On M and P estimators that have breakdown point equal to 1/2 (Q1077106)

In this interesting paper the author considers M- and P-estimators and tries to establish the conditions on the function \(\rho\) generating them in order to obtain estimators such that the corresponding breakdown point \(\epsilon^*(.)=1/2\). For M-estimators \(\hat M\) the following theorem is proved: Theorem: Let \(\rho:E^ p\to E^ 1\), \(\rho (x)=\rho (-x)\) and \(\rho\) (x)\(\geq 0\) for all \(x\in E^ p\), \(\rho (0)=0\) and \(\rho (x)\to +\infty\) as \(| x| \to \infty\). Let \(\hat M\) be defined by \(\sum_{X}\rho (x-\hat M)=\min !\) for a data cloud \(X=\{x_ 1,...,x_ n\}\subset E^ p.\) (a) \(\epsilon^*(\hat M)=1/2\) if \(\rho (x+y)\leq \rho (x)+\rho (y)\) for all \(x,y\in E^ p.\) (b) If \(\epsilon^*(\hat M)=1/2\) then for all \(x,y\in E^ p\) there exists a finite constant g(y) such that \(\rho (x+y)-\rho (x)\leq g(y).\) Similar necessary and sufficient conditions are also proved for P- estimators. All these conditions are useful tools for checking whether \(\epsilon^*(.)=1/2\).