Ask where it is heading, not where it is
2026-07-25 · Dynamics & Unpredictability
The same number can be good news or bad news. What settles which is not the number itself, but the direction it is moving in — and how close it is sitting to a dividing line you cannot see.
One ball rests at the bottom of a bowl. Another rests on top of a hill. Photograph them and the pictures are identical: both still, both going nowhere. Give each the smallest nudge and one rolls back home while the other never returns.
We describe things by their current value: revenue this month, weight this morning, how many people are left on the team, how you feel. The trouble with that description isn't that it's inaccurate. It's that it cannot answer the only question that matters — what happens next.
This issue swaps in a different notation. Take each key variable as an axis; the whole system at any instant collapses to a single point, and as time passes that point traces a line. Something strange follows: questions that used to require calculation become questions you answer by looking. Will it settle or keep swinging? Will it come back to where it was? On this picture, those questions have shapes.
Start with something as dumb as possible: a swinging pendulum — the kind in an old clock, or a bunch of keys swinging from your finger.
The angle alone will not describe its situation. Thirty degrees might mean it is swinging outward and about to go further, or swinging back. Same angle, two completely different futures. One more number fixes it: the angular velocity — how fast it is turning right now, and which way. Give me the angle and the angular velocity and everything the pendulum does next is settled.
That pair is the system's state. State is not "everything about it" but the smallest set of numbers that pins down the future — the pendulum's colour, whose house it hangs in, how many times it swung yesterday, none of that belongs.
Now the move that this whole issue rests on: treat those two numbers as the horizontal and vertical axes of a picture. Angle across, angular velocity up. At any instant the pendulum is one point on that picture; as it moves, the point drags out a curve called a trajectory. The picture is the state space (physicists say phase space), and once a representative set of trajectories has been drawn on it, the whole thing is a phase portrait. → ref · Phase space & attractors
Look at the right-hand picture. Each swing of the pendulum is one loop around the centre; friction makes every loop a little smaller, so the trajectory becomes a spiral winding inward until it reaches the point where the pendulum hangs still. Without friction (the blue dashed curve) the trajectory closes: one loop returns exactly to its starting point, and then repeats forever.
On the left-hand plot those two fates take a long stare to tell apart. On the right they are shapes: spiral means it stops, closed loop means it repeats. That is the entire payoff of changing notation.
One constraint is worth carrying forward: in a deterministic system, two trajectories never cross. A crossing would mean one state with two different futures, and the definition of state rules that out. This sounds like a triviality and it bites later — it means that on a flat sheet (two variables) a trajectory must either wind into a point, close into a loop, or run away. It cannot wander forever without repeating itself. For that kind of endless non-repetition you need at least three axes. That is why chaos does not start below three dimensions.
The special places in a state space are the ones where the flow is zero: land there and the system stops moving. These are fixed points (also called equilibria). A pendulum hanging straight down is one.
But the pendulum has a second fixed point: balanced upside down, perfectly vertical, pointing up. In principle it can indeed stay there. You know perfectly well that it won't — and it is worth being precise about what "won't" means. It isn't that this is not a fixed point. It is an unstable one.
The whole distinction is one test: nudge it, and see whether it comes back. If it does, the point is stable (nearby trajectories head toward it). If it doesn't, it is unstable (nearby trajectories run away). And there is an in-between case where a nudge one way comes back and a nudge the other way leaves for good: semi-stable.
That dismantles a very common inference: "it has been stable for years" is not evidence of stability. The ball on the hilltop can also sit still indefinitely — as long as nobody touches it. There is only one kind of evidence: it was disturbed and it came back. An untested calm and a structurally stable state look identical in a photograph.
Stop treating "nothing has gone wrong" as evidence of stability; that is the absence of an experiment. Run a controlled small disturbance and watch whether it recovers on its own, and how long it takes. Cut one dependency, take one person out for a week, simulate a small run on the bank, switch off an automation. What you record is not "did anything break" but the recovery time with no human intervention. A lengthening recovery time means the fixed point is getting shallower, even while every current reading still looks normal.
A fixed point is only the simplest ending. Pull back and ask where a system settles after a long time, and the answer is its attractor: a region that nearby trajectories are drawn into and do not leave.
There are only a few kinds. The picture is faster than the definitions.
① Fixed point. The system stops. A cup of coffee gone cold, a project that has spent its money, a meeting where nobody speaks any more.
② Limit cycle. The system runs round a closed loop forever. The important part is that the loop attracts: trajectories starting outside are drawn in, trajectories starting inside are pushed out, and both end up on the same loop — so the amplitude and the period are set by the system, not by where you started. Pacemaker cells in the heart, fireflies flashing in unison, predator and prey numbers rising and falling: all this family. This is the one most often misread: an oscillation does not need a metronome outside it. Regular ups and downs are not evidence that someone is beating a drum; the structure keeps its own time.
③ Torus (also: quasi-periodic). Two rhythms run at once and their periods are not in whole-number ratio, so the curve winds on forever without ever exactly repeating itself — while being perfectly predictable and not the least bit wild.
④ Strange attractor. The trajectory is confined to a finite shape, never repeats, and two trajectories that start almost on top of each other pull further and further apart. That is chaos, and it gets a whole issue next time.
Incidentally, this vocabulary is not restricted to smoothly varying things. Rules that jump cell by cell work the same way: in → ref · Game of Life, a "still life" is a fixed point, a "blinker" is a limit cycle of period 2, and a "glider" is a cycle up to translation.
