Von Neumann's minimax theorem through Fourier-Motzkin elimination

Mark Voorneveld

2024-08-21

Abstract:Fourier-Motzkin elimination, a standard method for solving systems of linear inequalities, leads to an elementary, short, and self-contained proof of von Neumann's minimax theorem.

Computer Science and Game Theory,Theoretical Economics

What problem does this paper attempt to address?

This paper attempts to provide a concise and self - contained proof by the Fourier - Motzkin Elimination (FME) method to solve the problem of the von Neumann minimax theorem.

### Background and Problem Description

The von Neumann minimax theorem is a fundamental result in game theory. It shows that in a zero - sum game, the optimal strategies of the two players can be found by the method of linear programming, and the expected payoffs corresponding to these two strategies are equal. Specifically, for a given payoff matrix \(A\in\mathbb{R}^{m\times n}\), there exist probability vectors \(p\) and \(q\) such that:

\[
\max_{p}\min_{q}p^{\top}Aq = \min_{q}\max_{p}p^{\top}Aq
\]

where:
- \(p\) is the probability distribution of player 1's choices, satisfying \(p\geq0\) and \(1^{\top}p = 1\)
- \(q\) is the probability distribution of player 2's choices, satisfying \(q\geq0\) and \(1^{\top}q = 1\)

### Main Contributions of the Paper

This paper directly proves the von Neumann minimax theorem by the Fourier - Motzkin Elimination method without relying on other intermediate results (such as Farkas' Lemma or the Linear Programming Duality Theorem). The Fourier - Motzkin Elimination method is a method for solving systems of linear inequalities. It can gradually eliminate variables and finally obtain an inequality system regarding the remaining variables.

### Specific Steps

1. **Introduction to the Fourier - Motzkin Elimination Method**:
- Given a system of linear inequalities \(Ax\leq b\), by gradually eliminating variables \(x_{1},x_{2},\ldots,x_{k}\), finally obtain an inequality system regarding the remaining variables.
- Each time a variable is eliminated, new inequalities are generated. These inequalities are in the form of non - negative combinations \(y^{\top}Ax\leq y^{\top}b\), where \(y\) is a non - negative vector.

2. **Application to the Minimax Theorem**:
- Write the constraint conditions in the minimax theorem in the form of linear inequalities, for example:
\[
\hat{A}\begin{bmatrix}q\\v\end{bmatrix}\leq\hat{b}
\]
where \(\hat{A}\) and \(\hat{b}\) are specific matrices and vectors.
- Eliminate the variable \(q\) by the Fourier - Motzkin Elimination method to obtain an inequality system that only contains \(v\).
- Find the optimal value \(v^{*}\) from these inequalities and verify its feasibility.

3. **Conclusion**:
- Prove that there exist \((p^{*},q^{*},v^{*})\) such that \((p^{*},v^{*})\) is feasible in the maximization problem, \((q^{*},v^{*})\) is feasible in the minimization problem, and the optimal values of the two problems are the same.

### Summary

This paper provides a concise, self - contained proof method without additional assumptions by the Fourier - Motzkin Elimination method, which solves the proof problem of the von Neumann minimax theorem. This method is not only applicable to the real number field but also can be extended to other ordered fields (such as the rational number field), and does not require the use of topological arguments (such as the compactness of the strategy space or the continuity of the payoff function).