Paper 14
Strange attractors and fractal dimension
A whitepaper on the geometry that chaotic motion settles into: why some attractors are called strange, how stretching and folding builds them, what non-integer dimension means, and how they differ from points, loops, and tori.
Where a system ends up
Start a dissipative system running and let it go. It will forget its exact starting point. Whatever transient wobble it began with dies away, and the long-run motion collapses onto a smaller set of states that the system returns to and keeps visiting forever. That set is the attractor. It is the answer to a simple question: after everything settles, where does the motion live?
The companion paper in this corpus, "The limit of prediction," used the Lorenz system to explain why deterministic rules can still be unpredictable, and it named the Lyapunov exponent that measures the divergence. This paper is about the shape that divergence happens on. Sensitive dependence is a fact about neighboring trajectories pulling apart. An attractor is the object they pull apart on. To see why a strange attractor is genuinely a new kind of geometric thing, it helps to first see the ordinary kinds it is not.
The three tame attractors
Most attractors are simple, and there are three simple families. Understanding them is what makes the fourth one look strange.
A fixed point is a single state the system falls into and stays. A marble rolling in a bowl comes to rest at the bottom; a hot cup of coffee cools to room temperature and stops. Perturb it a little and it returns. In state space the attractor is one point, a zero-dimensional object.
A limit cycle is a closed loop the system orbits forever with a single repeating period. A heartbeat, a steady clock pendulum with a driving spring, a stable predator-prey oscillation. The motion is not at rest, but it is perfectly periodic: wait one period and everything is exactly as it was. In state space the attractor is a one-dimensional closed curve.
A torus is what you get when two independent periodic motions combine and their frequencies do not divide evenly into each other. Think of a point moving around a small circle while that whole circle is dragged around a big circle, with the two periods incommensurate. The trajectory wraps around the doughnut surface forever, never quite closing, densely covering a two-dimensional surface. This motion is called quasiperiodic. It looks complicated, but it is not chaotic: two nearby starts stay close, because the surface never stretches them apart.
| Attractor | Dimension | Long-run motion | Nearby trajectories |
|---|---|---|---|
| Fixed point | 0 | At rest | Converge to the point |
| Limit cycle | 1 | Periodic | Stay parallel on the loop |
| Torus | 2 (or higher) | Quasiperiodic | Stay close, no stretching |
| Strange attractor | Non-integer | Bounded, non-repeating | Diverge exponentially |
The three tame attractors share a property: they are predictable. Errors do not grow. That is what a strange attractor breaks.
What makes an attractor strange
A strange attractor holds two facts together that seem to contradict each other.
The motion is bounded. The trajectory never runs off to infinity; it stays inside a finite region of state space forever. In the Lorenz system it circulates around two lobes and never leaves them.
The motion is also non-repeating. It never settles to a point, never closes into a cycle, and never lands on a smooth torus. It keeps producing genuinely new states for all time, threading through the same region again and again without ever exactly retracing its path.
On top of those two, a strange attractor carries sensitive dependence living on the attractor itself. Pick two points that both sit on the attractor, as close together as you like, and their trajectories separate exponentially, even though both are confined to the same bounded set forever. This is the crucial move. The divergence is not the trajectory escaping to somewhere else. It is happening inside the object. Two paths on the same finite shape drift apart while both remain trapped on it.
That combination is what forces the geometry to be strange. If you must pack exponentially-diverging trajectories into a bounded region and never let any of them cross or repeat, an ordinary smooth surface will not do. The set has to be intricate at every scale. The word "strange," introduced by David Ruelle and Floris Takens in 1971, names exactly this: an attractor that is not a point, a curve, or a smooth surface, but a fractal.
Stretching and folding: how it is built
There is a physical mechanism behind the intricacy, and it is worth picturing directly because it is the whole story. A strange attractor is built by two operations repeated forever: stretch, then fold.
Stretching pulls nearby points apart. This is sensitive dependence in action: the local separation of trajectories, the positive Lyapunov exponent. Stretching alone would send everything to infinity and there would be no bounded attractor.
Folding bends the stretched-out sheet back onto itself so the motion stays inside the finite region. This is what keeps the attractor bounded despite the relentless stretching.
The standard image is taffy or dough. A baker stretches the dough to twice its length, folds it back over, stretches again, folds again. Two grains of flour that started touching end up far apart after enough passes; that is the stretching. Nothing ever leaves the mixing bowl; that is the folding. Now do this infinitely many times. The dough develops layers within layers within layers, structure at every magnification, never smooth no matter how far you zoom in. That infinitely-layered object is a strange attractor. The stretching gives you unpredictability, the folding gives you boundedness, and together they manufacture the fractal.
Fractal dimension: what non-integer means
We are used to dimensions being whole numbers. A point is 0, a line is 1, a surface is 2, a solid is 3. A strange attractor has a dimension that falls between these, and the number is a real, measurable property of the shape.
The cleanest way to see why comes from the box-counting idea. Cover the object with a grid of boxes of side length r, and count how many boxes N(r) contain any part of it. For familiar shapes, N grows as a whole-number power of 1/r: a line needs about (1/r) boxes to cover, a filled square needs about (1/r) squared. The exponent is the dimension:
dimension = log N(r) / log (1 / r) as r shrinks toward zero
For a fractal, this exponent comes out non-integer. The classic teaching example is the Cantor set: take a line segment, delete the middle third, then delete the middle third of each remaining piece, forever. What is left is more than nothing but less than a line. Box-counting gives it dimension log 2 / log 3, about 0.63. It sits genuinely between a point and a line.
