The Figure Poincaré Would Not Draw

In the 1890s, wrestling with the three-body problem, Henri Poincaré found the object that would come to define chaos. Two curves — the paths along which orbits asymptotically leave and approach an unstable equilibrium — did not, as everyone had assumed, join smoothly. They crossed. And once they crossed once, the geometry forced them to cross infinitely often, weaving a mesh of such intricacy that he wrote [1]:

"One will be struck by the complexity of this figure, which I shall not even attempt to draw."

He had seen deterministic chaos two generations before it had a name. The figure he would not draw is the homoclinic tangle — and it is easiest to meet not in a cavity or a solar system, but in the simplest map that has one.

In Part I the circle kept every ray's angle forever; in Part II we deformed it and a chaotic sea opened up. Here we set the cavity aside and take up the kicked rotor: a rigid rotor that gets a periodic shove. Watched once per kick, its entire dynamics collapses to two lines of arithmetic — and inside those two lines live the tangle, the cantorus, and the quantum localization that Part IV will carry back to light.

The kicked rotor

A rotor is the barest mechanical system: a mass free to swing around a pivot, its state one angle $\theta$ and its conjugate momentum $p$. Left alone it turns at constant rate forever — $p$ is conserved, exactly as the circle of Part I conserved each ray's angle of incidence. Now kick it: once every period $T$, give it a jolt whose strength depends on where it is, $\propto K\sin\theta$. Between kicks it coasts; at each kick $p$ jumps. Sampling the motion once per kick — a stroboscopic Poincaré map — collapses the whole history to a single area-preserving map, the standard map:

$$ p_{n+1} = p_n + K\sin\theta_n, \qquad \theta_{n+1} = \theta_n + p_{n+1}. $$

The one parameter $K$ is the kick strength — the amount of nonlinearity. At $K=0$ the rotor is integrable: $p$ never changes, and the phase plane is a stack of horizontal lines, each a conserved-momentum torus (the standard map's whispering-gallery limit). Turn $K$ up and those lines begin to buckle, break, and finally dissolve into a chaotic sea. Everything in this notebook is one figure of that transition.

The integrable skeleton: a pendulum

Before the kicks tangle anything, it helps to meet the clean object underneath. Between kicks the rotor merely coasts, and averaged over a period the central resonance of the standard map is, to leading order, an ordinary pendulum:

$$ H = \frac{p^2}{2} + K\cos\theta. $$

Its phase space has three kinds of motion. Around the stable equilibrium at $\theta=\pi$ the pendulum librates — swings back and forth — on nested closed loops. With enough momentum it instead rotates, swinging all the way over the top again and again. Dividing the two is the separatrix: the single orbit of exactly the critical energy, which creeps toward the unstable equilibrium at $\theta=0$ and takes infinite time to arrive. It is a smooth closed curve through that saddle — the integrable ancestor of everything that follows. A finite kick is what shreds it into the tangle.

Now kick it

That pendulum is only an average — the smooth, integrable approximation to the motion. The real rotor takes its energy in sudden kicks, and sampling it once per kick gives back the standard map from the top of the page. The gap between the two is the whole story: the pendulum's separatrix is a single clean curve, but the kick forces the true stable and unstable manifolds slightly apart — and once apart, they have no choice but to tangle.

So turn the knob. The slider below sets the kick strength $K$; the panel under it is the standard map's phase portrait, recomputed live. Start near $K=0$ (the pendulum, essentially untouched) and turn it up: watch the tori buckle into wavy KAM curves, island chains open, the golden torus hold out to the last, and a thin chaotic layer fray along the old separatrix. Everything that follows is a magnification of that layer.

Drive it yourself.

