Phase Portraits and the Phase Plane

Here is the confession that opens every serious course on nonlinear dynamics: most differential equations cannot be solved. Not "are hard to solve" — cannot be, in any closed form, ever. The pendulum driven past small angles, two species competing for food, a transistor circuit with feedback: write down $\dot{\mathbf{x}}=\mathbf{f}(\mathbf{x})$ and the formula for $\mathbf{x}(t)$ simply does not exist. For a century this looked like a wall. Poincaré's insight, the founding move of this whole chapter, was to stop asking for formulas and start asking for pictures: treat the state of the system as a point in a plane, the equation as a wind field blowing that point around, and study the geometry of all possible journeys at once. That picture is the phase portrait, and it answers the questions you actually care about — does the motion settle, oscillate, or run away? — without integrating anything. You have already met phase portraits in the Hamiltonian articles of this course, where energy conservation pins every trajectory to a level curve of $H$ . This article deliberately drops that crutch. Real systems leak energy, and the moment they do, the level-curve picture shatters: closed orbits break into spirals, new objects called attractors appear, and the plane divides into basins competing for trajectories. The general theory — vector fields, nullclines, no-crossing, area contraction — is the language every remaining article in this chapter is written in, from limit cycles to bifurcations to the doorstep of chaos.

Key takeaways

A phase portrait displays every solution of $\dot{\mathbf{x}}=\mathbf{f}(\mathbf{x})$ at once: each state is a point in the plane, $\mathbf{f}$ assigns it a velocity arrow, and trajectories are the curves that follow the arrows.
The existence–uniqueness theorem forbids trajectories from crossing; in two dimensions this single fact constrains the possible long-term behaviours to equilibria and closed orbits (the content of the Poincaré–Bendixson theorem).
Nullclines — the curves where $\dot{x}=0$ or $\dot{y}=0$ — carve the plane into regions of definite flow direction and intersect at the fixed points; they are the working tool for sketching portraits by hand.
The divergence $\nabla\cdot\mathbf{f}$ measures how phase area evolves: conservative systems preserve it ( $\nabla\cdot\mathbf{f}=0$ , closed level-curve orbits, no attractors), while dissipative systems contract it ( $\nabla\cdot\mathbf{f}<0$ ), collapsing trajectories onto attractors.
Damping turns the pendulum's centres into stable spirals, keeps its saddles, and breaks its separatrix into basin boundaries — the archetype of how dissipation redraws a conservative portrait.
Key idea: The equations of motion define a vector field on the space of states; the phase portrait — fixed points, nullclines, separatrices, attractors — is the geometry of that field, and it delivers the qualitative fate of every initial condition without solving the equations.

Where this lives today

The phase portrait outgrew mechanics long ago and became the shared dialect of every field that models feedback. Neuroscience reads the FitzHugh–Nagumo portrait — a cubic nullcline crossed by a line — to explain why a neuron fires once and rests, or fires rhythmically; the geometry, not the biochemical detail, carries the explanation. Power-grid engineers live inside a phase plane too: each generator's rotor angle and frequency form a state, the synchronous grid is a stable equilibrium, and the practical question after a lightning strike — will the grid resynchronize or cascade? — is precisely "did the fault kick the state outside the basin of attraction?" With inverter-based renewables shrinking grid inertia, computing those basin boundaries has become a live 2026 engineering problem, run daily by transmission operators. The newest twist inverts the workflow. Instead of drawing portraits from known equations, data-driven methods — sparse regression on measured trajectories (SINDy and descendants), neural ODE surrogates for climate and fusion-plasma models — reconstruct the vector field itself from experimental time series, then interrogate the learned portrait for fixed points, basins, and impending bifurcations. The qualitative theory Poincaré built for want of solutions now serves as the interpretability layer for models that were never written down by humans in the first place.

The physical picture

Picture the phase plane as a pond with a permanent current: at every point an arrow tells a drifting leaf which way to move and how fast. The arrow field is $\mathbf{f}(\mathbf{x})$ — it never changes, because the system is autonomous. Drop a leaf anywhere and its path is fully determined; that path is a trajectory. The phase portrait is the whole current pattern seen at once: the still points where a leaf never moves, the whirlpools that swallow leaves, the ridgelines that send neighbouring leaves to opposite fates.

