For all solvers, the solution for the x_2 component is inaccurate.
Using KNITRO and the suggested option, the accuracy is to machine precision.
---
2.2
---
There are at least two solutions, one for the suggested starting guess ( x_i = 1 / n ) and one for x_i = (-1)^i / n
Both solutions have the same objective, i.e. 0.
---
2.3
---
Most of solvers give very bad fits. Only CONOPT gives the same solution as the certified values.
The problem is ill-conditioned, with 9 orders of magnitude between the largest and the smallest coefficients.
---
2.4
---
Orthogonal distance regression improves the solution for most solvers (in that the residuals are smaller), but still CONOPT is the only one getting the same solution as the certified values.