Classification of Fixed Points: Nodes, Saddles, Spirals, Centers

Two numbers you can read off a matrix in five seconds — the trace and the determinant — sort every equilibrium of every planar system into a small zoo of local portraits. That is an outrageous compression of information. A fixed point of $\dot{\mathbf{x}}=\mathbf{f}(\mathbf{x})$ could sit inside any smooth vector field whatsoever, yet once the Jacobian is in hand, the entire local geometry is fixed by $\tau=\mathrm{tr}\,J$ and $\Delta=\det J$ : whether trajectories pull straight in, split along a mountain-pass, swirl inward, or circulate forever. This article draws the master map. The $(\tau,\Delta)$ plane is a literal atlas of destinies: the lower half is saddle country, the parabola $\tau^2=4\Delta$ is the frontier between straight-pulling nodes and swirling spirals, the positive vertical axis is the knife-edge ridge of centers. Every planar equilibrium you will ever compute — in mechanics, ecology, circuits, chemistry — is a pin stuck somewhere on this one diagram, and moving a physical parameter moves the pin. Watch a damped oscillator's pin march left across the parabola as you strengthen the damping, and "underdamped versus overdamped" stops being a formula and becomes geography. The map also encodes its own fine print. Its interior regions are robust: nudge the system and a spiral stays a spiral. Its boundary curves — the parabola, the axes — are measure-zero borderlines where the type is structurally fragile and nonlinear terms or parameter changes take over. Conservative mechanics lives permanently on one such borderline, the $\tau=0$ axis, which is why the Hamiltonian world of the earlier phase-space articles knows only centers and saddles. Crossing any borderline is a bifurcation; the map tells you, before you compute anything, where qualitative change is even possible.

Key takeaways

Near a planar fixed point the Jacobian $J$ governs the flow, and its two eigenvalues are fixed entirely by two invariants: $\tau=\mathrm{tr}\,J=\lambda_1+\lambda_2$ and $\Delta=\det J=\lambda_1\lambda_2$ .
The eigenvalues $\lambda_{1,2}=\tfrac12\big(\tau\pm\sqrt{\tau^2-4\Delta}\big)$ are real when $\tau^2>4\Delta$ (nodes, saddles) and complex when $\tau^2<4\Delta$ (spirals, centers).
The map: $\Delta<0$ — saddle, always; $\Delta>0$ with $\tau^2>4\Delta$ — node; $\tau^2<4\Delta$ — spiral; $\tau=0$ , $\Delta>0$ — linear center. Stability follows the sign of $\tau$ .
The borderlines are the interesting real estate: the parabola $\tau^2=4\Delta$ carries degenerate and star nodes, the line $\Delta=0$ carries non-isolated equilibria, and the ray $\tau=0,\ \Delta>0$ carries centers — all structurally fragile, all bifurcation sites.
Conservative systems are pinned to $\tau=0$ by phase-space incompressibility, so Hamiltonian flows admit only centers and saddles — the general map shows exactly how special that is.
Spiral rotation sense is not in $\tau$ or $\Delta$ : read it from the off-diagonal entries (the sign of $J_{21}$ at a spiral).
Key idea: Eigenvalues of the Jacobian, encoded by $(\tau,\Delta)$ : real same-sign → node, real opposite-sign → saddle, complex with $\tau\neq0$ → spiral, complex with $\tau=0$ → center. Stability follows the sign of $\tau$ (equivalently $\mathrm{Re}\,\lambda$ ).

Where this lives today

The map is quietly load-bearing in machine learning. Gradient descent with momentum on a quadratic loss is exactly a damped oscillator per eigendirection of the Hessian: the momentum coefficient sets $\tau$ , the curvature sets $\Delta$ , and the practitioner's "ringing versus sluggish convergence" dial is Example 1's walk across the parabola — optimal hyperparameters sit near critical damping, a fact exploited explicitly in the analysis of accelerated methods. Meanwhile in computational ecology and epidemiology, the same census ritual (find equilibria, pin them on the map, watch pins approach borderlines as parameters drift) underlies resilience dashboards: a disease-free equilibrium crossing $\Delta=0$ is an outbreak threshold crossing $R_0=1$ . Mechanics keeps its own claim. Spacecraft mission design classifies the Lagrange-point equilibria of the circular restricted three-body problem — saddle-center types whose rail structure generates the invariant-manifold "superhighways" used by low-fuel trajectories; power-grid engineers watch generator equilibria spiral-versus-node character to set damping controllers; and every MEMS accelerometer datasheet quoting a $Q$ factor is reporting where its pin sits relative to the parabola. Wherever a system has an operating point and two effective dimensions, someone is — knowingly or not — reading this map.

