Lyapunov Stability Theory

⚡ Key Takeaways

Lyapunov Stability Theory studies whether solutions of a dynamical system that start near an equilibrium remain near it (stability) and whether they converge to it (asymptotic stability). Key result: if there exists a scalar "energy-like" function $V$ that is positive definite and whose derivative along trajectories is negative (semi)definite, then the equilibrium is (asymptotically) stable — without solving the differential equation. Applies to ODEs, control systems, mechanics, ecology, and numerical analysis. Introduced by Aleksandr Mikhailovich Lyapunov in his 1892 doctoral thesis.

💡 The Core Insight

Replace the question "where does every trajectory go?" with the simpler question "does some scalar function decrease along every trajectory?" If a positive-definite function $V(x)$ — a generalized energy — can only decrease as the system evolves, the state cannot escape the bowl-shaped level sets of $V$ . Monotone decrease of $V$ forces the trajectory toward the bottom of the bowl, i.e., toward the equilibrium. The genius is that $V$ need not be the physical energy and the system need not be solved.

DefinitionStability in the Sense of Lyapunov

The equilibrium $x=0$ of [1] is:

Stable if for every $\varepsilon > 0$ there exists $\delta = \delta(\varepsilon) > 0$ such that

2

\quad \|x(0)\| < \delta \;\Longrightarrow\; \|x(t)\| < \varepsilon \quad \forall\, t \ge 0.

Derivation hint: This is the $\varepsilon$ – $\delta$ definition of continuity transplanted from analysis: small initial deviations produce uniformly small future deviations; $\delta$ may shrink as $\varepsilon$ shrinks.

Asymptotically stable if it is stable and $\delta$ can be chosen so that

3

\quad \|x(0)\| < \delta \;\Longrightarrow\; \lim_{t\to\infty} x(t) = 0.

Derivation hint: Add convergence to mere boundedness; the basin of attraction is the set of all $x(0)$ for which this limit holds.

Exponentially stable if there exist constants $c, \lambda, \delta > 0$ with

4

\quad \|x(t)\| \le c\,\|x(0)\|\,e^{-\lambda t}, \qquad \|x(0)\| < \delta.

Derivation hint: Strengthen convergence to a guaranteed decay rate $\lambda$ ; this is the quantitative counterpart of asymptotic stability.

Unstable if it is not stable.

⚠️ Common mistake

Stability and asymptotic stability are logically distinct. A center (e.g., the undamped harmonic oscillator) is stable but not asymptotically stable: orbits stay near the origin forever without ever converging. Conversely, "all trajectories converge to 0" does not imply stability — there exist systems where trajectories first wander arbitrarily far before returning, violating [2] despite satisfying [3]. Both clauses matter.

DefinitionLyapunov Function Candidate and Its Orbital Derivative

Let $V:\mathcal{D}\to\mathbb{R}$ be continuously differentiable with $V(0)=0$ . Define its derivative along trajectories of [1] by the chain rule:

5

\quad \dot{V}(x) \;=\; \frac{d}{dt}V(x(t)) \;=\; \nabla V(x)\cdot f(x) \;=\; \sum_{i=1}^{n}\frac{\partial V}{\partial x_i}\,f_i(x).

Derivation hint: Differentiate $V(x(t))$ via the multivariable chain rule and substitute $\dot{x}_i = f_i(x)$ from [1]; note $\dot V$ depends only on $x$ , not on the unknown solution. $V$ is positive definite on $\mathcal{D}$ if $V(0)=0$ and $V(x)>0$ for $x\neq 0$ ; positive semidefinite if $V(x)\ge 0$ . Negative (semi)definiteness is defined analogously for $-V$ .

TheoremLyapunov's Direct (Second) Method

Let $x=0$ be an equilibrium of [1] and let $V:\mathcal{D}\to\mathbb{R}$ be $C^1$ and positive definite on a neighborhood $\mathcal{D}$ of the origin. 1. If $\dot V(x) \le 0$ on $\mathcal{D}$ (negative semidefinite), then $x=0$ is stable. 2. If $\dot V(x) < 0$ for all $x \in \mathcal{D}\setminus\{0\}$ (negative definite), then $x=0$ is asymptotically stable. 3. If additionally $\mathcal{D}=\mathbb{R}^n$ , $V$ is radially unbounded ( $V(x)\to\infty$ as $\|x\|\to\infty$ ), and $\dot V<0$ for all $x\neq 0$ , then $x=0$ is globally asymptotically stable.