Figure

One picture, every fate — a dissipative phase portrait with a spiral attractor, a saddle, and the basin boundary that decides which trajectories are captured

Two habits make the picture quantitative. First, when the system is mechanical — $\ddot{x}=g(x,\dot{x})$ recast with $y=\dot{x}$ — the plane has a built-in circulation rule: in the upper half $y>0$ so the state drifts right, in the lower half it drifts left. Loops therefore always run clockwise, and a glance tells you whether a curve can be a legal trajectory. Second, time is invisible but not lost: arrows shorten near fixed points, so the state lingers there. A trajectory passing close to a saddle can take arbitrarily long to clear it — the source of the critical slowing down we will meet again at bifurcations. The deepest divide in the subject is drawn by energy. The companion article Phase Portraits and Trajectories develops the conservative case, where $H(q,p)$ is constant and every trajectory is imprisoned on a level curve of the energy — the portrait is a contour map, drawable without integration. This article's business is everything that happens when that protection fails. Add friction and trajectories cut across the old energy contours, always downhill; closed orbits cease to exist near equilibria, and whole regions of the plane drain into attracting points like watersheds draining into lakes.

Figure

Dissipation redraws the portrait — the conservative well's closed loops become a spiral that cuts across the old energy contours, always inward

That downhill rule is worth internalizing before any formalism: for a damped mechanical system the energy is a strictly decreasing clock, so a trajectory can revisit an energy contour only if it never left it. Once you see the portrait as "contours plus leakage," the taxonomy of the whole chapter — spirals, nodes, basins, and the limit cycles that manage to sustain themselves against leakage — falls into place.

The core equations

Everything starts by writing the dynamics as a first-order autonomous system. Whatever the physics — mechanics, circuits, populations — the state is a vector $\mathbf{x}$ and the model is its velocity field:

\dot{\mathbf{x}} = \mathbf{f}(\mathbf{x}), \qquad \mathbf{x}=(x,y)\in\mathbb{R}^2 \qquad (1)

where: $\mathbf{x}$ — the state, a single point encoding everything needed to predict the future; $\mathbf{f}$ — the vector field, assigning each state its instantaneous velocity. Autonomy ( $\mathbf{f}$ independent of $t$ ) is what makes the portrait a fixed picture: the current pattern never changes, only the leaf's position on it. Mechanics enters through a standard reduction. Newton's law is second order, but one substitution converts it:

\ddot{x}=g(x,\dot{x}) \;\Longrightarrow\; \begin{cases}\dot{x}=y\\[1mm] \dot{y}=g(x,y)\end{cases} \qquad (2)

where: $y=\dot{x}$ — velocity promoted to an independent coordinate. The two initial data a second-order equation needs, $x(0)$ and $\dot{x}(0)$ , become the two coordinates of a single starting point: that is why one degree of freedom lives in a two-dimensional phase plane. Note the built-in structure: the first row forces upper-half states rightward, lower-half states leftward — the clockwise rule. Because the field is time-independent, time can be eliminated entirely. Dividing the two components gives the slope of the orbit through any point:

\frac{\mathrm{d}y}{\mathrm{d}x}=\frac{\dot{y}}{\dot{x}}=\frac{f_2(x,y)}{f_1(x,y)} \qquad (3)

where: the left side is pure geometry — no clock. This is the sense in which the portrait shows shapes of motion, not schedules; to recover timing along an orbit you reinstate $\mathrm{d}t=\mathrm{d}x/f_1$ . Where $\mathbf{f}$ is small the state crawls, which is why portraits are read with an eye on arrow lengths near special points. The skeleton of the portrait is the set of states where the current stops:

\mathbf{f}(\mathbf{x}^*)=\mathbf{0} \qquad (4)

where: $\mathbf{x}^*$ — a fixed point: a genuine solution that sits still forever. The article Fixed Points and Equilibria is devoted to these; here they matter as the portrait's organizing anchors. Everything between fixed points must keep moving — and, by uniqueness, along exactly one curve. Finding fixed points and flow directions by hand uses the workhorse tool of the subject, the nullclines:

x\text{-nullcline: } f_1(x,y)=0, \qquad y\text{-nullcline: } f_2(x,y)=0 \qquad (5)

where: on the $x$ -nullcline the flow is purely vertical (no horizontal velocity), on the $y$ -nullcline purely horizontal — and the fixed points are exactly their intersections. The nullclines partition the plane into regions inside which the signs of $(\dot{x},\dot{y})$ cannot change, so one arrow per region pins down the global circulation. Do not confuse nullclines with trajectories: they are construction lines, crossed by the flow everywhere except where they happen to align with it.

