the tail sigma-algebra and Kolmogorov's zero-one law
/ kol-muh-GOR-off /
Some questions about an infinite sequence of random variables do not depend on any finite number of them — only on their long-run, far-out behavior. Does the series X1 + X2 + X3 + ... converge? Does the running average settle to a limit? Do infinitely many of the variables exceed 100? You can change the first million terms entirely without affecting any of these. Such questions are tail events, and Kolmogorov's zero-one law makes a startling promise about them: if the variables are independent, every tail event has probability either 0 or 1 — never anything in between.
The tail sigma-algebra is the precise home of these questions. For a sequence X1, X2, X3, ..., form the sigma-algebra generated by all the variables from index n onward, then take the intersection of these over all n. An event survives this intersection exactly when it is unaffected by dropping any finite prefix — that is the formal meaning of depending only on the tail. The zero-one law then states: if the Xn are independent, any event in the tail sigma-algebra has P equal to 0 or 1. The intuition behind the proof is a clean paradox: a tail event is determined by the far-future variables, yet it is also independent of every finite block of early variables; pushing this all the way, the event ends up independent of itself, and the only numbers p with p = p^2 are 0 and 1.
The consequence is sharp and useful: for independent sequences, any genuinely tail-dependent quantity is deterministic in its yes/no verdict, even though which way it falls may take real work to discover. Whether a random series converges, whether a random walk returns to the origin infinitely often, whether a long-run limit exists — each is a sure thing or an impossibility, never a coin flip. The law tells you the answer is 0 or 1; it does not tell you which, and pinning that down is a separate task (often via the Borel-Cantelli lemmas).
For independent random variables X1, X2, X3, ..., the event the series sum of Xn converges is a tail event: changing any finite number of terms cannot alter convergence. By the zero-one law its probability is 0 or 1 — the series converges almost surely, or it diverges almost surely, with no middle ground.
For independent sequences, anything depending only on the far tail is settled with probability 0 or 1.
Independence is essential, and the law only certifies the probability is 0 or 1 — it never tells you which. Determining the actual verdict (e.g. via Borel-Cantelli) is a separate problem.