• 🎬 tour — a guided walk through the breakup sequence. It drops to a sparse view so the tracked tori stand out, sweeps K upward slowing down as each irrational torus dissolves in turn, then eases into the golden torus, zooms into the fine detail as it breaks, and pulls back out to watch the whole plane go chaotic (ends near K≈1.25). Touch any control to take over.
• K slider — the kick strength (nonlinearity). Drag it slowly and watch the smooth KAM tori wrinkle, break into island chains, and dissolve into a chaotic sea. Its own full-width row gives fine control.
• ▶ play — auto-scans K up to 6 and back; speed sets the pace.
• iterations — points per orbit: more fills the curves in (crisper tori) at some cost to responsiveness.
• scroll to zoom about the cursor, drag to pan — the torus tiles by $2\pi$ in both $\theta$ and $p$, so panning wraps seamlessly. reset view returns to one cell.
• The two gold threads track the golden torus — the last KAM barrier to survive (rotation number $\gamma=(\sqrt5-1)/2$) and its mirror partner. They glow gold with a bright multi-colour wave running along them while intact; watch them shatter into a shimmering multi-colour dust cloud as K crosses $K_c\approx0.972$. The badge (bottom-right) reads intact / critical / BROKEN and reports $\Delta p$, how far that orbit wanders — near zero on a torus, large once the barrier is gone.
• The other coloured curves (legend, below the plot) are tori of other famous irrationals — $\sqrt2-1$ (silver), $e-2$ (magenta), $\pi-3$ (violet). Because the golden mean is the "most irrational" number, its torus is the last to break: these shatter into dust at lower K, one by one, while the gold thread still spans the plane.

Two fixed points, two fates

The simplest orbits are the ones that return after a single kick — the fixed points. Setting $\theta_{n+1}=\theta_n$ and $p_{n+1}=p_n$ needs $\sin\theta=0$ and $p=0$: two points, $(\theta,p)=(\pi,0)$ and $(0,0)$. They could not behave more differently, and the difference is read straight off the linearised map. Its Jacobian has unit determinant (area preservation) and trace

$$ \operatorname{tr} M = 2 + K\cos\theta. $$

At $(\pi,0)$ the trace is $2-K$: for $0<K<4$ it lies in $(-2,2)$, so the eigenvalues are complex on the unit circle — an elliptic point, the centre of the island you can see in the portrait. At $(0,0)$ the trace is $2+K>2$: the eigenvalues are real, $\Lambda$ and $1/\Lambda$ — a hyperbolic point, a saddle. This is Poincaré–Birkhoff in miniature: the broken chain leaves stable and unstable orbits in equal measure, and it is the saddle at the origin whose manifolds we now follow.

Reading the saddle

Close to the hyperbolic point the map is its linearisation, the $2\times2$ monodromy matrix $M$. Area preservation forces its eigenvalues to multiply to one,

$$ \Lambda\,\cdot\,\Lambda^{-1} = 1, \qquad \Lambda = e^{\lambda} > 1, $$

so a displacement along the unstable eigenvector $\mathbf e_u$ is stretched away from the saddle by $\Lambda$ each kick, while one along the stable eigenvector $\mathbf e_s$ is squeezed in by $\Lambda^{-1}$. The rate $\lambda=\ln\Lambda$ is the orbit's Lyapunov exponent — the local clock of the chaos. These two eigen-directions seed the two curves Poincaré was drawing: run $\mathbf e_u$ forward to trace the unstable manifold $W^u$, run $\mathbf e_s$ backward to trace the stable one $W^s$.

Growing the manifolds

The unstable manifold $W^u$ is the set of points that stream away from the saddle under forward iteration. We grow it the natural way: seed a short segment of points along $\mathbf e_u$, a distance $\sim\!\delta$ from the fixed point, and iterate the map. Each kick stretches the segment by $\Lambda$ along its length, so a handful of iterations sweeps out a long curve — but a uniformly seeded segment thins out as it stretches, so we resample adaptively, inserting points wherever neighbours pull apart, to keep the curve smooth however far it runs. The stable manifold $W^s$ is the same construction run backward (equivalently, by the standard map's time-reversal symmetry, the mirror image of $W^u$).

For a few iterations the two curves look like the innocent separatrix of an integrable pendulum. The trouble starts only when they run far enough to find each other.

Where the threads cross

Run the manifolds out. In a perfectly integrable rotor the unstable curve leaving the saddle would arrive exactly on the stable curve of its neighbour, joining into a smooth separatrix. That coincidence is infinitely fragile: at any $K>0$ the two curves no longer align, and generically they cross transversally, at a single homoclinic point $h$.

One crossing begets infinitely many. Both manifolds are invariant — the map carries $W^u$ to $W^u$, $W^s$ to $W^s$ — so every image $h, T(h), T^2(h),\dots$ is a crossing too, marching toward the saddle and crowding up without limit as it approaches (each step shrinks by $\Lambda^{-1}$ along $W^s$). A manifold cannot cross itself and cannot simply stop, so between consecutive crossings it must bulge into a lobe; as the crossings crowd together the lobes stretch and thin into an endlessly folded mesh — the homoclinic tangle. This forced stretch-and-fold is exactly the Smale horseshoe [2], the geometric engine of deterministic chaos, and it is the figure Poincaré would not draw. We draw it.