Figure

The nullcline method — dashed construction curves carve the plane into sign regions, one arrow per region, and their intersection is the fixed point

Near a fixed point the field flattens into its own best linear model. Writing $\mathbf{x}=\mathbf{x}^*+\boldsymbol{\eta}$ and Taylor-expanding,

\dot{\boldsymbol{\eta}} = J\,\boldsymbol{\eta}+\mathcal{O}(|\boldsymbol{\eta}|^2), \qquad J_{ij}=\left.\frac{\partial f_i}{\partial x_j}\right|_{\mathbf{x}^*} \qquad (6)

where: $J$ — the Jacobian matrix, the local linearization of the flow. Its eigenvalues sort fixed points into nodes, saddles, spirals, and centres. The machinery — how to compute, when to trust it, the full trace–determinant taxonomy — belongs to the next two articles, Linear Stability Analysis and the Jacobian and Classification of Fixed Points: Nodes, Saddles, Spirals, Centers; for portrait-sketching purposes you need only the output: local pictures to paste at each nullcline intersection. The final structural quantity is the one that separates the two halves of mechanics. The divergence of the field measures the fractional rate at which the flow inflates or deflates phase area:

\nabla\cdot\mathbf{f}=\frac{\partial f_1}{\partial x}+\frac{\partial f_2}{\partial y} \qquad (7)

where: $\nabla\cdot\mathbf{f}>0$ — local expansion; $<0$ — contraction; $=0$ — area preserved. Hamiltonian systems sit identically at zero (the incompressible "phase fluid" of Hamiltonian Flow and the Phase Fluid), which is why they can never own an attractor. A damped system has $\nabla\cdot\mathbf{f}<0$ and must lose area — Derivation B turns this into the exponential law

A(t)=A(0)\,e^{\,\overline{\nabla\cdot\mathbf{f}}\;t} \qquad (8)

where: for a region swept along the flow with $\nabla\cdot\mathbf{f}$ constant (the common case below), the exponent is exact: a patch of initial conditions for the damped oscillator shrinks by $e^{-2\gamma t}$ , funnelling uncountably many pasts toward one future.

You want to know	Reach for	Because
where the flow can rest	fixed points, eq. (4)	only solutions that sit still
flow direction everywhere, fast	nullclines, eq. (5)	signs of $(\dot x,\dot y)$ frozen between them
the local picture at each rest point	Jacobian, eq. (6)	eigenvalues classify the neighbourhood
whether attractors are even possible	divergence, eq. (7)	area contraction is the license for attraction
the shape of an orbit, no clock	slope equation (3)	time eliminated, geometry remains

Derivation AHow damping redraws the pendulum portrait

We take the best-loved conservative portrait in mechanics — the pendulum's nested loops, cat's-eyes, and separatrix, dissected in the article The Pendulum Phase Portrait and Separatrix — and switch on friction. The goal is to see, in one controlled example, every general mechanism of this article at work: energy becoming a strictly decreasing function, orbits crossing the old level curves inward, the centre converting to a spiral attractor while the saddle survives, and the separatrix reborn as a basin boundary. The strategy is not to solve anything: we compute the sign of one derivative and let geometry do the rest. Watch for the fact that the saddle's character is untouched by damping — instability of that kind is not a matter of energy budget.

Step 1. Write the damped pendulum $\ddot{\theta}=-\omega_0^2\sin\theta-2\gamma\dot{\theta}$ (with $\omega_0^2=g/\ell$ , damping rate $\gamma>0$ ) as a planar system: $\dot{\theta}=y$ , $\dot{y}=-\omega_0^2\sin\theta-2\gamma y$ . Fixed points still sit at $y=0$ , $\sin\theta=0$ : the hanging state $(0,0)$ and the inverted state $(\pi,0)$ . Damping moves no equilibria — it only changes traffic between them.

