Radiation Grid
                             --------- ----

To save computational resources, the radiative fluxes/heating are calculated 
on a 128x62 Gaussian grid (model is T62).  The dynamics, on the other hand, 
are calculated on a 192x94 Gaussian grid (to prevent quadratic terms from 
aliasing the results).  So the radiative fluxes and heating terms have to 
converted to the higher-resolution grid.  The model uses a bilinear interpol-
ation for the interior grid points and extrapolates the northern and southern-
most latitude band.

This procedure can introduce interpolation/extrapolation errors.  These errors 
appear in the NSWRS.  While the upward and downward short wave fluxes both 
look reasonable (reduced fluxes towards the unlit parts of the globe), 
combining these terms (with the interpolation errors) can give a small upward 
net short-wave flux.  Of course, this error appears near the unlit parts of 
the globe, where the interpolations errors are relatively large.

Another place where interpolation errors can occur is in the variance fields.  
In the interpolation to the new grid, some of the grid points may be roughly 
equidistant to 4 of the old grid points.  The value at that grid point will be 
approximately the mean of the 4 old grid points.  On the other hand, another 
new grid point may be very near an old grid point.  Its value will be the same 
as the old grid point.  When computing the temporal variance, the first new 
grid point is likely to have a smaller variance than the second.  Plots of the 
variances in the cloud fraction and radiation fluxes show a grid-cell 
structure from the interpolation scheme.  Obviously the correct method to 
calculate the variance is to compute the variance on the original grid 
(128x62) and later convert to a common grid.

In hindsight, we should have (1) used the 192x94 Gaussian grid for the 
radiation calculations or at least have saved the data on the 128x62 grid (it 
would have also saved space).  It is, however, possible to recover the 
original data on the 128x62 grid.  While it is not as simple as interpolating, 
it is possible to write the inverse of the bilinear interpolation scheme.  If 
the bilinear interpolation operator could be split up into two operators, 
say X and Y.  Suppose these operators where only dependent on x (latitude) or 
y (longitude) respectively.  This will allow the problem to be split up into 
two manageable parts (suggestion by Jim Purser).