The Poincare-Bendixson Theorem

Can a bounded, forever-flowing trajectory in the plane wander erratically without ever settling down? For flows confined to two dimensions, the answer is no. This single fact is one of the deepest structural truths in dynamics. The Poincaré–Bendixson theorem says that a planar trajectory trapped in a compact region, with no equilibrium in the way, must approach a periodic orbit. It converts a question about the infinite-time future of a trajectory into a question about topology. Closed curves in the plane separate it into an inside and an outside. Nothing can cross a curve without leaving a trace. This theorem matters because it draws a sharp line between the tame world of planar dynamics and the wild world of three (or more) dimensions. In the plane, trajectories cannot cross themselves or each other. So the Jordan curve theorem constrains where they can go. In three dimensions this constraint vanishes. Chaotic, aperiodic, sensitive-dependence behavior (as in the Lorenz system) becomes possible. Poincaré–Bendixson is the rigorous reason chaos is impossible for continuous-time autonomous systems with two state variables. This is a fact engineers and biologists lean on constantly. They use it when modeling oscillators, predator–prey cycles, and electrical circuits. The big picture: the theorem is an existence theorem for Limit Cycles and Closed Orbits that requires no explicit construction of the orbit. Instead of solving the differential equation, one builds a "trapping region." This is a closed, bounded set that the flow enters but never leaves. One then shows it contains no fixed points. Compactness then forces the trajectory's omega-limit set to be a nonempty, connected, closed invariant set free of equilibria. The planar topology forces that set to be a periodic orbit. This article gives the full statement and a complete proof. The proof is built from the Jordan curve theorem and basic properties of limit sets. It includes two fully worked applications: a van der Pol–type oscillator and a chemical/ecological model. It also gives a counterexample showing why planarity is essential. Finally, it tours where the theorem is used today. This ranges from electronic oscillator design to the very argument that rules out two-dimensional chaos.

Key takeaways
  • The theorem applies only to autonomous planar ODEs x˙=f(x)\dot x = f(x), x∈R2x \in \mathbb{R}^2, with ff continuously differentiable.
  • If a trajectory stays in a closed, bounded region for all t≥0t \ge 0 and that region contains no equilibrium points, the trajectory limits onto a periodic orbit.
  • The proof rests on two pillars: properties of the omega-limit set (ω\omega-limit) and the Jordan curve theorem, which forbids trajectories from crossing themselves or each other.
  • It is the precise reason chaos cannot occur in continuous-time autonomous systems with only two state variables.
  • Practical use: build a trapping annulus (inner repelling boundary, outer attracting boundary) around an unstable equilibrium to prove a limit cycle exists without solving the ODE.
  • Key idea: in the plane, a trajectory trapped in a compact, equilibrium-free region has nowhere to go but around and around — so it must spiral onto a closed loop.

Where this lives today

Electrical engineering and oscillator design. The van der Pol equation, originally modeling vacuum-tube triode circuits, is the textbook application: Poincaré–Bendixson proves the existence of a stable limit cycle corresponding to a self-sustained oscillation, the mathematical backbone of every electronic clock and radio-frequency oscillator built from a negative-resistance element.

Mathematical biology and ecology. Predator–prey cycles (Lotka–Volterra variants with saturating functional responses, as in Example 2), neuron spiking models (FitzHugh–Nagumo), and biochemical oscillators (glycolytic oscillations, circadian rhythm models) are all proved to sustain periodic behavior via trapping-region arguments descending directly from this theorem.

Chemical kinetics. The Belousov–Zhabotinsky reaction's planar reduced models use Poincaré–Bendixson-style arguments to prove sustained chemical oscillation exists before ever solving the reaction-rate equations explicitly.

Foundations of chaos theory. The theorem is the exact reason chaotic, aperiodic, sensitive dynamics require at least three continuous state variables (or two dimensions plus explicit time-dependence, or a discrete map). It sits as the boundary condition referenced whenever a text introduces Chaos and Sensitive Dependence in the Lorenz or Rössler systems — "chaos needs three dimensions, and here is why."

Bifurcation theory. When a trapping-region argument shows a limit cycle is born as an equilibrium loses stability, this is exactly the qualitative content underlying the Hopf Bifurcation: Poincaré–Bendixson supplies the global existence of the cycle, while Hopf's theorem supplies the local, parameter-dependent birth mechanism near the critical eigenvalue crossing.

Control theory and robotics. Limit-cycle-based gaits for legged robots and self-stabilizing feedback controllers are frequently verified using explicit trapping-region ("invariant set") constructions modeled on the annulus argument here.

Ask the tutor: How do I explicitly construct a Dulac function to prove uniqueness of the limit cycle once Poincaré–Bendixson guarantees existence?

DefinitionPlanar autonomous system, trapping region, and limit cycle

Let D⊆R2D \subseteq \mathbb{R}^2 be open and let f:D→R2f: D \to \mathbb{R}^2 be C1C^1 (continuously differentiable). Consider the autonomous planar system

x˙=f(x),x=(x1,x2)∈D.\dot{x} = f(x), \qquad x = (x_1,x_2) \in D.

By the Picard–Lindelöf theorem, for each x0∈Dx_0 \in D there is a unique maximal solution φt(x0)\varphi_t(x_0) with φ0(x0)=x0\varphi_0(x_0) = x_0, called the flow. A set K⊂DK \subset D is called a trapping region (or positively invariant compact set) if KK is closed and bounded (compact) and

x0∈K  ⟹  φt(x0)∈K for all t≥0.x_0 \in K \implies \varphi_t(x_0) \in K \text{ for all } t \ge 0.

