Making chi-squared fitting usable

[ewoudwempe]

Currently it is just a script that calculates chi-squared for some files that you put in by hand. Since each folding data point costs about 3 seconds on a virgo-machine for a 1GB file, I computed them on the virgo-cluster at Kapteyn (more points is more precise, although maybe not, but that is another issue).

Then, once enough trial periods are calculated, I fitted the chisq peaks by hand separately for each pulsar (because different fitfuncs were necessary, and checking if the fit actually converges etc.). I used ray for the cluster-computing. How to use that needs to be explained too I guess (it involves some setup to make it go smoothly...). And how do we want this to go into the analysis code? Making the users go through a jupyter notebook to actually check the fits? Just trying every function we have and selecting the one that fits best?

[Fonotec]

I think it needs some explaining how to set it up, I think we don't want the user to go through a jupyter notebook, I think it is best to just try every function. Computation power is not a problem, I think doing a fit on these points will take less than a second, so doing all functions won't be a problem. So I think the one that fits best is a good way to go, if the user didn't specified the function, but I would also output the results for the other functions, such that the user at least has the result for all the functions.

[ewoudwempe]

The performance issue is in the calculating the chi-squared part. For each (period, DM, chi-squared) datapoint, the data needs to be folded with a different period. If you want say 10k points (if you don't have a good guess and want good precision that is reasonable), this'd take 30k seconds = 8h, so using a cluster or at least all cores is nice.
Another thing I am not sure about is how to estimate uncertainties from this chi-square analysis. The fit uncertainty is not a good indicator, it turns out.