Running to stay in place
2026-08-17 · Adaptation & Evolution
The word "progress" quietly assumes a frame of reference that holds still. In a large class of systems the frame is assembled out of your opponents — take a step forward and it steps with you. The improvement is real; the payoff can be exactly zero.
In any neighbourhood, the first family to enrol their child in weekend tutoring genuinely gains something. Three years later everyone has enrolled, the rankings are back where they started, and every family is spending six extra hours a week plus the fees. Nobody cheated, nobody blundered, every single step was rational — and the collective result is a higher cost floor with the ordering untouched.
That structure is more common than it looks: antibiotics and bacteria, spam filters and spammers, advertising and attention, defence and offence, even you and your own tolerance to a drug. They all share one mechanism — how well you are doing depends not on how good you are, but on how much better you are than your opponent right now; and your opponent is changing too. That is coevolution.
The adaptation in the previous few topics faced a stationary opponent: Topic 28 climbed a fixed fitness landscape, Topic 29's payoff table was handed to the players. This topic removes that last assumption — the landscape is worn into shape by everybody else's adaptation, and it shifts the moment you move. With that one change, equilibrium goes from being the normal state to being the rare one.
Start with the word "fitness". It does not mean strong, good, or advanced. It means one thing: in this round, how many more offspring (or units of market share, or surviving machines) you leave than the others do. It is a ratio by construction, never an absolute quantity.
Topic 28 drew it as a terrain map: your strategy or trait on the horizontal axis, the fitness of that position on the vertical, and you walk uphill → ref · NK fitness landscape. That picture carries an inconspicuous assumption — the terrain was drawn once and holds still. As long as it holds still, reaching a summit means permanently owning the best position available.
Coevolution deletes that assumption. Your fitness depends on your opponent's current state, and your opponent is climbing its own slope; every step it takes redraws your terrain. The result is a distinctly uncomfortable situation: your traits have not regressed at all, your absolute capability is still rising, and yet the ground beneath you has dropped — because what changed was the hill, not you.
That picture is exactly what Lewis Carroll's much-quoted line means. In Through the Looking-Glass (1871) the Red Queen drags Alice along at a sprint, and when they stop Alice finds they have not moved: "Now, here, you see, it takes all the running you can do, to keep in the same place." In 1973 the palaeontologist Leigh Van Valen borrowed the line to name a statistical pattern — across large numbers of fossil taxa, how long a group had already existed barely changed its probability of going extinct in the next interval. Experience had not accumulated into safety. That is the Red Queen hypothesis: every species keeps getting better, and precisely because everyone is getting better, nobody ends up safer.
In 1926 the Italian mathematician Vito Volterra was handed a puzzle from the fisheries. During the First World War fishing in the Adriatic all but stopped, so every kind of fish should have become more plentiful; yet the statistics compiled by his son-in-law, the biologist Umberto D'Ancona, showed that the share of sharks and other predatory fish in the catch had gone up. Why should fishing less benefit the eaters rather than the eaten?
He wrote down two lines (Alfred Lotka in the United States had reached the same equations a year earlier, coming from chemical reactions). They are now called the Lotka-Volterra equations → ref · Lotka-Volterra model:
Four parameters, two variables, no randomness, no weather, no seasons. Run it and it starts going in circles all by itself: hares increase → lynx increase after them → hares get eaten down → lynx starve → lynx decline → hares recover.
Two results are worth keeping. The first is the lag: the predator peak always follows the prey peak by roughly a quarter of a cycle, which means the current predator count is a report on past prey. It is chasing a target that has already left. The second is stranger. Work out the long-run averages and the mean number of prey depends only on the predator's parameters (it equals γ/δ) — it has nothing to do with how fast the hares breed. If you want more hares on average, change something about the lynx; changing the hares does nothing.
That is how the fisheries puzzle resolves. Apply an indiscriminate kill to both sides (nets that take small fish and sharks alike) and the mean prey number falls while the mean predator number rises; stop that indiscriminate kill and the prey side is what gains — hence the wartime rise in the shark share. The result is now called Volterra's principle: an intervention applied without discrimination has opposite effects on parties at different levels of the chain.