For x0∈Dx_0 \in D whose forward orbit γ+(x0)={φt(x0):t≥0}\gamma^+(x_0) = \{\varphi_t(x_0): t\ge 0\} is bounded, the omega-limit set is

ω(x0)={y∈D:∃ tn→∞ with φtn(x0)→y}.\omega(x_0) = \{ y \in D : \exists\, t_n \to \infty \text{ with } \varphi_{t_n}(x_0) \to y \}.

A periodic orbit (or closed orbit) is the image of a nonconstant solution φt(x0)\varphi_t(x_0) satisfying φT(x0)=x0\varphi_{T}(x_0) = x_0 for some minimal T>0T>0. A limit cycle is a periodic orbit Γ\Gamma that is the ω\omega-limit set (or α\alpha-limit set) of some point not on Γ\Gamma itself.

Intuition

Picture a planar vector field as water flowing across a tabletop. Suppose you build a fence — some closed, bounded region KK — such that water flowing in never flows back out: a trapping region. If additionally there is no drain (no equilibrium point) anywhere inside the fence, the water inside can never come to rest and can never escape. It swirls forever inside a bounded area, guided by a flow that (in the plane) cannot cross itself: two solution curves of an autonomous ODE cannot intersect, since uniqueness of solutions rules out two different futures from the same point. This "no self-crossing" property is the entire reason two dimensions are special: a wandering, non-repeating trajectory would have to weave through a bounded plane region forever without ever touching its own past — and topology (the Jordan curve theorem) says a closed injective curve that keeps almost-closing on itself is eventually forced to close exactly. The classic picture is an annulus: an unstable spiral at the center (so trajectories are pushed outward near the origin) surrounded by a large circle that trajectories cannot escape (so trajectories are pushed inward at the boundary). Any trajectory starting in the annulus is trapped there forever, and since the annulus contains no equilibrium, Poincaré–Bendixson guarantees a periodic orbit lives inside it.

Figure
A trapping annulus forces a limit cycle: inner boundary repels, outer boundary attracts, and any trajectory started between them is squeezed onto a periodic orbit Γ\Gamma.
O (equilibrium) outer boundary K (inflow) inner boundary (outflow) limit cycle Γ
TheoremPoincaré–Bendixson Theorem

Let D⊆R2D \subseteq \mathbb{R}^2 be open, let f:D→R2f: D \to \mathbb{R}^2 be C1C^1, and consider x˙=f(x)\dot x = f(x) on DD. Suppose M⊂DM \subset D is a nonempty compact (closed and bounded) set such that: 1. MM is positively invariant: x0∈M  ⟹  φt(x0)∈Mx_0 \in M \implies \varphi_t(x_0) \in M for all t≥0t \ge 0; 2. MM contains no equilibrium point of ff (no x∗∈Mx^* \in M with f(x∗)=0f(x^*) = 0). Then for any x0∈Mx_0 \in M, the omega-limit set ω(x0)\omega(x_0) is a periodic orbit contained in MM. Consequently, MM contains at least one periodic orbit of the system.

LemmaBasic properties of omega-limit sets

Let γ+(x0)\gamma^+(x_0) be bounded and contained in DD. Then ω(x0)\omega(x_0) is: (a) nonempty, (b) compact, (c) connected, and (d) invariant: if y∈ω(x0)y \in \omega(x_0) then φt(y)∈ω(x0)\varphi_t(y) \in \omega(x_0) for every t∈Rt \in \mathbb{R} for which φt(y)\varphi_t(y) is defined.

Proof

(Proof technique: sequential compactness + continuous dependence, plus a connectedness-by-contradiction splitting argument.)

(a) Nonempty. Since γ+(x0)\gamma^+(x_0) is bounded, its closure γ+(x0)‾\overline{\gamma^+(x_0)} is compact (Heine–Borel, in R2\mathbb{R}^2). The sequence φn(x0)\varphi_n(x_0), n=1,2,…n=1,2,\dots, lies in this compact set, so by Bolzano–Weierstrass it has a convergent subsequence φtn(x0)→y\varphi_{t_n}(x_0) \to y with tn=n→∞t_n = n \to \infty. Hence y∈ω(x0)y \in \omega(x_0).

(b) Compact. Boundedness of ω(x0)\omega(x_0) follows since ω(x0)⊂γ+(x0)‾\omega(x_0) \subset \overline{\gamma^+(x_0)}, which is bounded. For closedness, let yk∈ω(x0)y_k \in \omega(x_0) with yk→yy_k \to y. For each kk choose tkt_k with ∣φtk(x0)−yk∣<1/k|\varphi_{t_k}(x_0) - y_k| < 1/k and tk→∞t_k \to \infty (possible by definition of ω\omega-limit point, choosing tkt_k large enough at each stage). Then φtk(x0)→y\varphi_{t_k}(x_0) \to y, so y∈ω(x0)y \in \omega(x_0). Thus ω(x0)\omega(x_0) is closed and bounded, hence compact.

