Linear M-estimation with bounded variables

From MaRDI portal





The author considers the problem of finding the M-estimator of the overdetermined system of linear equations \(Jh=-f\), where \(f\in{\mathbb{R}}^m\), \(h\in{\mathbb{R}}^n\) and \(J\in{\mathbb{R}}^{m\times n}\), when there is a bound \(|h|_2\leq\delta\) on \(h\). With the residual \(l\) defined by \(l=f+Jh\), the solution \(h\) of this problem is found by solving \[ \text{minimize} \sum_{i=1}^m\rho\left(\frac{l_i(h)}{\sigma}\right)\;\text{subject to }|h|_2\leq\delta, \] where \(\sigma\) and \(\delta\) are real-valued positive constants and \(\rho:{\mathbb{R}}^1\to{\mathbb{R}}^1\) is a positive function with \(\rho(0)=0\). The author uses the method of Lagrange multipliers with a Lagrangian function of the form \[ g_\lambda(h)=\sum_{i=1}^m\rho\left(\frac{l_i(h)}{\sigma}\right)+\lambda\bigl(h^Th-\delta^2\bigr).\tag{1} \] Instead of letting both \(\lambda\) and \(h\) vary simultaneously [as did \textit{H. Ekblom} and \textit{K. Madsen} in BIT 29, No. 1, 60-76 (1989; Zbl 0677.65148)], the author keeps \(\lambda\) constant while minimizing (1). The proposed algorithm is presented in details, possible implementations are thoroughly discussed and its properties are described. Finally, results of testing on well-known test problems are presented.











This page was built for publication: Linear M-estimation with bounded variables

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q678208)