Lower bounds for function estimation (Q4355987)

In this expository article the author develops a general method for the obtention of minimax bounds for the risk function in nonparametric estimation. These bounds are given by the quantity \(\displaystyle\inf_{\hat f\in\hat F}\sup_{f\in{\mathcal F}}R(\hat f,f)\) where \({\mathcal F}\) is a class of functions, \(\hat F\) is a class of estimators and \(R\) is a risk function. She considers the case of nonparametric density estimation, carefully developing the problem when the density belongs to \({\mathcal F}=W_p^s(r)\) (Sobolev condition), a set of smooth densities such that \(\int | f^{(s)}(\lambda)| ^pd\lambda\leq r\). To develop the method, the function that measures the discrepancy between two densities must verify certain conditions. Examples of such discrepancies are: the p-th order discrepancy \(D(f,g)=\| f-g\|_p^p\), \(p\geq 1\), the p-th order discrepancy of the derivatives \(D(f,g)=\| f^{(s)}-g^{(s)}\|_p^p\), and finally \[ D(f,g)=\| f^{1/q}-g^{1/q}\|_p^p\quad \text{with}\quad 1/p+1/q=1. \] She then exposes a method to obtain a minimax bound for the risk function. It can be simplified as follows: (i) Looking for a basic inequality separating two points; (ii) an inequality in a finite-dimensional subset of \({\mathcal F}\) called the cube or piramid; (iii) obtention of the risk bound for the original problem using that the class \({\mathcal F}\) can be ''approximated'' by a sequence of finite-dimensional subsets. For the entire collection see [Zbl 0861.00032].