The Parrondo's paradox calculator above computes the expected gain per play for two gambling games that each lose money on their own, and for the mixture that alternates between them at random — which makes money. The effect is real and it is not a trick of the arithmetic, but it only makes sense once the exact rules are on the table, so this page states them before it computes anything.
Arb Digital publishes it in the same spirit as our other paradox tools. Our boy or girl paradox calculator and Monty Hall calculator deal with puzzles where the wording hides an assumption. This one is different: nothing is hidden and there is no ambiguity. The result is simply surprising, and the surprise dissolves entirely once you see what game A is doing to the state that game B depends on.
The Exact Rules Of The Two Games
This section is the whole page. Without these rules the effect reads as magic, and with them it reads as a straightforward Markov chain calculation.
Game A is one biased coin. You win one unit with probability 0.5 − ε and lose one unit otherwise. With ε positive this is a losing game, and it stays losing however long you play it. Nothing about it depends on how much capital you hold.
Game B uses two coins and chooses between them according to your current capital. If your capital is divisible by M — three, in the classic construction — you must play the bad coin, which wins with probability 0.1 − ε. If your capital is not divisible by M, you play the good coin, which wins with probability 0.75 − ε. Played alone, game B is also a losing game, and that is the part which looks wrong at first: the good coin is very good, the bad coin is only reached one time in three by a naive count, and yet the game loses.
The mixture chooses A or B independently at each turn, each with probability one half in the classic case. Every play is drawn from a losing game, and the combined process wins. Both games use the same capital counter, which is the channel through which they interact.
Why The Mixture Wins
Game B is not a sequence of independent coin flips. Which coin you play depends on your capital, your capital depends on the previous outcomes, and the whole thing is a Markov chain on the residue of your capital modulo M.
The bad coin loses most of the time, and a loss takes your capital from a multiple of M down to one below it. The good coin then usually wins, taking you back up to a multiple of M and straight into the bad coin again. With the classic coins and no bias, the chain settles at five thirteenths of its time in the bad state rather than the one third an even split would give, so the good coin is played less often than it looks. That over-visiting is what tips game B into losing territory despite how favourable its second coin is.
Game A does nothing but move the capital up or down at nearly even odds. Mixed into game B, it randomises which residue class you land in and breaks the trap. The stationary distribution flattens, the bad coin is played closer to one time in M, and the good coin's advantage finally shows up in the average. The mixture is not combining two payoffs; it is changing the state distribution that determines which payoff you face.
The Formula and How It Is Calculated
Each game is a Markov chain on the M residue classes of your capital. From state i you move to state i + 1 with the win probability that applies in state i, and to state i − 1 otherwise, both taken modulo M.
The tool builds that transition matrix, finds its stationary distribution π — the long-run share of time spent in each residue class — and computes the expected gain per play as the sum over states of πᵢ multiplied by (2pᵢ − 1). The factor 2p − 1 is the expected change in capital for a single play at win probability p, since a win adds one and a loss subtracts one.
For the mixture, the win probability in state i is simply the weighted average of game A's probability and game B's state-i probability, weighted by how often each game is chosen. That averaged chain has its own stationary distribution, and it is the difference between that distribution and game B's that produces the whole effect. Wolfram MathWorld's Parrondo's paradox page describes the phenomenon as two losing gambling games becoming winning when played one after the other, and notes that the simplest construction uses three biased coins.
What The Paradox Is Not
Three misreadings are common enough to be worth naming.
It is not a way to beat a casino. Casino games are independent of your capital by design, and the effect requires a game whose rules depend on the state. Where the games are independent, mixing them just averages their expected values, and an average of two negative numbers is negative. No sequence, ordering or staking pattern changes that.
It is not the gambler's fallacy in disguise. The gambler's fallacy is a false belief about independent events. Here the events are genuinely not independent — the dependence is written into game B's rules — and the calculation is a correct treatment of a real dependence rather than an incorrect treatment of an imagined one. Our coin flip probability calculator covers the genuinely independent case, where none of this applies.
And it is not a violation of linearity of expectation. Expectation is still linear. What is not linear is the map from the mixing weight to the stationary distribution: changing which game you play changes where the chain spends its time, and the payoff is evaluated against that shifted distribution. The two chains being averaged are not the chain you get by averaging them.