The physical picture

Picture the velocity field as wind blowing across the plane, stalling at the fixed point. The Jacobian records which way the wind leans around the stall. Two real eigenvalues mean the wind owns two straight preferred lines — eigenvector rails — pulling inward or pushing outward: a node if the rails agree, a saddle if they fight. Complex eigenvalues mean no direction is preferred; the wind rotates everything, and the real part decides whether the swirl tightens (stable spiral), unwinds (unstable spiral), or circulates forever (center). The whole taxonomy fits in one drawing. Plot every possible planar equilibrium by its coordinates $(\tau,\Delta)$ and the plane partitions into destiny regions:

Figure

The trace-determinant plane — the master map of every planar fixed point, with the parabola dividing nodes from spirals, saddle country below the axis, and the fragile borderlines that carry centers and degenerate nodes

Reading the map is faster than any formula. Saddle? Just check $\Delta<0$ — one glance. Above the axis, stability is the sign of $\tau$ (left half stable, right half unstable), and node-versus-spiral is which side of the golden parabola you are on. The two numbers even carry physical units of bookkeeping: $\tau=\mathrm{Re}\,\lambda_1+\mathrm{Re}\,\lambda_2$ is the local divergence of the flow — the rate at which small areas grow or shrink — while $\Delta$ is the product of the rates, negative exactly when the two rails disagree. What the portraits in each region actually look like is worth having in your eye before the algebra:

Figure

The portrait gallery — the six local behaviours a planar fixed point can have, drawn with their eigenvector rails and correct arrow sense

Saddles deserve special respect: their eigenvector rails extend, in the full nonlinear system, into the stable and unstable manifolds — the separatrices that carve the plane into basins and organize every global portrait in this chapter. And one thing the map deliberately omits: it cannot tell you which way a spiral turns. Rotation sense is not a function of $\tau$ and $\Delta$ — two systems can share both numbers and swirl opposite ways — so Derivation C extracts it from the one place it lives, the off-diagonal entries.

The core equations

The setting, from the previous article: a planar flow with a fixed point, and its local linear model

\dot{\boldsymbol{\xi}}=A\,\boldsymbol{\xi},\qquad A=D\mathbf{f}(\mathbf{x}^\ast)=\begin{pmatrix}\partial_x f_1 & \partial_y f_1\\ \partial_x f_2 & \partial_y f_2\end{pmatrix} \qquad (1)

where: $\boldsymbol{\xi}=\mathbf{x}-\mathbf{x}^\ast$ — displacement from equilibrium; $A$ — the Jacobian, whose construction and Hartman–Grobman license Linear Stability Analysis and the Jacobian established. This article's job is the complete taxonomy of what (1) can do. A $2\times2$ matrix has exactly two similarity invariants, and the characteristic polynomial is built from them alone:

\det(A-\lambda I)=\lambda^2-\tau\lambda+\Delta=0,\qquad \tau=\mathrm{tr}\,A,\quad \Delta=\det A \qquad (2)

where: $\tau=\lambda_1+\lambda_2$ and $\Delta=\lambda_1\lambda_2$ (Vieta). Any change of coordinates that keeps the dynamics — any similarity transform — leaves $\tau$ and $\Delta$ untouched, which is why two numbers can classify: they are everything about $A$ that geometry is allowed to see. The quadratic formula then hands over the eigenvalues:

\lambda_{1,2}=\frac{\tau\pm\sqrt{\tau^2-4\Delta}}{2} \qquad (3)

where: the discriminant $\tau^2-4\Delta$ is the single hinge of the whole classification — positive gives two real rates (rail behaviour: nodes, saddles), negative gives a conjugate pair $\tfrac{\tau}{2}\pm i\omega$ (swirl behaviour: spirals, centers), zero gives the repeated root of the parabola's degenerate nodes. The map itself, region by region (proved as Derivation A):

\begin{cases}\Delta<0 & \text{saddle}\\ \Delta>0,\ \tau^2>4\Delta & \text{node, stable iff }\tau<0\\ \Delta>0,\ \tau^2<4\Delta & \text{spiral, stable iff }\tau<0\\ \Delta>0,\ \tau=0 & \text{linear center}\end{cases} \qquad (4)

where: the saddle condition needs no reference to $\tau$ at all — a negative product forces real eigenvalues of opposite sign, one rail in, one rail out, unstable regardless of everything else. Note what the case list omits: rotation direction (not $(\tau,\Delta)$ -measurable; see eq. 7) and the guarantee that a linear center survives nonlinearity (it does not; see the warning). For complex eigenvalues the two invariants translate directly into the spiral's two physical rates:

