02.12.08 · analysis / ode

Lyapunov stability (direct method)

shipped3 tiersLean: none

Anchor (Master): Lyapunov 1892 *The General Problem of the Stability of Motion* (Kharkov PhD thesis; English translation, Taylor & Francis, 1992) — originator; LaSalle 1960 *Some extensions of Liapunov's second method* (IRE Trans. Circuit Theory CT-7) — invariance principle; Krasovskii 1959 *Stability of Motion* (Stanford University Press, English translation 1963); Hahn 1967 *Stability of Motion*; Hartman *Ordinary Differential Equations* Ch. 9; Arnold *Ordinary Differential Equations* Ch. 5 §23

Intuition [Beginner]

A stable resting position is one you can lean on without falling over. Think of a marble in a bowl: nudge it gently, and it rolls back toward the bottom; the deeper the bowl, the harder it is to push the marble out. A marble balanced on top of an upside-down bowl is the opposite — any nudge sends it rolling away. Lyapunov stability is the precise way of saying that a resting state of a dynamical system behaves like a marble in a bowl rather than on a dome.

Lyapunov's idea is to find a single number — an energy-like quantity — attached to every state of the system, with two properties. First, the resting state has the smallest energy, and every other nearby state has strictly more. Second, the energy never increases as the system evolves: if you start near the rest state, the energy you started with bounds the energy at every later time, and the bound traps you near the bottom of the bowl. A function with these two properties is called a Lyapunov function. Its existence is enough to prove the rest state is stable.

The real triumph is that you do not need to solve the differential equation. You only need to find one good function and check two inequalities. This is why Lyapunov called his approach the direct method: it sidesteps the explicit solution. The same idea builds toward control theory, mechanics, and modern dynamical systems, where finding a suitable function is the single technique that proves stability for systems too complicated to solve directly.

Visual [Beginner]

A simple picture: a bowl-shaped surface representing the Lyapunov function $V$ , with its lowest point at the equilibrium. A trajectory of the dynamical system is drawn on the horizontal plane below; arrows on the surface show the height $V$ decreasing along the trajectory. The trajectory stays inside the bowl because, once the height drops, the rules of the system do not allow it to climb back up. A second smaller bowl is nested inside — the level set of $V$ at the chosen starting height — and the trajectory stays inside that bowl for all future time.

The picture summarises the whole strategy: a positive function with its minimum at the equilibrium, decreasing along the flow, traps every nearby trajectory inside a small bowl around the rest point.

Worked example [Beginner]

Consider a pendulum with friction. Let $θ$ be the angle from straight down and let $\dot{θ}$ be the angular velocity. The motion satisfies $\ddot{θ} + α \dot{θ} + sin θ = 0$ , where $α > 0$ is the friction coefficient. The resting state is $θ = 0$ , $\dot{θ} = 0$ — the pendulum hanging straight down at rest.

Step 1. Pick the total mechanical energy as the candidate Lyapunov function: $$ V(\theta, \dot\theta) = \tfrac{1}{2} \dot\theta^2 + (1 - \cos\theta). $$ Read off two facts. First, $V (0, 0) = 0$ . Second, for any state different from the rest state and within the range $∣ θ ∣ < π$ , the value is strictly positive: $1 - cos θ > 0$ when $θ$ is non-zero, and $\dot{θ}^{2} \geq 0$ with equality only when $\dot{θ} = 0$ . So $V$ is positive definite around the equilibrium.

Step 2. Compute the rate of change of $V$ along the motion. Using the chain rule and the equation of motion: $$ \dot V = \dot\theta \cdot \ddot\theta + \sin\theta \cdot \dot\theta = \dot\theta (\ddot\theta + \sin\theta) = \dot\theta \cdot (-\alpha \dot\theta) = -\alpha \dot\theta^2. $$ The result is $\leq 0$ at every state, with equality only when $\dot{θ} = 0$ . So $V$ is non-increasing along the motion.

Step 3. Read off the conclusion. The function $V$ is positive away from the equilibrium and non-increasing along the motion, which means any trajectory starting close to the rest state stays close. The pendulum hanging straight down is Lyapunov stable.

Step 4. Friction is doing real work. Without friction ( $α = 0$ ), the calculation gives $\dot{V} = 0$ — energy is conserved, and the pendulum swings forever. With friction, $\dot{V}$ becomes negative when $\dot{θ} \neq = 0$ , which is what eventually drains the energy and brings the pendulum to rest. The full statement that the friction case is asymptotically stable needs a finer argument, but the basic Lyapunov stability falls straight out of the two-line calculation above.

What this tells us: a single function — total energy — plus one short computation closes the question of stability for the pendulum. No need to solve the equation. No need to draw the phase portrait. Just check positive definiteness of $V$ and non-positivity of $\dot{V}$ .

Check your understanding [Beginner]

Formal definition [Intermediate+]

Let $U \subseteq R^{n}$ be an open neighbourhood of a point $x_{0}$ , and let $f : U \to R^{n}$ be a $C^{1}$ vector field with $f (x_{0}) = 0$ . The autonomous ordinary differential equation $\overset{x}{˙} = f (x)$ has $x_{0}$ as an equilibrium (also called a fixed point or singular point). Write $φ_{t} (x)$ for the local flow of $f$ — the unique solution to $\overset{x}{˙} = f (x)$ , $x (0) = x$ , defined on a maximal time interval.

Definition (Lyapunov stability). The equilibrium $x_{0}$ is Lyapunov stable if for every $ε > 0$ there exists $δ > 0$ such that $$ |x - x_0| < \delta \implies |\varphi_t(x) - x_0| < \varepsilon \text{ for every } t \geq 0 $$ whenever $φ_{t} (x)$ is defined. The equilibrium is **asymptotically stable** if it is Lyapunov stable and additionally there exists $η > 0$ such that $∥ x - x_{0} ∥ < η$ implies $φ_{t} (x) \to x_{0}$ as $t \to \infty$ . The equilibrium is unstable if it is not Lyapunov stable.

Definition (Lyapunov function). A function $V : U \to R$ of class $C^{1}$ is a Lyapunov function for the equilibrium $x_{0}$ if

$V (x_{0}) = 0$ and $V (x) > 0$ for every $x \in U$ with $x \neq = x_{0}$ ;
the orbital derivative $\dot{V} (x) := \nabla V (x) \cdot f (x)$ satisfies $\dot{V} (x) \leq 0$ for every $x \in U$ .

The Lyapunov function is strict if additionally $\dot{V} (x) < 0$ for every $x \in U$ with $x \neq = x_{0}$ .

The orbital derivative records the rate of change of $V$ along trajectories: if $γ (t) = φ_{t} (x)$ is an integral curve of $f$ , the chain rule gives $\frac{d}{d t} V (γ (t)) = \nabla V (γ (t)) \cdot \overset{γ}{˙} (t) = \nabla V (γ (t)) \cdot f (γ (t)) = \dot{V} (γ (t))$ . The sign condition on $\dot{V}$ is therefore the sign condition on the time derivative of $V$ along every trajectory.

Definition (Hurwitz matrix). A real $n \times n$ matrix $A$ is Hurwitz (or stable) if every eigenvalue $λ$ of $A$ has $Re λ < 0$ .