(c) Connected. Suppose for contradiction ω(x0)=A∪B\omega(x_0) = A \cup B with A,BA, B nonempty, disjoint, and closed (hence compact, being closed subsets of a compact set). Since A,BA,B are disjoint compact sets, d(A,B)=δ>0d(A,B) = \delta > 0 for some δ\delta (the distance function x↦dist⁡(x,B)x \mapsto \operatorname{dist}(x,B) is continuous and positive on the compact set AA, so it attains a positive minimum). Since A,B⊂ω(x0)A, B \subset \omega(x_0), the trajectory φt(x0)\varphi_t(x_0) returns infinitely often within δ/3\delta/3 of AA and infinitely often within δ/3\delta/3 of BB, at arbitrarily large times. By continuity of t↦φt(x0)t \mapsto \varphi_t(x_0), as the trajectory alternates between these two neighborhoods, it must pass through the intermediate annular region {δ/3≤dist⁡(x,A)≤2δ/3}\{\delta/3 \le \operatorname{dist}(x,A) \le 2\delta/3\} at a sequence of times sn→∞s_n \to \infty (intermediate value theorem applied to the continuous function t↦dist⁡(φt(x0),A)t \mapsto \operatorname{dist}(\varphi_t(x_0), A)). Since γ+(x0)\gamma^+(x_0) lies in a compact set, the sequence φsn(x0)\varphi_{s_n}(x_0) has a convergent subsequence, whose limit zz satisfies δ/3≤dist⁡(z,A)≤2δ/3\delta/3 \le \operatorname{dist}(z, A) \le 2\delta/3, so z∉Az \notin A and z∉Bz \notin B (as dist⁡(z,B)≥δ−2δ/3=δ/3>0\operatorname{dist}(z,B) \geq \delta - 2\delta/3 = \delta/3 >0). But z∈ω(x0)z \in \omega(x_0) by construction, contradicting ω(x0)=A∪B\omega(x_0) = A \cup B. Hence ω(x0)\omega(x_0) is connected.

(d) Invariant. Let y∈ω(x0)y \in \omega(x_0), so φtn(x0)→y\varphi_{t_n}(x_0) \to y for some tn→∞t_n \to \infty. Fix t∈Rt \in \mathbb{R}. By continuous dependence of solutions on initial conditions (a consequence of the Picard–Lindelöf theorem and Grönwall's inequality), φt(⋅)\varphi_t(\cdot) is continuous where defined, so

φtn+t(x0)=φt(φtn(x0))→φt(y).\varphi_{t_n + t}(x_0) = \varphi_t(\varphi_{t_n}(x_0)) \to \varphi_t(y).

Since tn+t→∞t_n + t \to \infty, this shows φt(y)∈ω(x0)\varphi_t(y) \in \omega(x_0). ■\blacksquare

∎
Proof

Proof of the Poincaré–Bendixson Theorem. (Proof technique: transversal (Poincaré section) construction + Jordan curve theorem to force monotonicity of return points, then a limiting argument.) Fix x0∈Mx_0 \in M. Since MM is compact and positively invariant, γ+(x0)⊂M\gamma^+(x_0) \subset M is bounded, so by the Lemma above ω:=ω(x0)\omega := \omega(x_0) is nonempty, compact, connected, and invariant; moreover ω⊂M\omega \subset M (as MM is closed and contains γ+(x0)‾\overline{\gamma^+(x_0)}). Since MM has no equilibria, ω\omega contains no equilibrium of ff.

Step 1: Pick a point p∈ωp \in \omega and use invariance. Since ω≠∅\omega \ne \emptyset, pick p∈ωp \in \omega. By invariance, the entire orbit φt(p)\varphi_t(p) for t∈Rt \in \mathbb{R} stays in ω\omega. Since ω\omega is bounded, this orbit is bounded for all time (both directions), so its own omega-limit set ω(p)\omega(p) is nonempty (by part (a) of the Lemma applied to the point pp); moreover ω(p)⊆ω\omega(p) \subseteq \omega since ω\omega is closed and invariant (any limit point of {φt(p)}\{\varphi_t(p)\} is a limit of points in ω\omega, hence in ω\omega by closedness). Pick q∈ω(p)⊂ωq \in \omega(p) \subset \omega.

Step 2: Build a transversal at qq. Since q∈Mq \in M and MM has no equilibria, f(q)≠0f(q) \ne 0. Choose a short line segment Σ\Sigma through qq, transverse to ff: that is, Σ\Sigma is a smooth arc with qq in its interior such that f(x)f(x) is never tangent to Σ\Sigma for x∈Σx \in \Sigma (this is possible by continuity of ff: take Σ\Sigma perpendicular to f(q)f(q) and small enough that ff stays within a small angle of f(q)f(q) along Σ\Sigma, using C1C^1-continuity of ff). A standard consequence (the "flow-box" or tubular neighborhood theorem, itself proved via the implicit function theorem applied to the flow map) is that all trajectories crossing Σ\Sigma near qq cross it in the same direction, and any trajectory returning near Σ\Sigma does so in a sequence of points that is monotonic along Σ\Sigma — this monotonicity is the key topological fact and is justified in Step 3.

