The Minimax Theorem

In the [YOU] on AI Field Guide

The cycle framed by [YOU] on AI is centrally concerned with what happens when AI systems become agents—entities that pursue goals, model other agents, and choose strategies in light of what others will do. The minimax theorem and its extensions are the mathematical skeleton of that agency. The entire apparatus of reinforcement learning, in which an agent learns to maximize cumulative reward against an environment or adversary, is the minimax framework instantiated in code and set loose to optimize. When AI safety researchers worry about a capable system that treats obstacles as adversaries and finds the strategy that guarantees its ends against opposition, they are describing minimax play by a sufficiently powerful optimizer—and the mathematics says that such play is rational, in the precise sense the theorem establishes.

The theorem also frames the gap between rationality and desirability that runs through the cycle. Von Neumann proved that rational play in zero-sum conflict is possible and computable. But the framework is silent on whether the game being played is one we want to play, whether the objectives being maximized are ones we endorse, and whether the optimal strategy against an adversary is compatible with the broader requirements of a good society. An AI system that models its regulatory environment as an adversary and plays minimax against it will, by the theorem's logic, find the optimal strategy for defeating regulation—and this is a consequence of applying rational-agent theory to the wrong game, not a malfunction of the mathematics. The minimax theorem describes what rationality requires; wisdom must specify when rationality's requirements are to be constrained.

Origin

Von Neumann published the minimax theorem in “Zur Theorie der Gesellschaftsspiele” (On the Theory of Parlor Games) in 1928. The proof had two key components: the recognition that optimal play might require randomization—mixed strategies—and the application of a fixed-point theorem to guarantee that optimal strategies exist. Von Neumann himself reported that he had the insight in the mid-1920s but required several years to find a rigorous proof; the fixed-point machinery was new and powerful, borrowed from topology.

With Oskar Morgenstern he developed the result into a full theory of strategic interaction in the Theory of Games and Economic Behavior (1944). The most important extension was the expected-utility theorem: they showed that an agent whose preferences satisfy a few apparently reasonable axioms must act as though maximizing the expected value of some numerical utility function. This converted the vague notion of rational choice into a precise mathematical prescription and became the standard formal model of the rational agent in economics, decision theory, and AI.

Key Ideas

Minimax and the value of a game. The theorem establishes that in any finite two-person zero-sum game, there exists a unique value V such that the first player can guarantee an expected payoff of at least V, and the second player can hold the first player to at most V, regardless of what either does. The strategies that achieve this are the minimax strategies. The existence of a unique value means that rational play in pure conflict is determinate: there is a fact about what each side should do, and the outcome is predictable once you know the players are rational. This determinacy makes the minimax framework enormously useful for AI: it tells you what a rational adversary will do, so you can design systems that are robust against optimal opposition.

Mixed strategies and the necessity of randomization. The minimax strategies often require randomization. In rock-paper-scissors, no deterministic strategy is optimal; the only equilibrium is to choose each option with equal probability. This is not a curiosity but a fundamental fact about strategic interaction: against a sufficiently capable adversary, any predictable strategy can be exploited. The implication for AI security is direct—deterministic systems can be gamed; robust ones must inject controlled randomness. Adversarial robustness in machine learning, the design of models that resist adversarial inputs, is in part an application of this insight: a model that behaves deterministically is vulnerable in a way that a model with calibrated uncertainty is not.

From zero-sum to general games: the rational agent. Von Neumann and Morgenstern's extension from zero-sum games to general games of mixed motive produced the expected utility framework. A rational agent, on this account, is an entity that acts so as to maximize the expected value of some utility function, given its beliefs about the world. This is now the standard formal model of intelligence in AI: a reinforcement learning agent maximizes cumulative reward; a planning agent maximizes expected utility; a language model fine-tuned by human feedback is trained to maximize a proxy for human approval. All are instantiations of the rational agent von Neumann and Morgenstern formalized.

The alignment implication. The rational agent takes its utility function as given and optimizes toward it without evaluating whether the function is the right one. This is simultaneously the framework's power—any objective can be pursued with optimal efficiency—and the source of AI alignment's hardest problem. A highly capable rational agent will pursue its specified objective with a thoroughness that exposes every gap between the specified objective and the true goal. The problem of specifying objectives that capture what we actually want has turned out to be among the hardest in the field. The machines force us to articulate what we value, because they will pursue exactly what we say. Von Neumann gave us the agents that make the specification necessary, and the framework that shows why it is so hard to get right.