For a $C^{1}$ vector field $f$ with $f (x_{0}) = 0$ , the linearization at $x_{0}$ is the linear map $A = D f (x_{0}) : R^{n} \to R^{n}$ . The linearized equation $\overset{y}{˙} = A y$ is the leading-order approximation of $\overset{x}{˙} = f (x)$ near $x_{0}$ under the change of variables $y = x - x_{0}$ .

Counterexamples to common slips

The condition $V (x_{0}) = 0$ matters: a function that is positive everywhere on $U$ — including at $x_{0}$ — fails to certify stability of $x_{0}$ specifically, because its level sets do not shrink to $x_{0}$ . The vanishing at the equilibrium is what makes the sub-level sets ${V \leq c}$ shrink to ${x_{0}}$ as $c \to 0$ .
$\dot{V} \leq 0$ gives Lyapunov stability but not asymptotic stability. The undamped pendulum has $\dot{V} \equiv 0$ along trajectories — energy is conserved — so the equilibrium $θ = \dot{θ} = 0$ is Lyapunov stable but not asymptotically stable.
The linearization criterion has a hyperbolic-only converse. If $D f (x_{0})$ has eigenvalues with both negative and positive real parts, the equilibrium is unstable; if every eigenvalue has negative real part, the equilibrium is asymptotically stable. But if some eigenvalue has zero real part and the rest are non-positive, the linearization is inconclusive — the nonlinear system can be stable, unstable, or asymptotically stable depending on the higher-order terms.

Key theorem with proof [Intermediate+]

Theorem (Lyapunov's direct method, stability case). Let $f : U \to R^{n}$ be $C^{1}$ on an open neighbourhood $U$ of an equilibrium $x_{0}$ . If there exists a Lyapunov function $V : U \to R$ for $x_{0}$ — that is, $V \in C^{1} (U)$ with $V (x_{0}) = 0$ , $V (x) > 0$ for $x \neq = x_{0}$ in $U$ , and $\dot{V} (x) := \nabla V (x) \cdot f (x) \leq 0$ on $U$ — then $x_{0}$ is Lyapunov stable. If $V$ is strict, then $x_{0}$ is asymptotically stable. (See ^{[Lyapunov 1892]}, ^{[Arnold *ODE* §23]}.)

Proof. Fix $ε > 0$ small enough that the closed ball $\overline{B}_{ε} (x_{0}) = {x : ∥ x - x_{0} ∥ \leq ε}$ is contained in $U$ .

Sub-level-set construction. Let $S_{ε} = {x : ∥ x - x_{0} ∥ = ε}$ be the boundary sphere, a compact set on which the continuous function $V$ attains a positive minimum $m_{ε} := min_{x \in S_{ε}} V (x) > 0$ — positive because $V > 0$ off $x_{0}$ and $x_{0} \in / S_{ε}$ . Define $$ \Omega_\varepsilon = {x \in B_\varepsilon(x_0) : V(x) < m_\varepsilon}. $$ This is an open neighbourhood of $x_{0}$ contained in the interior of the ball $B_{ε} (x_{0})$ , because every point on the boundary sphere has $V$ -value at least $m_{ε}$ and is excluded by the strict inequality.

Invariance of $Ω_{ε}$ . For $x \in Ω_{ε}$ , the trajectory $φ_{t} (x)$ exists on some maximal forward interval $[0, T)$ . Along this trajectory, the orbital-derivative condition gives $$ \frac{d}{dt} V(\varphi_t(x)) = \dot V(\varphi_t(x)) \leq 0, $$ so $V (φ_{t} (x)) \leq V (x) < m_{ε}$ for every $t \in [0, T)$ . Suppose $φ_{t} (x)$ exits the closed ball $\overline{B}_{ε} (x_{0})$ at some $t^{*} < T$ . By continuity, $φ_{t^{*}} (x) \in S_{ε}$ (the trajectory crosses the boundary), so $V (φ_{t^{*}} (x)) \geq m_{ε}$ , contradicting $V (φ_{t} (x)) < m_{ε}$ for all $t \in [0, T)$ . Therefore $φ_{t} (x) \in B_{ε} (x_{0})$ for every $t \in [0, T)$ , the trajectory stays bounded in $U$ , and by standard ODE theory $T = \infty$ .

Selecting $δ$ . By continuity of $V$ at $x_{0}$ with $V (x_{0}) = 0$ , there exists $δ > 0$ with $δ \leq ε$ such that $∥ x - x_{0} ∥ < δ$ implies $V (x) < m_{ε}$ , that is, $x \in Ω_{ε}$ . By the invariance just proved, $∥ φ_{t} (x) - x_{0} ∥ < ε$ for every $t \geq 0$ . This is Lyapunov stability.

Asymptotic case. Assume in addition $\dot{V} (x) < 0$ for every $x \neq = x_{0}$ in $U$ . Take $η = δ$ for the same $ε$ as above, and fix $x$ with $0 < ∥ x - x_{0} ∥ < η$ . The function $t \mapsto V (φ_{t} (x))$ is non-increasing and bounded below by $0$ , so it converges to some limit $V_{\infty} \geq 0$ as $t \to \infty$ . Suppose $V_{\infty} > 0$ . The set $K = {y \in \overline{B}_{ε} (x_{0}) : V_{\infty} \leq V (y) \leq V (x)}$ is compact and excludes $x_{0}$ (because $V (x_{0}) = 0 < V_{\infty}$ ), so on $K$ the continuous function $- \dot{V}$ attains a positive minimum $μ > 0$ . The trajectory stays inside $K$ from some time on, because $V (φ_{t} (x)) \in [V_{\infty}, V (x)]$ for every $t \geq 0$ by monotonicity. Then $$ V(\varphi_t(x)) = V(x) + \int_0^t \dot V(\varphi_s(x)) , ds \leq V(x) - \mu t \to -\infty, $$ contradicting $V \geq 0$ . Therefore $V_{\infty} = 0$ . Continuity of $V$ at $x_{0}$ and the fact that $V_{\infty} = 0$ force $φ_{t} (x) \to x_{0}$ : for every $ρ > 0$ , the value $V (φ_{t} (x))$ eventually drops below $min_{∥ y - x_{0} ∥ = ρ} V (y) > 0$ , which traps the trajectory inside $B_{ρ} (x_{0})$ for all large $t$ . The equilibrium is asymptotically stable. $□$

Bridge. Lyapunov's direct method builds toward every modern stability technique in dynamical systems and control theory. The foundational reason it works is the sub-level-set construction: positive definiteness of $V$ makes the sets ${V \leq c}$ shrink to ${x_{0}}$ as $c \to 0$ , and non-positivity of $\dot{V}$ makes them invariant under the forward flow, so each pair $(c, Ω)$ produces a trapping neighbourhood. This is exactly the same principle that appears again in 02.12.10 (Poincaré-Bendixson) as the requirement that planar trajectories cannot escape bounded invariant regions, and identifies energy with stability in 05.00.04 (Noether): a conserved quantity is a Lyapunov function with $\dot{V} = 0$ , the boundary case between the strict and non-strict statements.

