Minimax Theorem¶
A theorem family giving hypotheses under which opposed max–min and min–max values coincide, thereby certifying a saddle value.
Core Idea¶
A minimax theorem states conditions under which the universal max–min inequality closes to equality:
The left side is the best value the maximizing chooser can secure before the minimizing chooser responds; the right side is the least upper value the minimizer can enforce when commitment order is reversed. The inequality from left to right always holds. Equality is the substantive theorem: it proves that neither order of optimization has an advantage and identifies a common saddle value.
Scope of Application¶
In finite zero-sum games, let \(A\in\mathbb R^{m\times n}\) be the row player's payoff matrix. Mixed strategies \(p\in\Delta_m\) and \(q\in\Delta_n\) produce expected payoff \(p^{\mathsf T}Aq\). Von Neumann's theorem gives
The compact simplexes ensure attainment, while bilinearity supplies both convex and concave structure. Each player therefore has an optimal mixed strategy and the game has a value.
Clarity¶
Three propositions must be kept separate:
- The weak inequality \(\sup\inf\le\inf\sup\) is formal and unconditional.
- Equality requires a minimax theorem and its hypotheses.
- Existence of optimizers requires attainment conditions in addition to equality of extended-real values.
For matching pennies with
Manages Complexity¶
The theorem compresses an infinite response analysis into one equality certificate. Without it, one must track two generally different values depending on who commits first. With it, a single number \(v\) summarizes opposed optimization and supports independent computation from either side.
In matrix games, this compression also links the two players' problems to primal-dual linear programs.
Abstract Reasoning¶
The weak inequality follows pointwise. For every \(x\in X\), \(y\in Y\),
Taking the supremum over \(x\) on the left and the infimum over \(y\) on the right yields \(\sup_x\inf_y f\le\inf_y\sup_x f\).
Knowledge Transfer¶
The finite game form transfers to convex-concave optimization by replacing probability simplexes with convex feasible sets and a bilinear payoff with a suitably quasi-concave/quasi-convex, semicontinuous function. Sion's theorem requires one choice set compact and both sets convex under an appropriate one-sided version; symmetry is not automatic.
The same quantifier exchange appears in strong duality, adversarial learning, and statistical minimax risk, but those applications must reconstruct the theorem's roles explicitly. The transfer fails when the action spaces are nonconvex without randomization, when regularity is absent, or when extrema escape to infinity.
Relationships to Other Abstractions¶
Current abstraction Minimax Theorem Domain-specific
Parents (1) — more general patterns this builds on
-
Minimax Theorem presupposes Minimax Strategy Prime
Minimax Strategy is the proposed minimal parent by composition: the theorem presupposes the two-sided worst-case optimization pattern and proves when its lower and upper values agree.
Hierarchy path (1) — routes to 1 parentless root
- Minimax Theorem → Minimax Strategy → Optimization
Neighborhood in Abstraction Space¶
Minimax Theorem sits in a sparse region of the domain-specific corpus (93rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Decision Under Risk & Ambiguity (13 abstractions)
Nearest neighbors
- Maximum Theorem — 0.79
- Decoy Effect — 0.78
- Fourier–Motzkin Elimination — 0.78
- Monotonic Function — 0.77
- Median Voter Theorem — 0.77
Computed from structural-signature embeddings · 2026-09-08