Imagine you’re playing a game of chess, and you’re down to the final few moves. How do you decide what to do? You think ahead-“If I move here, my opponent will likely move there, and then I’ll respond like this.” This kind of reverse thinking, starting from the end and working backwards, is at the heart of finding a subgame-perfect Nash equilibrium, particularly when players don’t have complete information about every prior move. It’s a powerful way to cut through complex strategic situations and identify which strategies are truly rational, and which are just empty threats.
Table of Contents
What makes a strategy “subgame perfect”?
Before diving into the solving process, let’s clarify what we’re actually looking for. A subgame-perfect Nash equilibrium is a special kind of Nash equilibrium that requires players to behave rationally not just at the start of the game, but at every possible point along the way-even at decision points that might never actually be reached if everyone plays optimally. Think of it as ensuring your strategy remains sensible throughout the entire game tree, regardless of what has happened before.
The key difference from a regular Nash equilibrium is that SPNE eliminates strategies that depend on non-credible threats. For instance, a player might threaten to do something that would actually hurt them more than it helps, hoping this threat will scare opponents into compliance. While such threats might seem reasonable when you look at the game as a whole, they fall apart under scrutiny because rational players know these threats won’t actually be carried out when the moment arrives.
The backward induction approach
The most common method for finding SPNE is backward induction, a technique that starts at the end of the game and works its way back to the beginning. Picture a tree where you start at the leaves and move toward the trunk. At each final decision point, you determine what a rational player would choose to maximize their payoff. Then, you replace that entire decision point with the expected outcome, effectively making the game shorter. You repeat this process, moving backward through the game tree, until you reach the very first decision.
Consider a simple business scenario: An established company (the incumbent) enjoys a monopoly, and a new company (the entrant) is deciding whether to enter the market. If the entrant stays out, the incumbent keeps its high monopoly profits. But if the entrant comes in, the incumbent can either “fight” by slashing prices-which hurts both companies-or “accommodate” by maintaining higher prices and sharing the market. Using backward induction, we look at the incumbent’s decision first. If the entrant does enter, fighting yields losses while accommodating yields some profit. Knowing the incumbent will rationally choose to accommodate, the entrant realizes entry will be profitable. The SPNE here is that the entrant enters, and the incumbent accommodates. The incumbent’s threat to fight is simply not credible, because following through would be irrational.
Working with proper subgames
When applying this method systematically, you need to identify all the proper subgames in your extensive-form game. A proper subgame is any part of the game that could stand alone as its own complete game, starting from a single decision point where all players know the full history up to that moment. For each proper subgame, you solve for its Nash equilibrium. Once you have that solution, you can treat it as a fixed outcome and substitute those payoffs back into the larger game.
Here’s where imperfect information complicates things. In games where players don’t have complete knowledge of previous moves-represented by information sets in the game tree-you can’t always apply pure backward induction. Why? Because backward induction requires cutting the game into subgames at specific decision points, but a proper subgame cannot “cut through” an information set where a player is uncertain about which exact point they’re at. If doing so would split an information set, then you don’t have a proper subgame there.
Handling imperfect information
Let’s say you’re analyzing a game where Player 1 makes a choice, and then Player 2 responds-but Player 2 doesn’t know what Player 1 chose. This creates an information set for Player 2 that spans multiple possible decision points. The entire section following Player 1’s initial move cannot be isolated as a proper subgame because Player 2’s information set ties those points together. In this case, the only proper subgame might be the entire game itself.
Does this mean SPNE becomes useless here? Not at all. Even when you can’t apply backward induction step-by-step through every node, you can still find the SPNE by solving for the Nash equilibrium within each identifiable proper subgame. You start with the subgames that don’t contain any other subgames nested within them, find their equilibria, and work your way outward. Each subgame’s equilibrium payoffs get plugged back into the larger game, gradually simplifying it until you’ve constructed a complete strategy profile that specifies rational play at every information set.
An example with simultaneous moves
Imagine a game where Player 1 first chooses to either “Enter” or “Exit.” If they exit, the game ends immediately. If they enter, both players simultaneously play a smaller two-by-two game-perhaps choosing between aggressive or cooperative strategies. Here, the simultaneous-move portion forms a proper subgame, since it begins after a known history (Player 1 entered) and all subsequent moves belong to that subgame.
To solve this, you first analyze the two-by-two simultaneous game and find its Nash equilibrium. Let’s say the equilibrium yields a payoff of (3, 2). Now, roll back to Player 1’s initial choice. Knowing that entering leads to (3, 2), Player 1 compares that outcome to the payoff from exiting. If exiting gives, say, (2, 6), Player 1 prefers to enter since 3 is greater than 2. The SPNE is: Player 1 enters, and then both players play the equilibrium strategies in the simultaneous subgame.