The central insight is that stability of an equilibrium follows from the existence of a single auxiliary function whose level sets trap trajectories — there is no need to solve the equation. This pattern recurs in the linearization theorem proved below, where a quadratic Lyapunov function constructed from a Hurwitz linearization certifies asymptotic stability of the nonlinear system. The bridge is the recognition that the Lyapunov-equation construction $A^{⊤} P + P A = - I$ generalises the energy-function picture to any linear system with spectrum in the open left half-plane, identifying the asymptotic-stability condition with the existence of a positive-definite quadratic form whose orbital derivative is negative definite. Putting these together, one direct-method framework — pick a positive-definite function, check the orbital derivative — handles mechanical systems with energy, linearized systems with a quadratic form from the Lyapunov equation, and a wide class of nonlinear systems treated in control theory and nonlinear stability analysis.

Exercises [Intermediate+]

Exercise 4 (medium, short-answer).

Consider the Hamiltonian system $\overset{x}{˙} = y$ , $\overset{y}{˙} = - x$ on $R^{2}$ . Show that $V (x, y) = \frac{1}{2} (x^{2} + y^{2})$ is a Lyapunov function but not a strict Lyapunov function, and explain what conclusion follows about the origin.

Hint

Compute $\dot{V}$ along the system. The Hamiltonian flow conserves energy.

Answer

Computing: $\dot{V} = x \overset{x}{˙} + y \overset{y}{˙} = x \cdot y + y \cdot (- x) = 0$ everywhere. So $V$ is conserved along trajectories: $\dot{V} \equiv 0$ . This means $V$ is a Lyapunov function (the conditions $V (0) = 0$ , $V > 0$ elsewhere, $\dot{V} \leq 0$ all hold) but not strict ( $\dot{V}$ is not negative away from the origin). The conclusion is Lyapunov stable but not asymptotically stable: trajectories stay on circles of constant radius, so they remain close to the origin without converging to it. This is the harmonic-oscillator picture, where every level set of the energy is an invariant orbit. Rubric: full credit for $\dot{V} = 0$ , the distinction between Lyapunov stable and asymptotically stable, and the circular-orbit picture.

Exercise 5 (medium, short-answer).

For the system $\overset{x}{˙}_{1} = - x_{1} + x_{2}^{2}$ , $\overset{x}{˙}_{2} = - x_{2}$ , show that the origin is asymptotically stable using $V (x_{1}, x_{2}) = x_{1}^{2} + x_{2}^{2}$ or a modification thereof.

Hint

The naive choice $V = x_{1}^{2} + x_{2}^{2}$ gives a sign-indefinite orbital derivative because of the coupling term $x_{2}^{2}$ . Modify the function — try $V = x_{1}^{2} + c x_{2}^{4}$ for some $c > 0$ , or use a weighted form.

Answer

Try $V (x_{1}, x_{2}) = x_{1}^{2} + x_{2}^{2}$ . Then $\dot{V} = 2 x_{1} (- x_{1} + x_{2}^{2}) + 2 x_{2} (- x_{2}) = - 2 x_{1}^{2} - 2 x_{2}^{2} + 2 x_{1} x_{2}^{2}$ . The last term has sign depending on $x_{1}$ , so $\dot{V}$ is not globally negative. But near the origin, $∣2 x_{1} x_{2}^{2} ∣ \leq ∣ x_{1} ∣^{2} + x_{2}^{4} \leq x_{1}^{2} + x_{2}^{2} \cdot x_{2}^{2}$ , which is dominated by $2 x_{1}^{2} + 2 x_{2}^{2}$ inside the small ball ${x_{2}^{2} < 1}$ . So $\dot{V} \leq - 2 x_{1}^{2} - 2 x_{2}^{2} + (x_{1}^{2} + x_{2}^{2}) = - (x_{1}^{2} + x_{2}^{2}) < 0$ on the punctured ball ${0 < x_{1}^{2} + x_{2}^{2} < 1} \cap {x_{2}^{2} < 1}$ . Strict Lyapunov function locally — asymptotic stability. The linearization $A = diag (- 1, - 1)$ has eigenvalues $- 1, - 1$ , both negative; the linearization theorem (below) also yields asymptotic stability immediately. Rubric: full credit for the local sign-bound on $\dot{V}$ plus the asymptotic-stability conclusion.

Exercise 7 (hard, short-answer).

Prove that if $A$ is a real $n \times n$ Hurwitz matrix, then the integral $P = \int_{0}^{\infty} e^{A^{⊤} t} e^{A t} d t$ converges absolutely and is the unique symmetric positive-definite solution of $A^{⊤} P + P A = - I$ .

Hint

Convergence uses the exponential decay $∥ e^{A t} ∥ \leq C e^{- α t}$ for some $α > 0$ depending on the spectral abscissa of $A$ . To verify the Lyapunov equation, differentiate $e^{A^{⊤} t} e^{A t}$ with respect to $t$ and integrate from $0$ to $\infty$ .

Answer

Convergence: since $A$ is Hurwitz, the spectral abscissa $α = - max_{λ} Re λ > 0$ , and there exists $C > 0$ with $∥ e^{A t} ∥ \leq C e^{- α t /2}$ for $t \geq 0$ (the Jordan-block growth is absorbed into the slack). Then $∥ e^{A^{⊤} t} e^{A t} ∥ \leq C^{2} e^{- α t}$ , integrable on $[0, \infty)$ . So $P$ is well-defined and symmetric: $P^{⊤} = \int_{0}^{\infty} e^{A^{⊤} t} e^{A t} d t = P$ . Positive-definiteness: for any $v \neq = 0$ , $v^{⊤} P v = \int_{0}^{\infty} ∥ e^{A t} v ∥^{2} d t > 0$ because $e^{A t} v \neq = 0$ for almost every $t$ . Lyapunov equation: differentiate inside the integral, $\frac{d}{d t} (e^{A^{⊤} t} e^{A t}) = A^{⊤} e^{A^{⊤} t} e^{A t} + e^{A^{⊤} t} e^{A t} A$ , and integrate by the fundamental theorem of calculus: $\int_{0}^{\infty} \frac{d}{d t} (e^{A^{⊤} t} e^{A t}) d t = lim_{T \to \infty} e^{A^{⊤} T} e^{A T} - I = 0 - I = - I$ . The integrand equals $A^{⊤} P + P A$ after integrating, giving $A^{⊤} P + P A = - I$ . Uniqueness: if $Q$ also satisfies $A^{⊤} Q + Q A = - I$ , then $A^{⊤} (P - Q) + (P - Q) A = 0$ , and the Sylvester-equation lemma (or a direct spectral argument using the Kronecker-sum eigenvalues $λ_{i} + λ_{j}$ being non-zero for a Hurwitz $A$ ) forces $P - Q = 0$ . Rubric: full credit for convergence, symmetry, positive-definiteness, the Lyapunov equation, and uniqueness.

Exercise 8 (hard, short-answer).

State and prove Chetaev's instability theorem: if there exists a $C^{1}$ function $W : U \to R$ with $W (x_{0}) = 0$ , an open set $Ω$ with $x_{0} \in \overline{Ω}$ where $W > 0$ , $W = 0$ on the boundary part of $\partial Ω$ inside $U$ , and $\dot{W} > 0$ on $Ω$ , then $x_{0}$ is unstable.

Hint

The argument is dual to the Lyapunov proof: $W$ grows along trajectories starting in $Ω$ , so a trajectory cannot stay in a bounded sub-region.