Step 2. Recall the conservative skeleton. At $\gamma=0$ the scaled energy $E(\theta,y)=\tfrac12 y^2-\omega_0^2\cos\theta$ is constant on orbits, giving closed loops for $E<\omega_0^2$ , whirling curves for $E>\omega_0^2$ , and the separatrix $y=\pm 2\omega_0\cos(\theta/2)$ at exactly $E=\omega_0^2$ . Why quote rather than re-derive: that level-curve analysis is the Hamiltonian story, owned by the companion articles; our subject is precisely what invalidates it.

Step 3. Differentiate the energy along the damped flow: $\dot{E}=y\,\dot{y}+\omega_0^2\sin\theta\;\dot{\theta} = y(-\omega_0^2\sin\theta-2\gamma y)+\omega_0^2 y\sin\theta = -2\gamma y^2$ . The conservative terms cancel identically; only the friction term survives. Why this is the whole theorem: $\dot{E}\le 0$ everywhere, and $\dot{E}<0$ whenever the pendulum is actually moving ( $y\neq 0$ ). Energy is now a one-way clock.

Step 4. Read the geometry. A trajectory can only descend through the family of old energy contours; each crossing is transversal and inward. No closed orbit can exist in the plane region where $y\neq0$ somewhere on it — going around a loop would return $E$ to its starting value, impossible when $E$ strictly decreases along every moving arc.

Step 5. Check the fixed points with the Jacobian $J=\begin{pmatrix}0&1\\-\omega_0^2\cos\theta^*&-2\gamma\end{pmatrix}$ . At the hanging state: $\operatorname{tr}J=-2\gamma<0$ , $\det J=\omega_0^2>0$ — the conservative centre becomes a stable spiral (underdamped) or stable node (overdamped). At the inverted state: $\det J=-\omega_0^2<0$ — a saddle, exactly as without damping. Why the saddle is immune: a saddle's instability comes from the sign of $\det J$ , which damping (entering only the trace) cannot touch. Energy arguments confirm it — the inverted state is an energy maximum, and losing energy cannot pin you to a maximum.

Step 6. Assemble the global portrait. Whirling solutions lose $2\gamma y^2$ per unit time, so over-the-top motion always decays until capture; every trajectory except a measure-zero set ends at a hanging state. The old separatrix is replaced by the stable manifolds of the saddles, which now spiral outward in backward time and partition the cylinder into basins: initial conditions on one side settle after $n$ full turns, on the other side after $n+1$ . Verified: at $\gamma\to 0$ every statement degenerates correctly — $\dot E\to 0$ restores level-curve orbits, the spiral's eigenvalues $-\gamma\pm i\sqrt{\omega_0^2-\gamma^2}$ collapse to the centre's $\pm i\omega_0$ , and the basin boundaries fuse back into the separatrix $y=\pm2\omega_0\cos(\theta/2)$ . Matches the damped-pendulum portrait of Strogatz section 6.7.

Go deeper · Derivation B · 2 minThe area transport law from the flow map

Derivation BThe area transport law from the flow map

We prove the claim hiding inside equations (7) and (8): the divergence of the vector field is the exact fractional growth rate of phase-plane area carried by the flow. This is the planar, general- $\mathbf{f}$ version of Liouville's theorem, and it is the single computation that splits dynamics into conservative and dissipative worlds. The route: follow an infinitesimal patch of initial conditions, express its area through the Jacobian determinant of the flow map, and differentiate that determinant with Abel's identity. The subtlety to watch is that two different Jacobians appear — the Jacobian of the flow map at finite time, and the Jacobian $J$ of the field itself — connected by a differential equation.

Step 1. Let $\Phi_t(\mathbf{x}_0)$ be the flow map: the state at time $t$ of the trajectory that started at $\mathbf{x}_0$ . A small patch of initial conditions of area $A(0)$ around $\mathbf{x}_0$ is carried to a patch of area $A(t)=|\det M(t)|\,A(0)$ , where $M(t)=\partial\Phi_t/\partial\mathbf{x}_0$ is the deformation matrix of the map.

