[daip] CALIB

[Eric Greisen]

The gridded subtraction method places all of the CCs on a grid of the
size of the image or one a factor of 2 larger in each dimension.  This
allows for some interpolation in the uv plane.  Due to (old) limits in
the FFTs in AIPS it doubles images of 2048 and smaller.  The 8192
images are then not interpolated and so even less accurate.  This
matters more, as you have found, when the source visibility is a rapid
function of uv (due to being well off center) rather than a smooth
function (centered source).  I am suprised that the 4096 images work
well since they too are not interpolated.

We could now double even 8192 images (although an FFT of 16K on a side
would take "forever").  Alternatively, we could double the 4096 and
force 8192 to use DFT unless the user has explicitly said 'GRID'.I
think that might be a better solution.  Then you could say GRID when
the source is mostly in the center of the field and live with DFT in
the other case.

Have you considered using facets instead of such monolithic images?

Eric Greisen