Answer

For any $δ > 0$ , pick $x_{1} \in Ω \cap B_{δ} (x_{0})$ with $W (x_{1}) > 0$ — possible because $x_{0} \in \overline{Ω}$ and $W$ is continuous with $W > 0$ on $Ω$ . The trajectory $φ_{t} (x_{1})$ stays in $Ω$ as long as it remains in $\overline{B}_{ε} (x_{0})$ , because $W$ is strictly increasing along the trajectory (so $W (φ_{t} (x_{1})) \geq W (x_{1}) > 0$ rules out hitting the part of $\partial Ω$ where $W = 0$ ). On the compact set $K = \overline{Ω} \cap \overline{B}_{ε} (x_{0}) \cap {W \geq W (x_{1})}$ the continuous function $\dot{W}$ has a positive minimum $μ > 0$ . So $W (φ_{t} (x_{1})) \geq W (x_{1}) + μ t$ as long as the trajectory remains in $K$ . But $W$ is bounded above on the compact $K$ , contradicting linear growth. The trajectory must leave $K$ , and since it cannot cross ${W = 0}$ , it must exit through $\partial B_{ε} (x_{0})$ — that is, $∥ φ_{t} (x_{1}) - x_{0} ∥ \geq ε$ for some $t$ . Since $δ$ was arbitrary and $ε$ fixed, $x_{0}$ is not Lyapunov stable. Rubric: full credit for the choice of $x_{1}$ , the growth bound on $W$ , the trajectory's forced exit, and the instability conclusion.

Advanced results [Master]

Theorem (Lyapunov's linearization theorem, asymptotic-stability case). Let $f : U \to R^{n}$ be $C^{1}$ on an open neighbourhood of an equilibrium $x_{0}$ , and let $A = D f (x_{0})$ be the linearization. If $A$ is Hurwitz — that is, every eigenvalue of $A$ has negative real part — then $x_{0}$ is asymptotically stable for $\overset{x}{˙} = f (x)$ . If $A$ has at least one eigenvalue with positive real part, then $x_{0}$ is unstable.

The proof of the asymptotic-stability case uses the Lyapunov equation. Since $A$ is Hurwitz, the matrix equation $A^{⊤} P + P A = - I$ has a unique positive-definite symmetric solution $P$ , given explicitly by $P = \int_{0}^{\infty} e^{A^{⊤} t} e^{A t} d t$ . The quadratic form $V (x) = ⟨ P (x - x_{0}), x - x_{0} ⟩$ is then a strict Lyapunov function for the nonlinear system in a neighbourhood of $x_{0}$ : positive-definite from $P ≻ 0$ , and the orbital-derivative bound $\dot{V} \leq - \frac{1}{2} ∥ x - x_{0} ∥^{2}$ holds on a small enough ball. The full proof appears in the Full proof set below. The instability case (some eigenvalue with positive real part) is handled by Chetaev's theorem applied to the unstable eigenspace; details are in Hartman ODE §IX ^{[Hartman *ODE* Ch. 9]}.

Theorem (Hartman-Grobman, sharper statement). Let $f : U \to R^{n}$ be $C^{1}$ on an open neighbourhood of an equilibrium $x_{0}$ , and assume the linearization $A = D f (x_{0})$ is hyperbolic — every eigenvalue has nonzero real part. Then there exists a homeomorphism $h$ defined on a neighbourhood of $x_{0}$ that conjugates the nonlinear flow $φ_{t}$ to the linear flow $e^{A t}$ : $h (φ_{t} (x)) = e^{A t} h (x)$ for every $x$ near $x_{0}$ and $t$ sufficiently small. (See Hartman 1960 ^{[Hartman 1960 *PAMS* 11]}, Grobman 1959.)

The Hartman-Grobman theorem strengthens the Lyapunov linearization theorem: not only do stability conclusions transfer between $\overset{x}{˙} = f (x)$ and its linearization at a hyperbolic equilibrium, but the entire phase portrait near the equilibrium is topologically equivalent to that of the linearization. The homeomorphism is $C^{0}$ in general but not $C^{1}$ — the smooth-conjugacy refinement requires Sternberg's resonance-free condition, a strictly stronger hypothesis.

Theorem (LaSalle's invariance principle). Let $V : U \to R$ be $C^{1}$ with $V (x_{0}) = 0$ , $V > 0$ off $x_{0}$ , and $\dot{V} \leq 0$ on $U$ . Let $E = {x \in U : \dot{V} (x) = 0}$ and let $M \subseteq E$ be the largest forward-invariant subset of $E$ under the flow. Then every bounded trajectory starting in a sub-level set of $V$ that lies in $U$ converges to $M$ as $t \to \infty$ . (LaSalle 1960 ^{[LaSalle 1960 *IRE CT-7*]}.)

LaSalle's principle recovers asymptotic stability when $\dot{V} \leq 0$ is only semidefinite, provided the only invariant subset of $E$ is the equilibrium itself. For the pendulum with friction, $\dot{V} = - α \dot{θ}^{2}$ vanishes on the set ${\dot{θ} = 0}$ , and the only forward-invariant subset of that set is the equilibrium $(θ, \dot{θ}) = (0, 0)$ (because once $\dot{θ} = 0$ at a point with $θ \neq = 0$ , the equation $\ddot{θ} = - sin θ \neq = 0$ immediately drives $\dot{θ}$ away from zero). LaSalle then upgrades the Lyapunov-stable conclusion to asymptotic stability, completing the analysis of the damped pendulum.

Theorem (Krasovskii's converse stability theorem). If $x_{0}$ is an asymptotically stable equilibrium of $\overset{x}{˙} = f (x)$ with $f$ of class $C^{k}$ for some $k \geq 1$ , then there exists a $C^{k}$ strict Lyapunov function $V$ defined on a neighbourhood of $x_{0}$ . (Krasovskii 1959 ^{[Krasovskii 1959]}, with the full smooth converse refined by Massera 1956 and Kurzweil 1955.)

Krasovskii's converse theorem closes the loop: the direct method is not just a sufficient condition for stability — for asymptotically stable equilibria of smooth systems, a Lyapunov function always exists. This converse legitimises searching for Lyapunov functions as a general technique, since one is guaranteed to be there. Construction is via a flow-time integral: $V (x) = \int_{0}^{\infty} ∥ φ_{t} (x) ∥^{2} d t$ is well-defined for $x$ in the domain of attraction by exponential decay of trajectories, and a routine check verifies the Lyapunov conditions.

Theorem (Center manifold reduction at non-hyperbolic equilibria). Let $f : U \to R^{n}$ be $C^{k}$ with $f (x_{0}) = 0$ and let $A = D f (x_{0})$ . Decompose $R^{n} = E^{s} \oplus E^{c} \oplus E^{u}$ into the stable, centre, and unstable eigenspaces of $A$ (with $E^{s}$ the span of eigenvectors with negative real part, $E^{c}$ with zero, $E^{u}$ with positive). Then there exists a locally invariant $C^{k}$ centre manifold $W^{c}$ tangent to $E^{c}$ at $x_{0}$ , and the stability of $x_{0}$ is determined by the reduced dynamics on $W^{c}$ . (Pliss 1964; Carr 1981, Vanderbauwhede 1989; see ^{[Khalil Ch. 4]}.)

