the ground / stratum · the physical seam
Two gambling games. Each one, played alone forever, bleeds your money away at a steady rate. Mix them — alternate, or just flip a coin each round to choose which to play — and the same money grows. Two certain losers average into a winner. This is Parrondo's paradox, and nothing on this page is asserted: you can race all three, and take apart exactly where the winnings come from.
research/parrondo/verify.mjs (17/17).
Here is something that sounds like a swindle, or a perpetual-motion machine. I hand you two games of chance. Game A is a single slightly-unfair coin: you win a dollar a touch less than half the time, so play it and you slide downhill — about a penny lost per round. Game B is stranger, a game that reads your current fortune, but it too is rigged to lose, a little under a penny a round. Two games, both downhill. Sit at either table all night and you go home poorer.
Now do the one thing that should not matter. Each round, before you play, flip a fair coin to decide which of the two losing games to sit at. That's it — no new game, no skill, no information about the future; just a random choice between two known losers. Your fortune turns around and climbs, about a cent and a half won per round. The two falling lines, averaged, rise. Below you can run it — thousands of rounds, live — and then we'll find exactly where the money is hiding.
Three players start with nothing and play ten thousand rounds. One plays only Game A, one only Game B, one flips a fair coin each round to pick A or B. Watch their fortunes. The running drift below each line is its average gain per round so far; it settles onto the exact value the page computed before any coin was thrown.
The two losers really do lose — let them run and both red-tinted lines wander down. The green line, built from nothing but a coin-flip choice between those same two losers, climbs. The drifts converge to numbers that were fixed in advance: −0.0100, −0.0087, and +0.0157. So the winnings are real. Where do they come from?
The secret is all in Game B, which is not one coin but two, and chooses between them by looking at your fortune. The exact rule is the classic one (Harmer & Abbott, 1999), built from a tiny bias ε = 0.005:
Glance at that and Game B looks like a winner: a great coin two-thirds of the time, a terrible coin only a third of the time. Average them as if your fortune visited each remainder (0, 1, 2 mod 3) equally often, and the arithmetic says you should gain +0.057 a round. That number is a trap. Your fortune does not visit the three remainders equally — the game's own dynamics herd it onto the bad one. The instrument below solves the long-run fraction of time spent at each remainder exactly (the stationary distribution of a 3-state Markov chain, computed in BigInt fractions in your browser), and shows the honest drift it implies.
There it is. Left to itself, Game B parks your fortune on a multiple of 3 more than a third of the time — the dashed line is the even 1/3 you'd naively assume. The reason is a little engine: the bad coin at ≡ 0 almost always knocks you down to ≡ 2; from ≡ 2 the good coin almost always carries you back up to ≡ 0. You get caught bouncing across the line into the bad state, and that over-visited bad coin is just heavy enough to sink the whole game: a true drift of −0.0087, not the +0.057 the even-split daydream promised.
So why does adding Game A — itself a loser — help? Because Game A doesn't care about your fortune at all. Slipping it in at random stirs the distribution: it breaks up the tidy bounce that kept you pinned on the bad state, loosening the grip on ≡ 0 (from 0.384 down toward 0.345). With your fortune spread more evenly, the good coin finally gets played as often as it should, and the mixture tips into profit. Game A pays a small, certain toll to spring the trap that was costing Game B much more. Two losses, but they are losses of different kinds, and one cancels the other's cause.
The random choice was never the point — any schedule that keeps stirring works. Set a fixed, repeating pattern of A's and B's and the page solves its exact long-run drift (a larger Markov chain over your remainder and your place in the pattern). Some win, some lose; the schedule itself is the dial.
Plain AB alternation actually loses here (−0.0067) — it stirs the wrong way. But AABB wins (+0.0147), with no randomness anywhere; ABBAB wins handsomely (+0.0652). Whether two losers make a winner depends on how you interleave them — proof that the gain is a property of the switching, not of luck, and not of either game alone.
It is not magic and not universal. Everything hangs on the bias ε. Turn it to zero and each game alone is exactly fair — Game A and Game B both drift nothing — and yet, remarkably, the mixture still wins (+0.0254 a round at ε = 0): the paradox is sharpest precisely where the parts are fairest. Turn ε too high and Game A's bleed overwhelms the stirring. Between is a window where both games lose yet the mixture wins, and the dial computes its exact edges.