TheoremLyapunov's Indirect (First) Method

Let $A = \left.\dfrac{\partial f}{\partial x}\right|_{x=0}$ be the Jacobian of $f$ at the origin. Then:

6

\quad \text{If } \operatorname{Re}\,\lambda_i(A) < 0 \;\forall i \;\Rightarrow\; \text{origin asymptotically stable; if some } \operatorname{Re}\,\lambda_i(A) > 0 \;\Rightarrow\; \text{unstable.}

Derivation hint: Write $f(x)=Ax+g(x)$ with $\|g(x)\|/\|x\|\to 0$ ; the linear part dominates the dynamics in a neighborhood, so eigenvalue placement of $A$ governs local behavior. The case $\operatorname{Re}\lambda_i = 0$ is inconclusive.

ProofProof of Stability — Theorem Part 1

Step 1 (Choose a sublevel set). Fix $\varepsilon>0$ small enough that the closed ball $\overline{B}_\varepsilon \subset \mathcal{D}$ . Let

7

\quad m = \min_{\|x\|=\varepsilon} V(x) > 0,

Derivation hint: $V$ is continuous and positive on the compact sphere $\{\|x\|=\varepsilon\}$ , so the minimum exists and is strictly positive by positive definiteness.

Step 2 (Pick a trapping level). Choose $0<\beta<m$ and define $\Omega_\beta = \{x\in B_\varepsilon : V(x)\le\beta\}$ . Because $V$ is continuous with $V(0)=0$ , there is $\delta>0$ with $B_\delta\subset\Omega_\beta$ .

Step 3 (Invariance). For $x(0)\in\Omega_\beta$ , since $\dot V\le 0$ ,

8

\quad V(x(t)) \le V(x(0)) \le \beta < m \qquad \forall\, t\ge 0.

Derivation hint: Integrate $\dot V \le 0$ from $0$ to $t$ ; $V$ is nonincreasing along trajectories so the trajectory cannot leave the sublevel set $\Omega_\beta$ .

Step 4 (Confinement). Were $\|x(t_1)\| = \varepsilon$ for some $t_1$ , then $V(x(t_1))\ge m > \beta$ , contradicting [8]. Hence $\|x(t)\|<\varepsilon$ for all $t\ge 0$ . Thus $\|x(0)\|<\delta \Rightarrow \|x(t)\|<\varepsilon$ , which is exactly Lyapunov stability. $\blacksquare$

∎

ProofProof of Asymptotic Stability — Theorem Part 2

Step 1. Stability holds by Part 1, so $x(t)$ stays in $\Omega_\beta$ , a compact set. $V(x(t))$ is nonincreasing and bounded below by $0$ , hence converges to some limit $V(x(t))\to c\ge 0$ .

Step 2 (Show $c=0$ ). Suppose $c>0$ . By continuity there is $d>0$ with $B_d \subset \{V<c\}$ . Then $x(t)$ remains in the annulus $\{d \le \|x\| \le \varepsilon\}$ , on which $\dot V$ attains a negative maximum $-\gamma<0$ :

9

\quad V(x(t)) = V(x(0)) + \int_0^t \dot V(x(\tau))\,d\tau \le V(x(0)) - \gamma t.

Derivation hint: On the compact annulus $\dot V$ is continuous and strictly negative, hence bounded above by some $-\gamma<0$ ; integrate this bound.

Step 3 (Contradiction). The right side of [9] $\to -\infty$ , contradicting $V\ge 0$ . Therefore $c=0$ , forcing $V(x(t))\to 0$ , and by positive definiteness $x(t)\to 0$ . $\blacksquare$

∎

TheoremLaSalle's Invariance Principle

Let $\Omega$ be a compact positively invariant set, and let $V$ be $C^1$ with $\dot V(x)\le 0$ on $\Omega$ . Let $E = \{x\in\Omega : \dot V(x)=0\}$ and let $M$ be the largest invariant set contained in $E$ . Then every solution starting in $\Omega$ approaches $M$ as $t\to\infty$ . This rescues asymptotic-stability conclusions when $\dot V$ is only negative semidefinite: if $M=\{0\}$ , the origin is asymptotically stable even though $\dot V$ vanishes off the equilibrium.

AM

Aleksandr Mikhailovich Lyapunov

1857–1918

A student of Chebyshev at St. Petersburg, Lyapunov defended his doctoral dissertation The General Problem of the Stability of Motion (1892), creating both the direct and indirect methods at a stroke and rigorously framing the modern definitions of stability. He also founded the theory of figures of equilibrium of rotating fluids and proved a central limit theorem under the condition that bears his name (the Lyapunov condition). His work was largely unknown in the West until the 1940s–1960s, when control engineers rediscovered it and made it the backbone of nonlinear systems theory. Following the death of his wife, Lyapunov took his own life in Odessa in 1918, days after she died of tuberculosis.