The centre manifold theorem extends the linearization-based stability analysis to non-hyperbolic equilibria, the case where Hartman-Grobman fails. The dimension of the centre manifold equals $dim E^{c}$ , and only that part of the dynamics matters for stability when $E^{u} = {0}$ . Combined with Lyapunov's direct method applied on the centre manifold, this gives a complete framework for stability of any equilibrium of a smooth system.

Theorem (input-to-state stability of nonlinear control systems). For a control system $\overset{x}{˙} = f (x, u)$ with $f (0, 0) = 0$ , the unforced system is input-to-state stable (ISS) at the origin if there exist a $C^{1}$ ISS-Lyapunov function $V$ and class- $K$ functions $α_{1}, α_{2}, α_{3}, χ$ such that $α_{1} (∥ x ∥) \leq V (x) \leq α_{2} (∥ x ∥)$ and $∥ x ∥ \geq χ (∥ u ∥) ⟹ \dot{V} (x, u) \leq - α_{3} (∥ x ∥)$ . (Sontag 1989 ^{[Sontag 1989 *IEEE TAC 34*]}.)

ISS is the modern generalisation of Lyapunov's direct method to systems with disturbances or inputs. Sontag's ISS framework underlies control-theoretic robustness analysis, adaptive control, and a substantial fraction of the nonlinear-control literature since 1990. The construction is the direct method augmented with a comparison condition on $∥ u ∥$ .

Theorem (Lyapunov functions in stochastic stability). For an Itô stochastic differential equation $d X_{t} = f (X_{t}) d t + g (X_{t}) d W_{t}$ with $f (x_{0}) = 0$ , $g (x_{0}) = 0$ , the equilibrium is stochastically stable if there exists a $C^{2}$ function $V$ with $V (x_{0}) = 0$ , $V > 0$ off $x_{0}$ , and the stochastic orbital derivative $$ \mathcal{L} V(x) := \nabla V(x) \cdot f(x) + \tfrac{1}{2} \mathrm{tr}\bigl(g(x)^\top \mathrm{Hess},V(x) , g(x)\bigr) $$ satisfies $L V \leq 0$ near $x_{0}$ . (Khasminskii 1969 ^{[Khasminskii 1969]}.)

The stochastic extension replaces the orbital derivative with the infinitesimal generator of the Itô diffusion, picking up the half-Hessian correction familiar from Itô's formula. Otherwise the direct-method picture is identical: positive-definite $V$ , generator $\leq 0$ , sub-level sets trap the diffusion.

Synthesis. The direct method is the foundational reason stability questions for finite-dimensional dynamical systems can be answered without solving the differential equation. The central insight is that a single positive-definite function $V$ , monitored along trajectories via the orbital derivative $\dot{V}$ , encodes everything: positive definiteness makes sub-level sets shrink to the equilibrium, non-positivity of $\dot{V}$ makes them forward-invariant, and the two conditions together trap nearby trajectories near the rest state.

This is exactly the structure that appears again in the linearization theorem, where the Lyapunov equation $A^{⊤} P + P A = - I$ produces a canonical quadratic Lyapunov function for any Hurwitz linearization, identifying the spectral condition with the variational condition: eigenvalues of $A$ in the open left half-plane is equivalent to existence of a positive-definite quadratic form whose orbital derivative is negative definite. Putting these together, the direct method generalises to LaSalle's invariance principle when $\dot{V}$ is only semidefinite, to Chetaev's theorem for instability, to Krasovskii's converse theorem certifying that a Lyapunov function exists whenever asymptotic stability holds, to centre-manifold reduction at non-hyperbolic equilibria, to input-to-state stability for control systems, and to the Khasminskii formalism for stochastic differential equations. The bridge between 05.00.04 (Noether's theorem) and the stability viewpoint here is precise: a Noether conserved quantity is exactly a Lyapunov function with $\dot{V} \equiv 0$ , the boundary case between strict and non-strict, which is why energy methods in classical mechanics produce Lyapunov stability — friction is what tips the boundary case into strict negativity and recovers asymptotic convergence.

The framework identifies several stability notions that look distinct at first inspection. The geometric notion (trajectories stay near the equilibrium) is identified with the analytic notion (a positive-definite function decreases along the flow). The spectral notion (eigenvalues in the open left half-plane) is identified with the variational notion (the Lyapunov equation has a positive-definite solution). The topological notion (the Hartman-Grobman conjugacy) is identified with the analytic notion at hyperbolic equilibria. The control-theoretic notion (input-to-state stability) is the direct method augmented with a comparison condition on the input norm. The bridge is that all of these are different presentations of the same picture: a positive-definite function on phase space whose level sets trap trajectories. The Lyapunov direct method is dual to itself in a precise sense: the existence of a strict Lyapunov function and the asymptotic stability of the equilibrium are equivalent for smooth systems (Krasovskii's converse), so the direct method is a complete characterisation, not merely a sufficient condition.

Full proof set [Master]

Theorem (Lyapunov stability), proof. Given in the Intermediate-tier section: the compact-boundary minimum $m_{ε} = min_{S_{ε}} V > 0$ produces the trapping sub-level set $Ω_{ε} = {V < m_{ε}} \cap B_{ε} (x_{0})$ ; the sign condition $\dot{V} \leq 0$ makes $Ω_{ε}$ forward-invariant; continuity of $V$ at $x_{0}$ with $V (x_{0}) = 0$ produces $δ > 0$ with $B_{δ} (x_{0}) \subseteq Ω_{ε}$ ; together this gives the $ε$ - $δ$ statement of Lyapunov stability. For the strict case, monotonicity of $V$ along trajectories plus the compactness argument for $- \dot{V} \geq μ > 0$ on annular sub-level shells forces $V (φ_{t} (x)) \to 0$ , which by continuity of $V$ at $x_{0}$ forces $φ_{t} (x) \to x_{0}$ . $□$

Theorem (Lyapunov linearization, asymptotic case), proof. Translate the equilibrium to the origin and write $f (x) = A x + r (x)$ with $A = D f (0)$ Hurwitz and $r (x) = o (∥ x ∥)$ as $x \to 0$ . Since $A$ is Hurwitz, there exist constants $C, α > 0$ with $∥ e^{A t} ∥ \leq C e^{- α t}$ for every $t \geq 0$ (this is the operator-norm form of the spectral-radius bound, with the Jordan-block polynomial growth absorbed into the constants). Define $$ P = \int_0^\infty e^{A^\top t} e^{A t} , dt. $$ Convergence: $∥ e^{A^{⊤} t} e^{A t} ∥ \leq C^{2} e^{- 2 α t}$ , integrable on $[0, \infty)$ . Symmetry: $P^{⊤} = P$ from the integrand. Positive-definiteness: $v^{⊤} P v = \int_{0}^{\infty} ∥ e^{A t} v ∥^{2} d t > 0$ for $v \neq = 0$ . Lyapunov equation: from $\frac{d}{d t} (e^{A^{⊤} t} e^{A t}) = A^{⊤} e^{A^{⊤} t} e^{A t} + e^{A^{⊤} t} e^{A t} A$ and the fundamental theorem of calculus, $\int_{0}^{\infty} \frac{d}{d t} (e^{A^{⊤} t} e^{A t}) d t = 0 - I = - I$ (the limit at infinity vanishes by the decay bound), which gives $A^{⊤} P + P A = - I$ .

