classical nonlinear examples.

7. Quantum, Stochastic, and Application-Specific CLF Functionals

CLF functionals have been extended to quantum systems, where the state is represented by a density matrix and the CLF is a trace functional of the form $V(\rho) = \text{tr}(P \rho)$. Rapid state preparation can be achieved by synthesizing CLF-based controls employing switching or approximate bang-bang strategies, balancing speed with robustness and guaranteed convergence (Kuang et al., 2015). In the presence of stochastic perturbations, CLF methods accommodate nonsmooth Lyapunov functionals; for sample-and-hold or discrete-time controllers, regularization via Moreau–Yosida convolution yields differentiable surrogates suitable for stability analysis under bounded noise (Osinenko et al., 2022).

• "Control Lyapunov Functions for Optimality in Sontag-Type Control" (Bongard et al., 4 Feb 2026)
• "Robust Control using Control Lyapunov Function and Hamilton-Jacobi Reachability" (Yang et al., 2024)
• "Computing Control Lyapunov-Barrier Functions: Softmax Relaxation and Smooth Patching with Formal Guarantees" (Liu et al., 2 Oct 2025)
• "Synthesis of control Lyapunov functions and stabilizing feedback strategies using exit-time optimal control" (Yegorov et al., 2019)
• "Formally Verified Physics-Informed Neural Control Lyapunov Functions" (Liu et al., 2024)
• "Neural Lyapunov Control" (Chang et al., 2020)
• "Learning Control Lyapunov Functions from Counterexamples and Demonstrations" (Ravanbakhsh et al., 2018)
• "Rapid Lyapunov control of finite-dimensional quantum systems" (Kuang et al., 2015)
• "Lyapunov-like functions involving Lie brackets" (Motta et al., 2016)
• "On stochastic stabilization via non-smooth control Lyapunov functions" (Osinenko et al., 2022)