convergence in probability

Picture a measurement that gets better and better as you collect more data — the running average of a long survey, say. You would like to say it 'converges' to the truth, but each individual run is still random, so you cannot promise it lands exactly on target. Convergence in probability is the honest middle ground: it says the chance of being noticeably wrong shrinks to zero as n grows.

Precisely: X_n converges in probability to X if for every tolerance epsilon > 0, P(|X_n - X| > epsilon) approaches 0 as n approaches infinity. Read it slowly. Pick any error bar you like, however small. As n grows, the probability that X_n misses X by more than that error bar goes to zero. It does NOT say X_n eventually stays within the bar forever — for any fixed large n there is still a small chance of a big miss; it just says that chance is tiny. The misses become rare, but they never have to stop entirely.

This is the mode in which the weak law of large numbers speaks: the sample mean converges in probability to the true mean. It is stronger than convergence in distribution (it implies it) but weaker than almost sure convergence. The standard subtle point: a sequence can converge in probability while, for almost every outcome, the actual path X_1, X_2, ... keeps jumping and never settles — the typewriter sequence is the classic example. Probabilistic closeness at each fixed n is a weaker promise than the whole trajectory settling down.