A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
A sequence {Mₙ} is a martingale if E[Mₙ₊₁ | ℱₙ] = Mₙ almost surely, where ℱₙ is the sigma-algebra of information up to time n. Martingales have zero expected change given current information—they are 'fair games.' The optional stopping theorem, martingale convergence theorem, and inequalities (Doob, Markov) are powerful tools for analyzing random processes.
A martingale formalizes the idea of a "fair game." Imagine a gambler whose fortune after n rounds is Mₙ. In a fair game, no matter what has happened so far, your best prediction for your fortune tomorrow is your fortune today: E[M_{n+1} | everything up to now] = Mₙ. This is precisely the martingale condition. Your prerequisite, conditional expectation, is exactly the tool that gives meaning to "expected value given current information." The filtration ℱₙ is just the mathematical object representing "everything knowable up to time n" — the sigma-algebra generated by M₁, M₂, …, Mₙ.
The simplest example is a symmetric random walk: a player wins or loses $1 on each fair coin flip. If Mₙ is the player's wealth, then E[M_{n+1} | ℱₙ] = (1/2)(Mₙ+1) + (1/2)(Mₙ-1) = Mₙ. The walk is a martingale. A supermartingale satisfies E[M_{n+1} | ℱₙ] ≤ Mₙ — the process tends to decrease (like a gambler at a casino with a house edge). A submartingale satisfies ≥ — the process tends to increase. Many important processes are martingales after appropriate centering: Sₙ - n·μ (a random walk minus its drift), or M²ₙ - n·σ² (the square of a centered walk minus a correcting term). Recognizing and constructing martingales from known processes is a core skill.
The connection to Markov chains (your soft prerequisite) is informative: every Markov chain generates martingales through harmonic functions. If h satisfies h(x) = Σ P(x,y)h(y) for all states x, then h(Xₙ) is a martingale. This bridges the two frameworks and lets you use martingale tools to analyze hitting times and absorption probabilities in Markov chains.
The power of the martingale framework lies in its theorems. The optional stopping theorem says E[M_T] = E[M₀] for a bounded stopping time T — you cannot gain an expected advantage by deciding when to stop a fair game, no matter how clever your stopping rule. The martingale convergence theorem says that a martingale bounded in L¹ converges almost surely to a limit — the fair game eventually settles. Doob's maximal inequality and Doob's Lp inequality provide moment and tail bounds on the supremum of a martingale, analogous to Markov's inequality but much sharper. Together, these make martingales the central tool in modern probability theory for proving convergence results, analyzing algorithms, and bounding stochastic processes.