A strange attractor is the same situation in higher dimensions. The infinite folding leaves a set that is more than a surface but less than a solid, and its dimension records how thoroughly the folding has packed structure into the space.
| System | State-space dimension | Attractor dimension (approximate) |
|---|---|---|
| Lorenz | 3 | about 2.06 |
| Henon map | 2 | about 1.26 |
| Rossler | 3 | about 2.01 |
The Lorenz number is telling. The attractor lives in three-dimensional state space, but its dimension is just above 2. It is nearly a surface, but not quite, and that "not quite" is the fractal thickness that lets exponentially-diverging trajectories coexist without crossing. The tabulated values are standard numerical estimates, not exactly known constants; different definitions of dimension give slightly different numbers.
Two worked examples: Lorenz and Henon
The Lorenz attractor comes from three coupled differential equations modeling convection, written out in full in the companion paper. Its trajectory circulates around two lobes, switching from one to the other in a sequence that never repeats. If you slice the attractor with a plane and look at where trajectories pierce it, you do not see a smooth curve. You see a curve that, magnified, reveals itself to be many nearly-parallel strands, and each strand under further magnification splits into more strands. That endless layering is the fold structure made visible. It took a long time to establish as pure mathematics: the claim that the Lorenz equations really do produce a strange attractor was proved rigorously by Warwick Tucker in 2002, decades after Lorenz first drew it.
The Henon map strips the phenomenon down to arithmetic. It is not a flow in continuous time but a two-line rule you iterate step by step:
x_next = 1 - a*x^2 + y
y_next = b*x
with the standard chaotic values a = 1.4 and b = 0.3. Michel Henon introduced it in 1976 to make a strange attractor you could examine closely with a simple computer. Iterate the map and the points trace a shape that looks, at a glance, like a few smooth boomerang-shaped arcs. Zoom into any arc and it resolves into several finer arcs. Zoom into one of those and again it splits. The self-similar layering is the same Cantor-set structure, riding along the folding direction of the map. The Henon attractor is the cleanest demonstration that strangeness needs neither many variables nor continuous time; two numbers and one quadratic term produce a full fractal attractor.
The connection to memory, kept honest
This corpus uses dynamical systems as a disciplined metaphor for memory, and the companion paper spelled out the bridge: a consolidated memory is like a trajectory settled onto an attractor, stable and self-restoring, and the labile window of reconsolidation is a moment when that settledness is briefly suspended. The geometry in this paper sharpens one word in that picture. Stability is not one thing. A fixed-point attractor is rigidly stable; nudge it and it returns to the exact same state. A strange attractor is stable as a set but never stable as a state, because it never sits still and neighboring points always diverge.
This is an analogy and nothing more. No memory has a measured fractal dimension, no basolateral-amygdala circuit has been shown to be a strange attractor, and the animal-to-human translation gap that runs through the reconsolidation evidence is untouched by any of this mathematics. The strong evidence that a reactivated memory becomes briefly labile comes from rodent work, established through amygdala studies such as Nader, Schafe and LeDoux (2000); the extension to human memory, through reactivation-extinction and clinical reconsolidation, remains emerging and in places contested. What the geometry offers is vocabulary with real content. It lets "stable" mean two genuinely different things, a resting point versus a bounded-but-restless set, and that distinction is a cleaner way to hold the idea that a memory can be both durable and, at the right moment, movable. The mathematics is precise. The bridge to memory is a metaphor, and it stays labeled as one.
What to take away
- An attractor is the set a dissipative system settles onto after transients die. Fixed points, limit cycles, and tori are the three tame kinds, all predictable, with whole-number dimensions of 0, 1, and 2.
- An attractor is strange when it holds three things at once: bounded motion, non-repeating motion, and sensitive dependence living on the attractor itself, so trajectories diverge while staying trapped on the same finite set.
- Strange attractors are built by stretching, which gives unpredictability, and folding, which gives boundedness, repeated forever. The result is a fractal.
- Fractal dimension is non-integer and measurable by box-counting. Standard estimates put the Lorenz attractor near 2.06, just above a surface, and the Henon attractor near 1.26, just above a line.
- Lorenz and Henon are the standard examples. Henon shows that two numbers and one iterated equation are enough for a full strange attractor.
- For this corpus, the payoff is one sharpened word: strange attractors let "stable" mean a restless bounded set rather than a resting point, a distinction the reconsolidation metaphor leans on. This is analogy, not mechanism.
Sources
- Lorenz (1963). Deterministic Nonperiodic Flow. Journal of the Atmospheric Sciences.
- Henon (1976). A two-dimensional mapping with a strange attractor. Communications in Mathematical Physics.
- Ruelle and Takens (1971). On the Nature of Turbulence. Communications in Mathematical Physics. Origin of the term "strange attractor."
- Tucker (2002). A Rigorous ODE Solver and Smale's 14th Problem. Foundations of Computational Mathematics. Rigorous proof that the Lorenz attractor exists.
- Strogatz. Nonlinear Dynamics and Chaos. For the classification of attractors, stretching and folding, and fractal dimension.
- Nader, Schafe and LeDoux (2000). Fear memories require protein synthesis in the amygdala for reconsolidation after retrieval. Nature. For the animal reconsolidation evidence referenced in the memory section.
- Companion paper in this corpus: "The limit of prediction" (02), for the Lorenz equations, Lyapunov exponents, and the memory-in-motion bridge.
- Project catalog: docs/chaotic-systems-catalog.md, for the Henon and Lorenz entries and attractor properties.