TOPIC 5 · PHASE A

Nonlinearity

Twice the cause, seldom twice the effect

2026-07-23 · What Counts as "Complex"

We are trained from childhood to be linear animals: spend twice the effort, expect twice the reward; when something big happens, go hunt for a proportionately big cause. But most relationships that matter in the real world are not shaped like that — disproportion is the rule, and the straight line we trust most is the rare exception.

Stir sugar into coffee: the first spoon is distinctly sweet, the fifth is barely noticeable. Now oversalt a pot of soup by a pinch and you ruin the whole thing. Both are "a little bit more," yet one barely responds and the other flips everything — and the difference isn't how much you add, it's where you started.

This is nonlinearity: output does not track input in proportion. It sounds like a truism, but the consequences are heavy — it means you cannot infer the whole from a small sample, cannot project "at the current rate," and cannot understand a system by pairing causes with effects. The linear-extrapolation machine that runs by default in our heads fails these relationships systematically, and in a predictable direction.

This issue covers just one basic thing: what "not proportional" actually buys you, and why it makes prediction so hard. Later topics — chaos, criticality, power laws — are all different faces of nonlinearity. This one is the ground they stand on.

01Double the Cause, Not Double the Effect

First, what "linear" means, so we can see how it breaks. A linear relationship has two well-behaved properties. One is proportionality: double the input, double the output. The other is superposition: the result of doing two things together equals the sum of doing each alone. Put two identical lamps on a desk and the total brightness is exactly twice one — the world really does hold a stock of relationships this obedient.

Nonlinearity breaks at least one of these. Once broken, the relationship takes on a few typical shapes: ① diminishing returns (steep then flat) — the first three months of training bring dramatic gains, and afterward the same workload buys less and less; ② accelerating (flat then steep) — compound interest, word-of-mouth crossing over, the interest on technical debt: each later step covers more ground; ③ the S-curve (flat, then steep, then flat) — many real processes look like this: nothing at the start, a sudden violent response in some band, then saturation.

Here is the crux: on a nonlinear curve, the effect of "one more unit of input" depends on where you currently sit on the curve. On the flat part you can push as hard as you like with almost nothing to show; on the steep part a light nudge is a big jump. So "return on investment" is not a fixed number — it is a function of your position.

The same +Δ of input differs by an order of magnitude — just from where it lands input → output linear (proportional) S-curve same +Δ ≈ nothing same +Δ big jump flat toe steep part
Same curve, same-sized input increment: on the flat toe almost nothing happens; on the steep part it is a big jump. "Return on investment" is a function of position, not a constant.

So a counterintuitive line holds here: a small input levering a big reward and a small slip triggering a big disaster are two faces of the same thing. Both say "you happen to be standing on the steep part" — the only difference is whether that jump goes your way or against you. If you want the upside, you have to accept the downside that comes with the same slope.

🎯 DECISION LINE

Before any push, judge which part of the curve you are on. Flat part: piling on more is waste — change the arena, don't double the effort. Steep part: this is leverage and minefield at once — same spot, push with it for outsized gain, push against it for outsized loss — so the right move on the steep part isn't "harder," it's controlling direction and step size. Replace the default reflex "I need to try harder" with the question "which part of the curve am I on?"

🌀 Economics · Diminishing marginal utility The first glass of water saves your life; the tenth is for washing your hands — the same glass loses value as you already hold more. This two-hundred-year-old law of economics is exactly a saturating nonlinearity. What it overturns is a linear assumption we use daily: "more is always better." More isn't always better; the slope of "better" keeps falling, and past some point it turns negative (drink water to the point of poisoning). Whether "a bit more" is worth it can never be answered apart from "how much do you already have."

02Nothing Happens, Until Everything Does

One kind of nonlinearity is especially deceptive: the threshold. Before you reach a certain point, no matter how you push, the system doesn't budge, as if it isn't listening at all; the moment you cross that point it flips over entirely — and often cannot go back.

The cleanest example is water. Heat it from 99°C up to 99.9°C and it is still water, with no sign of boiling; cross 100°C and it becomes vapor. The temperature was pushed up smoothly, one degree at a time (the continuous cause), but the phase snaps across (the discontinuous effect). Continuous cause, broken effect — that is the sharpest feature of a threshold.

More extreme: supercooled water can be carefully cooled below 0°C and still not freeze. Drop in a tiny ice crystal, or tap the glass, and the whole thing freezes solid in a second or two. The trigger is laughably small, the result is a full phase change — because the system had already been pushed to the edge of the threshold, needing only for someone to touch it.

threshold Cause: rising smoothly, one degree at a time time → Effect: no response, then it all flips over nothing happens everything does time →
Drive a cause up continuously and smoothly (top); the effect jumps discontinuously at the threshold (bottom). The "last push" that crosses the threshold is no different from every push before it.

