Parrondo's paradox explained: how two losing games can produce positive drift

Write the games before interpreting the paradox

Game A in Modic wins with probability 0.495. Game B wins with probability 0.095 when capital is divisible by three and 0.745 otherwise. Each win adds one unit and each loss subtracts one. Capital is allowed to be negative, and divisibility is handled consistently for negative integers too. Those details define a mathematical process rather than a finite-bankroll betting recommendation.

The primary Nature report on losing strategies introduced this counterintuitive combination to a broad audience. The interesting claim is conditional on specified games and switching. It does not say any two strategies with negative returns can be averaged into a positive one. The mechanism requires more than labelling each standalone result a loss.

Game B's average odds depend on where it spends time

Capital modulo three gives three states: remainder zero, one or two. Game B's low probability applies only in the zero state. You cannot average its two listed probabilities as if each occurred equally often. The frequency of visits is generated by the game's own transitions. That state distribution is part of the calculation.

The Parrondo experiment shows both a realised path and a stationary expected drift. Its three-state calculation repeatedly updates the state probabilities under the chosen policy. The random mixture changes transition probabilities and therefore changes state occupancy. That is why its long-run drift is not simply the average of the standalone games' already-averaged drifts.

Worked example: inspect the mixed odds at each state

Start with the transition-level arithmetic before looking at a simulated final balance. This calculation does not prove the stationary result by itself; it identifies the state-dependent probabilities that the Markov calculation must use. A single path will still fluctuate around the process's longer-run behaviour.

1. At a capital divisible by three, a fifty-fifty choice of A and B has win probability (0.495+0.095)/2=0.295, making this state unfavourable.
2. At the other two remainders, the same mixture wins with probability (0.495+0.745)/2=0.620. These states have favourable one-step drift.
3. Compare the simulator's stationary drift for A, B and the mixture. The mixture's transition rule changes time spent in the three states, allowing positive average drift under these exact parameters.

One winning path is different from positive expected drift

Run each policy and inspect the three-hundred-round path. A negative-drift policy can finish above zero by chance, and a positive-drift policy can finish below it. The hundred-path mean reduces some sampling variability but is still an estimate from a finite batch. The stationary calculation is a different object: a property of the specified transition model.

Harmer and Abbott's fuller treatment develops the games' mathematical structure. Modic keeps the state space small enough to expose the rule. It does not fit a real time series or estimate whether a practical strategy has positive drift. Agreement between a batch mean and the reference is a useful check, not empirical evidence about an outside market.

Connect the paradox with other dependent processes

The random-walk experiment is a useful starting contrast. A walk with fixed odds does not need to track a state-dependent change in its next-step win probability. Parrondo's capital-dependent game adds exactly that feature. The current state influences what happens next, so transition structure matters as well as an overall summary rate.

The Pólya-urn model provides another dependence example, though with a different mechanism. Drawing a colour changes future composition through reinforcement. It should not be called the same paradox; the useful connection is that a sequence's history can change future probabilities. When mixing processes, ask what state one leaves for the next, rather than assuming their standalone averages contain all the necessary information.

How to audit a surprising combination claim

A surprising result becomes easier to assess when the process is written as states, actions and transitions. Make the standalone policies and the mixture comparable under the same starting conditions, and specify which quantity is supposed to improve. A visually appealing path is not an adequate substitute for those definitions.

• State whether odds depend on capital, history or another hidden state.
• Write the switching policy explicitly.
• Compare one path, a batch mean and a model expectation separately.
• Check boundary conditions such as stopping at zero capital.
• Identify which state distribution changes under the mixture.
• Avoid extending the result to unrelated losing strategies.

Why this is not a casino or investment system

The model permits negative capital and has no absorbing ruin threshold. Adding a finite bankroll changes what can happen and when play stops. Real games and investments also involve costs, constraints and uncertain or changing probabilities. A positive drift derived from a fictional transition table cannot be transferred to them by analogy.

The valuable idea is structural: combining processes can change the distribution of conditions under which each operates. That possibility is worth investigating in a concrete model, while the burden remains to identify the transitions and evidence. The paradox does not suspend arithmetic. It warns against averaging summaries that were calculated under different state distributions and assuming the combination preserves those distributions.