When you see a rival's scale peak, do not treat it as news. In a lag-coupled system it is most likely the echo of your own expansion one round ago. A concrete test: correlate the rival's scale metric against your own metric at the same time and against your metric a quarter-cycle earlier. If the lagged correlation is clearly higher, you are looking at an echo — and reacting to it means firing at something that no longer exists.
Biology has an embarrassingly blunt problem: sexual reproduction is a terrible deal. An asexual population that produces only daughters reproduces about twice as fast, because males do not bear young — half the productive capacity is spent maintaining individuals who do not produce. On arithmetic alone, asexual lineages should have swept the board long ago. They did not.
The Red Queen's answer is that diversity in your offspring is not a luxury, it is fortification. Parasites (and bacteria, and viruses) reproduce far faster than their hosts, and they track whichever host genotype is commonest, cutting a key to fit it. Copy yourself exactly and you are the lock that key already opens. Sex reshuffles the genome every generation, so the lock you hand over is a new one.
The mechanism has a name: negative frequency-dependent selection — a type's advantage falls as it becomes common and returns as it becomes rare. Note how this differs from a vague "diversity is good": here the causal claim is sharp and measurable — a type's fitness is a decreasing function of its own frequency one step earlier.
This is not just paper reasoning; two bodies of evidence carry it. One comes from a New Zealand freshwater snail (Potamopyrgus antipodarum) in which sexual and asexual individuals coexist. Curtis Lively found in 1987 that the heavier the trematode parasite load in a body of water, the higher the proportion of sexual individuals; where parasite pressure is low, asexual clones nearly monopolise — exactly what the twofold advantage predicts. The other is a causal test in the lab: in 2011 Morran and colleagues coevolved nematodes with a bacterial pathogen and found that when the pathogen was allowed to evolve alongside, the worms shifted to outcrossing, while populations forced to self-fertilise were driven extinct. Freeze the pathogen so it cannot evolve, and the effect disappears.
Anywhere an opponent is watching you, standardising everything onto the single currently-best option is slow suicide — it hands the whole system over as one lock. Concretely: in critical positions (suppliers, crop varieties, mirrored systems, authentication methods) keep at least one option that is worse but different, and book its cost as an insurance premium rather than as waste. The warning sign is easy to spot: the moment you can say "we use X for everything," you are the commonest genotype. Bananas are the standing lesson — a fungus erased the previous dominant cultivar, and its replacement, the Cavendish, is again a single clone, so the same play is running again.
The snails and their parasites never got more "powerful"; they only took turns being the common type. But other coevolution looks nothing like that — both sides raise the stakes step by step and never go back. These are two regimes, and applying the wrong one leads to opposite decisions.
The escalating kind is an arms race. The rough-skinned newt of western North America carries tetrodotoxin (TTX, the pufferfish poison) in quantities that could kill a roomful of people, and the local garter snakes eat them anyway, thanks to a handful of amino-acid changes in a sodium channel. Where the newts are more toxic, the snakes are more resistant, each side ratcheting the other upward. The signature of this regime: the trait has a direction, the cost rises monotonically, and retreat is not available.
The other kind is fluctuating selection — the previous section's picture, where frequencies oscillate with no net direction, and an old model can win again a few generations later. Coevolution experiments with bacteria and their phages have produced both regimes, depending on conditions.
One more thing that often gets skipped: coevolution has a geography. John Thompson's geographic mosaic theory holds that the same pair of species is coupled with wildly different intensity in different places — some sites are hot spots where the two are locked together, others are cold spots where they barely interact. Broad sampling of the newts and snakes bears this out: in many populations toxicity and resistance simply do not match, and snakes occasionally leap, via a single decisive mutation, far above the local toxin level and escape the race entirely (while the prey side is never observed to get ahead). So "we are in fierce competition" is a claim that first has to say where.
