the impossibility of beating a fair game
Generations of gamblers have invented 'systems' — double up after losses, quit while ahead, press your luck on a streak — convinced they can squeeze a profit from a fair game. The martingale framework delivers a clean, deflating verdict: against a truly fair game, no strategy that decides each bet using only past information can change your expected fortune. Whatever staking scheme you dream up, your expected winnings remain exactly zero. This is not a heuristic; it is a theorem, and it is one of the most quietly profound results in all of probability.
Here is the argument in plain steps. Model your fortune as a martingale M with fair increments DeltaM_k. Model your strategy as a predictable process H — your stake on round k, chosen before the round (an honest, non-clairvoyant rule). Your accumulated winnings are the martingale transform (H . M)_n = sum of H_k times DeltaM_k. The key theorem says a predictable transform of a martingale is again a martingale, so E[(H . M)_n] = E[(H . M)_0] = 0 for every n. In words: stake-weighting a fair game cannot tilt its expectation. Even if you cleverly stop at a (bounded) stopping time T, the optional stopping theorem gives E[(H . M)_T] = 0 too. So no betting rule, and no stopping rule, beats the house when the house has no edge.
Two honest qualifications keep this from being misunderstood. First, it bans changing the EXPECTED outcome, not the SHAPE of the distribution: the famous doubling system genuinely wins a small amount with very high probability — it just balances that against a rare, devastating loss, leaving the mean at zero (and in the real world, finite wealth and table limits make even that small edge vanish). Second, the result needs the game to be fair (a martingale) and your strategy to be predictable; against an unfavourable game (a supermartingale) you do strictly worse, and if you could see the future the theorem dissolves. The deep moral, echoed everywhere from finance to information theory: you cannot create expected gain out of pure fairness.
The doubling system on a fair coin: bet 1, double after each loss, stop at the first win. You almost always walk away +1 dollar, which feels like beating the game. But there is a small chance of a long losing streak that costs 1 + 2 + 4 + ... = a fortune, and the expectation works out to exactly 0. With real finite wealth, the rare ruin dominates, and the strategy is a losing one once a table limit caps your doubling.
A predictable stake on a martingale keeps expected winnings at zero — no system beats a fair game.
The theorem forbids changing the EXPECTED winnings, not the distribution's shape: doubling really does usually win a little, while risking a rare huge loss that holds the mean at zero.