Exponential growth always bends in a finite system
2026-08-21 · Scale
A pond of water lilies doubles its covered area every day and fills the pond on day 30. What does the pond look like on day 29? Half empty. Every judgement of the form "we've got plenty of room" is made on day 29.
That exponential growth stops is not in doubt: nothing bounded can keep doubling. The question worth asking is what the stopping looks like — because it has at least three quite different shapes, and the difference between them matters far more than how fast the growth was.
Worse: you almost never can tell, from inside the curve, which segment you are on. Slowing down near a ceiling and pausing between two curves look identical in the data. In 1920 two statisticians took every scrap of US census data then available and computed an upper limit; their curve tracked reality beautifully for the next thirty years — and then the limit was crossed in 1967 and has since been exceeded by about 70%. Their arithmetic was fine. What was wrong was the premise that a fixed limit existed at all.
This site has crossed this thread twice already, from different angles. Topic 9 was about the logistic map — the discrete, one-generation-per-year version, which leads to chaos. Topic 10 was about a system jumping from one steady state to another. This issue is about the other object with the same name: the continuous-time logistic equation, and the three things that happen after you hit a ceiling. Same name, different question.
Start by unpacking the word. Exponential growth means one thing: the amount added in each interval is proportional to the amount already there. Money earning interest, people having children, one user bringing in one more user. Its signature is a fixed doubling time — 7% a year doubles in about ten years; 10% a year doubles in about seven.
A fixed doubling time has a consequence the human mind is almost incapable of feeling: nearly all the drama of exponential growth happens in the last one or two doublings. Half the work of covering the pond gets done on the final day. On day 25 only 3% of the water is covered, and anyone standing at the edge would say this is clearly not a problem.
But in a finite system the growth rate cannot stay constant. Each additional individual is one more claim on the same food, the same land, the same bandwidth. In 1838 the Belgian mathematician Pierre-François Verhulst wrote this down in the simplest possible way, multiplying the exponential by a single discount factor:
growth rate = r × N × (1 − N/K). Here N is the current amount, r is the unconstrained growth rate, and K is the largest amount the environment can sustain — the carrying capacity. The bracket is the whole trick: when N is small it is close to 1 and the discount barely exists, so growth is purely exponential; as N approaches K it goes to zero and growth is snuffed out.
The curve this produces is a stretched S, called logistic growth. It has one landmark worth memorising: the moment of fastest growth is exactly the halfway moment (N = K/2). That turning point is the inflection point — past it the total is still rising, but the rate of rise has begun to fall.
A trap in the naming, while we are here: take this same rule and change time from "continuously flowing" to "settled once a year" and it becomes a different object — the logistic map, whose behaviour is nothing like this and which runs all the way into chaos → ref · The Logistic Map. One formula, two worlds, depending on whether time is continuous or discrete.
The textbook S-curve is gentle: it slides up to the carrying capacity and stays there. In the real world that is the rarest of three outcomes.
Which one you get is decided not by how violent the growth was, but by two unglamorous quantities:
The first is delay — how long it takes to get from "already overloaded" to "you notice you are overloaded" to "you actually stop". A system is always reacting to a state that is already in the past. However hard you brake, if the signal arrives two years late, those two years of growth happen anyway.
The second is how fast the consumed thing regenerates. Grazed grassland grows back in a few years; grazed lichen takes decades. If consumption outruns regeneration, K itself is dragged down — you have not reached the ceiling, you have broken a piece of it.
The combinations give three outcomes: an immediate signal and you glide up to K; delay, but a resource that recovers and you oscillate around K, converging slowly; delay plus a resource that recovers more slowly than it is consumed and you get overshoot and collapse — you sail past, then fall to a level lower than where you started.
The third case has a real-world version of almost cruel clarity. In 1944, to leave an emergency meat supply for personnel stationed there, the US released 29 reindeer on St. Matthew Island in the Bering Sea. No wolves, no bears, no hunters, and a thick mat of lichen that had taken decades to grow. In 1957 the biologist David Klein walked the island and counted 1,350 animals, all fat and healthy. He returned in 1963: 6,000.
Then came the winter of 1963–64. When Klein landed a third time in 1966, 42 reindeer were left — all but one of them female, bones everywhere, the lichen grazed down to bedrock. A 99% loss in three years.
The point is not "there were too many deer". The herd never knows there are too many: a reindeer's reproductive condition tracks its own body state, and its body state reflects last summer's forage — the signal is a full year late. By the time hunger arrived as information, the lichen had already been eaten past the point of recovery. Lichen grows a few millimetres a year. K collapsed, so the curve did not oscillate back down to 6,000; it fell to a few dozen.