\lambda=\frac{\tau}{2}\pm i\omega,\qquad \omega=\frac{\sqrt{4\Delta-\tau^2}}{2} \qquad (5)

where: $\tau/2$ — the radial e-folding rate of the spiral (negative tightens, positive unwinds); $\omega$ — the angular frequency of the swirl. A spiral's pitch is the competition $|\tau|/2\omega$ : pins near the parabola wind tightly inward like an over-steeped node, pins near the $\Delta$ -axis circulate many times per e-folding, grading continuously into the center at $\tau=0$ . On the parabola itself the two eigenvalues collide, $\lambda_1=\lambda_2=\tau/2$ , and the matrix generically runs out of eigenvectors:

\tau^2=4\Delta:\qquad \boldsymbol{\xi}(t)=e^{\lambda t}\big(\mathbf{c}_1+t\,\mathbf{c}_2\big) \qquad (6)

where: the secular factor $t$ marks a degenerate (improper) node — one rail only, all trajectories curling around to enter along it (Derivation B). The exception is $A=\lambda I$ exactly, the star node, where every direction is a rail and trajectories fly straight in. Critical damping is the physicist's daily encounter with this line. Rotation sense — the datum the invariants cannot carry — comes from one matrix entry:

\text{at a spiral or center:}\qquad J_{21}>0\ \Rightarrow\ \text{counterclockwise},\qquad J_{21}<0\ \Rightarrow\ \text{clockwise} \qquad (7)

where: $J_{21}=\partial f_2/\partial x$ — evaluated at the fixed point; Derivation C shows why one sign suffices when the eigenvalues are complex. For the mechanical form $\dot x=y,\ \dot y=g(x,y)$ this entry is the (negative) stiffness, negative at any restoring equilibrium — which is why every mechanical oscillation runs clockwise in the $(x,\dot x)$ plane. Finally, the conservative special case. Hamilton's equations force $J_{11}=-J_{22}$ (mixed partials of $H$ ), so

\text{Hamiltonian flow:}\quad \tau=0\ \text{identically}\quad\Longrightarrow\quad \text{centers}\ (\Delta>0)\ \text{or saddles}\ (\Delta<0)\ \text{only} \qquad (8)

where: the pin never leaves the vertical axis of the map, whatever the potential, whatever the parameters. The derivation and its physical reading — incompressible phase fluid, Liouville, the Lagrange–Dirichlet stability theorem — belong to the Hamiltonian thread of this chapter: see Fixed Points and Stability in Phase Space for the full story and Conservative Systems and Hamiltonian Phase Flow for the energy-based view. From this article's vantage the message is structural: conservative mechanics lives permanently on a borderline that generic systems only visit at bifurcation instants, and the smallest friction ( $\tau<0$ ) drops every center off the axis into a spiral.

You want	Reach for	Because
instant saddle test	sign of $\Delta$	$\Delta<0$ settles it, no $\tau$ needed
node vs spiral	sign of $\tau^2-4\Delta$	real rails vs complex swirl
stable vs unstable (when $\Delta>0$ )	sign of $\tau$	trace = sum of growth rates
spiral's tightness and pitch	$\tau/2$ and $\omega$ from eq. (5)	radial rate vs angular rate
swirl direction	sign of $J_{21}$ , eq. (7)	invisible to $\tau,\Delta$
where bifurcations can strike	the borderlines: $\Delta=0$ , $\tau=0$ , $\tau^2=4\Delta$	type change requires crossing one

Derivation AThe trace-determinant map from the two invariants

We derive the full region-by-region classification, equation (4), from nothing but Vieta's formulas — the statement that $\tau$ and $\Delta$ are the sum and product of the eigenvalues. The strategy is a case split on the discriminant and the signs, and the point to watch is economy: at no step do we need eigenvectors, coordinates, or the matrix entries themselves. Two similarity invariants in, complete local topology out.

Step 1. From the characteristic polynomial $\lambda^2-\tau\lambda+\Delta=0$ , Vieta gives the dictionary both ways:

\tau=\lambda_1+\lambda_2,\qquad \Delta=\lambda_1\lambda_2

Why invariance matters: a change of basis $A\to P^{-1}AP$ is the algebra of redrawing the portrait in new coordinates; $\tau$ and $\Delta$ survive it, so whatever they determine is a property of the flow itself, not of the axes chosen.

