Exponential bounds for the probability of wrong determination of the order of a Markov chain by using the EDC criterion (Q2499105)

An \(r\)-th order Markov time series with finite number (\(N\)) of states is observed. The authors consider the ``efficient determination criterion''(EDC) for the estimation of the order \(r\) by observations: \(EDC(t)=-2\log L_t+\gamma(t)\alpha_n\), where \(L_t\) is the likelihood of \(r=t\), \(\gamma(t)\) is a penalty term and \(\alpha_n\) is a sequence of positive numbers. (AIC corresponds to \(\gamma(t)=N^t(N-1)\), \(\alpha_n=2\) and BIC to \(\alpha_n=2\log n\)). The estimate for \(r\) is \(\hat r=\arg\min_{0\leq r\leq R} EDC(t)\) The main result of the paper is that for some \(C_1\), \(C_2\), \[ \mathbf{P}\{\hat r_n\not=r\}\leq C_1\exp(-C_2\alpha_n). \]