Paper 02
The limit of prediction: chaos, sensitive dependence, and memory in motion
A whitepaper on deterministic chaos, why a fully determined system can still be unpredictable, and what a trajectory reshaped by a small nudge has to do with a memory reshaped in its labile window.
The puzzle chaos solves
Two facts seem to contradict each other. First, many physical systems are deterministic: give the same starting state and the same rules, and you get the same future, every time, with no randomness anywhere. Second, some of those same systems are unpredictable in practice: the weather, a double pendulum, a dripping faucet near the wrong flow rate. How can something be fully determined and still unpredictable?
Chaos is the resolution. A chaotic system is deterministic, but it amplifies the tiniest difference in starting conditions until two nearly identical starts diverge completely. Since you can never specify a starting state to infinite precision, and the system magnifies whatever imprecision you have left, prediction has a horizon. Beyond it, determinism buys you nothing.
This is not a story about noise or complexity. The systems can be simple, three numbers and three equations, and there is no randomness in them at all. The unpredictability is manufactured by the dynamics themselves.
Lorenz and the three equations
Edward Lorenz found this in 1963 while modeling convection with a drastically simplified set of three coupled differential equations:
dx/dt = sigma*(y - x)
dy/dt = x*(rho - z) - y
dz/dt = x*y - beta*z
with the canonical chaotic parameters sigma = 10, rho = 28, beta = 8/3. Here x, y and z track the state of a rolling convection cell. The system is deterministic and has only one nonlinear pair of terms (x times z, x times y), yet its trajectory never repeats and never settles.
The famous origin story is real: Lorenz restarted a simulation from a printout that had rounded the state from six decimals to three, expected the run to retrace the original, and watched it diverge into a completely different weather. The rounding in the fourth decimal, one part in a thousand, grew until the two runs had nothing in common. That is sensitive dependence on initial conditions, the defining property of chaos, and it is why the interactive weather tool in this project lets you set a starting value to only a few decimals and then watch your forecast track the truth and peel away.
The trajectory traces the Lorenz attractor: a bounded, two-lobed shape that the state orbits forever without ever crossing itself or repeating. It is a strange attractor, strange because it is not a point or a loop but a fractal set of intermediate dimension. The motion is confined (it never runs off to infinity) and unrepeating (it never closes into a cycle) at the same time.
Measuring the divergence: the Lyapunov exponent
Sensitive dependence can be made quantitative. Take two starting points separated by a tiny distance, call it delta-zero. In a chaotic system their separation grows on average exponentially:
delta(t) ~ delta_0 * e^(lambda * t)
The number lambda is the largest Lyapunov exponent. If lambda is positive, nearby trajectories separate exponentially and the system is chaotic. If lambda is zero or negative, errors stay bounded or shrink and the system is predictable. The size of lambda is the rate at which the future becomes unknowable.
This gives a clean formula for the prediction horizon. If you can tolerate an error of size delta-tolerate and you start with uncertainty delta-zero, you stay accurate until roughly:
T_horizon ~ (1 / lambda) * ln(delta_tolerate / delta_0)
Two consequences fall out of that logarithm, and both are counterintuitive.
Precision has diminishing returns. Because the tolerance enters through a logarithm, cutting your initial error in half does not double your usable forecast. Every factor of ten of extra precision buys the same fixed amount of extra time, ln(10) / lambda. That is why the weather tool tells you each additional decimal of precision buys a fixed number of model-time units and no more. You can delay the drift. You cannot prevent it.
The horizon is a wall, not a slope. Since the growth is exponential, error stays negligible for a while and then explodes over a short span near the horizon. The forecast looks perfect, then good, then useless, in quick succession. This is exactly the on-screen behavior of the weather forecast: it tracks, then drifts, then diverges hard.
Chaos you can build: the Malkus waterwheel
Chaos is not a numerical artifact. The Malkus waterwheel is a physical machine whose equations are exactly the Lorenz equations. Water pours into cups around the rim of a tilted wheel, each cup leaks, and gravity turns the wheel from whichever side is momentarily heavier. Under steady inflow the wheel does not settle into steady spin. It speeds up, slows, stalls, and reverses, never repeating, because the wheel's spin rate is precisely the Lorenz x variable.
The waterwheel matters because it shows the unpredictability is in the dynamics, not in the model or the computer. A real object, fed a perfectly constant input, produces a spin history you cannot forecast past its horizon. The interactive Malkus tool in this project drives the same equations and shows the reversals directly.