Classify every competition you are funding, using one question: "If we stopped this for two years, how long would it take to get back to where we are today?" If a few weeks, it is fluctuating — you can start and stop with the rhythm. If more than a year, it is an arms race: it will only get more expensive, and winning it does not buy safety. For that kind, what ends it is never running faster. It is moving to a dimension the coupling does not cover, or a mutual limit both sides can verify.
"Red Queen" is one of the most useful and most abused phrases in complexity science — it slides very easily into being a decorative way of saying "competition is fierce, so keep innovating," which is empty. Here are its four real boundaries.
First, Van Valen's law of constant extinction is not uniformly supported by later data. Palaeontology now runs two explanations side by side: the Red Queen (organisms pressing on each other) and the Court Jester (climate, tectonics, and other indifferent external events). Michael Benton's 2009 review concluded that they operate at different scales in space and time: biotic interaction dominates locally and over short spans, while climatic and geological events dominate regionally and over millions of years. Explaining a mass extinction with the Red Queen is usually reaching for the wrong level.
Second, Lotka-Volterra's "eternal cycle" is an artefact. Those closed orbits are structurally unstable: add one realistic touch — a ceiling on the hares' food supply — and the neutral loops immediately collapse into a spiral converging on the fixed point, or into a limit cycle of fixed amplitude. Do not read the loops as a prediction. The textbook hare-lynx cycle is not evidence for them either: Charles Krebs's team ran an eight-year field manipulation in the Yukon and found that excluding predators doubled hare density, adding food tripled it, and doing both raised it elevenfold — far more than additive, which says the cycle involves at least three trophic levels and does not fit inside two equations.
Third, not all coevolution is a fight. Mutualism is coevolution too — figs and fig wasps, legumes and rhizobia — and its direction is not escalation but deepening mutual dependence. Treating coevolution as a synonym for arms race is where the concept is most often misapplied. The test is the same one: look at whether each side's cost curve is climbing or falling.
Fourth, and most important: it has to be falsifiable. What the Red Queen actually claims is narrow — absolute improvement fails to convert into relative position, an advantage decays as it spreads, costs rise monotonically. All three are measurable. "Competition makes everyone better" excludes nothing; any outcome can be fitted into it. A concept that answers every counterexample with one more auxiliary clause ("that region was Court Jester", "that site is a cold spot") while issuing no new predictions has stopped being a tool and become rhetoric.
Before saying "this is a Red Queen situation", clear three bars: ① is your fitness genuinely a function of your opponent's current state rather than of a fixed external environment (if you are not sure, go look — did your results change during the year the opponent sat still?); ② is there a measurable quantity in motion — a frequency, a cost, a relative gap, any one of them; ③ what new prediction does the claim make, and what result would count as refuting it. If you cannot answer the third, delete the phrase and describe your mechanism in plain words.
Be careful not to read "the outcome is unchanged" as "nobody is running." The Red Queen describes exactly this case: stable positions with continually rising inputs. So the test is not in the outcome but in the cost. Plot relative share and absolute investment on one chart over several years: share flat, investment climbing, is a textbook Red Queen, not a steady state. A real steady state has both lines flat.
Often it is. The meaningful difference lies in what the new dimension's coupling strength and cost slope look like: how fast the opponent can follow, and what following costs them. If both match the old dimension, you have only changed the venue. A genuinely valuable move finds a dimension where the opponent's adaptation is hard-constrained — the way the measles virus cannot move its critical sites.
It does happen, and that is one of the Red Queen's falsifiable predictions — the New Zealand snail data broadly fits. But other explanations for sex are competing for the same job: purging the accumulation of harmful mutations (Muller's ratchet), and recombining beneficial mutations into one lineage. These are not mutually exclusive and probably stack. It is also a good illustration of what makes an explanation worth having: one that predicts a geographic difference is worth far more than "diversity is good."
Not quite. The Red Queen is descriptive: absolute improvement does not convert into relative position. "Rat race" usually carries two extra claims — that participants cannot exit, and that total output has not increased. The second is often false: arms races frequently do produce real technology (radar, antibiotics, the dreadnought's propulsion). Keeping them apart clarifies a great deal: mechanism is one thing, value judgement is another.