Identify the type first, then pick the forecasting method — doing it the other way round is where most forecasting disasters come from. Fixed point: point forecasts mean something; ask what the steady state is. Limit cycle: do not forecast "when it will rise", forecast the phase (where in the cycle you are), and stop hunting for an external cause behind every swing. Strange attractor: give up point forecasts, work on bounds and distributions instead. A ten-minute test: scatter-plot your historical series with the current value on one axis and the change since last period on the other. If the cloud of points wraps into a loop, you are looking at a limit cycle — and "who caused this downturn?" was the wrong question from the start.
A system often has more than one attractor. Which raises the question: starting from here, which one do I end up in?
Collect every starting point that eventually lands in attractor A and you have A's basin of attraction. The line between two basins is the separatrix. Every point on that line has the property that a nudge one way and a nudge the other way lead to entirely different endings.
The thing to stare at is that pair of nearly coincident starting points. Their current states are all but identical: positions differ by a hair, velocities are the same. Any dashboard reading current values calls them the same situation. Their endings are different worlds.
Which gives a distinction that is usually hard to state and is one sentence here: distance to the endpoint and distance to the divide are unrelated quantities. A company can look healthy on paper (far from the endpoint called insolvency) while sitting one step from the divide. A person can have every test result inside the normal range while being one infection away from crossing. Numerical headroom is not structural headroom.
It also explains something that happens over and over and confuses people every time: the same effort produces results differing by orders of magnitude depending on where it is applied. Deep inside a basin the system eats your effort — let go and it slides back. Near the separatrix a small push buys an entirely different ending. "Where there's a will there's a way" and "effort is pointless" are both true accounts; they just come from different places on the same picture.
List the few things you are currently pushing on and rank them by proximity to a divide, not by size of impact. How to find the divide: go through history for the cases that nearly flipped and didn't — what they have in common are the coordinates of the boundary. Then do two things: move resources toward the top of that list, and admit that the ones at the bottom are feeding the system effort rather than changing its ending. The test is blunt: does it slide back when you stop? If it slides back, that is an operating cost, not an achievement.
State space is one of the twentieth century's most useful thinking tools, which is exactly why it is such comfortable material for abuse: its vocabulary — attractor, divide, steady state — sounds like an explanation while promising nothing. Draw the boundaries before using it.
First, you have to know what the axes are. The pendulum is easy to draw because we know for certain it needs two numbers. How many numbers does a team need, or a mood, or a market? Nobody knows. Choose the wrong variables, or too few, and the "phase portrait" you produce is a self-soothing picture. There is a technical remedy called delay embedding (Takens, 1981): use lagged copies of a single time series as your axes and reconstruct the shape of the attractor. Its preconditions are strict, though — the system must be deterministic, of modest dimension, the series long, the noise small. Almost no real social or economic series qualifies.
Second, a phase portrait holds only while the parameters hold still. The entire picture — where the fixed points are, where the divide runs — is set by the system's parameters. Change them and the picture deforms: attractors move, split, or vanish outright, which is the subject of Topic 10. So "wait for it to return to normal" carries a hidden premise: that the normal is still there. Often the attractor you are waiting for has already ceased to exist, and the current reading cannot show you that.
Third, noise is not a detail. In textbooks trajectories glide cleanly into their attractor. Real systems are perturbed constantly, and if a basin is shallow the system will hop between basins by itself — noise-induced transitions. The consequence: a state that is "mathematically stable" may in a noisy world leave three times a year. Stability has to be quoted together with the size of the noise, or it is not a claim.
Fourth, and most important: do not use this vocabulary as an explanation. "His personality is an attractor", "this company's culture has inertia" — these sound dynamical and forbid nothing. Real phase-portrait talk costs something: you must name which two or three measurable quantities are the axes, over what timescale, and what observable says you are near the divide. If you cannot pay that, say "it has inertia" — three words, equally accurate, and not pretending to be science.
Before saying "attractor", write down three things: ① the axes — which two or three measurable quantities; ② the timescale — how long counts as "long run"; ③ an observable proxy for the divide — what shows up when you are getting close. If you cannot assemble all three, delete the word and use plain language. This rule applies to this site as much as to anyone.
It does, and that is precisely the trouble. A system with memory — where today's behaviour depends on the past week — only fits standard state-space form if the whole past week becomes extra axes, and the dimension jumps. Real people and organisations carry long memories, which is one of the hardest technical obstacles to drawing phase portraits of social systems, and the reason methods like delay embedding were invented.
No, and this issue's language pins the difference. Stability is a property of the fixed point: nudge it and does it return. Resilience is about the size of the basin: how big a shove can it take and still be inside. A system can be extremely stable (small disturbances damp out fast) and extremely fragile (the basin is narrow, so a slightly larger disturbance takes it over the edge). Topic 38 is devoted to this distinction.
Yes. This is the watershed between complex adaptive systems and physical ones: the pendulum doesn't know which picture it is drawn on, and people do. Once "which basin I am in" becomes a public description, it becomes a new input — a measure that becomes a target stops working (Topic 26). So state space applied to people is best used as after-the-fact description, not as coordinates announced in advance.
Because trajectories cannot cross. On a plane, a curve that keeps moving forward and cannot pass through itself gets trapped by the region it has enclosed, and can only wind into a point or close into a loop — the intuitive content of the Poincaré-Bendixson theorem. With a third axis the curve can pass "above" itself, so it can fold forever inside a bounded space without repeating. That folding is exactly what chaos requires.