Convergence rates for steady-state derivative estimators (Q1207840)

The author obtains the various convergence rates for various derivative estimators of steady state performance measures (derivative with respective to a continuous parameter) under a discrete-time Markov model as well as some appropriate mild assumptions. The estimators used include various schemes of finite difference (forward or central, with or not with common random numbers), infinitesimal perturbation analysis and likelihood. These estimators do not exploit the regenerative structure of the system explicitly. The results proved show that the convergence rates of some derivative estimators is poor than the canonical rate \((\sim n^{-1/2})\), which is due to that decreasing the bias implies a variance increase in these cases. Some numerical examples are also presented for illustrating these theoretical convergence rates obtained.