Performance under realistic spectroscopic noise

Determine how the probabilistic autoencoder's denoising and galaxy-population calibration performance holds up when trained on spectra with realistic heteroscedastic, wavelength-dependent, correlated, and observationally systematic noise rather than independent Gaussian noise of constant amplitude.

Background

The study evaluates the probabilistic autoencoder using a simplified noise model consisting of independent Gaussian noise with a constant amplitude across the dataset. Real spectroscopic observations instead exhibit heteroscedastic noise, wavelength-dependent uncertainties, detector correlations, observational systematics, and contamination from sky lines.

Because the reported autoencoder performance benefits from the assumed Gaussian likelihood, the authors leave unresolved whether the method retains its accuracy under realistic observational conditions. They note that full forward modelling of these effects might be needed, but its net impact has not been established.