Lobes and the turnstile

The tangle is not just ornament, it is bookkeeping. The lobes bounded between successive homoclinic crossings tile the region around the broken separatrix, and one pair acts as a turnstile: each kick, the map carries the area of the entering lobe across the pseudo-separatrix into the chaotic sea, and an equal area of the exiting lobe back in. Area is conserved, so the two lobe areas are equal, and that shared area is the flux — the exact rate at which orbits cross between order and chaos per kick [3]:

$$ \Phi \;=\; A_{\text{lobe}}. $$

This is the machinery behind stickiness: an orbit launched just outside the last island torus is neither free nor trapped — it is caught in the lobes, shuttled through the turnstile over many kicks before finally escaping. Count the flux and you have measured how leaky the barrier is.

The cantorus

So far, one saddle and its tangle. As $K$ grows, the invariant tori that still span the phase plane — the ones that block a rotor from ever changing its momentum — break one by one. The most robust is the torus with the most irrational winding number, the golden mean; it survives until $K_c \approx 0.971635$ [4]. Just past $K_c$ it does not vanish. It shatters into a cantorus: a Cantor-set remnant of the torus, full of gaps, built from a whole family of unstable periodic orbits of ever-higher period packed into a thin band of phase space [5].

A cantorus is a partial barrier. A torus lets nothing through; a cantorus lets a trickle through the gaps, and the size of that trickle is — again — a turnstile flux, now summed over the family's overlapping tangles. This is the concrete meaning of a family of unstable orbits concentrated in a region of phase space: not one saddle and one tangle, but a fractal ladder of them forming a leaky wall — and it is what suppresses diffusion.

An aside — why the golden torus?

Of all the tori, why is the one at the golden mean the last to survive? Because the golden mean is, in a precise sense, the most irrational number there is — and irrationality is exactly what protects a torus from breaking.

The tool is the continued fraction. Any number can be written $$ x = a_0 + \cfrac{1}{a_1 + \cfrac{1}{a_2 + \cfrac{1}{a_3 + \cdots}}}, $$ and truncating it gives the best possible rational approximations $p_n/q_n$ (the convergents). A large partial quotient $a_{n+1}$ means the previous truncation was already excellent — the number sits very close to a rational. So a number is hard to approximate precisely when all its $a_i$ are as small as possible.

The smallest they can be is $1$. And the number whose continued fraction is all ones, $$ \varphi - 1 = \cfrac{1}{1 + \cfrac{1}{1 + \cfrac{1}{1 + \cdots}}} = \frac{\sqrt5 - 1}{2} = 0.6180\ldots, $$ is the golden mean. Its convergents are the ratios of consecutive Fibonacci numbers — $\tfrac11, \tfrac12, \tfrac23, \tfrac35, \tfrac58, \tfrac{8}{13},\dots$ — and because the denominators grow only as fast as the Fibonacci numbers (the slowest possible), these are the worst rational approximations any number can have. Hurwitz's theorem makes it exact: every irrational $x$ has infinitely many $p/q$ with $|x - p/q| < 1/(\sqrt5\,q^2)$, and the constant $\sqrt5$ is the best possible — saturated only by the golden mean and its cousins. It sits right at the edge of approximability: the most irrational irrational.

That is why the golden torus is the most robust. A KAM torus is destroyed by its resonances with the rationals; the further its winding number sits from every $p/q$, the harder it is to break. The golden torus, maximally far from all of them, holds out longest — it is the very last to go, at $K_c$. And when it finally breaks, the Fibonacci periodic orbits — its rational approximants — are precisely the family of unstable orbits whose closure is the cantorus.

Diffusion, suppressed

Put the two acts together. Above $K_c$ no full torus spans the plane, so a rotor's momentum is free to wander — it diffuses, $\langle p^2\rangle \approx D(K)\,n$, with $D(K)\approx K^2/2$ for large $K$. But the cantori do not disappear; they linger as partial barriers, and their turnstile flux is a bottleneck. Momentum leaks across each cantorus only as fast as its lobes turn over, so transport that would be free in a fully chaotic plane is instead throttled — the classical face of dynamical localization. Below $K_c$ the golden torus is intact and diffusion is blocked outright; just above, it is slow and cantorus-limited; far above, the barriers are porous and diffusion runs near its free rate. The suppression is the family of unstable orbits doing its work.

