On Solving Lq-Penalized Regressions

T. Z. Wu,Yingyi Chu,Yan Yu
DOI: https://doi.org/10.1155/2007/24053
2007-10-30
Advances in Decision Sciences
Abstract:L q -penalized regression arises in multidimensional statistical modelling where all or part of the regression coefficients are penalized to achieve both accuracy and parsimony of statistical models. There is often substantial computational difficulty except for the quadratic penalty case. The difficulty is partly due to the nonsmoothness of the objective function inherited from the use of the absolute value. We propose a new solution method for the general L q -penalized regression problem based on space transformation and thus efficient optimization algorithms. The new method has immediate applications in statistics, notably in penalized spline smoothing problems. In particular, the LASSO problem is shown to be polynomial time solvable. Numerical studies show promise of our approach.
What problem does this paper attempt to address?