We compute accurate redshift distributions to I _ { \mbox { \small { AB } } } = 24 and R _ { \mbox { \small { AB } } } = 24.5 using photometric redshifts estimated from six–band U BV RI Z photometry in the Canada–France Deep Fields–Photometric Redshift Survey ( CFDF–PRS ) .
Our photometric redshift algorithm is calibrated using hundreds of CFRS spectroscopic redshifts in the same fields .
The dispersion in redshift is \sigma / ( 1 + z ) \lesssim 0.04 
to the CFRS depth of I _ { \mbox { \small { AB } } } = 22.5 , rising to \sigma / ( 1 + z ) \lesssim 0.06 at our nominal magnitude and redshift limits of I _ { \mbox { \small { AB } } } = 24 
and z \leq 1.3 , respectively .
We describe a new method to compute N ( z ) that incorporates the full redshift likelihood functions in a Bayesian iterative analysis and we demonstrate in extensive Monte Carlo simulations that it is superior to distributions calculated using simple maximum likelihood redshifts .
The field–to–field differences in the redshift distributions , while not unexpected theoretically , are substantial even on 30′ scales .
We provide I _ { \mbox { \small { AB } } } and R _ { \mbox { \small { AB } } } redshift distributions , median redshifts , and parametrized fits of our results in various magnitude ranges , accounting for both random and systematic errors in the analysis .