Estimation in the presence of infinitely many nuisance parameters - Geometry of estimating functions (Q1116223)

The present paper treats the problem of estimating a structural parameter \(\theta\) when there exist nuisance parameters whose number increases proportional to the number n of independent observations \(x_ 1,...,x_ n\). Consider the class \(C_ 1\) of all consistent estimators which are solutions of the equation \(\sum^{n}_{j=1}y(x_{ij}\theta)=0\) where y(x;\(\theta)\) is called estimating function. The aim of this paper is to study the structure and the asymptotic behavior of all \(C_ 1\)-estimators or estimating functions, providing a geometric theory of estimating functions. The authors solve the following three problems for \(C_ 1\)-estimators by constructing a differential geometric theory: (i) To give a necessary and sufficient condition for the existence of \(C_ 1\)-estimators, (ii) to elucidate structures of all \(C_ 1\)-estimators and (iii) to give a necessary and sufficient condition for the existence of the optimal \(C_ 1\)-estimator and to obtain it when it exists. The theory is applied to statistical models in which there exists a sufficient statistic for the nuisance parameter when \(\theta\) is fixed, where the abstract results of the paper become very transparent. The present theory also explains why the conditional score plays an important role.