Scan it: at ε = 0 the two games are each fair (drift exactly 0), but the mixture is already a winner — so the window's lower edge isn't some small positive number, it is ε = 0 itself, open: the instant ε clears zero, both games tip into loss while the mixture is still winning. Nudge ε up and that band holds — Game A loses, Game B loses, the mixture wins — until the mixture's win finally runs out at a single exact value, ε* ≈ 0.0131093 (the lone zero of the mix's drift, bisected over exact fractions; Game B is still losing there). So the true window is ε ∈ (0, 0.0131093) — not the ragged [0.001, 0.013] a coarse 0.001-grid scan reports, but the exact interval, and the classic ε = 0.005 sits well inside it. Past ε* the mixture loses too. The paradox is real but bounded, and the dial shows you its edges exactly instead of estimating them.
This is not a money machine, and it does not break any law. The honest control is quick: keep Game B's two coins but choose between them with an independent fair flip instead of by your fortune — capital-blind. Now there is no trap to spring, the games are uncorrelated, and the paradox vanishes: the mix just loses, the way averaging two losers should. verified — control mix drifts −0.085 The whole effect lived in the correlation between the games through the one thing they share, your running capital. Take that link away and "something from nothing" becomes nothing from nothing.
So where does the gain come from? From switching as a resource. Game B builds an asymmetry — a ratchet — that, left alone, runs the wrong way for you. Game A carries no information and no winnings of its own, but injecting it desynchronises B from its own trap, and the ratchet's rectifying asymmetry is freed to push the other way. The energy, so to speak, was always in B's structure; A is the flick that lets it out.
That picture is not a metaphor bolted on afterward — it is the paradox's origin. Juan Parrondo built these games in 1996 as a discrete, countable cartoon of the flashing Brownian ratchet: a particle in a sawtooth potential that is switched on and off. With the potential on, the particle has no net force; with it off, it just diffuses; yet alternating the two — flashing — pumps the particle steadily in one direction, because the sawtooth's broken left–right symmetry rectifies the otherwise-aimless diffusion. Two dynamics that each go nowhere, alternated, go somewhere. Parrondo's games are that physics with the calculus removed: the capital plays the particle, the modulo-3 rule plays the sawtooth's asymmetry, and the switch between A and B plays the flashing.1 The Second Law is never threatened — a flashing ratchet needs an external agent to do the flashing, and that switching is exactly the thermodynamic price. Here, you are that agent, and the coin you flip to choose the game is the work you put in.
The clean moral, the one worth carrying out of the toy: a property of an average need not be a property of the things averaged, once those things are coupled. Each game's losing-ness is real, but it is a fact about that game running undisturbed, free to settle into its own trap. Disturb it — interleave, stir, switch — and the trap never sets, and the sign can flip. It is the same shape as Simpson's paradox in statistics and the same shape as why a portfolio of volatile losers can be rebalanced into a gainer: the combination has access to a structure none of the parts, alone, could use.
These are exactly-solvable toy games — finite, noiseless, with the generative process known in full. That cleanness is what lets the paradox prove itself so sharply, and it is also what real money games lack: a casino's edge does not hide a springable ratchet, and no schedule of two house-edge bets beats the house. The result is not investment advice and not a gambling system; it is a precise demonstration that combination can reverse a sign when the parts are correlated — a fact with real teeth in stochastic control, molecular motors, and population dynamics, and none at the roulette table. The specific numbers are tied to these specific coins; what transfers is the mechanism, not the payout.
G. P. Harmer & D. Abbott, "Losing strategies can win by Parrondo's paradox," Nature 402 (1999), 864 — the games and the classic parameters used here.
G. P. Harmer & D. Abbott, "Parrondo's paradox," Statistical Science 14 (1999), 206–213 — the full Markov-chain analysis this page reproduces.
J. M. R. Parrondo, G. P. Harmer & D. Abbott, "New paradoxical games based on Brownian ratchets," Physical Review Letters 85 (2000), 5226.
A. Ajdari & J. Prost, "Drift induced by a spatially periodic potential of low symmetry: pulsed dielectrophoresis," C. R. Acad. Sci. Paris II 315 (1992), 1635 — the flashing ratchet the games discretise.
R. D. Astumian, "Thermodynamics and kinetics of a Brownian motor," Science 276 (1997), 917 — why the ratchet pays its thermodynamic toll.
Every figure here is recomputed in your browser by the same exact-rational engine checked offline in /research/parrondo/verify.mjs (run it: node research/parrondo/verify.mjs → 26 / 26 checks pass). The stationary distributions are solved over BigInt fractions; the drifts are cross-checked against float power-iteration and a 4-million-step Monte-Carlo walk; the ε = 0 mix-drift is pinned to the exact fraction 18/709, and the paradox window's upper edge ε* ≈ 0.0131093 is certified by a rational sign-change bracket (mix > 0 at 0.0131092, mix < 0 at 0.0131093) and pinned by bisection — the same bisection the dial runs live.