A Cold War Math Problem's Lesson for Auctions, Poker, and Evolution: The Way to Win Is to Stop Trying To

Most branches of mathematics don't come with a body count of ruined dinner parties. Game theory does, quietly, every time someone realizes that the "rational" move in a shared decision isn't the nice one. This is the story of how a handful of mathematicians formalized that discomfort into one of the most widely applied ideas in modern science, one that now prices your spectrum licenses, folds your poker hands, and explains why two hawks fighting over a nest looks a lot like two superpowers fighting over a border.

Game theory asks a narrow question with enormous reach: what's the best thing to do when the best thing to do depends on what someone else decides to do? Here's how it got built, what the math actually says, the different flavors of "game" it had to invent along the way, and where it's headed now that machines have started solving it faster than the humans who invented it.

1838 Cournot solves a duopoly problem using what we'd now call a Nash equilibrium, a century before Nash exists.

1913 Zermelo proves chess has a determined outcome, the field's first formal theorem.

1928 Von Neumann proves the minimax theorem, giving zero-sum games their first rigorous foundation.

1944 Von Neumann and Morgenstern publish Theory of Games and Economic Behavior, the field's founding text.

1950 John Nash publishes two separate results in the same year, a general equilibrium concept for competitive games, and a completely different solution for cooperative bargaining.

1953 Lloyd Shapley works out how coalitions should fairly split their joint winnings.

1961-1974 Vickrey, Selten, Harsanyi, and Aumann each patch a different gap: auctions, credible threats, hidden information, and coordination without contracts.

1973 Biology borrows the math. Maynard Smith and Price show evolution performs the same optimization, without anyone doing the reasoning.

1980 Axelrod's tournament proves "nice" strategies can out-earn ruthless ones, under the right conditions.

2007-2020 The Nobel committee spends over a decade recognizing game theory's move from describing markets to designing them, mechanism design, matching markets, spectrum auctions.

Now Machines compute equilibria no human ever could by hand, and the field's newest questions are about where the theory of "rational" play stops matching how anyone, human or AI, actually behaves.

The Different Flavors of a Game

"Game theory" isn't one theory, it's a family of frameworks, and which one applies depends on how you answer a set of structural questions about the game itself. This table doubles as a map for the rest of the article, each row points to the section that solves it.

Almost the entire history of the field is the story of relaxing one of these eight restrictions at a time. Von Neumann solved the narrowest cell in this table first, two-player, zero-sum, non-cooperative, perfect information, simultaneous, one-shot, because it's the cleanest possible case. Everything that follows is really an argument about what breaks, and what has to be rebuilt, once you let go of one restriction and move to the next row.

Before the Formal Theory: A Duopolist, a Chess Player, and a Skeptical Mathematician

The core idea showed up piecemeal, decades before anyone called it game theory:

• 1838: Antoine Augustin Cournot, working on spring water duopolies, solved for a stable output level where neither firm wants to change what it produces given what the other produces
• 1881: Francis Ysidro Edgeworth extended the thinking to bargaining between two parties
• 1913: Ernst Zermelo proved that in chess, exactly one of three things must be true, White can force a win, Black can force a win, or either player can force a draw
• Early 1920s: Émile Borel described mixed strategies for card games and conjectured a minimax theorem, guessing incorrectly that it might fail for games with enough possible strategies

It took a young Hungarian mathematician to prove Borel wrong and build the field's actual foundation.

Von Neumann, Morgenstern, and a Theorem Built for the Wrong War

John von Neumann proved the minimax theorem in 1928: in any two-player zero-sum game, a stable strategy always exists for each player, even if that strategy has to involve randomization.

See it with numbers. Two players secretly choose heads or tails. If the choices match, A wins ₹1 from B. If they don't match, B wins ₹1 from A.

No pure strategy survives, whatever A commits to, B exploits, and vice versa. Von Neumann's answer: both players randomize 50/50. Once both do, neither can improve their expected outcome by switching. The game's value settles at zero.

