Abstract
A method is proposed for least absolute deviation curve fitting. It may be used to obtain least absolute deviations fits of general linear regressions. As a special case it includes a minor variant of a method for fitting straight lines by least absolute deviations that was previously thought to possess no generalization. The method has been tested on a computer and was found on a range of problems to execute in 1/4 to 1/20 of the cpu time required by a published algorithm based on linear programming.