Fixed Points and Equilibria

Every dynamical system, however wild, is organized around its places of perfect rest. A marble motionless at the bottom of a bowl, a rocket balanced on its exhaust plume, a population holding steady at the carrying capacity of its habitat — each is a fixed point: a state where the velocity field of the dynamics vanishes and the system, placed there exactly, stays forever. They are the only solutions you can write down for a general nonlinear system without integrating anything, and yet they carry astonishing authority: the fate of every nearby motion, and often of every motion anywhere, is scripted by the handful of points where nothing happens. The physics lives in the second question: what does the flow do near a fixed point? A hanging pendulum and a balanced inverted pendulum are both equilibria — both have zero velocity and zero acceleration — but a millimetre of displacement dooms one to gentle oscillation and the other to catastrophic escape. Making that distinction precise turns out to require real care. "Stable" splits into inequivalent notions — staying nearby, versus actually returning — and the workhorse tool of linearization comes with fine print exactly where the most interesting physics (bifurcations) happens. This article builds the general theory the nonlinear-dynamics way: first on the line, where flows $\dot{x}=f(x)$ can be read off a single graph and oscillation is provably impossible; then in the plane, where basins of attraction appear. The conservative special case — equilibria of potentials, centres and saddles, the Lagrange–Dirichlet theorem — has its own article, Fixed Points and Stability in Phase Space; here it enters only as one province of a much larger empire that includes everything that rusts, radiates, and drains.

Key takeaways

A fixed point of $\dot{\mathbf{x}}=\mathbf{f}(\mathbf{x})$ is a state $\mathbf{x}^*$ with $\mathbf{f}(\mathbf{x}^*)=\mathbf{0}$ : every time derivative vanishes, and the constant function $\mathbf{x}(t)=\mathbf{x}^*$ is an exact solution.
On the line, the graph of $f$ is the whole theory: zeros are equilibria, arrows point right where $f>0$ and left where $f<0$ , stability is read from the slope $f'(x^*)$ , and no one-dimensional flow can oscillate.
Stability comes in inequivalent grades — Lyapunov stable (nearby states stay nearby), asymptotically stable (they also converge), neutrally stable, unstable — and conservative mechanics only ever achieves the neutral grade.
Linearization $\dot{\boldsymbol{\eta}}=D\mathbf{f}(\mathbf{x}^*)\boldsymbol{\eta}$ decides the local picture whenever the fixed point is hyperbolic (no eigenvalue with zero real part); in the marginal cases the nonlinear terms rule, and that is where bifurcations are born.
Attracting fixed points own basins of attraction — the sets of initial conditions they capture — and the plane is parcelled among attractors along boundaries formed by saddle stable manifolds.
Key idea: A fixed point is where the velocity field of the dynamical system vanishes, $\mathbf{f}(\mathbf{x}^*)=0$ ; the system frozen there stays there, and the eigenvalues of the local Jacobian $D\mathbf{f}(\mathbf{x}^*)$ tell you everything about how nearby states behave.

Where this lives today

Reusable rockets land by making an unstable equilibrium behave like a stable one. A booster descending on its exhaust is an inverted pendulum on a thrust vector — a saddle whose unstable eigenvalue is set by the vehicle's length scale, exactly as in Worked Example 2 — and the guidance computer's job is to close a feedback loop far inside the doubling time. The stability certificates for such controllers are Lyapunov functions, now synthesized numerically by sum-of-squares optimization: Derivation C's energy argument, industrialized. The same mathematics balances Segways, humanoid robots, and the vertical position of tokamak plasmas, which sit at a saddle of the magnetic field energy with growth times of milliseconds. The marginal case earns its keep in the Earth system. As a slowly forced system approaches a bifurcation, its leading eigenvalue drifts toward zero, so recovery from natural kicks slows and their variance swells — critical slowing down, the observable shadow of $f'(x^*)\to0$ in Derivation A. Climate scientists mine satellite records of the Atlantic overturning circulation, ice-sheet flow, and Amazon vegetation for exactly these fingerprints, publishing early-warning indicators of tipping points as routinely as weather anomalies. When those papers speak of an equilibrium "losing resilience," they mean this article's relaxation time $1/|f'|$ stretching toward infinity — the quiet before a fixed point disappears.

