Turn one knob of a simulation slowly: a harvest rate, a Reynolds number. For a long time only smooth changes happen, then the answer jumps, and turning the knob back does not undo it. Nothing is broken; the equation has a place where its set of long-run states changes.
3.8 Autonomous Equations, Equilibria & Stability ended on such a place: the harvested logistic equation had two equilibria that merged at and vanished. This lesson studies that event, finds its cousins, and asks what a numerical method sees next to one.
Transcript
- Turn a knob slowly and the answer can suddenly jump. Last lesson's harvest model did: as h grows, the graph of f slides down, the equilibria merge at one quarter and vanish. With u equals y minus a half and r equals a quarter minus h, that is u prime equals r minus u squared: a saddle-node.
- Stack the phase lines. For r below zero, f is negative everywhere: no equilibria. At zero, one. Above zero, plus and minus root r, with arrows from the sign of f. Let the equilibria trace out branches, solid where stable, dashed where unstable: the bifurcation diagram.
- Near the fold the stable slope is minus two root r, so the relaxation time is one over two root r: 2.2361 at h equals 0.2, 5 at 0.24, 15.8114 at 0.249, growing without bound. This critical slowing down warns that a fold is near. It is not a clock: noise can tip the system sooner.
- Transcritical: y prime equals r y minus y squared. The equilibria zero and r pass through each other. The slope at zero is r, at r it is minus r, so at r equals zero they swap stability, and the solid and dashed styles switch.
- Supercritical pitchfork: y prime equals r y minus y cubed. Zero loses stability at r equals zero, though even there y prime is minus y cubed, still stable. Two new stable branches, plus and minus root r, are born, with slope minus two r.
- Subcritical: y prime equals r y plus y cubed minus y to the fifth. Nonzero equilibria have y squared solving z squared minus z minus r equals zero. At r equals minus 0.2 that gives 0.7236 and 0.2764, the harvest equilibria at h equals 0.2: the same quadratic. The large branch is stable, the small one unstable. The fold lies left of the pitchfork.
- Start at zero and raise r slowly. It stays at zero through the interval from minus a quarter to zero, because zero is still stable, then at zero jumps up. Lower r: it stays on the large branch until the fold at minus a quarter, then falls. Two stable states coexist over an interval, so the state depends on history: hysteresis.
- Another loop: y prime equals r plus y minus y cubed, with folds at plus and minus 0.3849. A real sweep lags. At rate 0.002 the jumps come near plus and minus 0.416, at 0.0005 near 0.397. The lag shrinks as the sweep slows; the loop does not.
- A tipping point is a fold whose jump is not undone by returning the parameter. After the jump at 0.3849, bring r back to 0.30 and the state stays up; it falls only at minus 0.3849. These are caricatures; nothing here shows that any real system is near one.
- A time-stepper cannot follow unstable branches: at r equals 0.30 the arrows push away from the middle one. Continuation solves f equals zero by Newton's method, stepping r. It follows the middle branch: minus 0.101 at 0.10, minus 0.339 at 0.30, minus 0.429 at 0.35. At 0.40 it fails: the fold is behind us.
- The ghost. Past a fold, y prime equals minus epsilon minus y squared has no equilibrium, but y prime is tiny near zero. With epsilon 0.0001, a rule that stops when y prime is under 0.001 fires at t equals 31.2, y equals 0.03, and reports steady. The solution really reaches minus one at 312.16: pi over root epsilon, minus two.
- Flow past a 2D circular cylinder is steady below a critical Reynolds number, about 47, from memory, and sheds a periodic vortex street above it. That is a Hopf bifurcation: it cannot occur in one dimension, where solutions never oscillate, so it needs two. A preview of systems, not an example from this lesson.
- Recap: a bifurcation is where the phase line changes as a parameter moves. Solid stable, dashed unstable. Slowing down warns of a fold, hysteresis is memory, a ghost fools stopping rules.
Parameter-dependent equations and the bifurcation diagram
Let the right side depend on a parameter :
For each fixed this is an autonomous equation with its own phase line. A bifurcation is a value at which the phase line changes qualitatively: equilibria appear, vanish, or swap stability.
The bifurcation diagram is the whole family at once: horizontal, equilibria vertical, each branch solid where stable and dashed where unstable. A vertical slice at one is that ‘s phase line, with arrows read off the sign of . The sim below draws it beside the graph of and the solutions.
