forum.alglib.net

This is a long shot, but I'm wondering if there is anything I can do to optimise/improve the performance of the "lsfitlinear" method? I'm developing a .Net application that runs many different calculations on thousands of data points. One of these calculations involves a call to "lsfitlinear", and a .Net profiler tool shows these calls as accounting for half of the overall time spent in my calculation routine.

Typical usage scenario: my routine calls lsfitlinear around 5000 times, most times passing it just two values. The following is a typical example:-

var fmatrix = new double[,] { { 1, 0 }, { 1, 5 } };
var y = new double[] { 2623.203, 7202.05 };

Do you have guarantee that:
* you always have 2 points, 2 unknowns?
* fmatrix is non-degenerate and well conditioned?

If so, you may just solve it as "conventional" linear system with 2 unknowns and 2 equations. lsfitlinear() is optimized for medium and large problems (10-100-1000 points, 10-100-1000 parameters). It works for such tiny systems, but it is overkill.