Step 3: Monotonicity via the Jordan Curve Theorem. Since q∈ω(p)q \in \omega(p), the orbit of pp meets Σ\Sigma in a sequence of points y1,y2,y3,…y_1, y_2, y_3, \dots converging to qq, at increasing times t1<t2<⋯→∞t_1 < t_2 < \cdots \to \infty. Consider consecutive crossings yn,yn+1y_n, y_{n+1}. Join yny_n to yn+1y_{n+1} by the arc of the orbit of pp between times tnt_n and tn+1t_{n+1}, and close this arc up with the sub-segment of Σ\Sigma from yn+1y_{n+1} to yny_n. This produces a Jordan curve JJ (a closed curve without self-intersection: the orbit arc cannot self-intersect because solutions of an autonomous ODE are unique — two trajectory pieces through the same point must coincide — and the orbit arc meets Σ\Sigma only at its two endpoints by the transversality and monotonic-crossing property of Σ\Sigma established via the flow-box theorem). By the Jordan curve theorem, JJ separates the plane into a bounded interior int⁡(J)\operatorname{int}(J) and unbounded exterior. Because ff is transverse to Σ\Sigma and points consistently to one side, the vector field along the segment portion of JJ points entirely into (or entirely out of) int⁡(J)\operatorname{int}(J); along the orbit-arc portion of JJ, the vector field is tangent to JJ by construction (it is the flow direction). Hence trajectories crossing Σ\Sigma can only cross from the exterior side to the interior side (or vice versa, depending on orientation) at every subsequent crossing, since to reverse direction the trajectory would have to cross the orbit-arc part of JJ transversally, which is impossible (uniqueness of solutions: it cannot cross its own earlier path). This forces the sequence of crossing points y1,y2,y3,…y_1, y_2, y_3, \dots to be monotonic along Σ\Sigma (each subsequent crossing lies further along Σ\Sigma in the same direction, trapped between the previous crossing and the boundary of the region carved out).

Step 4: The monotone sequence converges, forcing periodicity. A monotonic sequence of points on the compact segment Σ\Sigma converges; here it converges to qq by hypothesis. Now reconsider the original trajectory φt(x0)\varphi_t(x_0). Since q∈ω(x0)q \in \omega(x_0) as well (because ω(p)⊂ω(x0)\omega(p) \subset \omega(x_0), as p∈ω(x0)p\in\omega(x_0) and ω(x0)\omega(x_0) is closed and invariant, by the same argument as in Step 1), the orbit through x0x_0 also crosses Σ\Sigma in a sequence of points converging monotonically to qq. We now show the orbit through qq itself is periodic. Since q∈ω(p)⊆ωq \in \omega(p) \subseteq \omega, and ω\omega is invariant, the full orbit of qq lies in ω⊆M\omega \subseteq M, which is equilibrium-free, so f(q)≠0f(q) \neq 0 and the orbit of qq is non-constant. Consider the forward orbit of qq: since q∈ω(p)q\in\omega(p), the orbit of pp returns arbitrarily close to qq infinitely often, meaning q∈ω(q)q \in \omega(q) is forced as follows. If the orbit of qq were not periodic, it would cross Σ\Sigma again at some point q′≠qq' \ne q; but by the same monotonicity argument applied to the orbit of qq (using that qq's orbit lies in ω\omega, whose omega limit set relative to itself must also be nonempty and reachable through Σ\Sigma-crossings converging to a single point by the squeezing/monotonicity established in Step 3), any subsequent crossing of the orbit of qq with Σ\Sigma must equal qq exactly — otherwise the strict monotonicity forces the crossings of pp's orbit (which interlace with and are squeezed by the orbit of qq, since q∈ω(p)q\in \omega(p) means pp's orbit accumulates on qq's entire orbit by invariance) to converge to two different points, contradicting convergence to the single point qq. Hence the orbit of qq returns to qq itself: there is T>0T>0 with φT(q)=q\varphi_T(q) = q, i.e. the orbit of qq is a periodic orbit Γ\Gamma.

Step 5: ω(x0)=Γ\omega(x_0) = \Gamma. We have shown Γ⊂ω⊂M\Gamma \subset \omega \subset M. It remains to show ω(x0)\omega(x_0) is exactly Γ\Gamma and not larger. Since the crossings of x0x_0's orbit with Σ\Sigma converge monotonically to q∈Γq \in \Gamma, and between consecutive crossings the orbit stays trapped in the shrinking Jordan-curve regions built in Step 3 (which shrink down onto Γ\Gamma by the same squeezing argument), the distance from φt(x0)\varphi_t(x_0) to Γ\Gamma tends to 00 as t→∞t \to \infty. Formally: for any ε>0\varepsilon>0, the flow-box neighborhood of Σ\Sigma combined with continuous dependence on initial conditions shows that once a crossing point yny_n is within ε′\varepsilon' of qq (for suitably small ε′\varepsilon' depending on ε\varepsilon and the period TT, by continuity of the time-TT flow map), the entire arc of orbit from yny_n to yn+1y_{n+1} stays within ε\varepsilon of Γ\Gamma. Hence every limit point of φt(x0)\varphi_t(x_0) as t→∞t\to\infty lies on Γ\Gamma, i.e. ω(x0)⊆Γ\omega(x_0) \subseteq \Gamma. Combined with Γ⊆ω(x0)\Gamma \subseteq \omega(x_0) shown in Step 4 (since Γ\Gamma's points, in particular qq, arose as limit points of x0x_0's own orbit through the chain q∈ω(p)⊂ω(x0)q\in\omega(p)\subset\omega(x_0), and invariance of ω(x0)\omega(x_0) then forces the whole orbit Γ\Gamma of qq into ω(x0)\omega(x_0)), we conclude

ω(x0)=Γ,\omega(x_0) = \Gamma,

a periodic orbit contained in MM. ■\blacksquare

∎

Worked examples

Example 1: A van der Pol–type oscillator via Liénard's construction. Consider the system on R2\mathbb{R}^2:

x˙=y,y˙=−x+y(1−x2−2y2).\dot x = y, \qquad \dot y = -x + y(1 - x^2 - 2y^2).

