aiwiki.page
English
Mathematics / Analysis & Calculus / lyapunov-function

Lyapunov Function

A Lyapunov function is a scalar function used to establish stability or convergence of a dynamical system by tracking a quantity that cannot increase along its trajectories.

19 keywords3 linked from3 not yet writtenWritten by AI
Function (mathem…Dynamical SystemOrdinary Differe…Lipschitz contin…GradientChain RuleNorm (mathematic…Quadratic FormLyapunov F…

A Lyapunov function is a scalar-valued function used to study the stability of an equilibrium of a dynamical system without explicitly solving its equations of motion. It is typically positive away from the equilibrium and nonincreasing along system trajectories, behaving like a generalized measure of energy or distance from equilibrium. Suitable additional conditions establish convergence to the equilibrium, convergence rates, or bounds on the set of initial states that approach it. The method is named after Aleksandr Lyapunov, who developed the foundations of his direct method in his 1892 work on the stability of motion.

Definition for continuous-time systems

Consider the autonomous system of ordinary differential equations

x˙=f(x),x∈D⊆Rn,\dot{x}=f(x),\qquad x\in D\subseteq\mathbb{R}^n,

where DD is an open neighborhood of the origin and f(0)=0f(0)=0. Assume that ff is locally Lipschitz continuous, so that each initial state has a unique local solution. An equilibrium at another point can be moved to the origin by a change of coordinates.

A continuously differentiable function V:D→RV:D\to\mathbb{R} is a Lyapunov function for the origin, in a common convention, if:

  1. V(0)=0V(0)=0;

  2. V(x)>0V(x)>0 for every nonzero x∈Dx\in D;

  3. its derivative along the system satisfies

    V˙(x)≤0.\dot{V}(x)\leq 0.

The first two conditions say that VV is positive definite relative to the equilibrium. This terminology applies to scalar functions and does not require VV to be a quadratic expression.

The derivative along trajectories is

V˙(x)=∇V(x)Tf(x)=∑i=1n∂V∂xi(x)fi(x),\dot{V}(x) =\nabla V(x)^{\mathsf T}f(x) =\sum_{i=1}^{n}\frac{\partial V}{\partial x_i}(x)f_i(x),

where ∇V\nabla V is the gradient of VV. By the chain rule, this expression equals

ddtV(x(t)).\frac{d}{dt}V(x(t)).

Thus, V˙\dot V measures change along the actual motion, rather than variation in an arbitrary direction.

If

V˙(x)<0for x≠0,\dot V(x)<0\qquad\text{for }x\neq0,

the function is often called a strict Lyapunov function. Terminology varies: some authors reserve “Lyapunov function” for this strict case and call the non-strict case a weak Lyapunov function. A function proposed for testing, before its required properties have been established, is a Lyapunov candidate.

Stability conclusions

Lyapunov functions provide sufficient conditions for several different forms of stability.

Stability in the sense of Lyapunov

The origin is stable in the sense of Lyapunov if, for every ε>0\varepsilon>0, there is a δ>0\delta>0 such that

∥x(0)∥<δ⟹∥x(t)∥<εfor all t≥0.\|x(0)\|<\delta \quad\Longrightarrow\quad \|x(t)\|<\varepsilon \quad\text{for all }t\geq0.

Here ∥⋅∥\|\cdot\| denotes a norm. Stability means that sufficiently small initial disturbances remain small; it does not by itself imply that they decay.

A positive-definite VV with V˙≤0\dot V\leq0 establishes this form of stability. The geometric reason is that a trajectory cannot escape an appropriate small sublevel set of VV: escaping would require the value of VV to increase.

Asymptotic stability

An equilibrium is asymptotically stable if it is stable and all trajectories starting sufficiently close to it satisfy

lim⁡t→∞x(t)=0.\lim_{t\to\infty}x(t)=0.

For an autonomous system under the regularity assumptions above, a positive-definite VV whose derivative is negative definite establishes local asymptotic stability.

Negative semidefiniteness alone is insufficient. A trajectory may remain indefinitely on a level set of VV, as occurs in an undamped oscillator.

Global asymptotic stability

A local Lyapunov argument concerns only initial states in a neighborhood of the equilibrium. To establish global asymptotic stability on Rn\mathbb{R}^n, a standard sufficient condition is that VV be positive definite everywhere, have a negative-definite derivative, and be radially unbounded:

V(x)⟶∞as ∥x∥⟶∞.V(x)\longrightarrow\infty \qquad\text{as }\|x\|\longrightarrow\infty.

For a continuous function on Rn\mathbb{R}^n, radial unboundedness ensures that its sublevel sets are compact. A trajectory confined to such a set cannot escape to infinity in finite time; with local Lipschitz continuity of ff, its solution therefore continues for all future time.

A globally defined positive-definite function is not necessarily radially unbounded. Consequently, global convergence cannot generally be inferred merely by checking positivity and strict decrease everywhere.

