Deleting dominated moves is really about best responses. A move you would never choose against any belief about the other player is a move worth throwing out.
Rationalizability and best response
Your best response is the move that pays most given what you expect the other to do. A strategy is rationalizable if it survives repeatedly deleting every strategy that is never a best response to a belief supported on the other player’s surviving strategies.
A strictly dominated strategy should be...
Predict, then play
Mark the best responses. For each column the other player might pick, reveal shows which of your rows pays most (the accent dot). Watch which of your three moves never earns a dot.
Middle is not beaten by any single one of your other moves (against Right it beats Top, and against Left it beats Bottom). But a fifty-fifty mix of Top and Bottom beats it in every column, so Middle is never a best reply to any belief. That is what rationalizability throws out.
Best responses, all the way down
A belief assigns probabilities to the other player’s possible strategies. For each such belief, some move of yours pays most in expectation. That is a best response. A move that is a best response to no belief gets deleted. Once it is gone, beliefs can put weight only on strategies still supported by rational play, so other moves may lose their justification too. Then you repeat.
Take a guess, then reveal the reason
Is rationalizable the same as Nash? No. Every Nash strategy is rationalizable, but the rationalizable set is usually larger. Rationalizability asks only that your belief be consistent with the others being rational; Nash adds that everyone’s beliefs are actually correct.
Work it out
Show that Middle is never a best reply
First check the easy test and watch it fail. Middle is not beaten by Top alone: against Right, Middle pays and Top pays . It is not beaten by Bottom alone either: against Left, Middle pays and Bottom pays . So no single row of yours dominates it, and a domination test that looks only at pure strategies keeps Middle.
Now allow a mix. Play Top and Bottom with probability each. Against Left it pays ; against Centre, ; against Right, . Every one of those beats Middle’s flat , so the mix strictly dominates Middle and no belief about the other player can make Middle a best reply. Delete it.
Faded. Suppose Middle paid in all three columns instead of . The same fifty-fifty mix now pays , , : it no longer beats Middle everywhere. Look for a belief that makes Middle best: put probability on Left and on Right, and Top pays , Bottom pays , Middle pays . Middle survives, because a strategy that is mediocre everywhere is the right hedge against a genuinely uncertain opponent.
Solo. Back to the original numbers, and now run the same test on the other player. Their Centre column pays them whatever you do, exactly mirroring your Middle. A fifty-fifty mix of Left and Right pays them against Top and against Bottom, so Centre is never their best reply either and it goes too. What survives on both sides is Top and Bottom against Left and Right: the rationalizable set of this game.
Key idea
Throw out any move that is never your best reply to anything. What survives is the set of choices rational players could actually justify.
Check yourself
Sources
- Bernheim, B. D. (1984). Rationalizable strategic behavior. Econometrica, 52(4), 1007-1028.
- Pearce, D. G. (1984). Rationalizable strategic behavior and the problem of perfection. Econometrica, 52(4), 1029-1050.