Sixteen years later, von Neumann and the economist Oskar Morgenstern published Theory of Games and Economic Behavior (1944). It arrived just as governments were starting to think seriously about zero-sum conflict of the literal, nuclear kind. Von Neumann became a RAND Corporation consultant in 1948, and RAND's early game-theoretic work on nuclear deterrence became one of the doctrine's intellectual foundations, one account even credits von Neumann himself with coining "MAD," mutually assured destruction, a taste for grim acronyms he shared with his earlier computer, MANIAC. One limitation remained: the framework only worked cleanly for zero-sum games. It took a twenty-two-year-old Princeton graduate student to fix that, twice over, in the same year.

Nash: Two Solutions in One Year, for Two Different Kinds of Game

John Nash published two separate, equally consequential results in 1950, and it's worth being precise about the difference, because they solve different problems and get conflated constantly.

The equilibrium concept. Nash's short paper "Equilibrium Points in N-Person Games" generalized von Neumann's zero-sum result to any competitive game, cooperative agreements not allowed.

The clearest illustration remains the Prisoner's Dilemma:

Betrayal is the better choice regardless of what the other player does, a dominant strategy, and both players reason their way to a worse outcome than mutual silence would have given them.

Not every game resolves this cleanly with pure strategies, sometimes the only stable equilibrium requires randomizing, exactly like matching pennies above, but in a general-sum game rather than a zero-sum one. Take Battle of the Sexes: two partners want to spend the evening together, one prefers Opera, the other prefers Football, but both prefer being together to being apart.

There are two pure equilibria (both go to Opera, or both go to Football), but there's also a mixed equilibrium, found using what's called the indifference principle: each player mixes their strategies in exactly the proportion that leaves the other player indifferent between their own two options.

Let p be the probability Partner 1 plays Opera, and q be the probability Partner 2 plays Opera.

So the mixed equilibrium has Partner 1 choosing Opera two-thirds of the time and Partner 2 choosing Opera only one-third of the time. Plug those probabilities back in and each partner's expected payoff comes out to exactly 2/3, worse for both of them than either pure equilibrium's payoff of 2 or 1. Randomizing resolves the coordination problem, but it resolves it badly, both partners frequently end up in different places entirely, which matters again later.

Nash's existence proof, that every finite game has at least one such equilibrium, mixed strategies allowed, rests on Kakutani's fixed-point theorem, a result from topology that guarantees a certain kind of "best response map" must cross itself somewhere. It's a genuinely elegant piece of mathematics doing quiet, load-bearing work behind a two-line definition.

The bargaining solution. In the same year, in a completely different paper, "The Bargaining Problem," Nash solved a cooperative question: if two parties can make a binding agreement, how should they split the surplus?

Nash proved this is the unique split satisfying four reasonable-sounding axioms: efficiency, symmetry, invariance to how utility is measured, and independence of irrelevant alternatives. Work it with numbers: two parties splitting a ₹100 surplus, where Party A's fallback option is worth ₹20 and Party B's is worth ₹10.

Maximizing (x−20)(90−x) across the feasible split resolves to x = 55. Notice the gains above each party's disagreement point come out identical, ₹35 each, out of a total surplus of ₹70 available above both fallback positions. That's the Nash bargaining solution's defining property with equal bargaining power: split the actual gains from trade evenly, so a party with a stronger outside option walks away with more overall, but the improvement each side gets from reaching a deal at all is shared equally.

Nash's own story took a genuinely difficult turn shortly after. He was diagnosed with paranoid schizophrenia in the late 1950s, spending much of the following three decades largely away from mathematical work. His gradual recovery in later years, without ongoing medication, is clinically unusual and was documented in Sylvia Nasar's 1998 biography A Beautiful Mind, later an Oscar-winning film. His 1994 Nobel, shared with Harsanyi and Selten, was genuinely contested inside the committee over his history of illness, and the episode reportedly changed the committee's selection procedures afterward. He recovered enough to attend the ceremony, and died alongside his wife Alicia in a 2015 car accident.