Step 2. Saddle country. Suppose $\Delta<0$ . A complex-conjugate pair has product $|\lambda|^2>0$ , so negative product forces two real eigenvalues of opposite sign: exponential escape along one rail, decay along the other — a saddle, and automatically unstable. No condition on $\tau$ enters.

Step 3. Real same-sign country. Suppose $\Delta>0$ and $\tau^2>4\Delta$ : the discriminant is positive, both roots real, and their positive product means a shared sign — the sign of their sum $\tau$ . Both negative ( $\tau<0$ ): every direction decays at its own rate, a stable node. Both positive: its time-reverse, an unstable node.

Step 4. Swirl country. Suppose $\tau^2<4\Delta$ (which forces $\Delta>0$ ): the roots are $\tfrac{\tau}{2}\pm i\omega$ with $\omega=\tfrac12\sqrt{4\Delta-\tau^2}$ . The common real part $\tau/2$ sets radial fate — stable spiral, unstable spiral, or, exactly at $\tau=0$ , the closed orbits of a linear center with period $2\pi/\sqrt{\Delta}$ .

Step 5. The boundaries collect the leftovers: $\tau^2=4\Delta$ (repeated root — degenerate or star nodes, Derivation B), $\Delta=0$ (a zero eigenvalue — a line of non-isolated equilibria in the linear system, the saddle-node staging ground), and $\tau=0,\Delta>0$ (centers). Each is a measure-zero curve: nudge $(\tau,\Delta)$ and you fall off into an open region, which is the precise sense in which borderline types are structurally fragile and interior types robust. Verified: consistency checks — at $\tau=0,\ \Delta>0$ equation (3) gives $\lambda=\pm i\sqrt{\Delta}$ , pure imaginary, matching Step 4's center; at $\Delta<0$ the product of roots is negative, confirming opposite signs; the four open regions and three boundary curves reproduce Strogatz Fig. 5.2.8 exactly.

Go deeper · Derivation B · 2 minOn the parabola — degenerate and star nodes from a defective matrix

Derivation BOn the parabola — degenerate and star nodes from a defective matrix

The parabola $\tau^2=4\Delta$ is where the discriminant dies and the eigenvalues merge, $\lambda_1=\lambda_2=\lambda=\tau/2$ . We solve the linear flow exactly there and extract the two possible portraits: the generic degenerate node with its single rail and secular $t\,e^{\lambda t}$ creep, and the exceptional star node. The strategy is Jordan's: a repeated eigenvalue leaves only two possibilities for the matrix, and each is integrable in two lines.

Step 1. Split the cases by counting eigenvectors. With repeated $\lambda$ , either $A=\lambda I$ (every vector an eigenvector) or $A$ is similar to the Jordan block

A\sim\begin{pmatrix}\lambda & 1\\ 0 & \lambda\end{pmatrix}

Why only these two: if two independent eigenvectors exist, $A$ acts as $\lambda$ on all of them, hence on the whole plane, forcing $A=\lambda I$ ; if not, the matrix is defective and the Jordan normal form is the block above.

Step 2. Star node first. $A=\lambda I$ gives $\dot{\boldsymbol{\xi}}=\lambda\boldsymbol{\xi}$ : every ray through the origin is invariant, and $\boldsymbol{\xi}(t)=e^{\lambda t}\boldsymbol{\xi}(0)$ moves straight along it. The portrait is a starburst of radial lines, inward for $\lambda<0$ . This case requires the entries to conspire ( $J_{12}=J_{21}=0$ , $J_{11}=J_{22}$ ), not just the invariants — isotropic damping is the physical recipe.

Step 3. Degenerate node. In Jordan coordinates $(u,v)$ : $\dot v=\lambda v$ integrates to $v=v_0e^{\lambda t}$ , then $\dot u=\lambda u+v$ is driven at its own rate, and integrating factor $e^{-\lambda t}$ yields

u(t)=(u_0+v_0\,t)\,e^{\lambda t}

— equation (6). The lone eigenvector is the $u$ -axis.

Step 4. Read the geometry. For $\lambda<0$ every trajectory dies, but the ratio $v/u=v_0/(u_0+v_0t)\to0$ : all orbits enter tangent to the single rail, having first swung around in a characteristic half-curl — the fingerprint by which you recognize a degenerate node in a numerically drawn portrait. It is the boundary creature in every sense: squeeze a stable node's two rails together ( $\tau^2\to4\Delta^+$ ) and they merge into it; push further and the rails go complex and the curl completes into a spiral. Verified: critical damping check — $\ddot x+2\gamma\dot x+\gamma^2x=0$ sits exactly on the parabola ( $\tau=-2\gamma$ , $\Delta=\gamma^2$ ), and its standard solution $x=(c_1+c_2t)e^{-\gamma t}$ is precisely Step 3's form; the secular $t$ is why a critically damped gauge needle makes one swift approach with no overshoot. Matches Strogatz §5.2 and Hirsch–Smale–Devaney Ch. 3.

