The evolution of aggregated Markov chains (Q2573253): Difference between revisions

Let \((X_n)_{n\in \mathbb{Z}}\) be a stationary, two-sided Markov chain with a finite state space \(I\). Let \(I=\bigcup_{\nu=0}^{s-1}I_\nu\) be a partition of \(I\), \(s\geq 2\). Then we can define the sequence of random variables \((Y_n)_{n\in {\mathbb{Z}}}\) by setting \(Y_n=\nu\) when \(X_n\in I_{\nu}\). The sequence \((Y_n)_{n\in\mathbb{Z}}\) is stationary, but in general is not a Markov chain. The paper is devoted to the computation of the transition probabilities: \[ \mathbf{P}(Y_n=\nu\mid Y_{n-l}=\nu_{n-l},\dots,Y_{n-1}=\nu_{n-1})\tag{1} \] (transition probabilities given a finite past) and \[ \mathbf{P}(Y_n=\nu\mid Y_m=\nu_m \text{ for all } m\leq n-1)\tag{2} \] (transition probabilities given an infinite past). The author shows that the limit of (1) exists and is equal to (2). Then he introduces the \([0,1]\)-valued process \(U_n=\sum_{m=1}^\infty s^{-m}Y_{n-m}\) called the moving average process. He gives explicit formulas for (1) and (2) in terms of the distribution probabilities of \(U_n\). More precisely, he proves that \[ \mathbf{P}(Y_n=\nu\mid Y_{n-m}=\nu_m \quad\text{for}\quad m=1,\dotsc,l)= \frac{\sum_{i\in I_\nu}\pi_i(F_i(b_l)-F_i(a_l))}{F(b_l)-F(a_l)} \] where \(a_l=\sum_{m=1}^l s^{-m}\nu_m\), \(b_l=\sum_{m=1}^l s^{-m}\nu_m+s^{-l}\), \(\pi=(\pi_i)_{i\in I}\) is the stationary distribution of \(X_n\), \(F\) is the distribution function of \(U_n\) and \(F_i(x)=\mathbf{P}(U_n\leq x\mid X_n=i)\), for \(x\in\mathbb{R}\). Finally, he introduces the Radon-Nikodym derivative \(g_\nu(x)\) of the measure corresponding to \(\sum_{i\in I_\nu}\pi_iF_i\) with respect to the measure corresponding to \(F\) and he shows that \[ \mathbf{P}(Y_n=\nu\mid Y_m=\nu_m\;\text{for all}\;m<m)=g_\nu(x) \] where \(x=\sum_{m=1}^\infty s^{-m}\nu_m\).