Lyapunov Function

A Lyapunov function, denoted as V(x), must be positive definite around the equilibrium point (typically the origin). This means:

• V(0) = 0 (zero at the equilibrium).
• V(x) > 0 for all x ≠ 0 in a region around the origin.

This property ensures the function acts like an abstract measure of "energy" or "distance" from the equilibrium. Common examples include quadratic forms like V(x) = x₁² + x₂², which is clearly zero only at the origin and positive elsewhere.

The most critical property is that the time derivative of V(x) along the system's trajectories must be negative definite (or negative semi-definite). For a system ẋ = f(x), this derivative is computed as Ṽ(x) = ∇V(x) · f(x).

• Ṽ(0) = 0 at the equilibrium.
• Ṽ(x) < 0 for all x ≠ 0 in a region.

This condition guarantees that V(x) strictly decreases over time as the system evolves, mathematically proving that the state x(t) is being driven toward the equilibrium.

To prove global asymptotic stability (stability from any initial condition), a Lyapunov function must be radially unbounded. This means:

• V(x) → ∞ as ||x|| → ∞.

This property ensures that the level sets of V(x) (contours of constant V) are closed and bounded. It prevents trajectories from "escaping to infinity" while decreasing V, guaranteeing they will eventually converge to the origin regardless of how far they start. A function like V(x) = x₁² + x₂² is radially unbounded.

When the derivative Ṽ(x) is only negative semi-definite (Ṽ(x) ≤ 0), asymptotic stability can still be proven using LaSalle's Invariance Principle. This advanced property states that the system state will converge to the largest invariant set contained within the set where Ṽ(x) = 0.

• This is crucial for systems with conserved quantities (like mechanical energy in frictionless systems).
• It allows engineers to prove stability even when the Lyapunov function's derivative is not strictly negative everywhere.

Finding a valid Lyapunov function is more art than science, but common construction methods include:

• Quadratic Forms (V = xᵀPx): For linear systems ẋ = Ax, solving the Lyapunov Equation AᵀP + PA = -Q for a positive definite P is systematic.
• Sum-of-Squares (SOS) Programming: For polynomial systems, this provides an algorithmic search method.
• Energy-Based Functions: For mechanical/electrical systems, total energy (kinetic + potential) is a natural candidate.

The core challenge is that no universal method exists for nonlinear systems; failure to find a function does not prove instability.

Lyapunov functions are not just for analysis but are central to control synthesis. Techniques like:

• Control Lyapunov Function (CLF): A function where a controller u(x) can be explicitly designed to ensure Ṽ(x) < 0.
• Backstepping: A recursive method that constructs both the controller and the Lyapunov function step-by-step for cascaded systems.
• Lyapunov Redesign: A method for robust control that modifies a nominal controller to handle bounded uncertainties or disturbances.

This makes the Lyapunov function a blueprint for building stabilizing controllers.

A Lyapunov function provides a certificate of stability for a designed controller. For a robot balancing (like an inverted pendulum) or tracking a planned trajectory, engineers construct a candidate Lyapunov function V(x). By proving V(x) > 0 and its derivative dV/dt < 0 along the system's trajectories, they mathematically guarantee the robot will converge to the desired state and remain there despite small disturbances. This replaces heuristic tuning with verifiable guarantees.

Not all stable systems recover from large pushes. The region of attraction (ROA) is the set of all initial states from which the system will converge to the equilibrium. By finding a level set of the Lyapunov function where dV/dt remains negative, engineers can computationally estimate and visualize this safe basin. This tells operators how far a legged robot can be leaned or a drone blown off course before its controller can no longer recover it, defining operational limits.

A Control Lyapunov Function (CLF) is a more powerful tool used for controller synthesis, not just analysis. The goal is to find a control law u(x) that makes dV/dt negative. This framework is used in:

• Quadrotor flight controllers to stabilize aggressive maneuvers.
• Feedback linearization techniques for robotic arms.
• Real-time optimization in methods like Model Predictive Control (MPC), where ensuring dV/dt < 0 can be added as a stability constraint.