Thresholds break cause-and-effect pairing entirely. The "last push" that crosses looks like the culprit, but it is identical to every push before it. The straw that breaks the camel's back weighs the same as the first straw — pinning the blame on it mistakes a process of continuous accumulation for a cause that suddenly appeared.

🌀 Materials engineering · The yield point When a metal is stretched, it is elastic at first: let go and it springs back, and force is proportional to deformation — a stretch of clean linearity. But once past the "yield point" it turns plastic, deforms permanently, and soon fractures. The "safety factor" engineers build in is essentially keeping the working stress well away from the yield point. The upshot: "it's been fine all along" is not evidence of safety — the obedience of the elastic stretch is exactly what hides how close you are to the breaking point; linear behavior holds all the way up to the instant it doesn't.

03Your Brain Is a Linear-Extrapolation Machine

Now about us. The human brain has a deeply ingrained default for handling trends: linear extrapolation. See a quantity rising and we instinctively ask, "at this rate, where is it in X time?" — and "at this rate" has quietly assumed a constant rate of increase, i.e., linearity.

Fold an ordinary sheet of paper 42 times and it reaches the Moon. Almost no one guesses this; most people guess something the height of a building. The reason is that each fold doubles the thickness — that is exponential — while our intuition adds. Grains on a chessboard tell the same story: one grain on the first square, doubling each square after, and by the 64th square the wheat required exceeds everything humanity has ever grown.

And the error has a direction, a predictable one: for compound growth we habitually underestimate ("still early"); for threshold jumps we are habitually blindsided ("how did it suddenly…"). Both errors come from the same default — fitting a straight line to a curve that isn't straight, close at the start and wildly off later.

Same data, two continuations — indistinguishable early, worlds apart later time → amount actual (exponential) your linear guess early: nearly identical off by more and more later
Early on, the exponential curve and the linear guess almost coincide — you can't tell them apart. The gap opens with time; by the time it's visible to the eye, the window to act has usually closed.

This is why "the numbers are still small" is the most dangerous reassurance early in an epidemic, and why "only down 5%" looks utterly harmless the day before a leveraged account is wiped out — a linear eye is always half a step behind in an exponential, threshold-laden world, and it errs hardest exactly when it matters most.

🎯 DECISION LINE

Whenever you catch yourself saying "at this rate…", stop — that phrase assumes linearity. Ask two questions instead: is this quantity growing by addition or by multiplication (doubling)? Is there a threshold ahead of it? If it's multiplicative or has a threshold, your gut estimate is guaranteed to be too low and too late — so draw the warning line where it "still feels too early." By the time it "feels about right to act," it usually already is late.

🌀 Epidemiology · The illusion of early exponentials Early in an epidemic or a rumor, the comfort of "only a few hundred so far" rides entirely on linear extrapolation of the current absolute number. But the only informative property of an exponentially growing quantity is its doubling time, not its present size — a two-day doubling time means a few hundred today, tens of thousands in two weeks. The upshot: on the exponential stretch, "the numbers are still small" is not evidence of safety; it is precisely the signal that "you still have time, but the window is slamming shut fast." By the time the absolute number is frightening, the moment to intervene is long gone.

04Where This Breaks Down

"Nonlinear" is one of the most heavily abused incantations in complexity science. Merely saying a system is "nonlinear" carries almost no information — because strictly linear systems are the rarities in nature. What matters is knowing where the idea fails to give you what you want.

First, nonlinear does not mean unpredictable. A shell's trajectory is nonlinear (air resistance goes roughly as the square of speed), and we still compute it with great precision; planetary orbits are nonlinear, and astronomers can announce an eclipse thousands of years out. Nonlinear only means "not proportional" — it does not rule out precise prediction at all. Reading "nonlinear" as "can't be measured, up to fate" confuses it with chaos (sensitive dependence on initial conditions, a later topic). Not proportional ≠ not predictable.

Second, locally almost everything is linear — this is not a loophole, it is why we survive. Any smooth curve, viewed close enough, is approximately a straight line. That is exactly why "linear approximation" is so useful: over the short term and small ranges, treating the nonlinear as linear gives errors small enough to ignore, and engineering, economics, and everyday decisions do this constantly and legitimately. The mistake was never "using a linear approximation," but extrapolating it far past the range where the approximation still holds — and not noticing. The danger isn't the straight line; it is not knowing its expiry date.

