Game Theory and Decision Analysis

Game theory provides a rigorous mathematical framework for analysing strategic interactions where your outcome depends not just on your own choices, but on the decisions of others. Mastering its core principles is essential for modelling conflicts, negotiations, and competitive scenarios in fields ranging from economics and business to evolutionary biology and military strategy. This article focuses on the foundational class of two-player zero-sum games, where one player's gain is exactly the other's loss.

Payoff Matrices and the Structure of Conflict

A two-player zero-sum game is defined by three elements: two players (conventionally called Row and Column), a set of possible strategies for each, and a payoff function. The entire game can be represented compactly in a payoff matrix. By convention, the matrix entries show the payoff to the Row player. Since the game is zero-sum, the payoff to the Column player is simply the negative of that number.

Consider a simple game between two companies, A and B, competing for market share. Company A (Row) can choose to launch a new product (Strategy A1) or increase advertising (Strategy A2). Company B (Column) can respond by cutting prices (B1) or enhancing its own product (B2). A hypothetical payoff matrix, showing the change in market share percentage for Company A, might look like this:

Here, if A chooses A1 and B chooses B1, A gains 3% market share (and B loses 3%). If A chooses A2 and B chooses B2, A gains 2%. This matrix is the starting point for all subsequent analysis. Your first task is always to correctly identify the players, their strategies, and construct or interpret this payoff matrix from a described scenario.

The Minimax Theorem and Pure Strategy Solutions

In a zero-sum game, players are assumed to be rational and to anticipate their opponent's rationality. The Row player wants to maximize their minimum guaranteed payoff, while the Column player wants to minimize their maximum possible loss. This leads to the minimax principle.

To find the maximin value for Row, you examine each row and find the worst-case (minimum) payoff within that row, assuming Column chooses the most damaging response. Then, you choose the row that offers the best of these worst-case scenarios. For our matrix:

• Row A1: min(3, -1) = -1
• Row A2: min(-2, 2) = -2

The maximin for Row is the maximum of these row minima: max(-1, -2) = -1.

For the Column player, you examine each column and find the worst-case (maximum payoff to Row, meaning maximum loss for Column) within that column. Then, you choose the column that offers the least of these worst-case losses. This is the minimax value for Column:

• Column B1: max(3, -2) = 3
• Column B2: max(-1, 2) = 2

The minimax for Column is the minimum of these column maxima: min(3, 2) = 2.

In this game, the maximin (-1) does not equal the minimax (2). This indicates there is no saddle point, or stable pure strategy solution. A saddle point exists if the maximin value equals the minimax value. That single cell represents the outcome where neither player can unilaterally improve their position by changing strategy; it is the optimal pure strategy for both. The value at the saddle point is called the value of the game ($V$). When a saddle point exists, the game is solved: players should consistently play their corresponding strategies.

Mixed Strategies and Expected Payoff

When no saddle point exists, as in our example, the solution requires mixed strategies. This means players randomize their choices according to a specific probability distribution to make themselves unpredictable and maximize their expected payoff.

Suppose Row plays strategy A1 with probability $p$ and A2 with probability $(1-p)$. Suppose Column plays B1 with probability $q$ and B2 with probability $(1-q)$. The expected payoff to Row, $E(p,q)$, is calculated as a weighted average: $$E(p,q)=3pq+(-1)p(1-q)+(-2)(1-p)q+2(1-p)(1-q)$$

The fundamental theorem for finite zero-sum games (von Neumann's Minimax Theorem) states that there exists a pair of optimal mixed strategies such that the maximin expected payoff equals the minimax expected payoff, defining the value of the game. To find Row's optimal mix ($p$), we use the principle of indifference: Row chooses $p$ so that Column's expected payoff is the same regardless of which pure strategy Column chooses. This makes Column indifferent, removing any advantage from exploiting a predictable pattern.