The physical picture

Begin one dimension down from the phase plane, where the entire subject fits on a single graph. A flow on the line, $\dot{x}=f(x)$ , is a rule pushing a bead along a wire: where $f>0$ the bead slides right, where $f<0$ it slides left, and at the zeros of $f$ it rests. Plot $f$ and the dynamics is over: mark the zeros, draw arrows toward or away from each according to the sign of $f$ on either side, and you possess every solution's fate without integrating. A zero with $f$ falling through it (positive to the left, negative to the right) traps its neighbours — stable. A zero with $f$ rising through it repels them — unstable.

Figure

A flow on the line — the graph of f is the complete theory: arrows follow the sign of f, zeros with falling slope attract, zeros with rising slope repel

Two morals from the line survive every generalization. First, equilibria are found by algebra (solve $f=0$ ), their character by calculus (differentiate $f$ there): this division of labour is the whole method, in any dimension. Second, dimensionality is destiny: a bead on the line can only slide monotonically to a fixed point or to infinity — it can never oscillate, because turning around would require passing through a state twice in different directions, and one state has one future. Oscillation needs a second dimension; chaos will need a third. In the plane, rest points acquire richer courts. Around some, the flow circulates and decays inward — attractors, each ruling a basin of initial conditions that fall to it. Others repel everything, or attract along one special curve while repelling along another (saddles). The new global object is the basin boundary: the watershed, typically woven from saddle stable manifolds, that decides which attractor claims which history.

Figure

The grades of stability — nearby states may stay nearby, converge back, or escape; conservative mechanics tops out at the neutral middle grade

Beware the everyday word "equilibrium": in mechanics it often means only a point of zero force. In phase space the standard is stricter — every state variable must have zero rate. A pendulum at the top of its swing has zero velocity but nonzero acceleration; its phase point is still moving (vertically, in the $(x,\dot x)$ plane) and will be somewhere else in an instant. Rest, in this subject, means rest of the state, not of the body alone.

The core equations

The definition is one line, and it is worth reading actively:

\dot{\mathbf{x}}=\mathbf{f}(\mathbf{x}), \qquad \mathbf{f}(\mathbf{x}^*)=\mathbf{0} \qquad (1)

where: $\mathbf{x}^*$ — a fixed point (equilibrium, stationary point, critical point — the literature uses all four); the constant function $\mathbf{x}(t)\equiv\mathbf{x}^*$ solves the system exactly. Finding fixed points is algebra, not analysis: set every rate to zero simultaneously and solve. In the plane this means intersecting the nullclines of the article Phase Portraits and the Phase Plane. On the line, stability has a one-number answer. Perturb $x=x^*+\eta$ and keep the leading term:

\dot{\eta}=f'(x^*)\,\eta+\mathcal{O}(\eta^2) \;\Longrightarrow\; \eta(t)\approx\eta(0)\,e^{f'(x^*)t} \qquad (2)

where: $f'(x^*)$ — the slope of $f$ at the zero. Negative slope: perturbations decay with characteristic time $1/|f'(x^*)|$ — the single most used number in applied stability analysis. Positive slope: exponential escape with doubling time $\ln 2/f'(x^*)$ . Zero slope: the linear term is silent and the verdict passes to higher-order terms (Derivation A). Mechanics supplies the canonical two-dimensional family. A particle in a potential, written in phase form,

\dot{q}=\frac{p}{m}, \qquad \dot{p}=-V'(q) \qquad (3)

where: setting both rates to zero forces $p=0$ and $V'(q^*)=0$ : mechanical equilibria sit on the zero-velocity axis at the stationary points of the potential. Minima and maxima of $V$ give the centres and saddles whose complete conservative census — including the Lagrange–Dirichlet stability theorem — is the business of the companion article Fixed Points and Stability in Phase Space. Keep from it one export:

\omega=\sqrt{\frac{V''(q^*)}{m}} \qquad (4)

where: $\omega$ — the small-oscillation frequency about a potential minimum, obtained by linearizing the force: $m\ddot{u}\approx-V''(q^*)u$ for $u=q-q^*$ . If $V''(q^*)<0$ the same algebra returns a growth rate $\lambda=\sqrt{-V''(q^*)/m}$ instead — the inverted pendulum's ticking clock, computed numerically in Worked Example 2. In any dimension, the local analysis runs through the Jacobian. Displace $\mathbf{x}=\mathbf{x}^*+\boldsymbol{\eta}$ and Taylor-expand the field:

\dot{\boldsymbol{\eta}} = D\mathbf{f}(\mathbf{x}^*)\,\boldsymbol{\eta}+\mathcal{O}(|\boldsymbol{\eta}|^2), \qquad \big(D\mathbf{f}\big)_{ij}=\frac{\partial f_i}{\partial x_j} \qquad (5)

where: $D\mathbf{f}(\mathbf{x}^*)$ — the Jacobian matrix at the fixed point; the zeroth-order term died by the definition of $\mathbf{x}^*$ , so the leading behaviour of small displacements is a constant-coefficient linear system. Its eigenvalues $\lambda$ are growth rates; eigenvectors are the special directions along which growth is pure. When every eigenvalue has nonzero real part the fixed point is called hyperbolic, and the Hartman–Grobman theorem promises the honest nonlinear flow is a continuous distortion of the linear picture — the full statement and its abuse cases belong to the next article, Linear Stability Analysis and the Jacobian. In two dimensions the eigenvalues come from a quadratic you will use hundreds of times:

\lambda^2-\tau\lambda+\Delta=0, \qquad \lambda_\pm=\tfrac{1}{2}\left(\tau\pm\sqrt{\tau^2-4\Delta}\right) \qquad (6)

where: $\tau=\operatorname{tr}D\mathbf{f}$ , $\Delta=\det D\mathbf{f}$ — two numbers that between them sort planar equilibria into saddles ( $\Delta<0$ ), nodes, spirals and centres. The full map of that $(\tau,\Delta)$ plane is the next-but-one article, Classification of Fixed Points: Nodes, Saddles, Spirals, Centers; a first taste: $\tau=-3$ , $\Delta=2$ gives $\lambda=-1,-2$ — both negative, a stable node. Note also $\tau=\nabla\cdot\mathbf{f}$ : the trace doubles as the local phase-area growth rate, tying stability to the dissipation budget of the previous article. Fixed points are equally central to discrete dynamics, where they seed the route to chaos later in the chapter:

\mathbf{x}_{n+1}=\mathbf{g}(\mathbf{x}_n), \qquad \mathbf{x}^*=\mathbf{g}(\mathbf{x}^*) \qquad (7)

where: $\mathbf{g}$ — an iteration map, such as a Poincaré return map; the fixed point is a state the map sends to itself, and (in one dimension) it is stable when $|g'(x^*)|<1$ — the multiplier per step playing the role that $e^{f'}$ per unit time plays for flows.

You want	Reach for	Because
all equilibria	solve $\mathbf{f}(\mathbf{x})=\mathbf{0}$	algebra locates every candidate
stability on the line	sign of $f'(x^*)$ , eq. (2)	one derivative is the whole verdict
relaxation or escape time	$1/\|f'(x^*)\|$ or $1/\mathrm{Re}\,\lambda$	linear rates set the local clock
stability in the plane	$\tau,\ \Delta$ of the Jacobian, eq. (6)	two numbers classify the neighbourhood
verdict when linearization is silent	higher-order terms, Lyapunov functions	marginal cases are nonlinear property
stability of a periodic orbit	fixed point of the return map, eq. (7)	reduces a cycle to an equilibrium of a map

Derivation AStability from the slope and what happens when the slope is zero

We prove the one-dimensional stability criterion honestly and then push it to its breaking point. Goal: show that the sign of $f'(x^*)$ decides stability with an exponential rate, and that when $f'(x^*)=0$ no amount of linear thinking can decide — three systems with identical linearizations will be shown to have three different fates. The strategy is direct: expand the flow near the zero, solve the leading dynamics, and then treat the truncated terms as the leading dynamics in the marginal case. Watch how quickly "the linearization is silent" becomes "the geometry of $f$ decides."

Step 1. Let $x^*$ satisfy $f(x^*)=0$ and set $\eta=x-x^*$ . Taylor expansion gives $\dot{\eta}=f(x^*+\eta)=f'(x^*)\eta+\tfrac12 f''(x^*)\eta^2+\tfrac16 f'''(x^*)\eta^3+\cdots$ , with no constant term precisely because $x^*$ is a fixed point.

Step 2. Suppose $f'(x^*)\neq 0$ . For $|\eta|$ small enough the linear term dominates every other, and the comparison $\dot{\eta}=f'(x^*)\eta$ integrates to $\eta(t)=\eta(0)e^{f'(x^*)t}$ . Negative slope: every sufficiently small perturbation decays exponentially with time constant $1/|f'(x^*)|$ — asymptotic stability. Positive slope: perturbations grow until they leave the neighbourhood — instability. Why "sufficiently small" is not weasel wording: the quadratic term is smaller than the linear one whenever $|\eta|<|f'(x^*)/f''(x^*)|$ , so there is a definite trapping neighbourhood on which the linear verdict is rigorous, not heuristic.