🧠 Intuition

Picture the level sets of $V$ as nested, bowl-shaped contour lines around the origin, like a topographic map of a valley with the equilibrium at the lowest point. The condition $\dot V \le 0$ says the state, as it moves, can only step downhill or sideways across these contours — never uphill. A ball placed anywhere in the bowl can never climb out, so it stays near the bottom (stability). If the rule is strictly downhill, $\dot V < 0$ , the ball must keep descending until it settles at the bottom (asymptotic stability). LaSalle's principle handles the subtle case where the ball can roll along a contour (where $\dot V = 0$ ) for a while: as long as it can't stay on such a contour forever, it must eventually resume its descent.

Figure

Phase plane with a Lyapunov level set

\Omega_\beta

trapping a spiraling trajectory that crosses contours inward

ExampleDamped Pendulum — Asymptotic Stability via LaSalle

Given. A pendulum with unit mass and length, gravity $g=9.81~\mathrm{m/s^2}$ , length $\ell = 1~\mathrm{m}$ , damping $b = 0.5~\mathrm{kg\,m^2/s}$ . With $x_1=\theta$ (angle, rad) and $x_2=\dot\theta$ (angular velocity, rad/s):

10

\quad \dot x_1 = x_2, \qquad \dot x_2 = -\frac{g}{\ell}\sin x_1 - b\,x_2.

Derivation hint: Newton's second law for rotation $\ddot\theta = -(g/\ell)\sin\theta - b\dot\theta$ written as a first-order system.

Step 1 — Candidate (total energy). Take $V =$ kinetic + potential energy:

11

\quad V(x) = \tfrac{1}{2}x_2^2 + \frac{g}{\ell}\,(1-\cos x_1).

Derivation hint: Rotational kinetic energy $\tfrac12 \dot\theta^2$ plus gravitational potential $\tfrac{g}{\ell}(1-\cos\theta)$ ; $V$ is positive definite near the origin since $1-\cos x_1 \approx x_1^2/2 > 0$ .

Step 2 — Orbital derivative.

12

\quad \dot V = \frac{g}{\ell}\sin x_1\,\dot x_1 + x_2\,\dot x_2 = \frac{g}{\ell}\sin x_1\,x_2 + x_2\!\left(-\frac{g}{\ell}\sin x_1 - b x_2\right) = -b\,x_2^2 \le 0.

Derivation hint: Substitute [10] into [5]; the cross terms cancel, leaving only the dissipation term $-b x_2^2$ .

Step 3 — LaSalle. $\dot V = 0 \iff x_2 = 0$ . On the set $\{x_2=0\}$ , the dynamics give $\dot x_2 = -(g/\ell)\sin x_1$ , so to remain invariant we need $\sin x_1 = 0 \Rightarrow x_1=0$ (locally). The largest invariant set is $\{0\}$ .

Result.

\boxed{\text{The origin (pendulum hanging at rest) is asymptotically stable; numeric } -b x_2^2 = -0.5\,x_2^2~\mathrm{J/s}.}

ExampleLinear System — Solving the Lyapunov Equation

Given. The stable linear system $\dot x = Ax$ with

13

\quad A = \begin{pmatrix} 0 & 1 \\ -2 & -3 \end{pmatrix}, \qquad \text{eigenvalues } \lambda = -1, -2.

Derivation hint: $\det(A-\lambda I) = \lambda^2 + 3\lambda + 2 = (\lambda+1)(\lambda+2)$ ; both eigenvalues have negative real part.

Step 1 — Lyapunov equation. Seek $V = x^\top P x$ with $P=P^\top \succ 0$ solving

14

\quad A^\top P + P A = -Q, \qquad Q = I.

Derivation hint: $\dot V = x^\top(A^\top P + P A)x$ ; choosing $-Q$ negative definite forces $\dot V<0$ . For Hurwitz $A$ and any $Q\succ 0$ , a unique $P\succ 0$ exists.

Step 2 — Solve. Writing $P=\begin{pmatrix}p_{11}&p_{12}\\p_{12}&p_{22}\end{pmatrix}$ and matching entries of $A^\top P + PA = -I$ :

15

\quad -4p_{12} = -1,\quad p_{11}-3p_{12}-2p_{22}=0,\quad 2p_{12}-6p_{22}=-1.

Derivation hint: Expand the matrix product symbolically and equate the three independent (symmetric) entries to those of $-I$ .

Step 3 — Values. From these: $p_{12}=\tfrac14$ , then $p_{22}=\tfrac{1}{6}\!\left(1+2\cdot\tfrac14\right)=\tfrac14$ , and $p_{11}=3\cdot\tfrac14+2\cdot\tfrac14=\tfrac54$ .

\boxed{P = \begin{pmatrix} 1.25 & 0.25 \\ 0.25 & 0.25 \end{pmatrix},\quad \det = 0.25 > 0,\ p_{11}>0 \Rightarrow P\succ 0,\ V=x^\top P x \text{ certifies global asymptotic stability.}}

⚠️ Common mistake

A failed Lyapunov candidate does not prove instability. If you try some $V$ and find $\dot V$ is positive somewhere, you have learned nothing about stability — only that this particular $V$ does not work. Lyapunov's conditions are sufficient, not necessary, in the constructive sense. To prove instability you need a separate tool (Chetaev's theorem) or the indirect method with an eigenvalue in the right half-plane.