Where? If and , the implicit function theorem gives one smooth equilibrium through that point, and keeps its sign nearby, so nothing appears, vanishes or swaps. A bifurcation needs : the linear test of lesson 3.8 goes silent exactly at the interesting points. That is necessary, not sufficient: for , at , yet the single branch is smooth and stable for every .
Saddle-node: equilibria collide and annihilate
The normal form is
For the equilibria are with : is stable (), unstable (). At they merge at , where : semi-stable, . For , everywhere and every solution decreases for ever. The diagram is a sideways parabola, upper half solid, lower half dashed.
This is the harvest model in other coordinates. Put in : then , so with it is . At : , , and , the equilibria of lesson 3.8. Raising is lowering .
Why this is the generic fold. If at with and , then : after rescaling, or its mirror image.
Critical slowing down. At the stable equilibrium , so the relaxation time is
which diverges as . For the harvest model gives , the figures quoted in lesson 3.8 and two more:
Kick the system and time its recovery: a recovery time that grows as a parameter drifts is a known, measurable early-warning signal of a fold.
What it does not do. It says a bifurcation is approaching, not when the system tips: noise can push the state across the unstable equilibrium before reaches , since the gap shrinks like . Rising is a reason to look harder, not a clock.
- Saddle-node. Drag through : two equilibria, one (“inconclusive”), none. Sweep r: the state follows the stable branch, crawls past , escapes, and does not return.
- Transcritical, supercritical pitchfork. The sweep legs nearly retrace each other; the small gap is the lag of a finite sweep speed.
- Subcritical pitchfork, cusp. A loop, below.
Transcritical: two equilibria swap stability
The equilibria are and for every , with : , . For , is stable and unstable; for the roles swap; at they coincide and is semi-stable. Nothing is destroyed: the equilibria pass through each other and exchange stability.
The logistic connection. Write logistic growth as , carrying capacity . With crowding fixed and the growth rate as parameter, and this is the transcritical form. As falls through the equilibrium goes from unstable to stable: the line between persistence and extinction. The branch is a negative “population”, meaningless in ecology but visible in the mathematics.
Pitchfork: symmetry and its breaking
Supercritical. is unchanged by . For the only equilibrium is . For there are three, and , with :
At the equation is : stable by the sign of although . One stable state splits into an unstable one and two stable ones, and an arbitrarily small asymmetry in the start decides which: symmetry breaking. The relaxation times at and at diverge as . (Checked on values of in .)
Subcritical. : the cubic now pushes outward and the quintic catches the solution later. Away from , with , the equilibria solve :
The "" root is positive for every (the large branch); the "" root is positive only for (the small branch). At : and , so and . These are the harvest equilibria at : the same quadratic, with . The fold at is the harvest bifurcation, at .
Stability, exactly. On an equilibrium , so . With and this is :
- large branch (upper sign): , stable for every ;
- small branch (lower sign): for , that is for : unstable there;
- the origin: , stable for , unstable for . At , just above : unstable, although .
Check at : and . So: below only (stable); at a saddle-node creates the large and small pair; for there are three stable states ( and large) separated by unstable small branches; at the small branch collapses onto the origin, which turns unstable. The fold sits to the left of the pitchfork, and that ordering makes hysteresis.
Hysteresis
Definition. A system shows hysteresis when its long-run state depends on its history, because two stable states coexist over an interval of .
Subcritical pitchfork. Start at on and raise very slowly. The state stays at through the whole bistable interval , because is still stable. At it turns unstable and the state jumps to the large branch, . Now lower : the state stays on the large branch down to , where it folds away, and falls to . In the slow limit the jumps are at (up) and (down): a loop of width .
The cusp model. has equilibria on . Folds are where , so , at . For there are three equilibria, two stable and one unstable; at : (unstable, ) and (stable, ). Sweeping up from (where ) the state runs to the fold at and jumps to ; sweeping down it runs along the upper branch to and jumps to . Each jump has length .
What a real sweep does. The above is the infinitely slow limit. The sim and the playground cell integrate with RK4 (step ) while moves at speed . At the cusp state moves fastest at going up and going down; at , ; static, . For the subcritical pitchfork at (with a noise floor , since the exact equilibrium would hold the state for ever) it is up (static ) and down (static ). A finite sweep lags: the jump comes after the bifurcation. The lag shrinks as the sweep slows; the loop does not, converging to the width of the bistable interval. That is the test: a lag closes as the sweep slows, hysteresis does not.