A whole zoo, not one system
Lorenz is one entry in a large taxonomy. The catalog compiled for this project holds 177 distinct, source-verified chaotic systems, grouped by the kind of mathematical object they are:
| Family | Count | What they are |
|---|---|---|
| Continuous flows (ODEs) | 111 | Systems evolving smoothly in time: Lorenz, Rossler, Chua's circuit, the Sprott minimal flows |
| Discrete maps | 39 | Systems that jump in steps: the logistic map, Henon, the Chirikov standard map |
| Delay equations | 4 | Systems whose rate depends on their own past: Mackey-Glass |
| Spatiotemporal (PDEs, lattices) | 10 | Chaos spread over space as well as time: Kuramoto-Sivashinsky, Rayleigh-Benard convection |
| Hamiltonian and conservative | 10 | Energy-preserving chaos: Henon-Heiles, the double pendulum, the three-body problem |
| Billiards and other | 3 | Chaos from geometry alone: the Sinai and Bunimovich stadium billiards |
A few distinctions from that taxonomy are worth carrying:
Dissipative versus conservative. Dissipative systems (Lorenz, Rossler) lose volume in state space and collapse onto a strange attractor: whatever you start with, you end up on the attractor. Conservative or Hamiltonian systems (the double pendulum, billiards) preserve volume and have no attractor. Their state space is a mix of orderly islands and a chaotic sea, and the transition between them is governed by KAM theory. The double pendulum is the everyday example: a single rigid arm swinging is boring; add a second hinged segment and small differences in release explode into wildly different swings.
Flows versus maps. The logistic map, x-next = r times x times (1 minus x), is a single line of arithmetic, yet as you raise r it period-doubles into full chaos. Chaos does not need many variables or continuous time. One number iterated is enough.
Hyperchaos. Some higher-dimensional systems have more than one positive Lyapunov exponent, meaning they stretch in several independent directions at once. They are even harder to predict, and they are why four and five-variable systems are catalogued separately.
Memory in motion, stated carefully
The name of this work is a claim about motion, and dynamical systems are where the word earns its precision. Here is the bridge, drawn as an analogy and labeled as one.
A consolidated memory is like a trajectory that has settled onto an attractor. It is stable: perturb it a little and it returns to the same basin. Ordinary attempts to change it are fighting the system's own tendency to fall back. The labile window opened by reconsolidation is the moment when that stability is briefly suspended, when the trajectory is off the attractor and a small, well-timed nudge can redirect it toward a different basin entirely. Sensitive dependence is the reason a small input at the right moment can have a large downstream effect. In a locked, stable system a small nudge washes out. In a system that is momentarily labile and sensitive, a small nudge is decisive.
This is a metaphor, and the honest version keeps the seams visible. A brain is not literally the Lorenz system, memories are not literally points on a strange attractor, and no Lyapunov exponent has been measured for a fear memory. What the dynamical-systems picture provides is not a mechanism but a vocabulary with real content: stability and basins, labile intervals, sensitive dependence, and the difference between a nudge that washes out and a nudge that redirects. Those distinctions are exactly the ones the reconsolidation work turns on. The mathematics disciplines the metaphor, so "memory in motion" means something specific rather than something poetic.
What to take away
- Deterministic does not imply predictable. Chaos manufactures unpredictability from exact rules by amplifying unavoidable imprecision.
- The largest Lyapunov exponent lambda measures how fast. A positive lambda means exponential divergence and a finite prediction horizon.
- The horizon grows only as the logarithm of your precision, so more precision buys a fixed, small extension and never a cure.
- Chaos is common and physical, spanning a large taxonomy from one-line maps to fluid turbulence, and demonstrable in a real waterwheel.
- Used carefully, this gives "memory in motion" a precise meaning: a brief interval where a stable trace becomes sensitive, and a small timed nudge can redirect it.
Sources
- Lorenz (1963). Deterministic Nonperiodic Flow. Journal of the Atmospheric Sciences 20, 130 to 141.
- Strogatz, Nonlinear Dynamics and Chaos, for the Malkus waterwheel derivation and the Lyapunov-horizon treatment.
- Project catalog: docs/chaotic-systems-catalog.json and docs/chaotic-systems-catalog.md (177 source-verified systems), from which the taxonomy and counts above are drawn.
- Interactive tools in this project: /weather (the prediction horizon) and /malkus (the physical Lorenz system).