A frictionless pendulum swings forever at whatever amplitude you gave it. That is a center, and the last lesson showed that an integrator drifts off its level curve. A heartbeat, a violin string and the wake behind a cylinder do something else: their oscillation has one size, set by the system, and it recovers if you kick it. That is a limit cycle, and only a nonlinear system can have one. We solve one in polar coordinates, prove when one must exist, measure the van der Pol cycle, and watch a cycle being born as a parameter crosses a critical value: the two-dimensional cousin of 3.9 Bifurcations, and how vortex shedding starts.
Transcript
- A pendulum keeps whatever amplitude you gave it: a center, with a continuum of closed orbits. A limit cycle is an isolated closed orbit that nearby trajectories approach. Kick a center and the error stays for good; kick a stable cycle and the flow shrinks it.
- Polar coordinates give a cycle you can solve. For this system, r prime equals r times one minus r squared, and theta prime equals one. Every lap takes two pi. The radius grows inside r equals one and shrinks outside, so the unit circle attracts.
- The radius solves exactly with u equals r to the minus two, and RK4 agrees: from point two, r of three is 0.97152; from two, 1.000931. Reverse the sign and the circle repels. From 0.9 the radius is 0.00009 at t equals ten; from 1.1 it blows up at 0.8756.
- To show a cycle exists, build a trapping region. On r equals one half the radial speed is 0.375, inward; on r equals two it is minus six, also inward. No equilibrium in the ring, so Poincaré–Bendixson, stated, not proved, gives a closed orbit. No chaos in the plane: that needs three dimensions. A divergence of one sign forbids cycles.
- Van der Pol has negative damping for x between minus one and one. The origin has trace mu and determinant one: at mu equals one, 0.5 plus or minus 0.866 i; at five, 0.2087 and 4.7913. Everything leaves the origin and large orbits are pulled back, so a cycle exists.
- Why amplitude two? Energy obeys E prime equals mu times one minus x squared times y squared. Averaging over a lap gives a squared times one half minus a squared over eight, zero. So a equals two: pumping inside balances loss outside.
- Measured amplitudes: 2.0001, 2.0086, 2.0215, 2.0143, all within 1.1 percent of two. Periods: 6.2871, 6.6633, 11.6123, 19.0784, from two pi upward. The asymptote three minus two log two times mu is only a trend: at mu ten the period is 18 percent above it, at mu five, 44 percent.
- Large mu is fast and slow. With u equal to x prime over mu minus x plus x cubed over three, x races to the cubic and creeps along it. Creeping from x equals two to the fold takes mu times three halves minus log two. Then x jumps, and two creeps per lap give three minus two log two, times mu.
- That fast-slow structure is stiff: the fast eigenvalue is minus 3000 at x equals two when mu is 1000. RK45 takes 17977 steps, Radau 39, and both give x of twenty equal 1.986592. The mean RK45 step is the stability limit, not accuracy.
- The Hopf normal form: r prime equals mu r minus r cubed. The eigenvalues mu plus or minus i cross the imaginary axis. For mu positive, a cycle of radius root mu appears: measured 0.2000, 0.5000, 1.0000. Quadrupling mu doubles the radius.
- Flip the cubic's sign and add a quintic: subcritical. At mu equals minus 0.2 the origin is stable, with cycles at 0.52573, unstable, and 0.85065, stable. Seed 0.5 decays; 0.55 and 1.2 reach 0.85065. They merge at minus one quarter, radius 0.70711. Raising mu through zero the state jumps to a cycle of radius one: hysteresis.
- Why does a stable cycle absorb numerical error? A patch changes area by e to the integral of the divergence per lap. For van der Pol at mu one that integral is minus 7.052, a multiplier of 8.65 times ten to the minus four, matching the measured return map. At mu five it is about e to the minus 85.
- Vortex shedding is a Hopf bifurcation, near Reynolds 47, from memory, and numerical damping delta turns mu into mu minus delta. Recap: cycles are isolated and polar form solves them; Poincaré–Bendixson forbids planar chaos; van der Pol is relaxation and stiff; Hopf births cycles. Next: chaos in three dimensions.
Closed orbits: a continuum versus an isolated loop
A periodic orbit is a closed loop in the phase plane. In a conservative system (7.8 Conservative Systems & Lyapunov Functions) closed orbits come in a continuum, one per level curve. A limit cycle is an isolated closed orbit: nearby trajectories spiral toward it (stable) or away from it (unstable). For simulation that is the difference between accumulating and correcting errors: on an oscillator an error moves you to a neighbouring level curve for good, while on a stable cycle it is a distance from the loop that the flow shrinks. A consistent scheme ends on a slightly displaced cycle and stays there; we measure this below.
