a Doob martingale
/ Doob rhymes with 'tube' /
Suppose there is a single, fixed random quantity Y you care about — say the final score of a game that has already been decided but not yet announced, or the total length of a random network — and information about it leaks out gradually. At each stage you form your best current estimate of Y using only what you have learned so far. As more is revealed your estimate gets refined, jumping around as surprises arrive. That sequence of running best-guesses is a Doob martingale: it is built not from adding fair shocks, but from conditioning ONE fixed variable on a growing flow of information.
Concretely, fix an integrable random variable Y and a filtration F_0 included in F_1 included in F_2 ... The Doob martingale is M_n = E[Y given F_n], your best estimate of Y given everything known by time n. Why is this automatically a martingale? Apply the tower property of conditional expectation: E[M_(n+1) given F_n] = E[ E[Y given F_(n+1)] given F_n] = E[Y given F_n] = M_n, since F_n is contained in F_(n+1) and conditioning on less information overrides conditioning on more. So the very act of estimating a fixed target through accumulating data is, free of charge, a fair game. Remarkably, the converse is nearly true: a martingale that is uniformly integrable is exactly a Doob martingale of its own limit, M_n = E[M_infinity given F_n] — closing the loop between 'revealing a fixed Y' and 'a fair game that settles down'.
Doob martingales are the workhorse of the concentration-of-measure method. To control a complicated function f(Z_1, ..., Z_m) of many independent inputs, set Y = f and let F_n reveal the inputs one at a time; the resulting Doob martingale has bounded increments whenever changing one input changes f by little, and the Azuma-Hoeffding inequality then proves f is sharply concentrated around its mean. The caveat to remember: a Doob martingale exists for any integrable Y, but it only converges back to Y (rather than to some smaller F-measurable shadow of it) when the filtration eventually contains all the information about Y.
Deal a shuffled deck face down and let Y be the number of red cards in the top half. Reveal cards one at a time and let M_n = E[Y given the first n cards seen]. Before any reveal, M_0 = E[Y]; each card nudges the estimate; once 26 cards are exposed, M_26 = Y exactly. The running estimate is a Doob martingale.
Estimating a fixed Y through accumulating information is, for free, a fair game.
A Doob martingale M_n = E[Y given F_n] exists for any integrable Y, but it converges back to Y only when the filtration eventually captures all the information about Y.