search

JavGVastro / PhD.Paper

This repository contains updated methods for the project H II regions Structure functions (previously CSHR).
0 stars 2 forks source link

Some of the MCMC runs may not have converged #9

Closed will-henney closed 2 years ago

will-henney commented 2 years ago

I am writing the description of the MCMC estimation of the parameter uncertainties.

The original plan was to use chain lengths that were at least 50 times the autocorrelation time, to ensure that the walkers have converged. I have just done some checks that this is satisfied:

For Orion we used 5000 and the autocorrelation times are all less than 60, so that is OK.

For 30 Dor we used 10000 and autocorrelation times are less than about 100, also OK.

For Hubble X we used 5000 and autocorrelation times are as high as 270, so not OK. We should use at least 150000 and retry it. Note that the acceptance fraction is also very low in this case (around 0.35), which also indicates that we probably haven't converged.

Please check all the other sources to make sure they satisfy the criterion.

JavGVastro commented 2 years ago

Here I made a table for the walkers/autocorrelation-time ratio and the acceptance fraction showing that most regions, besides HX (as previously mentioned) and the EON, comply with the criterion. For the previous regions the calculations haven't converged yet. The values shown are from the confidence intervals determinations used in the current results showed in the paper. Any ratio value that may need a re-run let me know.

Region | Walkers/autocorrelation-time | Acceptance fraction Carina | 5000 / 62 = 80 | 0.49 30Dor | 5000 / 73 = 68 | 0.54 HV | 15 000 / 241 = 62 | 0.44 HX | 10 000 / 70 =142 | 0.48 M8 | 10 000 / 100 = 100 | 0.48 346 | 5 000 / 57 = 87 | 0.53 595 | 25 000 / 66 = 380 | 0.54
604 | 25 000 / 120 = 206 | 0.47 EON | Still running Orion | 5 000 / 57 = 87 | 0.55

Edit: I just added HX. For this region I adjust the limits on the parameters and wasn't necessary to increase the number of walkers; the advantage of this is shorter computation time and smaller results files. The results were pretty similar but the confidence intervals for r0 and sig2 decreased considerably

will-henney commented 2 years ago

This seems to be resolved now.