yuki.moritani Thank you for the information. I slightly updated my ticket with information on how the pfs offsets were calculated, just for clarification. They are calculated using the pfssingle.observations nominal/center positions. I wasn't aware of the 20-30um configuration error, which as you said may explain some of the observed behavior. I will check with Tanaka-san further about this.

My understanding from the technical telecon a while ago is that the distance used to set NOTCONVERGED is the distance between pfiNominal and pfiCenter. Maybe I misunderstood because I am not sure what Moritani-san meant by "this flag is based on cobra converged sequence itself".

I'm sorry for confusion. It s right that NOTCONVERGED is set based on distance between pfiNominal and pfiCenter, which is calculated cobra convergence sequence. And this flag is use to mark fibers with unexpectedly large offset. (Convergence itself has threshold, for which 5um is used now, and Cobra stops when its distance to pfiNominal gets below it.) What I mean is that you should be careful that convergence assumes pfiNominal has negligible error, but in reality, pfiNominal still has uncertainty. That's why raster scan data shows larger error than convergence error. So, if you see flux loss distance is larger than xx um, I'm not sure this means cobra is not converged.

Thank you, Moritani-san! I think the uncertainty in pfiNominal explains why (at least in part) Sadman's plot does not look quite similar to Arnauds's one. (1) Sadman's plot shows a sequence of objects at flux loss ~ 0 with ~ 0.2mag scatter. Most of these objects are located at offset < 30um. So, their offsets are dominated by the uncertainty in pfiNominal, and we see little correlation between the offset and flux loss. (2) At larger offsets, we start to see objects with a significant flux loss. I think this is the regime where the real offset becomes dominant over the uncertainty. Because we use the distance between pfiNominal and pfiCenter to set the UNCONVERGED flag, perhaps it makes sense to define the distance threshold based on the observed data (i.e., Sadman's plot)?

Thank you, Tanaka-san, for explanations. I see... it makes sense to determine by using observed data for larger offset. But I'm confused why you selected 25um  (<30um) was selected? And probably we should analyze all available data not limiting to one FoV or something... (And you should work under a new ticket, as this is already closed.)