Define $V (x) = x^{⊤} P x$ . The function is $C^{\infty}$ , positive-definite, and $V (0) = 0$ . Compute the orbital derivative along $\overset{x}{˙} = A x + r (x)$ : $$ \dot V(x) = x^\top P f(x) + f(x)^\top P x = x^\top (A^\top P + P A) x + 2 x^\top P , r(x) = -|x|^2 + 2 x^\top P , r(x). $$ Bound the perturbation term: $∣2 x^{⊤} P r (x) ∣ \leq 2∥ P ∥_{op} ∥ x ∥∥ r (x) ∥ \leq 2∥ P ∥_{op} ∥ x ∥ \cdot ε (∥ x ∥) ∥ x ∥$ , where $ε (s) \to 0$ as $s \to 0$ from $r (x) = o (∥ x ∥)$ . Choose $ρ > 0$ small enough that $ε (∥ x ∥) \leq 1/ (4∥ P ∥_{op})$ for $∥ x ∥ \leq ρ$ . Then on the ball $∥ x ∥ \leq ρ$ , $$ \dot V(x) \leq -|x|^2 + 2 |P|{\mathrm{op}} \cdot \frac{1}{4 |P|{\mathrm{op}}} |x|^2 = -\tfrac{1}{2} |x|^2 < 0 $$ for $x \neq = 0$ . So $V$ is a strict Lyapunov function on the small ball, and the Lyapunov direct-method theorem proves asymptotic stability of the origin. $□$

Theorem (Lyapunov linearization, instability case), proof sketch. Assume $A = D f (0)$ has at least one eigenvalue $λ$ with $Re λ > 0$ . Decompose $R^{n} = E^{u} \oplus E^{cs}$ into the unstable eigenspace $E^{u}$ (positive-real-part eigenvalues) and the centre-stable complement $E^{cs}$ . On $E^{u}$ , the matrix $A ∣_{E^{u}}$ is anti-Hurwitz (all eigenvalues in the open right half-plane), and the time-reversed Lyapunov equation $A ∣_{E^{u}}^{⊤} Q + Q A ∣_{E^{u}} = + I$ has a positive-definite solution $Q$ . The function $W (x) = x_{u}^{⊤} Q x_{u} - ∥ x_{cs} ∥^{2}$ (with $x = x_{u} + x_{cs}$ in the splitting) satisfies $W (0) = 0$ , $W > 0$ on the cone $Ω = {x_{u}^{⊤} Q x_{u} > ∥ x_{cs} ∥^{2}}$ , and $\dot{W} > 0$ on $Ω$ near the origin. Chetaev's theorem (Exercise 8) then forces instability. $□$

Theorem (LaSalle's invariance principle), proof. Let $V$ be the Lyapunov function and $E = {\dot{V} = 0}$ . For $x_{1}$ in a bounded sub-level set ${V \leq c} \subseteq U$ , the forward orbit $γ^{+} (x_{1}) = {φ_{t} (x_{1}) : t \geq 0}$ is bounded, so it has a non-empty omega-limit set $Ω (x_{1})$ . The function $V$ is non-increasing along $γ^{+} (x_{1})$ and bounded below, so $V (φ_{t} (x_{1})) \to V_{\infty}$ for some $V_{\infty} \geq 0$ . By continuity, $V \equiv V_{\infty}$ on $Ω (x_{1})$ . Now $Ω (x_{1})$ is forward-invariant (a standard property of omega-limit sets), so $\frac{d}{d t} V \equiv 0$ along trajectories in $Ω (x_{1})$ , that is, $\dot{V} \equiv 0$ on $Ω (x_{1})$ . So $Ω (x_{1}) \subseteq E$ and $Ω (x_{1}) \subseteq M$ (the largest forward-invariant subset of $E$ ). Every trajectory in the sub-level set converges to $M$ . $□$

Theorem (Hartman-Grobman), stated without proof — full argument in Hartman 1960 ^{[Hartman 1960 *PAMS* 11]} and Pugh 1969 American J. Math. 91. The proof constructs the conjugating homeomorphism via a fixed-point argument on a Banach space of continuous maps, exploiting the hyperbolic splitting to obtain a contraction. Topological — not smooth — conjugacy is the best one can achieve in general; the Sternberg theorem (1957) provides smooth conjugacy under a resonance-free hypothesis on the eigenvalues.

Theorem (Krasovskii's converse), proof sketch. Translate the equilibrium to the origin and let $Ω$ be the domain of attraction. For $x \in Ω$ , define $V (x) = \int_{0}^{\infty} ψ (∥ φ_{t} (x) ∥) d t$ , where $ψ : [0, \infty) \to [0, \infty)$ is a smooth bump function with $ψ (0) = 0$ , $ψ > 0$ off $0$ , and $ψ$ decaying fast enough that the integral converges (using exponential decay of trajectories near the origin). Then $V (0) = 0$ , $V > 0$ off the origin, and $\dot{V} (x) = - ψ (∥ x ∥) < 0$ off the origin by the fundamental theorem of calculus applied to $V (φ_{t} (x)) - V (x) = - \int_{0}^{t} ψ (∥ φ_{s} (x) ∥) d s$ . The decay rate of $ψ$ governs the smoothness class of $V$ ; sharp converse theorems (Massera, Kurzweil) achieve $C^{\infty}$ smoothness. $□$

Theorem (Chetaev's instability theorem), proof. Given in Exercise 8 above. The key step is the growth bound $W (φ_{t} (x_{1})) \geq W (x_{1}) + μ t$ inside the cone where $\dot{W} \geq μ > 0$ , which forces the trajectory out of every compact sub-region of the cone and out through the sphere $∥ x ∥ = ε$ . $□$

Connections [Master]