Step 2. Differentiate the defining ODE with respect to initial conditions: $\dot{M}=J(\Phi_t(\mathbf{x}_0))\,M$ , with $M(0)=I$ . The deformation matrix is dragged along the trajectory by the local Jacobian of the field. Why the chain rule gives this: $\frac{\mathrm{d}}{\mathrm{d}t}\frac{\partial \Phi_t}{\partial \mathbf{x}_0} = \frac{\partial \dot{\Phi}_t}{\partial \mathbf{x}_0} = \frac{\partial \mathbf{f}(\Phi_t)}{\partial \mathbf{x}_0} = J(\Phi_t)\,\frac{\partial \Phi_t}{\partial \mathbf{x}_0}$ — differentiation with respect to $t$ and $\mathbf{x}_0$ commute for smooth $\mathbf{f}$ .

Step 3. Apply Abel's identity (Jacobi's formula) for the determinant of a matrix solution: $\frac{\mathrm{d}}{\mathrm{d}t}\det M = \operatorname{tr}\!\big(J\big)\,\det M$ . Since $\operatorname{tr}J=\nabla\cdot\mathbf{f}$ evaluated at the moving point, the area obeys $\dot{A}=(\nabla\cdot\mathbf{f})\,A$ — the transport law.

Step 4. Integrate along the trajectory: $A(t)=A(0)\exp\!\int_0^t \nabla\cdot\mathbf{f}\,\mathrm{d}t'$ . When the divergence is a constant $c$ , this is a clean exponential $A(t)=A(0)e^{ct}$ .

Step 5. Evaluate the two flagship cases. Damped linear oscillator, $\dot{x}=y$ , $\dot{y}=-\omega_0^2x-2\gamma y$ : $\nabla\cdot\mathbf{f}=\partial_x(y)+\partial_y(-\omega_0^2x-2\gamma y)=-2\gamma$ , so $A(t)=A(0)e^{-2\gamma t}$ — every patch of pasts funnels exponentially toward the attractor. Hamiltonian system, $\dot q=\partial H/\partial p$ , $\dot p=-\partial H/\partial q$ : $\nabla\cdot\mathbf{f}=\partial_q\partial_p H-\partial_p\partial_q H=0$ identically — the incompressible phase fluid, no funnelling, no attractors, ever. Verified: for the damped oscillator the eigenvalues of $J$ are $\lambda_\pm=-\gamma\pm\sqrt{\gamma^2-\omega_0^2}$ , whose sum is $-2\gamma=\operatorname{tr}J$ — the two eigencontractions compound to exactly the area law. The Hamiltonian case reproduces Liouville's theorem as derived in Hamiltonian Flow and the Phase Fluid.

Figure

A patch of initial conditions carried by the flow — dissipative flows shrink it exponentially, Hamiltonian flows only reshape it

Limiting cases

Case	Result	Meaning
Near any hyperbolic fixed point	$\dot{\boldsymbol{\eta}}=J\boldsymbol{\eta}$	the local portrait is the linear one — nodes, saddles, spirals per the eigenvalues of $J$
Conservative limit, $\nabla\cdot\mathbf{f}=0$	orbits on level curves of $E$	contour-map portraits: centres and saddles only, no attractors — the regime of the Hamiltonian articles
Strong damping, $\gamma\gg\omega_0$	$\dot{x}\approx -V'(x)/2\gamma m$	inertia drops out; the plane effectively collapses onto a one-dimensional flow down the potential
Weak damping, $\gamma\ll\omega_0$	slow spiral through level curves	motion is quasi-conservative over one turn; energy drifts by $\Delta E\approx-2\gamma\langle y^2\rangle T$ per period
Three or more dimensions	Poincaré–Bendixson fails	trajectories can wander forever without settling — strange attractors and chaos become possible

Worked examples

Example 1 — the sketching workflow, end to end (confidence builder). Sketch the portrait of $\dot{x}=x+e^{-y}$ , $\dot{y}=-y$ .

Step 1. Fixed points: $\dot{y}=0$ forces $y=0$ ; then $\dot{x}=0$ forces $x=-e^{0}=-1$ . Unique fixed point $(-1,0)$ .

Step 2. Nullclines: the $y$ -nullcline is the line $y=0$ (flow purely horizontal there); the $x$ -nullcline is the curve $x=-e^{-y}$ (flow purely vertical). They cross only at $(-1,0)$ .

