Robustness to non-representative spectroscopic selection

Evaluate how the probabilistic autoencoder galaxy-population prior performs when the spectroscopic training sample is selected under a Rubin-like photometric survey selection and progressively subjected to magnitude cuts that reduce its representativeness.

Background

The analysis assumes that the spectroscopic calibration sample is representative of the photometric weak-lensing sample and uses a 4C3R2-like magnitude limit. The authors note that real weak-lensing samples are generally fainter than spectroscopic calibration samples, so selection effects may bias the learned prior and the resulting redshift distributions.

The paper proposes a concrete unresolved test: repeat the analysis with a Rubin-like selection and impose increasingly restrictive magnitude cuts on the spectroscopic training data. This would quantify how the method's photometric-redshift calibration accuracy degrades as the training sample becomes less representative.