How early accidents become late inevitabilities
2026-07-30 · Dynamics & Unpredictability
The thing you were trying to filter out is, in one class of systems, the only thing that lets the signal through — and in another class, the thing that decides who wins. These sound unrelated. They are two faces of one mechanism.
Two nearly identical coffee shops open on the same street. Same beans, same fit-out, same prices. Six months later one has a queue out the door and the other has closed. Afterwards you can always find a reason — one fewer step at the entrance, a friendlier owner. But if you could replay those six months a hundred times, it is entirely plausible that the other one survives in forty of them.
Everything so far has been about unpredictability in deterministic systems: chaos (Topic 8) has no random term at all — the equations are exact and the trouble is that you cannot measure the initial state precisely enough; computational irreducibility (Topic 11) does not even care about measurement, the problem is that no shortcut exists. This issue is about something else: a system that really does contain a random term. That term has two possible fates, and they could not be further apart — it either averages away to nothing, or it gets amplified into the outcome of the whole system.
And there is a stranger half to this: for a whole class of systems, turning the noise down to zero means detecting nothing at all.
First, the words. When a quantity wobbles around the value it "should" have, that wobble is a fluctuation; the collective name for those wobbles is noise. Noise is not error. It is the combined effect of everything you are not tracking — who happened to walk past today, what else the server was running, which way the wind blew this second.
There is a rule about noise that everyone already half-knows: do it enough times and it disappears. Flip a coin once and the result is anyone's guess; flip it ten thousand times and the fraction of heads sits practically on 0.5. One household's electricity use jumps around wildly; a whole city's load curve is smooth enough to draw by hand. This is the law of large numbers, and the relative fluctuation shrinks roughly as 1/√N — a hundredfold increase in samples leaves a tenth of the relative error.
It rests on one precondition, and that precondition is the one most often violated: each draw must be unaffected by the ones before it.
The moment an outcome changes the odds of the next outcome — a queue attracts more people to queue, a cited paper gets cited more easily — you no longer have independent sampling. You have positive feedback → ref · System Dynamics. Same coin, same randomness, completely different fate.
The left panel squeezes together; the right one fans out. And the right side has a harsher property still: it is not "temporarily off, will return". It cannot return. Once the urn holds 300 red against 20 blue, one more draw moves the ratio by a fraction of a percent. The system has pinned itself down. An early fluctuation has been written into the state, and from then on it is the initial condition for everything that follows.
So "does randomness matter in this system" is the wrong question. The right one is: does a random outcome here change the next probability? If not, it is noise. If so, it is history.
When a metric looks like it is converging, do not report the mean yet. Run a replication — N independent runs under the same conditions, or replay the history in segments — and watch whether the variance shrinks over time. If it shrinks, a mean with error bars is honest. If it does not, the mean means nothing: report the distribution of outcomes and its range, and say out loud that a rerun would come out differently. That single step blocks most of the post-mortems that turn luck into competence.
Now the reverse.
Many things have a threshold: below some level there is no response whatsoever, above it there is output. A smoke alarm needs enough particulate to sound. A neuron needs enough depolarisation to fire. An alerting rule needs the metric to cross a line before anyone is paged. These are threshold systems.
Now feed one a very weak signal — too weak ever to reach the threshold. What comes out? Nothing. The signal is there the whole time and is never heard.
Then do something that sounds foolish: add noise. Noise is random, but the chance that it pushes the signal over the line is larger when the signal itself is already high. Pulses start to appear at the output — and the timing of those pulses carries the signal. Add more noise and the output fires everywhere, burying the information again. Somewhere in between there is an optimum.
This is stochastic resonance. The name misleads slightly: "resonance" here has nothing to do with matching frequencies. What is matched is the noise intensity against how far the signal falls short of the threshold.
It is not a chalkboard curiosity. Crayfish tail mechanoreceptors detect weak water motion better with a moderate amount of hydrodynamic noise (Douglass et al., 1993). Juvenile paddlefish hunt plankton electrically, and they catch individual prey more successfully when a swarm of Daphnia is generating electrical noise around them (Russell et al., 1999). Human balance works the same way: add subsensory random vibration to the insoles and postural sway improves across the board in 27 subjects, with elderly participants gaining more than young ones (Priplata et al., 2003).
A cleaner way to picture it: think of two wells with a ball sitting in one of them → ref · Phase Space & Attractors. A weak periodic push cannot get the ball over the ridge on its own. Add noise and the ball gets kicked at random — but it gets over more easily during the half-cycle when the push is helping. The rhythm of the hopping ends up tracking a signal that could never have moved it alone.
Stop treating noise reduction as unconditionally good. Two tests: is there a hard threshold anywhere in your detection chain (alert level, review pass mark, KPI target, eligibility cut-off)? Is the signal you care about chronically below it? If both, measure whether detection rate versus noise intensity is an inverted U before you flatten the jitter. A concrete version: alongside a fixed-threshold alert, add a layer of random sampling, or let the threshold itself fluctuate slightly — so that weak anomalies which could never cross the line get an occasional chance to surface. Freezing the threshold and then removing all variation is a guarantee that you will only ever see problems that are already severe.
Back to the urn on the right of the first figure. It has a name — the Pólya urn, written down in 1923, and the rules take three lines → ref · Pólya Urn: one red ball and one blue ball; draw one at random, look at its colour, put it back, and add another ball of that colour. Repeat.
What happens is one of the loveliest counter-intuitive results in the field: the final share of red follows a uniform distribution — every ratio between 0 and 1 is equally likely. Not "usually near half, occasionally extreme". Five percent, fifty percent and ninety-five percent are equally common.