Where The Effect Turns Up Outside Gambling
The structure — two individually unfavourable dynamics combining into a favourable one through a shared state — appears well beyond coin games, and the coin version is best understood as the simplest model of it.
Parrondo derived the games from Brownian ratchets, physical systems where alternating between two potentials that each dissipate energy produces directed motion. The same alternation logic appears in models of population dynamics where two environments that each drive a population to extinction can sustain it when they alternate, and in switching-strategy models in evolutionary biology.
The physics-facing literature treats the games as disordered systems. J. M. Luck's paper "Parrondo games as disordered systems", published in the European Physical Journal B in 2019, maps both the capital-dependent games modelled here and the history-dependent variants onto one-dimensional random walks and analyses them with transfer-matrix methods, emphasising that the gains depend non-linearly on the parameters — which is precisely what makes the paradox possible.
Exploring The Parameter Space
The effect is not fragile, but it is not universal either, and the tool is most useful for finding its edges.
Raise ε and both games lose harder; past a threshold the mixture cannot overcome the drag and it loses too. Move game B's coins closer together and the state dependence weakens until game B is nearly a plain coin, at which point mixing achieves nothing. Change the modulus and the trap changes shape: larger M means the bad state is reached less often by chance, which alters both the size and sometimes the sign of the effect.
Set ε to zero and the cleanest version of the result appears. Game A becomes an exactly fair coin and game B becomes exactly fair too, each gaining nothing per play, while the classic mixture gains about 0.0254 units per play. That figure is the pure size of the mechanism, with no bias for it to overcome, and it is why the paradox is a statement about the state distribution rather than about the coins. Raising ε then drags all three figures down together, and the paradox proper — two losing games, one winning mixture — holds until the drag exceeds that 0.0254. Our expected value calculator handles the single-play arithmetic, and our probability calculator builds the event probabilities that feed it.
Arb Digital measures how channels interact rather than scoring each in isolation, because the interaction is where the number you care about actually lives.
Browse All Free Tools Talk To Our TeamCommon Mistakes to Avoid
- Dropping the capital-dependence rule — without it game B is a single coin, and mixing two independent losing coins always loses.
- Assuming the bad coin is played one time in M — the chain over-visits that state, and that over-visiting is the entire mechanism.
- Treating it as a betting system — casino games do not condition on your capital, so no mixture of them produces this effect.
- Expecting the effect at any parameters — a large bias, coins that are close together, or an extreme mixing weight all destroy it.
- Reading it as a failure of expectation — expectation is linear; what changes under mixing is the stationary distribution the payoff is evaluated against.
Related Free Tools From Arb Digital
Work single-play payoffs with the expected value calculator, build event probabilities with the probability calculator, check the independent-coin case with the coin flip probability calculator, compare a puzzle whose answer depends on its wording with the Monty Hall calculator, or see a paradox with three defensible answers in the Bertrand's paradox calculator. The free online tools hub lists every probability tool we publish.
Frequently Asked Questions
The result that two gambling games which each lose money on their own can produce a winning expectation when alternated, provided at least one of them depends on the player's current capital.
Game A wins one unit with probability 0.5 minus epsilon. Game B plays a coin winning with probability 0.1 minus epsilon when capital is divisible by three, and a coin winning with probability 0.75 minus epsilon otherwise.
Because the bad coin usually loses, which drops the capital to one below a multiple of three, and the good coin then usually wins and returns it there. The chain spends more time in the bad state than an even split would give.
Game A moves the capital at nearly even odds regardless of its value, which randomises the residue class and breaks the trap. The bad coin is then played closer to one time in three and the good coin's advantage shows up.
No. Casino games are deliberately independent of your capital, and mixing independent losing games only averages their expected values, which stays negative.
No. A large bias, coins that are close in value, or an extreme mixing weight all destroy the effect. The tool is most useful for finding where those edges lie.
No. Expectation stays linear. What is non-linear is the map from the mixing weight to the chain's stationary distribution, and the payoff is evaluated against that shifted distribution.
This page explains a probability result for educational purposes. It is not a betting system and produces no advantage in any game whose odds do not depend on the player's current capital.