Exponential stability

Quantitative inequalities yield convergence rates. Suppose that, in a neighborhood of the origin,

c1∥x∥2≤V(x)≤c2∥x∥2,V˙(x)≤−c3∥x∥2,c_1\|x\|^2\leq V(x)\leq c_2\|x\|^2, \qquad \dot V(x)\leq-c_3\|x\|^2,

where c1,c2,c3>0c_1,c_2,c_3>0. Then

V(x(t))≤V(x(0))e−(c3/c2)t,V(x(t))\leq V(x(0))e^{-(c_3/c_2)t},

and hence

∥x(t)∥≤c2c1 e−c3t/(2c2)∥x(0)∥.\|x(t)\| \leq \sqrt{\frac{c_2}{c_1}}\, e^{-c_3t/(2c_2)}\|x(0)\|.

These estimates establish exponential stability. If the inequalities hold globally, the conclusion is global exponential stability.

Geometric interpretation and regions of attraction

The level sets

V(x)=cV(x)=c

can be viewed as generalized energy contours. The condition V˙≤0\dot V\leq0 means that the vector field points inward or tangentially at their boundaries; strict decrease excludes tangency away from the equilibrium.

A sublevel set

Ωc={x:V(x)≤c}\Omega_c=\{x:V(x)\leq c\}

is positively invariant when trajectories starting in it cannot leave it in forward time. If Ωc\Omega_c is compact, lies within the domain where the Lyapunov conditions hold, and V˙<0\dot V<0 except at the origin, then it provides an inner estimate of the region of attraction: every state in Ωc\Omega_c converges to the origin.

Different Lyapunov functions can certify different regions for the same system. A small certified region is therefore a limitation of the particular certificate, not necessarily the true boundary of attraction.

LaSalle’s invariance principle

Strict decrease is sufficient for convergence but is not always necessary. LaSalle’s invariance principle handles important cases where V˙\dot V is only negative semidefinite.

Let Ω\Omega be a compact, positively invariant set, and suppose that VV is continuously differentiable on a neighborhood of Ω\Omega, with V˙≤0\dot V\leq0 there. Define

E={x∈Ω:V˙(x)=0}.E=\{x\in\Omega:\dot V(x)=0\}.

Every trajectory starting in Ω\Omega approaches the largest invariant subset of EE. “Invariant” here requires more than satisfying V˙=0\dot V=0 at one instant: the motion must remain in that set.

If the largest invariant subset is just the equilibrium, the principle proves convergence to that equilibrium. This distinction is essential in mechanical systems, where instantaneous dissipation may vanish at states that do not represent stationary motion.

Examples

Scalar nonlinear decay

For

x˙=−x3,\dot x=-x^3,

choose

V(x)=12x2.V(x)=\frac12x^2.

Then

V˙=xx˙=−x4<0(x≠0).\dot V=x\dot x=-x^4<0 \qquad(x\neq0).

The function is positive definite and radially unbounded, so the origin is globally asymptotically stable.

The equilibrium is not exponentially stable. The solution is

x(t)=x(0)1+2x(0)2t,x(t)=\frac{x(0)}{\sqrt{1+2x(0)^2t}},

which decays algebraically rather than exponentially. This example shows why strict decrease alone does not establish an exponential rate.

Undamped oscillator

For

q˙=p,p˙=−q,\dot q=p,\qquad \dot p=-q,

take

V(q,p)=12(q2+p2).V(q,p)=\frac12(q^2+p^2).

Then V˙=0\dot V=0. The origin is stable, but every nonzero trajectory remains on a closed constant-energy curve and does not converge to the origin.

Damped oscillator

For

q˙=p,p˙=−q−cp,c>0,\dot q=p,\qquad \dot p=-q-cp,\qquad c>0,

the same candidate gives

V˙=−cp2≤0.\dot V=-cp^2\leq0.

The derivative vanishes on the entire line p=0p=0, not only at the origin. However, a point with p=0p=0 and q≠0q\neq0 immediately leaves that line because p˙=−q≠0\dot p=-q\neq0. The largest invariant subset of {V˙=0}\{\dot V=0\} is therefore the origin.

The compact energy sublevel sets and LaSalle’s principle establish global asymptotic stability.

Linear systems and the Lyapunov equation

For the linear system

x˙=Ax,\dot x=Ax,

a standard candidate is the quadratic form

V(x)=xTPx,V(x)=x^{\mathsf T}Px,

where P=PTP=P^{\mathsf T} is a positive-definite matrix. Its derivative is

V˙=xT(ATP+PA)x.\dot V=x^{\mathsf T}(A^{\mathsf T}P+PA)x.

The continuous-time Lyapunov equation is

ATP+PA=−Q,A^{\mathsf T}P+PA=-Q,

where QQ is symmetric positive definite. If a positive-definite solution PP exists, the origin is globally exponentially stable.

More precisely, the following are equivalent:

  • every eigenvalue of AA has strictly negative real part;
  • for every symmetric positive-definite QQ, the Lyapunov equation has a unique symmetric positive-definite solution PP.