Third, and most important: "nonlinear" is not a fig leaf. Saying after the fact that "the system was too nonlinear to foresee" is often just a dodge. Many so-called black swans, seen in hindsight, were standing on a steep segment or a threshold that should have been known — nonlinearity makes precise point prediction hard, yet it often makes the direction and order of magnitude clearer, not murkier: you know full well you are on the steep part, you know there is a threshold ahead, you just don't know which exact day. Swapping "can't predict which day" for "utterly unforeseeable" uses genuine uncertainty to cover for preparation you could have made.

🎯 DECISION LINE

Before saying "this can't be predicted," finish the sentence: is it "I don't know the direction or the magnitude," or "I just don't know the exact timing"? The former is rare, the latter is common. The latter can be prepared for — you don't need to predict the exact moment of a blowout to decide to carry a spare tire. Don't let the word "nonlinear" help you turn "wasn't prepared" into "couldn't be prepared for."

🎒 Scenarios · BigCat

  1. Health & energyYou treat "more training = more progress" as a straight line, so you add volume when you feel good and add more when progress slows — until one day you suddenly strain something, or no amount of sleep brings you back for weeks. Training and recovery is actually an inverted U (steep, then a peak, then negative): the same extra set has a different sign of effect depending on which part of the curve it lands. What to change: don't make "weekly total volume" your only dashboard — add a readout that reflects which part of the curve you're on, such as resting morning heart rate or heart rate variability (HRV, which most sleep wearables can measure). When it turns down, that's your body telling you you're past the peak. The move is simple: cut volume the moment that readout worsens, rather than "gritting through one more push."
  2. Contemplative practiceYou assume meditation is linear: sit 20 minutes a day and calm should grow linearly, day by day. So after two weeks with "no feeling," you start tallying whether the "return on investment" is worth it and toy with quitting. But many inner changes are threshold-shaped — a long stretch of nothing, then one session where something suddenly loosens. The problem isn't the method; it's measuring a threshold process with a linear ruler: settling accounts on "did this single sit pay off" is guaranteed to talk you out of it before you ever cross the threshold. What to change: switch the unit of evaluation from "was this one sit worth it" to "did I hold a long enough unbroken window" — crossing a threshold rides on accumulated total, not the intensity of any single sit. A concrete test: if you're using "didn't feel much today" to vote on "sit again tomorrow or not," you're measuring the wrong thing.

🌀 Crossings · Across Disciplines

Going Deeper

If the local is always approximately linear, how do you know when you've stepped outside the range where the approximation holds?

Watch the second-order information. The error of a linear approximation grows roughly with the square of the distance from the base point, so just track "is the same action having a different effect than last time?" — the moment the effect starts changing, curvature is telling you you've left the straight segment. The real difficulty isn't whether there's a signal, it's that we too readily explain "the effect changed" as external noise rather than "I've left the linear zone."

If "small input, big reward" and "small slip, big disaster" share the same origin, why do we chase the former and ignore the latter?

Because convexity (the gain side) and concavity (the loss side) aren't psychologically symmetric: loss aversion makes us overweight the nearby small loss and underweight the distant large disaster, while compound gains are chronically underestimated. The result is that on the same steep curve we stare only at the upside half. To see it symmetrically, you have to force yourself to draw "what if, on this segment, I'm pushing against it" too.

Under a threshold, "no early response" could mean "far from the threshold, safe" or "about to cross, most dangerous," and they look identical. How do you tell them apart?

This is exactly what the later Topic 36, "critical slowing down," addresses: as a threshold nears, the system recovers more slowly from a disturbance and its fluctuations grow. So to tell the two apart, don't look at the size of the response — look at how long it takes to settle afterward. Which also warns: the calm before a threshold carries no distance information by itself; you have to measure something else.

If nonlinear doesn't mean unpredictable, what actually pushes a system from "nonlinear but computable" into "unpredictable in principle"?

The key step is sensitive dependence on initial conditions — a hair's difference in starting values, a thousand-mile divergence in trajectory, with error amplified exponentially. That's what Topic 8, chaos, will cover. Nonlinearity is a necessary condition for it (a linear system can't be chaotic), but far from sufficient: the vast majority of nonlinear systems aren't chaotic at all.

Does "use a linear approximation" hold in moral and policy judgment too? When does treating complex social consequences as "independent, additive" go badly wrong?

Superposition fails earliest in human systems: two policies each mild on their own can jointly trip a threshold that neither created alone (a bank run, a stampede, an opinion cascade). "Evaluate each item separately, each clears the bar, so the sum is fine" is precisely a linear assumption — and it fails hardest exactly where interaction is strongest and matters most.

Further Reading