What to do for underdetermined problems?

[jseabold]

What to do for the linear OLS models when k > nobs? See #1046, for example.

[josef-pkt]

I'm slowly tending towards the view that we should issue a lot more warnings or exceptions by default, that can be overridden with an option.

k == nobs
Example in least trimmed squares LTS, the articles I read where starting with an exactly identified system (nobs = k_params). In my branch I used one additional observation, nobs = k_params + 1.
Similar Theil-Sen #557 would also use exactly identified.

Another case would be when we start with regularization. But this would need special handling than just doing the pinv regularization. (sklearn case with k > nobs)

[jseabold]

Not actionable. We have regularization now. Can add issues for specific cases.