In the bistable interval click on both sides of the dashed small branch: the solution goes to or to the large branch. The unstable branch is the boundary between the basins, invisible in a time-stepper’s long-run output, and it merges with a stable branch at the fold.
Tipping points
A tipping point is a fold where a system on a vanishing branch must jump to a distant state, and returning the parameter to its old value does not undo the jump. In the cusp model the lower state disappears at and the state jumps up; at it is still on the upper branch, and to get back you must go to . That is hysteresis. The harvest model is the collapse version: past the population goes extinct, and lowering does not revive it because is absorbing.
What this does not say. These are caricatures, showing a mechanism: a fold plus a distant stable state to land in. Whether a particular lake, fishery or ice sheet has such a fold, and how close it is, is an empirical question. Nothing here shows that any real system is near a tipping point.
What a solver sees
The ghost of a vanished equilibrium. Take the normal form just past the fold, with , :
There is no equilibrium, but is tiny near : a bottleneck. Separating variables (lesson 3.3), the time to pass it is
Over the exact time is , and RK4 reproduces it:
| RK4, | exact | ||
|---|---|---|---|
The passage time grows like , the mirror of critical slowing down, .
This breaks “run until steady”. A solver that stops when meets , which dips below tol exactly when : possible only if . With and tol , from , the rule fires at with and the program reports a steady state. There is none: the true solution goes through and reaches at . Residual tests that work far from a bifurcation fail near one. A defence: evaluate the linearisation where the rule fires. Here , tiny, which signals a bottleneck, and says how long the run must be.
Euler near a fold. The limit of lesson 3.2 is set by the largest visited ( at , so ), not by the bottleneck where . Accuracy is the issue: for , takes steps and gets against ; gets .
Flow past a cylinder
Exercise: continuation along stable and unstable branches
Implement natural-parameter continuation. It must stop when Newton fails (a fold) or the solution jumps to another branch.
Problems
Consider . (a) Find the bifurcation of and decide whether it is super- or subcritical. (b) Estimate the nonzero equilibria near it and their stability. (c) Where do the nonzero equilibria fold?
Show worked solution
(a) , so is stable for and unstable for . Expanding, ; the cubic term at is , pushing outward: subcritical.
(b) gives , for , with : unstable. At solving gives , (estimate ), .
(c) Equilibria satisfy , which decreases until at , where and : a fold. Beyond it the branch is stable (: , ). The same structure as the lesson’s subcritical pitchfork, with the fold to the left.
A solver integrates from and stops when . (a) For which does the rule fire? (b) For , at which does it fire, and how long would the true solution need to reach from there? Compare with .
Show worked solution
(a) has minimum at , so the rule fires iff .
(b) It fires when , at . The remaining time is , confirmed by adaptive integration (); . The “steady state” is about time units from the end of the passage.
Let in , , . (a) For small , solve the linearised equation and find the at which returns to . (b) At which does reach ? (c) Repeat (b) for and say why the sim uses a noise floor.
Show worked solution
(a) , which is at , however slow the sweep: the state decayed while and must climb back.
(b) gives ; integrating the full equation gives .
(c) gives (integration: ). The instability starts at but is invisible until : a delayed bifurcation. A noise-free sweep from deep in the stable region would stay at far past the critical value, so the sim holds the state off ; real systems have noise.
Key Takeaways
- A bifurcation is an where the phase line of changes qualitatively; it needs , which is not sufficient (). The bifurcation diagram stacks the phase lines: stable branches solid, unstable dashed.
- Saddle-node : for merge and vanish; the harvest model is this with . Critical slowing down: , a warning that a fold is near, not a clock.
- Transcritical : and swap stability. Pitchforks: supercritical has stable; subcritical has a stable large branch for , left of the pitchfork.
- Hysteresis: two stable states coexist over an interval of (jumps at and ; cusp model at ). A finite sweep lags; the lag closes as the sweep slows, the loop does not.
- A tipping point is a fold whose jump is not undone by returning ; these models are caricatures.
- Time-stepping cannot follow unstable branches; continuation can. Near a fold the passage takes and a stopping rule can stop on the ghost.
- Vortex shedding behind a cylinder (critical Reynolds number about in 2D) is a Hopf bifurcation: it needs two dimensions, see 7.9 Limit Cycles & Hopf Bifurcations.