Cooperative Game Theory: When Players Are Allowed to Team Up?

Nash's bargaining solution handled exactly two parties splitting a surplus. Three years later, Lloyd Shapley took the same cooperative logic, binding agreements allowed, and generalized it to coalitions of any size: if a group's total winnings depend on who joins, how should the payoff be divided fairly among everyone?

Shapley's 1953 answer, the Shapley value, divides a coalition's payoff according to each player's average marginal contribution across every possible order in which players could join. Take three players who individually produce nothing, but generate value in pairs and more together:

Averaging each player's marginal contribution across all six possible joining orders:

The three values sum to exactly 120, the full value of the grand coalition, by design, the method always splits the whole pie with nothing left over. A related concept, the Core, asks a stricter question: which divisions are stable enough that no sub-group would be better off breaking away entirely. The Shapley value gives a fair split. The Core, when one exists, gives a split nobody has an incentive to walk away from.

That's cooperative game theory, binding deals allowed. Step back into a world where no such deals are possible, and a new problem appears, one that has nothing to do with fairness and everything to do with credibility.

Sequential Games and the Problem of Empty Threats

Nash's equilibrium works whether moves are simultaneous or sequential, but sequential games hide a trap: some Nash equilibria rely on threats nobody would actually carry out. Reinhard Selten fixed this in 1965 with subgame perfect equilibrium, solved by working backward from the end of the game.

Take a market entry game. A challenger decides whether to enter a market. If they enter, the incumbent decides whether to fight with a price war or accommodate and share the market.

Read as a simultaneous-move table, there appear to be two equilibria: Stay Out (if the entrant believes the incumbent will Fight) and Enter-Accommodate. Backward induction kills the first one. Work from the last decision: if entry actually happens, the incumbent compares −1 (Fight) against 2 (Accommodate) and picks Accommodate, fighting hurts the incumbent too. Knowing this, the entrant anticipates Accommodate, not Fight, and enters. The only subgame perfect equilibrium is (Enter, Accommodate). The incumbent's threat to fight was never credible, it only looked like an equilibrium because the simultaneous-move analysis never checked whether the threat would actually get carried out.

Games Where Nobody Knows Everyone's Cards

Every example so far assumes both players know each other's payoffs. Real negotiations rarely work that way. John Harsanyi solved this in 1967 to 1968 by introducing the idea of a player's "type," turning uncertainty about an opponent into a probability distribution a rational player can still optimize against.

A simple version: Player 1 negotiates with Player 2, who is privately either Tough (accepts only ₹70 or more) or Flexible (accepts ₹40), and Player 1 believes there's a 40 percent chance Player 2 is Tough.

Offering ₹40 has the higher expected value despite risking total failure against a Tough counterpart. This is the machinery underneath modern auction theory and insurance pricing. Harsanyi's players were at least consciously reasoning about probabilities and types, though. The next major leap in the field dropped consciousness from the model entirely.

Evolutionary Game Theory: When Biology Borrowed the Math

Nash equilibrium doesn't require conscious, calculating players, just that whatever strategy survives, survives because nothing else currently in play can beat it. John Maynard Smith and George Price formalized this for biology in a 1973 Nature paper, introducing the evolutionarily stable strategy, using the Hawk-Dove game: Hawks always escalate to a fight, Doves always back down.

With a resource worth 50 and an injury costing 100, p = 0.5, half Hawks, half Doves. Push the injury cost to 200 and the stable Hawk share drops to 0.25. The math doesn't care whether the "players" are lions or bacteria, it describes what any self-replicating strategy space settles into once costs and benefits are fixed.

Back among conscious, coordinating players, one problem from the non-cooperative world was still unsolved: what happens when binding contracts aren't allowed, but the players can still see a shared signal?