Polar coordinates: a limit cycle you can solve
With ,
For one finds and , so
The angle turns at unit speed, so every lap takes . The radius obeys a one-variable equation (3.8 Autonomous Equations, Equilibria & Stability): inside and outside, so is stable and the unit circle is an attracting closed orbit. The origin repels. It also solves exactly: gives , hence
Lab block A, RK4 on the Cartesian system, : from both RK4 and the formula give ; from both give . Inside and outside converge from opposite sides.
An unstable cycle. Reverse the sign: (the sim’s “unstable ring”). Now repels and the origin attracts: from the radius is at ; from it blows up at , and the sim stops with “BLEW UP” at , a little late because RK4 only notices a singularity within a step of it. An unstable cycle is a basin boundary, and a forward run never lands on it.
Trapping regions and the Poincaré–Bendixson theorem
How do you know a cycle exists when you cannot solve? Build a trapping region. For the ring above, the radial speed on the inner circle is and on the outer circle (lab A; the sim’s bottom panel plots ). The flow enters the ring on both sides, and the only equilibrium, the origin, is outside it.
So the ring must contain a cycle. The idea: a closed curve in the plane has an inside and an outside and orbits cannot cross (3.7 Existence & Uniqueness), so a trapped orbit can only wind onto a loop. Consequence: no chaos in the phase plane of an autonomous system. Chaos needs bounded trajectories that never settle, which needs room to cross over and under: three variables, as in the Lorenz equations of 7.10 Chaos & the Lorenz System. A forced 2D oscillator counts as three, since time is a variable.
No cycle: Bendixson’s criterion (from memory, one-line proof). If has one strict sign on a simply connected region , no closed orbit lies in . On a closed orbit is tangent, so the flux ; Green’s theorem (2.10 Green, Stokes & the Divergence Theorem) says it equals . Example: has divergence , so no periodic orbit. The test is silent when the divergence changes sign (van der Pol: ) or vanishes (Hamiltonian flows).
The van der Pol oscillator
with reads
The damping term has negative damping (energy input) for and ordinary damping for . The Jacobian at the origin is , trace , determinant : eigenvalues . That is an unstable spiral for (: ; sim readout) and an unstable node for (: and ). Everything leaves the origin and large trajectories are pulled back, so Poincaré–Bendixson gives a cycle (unique and stable by Liénard’s theorem; from memory).
Why the amplitude is about 2. For small the motion is nearly , and obeys . Over a lap the net change must vanish. Averaging with and gives , so : pumping inside balances loss outside.
Measured (RK4, , seed , last lap; lab block B and the sim readout):
| amplitude | period | |||
|---|---|---|---|---|
| 0.1 | 2.0001 | 6.2871 | 6.2832 | 0.1614 |
| 1 | 2.0086 | 6.6633 | 6.2832 | 1.6137 |
| 5 | 2.0215 | 11.6123 | 6.2832 | 8.0685 |
| 10 | 2.0143 | 19.0784 | 6.2832 | 16.1371 |
The amplitude stays within of over two decades of . The period is at small (the linear frequency ) and grows with , but the large- formula is an asymptotic trend: at the measured period is above it (gap ) and at it is above (gap ). The gap shrinks slowly (like ; from memory), so quote the formula as a trend and measure when it matters.
Relaxation oscillations, and the origin of . Put . The equation gives and : for large , is fast and is slow. So races to the curve and creeps along it, with , i.e. . Creeping from to the fold at takes
then jumps (time of order ) to (the line meets the curve again there) and creeps back. Two creeps per lap give , and the jump explains the amplitude at .
Stiffness. On the slow curve the fast eigenvalue is , which is at for : a quickly decaying mode with a slow solution riding on it, the definition of stiffness in 4.5 Stability Regions & Stiffness. Lab block G, on , : RK45 takes steps ( evaluations), Radau steps ( evaluations), and both give . The mean RK45 step is essentially the stability limit of that lesson: accuracy did not set it (4.4 Adaptive Step Size & RK45).
The Hopf bifurcation
In 3.9 Bifurcations equilibria appeared, vanished or swapped stability, but nothing could oscillate. In two dimensions an equilibrium can lose stability to an oscillation. The normal form is
and the polar computation above gives . The Jacobian at the origin, , has eigenvalues . For that is a stable spiral; at the pair sits on the imaginary axis with imaginary part , not (that is what separates Hopf from the real crossing of a saddle-node or pitchfork); for it is unstable, and has a stable root: a limit cycle of radius and period . The radial equation is exactly the supercritical pitchfork of the bifurcations lesson (for ), and the eigenvalue crosses at unit speed, . At this is the borderline case of the last lesson: decays only like , so transients near onset are very long.
Measured radius (lab C, sim; RK4, ): give to four decimals. The square-root law: quadrupling from to multiplies the radius by , and a factor (from to ) multiplies it by . The sim’s bottom panel sweeps and plots measured radii on the parabola.
