Zero-Sum Games: Incentives, Strategy, Equilibria

Build a working intuition for zero-sum and constant-sum games, learn to spot pure-strategy solutions fast, and understand why minimax and mixed strategies are about guarantees under worst-case play. You will also learn when the model breaks and what to use instead.

Zero-sum games are the places where being nice does not help, because every advantage you give away becomes someone else’s advantage. The moment payoffs are locked together, good strategy stops meaning what feels fair and starts meaning what stays safe when the other side actively tries to hurt you. That shift, from cooperate to outguess, is the whole point. Once you see incentives as mirrored constraints, equilibrium becomes less mysterious. It becomes the point where neither player can do better by changing their own move, given the other is trying just as hard.

Zero-sum incentives in one picture

The cleanest way to recognize a zero-sum game is that one player’s payoff determines the other’s, typically as the negative. If the row player gets $+2$, the column player gets $-2$. That single fact turns strategy into a contest of anticipating best responses rather than coordinating on a mutually good outcome.

See how the incentives mirror across players in a simple payoff table.

Once you internalize the mirror, lots of familiar instincts become unreliable. A move that looks generous is just a move that lowers your own payoff. A move that looks risky might be correct if it prevents the opponent from steering you into your worst-case outcome.

Mirrored incentives
In a true zero-sum setting, any improvement for you is, by definition, a deterioration for them, so stable outcomes come from constraint and leverage, not goodwill.

Payoff matrices and perspectives

A payoff matrix is a perspective, not the game itself. You can write payoffs as a pair in each cell, or as a single matrix for one player and infer the other by negation. You can also swap the labels row and column and the strategic content is unchanged, even though the picture looks different.

Constant-sum is the quiet cousin. If every outcome sums to the same constant $c$, you can convert it to zero-sum by subtracting $c/2$ from each player’s payoff in every cell. The strategic choices do not change because you did not change anyone’s incentives across actions, you only shifted the baseline. Best responses stay best responses.

Explore how shifting from constant-sum to zero-sum changes the numbers but not the strategic logic.

This matters in practice because people often argue about whether a situation is zero-sum when it is merely constant-sum under a different accounting. If the shift is truly uniform across outcomes, equilibrium analysis is invariant. If it is not uniform, the game is different.

Dominance, best responses, and saddle points

A fast way to solve many zero-sum matrices is to look for structure before you calculate anything.

What you check first

• Dominated strategy: an action that is never better and sometimes worse than another, so it can be eliminated.
• Best response: for each opponent action, the move that gives you the highest payoff.
• Saddle point: a cell that is simultaneously the row player’s best response to the column choice and the column player’s best response to the row choice.

In a zero-sum matrix, a saddle point corresponds to a pure-strategy equilibrium and a game value you can read directly off the table. If it exists, neither player needs randomness. If it does not, that absence is already a diagnosis. The opponent can exploit any fixed move.

Use the annotated matrix to practice eliminating dominated strategies and spotting a saddle point when one exists.

The deeper idea is that dominance simplifies the game without losing anything strategically. You are removing options that a rational player would never pick, which shrinks the search space and makes the remaining incentives easier to see.

Solve before math
If you can eliminate even one dominated row or column, you often turn a messy matrix into a near-obvious best-response picture.

Minimax theorem and worst-case guarantees

In a finite two-player zero-sum game, the minimax theorem says the row player’s best guaranteed payoff equals the column player’s best way to hold the row player down. Formally, for payoff matrix $A$ to the row player,

$$\max_{p}\min_{q}p^\top Aq=\min_{q}\max_{p}p^\top Aq.$$

That shared quantity is the value of the game.

The point is not that the opponent is always adversarial in real life. The point is that minimax tells you what you can guarantee even if they are. Your strategy is chosen to maximize your worst case, and the opponent’s is chosen to minimize your best case. When those meet, neither side can improve their outcome by changing strategies unilaterally.

Try adjusting a mixed row strategy and see what payoff you can guarantee against the worst-case response.

Minimax is a mental model for stakes. If you are building a system that must not fail, or pricing a contract with a hostile counterparty, guarantee can matter more than average-case performance. In other contexts, worst-case thinking can be overly conservative, so it is something you choose, not something you worship.

Mixed strategies and deliberate randomness

When there is no saddle point, any fixed action can be punished. Mixed strategy play fixes that by making your behavior unpredictable in a controlled way. The goal is not to be random for its own sake. The goal is to choose probabilities that remove the opponent’s ability to gain by tailoring a response to your pattern.

A useful featured-snippet way to say it is this. Mixed strategies in zero-sum games work by making the opponent indifferent among the actions you want them to consider. You pick probabilities so their expected payoff from each of those responses is equal. If they are indifferent, they cannot exploit you by switching.

For a $2\times2$ game, you often solve by setting two expected payoffs equal and solving for the mixing probability. For larger games, the same idea generalizes into linear constraints, which is one reason linear programming shows up naturally in zero-sum analysis.

Play with the mixing probability and watch where the opponent becomes indifferent.

Common intuition traps show up here. People think mixing means you do not know what to do. In reality, mixing is what you do when you know exactly what the opponent would do against each pure move. You are choosing probabilities to control the opponent’s incentives.

Randomness with purpose
A good mix is not noisy. It is calibrated to deny the opponent a profitable adjustment.

Sequential zero-sum games and policies

Matrices hide time. Many real zero-sum situations are sequential. You move, the opponent observes something, they move, chance happens, information is revealed, and so on. In a sequential setting, optimal play is not a single move, it is a contingent plan. That plan is a policy, meaning what you will do at each decision point given what you know then.

With perfect information, you can compute optimal play by backing up values from terminal outcomes to earlier decisions using minimax. With hidden information, you reason over information sets, where different underlying states look the same to the player. You still back up values, but now you are choosing actions that are robust to uncertainty about where in the tree you really are.

Look at how minimax values back up through a simple game tree with a hidden-information node.

This framing matters because people often apply matrix intuition to sequential problems and get burned. A move that is weak in isolation may be strong because it sets up future constraints. In sequential games, threat credibility and information flow can be as important as the immediate payoff.

When it is not actually zero-sum

The most common modeling mistake is treating a conflict as zero-sum when payoffs are not coupled that tightly. Externalities can create value or destroy value. Side payments can transform incentives. Repeated play can support cooperation because future consequences matter. Even when there is a fixed resource, the way it is measured can make it look constant-sum when it is not.

A practical test is to ask what changes the size of the pie. Innovation, trade, reputation, regulation, and risk can all move total value. If total value moves differently across outcomes, a constant shift does not preserve incentives and you are no longer in the tidy minimax world.

Compare the telltale signs of zero-sum, positive-sum, and negative-sum situations.

If you choose the wrong model, you choose the wrong strategy. Treating a positive-sum negotiation as zero-sum can make you leave value on the table. Treating a zero-sum security contest as cooperative can make you fragile.

Choosing minimax without becoming paranoid

Minimax is a stance toward uncertainty and adversaries, not a personality trait. Use it when the downside is unacceptable, when the opponent’s incentives are truly opposed, or when you cannot trust the environment to be kind. Switch to models that include learning, repeated interaction, or value creation when the situation actually has room for it.

A good next step is to take one conflict you care about and rewrite it twice. First as zero-sum, explicitly stating what is conserved. Then as non-zero-sum, explicitly naming what can expand or shrink. The gap between the two versions tells you where strategy is about defense and where it is about design.