Vector field on phase space 02.12.01. The Lyapunov stability theorem is a statement about an equilibrium of a vector field $f$ on a phase space $U$ ; without the vector-field framing there is no orbital derivative $\dot{V} = \nabla V \cdot f$ and no notion of an equilibrium to study. The phase-space picture is the setting in which the entire stability analysis takes place.
Phase flow / one-parameter group $g^{t}$ 02.12.02. The Lyapunov theorem is a statement about the forward flow $φ_{t}$ generated by the vector field: the sub-level set $Ω_{ε}$ is forward-invariant under $φ_{t}$ , and the asymptotic-stability conclusion is convergence under $φ_{t}$ as $t \to \infty$ . The flow is the structure that turns the sign condition on $\dot{V}$ into the time-dependent inequality $V (φ_{t} (x)) \leq V (x)$ used throughout the proof.
Noether's theorem 05.00.04. Noether identifies conserved quantities with continuous symmetries of a Lagrangian. A conserved quantity is exactly a Lyapunov function with $\dot{V} \equiv 0$ — the boundary case between strict and non-strict in Lyapunov's theorem. This is why energy methods in classical mechanics produce Lyapunov stability (the boundary case) and why dissipation — friction, viscosity, resistance — is what tips the boundary case into strict negativity and recovers asymptotic stability. The pendulum-with-friction worked example is the cleanest demonstration of the bridge.
Implicit and inverse function theorems 02.05.04. The linearization theorem reduces the nonlinear stability question to the linear one via the Taylor expansion $f (x) = A (x - x_{0}) + r (x)$ with $r (x) = o (∥ x - x_{0} ∥)$ . The local analysis of this expansion — the bounding of the perturbation term $r$ by an $ε$ times a power of $∥ x - x_{0} ∥$ — is exactly the technique of the implicit-and-inverse function machinery, and the centre-manifold theorem at non-hyperbolic equilibria depends on a Banach-space inverse function theorem.
Normed vector space 02.11.06. The operator-norm bound $∥ e^{A t} ∥ \leq C e^{- α t}$ used in the Lyapunov-equation proof — and in every quantitative statement about Hurwitz matrices — is a norm-theoretic statement on the Banach algebra $R^{n \times n}$ with the operator norm inherited from any norm on $R^{n}$ . The whole quantitative theory of asymptotic stability sits on this normed-space foundation.
Limit cycle and Liénard / Van der Pol systems 02.12.14. The Hopf bifurcation criterion — a planar equilibrium loses stability as a complex-conjugate eigenvalue pair of the linearization crosses the imaginary axis, and a limit cycle of amplitude $μ - μ_{0}$ emerges from the destabilised spiral focus — is a direct application of Lyapunov's linearization theorem at the bifurcation point. Below the bifurcation, the linearization has eigenvalues with negative real part and the present unit's linearization theorem certifies asymptotic stability of the equilibrium; at the bifurcation, the eigenvalues are pure imaginary and the linearization theorem is silent (the boundary case where higher-order terms decide); above the bifurcation, the eigenvalues have positive real part and the equilibrium is Lyapunov-unstable, with a limit cycle absorbing the destabilised orbits. The successor unit develops the Van der Pol equation as the canonical supercritical Hopf bifurcation at $μ = 0$ , using the Liénard energy $H = G (x) + y^{2} /2$ as a non-strict Lyapunov-style function whose level-set trapping picks out the limit cycle as the unique attractor. Connection type: bridging-theorem — Lyapunov's linearization controls the equilibrium side of the bifurcation, and the Liénard energy is the limit-cycle analogue of a Lyapunov function for the periodic orbit that emerges.
Receptor tyrosine kinases and the MAPK signaling cascade 17.07.02. Multi-stable systems appear in biology — see the MAPK cascade with positive feedback (17.07.02), in which stacked Hill-function ultrasensitivity (Huang-Ferrell) combined with Raf-Ras or ERK-Raf positive-feedback loops produces a saddle-node bifurcation creating coexisting active / inactive fixed points. Stability of each cascade fixed point is analysed via the present unit's linearization theorem on the kinase-cascade ODEs, and the strong-feedback limit admits a Lyapunov-like potential function whose level sets organise the bistable phase portrait. Cross-domain instance: the same direct-method machinery developed here for mathematical ODEs reappears as the qualitative-dynamics toolkit of systems biology.
Bifurcation theory pointer 02.12.17. Bifurcation theory is the perturbation theory of the failure of the Lyapunov linearization criterion. A bifurcation occurs precisely when the spectrum of the linearization $D f (x_{0})$ touches the imaginary axis — exactly the boundary case where the linearization theorem of the present unit is silent. The codim-one bifurcations are organised by which eigenvalue crosses: one real eigenvalue through zero gives the saddle-node, transcritical, and pitchfork normal forms; one complex-conjugate pair through the imaginary axis gives the Hopf normal form; and the higher-codimension bifurcations involve simultaneous or nested crossings. The Lyapunov-stability machinery survives the bifurcation by passing to the centre manifold: on the centre manifold the linearization is silent, but the Lyapunov direct-method's higher-order argument continues to apply via the first Lyapunov coefficient $ℓ_{1}$ , which controls supercritical versus subcritical Hopf and is the structural successor of Lyapunov's positive-definite-function technique. Connection type: perturbation-theoretic complement — bifurcation theory studies exactly the parameter values at which Lyapunov stability fails, with higher-order Lyapunov-style analysis taking over on the centre manifold.

Historical & philosophical context [Master]

Aleksandr Lyapunov stated and proved the direct method in his 1892 Kharkov PhD thesis Obshchaya zadacha ob ustoichivosti dvizheniya (The General Problem of the Stability of Motion) ^{[Lyapunov 1892]}. The thesis introduced both methods that bear his name: the direct method, in which a single auxiliary function plus a sign condition on its orbital derivative certifies stability without integrating the equation, and the first method (now subsumed by what is called the linearization theorem), in which the spectrum of the Jacobian at an equilibrium controls the local stability for hyperbolic equilibria. Lyapunov's thesis received limited attention outside Russia for several decades; the 1907 French translation by Davaux in the Annales de la Faculté des Sciences de Toulouse made the work accessible in Western Europe, and the 1992 English translation in Taylor & Francis brought it into wider circulation. The thesis itself is striking in its modern flavour: the precise $ε$ - $δ$ definition of stability, the proof of the linearization theorem via the Lyapunov equation, and the introduction of what is now called the Lyapunov function are all present, with arguments very close to modern textbook treatments.

The mid-twentieth century saw the systematic absorption of Lyapunov's direct method into control theory and nonlinear dynamics. Nikolay Krasovskii's 1959 monograph Stability of Motion (English translation Stanford University Press 1963) ^{[Krasovskii 1959]} consolidated the Soviet school's work on converse theorems, time-varying systems, and applications to control. José Massera's 1956 paper Contributions to stability theory (Annals of Mathematics 64) ^{[Massera 1956]} established the smooth converse theorem for asymptotically stable equilibria, and Joseph LaSalle's 1960 paper Some extensions of Liapunov's second method (IRE Transactions on Circuit Theory CT-7) ^{[LaSalle 1960 *IRE CT-7*]} introduced the invariance principle that recovers asymptotic stability from semidefinite orbital derivatives. Wolfgang Hahn's 1967 Stability of Motion (Springer Grundlehren 138) and Hassan Khalil's Nonlinear Systems (3rd edition 2002) ^{[Khalil Ch. 4]} are the canonical modern textbook syntheses.

Philip Hartman's 1960 paper A lemma in the theory of structural stability of differential equations (Proceedings of the AMS 11) ^{[Hartman 1960 *PAMS* 11]} established the topological-conjugacy theorem now called Hartman-Grobman (independently discovered by David Grobman in 1959 in the Soviet literature), strengthening the linearization picture at hyperbolic equilibria. Eduard Sontag's 1989 paper Smooth stabilization implies coprime factorization (IEEE Transactions on Automatic Control 34) ^{[Sontag 1989 *IEEE TAC 34*]} introduced input-to-state stability, the modern generalisation of Lyapunov's direct method to control systems with disturbances; ISS-Lyapunov functions are now a standard tool in nonlinear control. Rafail Khasminskii's 1969 monograph Stochastic Stability of Differential Equations ^{[Khasminskii 1969]} extended the direct method to Itô stochastic differential equations, replacing the orbital derivative with the infinitesimal generator. Throughout, the underlying recipe — pick a positive-definite function, check the sign of its orbital derivative or generator — has remained Lyapunov's.