Go deeper · Derivation C · 1 minReading the rotation direction — what the invariants cannot tell you

Derivation CReading the rotation direction — what the invariants cannot tell you

Two spirals can share $\tau$ and $\Delta$ and turn opposite ways, so the map is silent about swirl sense. We derive the one-entry test of equation (7). The strategy is embarrassingly direct: evaluate the linear velocity field at a single convenient point and see which way it pushes — a trick worth internalizing, because it works on any portrait you will ever sketch.

Step 1. Take the linearization $\dot{\boldsymbol{\xi}}=J\boldsymbol{\xi}$ with complex eigenvalues, and probe the field at the point $\boldsymbol{\xi}=(1,0)$ on the positive $x$ -axis:

\dot{\boldsymbol{\xi}}\big|_{(1,0)}=\big(J_{11},\ J_{21}\big)

Step 2. The vertical component decides. If $J_{21}>0$ the flow crosses the positive $x$ -axis moving upward — counterclockwise circulation; if $J_{21}<0$ , downward — clockwise. The horizontal component $J_{11}$ only slides the point along the axis and cannot circulate it. Why one probe suffices: the discriminant is $\tau^2-4\Delta=(J_{11}-J_{22})^2+4J_{12}J_{21}$ , so complex eigenvalues force $J_{12}J_{21}<0$ — in particular $J_{21}\neq0$ , the field never crosses the $x$ -axis horizontally, and the sign read at one probe point is the sign at every crossing. The two off-diagonal entries always disagree, so either determines the sense (with opposite conventions).

Step 3. Specialize to mechanics. For $\dot x=y,\ \dot y=g(x,y)$ : $J_{21}=\partial g/\partial x=-V''(x^\ast)/m$ at a potential minimum — negative. Every oscillatory mechanical equilibrium therefore swirls clockwise in the $(x,\dot x)$ plane: moving right at the top ( $\dot x>0$ above the axis), falling on the right, moving left at the bottom. One memorized fact replaces endless arrow-checking. Verified: damped oscillator $J=\begin{pmatrix}0&1\\-\omega_0^2&-2\gamma\end{pmatrix}$ : $J_{21}=-\omega_0^2<0$ — clockwise, agreeing with the physical cycle (position increasing while velocity positive). The counterexample pair $\dot x=-y,\dot y=x$ versus $\dot x=y,\dot y=-x$ shares $\tau=0,\ \Delta=1$ and swirls opposite ways, confirming the invariants' silence.

Limiting cases

Case	Result	Meaning
$\Delta<0$ , any $\tau$	saddle	Opposite-sign rails; always unstable, always hyperbolic — the most robust verdict on the map.
$\tau<0$ , $0<\Delta<\tau^2/4$	stable node	Two decaying rails; arrival tangent to the slow one.
$\tau<0$ , $\Delta>\tau^2/4$	stable spiral	Decay with rotation; pitch set by $\|\tau\|/2\omega$ .
$\tau=0$ , $\Delta>0$	linear center	Non-hyperbolic knife-edge; nonlinear terms adjudicate (warning below).
$\tau^2=4\Delta$ , $A\neq\lambda I$	degenerate node	Single rail, secular $t\,e^{\lambda t}$ ; critical damping lives here.
$\tau^2=4\Delta$ , $A=\lambda I$	star node	Straight rays; needs entry-level fine-tuning, not just invariants.
$\Delta=0$ , $\tau\neq0$	line of equilibria (linear)	Zero eigenvalue; nonlinearity typically resolves it into a saddle-node event.
$\tau\to0^-$ at fixed $\Delta>0$	spiral tightening toward center	The Hopf corridor: a complex pair marching to the axis, oscillation about to be born.

Common trapA center on the map is a conjecture, not a verdict

The ray $\tau=0,\ \Delta>0$ is the one region of the trace-determinant plane whose label you may not believe outright. Everywhere else the fixed point is hyperbolic and Hartman–Grobman transfers the linear portrait to the true nonlinear flow. On the center ray the license is void: the closed orbits of the linearization can be unwound into slow spirals by cubic terms the Jacobian cannot see — the previous article's Derivation C exhibits one system whose "center" is a stable spiral, a true center, or an unstable spiral depending on an invisible parameter. A linear center is confirmed only by extra structure: a conserved energy (the conservative systems of this chapter, pinned to the axis by equation 8), a time-reversal symmetry, or an explicit first integral. The same skepticism applies on $\Delta=0$ : the linear "line of equilibria" is almost always an artifact that quadratic terms collapse into isolated fixed points about to be born or destroyed. Interior regions of the map are testimony; the borderlines are only invitations to look closer.