Step 3. Now let $f'(x^*)=0$ . The expansion begins at the first non-vanishing derivative, and that term owns the dynamics. Take the three cleanest representatives at $x^*=0$ : $\dot{x}=-x^3$ , $\dot{x}=+x^3$ , and $\dot{x}=x^2$ . All three have the same linearization, $\dot{\eta}=0$ .

Step 4. Solve $\dot{x}=-x^3$ by separation: $x(t)=x(0)/\sqrt{1+2x(0)^2 t}$ . Every initial condition decays to zero — stable — but only algebraically, like $t^{-1/2}$ , with no characteristic time scale at all. Why this matters physically: marginal equilibria relax without a clock. In experiments, algebraic relaxation near a threshold is a fingerprint that a control parameter sits at a bifurcation point.

Step 5. Solve $\dot{x}=+x^3$ : $x(t)=x(0)/\sqrt{1-2x(0)^2t}$ , which diverges in finite time $t^*=1/(2x(0)^2)$ . Unstable — more violently than any exponential.

Step 6. Solve $\dot{x}=x^2$ : for $x(0)>0$ solutions blow up in finite time, while for $x(0)<0$ they relax to zero. The point is half-stable — attracting from one side, repelling from the other. This hybrid is not a curiosity: it is precisely the local picture at the moment a saddle-node bifurcation annihilates a stable–unstable pair, as the article on the standard bifurcations shows. Verified: differentiating each closed-form solution reproduces its ODE; all three share $f(0)=f'(0)=0$ , confirming that the linearization genuinely cannot distinguish them. Cross-check with eq. (2): for $\dot x=-kx$ with $k>0$ the same separation method returns $e^{-kt}$ , matching the exponential law.

Go deeper · Derivation B · 2 minWhy nothing can oscillate on a line

Derivation BWhy nothing can oscillate on a line

We establish the strongest structural theorem available this early in the chapter: a one-dimensional flow $\dot{x}=f(x)$ admits no periodic solutions whatsoever — every motion is monotone, marching to a fixed point or to infinity. The proof has two independent routes, and both teach a general-purpose weapon: the uniqueness/monotonicity argument, and the gradient (energy-function) argument that returns as the Lyapunov method in Derivation C. The result explains, at one stroke, why oscillators require a phase plane and why the overdamped limit of any mechanical system is dynamically so tame.

Step 1. Monotonicity route. Along a solution, $\dot{x}(t)=f(x(t))$ : the sign of the velocity at any instant is a function of position alone. Suppose a trajectory reverses direction at some time $t_1$ . Continuity forces $\dot x(t_1)=0$ , i.e. $f(x(t_1))=0$ — the reversal point is a fixed point.

Step 2. But the constant solution $x(t)\equiv x(t_1)$ also occupies that state, and the uniqueness theorem allows one solution per state. A trajectory reaching a fixed point in finite time and leaving it again would be a second solution through that state. Hence no moving trajectory ever reaches, or turns around at, a fixed point in finite time: it can only approach one asymptotically. Every solution is monotone in $t$ .

Step 3. A periodic solution must return to its starting value, which for a monotone function is impossible unless the solution is constant. Therefore the only recurrent motions on the line are the fixed points themselves — no oscillations, no cycles, full stop. Why the theorem dies in the plane: with two dimensions, "turning around" no longer requires revisiting a state — the trajectory swings around in a new direction. Oscillation is not forbidden, and the pendulum exists.

Step 4. Gradient route. Define $V(x)=-\int^x f(s)\,\mathrm{d}s$ , so the flow is $\dot{x}=-V'(x)$ : every one-dimensional system is an overdamped particle sliding on a potential landscape. Along any solution, $\frac{\mathrm{d}}{\mathrm{d}t}V(x(t))=V'(x)\dot{x}=-\big(\dot{x}\big)^2\le 0$ .

Step 5. $V$ decreases strictly whenever the state moves, so the motion can never revisit a state (which would require $V$ to return to a previous value). Fixed points are the stationary points of $V$ ; minima attract, maxima repel — the bead-on-a-graph picture of the hero figure, now derived rather than asserted. Verified: for $f=-kx$ , $V=\tfrac12 kx^2$ and the routes agree: monotone exponential decay into the minimum. The gradient argument is the $m\to 0$ limit of mechanics: $m\ddot q+\gamma\dot q=-U'(q)$ reduces to $\dot q=-U'(q)/\gamma$ , matching the overdamped row of the limiting-cases table.

