Most people know it as the observation that plurality elections in single-seat districts tend to come with just two parties. Except that the correlation is weak. As for causation, see Riker (1982):
The direction one must go, I believe, is to turn attention away from the expected utility calculus of the individual voter and to the expected utility calculus of the politician and other more substantial participants in the system. The groups and individuals who buy access and the politicians who buy a future have substantial interests, and it is their actions to maximize expected utility that have the effect of maintaining the two-party system under plurality voting.
So the answer to the question of failure is that third parties are rejected in the rational calculus of expected utility especially by leaders, though also in the calculus by many simple voters. Any adequate theory to subsume Duverger’s law must, I believe, begin there, which is a task for scholars in the next decade.