Step 3. Jacobian: $J=\begin{pmatrix}1&-e^{-y}\\0&-1\end{pmatrix}$ , so at $(-1,0)$ : $J=\begin{pmatrix}1&-1\\0&-1\end{pmatrix}$ . Triangular, so the eigenvalues are read off the diagonal: $\lambda_1=+1$ , $\lambda_2=-1$ . Opposite signs — a saddle.

Step 4. Eigenvectors: for $\lambda_1=1$ : $(J-I)\mathbf{v}=0$ gives $-v_y=0$ — the unstable direction is $(1,0)$ , along the $x$ -axis. For $\lambda_2=-1$ : $(J+I)\mathbf{v}=0$ gives $2v_x-v_y=0$ — the stable direction is $(1,2)$ .

Step 5. Global assembly: far to the right $\dot{x}\approx x>0$ (escape east); $y$ always relaxes toward $0$ since $\dot y=-y$ . Trajectories slide down toward the $x$ -axis while being expelled along it, entering along $(1,2)$ and leaving along $(1,0)$ . The portrait is a single saddle at $(-1,0)$ with stable direction $(1,2)$ and unstable direction $(1,0)$ — matching Strogatz's Example 6.1.1.

Example 2 — damped double-well oscillator, full census (exam level). A particle obeys $\ddot{x}+2\gamma\dot{x}+x-x^3=0$ with $\gamma=0.1$ . Locate and classify all fixed points, give the spiral's period and per-turn decay, and describe the basins.

Step 1. Planar form: $\dot{x}=y$ , $\dot{y}=-x+x^3-2\gamma y$ . Fixed points at $y=0$ , $x-x^3=0$ : three of them, $(0,0)$ and $(\pm1,0)$ .

Step 2. Jacobian: $J=\begin{pmatrix}0&1\\-1+3x^2&-0.2\end{pmatrix}$ .

Step 3. At the origin: $\det J=\Delta=1>0$ , $\operatorname{tr}J=\tau=-0.2<0$ , and $\tau^2-4\Delta=0.04-4<0$ : a stable spiral. Eigenvalues $\lambda=-0.1\pm i\sqrt{1-0.01}=-0.1\pm0.995\,i$ .

Step 4. At $(\pm1,0)$ : $-1+3x^2=2$ , so $\Delta=-2<0$ : saddles regardless of damping. Eigenvalues $\lambda=\tfrac12\left(-0.2\pm\sqrt{0.04+8}\right)=\tfrac12(-0.2\pm2.835)$ , i.e. $\lambda_+=+1.318$ , $\lambda_-=-1.518$ .

Step 5. Spiral clock: angular frequency $0.995$ , so the period is $T=2\pi/0.995=6.32$ time units. Amplitude decay per turn: $e^{-0.1\times6.32}=e^{-0.632}=0.53$ — each loop comes in at 53% of the previous radius.

Step 6. Area budget: $\nabla\cdot\mathbf{f}=-2\gamma=-0.2$ everywhere, so any patch of initial conditions shrinks by $e^{-0.2\times6.32}=0.28$ per spiral period — consistent with $\lambda_++\lambda_-$ at each point summing to $-0.2$ .

Step 7. Global structure: the potential $V(x)=x^2/2-x^4/4$ is a well guarded by two hilltops at $x=\pm1$ . The origin is the only attractor; its basin is the region bounded by the stable manifolds of the two saddles (the damped remnant of the conservative separatrix), and initial conditions outside that boundary escape to infinity past a hilltop. In the $\gamma\to0$ limit the origin's eigenvalues become $\pm i$ (a centre) and the saddles' become $\pm\sqrt2=\pm1.414$ — the conservative census of $\ddot{x}+x-x^3=0$ .

Origins · 1 minHenri Poincaré and how this was discovered

Henri Poincaré

1854–1912 · Sorbonne and the Bureau des Longitudes, Paris

Poincaré invented this article. His 1881 memoir on curves defined by differential equations proposed, for the first time, studying the qualitative geometry of all solutions at once — fixed points, their classification into nodes, saddles, foci and centres, limit cycles, the index of a curve — when explicit solution is impossible. He built the machinery for a purpose: celestial mechanics, where his work on the three-body problem (the prize memoir of 1889 and the three volumes of Méthodes nouvelles) uncovered homoclinic tangles, the first glimpse of deterministic chaos, seventy years before computers could draw one. Nearly every object in this chapter — the phase portrait, the return map, the recurrence theorem, the very idea that topology constrains dynamics — carries his fingerprints. Mathematics' last universalist, he wrote with equal authority on non-Euclidean geometry, topology (the Poincaré conjecture), the foundations of special relativity, and the philosophy of science.