(In fairness: later work noted that the snowfall that winter was exceptionally deep, so weather had a hand in it. That does not overturn the structure of overshoot, but it sharpens the lesson — overshoot removes the margin a system needs to survive one bad winter, rather than causing the collapse single-handedly.)
Stop asking "how far are we from the limit" — that number usually cannot be estimated (see section 04). Measure two numbers that can be: ① how long it takes for overload to become visible to you, and ② how long one recovery cycle of the resource you are consuming takes. Whenever ① is clearly larger than ②, your system is structurally committed to overshoot — willpower, discipline and effort do not change that conclusion. The only things that do are those two numbers themselves: shorten the measurement cycle, or switch to a faster-regenerating resource.
The last section treated K as given. But K is never a constant of nature. It measures whichever constraint is currently binding — replace the constraint and you have a different curve.
Agriculture shows this most plainly. For a long time the ceiling on food output was arable land: more people on less land was a hard stop. In the early twentieth century the synthesis of ammonia pulled nitrogen out of the air and the ceiling became fertiliser; later, short-stalked high-yield varieties and irrigation made it seed and water. Each time the old K was not nudged upward — a completely different quantity took over as the ceiling. The old curve still flattened out on schedule; another curve had simply started before it did.
Join the upper edges of a series of such relayed S-curves and you get an envelope curve. From a distance it looks like a smooth exponential; up close it is made of discrete handoffs. Moore's law is the most famous envelope there is: bipolar transistors, CMOS, strained silicon, FinFET, three-dimensional stacking. Every prediction that "Moore's law is about to end" was right about the curve then in play, and wrong about the envelope.
In 1986 the management scholar Richard Foster turned this into an investment criterion, and incidentally explained why incumbents keep missing transitions. His observation: near the top of the old curve, the performance you buy with each additional dollar keeps falling, while the new curve is still in its own slow beginning and is necessarily lower in absolute terms. So at the moment when switching is genuinely correct, the new option on the table necessarily looks like the worse choice. What fails is not judgement, it is the criterion. Compare absolute levels and you will always be late; how late depends on how long the old curve's tail is.
To decide between doubling down and switching, do not compare levels. Compare two things: ① the old option's marginal return — has the performance bought per additional unit of input been falling for several periods running; and ② whether the new option's slope has passed the old option's slope (the slope, not the height). The slopes cross years before the heights do, and those years are the entire window for switching. Concretely: replace the line in your reporting that reads "how far ahead of the competition are we" with "how much did we improve this year versus how much did they improve this year".
The last three sections assumed you know which part of the curve you are on. This one says: during the phase when you most need to know, you cannot.
The reason is hidden in the formula. While N is much smaller than K, the discount term (1 − N/K) is almost exactly 1, and logistic growth is exponential growth — not "close to", but indistinguishable within the noise of the data. And K can only be estimated from the stretch where the curve begins to bend. So when you want to know K there is no data, and by the time the data exists the answer no longer matters.
In 1920 Raymond Pearl and Lowell Reed gave the textbook demonstration. They fitted an S-curve to US census figures from 1790 to 1910 and obtained a carrying capacity of 197 million. This was not a careless exercise — the curve hugged the actual data all the way to about 1950 and was the best population forecast of its era. Then the US population crossed that "absolute upper limit" in 1967, and today stands at 340 million, roughly 70% above their K.
Their error was not in the arithmetic, nor in the data. It was that 197 million was never the location of a wall; it was a projection of where things would settle if the agriculture, urbanisation and immigration policy of the time never changed. The constraint changed, and the projection expired.
This has a direct consequence. Every argument of the form "this time is different / this is a new paradigm / the ceiling is right there" is, before the inflection point, undecidable in principle from the data. It is not that the arguers are insufficiently clever; the data in that stretch genuinely does not contain the information needed to separate the hypotheses. Continuing to argue is just an exchange of priors.
The way out is not a better estimate of K but a different observable. Carrying capacity is an abstract limit and invisible; the specific thing blocking you right now is not. Change the question from "how much longer can we grow" to "which input is binding us right now, and is the output bought by one more unit of it rising or falling this year" — the second can be measured every quarter.
Kill the "how much longer can we grow" debate — it is undecidable before the inflection point, and holding it only burns trust. Replace it with a two-column table: on the left, the 3–5 inputs current growth depends on (headcount, channels, compute, capital, attention); on the right, the change since last period in output per unit of that input. K is unmeasurable; marginal return is measurable. Whichever line falls two periods running is your actual K — and it is usually not the one you assumed.
The S-curve is the most widely circulated shape in complexity science, and therefore the most abused. Four things to know before using it.