In modern safe robotics, Lyapunov functions (for stability) are combined with Control Barrier Functions (CBFs) (for safety). This creates a quadratic program (QP)-based controller that:

1. Seeks stability by minimizing deviation from a CLF condition.
2. Enforces safety by ensuring the system state remains within a defined safe set (e.g., away from obstacles, within joint limits). This allows a mobile robot to stabilize to a goal while dynamically avoiding collisions, with both properties formally verified.

When reinforcement learning (RL) or neural networks are used for robot control, stability is often lost. Lyapunov functions are used to:

• Analyze and certify the stability of a learned policy post-training.
• Guide the training process itself by incorporating a Lyapunov-derived stability loss into the RL objective.
• Provide stability guarantees for adaptive controllers that adjust parameters online, ensuring the robot remains stable even as its model is updated.

Robots often operate as hybrid systems with discrete mode switches (e.g., a walking robot's feet making/breaking contact). Proving stability requires a multiple Lyapunov function approach. Engineers design a function for each mode (e.g., left-foot stance, right-foot stance) and show that the function's value decreases at the switching instants. This proves the entire walking gait is stable, not just individual phases.

A system's equilibrium point is asymptotically stable if trajectories starting nearby not only stay nearby but also converge to the equilibrium point over time. A Lyapunov function with a negative definite derivative provides a formal proof of this property.

• Local vs. Global: Asymptotic stability can be local (for initial conditions in a region) or global (for all initial conditions).
• Exponential Stability: A stronger form where convergence to the equilibrium occurs at an exponential rate.

A Control Lyapunov Function (CLF) is a Lyapunov function where the negative definiteness of its derivative is not inherent to the system but is achieved through the design of a feedback control law. It is a central tool in nonlinear control design.

• Core Problem: Find a control input u such that the derivative V̇(x, u) < 0 for all x ≠ 0.
• Application: Enables the synthesis of stabilizing controllers for complex systems, such as robotic walkers or aircraft, by solving the CLF condition.

A Control Barrier Function (CBF) is a mathematical tool complementary to the CLF, used to enforce safety constraints (e.g., collision avoidance, joint limits). It defines a "safe set" and synthesizes controllers that keep the system within it.

• Safety vs. Stability: A CLF drives the system to a goal (stability), while a CBF keeps it away from danger (safety).
• Quadratic Programs (QPs): CLFs and CBFs are often combined and solved together in real-time via a CLF-CBF Quadratic Program, balancing performance and safety.

LaSalle's Invariance Principle extends Lyapunov's direct method for systems where the Lyapunov derivative is only negative semi-definite (V̇(x) ≤ 0). It proves asymptotic stability by analyzing the largest invariant set within the region where V̇(x) = 0.

• Key Insight: Even if energy (V) doesn't strictly decrease everywhere, the system may still converge to the equilibrium if it cannot stay in places where energy is constant.
• Common Use: Essential for analyzing systems with conservative forces or for proving stability of adaptive control laws.

For linear time-invariant (LTI) systems of the form ẋ = Ax, stability analysis simplifies. The Lyapunov equation AᵀP + PA = -Q provides a direct method to find a quadratic Lyapunov function V(x) = xᵀPx.

• Procedure: Choose any positive definite matrix Q (e.g., the identity I). Solve the linear matrix equation for P. If P is positive definite, the system is globally asymptotically stable.
• Connection to LQR: The cost-to-go matrix from the Linear Quadratic Regulator (LQR) solution satisfies the Lyapunov equation, directly linking optimal control to stability.

Sum-of-Squares (SOS) Programming is a computational method, based on convex optimization, to algorithmically search for polynomial Lyapunov functions for nonlinear polynomial systems. It transforms the stability proof into a feasibility problem.

• Core Idea: Restrict the search for V(x) to polynomials that can be expressed as a sum of squares of other polynomials, a condition that can be checked via semidefinite programming.
• Application: Enables automated stability verification and controller synthesis for complex robotic systems with polynomial dynamics, where analytical Lyapunov functions are intractable to find by hand.