In astrophysics and cosmology, it is necessary to estimate redshifts (z)
of astronomical objetcs, such as galaxies, in order to study the expansion
history and the growth of structure in the Universe, or to simply know more
about these objects and their physics.
The most accurate way to compute redshifts is to obtain spectra for all objects of interest
and look at the shift in the spectral lines due to cosmic expansion.
These are called spectroscopic redshifts.
However it is very expensive, time-consuming and sometimes impossible to obtain
spectra for large numbers of galaxies. In such cases, we estimate redshifts by
using broad-band photometry, i.e. we only use the information in the galaxy
magnitude/colors through a few broad filters.
Redshifts estimated in this way are called photometric redshifts (photo-z's).
There are two classes of methods to compute photo-z's. The first class, called
template fitting methods, uses a set of standard
galaxy spectral energy distributions (SED), the templates. These templates are supposed to
cover all possible galaxy types observed, and can be based on theoretical models
(synthetic templates), or based on real data (empirical templates).
Through a chi-squared minimization,
one determines the redshifted template whose broad-band colors best matches those observed in the
real galaxy of interest, simultaneously determining the galaxy type and redshift.
Another class of methods are the empirical or training-set methods.
These use a training set, i.e. a set of galaxies for which it is known not only
their broad-band colors, but also their true spectroscopic redshifts.
The training set is used to determine a functional relationship between
redshifts and colors, which is then applied to the photometric galaxies of interest
to estimate their redshifts. These methods include polynomial fits,
nearest neighbors, neural networks, gaussian processes, etc.
When a representative and dense training set is available, empirical methods typically
perform better than template methods. However, template methods can extrapolate
redshifts better when such training set is not available. A hibrid method that
includes both techniques
is probably the best way to ensure one has the most precise redshifts possible
without overfitting to the training set.
With Hiro Oyaizu, Carlos Cunha, Josh Frieman and Huan Lin,
I have worked on various aspects of photo-z estimation, including different photo-z
methods, estimates of photo-z errors, estimates of the redshift distribution,
with applications to both real data from the Sloan Digital Sky Survey (SDSS)
and simulated catalogs of the Dark Energy Survey (DES). With Wayne Hu, I also
worked on the requirements on photo-z precision in order to use galaxy
clusters to contrain cosmological parameters.
With Raul Abramo and collaborators, I have computed photo-z's for quasars and shown that they are potentially useful tracers of large-scale structure in narrow-band filter surveys, such as the Javalambre Physics of the Accelerating Universe Astrophysical Survey (J-PAS).
With the DES collaboration, I continue to be involved in the improvements of
photo-z methods for DES and their implementation in the DES Science Portal.