History

1881

— Poincaré's memoir on curves defined by differential equations founds the qualitative theory: portraits instead of formulas

1892

— Lyapunov's doctoral thesis gives stability its rigorous definitions, independent of linearization

1899

— Poincaré's Méthodes nouvelles closes with homoclinic tangles in the three-body problem — chaos glimpsed on paper

1927

— van der Pol observes stable self-sustained oscillations in triode circuits; Andronov soon identifies them with Poincaré's limit cycles

1963

— Lorenz computes a three-dimensional flow whose trajectories settle onto a strange attractor, ending the plane's monopoly on intuition

Going further

References. Strogatz, Nonlinear Dynamics and Chaos, 2nd ed., Ch. 5–6 · Taylor, Classical Mechanics, Ch. 12 · Morin, Introduction to Classical Mechanics, oscillations chapters · Goldstein, Classical Mechanics, 3rd ed., Ch. 8 (Hamiltonian phase flow) · Landau & Lifshitz, Mechanics (Vol. 1), §11–14

Provenance. Equations and worked examples follow Strogatz (Nonlinear Dynamics and Chaos, Ch. 5–6), Taylor (Classical Mechanics, Ch. 12), and Goldstein (Ch. 8–9). A verifier should recompute: the pendulum energy integral and separatrix $y=\pm2\omega_0\cos(\theta/2)$ ; the Jacobian eigenvalues $\pm i\omega_0$ (center) and $\pm\omega_0$ (saddle); the divergence $\nabla\cdot\mathbf{f}=-2\gamma$ and area decay $A(t)=A(0)e^{-2\gamma t}$ ; and the fixed-point classification of $\ddot{x}+x-x^3=0$ . Small-oscillation frequency $\omega_0=\sqrt{g/\ell}$ is standard.

Builds on: Fixed Points and Equilibria · Ordinary differential equations and the existence–uniqueness theorem · Energy and potential in 1-D motion Leads to: Linear Stability Analysis and the Jacobian · Classification of Fixed Points: Nodes, Saddles, Spirals, Centers · Conservative Systems and Hamiltonian Phase Flow · Limit Cycles and Self-Sustained Oscillations · Poincaré–Bendixson Theorem

If phase portraits contain no time axis, how can they tell me anything about how long motions take?

Timing hides in the arrow lengths. The speed of the state along its orbit is $|\mathbf{f}(\mathbf{x})|$ , so segments of an orbit through regions of small $|\mathbf{f}|$ are traversed slowly; near a fixed point the crawl becomes arbitrarily slow, and a trajectory approaching a saddle along its stable direction takes logarithmically diverging time to pass it. You can always reconstruct elapsed time explicitly as $t=\int \mathrm{d}x/f_1(x,y)$ along the orbit. What the portrait genuinely discards is phase — which point of the journey you occupy at $t=0$ — and that is exactly the information that qualitative questions (settle? oscillate? escape?) never needed.

Why is the reduction to first-order systems more than a notational trick?

Because it converts the geometry of solutions into the geometry of a static object. A second-order equation's solutions live in a space of functions; after setting $y=\dot x$ , all solutions become integral curves of one fixed vector field in a finite-dimensional space, where theorems of topology apply: uniqueness forbids crossings, planarity constrains limit sets (Poincaré–Bendixson), the index of a closed curve counts equilibria inside it. Every hard theorem of this chapter is proved on the first-order form. It also exposes the true dimensionality of a problem: the driven pendulum, which looks two-dimensional, is genuinely three-dimensional once the drive phase is included as a state — which is precisely why it can be chaotic while the undriven one cannot.

The damped pendulum's energy argument shows no closed orbits can exist. But the van der Pol oscillator is dissipative and has one. What gives?