An interactive tangle

Static figures freeze the tangle at one iteration count and one scale. Its real character is dynamical and self-similar — the manifolds elongating by $\Lambda$ each kick, folding, and crowding into ever-finer lobes as they approach the saddle. The explorer below grows the manifolds live in the browser: scrub the iteration count $n$ to watch $W^u$ and $W^s$ stretch from smooth separatrix into full tangle, and turn $K$ to watch the saddle sharpen, the island shrink, and the mesh open. Then zoom in — as you dive toward the saddle the manifolds are re-grown with more iterations and finer sampling, so the lace keeps refilling and sharpening the deeper you go. This is the figure Poincaré would not draw, drawn as fast as you can move a slider.

Quantising the rotor

Everything so far is classical. Give the rotor $\hbar$ and it becomes the quantum kicked rotor, the model that first revealed dynamical localization in its quantum form [6]. Its one-period evolution is a Floquet operator — free rotation, then a kick — that we can iterate exactly by fast Fourier transform, alternating between the angle basis (where the kick is diagonal) and the momentum basis (where the rotation is). The effective Planck constant $\hbar_{\text{eff}}$ sets how finely the wavefunction can resolve the classical phase space.

Two things happen, both tied to what we built above. First, the momentum that classically diffuses across the cantori instead localizes: the quantum state spreads for a while, then freezes into an exponential profile $|\psi(p)|^2\sim e^{-|p|/\ell}$, pinned by the same partial barriers — quantum interference finishing the job the cantori started. Fishman, Grempel and Prange showed this is precisely Anderson localization, the map carried onto a 1-D disordered lattice [7]. Second, the Floquet eigenstates do not spread evenly over the chaotic sea; a subset piles up along the short unstable periodic orbits — scars [8] — their Husimi distributions settling right onto the tangle we drew. The wave that no ray can sit on, sitting there anyway.

Scars — the wave on the orbit no ray can hold

Dynamical localization was a statement about momentum. Turn now to the eigenstates themselves, in phase space. Most Floquet states of a chaotic map spread evenly over the sea, as ergodicity would suggest — but a special few do not. They pile up along a short unstable periodic orbit, exactly where no classical trajectory can linger: the anomalous enhancement Eric Heller named a scar [8].

The clearest one sits on our hyperbolic fixed point at $\theta=0$. Below is its Husimi distribution — the state's density smeared over coherent states, its portrait in phase space — as we make the problem steadily more semiclassical, shrinking $\hbar_{\text{eff}} = 2\pi/N$ toward zero. At large $\hbar$ the scar is a coarse blob; as $\hbar\to0$ it tightens onto the saddle and its density streams out along the unstable manifold (cyan), tracing the tangle out to the log-time $t_E \approx \ln(1/\hbar)/\lambda$ set by the Lyapunov exponent $\lambda$. The wave settles onto the very orbit a ray must flee — the quantum flesh finding the classical bone, and the same physics that Part IV will carry back to the folded modes of the cavity.

Coda

From two lines of arithmetic we drew the whole apparatus of chaos. A rotor that conserves its momentum, kicked, breaks its tori one by one; at its origin sits a hyperbolic saddle whose stable and unstable manifolds cross transversally and then infinitely often, weaving the homoclinic tangle Poincaré discovered in the heavens and declined to draw. The tangle's lobes form a turnstile that meters the flux between order and chaos; and when the last golden torus shatters into a cantorus — a fractal wall built from a whole family of unstable periodic orbits — that flux becomes a bottleneck, throttling transport into dynamical localization. Quantised, the rotor localizes harder still: its momentum freezes into an exponential, and its eigenstates scar onto the very orbits no ray can hold.

That last picture — a wave living on an unstable orbit family, holding where classical intuition says it cannot — is exactly what Part IV needs. There, the phase space is a real optical microcavity, and the classical ceiling is a theorem: Mather's [9] says that once a billiard boundary is pushed non-convex, its whispering-gallery caustics are destroyed and no ray can cling to the wall. And yet, in the wave regime, a class of WGM-like modes survives anyway, riding an unstable periodic-orbit family in a strongly non-convex cavity — the folded chaotic whispering-gallery modes [10]. The tangle and the cantorus we drew here in the kicked rotor are the skeleton those modes are built on. That defiance of the classical no-caustic theorem is the culmination of the series, and where Part IV begins.