Equilibrium selection
A Socratic walk-through of equilibrium selection — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #When a game offers several equally good solutions, what makes everyone pick the same one?
Game theory's founding result is reassuring: Nash proved around 1950 that every finite game has at least one equilibrium — some combination of strategies from which no player wants to deviate alone. Solve for it, and you know what rational players do.
But read that guarantee again. It promises at least one. It says nothing about there being only one. And the moment a game has two, the theory has done all it can and still not told anyone what to do. So the question is what does the rest of the work, and whether that remainder is rationality at all.
Reasoning it through
REASONING #Take the plainest case. Two drivers approach each other on an empty road. Both keep left, and all is well. Both keep right, and all is equally well. One of each, and they collide.
Two equilibria, identical payoffs, perfectly symmetric. Ask what a purely rational agent, knowing the payoff matrix and nothing else, should do. There is no answer — and not because the analysis is incomplete. The problem is symmetric, so any reason to choose left is mirrored by an equally good reason to choose right. Reasoning from the payoffs alone can never break a tie the payoffs themselves do not break.
Notice what this means: the thing that decides must come from outside the game as written.
So what actually decides? Ask the question the way each driver must: not "which is better?" but "which will the other one expect me to expect?" That reframing is the whole of it. What matters is not what is good but what is conspicuous — and conspicuous to both, and known by both to be conspicuous to both.
Schelling called such a cue a focal point, in The Strategy of Conflict in 1960. He asked people where they would go to meet a stranger in New York with no time or place agreed; a striking majority named Grand Central Station, at noon. Nothing in the payoffs prefers Grand Central to any other landmark. What made it work is that everyone could see that everyone could see it. (His surveys were informal, run on his own students rather than as controlled experiments, and later replications find the effect real but sensitive to which population is asked.)
Now press on the "known to be known" part, because it is where the mechanism actually lives. A cue salient to you alone is worthless. If you find noon obvious but suspect your partner is a night owl, you must reason a level further: what will they think I think? Coordination on a focal point rests on a shared culture supplying an expectation that recurses — not on the cue being good, only on it being commonly obvious.
Where does such shared salience come from? Several sources, and they are worth separating. Precedent is the strongest: whatever was done last time, or has been done for a century, needs no reasoning at all. Communication, where it is available, manufactures a focal point outright — and this is why cheap talk, which changes no payoff, can transform a coordination game while doing nothing to a competitive one. Authority does the same by decree; Sweden's Dagen H in September 1967 moved a whole country from left to right overnight by simply announcing which equilibrium was now the salient one.
Then a harder case, where the equilibria are not equally good. In the stag hunt, two hunters cooperating bring down a stag and eat well; either can instead take a hare alone, small but certain. Both hunting stag is an equilibrium, and it pays more. Both taking hare is also an equilibrium, and it pays less. Rationality now seems to have an answer — take the stag.
But ask what happens if you are unsure of your partner. Hunting stag alone yields nothing; taking the hare yields something regardless of what they do. The safer equilibrium is not the better one. Harsanyi and Selten formalised this tension in 1988 as payoff dominance against risk dominance, and could not reduce one to the other; experiments frequently find people, especially in larger groups, settling on the risk-dominant option. Which criterion should govern is genuinely unsettled.
The analogy
THE ANALOGY #Two strangers meet head-on in a corridor too narrow for both. Each must pick a side; the sides are equally good; picking differently produces the shuffling collision everyone has performed. Nothing about the corridor prefers left. What resolves it is a shared cue — the side the country drives on, a corridor's worn path, or one walker committing early and visibly enough that the other can read the commitment and go along.
in a corridor you can watch the other person and correct mid-stride, whereas the cases that matter are simultaneous and one-shot — and the corridor makes it look like a trivial nuisance, when the same structure settles currency standards, file formats and rail gauges, where the chosen equilibrium then locks in for a century.
Clarifying the model
THE MODEL #The most common misreading is that a focal point is a reason. It is not; it is a coincidence of attention that both parties can rely on. Whether the cue is arbitrary is irrelevant — what matters is that it is uniquely conspicuous.
The second is to suppose selection is a one-off act of choosing. Usually it is a history. Peyton Young's 1993 evolutionary account has conventions emerging from repeated play with occasional mistakes, and shows that over long enough horizons this process tends to settle on the risk-dominant equilibrium rather than the payoff-dominant one — selection by drift and accumulated precedent rather than by anyone deciding.
And it is worth being clear about the boundary. Equilibrium selection is not a defect in game theory that better mathematics will remove. Multiplicity is a real feature of the world: there genuinely is no fact of the matter about which side to drive on. What the theory can supply is the observation that some outside thing must do the choosing, and a vocabulary for what that thing tends to be.
A picture of it
THE PICTURE #How to readthe two settled conventions are both equilibria and neither is preferred by the payoffs, so every arrow into one of them is labelled with something outside the game — a salient cue, a precedent, an announcement; the middle state is the costly limbo that arises when no such cue is commonly held.
What became clearer
WHAT CLEARED #When a situation admits several equally rational outcomes, rationality cannot pick among them, and the choosing is done by whatever both parties can expect the other to expect — salience, precedent, announcement, habit. Coordination is therefore less a calculation than a shared reading of the same cue, which is why so much of it looks arbitrary in retrospect and is nonetheless very hard to change.
Where to go next
ONWARD #- Cheap talk: why costless, unenforceable messages reshape coordination games and barely touch competitive ones.
- Lock-in and path dependence, where an early-selected equilibrium survives long after a better one becomes available.
- Level-k and cognitive-hierarchy models, which try to measure how many steps of "what will they think I think" people actually take.
Key terms
TERMS #| Term | What it means |
|---|---|
| Nash equilibrium | a strategy profile from which no player gains by deviating alone. |
| Focal point | a cue that is conspicuous to all parties and commonly known to be so, used to coordinate. |
| Payoff dominance | the criterion preferring the equilibrium that pays everyone more. |
| Risk dominance | the criterion preferring the equilibrium that is safer against uncertainty about the other player. |
Every term the collection defines is gathered in the glossary.