The pendulum's friction has a fixed sign: $\dot E=-2\gamma y^2\le0$ everywhere, so energy is a strictly decreasing clock and no motion can return to a previous state. The van der Pol circuit's "friction" coefficient $\mu(x^2-1)$ changes sign: it drains energy at large amplitude but pumps energy in at small amplitude. A closed orbit becomes possible exactly where the pumping and draining balance over one cycle. That sign-changing dissipation is the defining mechanism of self-sustained oscillation, and the article Limit Cycles and Self-Sustained Oscillations builds the whole theory on it. The general lesson: monotone energy functions forbid recurrence; non-monotone ones license it.

How do I know my hand-sketched portrait is right, when I have not solved anything?

Run the consistency checklist the professionals use. Every fixed point must sit at a nullcline intersection, and its local sketch must match the Jacobian's eigenvalues. Arrows must be purely vertical on the $x$ -nullcline and purely horizontal on the $y$ -nullcline, and their sign must be constant within each region the nullclines bound. No two curves may cross; mechanical-plane loops must run clockwise; and if $\nabla\cdot\mathbf{f}<0$ everywhere, your sketch must contain no closed orbits and no unstable spirals feeding outward forever without arriving somewhere. Most sketching errors — an orbit crossing a separatrix, a spiral with the wrong handedness — die on one of these checks. When in doubt, integrate a single trajectory numerically and see whether it respects your skeleton.

Are nullclines the same as separatrices? Both get drawn as special curves.

No, and confusing them is the most common sketching error after crossed trajectories. Nullclines are construction lines: loci where one velocity component vanishes, generally crossed by the flow, meaningful only as bookkeeping for signs and fixed points. Separatrices are dynamical objects: actual trajectories (stable or unstable manifolds of saddles, or boundaries between basins) that the flow can never cross, because they are themselves solutions. A nullcline can be crossed by every orbit in the plane; a separatrix by none. They intersect each other freely, and only at fixed points do they both pass through the same distinguished location. On a finished portrait, dash your nullclines and draw separatrices solid — your future self will thank you.

Why do textbooks obsess over planar systems when every interesting real system has many more dimensions?

Three reasons, in ascending order of honesty. First, the plane is where geometry can be complete: Poincaré–Bendixson pins down every possible long-term behaviour, so the taxonomy you learn here is exhaustive rather than anecdotal. Second, high-dimensional systems constantly reduce to planar ones — near equilibria, on slow manifolds, after symmetry quotients, or on a Poincaré section — so planar fluency is the working skill even in large models. Third, the plane calibrates your surprise: once you know that bounded planar flow must settle to equilibria or cycles, the wandering of the Lorenz attractor in three dimensions registers as the genuinely new phenomenon it is, rather than as noise. You learn the plane precisely to know what only higher dimensions can do.

Numerically computed portraits are cheap now. Is hand-sketching with nullclines obsolete?

The computer draws trajectories; it does not tell you why the portrait looks as it does, and it can actively mislead. Numerical integration near saddles is unreliable (trajectories hug the unstable manifold and shoot off), stiff systems mis-render spirals, and a finite grid of initial conditions can miss a thin basin or a half-stable cycle entirely. The hand workflow — fixed points, nullclines, Jacobian types, sign regions — produces the skeleton that tells you where to point the computer and which numerical artifacts to distrust. In research practice the two run together: sketch to form hypotheses about the skeleton, integrate to confirm and decorate. Skipping the sketch stage is how published papers end up with phase portraits whose arrows violate their own equations.

Does the clockwise rule hold for every phase plane I will meet?

No — it is a property of mechanical coordinates $(x,\dot x)$ , not of phase planes as such. It follows from the kinematic row $\dot x=y$ of the reduction: above the axis the state must move right, below it left, and this no model can evade. But general planar systems — predator–prey, chemical kinetics, the FitzHugh–Nagumo neuron — have two independent components with no kinematic linkage, and their orbits may circulate either way or not at all. Related fine print: the clockwise rule fixes orientation but not shape, and it survives damping and forcing so long as the coordinates remain position and its own derivative. On the pendulum's cylinder the rule holds locally but whirling orbits never close, so "circulation" refers to loops around fixed points, not around the cylinder.

Phase Portraits and the Phase Plane

Where this lives today

The physical picture

The core equations

Limiting cases

Worked examples

Going further

Before this

Where this leads