This is an asymmetric variant (the coefficient 22 on y2y^2 breaks the usual van der Pol symmetry) designed so no shortcut symmetry argument applies. Step 1 — locate and classify the only equilibrium. Setting x˙=y˙=0\dot x = \dot y = 0: from x˙=0\dot x=0, y=0y=0; substituting into y˙=0\dot y = 0 gives −x=0-x = 0, so x=0x=0. The unique equilibrium is the origin. The Jacobian at the origin, linearizing y˙=−x+y−x2y−2y3\dot y = -x + y - x^2 y - 2y^3, is

Df(0,0)=(01−11),Df(0,0) = \begin{pmatrix} 0 & 1 \\ -1 & 1 \end{pmatrix},

with eigenvalues solving λ2−λ+1=0\lambda^2 - \lambda + 1 = 0, i.e. λ=1±i32\lambda = \frac{1 \pm i\sqrt3}{2}. Since Re⁡(λ)=1/2>0\operatorname{Re}(\lambda) = 1/2 > 0, the origin is an unstable spiral: nearby trajectories spiral outward. Step 2 — outward flux near the origin. Using polar coordinates x=rcos⁡θ,y=rsin⁡θx = r\cos\theta, y=r\sin\theta, compute

rr˙=xx˙+yy˙=xy+y(−x+y(1−x2−2y2))=y2(1−x2−2y2).r\dot r = x\dot x + y \dot y = xy + y(-x + y(1-x^2-2y^2)) = y^2(1 - x^2 - 2y^2).

For small rr (say r<0.5r < 0.5), x2+2y2≤2r2<0.5x^2 + 2y^2 \le 2r^2 < 0.5, so 1−x2−2y2>0.51 - x^2-2y^2 > 0.5, giving rr˙≥0r\dot r \ge 0 whenever y≠0y \ne 0, and rr˙=0r\dot r = 0 only momentarily when y=0y=0 (a measure-zero, non-accumulating set of times since y˙≠0\dot y \ne 0 there generically). Averaged over an orbit, rr strictly increases near 00; concretely, choose the inner circle C1:r=0.1C_1: r = 0.1. On C1C_1, rr˙=y2(1−x2−2y2)≥0r\dot r = y^2(1-x^2-2y^2) \ge 0 with equality only at isolated points y=0y=0, and one checks the flow is not tangent to C1C_1 along the entire circle, so trajectories starting on C1C_1 move strictly outward (this can also be seen directly: the origin is a repelling spiral, so a small circle around it is crossed strictly outward, a standard fact for spiral sources). Step 3 — inward flux on a large circle. Take C2:r=RC_2: r = R for RR large, say R=3R=3. Then

rr˙=y2(1−x2−2y2).r \dot r = y^2(1 - x^2 - 2y^2).

On C2C_2, x2+y2=9x^2+y^2 = 9. We bound x2+2y2=9+y2≥9x^2 + 2y^2 = 9 + y^2 \ge 9 (since x2+2y2=(x2+y2)+y2x^2+2y^2 = (x^2+y^2)+y^2), so 1−x2−2y2≤1−9=−8<01 - x^2-2y^2 \le 1-9 = -8 < 0 whenever y≠0y\neq0; at points where y=0y=0, rr˙=0r\dot r = 0 but r˙\dot r evaluated via xx˙+yy˙x\dot x + y\dot y at those instants is momentary and the flow immediately re-enters the region y≠0y\neq 0 pushing rr down (since ¨\ddot{} analysis or a direct check of x˙=y=0, y˙=−x=∓3≠0\dot x = y=0,\ \dot y = -x = \mp 3 \ne 0 shows the trajectory does not stay on y=0y=0, so it is not tangent to C2C_2 there either). Hence trajectories starting on C2C_2 move strictly inward. Step 4 — apply the theorem. Let M={(x,y):0.1≤r≤3}M = \{ (x,y): 0.1 \le r \le 3\}, the closed annulus. By Steps 2–3, MM is positively invariant: the outer boundary is crossed inward and the inner boundary is crossed outward, so no trajectory starting in MM can leave through either boundary (a trajectory attempting to cross would have to be tangent to the boundary circle at the crossing point, contradicting the strict inward/outward flux just shown, except at isolated instantaneous points which do not allow escape since the vector field on either side still points into MM). The only equilibrium, the origin, lies outside MM (since r≥0.1>0r \geq 0.1 > 0 throughout MM). All hypotheses of the Poincaré–Bendixson theorem hold for MM: compact, positively invariant, equilibrium-free.

Conclusion: MM contains at least one periodic orbit. Since the origin is an unstable spiral, orbits starting arbitrarily close to it (but inside MM) spiral outward and their ω\omega-limit set — by the theorem — must be a periodic orbit inside the annulus 0.1≤r≤30.1 \le r \le 3. This exhibits a genuine limit cycle for an asymmetric, non-standard Liénard-type system, without ever solving the ODE explicitly.

Example 2: A predator–prey style system with a saturating (Holling type) response. Consider

x˙=x(1−x)−xy1+x,y˙=y(x1+x−0.4),\dot x = x(1-x) - \frac{xy}{1+x}, \qquad \dot y = y\left(\frac{x}{1+x} - 0.4\right),

on the biologically meaningful domain D={x>0,y>0}D = \{x>0, y>0\}, where xx is prey density and yy is predator density. Step 1 — find the coexistence equilibrium. Setting y˙=0\dot y = 0 with y>0y>0 requires x1+x=0.4\frac{x}{1+x} = 0.4, giving x=0.40.6=23x = \frac{0.4}{0.6} = \frac{2}{3}. Setting x˙=0\dot x = 0 with x=2/3>0x = 2/3 >0:

23(1−23)=(2/3)y1+2/3  ⟹  23⋅13=(2/3)y5/3  ⟹  29=2y5  ⟹  y=59.\frac{2}{3}\left(1-\frac23\right) = \frac{(2/3) y}{1+2/3} \implies \frac{2}{3}\cdot\frac13 = \frac{(2/3)y}{5/3} \implies \frac{2}{9} = \frac{2y}{5} \implies y = \frac{5}{9}.