Set Column's expected payoff from playing B1 equal to that from playing B2:

• Expected loss for Column if playing B1: $E_{\text{B1}}=3p+(-2)(1-p)$
• Expected loss for Column if playing B2: $E_{\text{B2}}=(-1)p+2(1-p)$

Set $E_{\text{B1}}=E_{\text{B2}}$: $$3p-2(1-p)=-1p+2(1-p)$$ $$3p-2+2p=-p+2-2p$$ $$5p-2=2-3p$$ $$8p=4$$ $$p=0.5$$

Thus, Row's optimal strategy is to play A1 and A2 each with 50% probability. A similar process, setting Row's expected payoff from A1 equal to that from A2, yields Column's optimal mix ($q$). You would find $q=0.5$ as well. Substituting $p=0.5$ and $q=0.5$ into the expected payoff formula gives the value of the game, $V=0.5$. On average, Company A gains 0.5% market share under optimal play.

Reducing Complexity: Dominance Arguments

Before performing any calculations, you should simplify the payoff matrix using dominance arguments. A strategy is dominated if there exists another strategy that yields a payoff at least as good in all scenarios and strictly better in at least one. Rational players will never play a dominated strategy, allowing you to eliminate it from the matrix.

There are two types of dominance:

1. Strict Dominance: Strategy X strictly dominates Strategy Y if the payoff from X is strictly greater than the payoff from Y, regardless of the opponent's choice. Y can be deleted.
2. Weak Dominance: Strategy X weakly dominates Strategy Y if the payoff from X is at least as good as the payoff from Y for all opponent plays, and strictly better for at least one. Weakly dominated strategies can often be eliminated, but caution is needed as they might be part of an optimal mix in some edge cases.

For example, consider this matrix:

Compare R1 and R3. Against C1: 4 > 1. Against C2: 1 > -1. Against C3: 3 > 0. R1 strictly dominates R3. We can eliminate row R3. Next, examine the columns from Column's perspective (seeking low numbers, as they represent payouts to Row). After eliminating R3, compare C1 and C3. Against R1: 4 vs 3 (C3 is better for Column). Against R2: 2 vs 2 (they are equal). C3 is at least as good as C1 for Column, and strictly better against R1. Therefore, C3 weakly dominates C1. We can eliminate column C1. This reduces the game to a simpler 2x2 matrix for easier analysis.

Common Pitfalls

1. Assuming a Pure Strategy Solution Always Exists: A frequent error is to look at the matrix, pick what seems like a "best" move, and stop. You must systematically check for a saddle point by calculating the maximin and minimax values. If they are unequal, you must proceed to mixed strategies.

1. Misapplying the Principle of Indifference: When solving for mixed strategies, the indifference condition is used to find the opponent's optimal probabilities. A common mistake is to set the player's own expected payoffs from their strategies equal. Remember: Row chooses $p$ to make Column indifferent, and Column chooses $q$ to make Row indifferent.

1. Incorrectly Handling Dominance in Non-Zero-Sum Games: The rules of strict and weak dominance are logical and apply to all strategic games. However, the pitfall is assuming that eliminating a dominated strategy never affects the equilibrium outcome in more complex, non-zero-sum games (like the Prisoner's Dilemma). While safe in zero-sum games, in other game types the iterative elimination of dominated strategies requires more careful justification.

1. Confusing Payoff Perspective: In a zero-sum payoff matrix, always verify whose payoff is displayed. If the matrix shows payoffs to Row, then Column's goal is to minimize these numbers. When calculating dominance for Column, you must think from Column's perspective of minimizing the matrix entry, not maximizing it.

Summary

• Two-player zero-sum games are completely defined by a payoff matrix showing the reward to one player. The core solution concept is based on the minimax theorem, where each player optimizes their worst-case scenario.
• A saddle point, where the maximin equals the minimax, indicates an optimal pure strategy for both players. The value at this point is the value of the game ($V$).
• If no saddle point exists, the solution involves mixed strategies. Players randomize their moves with specific probabilities, found using the principle of indifference, to maximize their expected payoff.
• Dominance arguments (strict and weak) allow you to eliminate rationally inferior strategies before calculation, simplifying the payoff matrix and reducing computational complexity.
• These tools are directly applicable to modeling competitive scenarios in business, economics, and strategy, providing a mathematical basis for making robust decisions against a rational opponent.