Understanding how the physical properties of galaxies (e.g., their spectral
type or age) evolve as a function of redshift relies on having an accurate
representation of galaxy spectral energy distributions. While it has been
known for some time that galaxy spectra can be reconstructed from a handful
of orthogonal basis templates, the underlying basis is poorly constrained.
The limiting factor has been the lack of large samples of galaxies (coveri
ng a wide range in spectral type) with high signal-to-noise spectrophotomet
ric observations. To alleviate this problem we introduce here a new techniq
ue for reconstructing galaxy spectral energy distributions directly from sa
mples of galaxies with broadband photometric data and spectroscopic redshif
ts. Exploiting the statistical approach of the Karhunen-Loeve expansion, ou
r iterative training procedure increasingly improves the eigenbasis, so tha
t it provides better agreement with the photometry. We demonstrate the util
ity of this approach by applying these improved spectral energy distributio
ns to the estimation of photometric redshifts for the HDF sample of galaxie
s. We find that in a small number of iterations the dispersion in the photo
metric redshifts estimator (a comparison between predicted and measured red
shifts) can decrease by up to a factor of 2.