Two enemies, one number. In any finite two-person zero-sum game, what the row player can guarantee by playing well equals what the column player can hold them to — maxx miny x′Ay = miny maxx x′Ay. That shared value is the value of the game. Von Neumann proved it in 1928; there is no duality gap. Down the center: a payoff matrix goes in, the solver finds the value and the optimal mixed strategies, the proof comes out. Blue builds it; red tries to break the guarantee.
source von Neumann, Zur Theorie der Gesellschaftsspiele, Mathematische Annalen 100 (1928) 295–320 — eudml.org/doc/159291. Rendered, not quoted.
A payoff matrix A. The row player picks a mixed strategy x to maximize; the column player picks y to minimize. Two quantities:
maximin = the most the row player can secure no matter what column does. minimax = the least column can force no matter what row does. Weak duality gives maximin ≤ minimax always; von Neumann proved they are equal — once mixing is allowed.
The solver equalizes on the optimal support (a saddle when one exists) — no learning loop, an exact linear solve. Live, for the current game:
The minimax theorem is the Nash equilibrium specialized to pure conflict: in a zero-sum game every equilibrium yields the same payoff — the value — so "an equilibrium exists" collapses to "the game has a number." Nash (1950) restored existence for non-zero-sum play, but lost the clean value and the max=min interchange.
Downstream, this value is exactly what the Monte-Carlo tree search approximates: game-tree search is a hunt for the minimax value, here reached in closed form on tiny games. Each sphere is the next one's premise.
The blue team's live check: re-solve every preset game, confirm no duality gap and that x* secures the value against all replies — and confirm the guarantee fails the instant red abandons the mix. If tampering breaks the guarantee, this badge turns red.
Pick a game, or edit the numbers of a 2×2. Entry A[i][j] is what the row player (maximizer) wins and the column player (minimizer) pays when row plays i and column plays j. Zero-sum: one gain is the other's loss.
The panel plays x* and lets the opponent best-respond with every pure reply. Payoff x·A·ej must stay ≥ v:
Change the matrix — value and strategies are re-solved on the spot by linear algebra, never looked up.
What the machine produces: for every game a single number v with maximin = minimax = v (no gap), an optimal strategy x* that secures v against any opponent, and a y* that caps it there. Rock-Paper-Scissors comes out (⅓,⅓,⅓) with value 0; matching pennies (½,½), value 0; the saddle game a pure pair.
The blue witness (left) confirms these live; the red team (right) tries to make the guarantee break.
And it is an existence theorem: it says the value is there, and (via LP duality) is computable — but for astronomically large games you still cannot enumerate the matrix. Minimax tells you the game has an answer; it does not make a hard game cheap.
"The maximin strategy only works if the opponent plays optimally." Cut. The opposite: x* secures the value against any reply, even a foolish one — that is the whole point, and the panel checks it against every pure column.
"Von Neumann proved it with a fixed-point theorem, like Nash." Cut. The 1928 proof is elementary (topological in flavour, not Kakutani); Brouwer/Kakutani fixed points came with Nash in 1950 for the general case. Minimax also follows from LP duality.
"Every game has a saddle point in pure strategies." Kept, corrected. Matching Pennies and RPS have none — the value exists only because mixed strategies are allowed. Mixing is not a trick; it is the content of the theorem.
The red team's move: in Rock-Paper-Scissors, abandon the maximin mix and commit to a pure strategy — all Rock. The minimizing opponent best-responds with all Paper and drives the payoff to −1, strictly below the value 0. The blue witness (window 7) is watching.
Commit to one pure choice and the opponent exploits it — the guarantee check finds a reply below v, the witness recomputes and turns red. Nothing is faked; the attack is real and it is caught.