LP Optimum ↔ Game-Theoretic Optimal Strategy
Theorem. Let $A : I \times J \to \mathbb{R}$ be a finite matrix game with value $\operatorname{val}(A) \in \mathbb{R}$ ([[game_theory.strategic_game.zero_sum.core.value]]).
-
The pair $(x^*, v^*)$ is optimal for player 1's LP ([[math.linear_programming.minimax_bridge.player_1_lp]]) if and only if $x^*$ is a maxmin (optimal) mixed strategy for player 1 and $v^* = \operatorname{val}(A)$.
-
The pair $(y^*, w^*)$ is optimal for player 2's LP ([[math.linear_programming.minimax_bridge.player_2_lp]]) if and only if $y^*$ is a minmax (optimal) mixed strategy for player 2 and $w^* = \operatorname{val}(A)$.
-
Any pair $(x^*, y^*)$ obtained from LP optima is a saddle point of $A$ ([[game_theory.strategic_game.zero_sum.core.saddle_point]]); conversely every saddle point gives optimal LP solutions.
Proof
(⇒) LP optimum gives optimal strategy. If $(x^*, v^*)$ is optimal for player 1's LP, then by the constraint $xA \ge v^* \cdot \mathbf{1}_J$ at $(x^*, v^*)$, player 1's mixed strategy $x^*$ guarantees at least $v^*$ against every pure column, hence against every mixed column. By LP-duality with [[math.linear_programming.minimax_bridge.player_2_lp]] and [[math.linear_programming.strong_duality]], $v^* = w^*$ equals the LP common value, which is also $\operatorname{val}(A)$ by the standard maxmin-minmax characterisation.
(⇐) Optimal strategy gives LP optimum. If $x^*$ achieves $\max_x \min_j (xA)_j = \operatorname{val}(A)$, then $(x^*, \operatorname{val}(A))$ satisfies all LP constraints and attains the objective at $v = \operatorname{val}(A)$. Symmetric for player 2.
The saddle-point statement (3) follows by combining (1) and (2): optimal $(x^*, y^*)$ jointly satisfy $x^* A \ge \operatorname{val}(A) \cdot \mathbf{1}_J$ and $A y^* \le \operatorname{val}(A) \cdot \mathbf{1}_I$, which is exactly the saddle-point characterisation ([[game_theory.strategic_game.zero_sum.core.optimal_pairs_are_saddle_points]]).
Significance
This theorem is the clean conversion between the LP and game-theoretic viewpoints:
- Algorithmically, it lets us compute optimal mixed strategies with any LP solver (simplex, interior-point, ...) — the entire computational complexity of zero-sum games is exactly the LP complexity.
- Mathematically, it says LP-duality and the minimax theorem are equivalent statements ([[math.minimax.lp_duality_minimax_proof]]).
- Structurally, support and strong-complementarity properties ([[math.minimax.support_complementarity]], [[math.minimax.strong_complementarity]]) transfer cleanly between the two pictures.
References
- [AGT Section 1.4.2, Thm. 1.11] Nisan, Roughgarden, Tardos, and Vazirani, Algorithmic Game Theory.
- [MFoGT Section 2.5] Laraki, Renault, and Sorin, Mathematical Foundations of Game Theory.