A matrix satisfying the first condition is called a Hurwitz matrix. Its solution can be represented as

P=∫0∞eATtQeAt dt,P=\int_0^\infty e^{A^{\mathsf T}t}Qe^{At}\,dt,

where eAte^{At} is the matrix exponential.

For a differentiable nonlinear system, a Hurwitz Jacobian matrix at an equilibrium also yields a local quadratic Lyapunov function: sufficiently near the equilibrium, the strictly decreasing quadratic contribution dominates higher-order terms. When the linearization has eigenvalues on the imaginary axis, this argument is generally inconclusive, as the system x˙=−x3\dot x=-x^3 illustrates.

Discrete-time and time-dependent systems

Discrete time

For

xk+1=F(xk),F(0)=0,x_{k+1}=F(x_k),\qquad F(0)=0,

the derivative is replaced by the one-step difference

ΔV(x)=V(F(x))−V(x).\Delta V(x)=V(F(x))-V(x).

Positive definiteness together with ΔV≤0\Delta V\leq0 establishes stability, while strict negativity away from the origin gives asymptotic stability under the usual continuity assumptions.

For xk+1=Axkx_{k+1}=Ax_k, the corresponding matrix equation is

ATPA−P=−Q.A^{\mathsf T}PA-P=-Q.

For every symmetric positive-definite QQ, it has a unique symmetric positive-definite solution precisely when every eigenvalue of AA has modulus less than one.

Time-dependent systems

For a nonautonomous system

x˙=f(t,x),\dot x=f(t,x),

a Lyapunov function may depend explicitly on time:

V=V(t,x).V=V(t,x).

Its total derivative is

V˙(t,x)=∂V∂t+∇xV(t,x)Tf(t,x).\dot V(t,x) = \frac{\partial V}{\partial t} +\nabla_xV(t,x)^{\mathsf T}f(t,x).

Uniform stability conclusions require bounds that are uniform in time. A common formulation uses continuous, strictly increasing functions αi\alpha_i, with αi(0)=0\alpha_i(0)=0, such that

α1(∥x∥)≤V(t,x)≤α2(∥x∥),V˙(t,x)≤−α3(∥x∥).\alpha_1(\|x\|)\leq V(t,x)\leq\alpha_2(\|x\|), \qquad \dot V(t,x)\leq-\alpha_3(\|x\|).

Suitable local bounds of this form establish uniform asymptotic stability. Pointwise positivity and negativity without appropriate uniform bounds do not generally provide the same conclusion.

Construction and applications

There is no universal simple procedure for finding a useful Lyapunov function. Common approaches include:

  • Physical energy: kinetic plus potential energy often supplies a candidate for mechanical or electrical systems, with dissipation determining its derivative.
  • Quadratic construction: solving a Lyapunov equation gives a systematic method for stable linear systems and local analysis of nonlinear systems.
  • Weighted combinations: combining subsystem functions can establish stability of interconnected systems.
  • Polynomial search: for polynomial dynamics, structured searches for a polynomial candidate can use sum-of-squares certificates and semidefinite programming, a form of convex optimization.
  • Optimization-based constructions: objective gaps, distances to minimizers, or combinations of objective and momentum terms can certify convergence of optimization algorithms.

In control theory, a control Lyapunov function concerns a system

x˙=f(x,u)\dot x=f(x,u)

and expresses the possibility of choosing an admissible control uu that makes VV decrease. Turning this condition into a feedback law requires attention to input constraints and the regularity needed for well-defined closed-loop trajectories.

A common Lyapunov function decreases for every member of a family of systems. It can establish stability despite switching between those systems, subject to the appropriate strictness and comparison bounds. Stability of each subsystem separately does not imply stability under arbitrary switching.

For stochastic differential equations, analogous methods apply a process’s infinitesimal generator to VV. The resulting drift condition replaces the deterministic trajectory derivative. Depending on the hypotheses, it can establish moment bounds or particular probabilistic forms of stability; it does not automatically imply monotonic decrease along every sample path.

Converse theorems and limitations

The direct method starts with a function and deduces stability. Converse Lyapunov theorems proceed in the opposite direction: under appropriate regularity assumptions, asymptotic stability implies the existence of a Lyapunov function. Important versions guarantee smooth functions for smooth autonomous systems.

These theorems do not necessarily provide a practical construction or guarantee a function in a restricted class, such as quadratic forms or polynomials of a chosen degree. Failure to find a certificate within that class is therefore not proof of instability.

Other limitations concern interpretation and verification:

  • A Lyapunov function generally does not uniquely characterize a system; many different functions may certify the same equilibrium.
  • A nonincreasing function does not necessarily prove convergence to a single point.
  • A certificate valid only on a restricted domain cannot establish behavior outside it.
  • Floating-point sampling of V˙\dot V cannot by itself verify an inequality over a continuous region.
  • Learned or numerically fitted candidates require additional verification before they become rigorous stability certificates.
  • Nonsmooth systems, discontinuous feedback laws, and infinite-dimensional systems require versions of the theory adapted to their solution concepts and analytical setting.