So the interior equilibrium is E∗=(2/3,5/9)E^* = (2/3, 5/9), with both coordinates positive. Step 2 — linearize and check instability. Write x˙=xF(x)−xy1+x\dot x = xF(x) - \dfrac{xy}{1+x} with F(x)=1−xF(x)=1-x, and y˙=yG(x)\dot y = yG(x), G(x)=x1+x−0.4G(x) = \frac{x}{1+x}-0.4. Compute partials at E∗=(2/3,5/9)E^*=(2/3,5/9). ∂x˙/∂x=F(x)+xF′(x)−y(1+x)−xy(1+x)2=F(x)+xF′(x)−y(1+x)2\partial \dot x/\partial x = F(x) + xF'(x) - \frac{y(1+x) - xy}{(1+x)^2} = F(x)+xF'(x) - \frac{y}{(1+x)^2}. At x=2/3x=2/3: F=1/3F=1/3, F′=−1F'=-1, so F+xF′=1/3−2/3=−1/3F+xF' = 1/3 - 2/3 = -1/3. And y(1+x)2=5/9(5/3)2=5/925/9=15\frac{y}{(1+x)^2} = \frac{5/9}{(5/3)^2} = \frac{5/9}{25/9} = \frac15. So ∂x˙/∂x∣E∗=−13−15=−815\partial\dot x/\partial x|_{E^*} = -\frac13 - \frac15 = -\frac{8}{15}. ∂x˙/∂y=−x1+x=−2/35/3=−25\partial \dot x /\partial y = -\frac{x}{1+x} = -\frac{2/3}{5/3} = -\frac25. ∂y˙/∂x=yG′(x)\partial \dot y/\partial x = y G'(x), and G′(x)=1(1+x)2G'(x) = \frac{1}{(1+x)^2}, so at x=2/3x=2/3: G′=125/9=925G'=\frac{1}{25/9}=\frac9{25}. Thus ∂y˙/∂x∣E∗=59⋅925=15\partial\dot y/\partial x|_{E^*} = \frac59\cdot\frac9{25} = \frac15. ∂y˙/∂y=G(x)=0.4−0.4=0\partial \dot y/\partial y = G(x) = 0.4-0.4=0 (by construction of the equilibrium condition). So the Jacobian is

J(E∗)=(−8/15−2/51/50).J(E^*) = \begin{pmatrix} -8/15 & -2/5 \\ 1/5 & 0 \end{pmatrix}.

Trace =−8/15<0= -8/15 < 0... but we must check more carefully: for a genuine repelling spiral we'd want positive trace. Let's compute the actual determinant and trace:

tr⁡J=−815,det⁡J=(−815)(0)−(−25)(15)=225.\operatorname{tr} J = -\frac{8}{15}, \qquad \det J = \left(-\frac{8}{15}\right)(0) - \left(-\frac25\right)\left(\frac15\right) = \frac{2}{25}.

Here det⁡J=2/25>0\det J = 2/25 > 0 but tr⁡J=−8/15<0\operatorname{tr} J = -8/15 < 0, so eigenvalues have negative real part: E∗E^* is actually stable, not the repelling scenario we want for a limit cycle by this route. This shows the honest arithmetic: not every parameter choice gives an oscillator, and one must check. (This is itself instructive — see the Traps table below on verifying, not assuming, instability.) To keep the example genuinely illustrative of Poincaré–Bendixson, adjust the carrying-capacity parameter: replace 1−x1-x by 1−x/1.21 - x/1.2 so that the prey isocline is shallower (a standard destabilizing change, mimicking the paradox of enrichment). Recomputing with F(x)=1−x/1.2F(x) = 1-x/1.2, F′(x)=−1/1.2=−5/6F'(x)=-1/1.2 = -5/6:

F(2/3)=1−2/31.2=1−59=49,F+xF′=49+23(−56)=49−59=−19.F(2/3) = 1-\frac{2/3}{1.2} = 1-\frac{5}{9} = \frac49, \qquad F + xF' = \frac49 + \frac23\left(-\frac56\right) = \frac49-\frac59 = -\frac19.

Then ∂x˙/∂x∣E∗=−19−15=−545−945=−1445\partial \dot x/\partial x|_{E^*} = -\frac19 - \frac15 = -\frac{5}{45}-\frac9{45} = -\frac{14}{45}. This is still negative — but the key destabilizing mechanism in Rosenzweig–MacArthur models is a larger carrying capacity; take capacity K=3K=3 instead, i.e. F(x)=1−x/3F(x) = 1-x/3, F′=−1/3F'=-1/3:

F(2/3)=1−2/91=1−29=79,F+xF′=79+23(−13)=79−29=59.F(2/3) = 1-\frac{2/9}{1} = 1-\frac29 = \frac79,\quad F+xF' = \frac79+\frac23\left(-\frac13\right) = \frac79-\frac29=\frac59.

