From Jeff Newman -- 

To recap: I have taken the code Alison Coil used to produce redshift 
distributions for her recent 
<a href="http://arxiv.org/abs/astro-ph/0403423">w(theta) paper</a>  
and made a couple of modifications: adding a z^2 exp(-(z/z0)^1.2) fit, 
in addition to the existing z^2 exp(-(z/z0)) ; and doing fits as a 
function of R magnitude, as well as I (because the faintest I bins 
are incomplete, since redshifts are obtained only for objects with 
R_AB < 24.1, we can get robust results over a broader magnitude range 
in R than I).  Alison's code applies lowest-order corrections for 
target selection as a function of magnitude; it will be possible to 
do somewhat better in the long term, but that will take a fair amount 
of time.  I would not consider these numbers definitive, but I'd 
certainly expect them to be good to 10% at worst.

All magnitudes are AB in the native CFHT 12k filter system.  Note that 
these filters yield response curves somewhat different from standard 
R and I.

Note the strong agreement between the numbers for the two fits for 
mean and median z.  To go from z0 to mean z, I have used multipliers 
of 3 and 2.09 for power 1 and 1.2, respectively; to go from z0 to 
median z, I have used 2.67 and 1.91.

I have attempted to extrapolate these numbers to fainter magnitudes; 
those extrapolations need to be taken with a huge grain of salt.  A 
linear relation fits the trend of z0 vs. limiting magnitude extremely 
well (a linear fit has an RMS error in z0 corresponding to 0.01 in 
predicting mean or median z), which is certainly suggestive, but I 
would be absolutely shocked if those trends continued to R=28, for 
a variety of reasons.

The results are:

        mag                   ^1 fit                  ^1.2 fit
       range           <z>         median         <z>         median
     18<R<20         0.311         0.277         0.302         0.276
     18<R<21         0.408         0.363         0.393         0.359
     18<R<22         0.503         0.448         0.485         0.443
     18<R<23         0.621         0.553         0.596         0.545
     18<R<24         0.766         0.681         0.731         0.668

extrapolated:
   18<R<24.0         0.746         0.664         0.714         0.652
   18<R<24.5         0.802         0.714         0.767         0.701
   18<R<25.0         0.858         0.764         0.820         0.749
   18<R<25.5         0.915         0.814         0.873         0.797
   18<R<26.0         0.971         0.864         0.926         0.846
   18<R<26.5         1.027         0.914         0.979         0.894

extrapolation is based on a simple linear fit to z0 vs. limiting 
magnitude, which works extraordinarily well over the sampled magnitude 
range.


For I:

        mag                 ^1 fit                     ^1.2 fit
       range           <z>         median         <z>         median

     18<I<20         0.370         0.330         0.358         0.327
     18<I<21         0.504         0.448         0.486         0.444
     18<I<22         0.617         0.550         0.594         0.542
     18<I<23         0.725         0.645         0.695         0.635 *
     18<I<24         0.774         0.689         0.739         0.675 **

* slightly incomplete (objects around I~23 with R-I > 1.1 will be lost)
** highly incomplete (objects around I~24 with R-I > 0.1 (!!!) will be lost)

extrapolated:

   18<I<24.0         0.849         0.755         0.813         0.743
   18<I<24.5         0.908         0.808         0.869         0.794
   18<I<25.0         0.967         0.860         0.925         0.845
   18<I<25.5         1.026         0.913         0.981         0.896
   18<I<26.0         1.085         0.965         1.037         0.948
   18<I<26.5         1.143         1.018         1.093         0.999

extrapolation is based on a simple linear fit to z0 vs. limiting 
magnitude, which works extraordinarily well over the sampled magnitude 
range, except in the I=24 bin (omitted from the fit, note its deviation 
from the I=24 extrapolation).