⚠️ Common mistake

$\dot V < 0$ for all $x\neq 0$ is not enough for global asymptotic stability without radial unboundedness. Without $V(x)\to\infty$ , level sets can be unbounded and trajectories can escape to infinity along them while $V$ stays bounded. Forgetting this hypothesis is the single most common error in global stability claims.

CounterexampleConvergence without stability

Consider a planar system in polar coordinates where every trajectory eventually reaches the equilibrium but must first traverse a long excursion: trajectories from points just off the positive $x$ -axis travel almost fully around a circle before spiraling in. Formally one constructs $f$ so that $r(t)\to 0$ for all initial conditions (so [3] holds), yet for any $\delta$ there is a start with $\|x(0)\|<\delta$ whose trajectory reaches $\|x\|=1$ (so [2] fails for $\varepsilon<1$ ). Moral: attractivity does not imply Lyapunov stability; the full definition of asymptotic stability requires both clauses.

CounterexampleThe borderline case of the indirect method

For $\dot x = -x^3$ the Jacobian at $0$ is $A=0$ , with eigenvalue $\lambda=0$ — the indirect method is silent. Yet $V=x^2$ gives $\dot V = -2x^4 < 0$ , so the origin is asymptotically stable. For $\dot x = +x^3$ the Jacobian is also $0$ but the origin is unstable. Identical linearizations, opposite conclusions: when $\operatorname{Re}\lambda_i=0$ , nonlinear terms decide the outcome and only the direct method (or center-manifold analysis) resolves it.

connection

Lyapunov theory sits at the crossroads of several fields. The indirect method reduces to Linear Systems Theory and the eigenvalue (Hurwitz) criterion near equilibria. For linear systems, solving [14] is an instance of the algebraic <span class="wiki-link-missing">Lyapunov Equation</span>, intimately tied to controllability and observability Gramians. The notion of an energy-decreasing function generalizes the Hamiltonian Mechanics energy and connects to dissipativity and passivity in Control Theory. LaSalle's principle and $\omega$ -limit sets link to <span class="wiki-link-missing">Dynamical Systems</span> and ergodic theory, while converse theorems relate to the smoothness theory of flows. The discrete-time analog replaces $\dot V$ with $\Delta V = V(x_{k+1})-V(x_k)$ and connects to Markov Chains via stochastic Lyapunov (Foster–Lyapunov) drift conditions.

🔬 Where this appears

⏳ History

- 1788 — Lagrange's theorem: an equilibrium of a conservative mechanical system at a strict minimum of potential energy is stable — the conceptual seed of energy methods.

- 1892 — Lyapunov's doctoral dissertation The General Problem of the Stability of Motion (Kharkov) introduces both the direct and indirect methods and the modern stability definitions.

- 1907 — French translation by Édouard Davaux published in Annales de Toulouse, giving the work limited Western circulation.

- 1949 — N. G. Chetaev publishes Stability of Motion, adding the Chetaev instability theorem and consolidating the Soviet school.

- 1950s — Translation efforts and the work of Malkin, Krasovskii, and Barbashin extend converse theorems and global results (Barbashin–Krasovskii).

- 1960 — Joseph LaSalle's invariance principle generalizes asymptotic-stability conclusions to negative-semidefinite

\dot V

.

- 1961 — LaSalle & Lefschetz's monograph Stability by Liapunov's Direct Method brings the theory firmly into Western control engineering.

- 1960s–70s — Kalman, Yakubovich, and Popov connect Lyapunov functions to frequency-domain criteria (KYP lemma, Popov criterion, absolute stability).

- 1994–present — Boyd et al. recast Lyapunov conditions as linear matrix inequalities (LMIs), making computational search for quadratic and sum-of-squares Lyapunov functions tractable.

open-problem

The construction problem. Converse theorems (Massera, Kurzweil) guarantee that an asymptotically stable equilibrium admits a smooth Lyapunov function, yet there is no universal constructive procedure for nonlinear systems. Sum-of-squares (SOS) programming searches polynomial $V$ but is limited by polynomial degree and the gap between nonnegativity and SOS representability (Motzkin-type counterexamples). Estimating the region of attraction sharply is generically hard; Lyapunov sublevel sets typically give conservative under-estimates. For time-varying and switched systems, common vs. multiple Lyapunov functions raise deep existence questions, and deciding stability of switched linear systems under arbitrary switching is undecidable in general. Finally, data-driven / learned Lyapunov functions (neural Lyapunov functions) currently lack the rigorous, scalable verification guarantees that classical certificates enjoy.

Formal Statement

Why It Holds / Derivation

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also