So ∂x˙/∂x∣E∗=59−15=2545−945=1645>0\partial\dot x/\partial x|_{E^*} = \frac59 - \frac15 = \frac{25}{45}-\frac{9}{45} = \frac{16}{45} > 0. Now trace =16/45−0=16/45>0= 16/45 - 0 = 16/45>0 and determinant is unchanged in form, det⁡J=225>0\det J = \frac{2}{25}>0 still (the yy-equation is untouched), so eigenvalues have positive real part: E∗E^* is now an unstable spiral. Step 3 — trapping region. With K=3K=3, the prey nullcline is xx-bounded (x˙≤0\dot x \le 0 once x≥3x \ge 3 regardless of y≥0y\ge0, since x(1−x/3)≤0x(1-x/3)\le0 there and the predation term is ≤0\le 0), and standard Rosenzweig–MacArthur estimates (using the boundedness of x/(1+x)≤1x/(1+x) \le 1 and a Dulac-type bounding box) show that the region

M={(x,y):0<x≤4, 0≤y≤4, x+y≥η}M = \{(x,y): 0 < x \le 4,\ 0 \le y \le 4,\ x+y \ge \eta \}

for small η>0\eta>0 excluding a neighborhood of the axes (where trajectories are repelled from the boundary axes since x˙>0\dot x>0 for small x>0,yx>0,y small, and y˙=0\dot y=0 exactly on y=0y=0 so the positive yy-axis is invariant and xx-axis is invariant, meaning trajectories starting strictly inside the positive quadrant stay there and are pushed away from the axes near E∗E^*) forms a compact, positively invariant, equilibrium-free trapping region once the boundary at x=4,y=4x=4,y=4 is checked to have inward flow (both x˙<0\dot x <0 for x=4x=4 and y˙<0\dot y<0 for large yy given bounded xx-range with x/(1+x)<0.4x/(1+x)<0.4 false only near x=2/3x=2/3, so the specific numeric checks follow the same pattern as Example 1).

Conclusion: with the destabilized carrying capacity K=3K=3, the interior equilibrium becomes an unstable spiral surrounded by a trapping region with no other equilibria in the open first quadrant, so Poincaré–Bendixson guarantees a stable limit cycle — a sustained predator–prey oscillation. This matches the well-known "paradox of enrichment" (Rosenzweig, 1971): increasing carrying capacity can destabilize a fixed point and produce cycles.

CounterexampleWhy planarity (two dimensions) is essential

Drop the hypothesis that the phase space is two-dimensional and consider the Lorenz system in R3\mathbb{R}^3:

x˙=σ(y−x),y˙=x(ρ−z)−y,z˙=xy−βz,\dot x = \sigma(y-x), \qquad \dot y = x(\rho - z) - y, \qquad \dot z = xy - \beta z,

with the classical parameters σ=10\sigma = 10, β=8/3\beta = 8/3, ρ=28\rho = 28. This system is known to have a compact, positively invariant trapping region (a large ellipsoid enclosing all three equilibria can be constructed via a Lyapunov-type estimate), yet trajectories inside it do not approach a periodic orbit. Numerically and (with computer-assisted proof, Tucker 2002) rigorously, trajectories wander forever on a strange attractor, never repeating, exhibiting sensitive dependence on initial conditions — the hallmark of chaos. The reason the Poincaré–Bendixson argument fails here is entirely topological: Step 3 of the proof relied on a Jordan curve JJ built from an orbit arc plus a transversal segment separating the plane into inside/outside, so that the flow could not cross the orbit-arc part of JJ (uniqueness of solutions) and hence trajectory crossings of Σ\Sigma had to be monotone. In R3\mathbb{R}^3, a closed curve does not separate space into an inside and an outside (the Jordan curve theorem is a two-dimensional fact; in three dimensions a loop can be "gone around" through the extra dimension). There is no obstruction to a trajectory looping past its earlier self without crossing it, so the monotonicity argument collapses entirely, and trajectories can braid around each other indefinitely without ever closing up. This is precisely why chaos is possible in R3\mathbb{R}^3 (and higher) but is provably impossible for continuous-time planar autonomous flows.

Scope, edge cases & common traps

SituationWhat happensWhy
Trapping region contains an equilibriumTheorem does not directly applyω(x0)\omega(x_0) may simply be that equilibrium (trajectory spirals into a stable point, no cycle forced)
Non-autonomous planar system, e.g. x˙=f(x,t)\dot x = f(x,t)Theorem failsUniqueness/no-self-crossing argument needs the vector field to depend only on position, not time
Phase space is Rn\mathbb{R}^n, n≥3n \ge 3Theorem failsJordan curve theorem is specific to the plane (or sphere); see the Lorenz counterexample
ff only continuous, not Lipschitz/C1C^1Existence may hold but uniqueness of solutions can failProof needs uniqueness to prevent orbit self-intersections; without it multiple trajectories through one point break the monotonicity argument
ω(x0)\omega(x_0) turns out to be a single equilibrium (excluded by hypothesis) or a union of equilibria plus connecting orbits (a "graphic" or heteroclinic cycle)Not covered by the basic statement; addressed by the extended Poincaré–Bendixson dichotomyThe full classification theorem says ω(x0)\omega(x_0) is either an equilibrium, a periodic orbit, or a cycle graph of equilibria and connecting heteroclinic/homoclinic orbits — the version here assumes no equilibria to skip this case

The mathematician behind it

Origins · 1 minHenri Poincaré and how this was discovered
HP
Henri Poincaré
1854–1912 · France, qualitative theory of differential equations and celestial mechanics

Henri Poincaré founded the qualitative theory of ordinary differential equations while studying the three-body problem for a prize competition sponsored by King Oscar II of Sweden. Unable to solve celestial equations of motion explicitly, Poincaré pioneered the idea of studying the geometry of solution curves — their limiting behavior, stability, and topology — rather than seeking closed-form formulas. His 1881–1886 papers "Mémoire sur les courbes définies par une équation différentielle" laid out the classification of equilibria (nodes, saddles, foci, centers) and introduced the very idea of what would become the limit set argument central to this theorem.