Coordination Without Cooperation: Correlated Equilibrium and Focal Points

Robert Aumann and Thomas Schelling shared the 2005 Nobel for exactly this. Recall the Battle of the Sexes mixed equilibrium from earlier, both partners ended up with an expected payoff of only 2/3, worse than either pure equilibrium's payoff of 2 or 1, because independent randomizing means they frequently land in different places entirely. Aumann's 1974 fix, correlated equilibrium, introduces a shared external signal, a coin flip both partners can see, that recommends an action to each of them, heads means both go to Opera, tails means both go to Football. Neither partner benefits from ignoring the signal once it's given, so it's still an equilibrium, but now the outcome averages the two good pure-equilibrium payoffs (2 and 1) instead of collapsing to the worse mixed-equilibrium payoff (2/3 each). A shared coin flip, with no binding contract at all, strictly beats independent randomizing.

Schelling's contribution was less mathematical and, in a way, more surprising: people often coordinate without any signal or agreement, by independently converging on whatever option simply feels distinctive. In his famous informal experiment, he asked people where they'd go to meet a stranger in New York City with no way to communicate beforehand. Most independently answered Grand Central Terminal, at noon. Nothing in the payoff structure makes that answer special, it's special only because everyone expects everyone else to think it's special. Schelling called these focal points.

Repeated Games and the Discovery That Niceness Can Win

The Prisoner's Dilemma looks bleak played once. Almost nothing in real life is played only once, and that changes everything.

In 1980, Robert Axelrod ran a computer tournament for a repeated Prisoner's Dilemma, two hundred rounds, end unknown in advance. The winner, submitted by Anatol Rapoport, was strikingly simple: cooperate first, then copy whatever the opponent did last time. Tit for tat. It won again in a second tournament the following year, against sixty-two entrants who'd all seen the first result. Axelrod's 1984 book The Evolution of Cooperation credited its success to four properties:

• Nice: never defects first
• Retaliatory: punishes defection immediately
• Forgiving: returns to cooperation as soon as the opponent does
• Clear: predictable enough that cooperation is easy to sustain

A 2025 replication at Cardiff University, rerunning the tournament with today's far larger set of strategies, found tit for tat ranked sixteenth, beaten by modern reinforcement-learning-trained entrants. The core finding, that reciprocity and forgiveness sustain cooperation better than blind aggression or naivety, held. The claim that tit for tat is the single best strategy in any environment did not.

Mechanism Design: Running the Whole Thing in Reverse

Every concept covered so far describes how to solve a game as it's given. One question had been sitting unanswered since 1961: what if you could design the rules instead?

William Vickrey answered this for auctions that year. In an ordinary first-price auction, bidders shade their bids below their true value. Vickrey's second-price auction fixes this: the highest bidder wins, but pays only the second-highest bid.

Bidding your true valuation becomes a dominant strategy. There's no strategizing left to do, no shading, no bluffing about what you'd pay, the winning move is simply to stop trying to outsmart the mechanism and state what the item is actually worth to you. Leonid Hurwicz, Eric Maskin, and Roger Myerson formalized mechanism design broadly and won the 2007 Nobel. Alvin Roth and Lloyd Shapley won the 2012 Nobel extending this to markets where money isn't allowed, kidney donor matching, medical residency placement. Paul Milgrom and Robert Wilson won the 2020 Nobel scaling Vickrey's insight to US spectrum auctions, the first FCC auction using their format, in July 1994, sold ten licenses across forty-seven rounds for 617 million dollars, for spectrum previously handed out nearly free. Auctions built on their design have since allocated more than a hundred billion dollars worldwide.

