Distributing a computationally intensive estimator: the case of exact LMS regression (Q1965947)

Consider the usual linear regression model \(y_{i}=x'_{i}\beta+\varepsilon_{i}\), \(i=1,\dots,n\), where \(\beta\) is an unknown vector-valued parameter and \(\varepsilon_{i}\) are unknown errors. The least median of squares estimator \(\hat \beta_{LMS}\) is defined as \[ \underset\beta{\text{Arg min}} \underset{1\leq i\leq n}{\text{Median}} (y_{i}-x'_{i}\beta)^{2}. \] A set of Fortran subroutines for determining the exact \(\hat\beta_{LMS}\) is presented. Details are given for distributing the computations. Implementation of this algorithm for a network of computers is discussed. The proposed distribution techniques can be directly applied to any estimator whose calculation is based on repeated subsampling from the data.