Estimation for diffusion processes from discrete observation (Q1192000): Difference between revisions

The author treats the vector-valued stochastic differential equation \[ dX_ t = a(X_ t,\theta)dt + b(X_ t)\sigma dw_ t, \] where \(\theta\) and \(\sigma\) are unknown parameters and \(w(t)\) is an \(r\)- dimensional standard Wiener process. The observations from a realization consist of \(X_{t_ i}\), \(t_ i = ih\), \(i = 1, 2, \dots, N\). Using the likelihood function of the continuous-time process \[ \exp\{\int a'\overline{B} dX_ t - 2^{-1} \int a' \overline{B} a dt\},\quad \text{with}\quad \overline{B} = b\sigma(\sigma'b' b\sigma)^{-2} \sigma'b', \] the inverse being understood as a generalized inverse, the author gives an approximation for it in the form \[ \exp\{\sum a_{i - 1}' (\theta) \overline{B}_{i - 1} \overline{\Delta}_ i - 2^{-1} h\sum a_{i - 1}' \overline{B}_{i - 1} a_{i - 1}\}, \] where \(\overline {\Delta}_ i = X_{t_ i} - X_{t_{i - 1}}\). Maximizing this function with respect to \(\theta\) yields the approximate MLE. Sequentially estimating \(\theta\) and \(\sigma' \sigma\), the author proves the consistency and asymptotic efficiency of the estimators. The asymptotic normality of the joint distribution of the MLE is also proved. After preparing the notations and assumptions the main results and proofs are given. The Novikov moment inequality and the Burkholder-Davis-Gundy inequality are used (Novikov's paper is not cited). The AR(2) process is mentioned as an example.