Go deeper · Derivation C · 2 minEnergy as a Lyapunov function for the damped oscillator

Derivation CEnergy as a Lyapunov function for the damped oscillator

Linearization is local and sometimes inadmissible; the Lyapunov method is the global, nonlinear-honest alternative, and this derivation is its canonical first outing. Goal: prove — without solving anything and without linearizing — that the origin of the damped oscillator $\ddot{x}+\gamma\dot{x}+\omega_0^2x=0$ , $\gamma>0$ , attracts every initial condition. The strategy generalizes Derivation B's gradient argument: find a positive, bowl-shaped function of the state that the flow strictly drains. The physical energy is the natural candidate, and the one subtlety — the drain rate vanishes on a whole axis — introduces the invariance-principle reasoning used throughout modern stability theory.

Step 1. Phase form: $\dot{x}=v$ , $\dot{v}=-\gamma v-\omega_0^2 x$ . Candidate function: the mechanical energy per unit mass, $E(x,v)=\tfrac12 v^2+\tfrac12\omega_0^2x^2$ . It is positive everywhere except at the fixed point $(0,0)$ , where it vanishes — a bowl with its bottom at the equilibrium.

Step 2. Differentiate along the flow: $\dot{E}=v\dot{v}+\omega_0^2 x\dot{x}=v(-\gamma v-\omega_0^2x)+\omega_0^2xv=-\gamma v^2$ . The conservative terms cancel identically; only dissipation survives. Why this cancellation is not luck: energy is conserved by the undamped flow, so its derivative along the damped flow consists purely of the damping term — the same mechanism as the damped-pendulum computation in Phase Portraits and the Phase Plane.

Step 3. $\dot{E}\le 0$ everywhere: no trajectory can climb the bowl, so states starting inside the energy level set $E<E_0$ stay inside it forever. That alone certifies Lyapunov stability — with zero reference to eigenvalues.

Step 4. The drain vanishes on the axis $v=0$ (turning points), so strict decrease fails there and we must rule out motion hiding on that axis. Invariance argument: a trajectory remaining in $\{v=0\}$ must have $\dot{v}=0$ too, and the equation of motion then forces $\omega_0^2 x=0$ . The only complete trajectory on which $\dot{E}=0$ identically is the origin itself.

Step 5. Since $E$ decreases along every non-equilibrium arc and cannot stall on any invariant set except the origin, every trajectory drains to $E=0$ : the origin is asymptotically stable, globally — its basin is the entire plane. Verified: the linear check agrees where it applies — eigenvalues $\lambda=\tfrac12\left(-\gamma\pm\sqrt{\gamma^2-4\omega_0^2}\right)$ have negative real part for all $\gamma>0$ . But the Lyapunov argument proves more (a global basin) and survives nonlinear stiffening: replace $\omega_0^2 x$ by any restoring force $g(x)$ with $xg(x)>0$ and the same computation runs with $E=\tfrac12v^2+\int_0^x g$ , no re-derivation required.

Figure

Two attractors partition the plane — the damped double well's basins of attraction, with the saddle's stable manifold as the watershed between them

Limiting cases

Case	Result	Meaning
Hyperbolic fixed point	linearization decides	Hartman–Grobman: the nonlinear flow is a warped copy of $\dot{\boldsymbol{\eta}}=D\mathbf{f}\,\boldsymbol{\eta}$ near $\mathbf{x}^*$
Marginal case, $\mathrm{Re}\,\lambda=0$	nonlinear terms decide	the excluded case of the theorem — and the doorway through which every bifurcation of this chapter enters
Conservative limit	$\tau=0$ : centres and saddles only	neutral stability at best; no attractors, no basins — see Fixed Points and Stability in Phase Space
Overdamped limit, $m\to 0$	$\dot{q}=-V'(q)/\gamma$	dynamics collapses onto the line: potential minima become one-dimensional attractors, and Derivation B applies verbatim
Discrete map	stable iff all $\lvert\lambda_i\rvert<1$	the unit circle replaces the imaginary axis as the stability boundary

Common trapZero velocity at the turning point means the oscillator is momentarily at equilibrium

It does not — and this single confusion generates more wrong answers than any other in the chapter. At the turning point of an oscillation the velocity vanishes but the acceleration does not: in phase space the state sits on the $v=0$ axis with $\dot{v}=-\omega_0^2x\neq0$ , still in full flight, moving vertically. A fixed point requires every component of the state velocity to vanish simultaneously — $\dot{x}=0$ and $\dot{v}=0$ — which happens only at $x=0,\,v=0$ . The reliable test is always eq. (1) applied to the full state, never to the visible coordinate alone. Related trap: "the system passes through equilibrium" (the bottom of the swing) does not mean it is at a fixed point either — there $x=0$ but $\dot{x}\neq0$ . Rest of the state means rest of all of it.