First, Easter Island is not evidence. You have probably heard the story: the islanders cut down every tree to move their statues, lost their canoes and their soil, and the population collapsed — civilisational suicide. It has been written into countless popular accounts of limits to growth as the specimen case of overshoot and collapse. It has also been substantially revised. The archaeologists Terry Hunt and Carl Lipo argue that the Polynesian rat, which arrived with the settlers and ate palm seeds, played a large part in the deforestation; and a 2024 Nature study of 15 ancient Rapanui genomes found no genetic signal of a population crash before Europeans arrived. The real collapse followed the Peruvian slave raids of 1862–63 and the smallpox that came after. The lesson is not that ecological collapse never happens, but this: a story does not become evidence by being retold ten thousand times. The St. Matthew reindeer were counted. Easter Island was not.
Second, slowing down does not imply a ceiling. The global population growth rate peaked around 1963 at roughly 2% a year and has fallen ever since, to under 1% today; the UN's 2024 projection has the population peaking near 10.3 billion in the 2080s and then declining. The shape fits an S perfectly. But the mechanism is not Malthusian: food output grew faster than population across those same decades. The slowdown comes from falling fertility — the demographic transition, driven by income, education, infant survival and urbanisation changing how many children people want. That mechanism has nothing to do with carrying capacity. Hence: the S-shape is the common appearance of many different mechanisms, not the signature of any one of them. Seeing a curve bend and concluding "we are near the carrying capacity" is like seeing a fever and concluding influenza.
Third, the ledger of The Limits to Growth has two columns. In 1972 Meadows and colleagues ran twelve scenarios through the World3 model → ref · World3 and The Limits to Growth, and the book still serves both as scripture and as target practice. The honest account: the famous table of "years of this mineral remaining" was a demonstration of how fast exponential arithmetic eats a reserve, not a forecast, yet it has been attacked as a forecast for fifty years. On the other side, Turner (2008, 2014) and Herrington (2021) compared observed data against the scenarios and found that the aggregate variables — population, industrial output, pollution — do broadly track the "standard run". But that is not a victory either: the variables are aggregated extremely coarsely over an extremely long horizon, so "broadly tracks" is weak evidence, and many of the model's relationships cannot be independently calibrated. Its real contribution is structural (delay + overshoot + a K that can be damaged), not numerical.
Fourth, and most dangerous: the S-curve story is inherently unfalsifiable. Any slowdown can be relabelled after the fact as "approaching K", and any reacceleration as "a new curve starting". A framework that explains every outcome explains nothing. To make it a tool rather than a rhetoric, two things have to be done in advance: name the constraint (what exactly is holding you back), and state how it is measured (which number tells you whether it has loosened). Miss either one and all you are holding is a shape.
Before drawing any conclusion from an S-curve, write down a falsifiable sentence: "If the marginal return on ××× (the constraint I have named) does not fall within the next N periods, then my judgement that we are near the limit is wrong." If you can write it, the judgement is the kind of thing you can bet on. If you cannot, you have only fitted a shape to something that already happened. That sentence also settles which number to watch — it is the single new line on your dashboard.
Yes, but it means something else. It no longer denotes a fixed number but the ceiling under the current constraint. Accepting that raises a harder question: is there a rate limit on the relays themselves? The envelope looks exponential only because each handoff happened in time; the antibiotic envelope stopped when the handoff rate went to zero. So the question of limits shifts from "how much resource is left" to "how fast can we manufacture new curves".
An operational dividing line: if the falling marginal return comes from scale itself (bigger means harder to coordinate, more crowded), more input will not help, because scale is the constraint. If it comes from a specific replaceable input (a material, a channel, a kind of compute), replacing it is the start of a new curve. So the first job is not deciding whether to wait, but isolating which line is actually falling.
It depends on whether the consumed thing is renewable and whether an irreversible threshold was crossed. Lichen takes decades, topsoil centuries, extinction is permanent. This connects to the hysteresis loop of Topic 10: some systems will not retrace their path even if the pressure is removed entirely. The question to ask is "if consumption stopped now, how long until recovery" — not "can it recover".
Worth thinking about: voluntary reduction in movement during an epidemic (which happens before policy does), substitution triggered by rising prices in a market, hierarchy emerging spontaneously as an organisation grows. What they share is that the negative feedback comes from agents' expectations rather than from resource depletion — which also means they can fail abruptly when expectations change, a fragility that a pure resource constraint does not have.
Beyond the criterion problem (comparing absolute levels makes you late), there is a structural layer: marginal return on the old curve's tail is falling, but its total is still the largest in the business, and the organisation's costs, staffing and performance measures all hang off that total. So switching is not a cognitive problem but a question of whose budget gets cut — which is also why outsiders find the same judgement so much easier to make.