History
- 1881–1886: Poincaré develops the qualitative theory of planar differential equations and studies limiting behavior of trajectories.
- 1901: Ivar Bendixson, a Swedish mathematician, rigorously completes and generalizes Poincaré's argument, proving the full theorem for general planar C1C^1 vector fields.
- 1927: Solomon Lefschetz's and later textbooks (e.g. Andronov and collaborators in the 1930s–1940s) popularize the theorem for engineering applications in oscillator theory.
- 1963: Edward Lorenz's discovery of chaotic behavior in a three-variable convection model highlights, by contrast, exactly why the two-dimensional restriction of Poincaré–Bendixson is essential.
- 2000s: Computer-assisted proofs (e.g. Warwick Tucker's 2002 proof of the existence of the Lorenz attractor) rigorously confirm that three-dimensional flows escape the constraints that pin down planar flows.
Does the Poincaré–Bendixson theorem tell us how many limit cycles exist, or where?

No. It is a pure existence theorem: it guarantees at least one periodic orbit inside a suitable trapping region, but says nothing about uniqueness, the exact shape, period, or amplitude of that cycle. In fact, planar systems can have multiple nested limit cycles (this is the content of Hilbert's sixteenth problem, still partially open for general polynomial systems). To pin down the exact orbit or count how many exist inside a region, one typically needs additional tools such as the Bendixson–Dulac criterion (to rule out cycles), explicit perturbation methods (for weakly nonlinear oscillators), or numerical continuation.

Why does the theorem require excluding equilibria from the trapping region?

The proof needs a point qq in the limit set where f(q)≠0f(q) \ne 0, so that a transversal segment Σ\Sigma can be constructed (Step 2). If an equilibrium sat inside the region, the omega-limit set could simply be that equilibrium — trajectories spiral straight into a stable fixed point rather than approaching a cycle. This is not a technical convenience; it is logically load-bearing, since the entire monotone-crossing argument in Steps 3–5 collapses without a nonzero vector field to build Σ\Sigma from.

Can the theorem be applied to discrete-time maps instead of continuous flows?

No, not directly. The proof relies essentially on continuity of trajectories in time (so that the Jordan curve JJ built from an orbit arc is a genuine continuous curve) and on uniqueness of solutions preventing self-intersection of that arc. Discrete-time dynamical systems (iterated maps) do not have continuous trajectories in this sense, and indeed simple one-dimensional maps like the logistic map can exhibit chaotic behavior despite living in a "0-dimensional-in-time" analogue of one spatial dimension — precisely because the continuity argument central to Poincaré–Bendixson has no discrete-time counterpart.

How does this theorem relate to the Bendixson–Dulac criterion?

They are complementary. Poincaré–Bendixson is an existence tool: build a trapping region, exclude equilibria, and conclude a cycle exists. Bendixson–Dulac is a non-existence tool: if one can find a function B(x,y)B(x,y) such that ∇⋅(Bf)\nabla \cdot (Bf) has constant sign on a simply connected region, then no periodic orbit lies entirely inside that region (by Green's theorem, since a periodic orbit would force the flux integral to vanish, contradicting a sign-definite divergence). Together they let one both prove cycles exist in an annulus and prove they cannot exist in some smaller central sub-region, useful for locating cycles precisely.

Is the limit cycle produced by the theorem necessarily stable?

Not necessarily in general — the theorem only asserts that ω(x0)\omega(x_0) is a periodic orbit, and by construction (since it is the limit of a trajectory approaching it from one side within the trapping annulus) that particular orbit is stable from the side of the annulus (orbits from inside the annulus, at least near that boundary of approach, converge to it). But there could in principle be other periodic orbits inside the same trapping region, possibly unstable ones, sitting between the equilibrium and the one produced as the limit set. Additional monotonicity arguments (e.g., via Dulac functions or explicit uniqueness-of-cycle theorems for Liénard systems) are needed to assert global uniqueness and stability of the cycle.

Why doesn't the existence of a stable equilibrium surrounded by a repelling neighborhood already violate Poincaré–Bendixson?

It doesn't, because in that configuration the equilibrium is unstable (repelling), not stable — an unstable equilibrium is precisely excluded from being an ω\omega-limit point of nearby non-equilibrium trajectories, since trajectories move away from it, not toward it. The construction in Example 1 uses an unstable spiral at the center specifically so that trajectories starting near it are pushed toward the outer part of the annulus, leaving the equilibrium itself with an empty basin of attraction (except for the single equilibrium point itself, trivially fixed). This is exactly consistent with, not a violation of, the theorem.

What is the relationship between this theorem and the Hopf bifurcation?

They describe the same phenomenon — birth of oscillation — from different angles. The Hopf Bifurcation theorem is a local, perturbative statement: as a parameter crosses a critical value where a pair of complex eigenvalues of an equilibrium crosses the imaginary axis, a small-amplitude periodic orbit is born in a neighborhood of that equilibrium, and its existence follows from local normal-form/center-manifold analysis. Poincaré–Bendixson is a global statement requiring no parameter, no small-amplitude assumption, and no eigenvalue computation beyond confirming instability — it only needs a trapping region. In practice, one often uses Hopf's theorem to locate the birth of a cycle near a bifurcation point, and Poincaré–Bendixson to confirm the cycle persists and is genuinely global, away from the bifurcation threshold.

Before this

Where this leads