Worked examples

Example 1 — the logistic population, complete analysis in four lines (confidence builder). A fishery's stock obeys $\dot{N}=rN(1-N/K)$ with growth rate $r=0.5\ \mathrm{yr^{-1}}$ and carrying capacity $K=1000$ tonnes. Find the equilibria, their stability, and the relevant time scales.

Step 1. Equilibria: $f(N)=rN(1-N/K)=0$ gives $N^*=0$ (extinction) and $N^*=K=1000$ (carrying capacity).

Step 2. Slope: $f'(N)=r(1-2N/K)$ .

Step 3. At extinction: $f'(0)=r=+0.5\ \mathrm{yr^{-1}}>0$ — unstable. A small introduced population grows exponentially, doubling every $\ln 2/0.5=1.39$ years.

Step 4. At capacity: $f'(K)=-r=-0.5\ \mathrm{yr^{-1}}<0$ — asymptotically stable, with relaxation time $1/0.5=2.0$ years: after a one-off harvest or a good season, the stock returns to $1000$ tonnes with deviations decaying by $e^{-1}$ every two years. Two equilibria: $N^*=0$ unstable (doubling 1.39 yr), $N^*=1000$ t stable (relaxation 2.0 yr) — and by Derivation B the approach is monotone: a one-dimensional stock can never overshoot and oscillate about $K$ .

Example 2 — the two faces of the pendulum, with numbers (exam level). A rigid pendulum has length $\ell=0.49$ m in $g=9.8\ \mathrm{m\,s^{-2}}$ (so $g/\ell=20.0\ \mathrm{s^{-2}}$ ). Classify both equilibria and extract every physical time scale; then state the control requirement for balancing it inverted.

Step 1. Phase form with $\theta$ measured from the downward vertical: $\dot{\theta}=\omega$ , $\dot{\omega}=-(g/\ell)\sin\theta$ . Fixed points at $\omega=0$ , $\sin\theta=0$ : hanging $(0,0)$ and inverted $(\pi,0)$ .

Step 2. Jacobian: $D\mathbf{f}=\begin{pmatrix}0&1\\-(g/\ell)\cos\theta^*&0\end{pmatrix}$ .

Step 3. Hanging state: $\cos 0=1$ , so $\lambda^2=-g/\ell=-20.0$ , giving $\lambda=\pm i\sqrt{20.0}=\pm 4.47i\ \mathrm{s^{-1}}$ . A centre: small oscillations at $\omega_0=4.47\ \mathrm{rad\,s^{-1}}$ , period $T=2\pi/4.47=1.41$ s. (Neutral stability only — add any damping and the trace goes negative, making it a true attractor.)

Step 4. Inverted state: $\cos\pi=-1$ , so $\lambda^2=+20.0$ and $\lambda=\pm4.47\ \mathrm{s^{-1}}$ — a saddle. The unstable eigenvalue sets the escape clock: perturbations $e$ -fold every $1/4.47=0.224$ s and double every $\ln2/4.47=0.155$ s.

Step 5. Control requirement: a stabilizing feedback (reaction wheel, moving cart) must sense and correct faster than the instability grows — loop latency comfortably below $0.15$ s, which is why balancing robots run their control loops at tens to hundreds of hertz. Hanging: centre, $\lambda=\pm4.47i\ \mathrm{s^{-1}}$ , period 1.41 s. Inverted: saddle, $\lambda=\pm4.47\ \mathrm{s^{-1}}$ , doubling time 0.155 s. The general- $\ell$ formulas $\lambda=\pm i\sqrt{g/\ell}$ and $\pm\sqrt{g/\ell}$ reproduce the conservative census that the companion Hamiltonian article derives from the potential landscape.

