Abstract
We examine the impact of non-Gaussian photometry errors on photometric
redshift performance. We find that they greatly increase the scatter, but this
can be mitigated to some extent by incorporating the correct noise model into
the photometric redshift estimation process. However, the remaining scatter is
still equivalent to that of a much shallower survey with Gaussian photometry
errors. We also estimate the impact of non-Gaussian errors on the spectroscopic
sample size required to verify the photometric redshift rms scatter to a given
precision. Even with Gaussian {\it photometry} errors, photometric redshift
errors are sufficiently non-Gaussian to require an order of magnitude larger
sample than simple Gaussian statistics would indicate. The requirements
increase from this baseline if non-Gaussian photometry errors are included.
Again the impact can be mitigated by incorporating the correct noise model, but
only to the equivalent of a survey with much larger Gaussian photometry errors.
However, these requirements may well be overestimates because they are based on
a need to know the rms, which is particularly sensitive to tails. Other
parametrizations of the distribution may require smaller samples.