Hello, me again!

I received wonderful instructions last week on how to get the estimated impulse response function for my regressors, and the mathematician in me is wondering:

Since we're simultaneously estimating the IRFs and the solution to the GLM, what constraints are there on the IRF (e.g., what prevents us from simply generating an IRF for one regressor that matches the activation) other than the restriction of the order of the CSPLIN (e.g.)?

More generally, what does it mean (formally) to solve these two simultaneously?

I've tried to find the answer online to no avail (in fact, "AFNI" and "CSPLIN" together in Google gives virtually nothing).

Cheers,
Aditya