The sensitivity of GPz estimates of photo-z posterior PDFs to realistically complex training set imperfections

The accurate estimation of photometric redshifts is crucial to many upcoming galaxy surveys, for example the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST). Almost all Rubin extragalactic and cosmological science requires accurate and precise calculation of photometric redshifts; many diverse approaches to this problem are currently in the process of being developed, validated, and tested. In this work, we use the photometric redshift code GPz to examine two realistically complex training set imperfections scenarios for machine learning based photometric redshift calculation: i) where the spectroscopic training set has a very different distribution in colour-magnitude space to the test set, and ii) where the effect of emission line confusion causes a fraction of the training spectroscopic sample to not have the true redshift. By evaluating the sensitivity of GPz to a range of increasingly severe imperfections, with a range of metrics (both of photo-z point estimates as well as posterior probability distribution functions, PDFs), we quantify the degree to which predictions get worse with higher degrees of degradation. In particular we find that there is a substantial drop-off in photo-z quality when line-confusion goes above ~1%, and sample incompleteness below a redshift of 1.5, for an experimental setup using data from the Buzzard Flock synthetic sky catalogues.