Bibliography [Master]

@phdthesis{Lyapunov1892,
  author = {Lyapunov, Aleksandr M.},
  title  = {The General Problem of the Stability of Motion},
  school = {Kharkov University},
  year   = {1892},
  note   = {Russian original; French translation by Davaux, Annales de la Facult\'e des Sciences de Toulouse, 1907; English translation by A. T. Fuller, International Journal of Control 55, 531--773 (1992); reprinted Taylor \& Francis, 1992}
}

@article{Hartman1960,
  author  = {Hartman, Philip},
  title   = {A lemma in the theory of structural stability of differential equations},
  journal = {Proceedings of the American Mathematical Society},
  volume  = {11},
  year    = {1960},
  pages   = {610--620}
}

@article{Grobman1959,
  author  = {Grobman, David M.},
  title   = {Homeomorphism of systems of differential equations},
  journal = {Doklady Akademii Nauk SSSR},
  volume  = {128},
  year    = {1959},
  pages   = {880--881}
}

@article{LaSalle1960,
  author  = {LaSalle, Joseph P.},
  title   = {Some extensions of Liapunov's second method},
  journal = {IRE Transactions on Circuit Theory},
  volume  = {CT-7},
  year    = {1960},
  pages   = {520--527}
}

@article{Massera1956,
  author  = {Massera, Jos\'e L.},
  title   = {Contributions to stability theory},
  journal = {Annals of Mathematics},
  volume  = {64},
  year    = {1956},
  pages   = {182--206}
}

@book{Krasovskii1959,
  author    = {Krasovskii, Nikolay N.},
  title     = {Stability of Motion},
  publisher = {Stanford University Press},
  year      = {1963},
  note      = {English translation of the 1959 Russian original}
}

@book{Hahn1967,
  author    = {Hahn, Wolfgang},
  title     = {Stability of Motion},
  publisher = {Springer-Verlag},
  series    = {Grundlehren der mathematischen Wissenschaften},
  volume    = {138},
  year      = {1967}
}

@book{Khalil2002,
  author    = {Khalil, Hassan K.},
  title     = {Nonlinear Systems},
  publisher = {Prentice-Hall},
  edition   = {3},
  year      = {2002}
}

@book{HirschSmaleDevaney2013,
  author    = {Hirsch, Morris W. and Smale, Stephen and Devaney, Robert L.},
  title     = {Differential Equations, Dynamical Systems, and an Introduction to Chaos},
  publisher = {Academic Press},
  edition   = {3},
  year      = {2013}
}

@book{HartmanODE,
  author    = {Hartman, Philip},
  title     = {Ordinary Differential Equations},
  publisher = {SIAM Classics in Applied Mathematics},
  volume    = {38},
  year      = {2002},
  note      = {Reprint of the 1964 Wiley original}
}

@book{ArnoldODE,
  author    = {Arnold, Vladimir I.},
  title     = {Ordinary Differential Equations},
  publisher = {MIT Press},
  year      = {1973},
  note      = {Translated from the Russian; reissued by Springer in 1992}
}

@article{Sontag1989,
  author  = {Sontag, Eduard D.},
  title   = {Smooth stabilization implies coprime factorization},
  journal = {IEEE Transactions on Automatic Control},
  volume  = {34},
  year    = {1989},
  pages   = {435--443}
}

@book{Khasminskii1969,
  author    = {Khasminskii, Rafail Z.},
  title     = {Stochastic Stability of Differential Equations},
  publisher = {Sijthoff and Noordhoff},
  year      = {1980},
  note      = {English translation of the 1969 Russian original; second edition Springer 2012}
}

@article{Sternberg1957,
  author  = {Sternberg, Shlomo},
  title   = {Local contractions and a theorem of Poincar\'e},
  journal = {American Journal of Mathematics},
  volume  = {79},
  year    = {1957},
  pages   = {809--824}
}

@article{Pugh1969,
  author  = {Pugh, Charles C.},
  title   = {On a theorem of P. Hartman},
  journal = {American Journal of Mathematics},
  volume  = {91},
  year    = {1969},
  pages   = {363--367}
}

Prerequisites

02.05.04
02.11.06
05.00.04

Tier anchors

beginner: Strogatz *Nonlinear Dynamics and Chaos* Ch. 6 §6.5 (informal energy-function picture for the pendulum and similar mechanical examples)
intermediate: Arnold *Ordinary Differential Equations* Ch. 5 §22-§23; Hirsch-Smale-Devaney *Differential Equations, Dynamical Systems, and an Introduction to Chaos* Ch. 9; Khalil *Nonlinear Systems* Ch. 4
master: Lyapunov 1892 *The General Problem of the Stability of Motion* (Kharkov PhD thesis; English translation, Taylor & Francis, 1992) — originator; LaSalle 1960 *Some extensions of Liapunov's second method* (IRE Trans. Circuit Theory CT-7) — invariance principle; Krasovskii 1959 *Stability of Motion* (Stanford University Press, English translation 1963); Hahn 1967 *Stability of Motion*; Hartman *Ordinary Differential Equations* Ch. 9; Arnold *Ordinary Differential Equations* Ch. 5 §23

References

TODO_REF
Lyapunov 1892 — The General Problem of the Stability of Motion · Kharkov PhD thesis; English translation A. T. Fuller, ed., International Journal of Control 55 (1992), and Taylor & Francis monograph 1992; originator of the direct method, the linearization theorem with negative real-part eigenvalues, and the Lyapunov-equation construction of a quadratic Lyapunov function
TODO_REF
Arnold — Ordinary Differential Equations · Ch. 5 §22 (stability of equilibria), §23 (Lyapunov's theorem and the linearization theorem); MIT Press 1973 / Springer 1992 editions
TODO_REF
Hirsch-Smale-Devaney — Differential Equations, Dynamical Systems, and an Introduction to Chaos · Ch. 9 stability of equilibria; Academic Press, 3rd edition 2013
TODO_REF
Khalil — Nonlinear Systems · Ch. 4 (Lyapunov stability), Theorem 4.1 (stability theorem), Theorem 4.2 (asymptotic stability), §4.3 (linearization); Prentice-Hall, 3rd edition 2002
TODO_REF
LaSalle 1960 — Some extensions of Liapunov's second method · IRE Transactions on Circuit Theory CT-7, 520-527; originator of the invariance principle that recovers asymptotic stability when $\dot V$ is only semidefinite
TODO_REF
Krasovskii 1959 — Stability of Motion · Stanford University Press 1963 English translation; converse Lyapunov theorems and the systematic exploitation of the direct method for nonlinear control
TODO_REF
Hahn 1967 — Stability of Motion · Springer Grundlehren der mathematischen Wissenschaften 138; canonical textbook synthesis of Lyapunov-direct-method theory with attention to converse theorems and time-varying systems
TODO_REF
Hartman — Ordinary Differential Equations · Ch. 9 stability of equilibria, including the Hartman-Grobman theorem (Ch. 9 Theorem 7.1); SIAM Classics in Applied Mathematics 38, 2002 reprint of the 1964 Wiley original
TODO_REF
Hartman 1960 — A lemma in the theory of structural stability of differential equations · Proceedings of the American Mathematical Society 11, 610-620; originator of the Hartman-Grobman theorem in the form used today

Reviewer

TBD

Estimated time

beginner: 16m
intermediate: 38m
master: 75m