Where the Math Actually Shows Up in the World Today

• Online ad auctions: Google's and Meta's ad marketplaces run on descendants of Vickrey's logic, the Vickrey-Clarke-Groves mechanism, so advertisers bid their genuine value rather than gaming the system
• Professional football: Ignacio Palacios-Huerta's 2003 study of 1,417 penalty kicks found kickers and goalkeepers mix shot and dive direction in almost exactly the proportions a mixed-strategy equilibrium predicts, confirmed by a follow-up study of 12,399 kicks
• Poker AI: Carnegie Mellon's Libratus beat four professional players over 120,000 hands of heads-up hold'em in 2017 at 99.98 percent statistical confidence. Its successor Pluribus beat professionals at six-player poker in 2019, where a clean Nash equilibrium isn't even guaranteed to exist
• International relations: mutually assured destruction is formally a Nash equilibrium, a stable outcome, not necessarily a good one
• Everyday market design: ride-share surge pricing, kidney exchange programs, school choice algorithms, and blockchain consensus protocols all lean directly on mechanism design principles

Where Game Theory Stands Today, and Where It's Headed

Two shifts define the field's current frontier. The first is behavioral: real people don't reliably play Nash equilibria, precisely the gap Libratus and Pluribus learned to exploit against professionals who were already about as game-theoretically sharp as humans get. The second is algorithmic: computing an equilibrium for something as complicated as six-player poker isn't a by-hand exercise anymore, and for many classes of games it's known to be computationally hard even for a machine, pushing the field toward measuring how much worse things get when everyone acts selfishly without coordination, the price of anarchy, now showing up in traffic routing, network design, and competing AI systems.

The honest outlook is that game theory has stopped being one elegant theory and become a toolkit, deployed wherever multiple self-interested parties have to share an outcome. Less tidy than what von Neumann and Morgenstern set out to build in 1944. Eighty years on, considerably more useful.

Conclusion & Key Takeaways

The through-line across eighty years isn't that rational people cooperate, or that they don't. It's that the outcome depends entirely on the structure of the game, and this field's real achievement has been building a distinct, mathematically rigorous answer for nearly every way that structure can vary.

Sources

John von Neumann's 1928 minimax theorem and the origins of formal game theory, via arXiv Roger Myerson on Émile Borel and the foundations of game theory John Nash, "Equilibrium Points in N-Person Games," PNAS (1950), open access via PMC John Nash, "The Bargaining Problem," Econometrica (1950), via the Econometric Society John F. Nash Jr., Nobel Prize facts page, Nobel Prize committee Maynard Smith and Price, "The Logic of Animal Conflict" (1973), overview via Stanford Encyclopedia of Philosophy The 2005 Nobel Prize to Aumann and Schelling, popular science explanation, Nobel Prize committee The 2020 Nobel Prize in auction theory, popular science explanation, Nobel Prize committee Cardiff University's 2025 replication of Axelrod's Prisoner's Dilemma tournaments Carnegie Mellon University, "Carnegie Mellon and Facebook AI Beats Professionals in Six-Player Poker" Antoine Augustin Cournot, Recherches sur les Principes Mathématiques de la Théorie des Richesses (1838); Ernst Zermelo, "Über eine Anwendung der Mengenlehre auf die Theorie des Schachspiels" (1913); John von Neumann and Oskar Morgenstern, Theory of Games and Economic Behavior (1944); Reinhard Selten, "Spieltheoretische Behandlung eines Oligopolmodells mit Nachfragetragheit" (1965); John Harsanyi, "Games with Incomplete Information Played by Bayesian Players" (1967-68); Lloyd Shapley, "A Value for n-Person Games" (1953); Robert Aumann, "Subjectivity and Correlation in Randomized Strategies" (1974); Robert Axelrod, The Evolution of Cooperation (1984); William Vickrey, "Counterspeculation, Auctions, and Competitive Sealed Tenders" (1961); Ignacio Palacios-Huerta, "Professionals Play Minimax" (2003); Noam Brown and Tuomas Sandholm, "Superhuman AI for heads-up no-limit poker: Libratus beats top professionals," Science (2018); Nobel Prize committee citations (1994, 2005, 2007, 2012, 2020)