Origins · 1 minAleksandr Lyapunov and how this was discovered

Aleksandr Lyapunov

1857–1918 · St Petersburg school of Chebyshev; Kharkov and the Russian Academy of Sciences

Lyapunov gave stability its mathematics. Before his 1892 doctoral thesis, The General Problem of the Stability of Motion, arguments about equilibrium were a patchwork of linearizations and energy intuitions with no theory saying when they could be trusted. Lyapunov supplied the honest definitions — the $\varepsilon$ – $\delta$ grades of stability that still bear his name — and two methods: the "first method" justifying linearization through what we now call Lyapunov exponents, and the "second method" of certifying stability with a decreasing function of the state, no solution formulas required. The second method, a vast generalization of the energy argument of Derivation C, became the backbone of twentieth-century control theory: autopilots, power grids, and robot balance controllers are certified stable today by constructing Lyapunov functions, often automatically by convex optimization. He worked on the stability of rotating fluid masses — the shape of planets — proving results Poincaré had conjectured, and took his own life in 1918, days after his wife's death.

History

1788

— Lagrange argues in the Mécanique analytique that a potential minimum is a stable equilibrium

1846

— Dirichlet gives the first rigorous proof of Lagrange's criterion, the Lagrange–Dirichlet theorem

1892

— Lyapunov's doctoral thesis defines stability precisely and founds both the linearization and direct methods

1937

— Andronov and Pontryagin isolate structural stability, asking which portraits survive perturbation of the equations themselves

1960

— LaSalle's invariance principle completes the Lyapunov method, licensing conclusions from semi-definite energy drains

Going further

References. Strogatz, Nonlinear Dynamics and Chaos, 2nd ed., ch. 2, 5, 6 · Morin, Introduction to Classical Mechanics, ch. on small oscillations · Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §6.2 · Landau & Lifshitz, Mechanics, 3rd ed., §21 · Taylor, Classical Mechanics, ch. 5

Provenance. Definitions and classification follow Strogatz (Nonlinear Dynamics and Chaos, ch. 2,5,6), Morin, Goldstein §6.2, Landau-Lifshitz Mechanics §21, Taylor ch.5. A verifier should recompute: the pendulum Jacobian eigenvalues $\pm i\sqrt{g/\ell}$ (center, down) and $\pm\sqrt{g/\ell}$ (saddle, up); the trace-determinant eigenvalue formula $\lambda_\pm=\tfrac12(\tau\pm\sqrt{\tau^2-4\Delta})$ ; the small-oscillation frequency $\omega=\sqrt{V''(q^*)/m}$ ; the Lyapunov derivative $\dot E=-\gamma v^2$ for the damped oscillator; and the worked classification example ( $\tau=-3,\Delta=2\Rightarrow$ stable node). All numbers are standard textbook results.

Builds on: Phase Portraits and the Phase Plane · Ordinary differential equations as first-order systems · Potential energy and conservative forces Leads to: Linear Stability Analysis and the Jacobian · Classification of Fixed Points: Nodes, Saddles, Spirals, Centers · Conservative Systems and Hamiltonian Phase Flow · Bifurcations and the Loss of Stability

If fixed points are exact solutions, why does the system never actually reach one in finite time?

Uniqueness protects them. The constant solution occupies the fixed-point state for all time; if a moving trajectory arrived there at some finite $t_1$ , two distinct solutions would pass through one state, which the existence–uniqueness theorem forbids. So moving trajectories only converge asymptotically, closing the distance like $e^{\lambda t}$ but never to zero. Physically the approach is soon academic: after a few multiples of $1/|\lambda|$ the residual displacement is smaller than any noise floor. Watch the logic reverse for marginal fixed points, though — the finite-time blow-up of $\dot x=x^3$ in Derivation A shows nonlinearities can beat any exponential when the linear term is absent, and numerical integrators near such points need correspondingly small steps.

My mechanics course said equilibrium means net force zero. Why does this article keep insisting on more conditions?

Force balance is only half the state. In phase space the condition is $\mathbf{f}(\mathbf{x}^*)=0$ , which for a mechanical system reads $\dot q=p/m=0$ and $\dot p=-V'(q)=0$ : zero velocity as well as zero force. A coasting puck has zero net force but is nowhere near equilibrium — its state point marches steadily along the position axis. The distinction matters because stability theory is about the motion of the state, and only states satisfying both conditions stand still. When your statics course drew free-body diagrams and set forces to zero, it was implicitly working on the $p=0$ axis all along; dynamics just makes the second condition explicit.

Is a centre stable or not? Different books seem to disagree.