Figure

One oscillator walks the map — increasing the damping slides the pin left along a line of constant stiffness, crossing from spiral through the critical parabola into node country

Worked examples

Example 1 — one oscillator walks the whole map (confidence builder). The damped oscillator $\ddot x+b\dot x+x=0$ (mass and stiffness scaled to 1) has Jacobian invariants $\tau=-b$ , $\Delta=1$ . Classify it for $b=0.6$ , $b=2$ , and $b=3$ , and find where the type changes.

Step 1. The pin sits at $(-b,\ 1)$ : increasing damping slides it left along the horizontal line $\Delta=1$ of the map. The parabola $\tau^2=4\Delta$ is crossed when $b^2=4$ , i.e. $b=2$ — the critical damping value, now read off as geography.

Step 2. $b=0.6$ : discriminant $\tau^2-4\Delta=0.36-4=-3.64<0$ — stable spiral. Eigenvalues $\lambda=-0.3\pm i\sqrt{3.64}/2=-0.3\pm0.954\,i$ : e-folding time $1/0.3=3.3$ time units, swirl period $2\pi/0.954=6.59$ . Clockwise, by equation (7): $J_{21}=-1<0$ .

Step 3. $b=2$ : discriminant zero, repeated $\lambda=-1$ , and $A\neq\lambda I$ (the top row is $(0,1)$ ) — a degenerate node with solution $(c_1+c_2t)e^{-t}$ : the fastest overshoot-free return, which is why gauges and door closers are tuned here.

Step 4. $b=3$ : discriminant $9-4=5>0$ — stable node, $\lambda=(-3\pm\sqrt5)/2=-0.382,\ -2.618$ . Arrival is tangent to the slow rail ( $\lambda=-0.382$ ): an overdamped needle creeps in with time constant $1/0.382=2.6$ , slower than the critically damped case — more damping, lazier return.

Step 5. Collect the walk: spiral → degenerate node → node, entirely by sliding one pin. Underdamped/critical/overdamped is not three formulas but one straight-line stroll across the parabola.

Example 2 — full census of the damped pendulum (exam level). For $\dot\theta=y,\ \dot y=-\sin\theta-b\,y$ with $b=0.2$ : classify all equilibria, give numbers, and find the damping at which the bottom equilibrium stops spiralling.

Step 1. Fixed points: $y=0,\ \sin\theta=0$ : the hanging state $(0,0)$ and the inverted state $(\pi,0)$ (mod $2\pi$ ).

Step 2. Jacobian: top row $(0,1)$ ; bottom row $(-\cos\theta,\ -b)$ .

Step 3. Hanging state $(0,0)$ : $\tau=-0.2$ , $\Delta=\cos0=1>0$ ; discriminant $0.04-4<0$ — stable spiral. Eigenvalues $\lambda=-0.1\pm i\sqrt{4-0.04}/2=-0.1\pm0.995\,i$ : the pendulum rings at essentially its undamped frequency while amplitudes e-fold away in $1/0.1=10$ time units. Clockwise ( $J_{21}=-1<0$ ), as all mechanical oscillations are.

Step 4. Inverted state $(\pi,0)$ : $\Delta=-\cos\pi\cdot1$ ... compute carefully: $J_{21}=-\cos\pi=+1$ , so $\Delta=(0)(-b)-(1)(1)=-1<0$ — saddle, instantly, from the determinant alone. Eigenvalues $\lambda=(-0.2\pm\sqrt{0.04+4})/2=-0.1\pm1.005$ : escape rate $0.905$ , approach rate $-1.105$ . The saddle's rails extend into the separatrices dividing swinging from over-the-top rotation.

Step 5. Spiral-to-node threshold at the bottom: $\tau^2=4\Delta$ gives $b^2=4$ , $b=2$ (in units of $\sqrt{g/L}$ ; restoring units, $b_{\text{crit}}=2\sqrt{g/L}$ ). Below it the pendulum rings; above it, it oozes back without a single overshoot.

