Stochastic dynamics and Parrondo's paradox (Q2477690): Difference between revisions

The Spanish physicist Juan Parrondo has provided two stochastic losing games such that for certain stochastic combinations one may obtain a winning game. If a large number of players are involved and if they try to play such that their gain in the next round is maximized one arrives at the problem of investigating a random walk on a certain space of measures. The appropriate abstract setting is as follows. There is given a compact metric space \((M,d)\), and \(M\) is written as the union of certain closed subsets \(A_1,\dots, A_r\). For every \(\rho=1,\dots,r\) there is prescribed a strict contraction \(\Gamma_\rho: A_\rho\to M\). A random walk \((X_m)_{m\in \mathbb{N}_0}\) on \(M\) is then defined as follows. The starting position is \(X_0=x_0\), where \(x_0\in M\) is fixed, and if the walk at the \(m\)-th step is at position \(X_m\in M\), then one chooses a \(\rho\) among the \(\rho\) with \(X_m\in A_\rho\) (with equal probability, say) and defines \(X_{m+1}\) as \(\Gamma_{\rho}(X_m)\). Associated with the walk is a gain \(\varphi(X_m)\) in every round, where \(\varphi:M\to\mathbb{R}\) is a continuous function. The aim of the present investigations is the study of the expectation \(G_m\) of \(\varphi(X_m)\) as a function of \(m\). Our main result states that the sequence \((G_m)\) is ``eventually approximately periodic'' provided that all \(A_\rho\) are not only closed but also open in \(M\): for every \(\varepsilon\) there is an \(l_o\in\mathbb{N}\) such that \((G_m)\) is \(l_0\)-periodic up to an error of at most \(\varepsilon\) for sufficiently large \(m\). In fact it turns out that the behaviour of our process can be described well with a finite Markov chain. In the general case, however, the process might behave rather chaotically. We give an example where \(M\) is the unit interval. \(M\) is written as the union of two closed subsets \(A_1,A_2\), the contractions \(\Gamma_1,\gamma_2\) are rather simple, but the expectations of the gains are not even Cesáro convergent.