Both, once the vocabulary is precise — this is exactly why Lyapunov's grades exist. A centre is Lyapunov stable: start close enough and the orbit remains within any tolerance forever, circulating on its closed loop. It is not asymptotically stable: the orbit never returns to the equilibrium, and the memory of the initial kick persists undiminished. Books that say "stable" mean the first grade; books that say "not stable" usually reserve the bare word for the second. Conservative mechanics can only ever achieve the neutral grade — attracting equilibria need phase-area contraction, which Hamiltonian flow forbids — which is why the distinction was invented by people studying celestial mechanics, where nothing ever damps.

How do I choose a Lyapunov function when the system is not mechanical and there is no energy to guess?

There is no algorithm in general — that is the method's honest cost — but the working heuristics cover most textbook and many research cases. Try quadratic forms $V=\mathbf{x}^{\mathsf T}P\mathbf{x}$ first: for linear systems a suitable $P$ always exists (solve the Lyapunov equation $A^{\mathsf T}P+PA=-Q$ ), and the same $V$ often certifies the nonlinear system locally. For systems with a conserved quantity in some limit, perturb that quantity — energy plus a small cross-term like $\epsilon xv$ can convert a semi-definite drain into a strict one. For gradient-like systems, the potential itself works (Derivation B). Modern practice automates the search: sum-of-squares programming finds polynomial Lyapunov functions by convex optimization, which is how robot-balance and drone controllers ship with stability certificates.

The article says linearization fails when an eigenvalue has zero real part. How common is that, really?

As a random accident, measure-zero rare: a generic matrix has no eigenvalue exactly on the imaginary axis. But two forces make the marginal case central rather than exotic. First, structure: conservative systems have $\tau=0$ identically, so every centre in frictionless mechanics is permanently marginal — protected by energy conservation, not by genericity. Second, parameters: as you tune a knob, eigenvalues move continuously, and every qualitative change of the portrait must pass through the marginal wall. Bifurcation theory is nothing but the systematic study of these crossings. So the honest summary is: at a randomly chosen parameter value you will likely never meet a non-hyperbolic point; in the course of understanding a system's full behaviour you will meet one at every interesting threshold.

Are basins of attraction always simple regions with smooth boundaries?

No — and the failure modes are physically consequential. In the damped double well the two basins are separated by the smooth stable manifold of the saddle, but even there the basins spiral around each other near the wells' watershed, so a small error in initial condition near the boundary flips the outcome. Add periodic forcing or a third dimension and basin boundaries can become fractal: sets of initial conditions where attractor membership varies on every scale, making long-term outcomes effectively unpredictable despite deterministic dynamics and simple attractors. Magnetic pendulum toys over three magnets show this beautifully. Quantifying basin geometry — volumes, boundary dimension, "basin stability" — is a current tool in power-grid research, where it measures how large a disturbance the synchronous state can absorb.

Why do we bother with fixed points of maps when we already have the theory for flows?

Because maps are how flows are actually interrogated beyond the plane. A Poincaré section converts a three-dimensional flow into a two-dimensional return map; a periodic orbit of the flow becomes a fixed point of the map, and its stability — eigenvalues inside or outside the unit circle — is exactly the stability of the cycle. The translation dictionary is: imaginary axis for flows becomes unit circle for maps, $e^{\lambda T}$ are the map's multipliers. Maps also stand alone: population models with discrete generations, kicked rotors, and numerical integrators are intrinsically iterative, and the logistic map $x_{n+1}=rx_n(1-x_n)$ — unlike its tame flow cousin of Worked Example 1 — undergoes the period-doubling cascade into chaos that closes this chapter. One-dimensional maps can do what one-dimensional flows provably cannot.

What happens to a fixed point when I add a little noise to the system?

The fixed point stops being a destination and becomes a neighbourhood. With weak noise, the state jitters in a cloud around a stable equilibrium whose width balances noise strength against the restoring rate $|f'(x^*)|$ — for the linear case an Ornstein–Uhlenbeck cloud with variance proportional to (noise power)/(2|f'|). Two deterministic conclusions get softened. Escape becomes possible: rare fluctuations can kick the state over a basin boundary, at a rate exponentially suppressed by the barrier (Kramers' escape), which is how chemical reactions and magnetic bit flips happen. And marginality becomes measurable: as $|f'|\to0$ near a bifurcation, the cloud's variance and correlation time diverge — the critical-slowing-down signatures used as tipping-point early warnings in climate records.

Fixed Points and Equilibria

Where this lives today

The physical picture

The core equations

Limiting cases

Worked examples

Going further

Before this

Where this leads