Eliminating non-credible equilibria
One of the most valuable aspects of SPNE is its ability to rule out Nash equilibria that rely on irrational behavior off the equilibrium path. Consider a game where Player 1 can choose Left or Right, and if they choose Right, Player 2 must respond. Suppose there’s a Nash equilibrium where Player 1 goes Left, and Player 2’s strategy says they would punish Player 1 harshly if Player 1 ever went Right-even though this punishment would hurt Player 2 more than Player 1.
On paper, this might be a Nash equilibrium because Player 1 never goes Right, so Player 2’s threat is never tested. But this equilibrium isn’t subgame perfect. Why? Because if we isolated the subgame starting at Player 2’s decision point (after Player 1 goes Right), Player 2’s threatened action is not a Nash equilibrium of that subgame. Player 2 would be better off choosing a different action. Since the strategy profile doesn’t represent optimal play in every subgame, it fails the SPNE test. Backward induction naturally eliminates such strategies by ensuring that at every stage, even hypothetical or off-path stages, players are responding optimally given what they believe will happen next.
The chain store paradox revisited
A classic illustration of this is the chain store game, where a chain store operates in multiple cities, and in each city, a local competitor decides whether to enter the market. The chain store can either fight aggressively (at a cost) or accommodate. If the chain fights, the entrant loses money. But here’s the catch: fighting is costly for the chain store too, so it’s not actually in their best interest to fight when entry happens. By backward induction, starting from the last city, the chain should always accommodate. Knowing this, the competitor in the last city should enter. The same reasoning applies to the second-to-last city, and so on, all the way back to the first city.
So theoretically, the threat to fight is non-credible-it’s an empty promise that collapses under rational scrutiny. The SPNE predicts that all competitors should enter, and the chain should accommodate in every city. However, in reality, the chain might build a reputation by fighting early on, which shows the limits of purely formal game theory in capturing strategic reputation effects. Still, for one-shot or finitely repeated games, SPNE successfully identifies which threats are genuine and which are bluffs.
Practical steps to find SPNE
If you’re faced with a sequential game and want to find its subgame-perfect equilibrium, here’s a streamlined process:
First, draw the game in extensive form, showing the sequence of moves, information sets, and payoffs at each terminal node. This visual representation makes it easier to identify where subgames begin and end.
Second, locate all proper subgames. Remember, a proper subgame must start at a single decision node where the history is fully known, and it must include all subsequent moves without cutting through any information sets.
Third, start with the subgames that are furthest from the initial node-the ones that don’t contain any other subgames within them. For each of these, solve for the Nash equilibrium. If there are multiple equilibria, you may end up with multiple SPNE for the overall game, one for each possible combination of subgame equilibria.
Fourth, replace each solved subgame with its equilibrium payoffs, effectively pruning that part of the tree. Now the game is shorter. Repeat the process, solving the next layer of subgames, until you’ve worked your way back to the starting point of the game.
Finally, compile the strategies that specify what each player does at every information set along this path. This complete strategy profile is your SPNE. To verify it, check that at every subgame-including the whole game-the strategy profile forms a Nash equilibrium.
When backward induction isn’t enough
While backward induction is powerful, it has limitations. It works beautifully for finite games with perfect information, where every player knows exactly where they are in the game tree at every decision point. But with imperfect information, information sets can prevent you from isolating certain portions as proper subgames. In such cases, you might find that the only proper subgame is the entire game itself, which means SPNE is equivalent to a regular Nash equilibrium for that game.
Moreover, some games have infinitely long horizons, or involve continuous actions rather than discrete choices. In these situations, backward induction isn’t directly applicable. Instead, game theorists use other techniques like the one-shot deviation principle, which checks whether a player could benefit by deviating at any single stage while keeping all other future actions fixed. This principle can be used to verify SPNE even when the game structure is more complex.
There’s also the human factor. Experimental studies show that people don’t always play according to SPNE predictions, especially in games like the Centipede game, where the backward induction solution prescribes immediate exit, yet players often continue for several rounds. This suggests that social preferences, fairness considerations, or bounded rationality can override the cold logic of formal game theory. Still, SPNE remains a crucial benchmark for understanding what fully rational play would look like.
What do you think? Can you imagine a situation in your own experience-perhaps in business negotiations or strategic planning-where identifying non-credible threats helped you see through someone’s bluff? How might understanding subgame perfection change the way you approach sequential decision-making?
Leave a Reply