In 1989 the economist W. Brian Arthur rewrote this urn as a model of competing technologies: two options, adopters arriving one at a time, and each adoption raising the option's appeal to the next arrival. That "the more people use it the better it works" property is called increasing returns, and it usually comes from three sources — more people know how to use it (learning effects), more things plug into it (network externalities), and everyone else uses it so following is cheapest (coordination).
The model yields three conclusions, each blunter than the last: the market will lock in to one of them; which one is not predictable; and the option that was better beforehand is not guaranteed to win.
The useful part is the time structure. Early on, a handful of choices can flip the outcome. Late on, ten thousand cannot. The same resource spent inside the window and outside it differs by orders of magnitude. That is not a question of effort; it is a question of how many balls are already in the urn.
Two over-told examples: the QWERTY layout, supposedly designed to scatter common letters so mechanical typewriters would jam less, then locked in because everybody learned it; and VHS beating the technically better Betamax. The stories run smoothly — suspiciously so, and the next section takes them apart. But note in advance: what gets taken apart is those two examples, not the mechanism.
Where returns are increasing, "when" matters far more than "which". Diagnose the stage first: users and suppliers still switching frequently → still inside the window, front-load resources, one unit now beats ten later; switching rates already in the low single digits and the ecosystem self-sustaining → locked in, so kill the "we'll build something better and flip it" plan and pivot to interoperability, compatibility layers or living on top of it. The mirror image matters just as much: to keep your own options open you must deliberately sustain variety early — a second supplier, parallel designs, a standard you decline to freeze. Variety here is not a virtue signal; it is holding the urn back from rolling out of the range you can still return from.
Path dependence is the most easily abused idea in this field. Four things have to be said before using it.
First, the QWERTY case itself is disputed. Paul David's 1985 paper made "the market locked in to an inferior standard" famous, and its key evidence was a US Navy typing experiment from the Second World War — typists on the Dvorak layout were faster, and retraining costs were recovered within ten days. Liebowitz and Margolis's 1990 paper The Fable of the Keys turned that evidence over: the Navy study was run by Dvorak himself, who had a commercial stake in the keyboard, and later controlled studies failed to reproduce the margin reliably. The careful statement is therefore: the lock-in mechanism is real, but the most famous case of "locked in to something clearly worse" rests on soft evidence.
Three claims usually smuggled into one word need separating here — Scott Page made the distinction explicitly in 2006: history mattered, the outcome is inefficient, and the outcome is irreversible. The first is easy to establish. The other two each require their own evidence.
Second, "early accident" slides very easily into unfalsifiable hindsight. You have exactly one history. Seeing the winner lead early, you cannot tell luck from some real advantage you failed to observe. That is precisely why the music experiment below matters — it actually built parallel worlds. Without them, the honest sentence is "I cannot tell", not "this shows path dependence".
Third, stochastic resonance has hard preconditions, and all of them are required: the system must be nonlinear (a threshold is the classic case), the signal must genuinely be subthreshold, and the noise must be tunable into the right window. Miss one and added noise simply makes things worse — in a linear system noise only ever lowers the signal-to-noise ratio. Worth noting where the idea came from: stochastic resonance was proposed (Benzi et al., 1981) to explain the 100,000-year glacial cycle, with weak orbital forcing amplified by climate noise — and that particular application is largely not accepted today, with several competing explanations in play. "Noise helps under specific conditions" is a technical result, not a worldview about the benefits of chaos.
Fourth, drift explanations get over-used too. Judging a corpus time series to be "pure drift" is quite sensitive to sampling and to topical fluctuations — which is the thrust of Karjus and colleagues' critique of exactly the work cited in section 1. A neutral null hypothesis is a good default, but it is still a model and does not get a free pass.
Before saying something is "locked in to the worse option", answer the three questions above one at a time: what the better alternative is, what switching costs and what it would gain, and whether the cost is technical or a failure to coordinate. If all three are answerable you have a claim you can act on — and question ③ points straight at the remedy: coordination failures are solved by coordination devices (a common switch-over date, a compatibility layer, subsidising early movers), not by building a better product. If they are not answerable, say only "history mattered". It is a cheap sentence, but an honest one.
The only useful action inside the urn model is changing the draw probabilities, and that is very cheap when there are few balls and very expensive when there are many. So "does effort matter" is the wrong question; the question is how much subsequent probability this particular unit of effort moves. The same reasoning explains why "just build something better" so often fails in a mature market: it changes the product, not the probabilities.
No, and conflating them is this concept's most common abuse. Stochastic resonance comes with units: how high the threshold is, how weak the signal is, how intense the noise is — and their relationship can be drawn as a curve and refuted by experiment. "A little disorder is beneficial" has none of that and cannot say when it would fail, which makes it a metaphor rather than a claim. Simple test: can you state what the horizontal axis of that inverted U is?
The variety in this issue means the state space has not yet been compressed — several approaches still coexist, so the system retains the ability to reach different outcomes. It is not automatically the same as diversity of people, though the two correlate: if everyone was trained the same way, the space of proposals tends to collapse on its own. To tell which you have, look at behaviour rather than composition: in the last six months, was there a decision where two genuinely different proposals were compared seriously?
Because a post-mortem can have two goals. One is to explain the outcome, which in a high-fluctuation system is mostly impossible. The other is to evaluate the decision process — the information available, the candidate set, the size of the bet — which is independent of the outcome and needs no parallel worlds. Converting post-mortems from the first kind to the second is the one change in this issue that is completely free at the individual level.