Supercritical versus subcritical. Flip the cubic’s sign and add a quintic to keep it bounded: (the sim’s “subcritical”). Cycles satisfy . For there is one stable cycle. For there are two and the origin is also stable (eigenvalues at ): the radii are (unstable) and (stable). In lab D (to ) the seed ends at , while and both end at . At the cycles merge at and vanish. The consequence is hysteresis and a jump: as rises through the origin loses stability and the state leaps to a cycle that already has radius at ; lowering again it stays on the large cycle until . The sim’s sweep (small seed green, large seed blue) draws the loop. The supercritical form is gentle and reversible, the subcritical one abrupt (the hysteresis of the bifurcations lesson). That every Hopf bifurcation reduces to one of these forms, the sign of one coefficient deciding, is from memory.
Measuring a cycle on a computer
An upward zero crossing is a pair of samples with ; linear interpolation puts it at , with error of order . The period is the gap between successive crossings, the amplitude the largest between them. The sim reads the last two laps and prints “not settled” unless the lap before agrees to . For this matters: , so a seed at needs about time units merely to reach the cycle (the sim runs to ).
Step size and method (lab E, , ; reference RK4 with : amplitude , period ):
| method, | amplitude | period error |
|---|---|---|
| RK4, | 2.00862 | |
| Euler, | 2.02779 | |
| Euler, | 2.01052 |
RK4 at matches the reference to five decimals. Euler is first order and shows it: a tenfold smaller step cuts the error by . The step size bites harder for large (sim, , seed 1 = ; the outside seed is already unstable at , where ): gives period ( over ), no longer settles, and blows up at . The RK4 limit of the stiffness lesson (fast eigenvalue about ) is the right range; the observed threshold is a little lower.
The cycle does not accumulate error. Euler at moved the cycle (amplitude , period ) but lap and the last lap both take : nothing compounds. The reason is contraction. By Liouville’s formula (the linear case was in 7.5 The Matrix Exponential & Fundamental Matrices) a patch around the orbit changes area by per lap; one multiplier is (a shift along the orbit only moves the phase; from memory), so the other is . For van der Pol the divergence is , the lap integral is and the multiplier (lab F and the sim, seed 1). Check: nudge the orbit by on the line and run one lap; the return map has slope , within . At the sim reports about : one lap erases a disturbance to rounding error. For the Hopf form the lap integral is : the sim reads at against (the lap is rounded to the grid), a factor per lap.
Try it
What to try. (1) Van der Pol: , reading the period against the dashed asymptotes in the bottom panel. (2) Hopf: (radius ), , , then and watch every seed decay. (3) Subcritical at : click inside and outside the unstable cycle. (4) Unstable ring: seeds and . (5) Van der Pol : raise to , , .
Computing it
Now build the measuring tool yourself: an integrator, a crossing detector with interpolation, and the cycle’s amplitude and period.
Concept checks
Problems
For with , use to find and evaluate .
Show worked solution
, so with , and . This matches RK4 in lab block A. The solution decreases to from above without crossing it, as the sign of predicts.
For the supercritical normal form with give the eigenvalues at the origin, the cycle’s radius and period, and the factor by which a small distance from the cycle shrinks per lap.
Show worked solution
Eigenvalues : an unstable spiral. has the stable root , and the period is . The derivative of at is , so a lap multiplies a small distance by (by hand; the sim’s prediction line prints the same formula).
For find when cycles exist, their radii at , and their stability.
Show worked solution
Cycles satisfy , i.e. : two positive roots for , one for , merging at (). At : and , so and (lab D). With : at , , unstable; at , , stable. The origin () is stable, so the inner cycle separates two basins: lab seeds and end on opposite sides. (The values are by hand.)
Key Takeaways
- A limit cycle is an isolated closed orbit; a center has a continuum. A stable cycle corrects errors, so a consistent scheme shifts it but does not accumulate error (Euler at : period error , identical on lap and the last lap).
- Polar form turns , into , : unit circle, period , exact solution via ( from , RK4 and formula). The sign of decides attraction; is an unstable cycle.
- Poincaré–Bendixson (stated): trapped in a bounded region with no equilibrium means approaching a closed orbit, hence no chaos in the plane. Bendixson (stated, via Green): a divergence of one sign forbids cycles.
- Van der Pol: amplitude about (measured at ); period at against the trend . Large is fast-slow and stiff: RK45 steps, Radau at .
- Hopf: eigenvalues cross the imaginary axis; the supercritical cycle has radius (measured ; quadrupling doubles it). Subcritical at : stable origin and stable cycle () split by an unstable one (), hence hysteresis.
- Vortex shedding onset (about , from memory) is a Hopf bifurcation; uniform numerical damping shifts to . Next, three dimensions and chaos: 7.10 Chaos & the Lorenz System.