Bounds for the normal approximation of the maximum likelihood estimator (Q502863)

Let \(\mathbf{X}=\left( X_{1},X_{2},\ldots,X_{n} \right) \) be i.i.d. random variables with density or frequency function \( f\left( x_{i}\, |\, \theta\right) \), where \(\theta\) denotes an unknown scalar parameter found in a parametric statistical model. Assume that the maximum likelihood estimator (MLE) \(\hat{\theta}_{n}\left(\mathbf{X} \right) \) exists and is unique. Let \(\theta_{0}\) be the true value (still unknown) of the parameter \(\theta\). It is well known that under some regularity conditions, as \(n\to \infty\) \[ \sqrt{ni\left( \theta_{0}\right)} \left(\hat{\theta}_{n}\left( \mathbf{X}\right) - \theta_{0} \right)\to Z \] in distribution, where \(Z \sim N(0,1)\) and \(i\left(\theta \right) \) is the expected Fisher information of a random variable. Using Stein's method, the authors derive explicit bounds for \[ d_{bW}\left( \sqrt{ni\left( \theta_{0}\right)} \left(\hat{\theta}_{n} \left( \mathbf{X}\right) - \theta_{0} \right),Z \right), \] where \(d_{bW}\) is the bounded Wasserstein distance between random variables. The results are applied to the class of one-parameter exponential distributions. The authors also use a perturbation method to treat cases where the MLE has positive probability of being on the boundary of the parameter space.