Step 6. Structural summary on the map: the saddle pin sits deep in robust territory ( $\Delta=-1$ ) — no amount of small damping-tuning changes it; the bottom pin at $(-0.2,1)$ sits just left of the center axis, where the undamped pendulum's center used to be. Friction moved the pin off the Hamiltonian axis: the conservative center became a dissipative spiral, and the map shows that this is the generic fate of every conservative center under the smallest $\tau<0$ .

Origins · 1 minVladimir Arnold and how this was discovered

Vladimir Arnold

1937–2010 · Moscow

Arnold learned dynamics in Kolmogorov's Moscow seminar and, at 20, solved Hilbert's 13th problem; by 26 he had proved the "A" of KAM theory. But his fingerprints on this article come from a different service: Arnold was the great geometrizer of differential equations. His textbook Ordinary Differential Equations — with the trace-determinant plane, portrait galleries, and structural-stability reasoning at its heart — retrained generations to see an ODE not as a formula factory but as a field of arrows with a geography of robust regions and fragile borderlines, exactly the reading this article teaches. Notoriously sharp-tongued about formalism without pictures ("mathematics is the part of physics where experiments are cheap"), he fought to keep geometry at the center of the subject. The classification map, in the form you have just learned it, is largely his pedagogy — the local alphabet from which his global theory of dynamical systems is spelled.

History

1881

— Poincaré's memoir on curves defined by differential equations names and classifies nodes, saddles, foci and centers

1892

— Lyapunov's stability thesis separates what linearization decides from the marginal cases it cannot

1937

— Andronov and Pontryagin define structural stability, making "robust regions, fragile borderlines" a theorem

1963

— Arnold proves the KAM theorem's key case; his Moscow lectures begin geometrizing the ODE curriculum

1971

— Arnold's Ordinary Differential Equations spreads the trace-determinant portrait gallery worldwide

Going further

References. Morin, Introduction to Classical Mechanics (2008), Ch. on oscillations and phase plane · Goldstein, Poole & Safko, Classical Mechanics 3rd ed. · Landau & Lifshitz, Mechanics (Vol. 1), §§ on small oscillations · Taylor, Classical Mechanics (2005) · Strogatz, Nonlinear Dynamics and Chaos, 2nd ed., Ch. 5–6

Provenance. Standard linear-stability classification; trace–determinant diagram as in Strogatz Ch. 5–6 and Taylor. A verifier should reconfirm: (1) Vieta relations $\tau=\lambda_1+\lambda_2,\ \Delta=\lambda_1\lambda_2$ ; (2) discriminant sign rules for node/spiral; (3) the trace-free property of Hamiltonian Jacobians via equality of mixed partials; (4) the damped-oscillator mapping $\tau=-2\gamma,\ \Delta=\omega_0^2$ . Cross-check pendulum center/saddle alternation against Landau–Lifshitz Mechanics and Goldstein. Additional worked material in this upgrade: the Jordan-block solution $(c_1+c_2t)e^{\lambda t}$ on the parabola (Hirsch–Smale–Devaney Ch. 3), the rotation-sense test via the sign of $J_{21}$ (with $J_{12}J_{21}<0$ forced by complex eigenvalues), and the damped-pendulum census ( $b=0.2$ : spiral $-0.1\pm0.995i$ at the bottom, saddle $-0.1\pm1.005$ inverted; critical $b=2$ ) — each recomputed here.

Builds on: Fixed Points and Equilibria · Linear Stability Analysis and the Jacobian · Phase Portraits and the Phase Plane Leads to: Conservative Systems and Hamiltonian Phase Flow · Bifurcations and the Loss of Stability · Hopf Bifurcation and Birth of Limit Cycles · Limit Cycles and Self-Sustained Oscillations

Do I ever need the actual eigenvectors, or do trace and determinant always suffice?

The invariants fix the type; the eigenvectors fix its orientation in your coordinates. For a saddle they are non-negotiable: the rails become the separatrices partitioning the plane into basins. For a stable node you need at least the slow one, which dictates the arrival direction. Spirals and centers have no real eigenvectors — the only orientation datum, swirl direction, comes from the off-diagonal sign test. Practical workflow: classify from $(\tau,\Delta)$ first, then compute eigenvectors only for the saddles and nodes you intend to draw.

Is a star node physically common, or a textbook curiosity?

Rarer than a degenerate node, and the map explains why: the star needs $A=\lambda I$ — three entry conditions, not just the invariant condition $\tau^2=4\Delta$ — a symmetry rather than a tuning. It does occur honestly: isotropic drag ( $\dot{\mathbf{v}}=-\gamma\mathbf{v}$ ) is a star node in velocity space, as is any radially symmetric overdamped relaxation. The degenerate node, by contrast, appears every time one real parameter crosses critical damping — engineers meet it daily; star nodes mostly live in problem sets.

The map says nothing about how big the neighbourhood of validity is. How local is local?

The classification is exact for the linear system and topologically faithful for the nonlinear one in some neighbourhood — whose size the theorems refuse to quantify, since it depends on how fast quadratic terms grow relative to linear ones. Practical estimate: trust the picture out to displacements where neglected terms are a small fraction of kept ones, $|\boldsymbol{\xi}|\ll$ (linear coefficients)/(quadratic coefficients) — order one radian for the pendulum, generous; near a borderline it can shrink to nothing, the linear term itself being weak along some direction. Rule of thumb: the closer the pin sits to a boundary of the map, the smaller the neighbourhood the label governs.

Why is the saddle the only type that never needs a stability discussion?

Because its instability is categorical rather than quantitative. Every other verdict on the map involves comparing rates to zero — is $\tau$ negative, is the real part negative — and flips under time reversal. A saddle has one genuinely growing direction whatever the trace does: $\Delta<0$ forces opposite-sign real eigenvalues, so escape is guaranteed along one rail, approach along the other, and time reversal maps the saddle to another saddle. It is also maximally robust — the whole lower half-plane is one region, so no small perturbation of the system can reclassify it. That permanence is why saddles, via their manifolds, form the durable skeleton of global phase portraits.

How do I place a real system's pin on the map when the model has parameters?

Treat the pin as a curve parametrized by the physics. Compute $\tau(p)$ and $\Delta(p)$ symbolically and trace the path as the parameter $p$ varies — Example 1's damping walk is the simplest case, a straight horizontal line. The payoff moments are the crossings: through the parabola (response character changes: ringing to creeping), through $\tau=0$ at positive $\Delta$ (a Hopf event: spiral reverses stability and a limit cycle is typically born — see the bifurcation articles), through $\Delta=0$ (equilibria collide and annihilate: saddle-node). Reading bifurcations as road-crossings on this one map, before computing any normal form, is the single most useful habit this chapter installs.

Complex eigenvalues came out of a real matrix. Where is the imaginary motion physically?

Nowhere — and everywhere in the geometry. A conjugate pair $\tfrac{\tau}{2}\pm i\omega$ is the plane's compact notation for a real block that rotates by $\omega$ while scaling by $e^{\tau t/2}$ : two real dimensions doing a real spiral. Nothing physical is imaginary; rather, no single real direction is preserved, so the "modes" cannot be real rays and pair up into a rotating plane instead. The observable content is concrete: $\omega$ is the ringing frequency of the return to equilibrium, $\tau/2$ the decay rate of its envelope, and the ratio sets how many swings you see before the motion dies — the $Q$ factor, read directly off the pin's position.

Equation 8 pins conservative systems to the tau equals zero axis. Does adding ANY friction really change the type instantly?

Formally yes, and the map makes the suddenness precise: the center label lives only on the axis, so any $\tau<0$ , however tiny, reclassifies the equilibrium as a stable spiral. What changes gradually is the quantitative portrait — for minuscule friction the spiral's pitch $|\tau|/2\omega$ is minuscule and one orbit looks closed to any finite-resolution plot. This is the resolution of an apparent paradox: qualitatively, dissipation is a discontinuous edit (closed orbits cease to exist, an attractor appears, time-reversal symmetry breaks); quantitatively, everything deforms continuously. The Hamiltonian articles' phase-fluid incompressibility is exactly what holds the pin on the axis, and its violation — however small — is what lets phase-space area leak and orbits terminate at the fixed point.

Does the trace-determinant classification survive in three dimensions?

The philosophy survives; the tidy plane does not. A $3\times3$ Jacobian has three invariants ( $\tau$ , the sum of principal $2\times2$ minors, and $\Delta$ ), the type space becomes three-dimensional, and new hybrid creatures appear — saddle-foci with a spiral in one plane and a real rail transverse to it, the fixed points at the heart of Lorenz-type chaos. Stability still reduces to "all real parts negative," testable by Routh–Hurwitz inequalities on the invariants without solving the cubic. But no flat map exhibits all the regions at a glance, and the planar constraint that made this chapter's global theory possible (trajectories cannot cross in two dimensions) is gone — which is precisely why the next articles' Poincaré–Bendixson world ends at the plane's edge.

Classification of Fixed Points: Nodes, Saddles, Spirals, Centers

Where this lives today

The physical picture

The core equations

Limiting cases

Worked examples

Going further

Before this

Where this leads