The forward step (aka flux mop in the clean worker) is sometimes unstable when using a PSF that is less than twice the size of the image in both dimensions. I suspect this is due to the use of PCG algorithm which is very fast but not always stable. Need to investigate ways to stabilize the inversion in this case. Possible avenues to investigate:
more stable optimisation routines (eg. the primal dual),
shaped Tikhonov regularisation,
BFGS-like rank 2 updates to the PSF convolution operator,
w-binned PSFs.
The latter two, which would improve the Hessian approximation, will also make the forward step more accurate when imaging a wide field of view.
The forward step (aka flux mop in the clean worker) is sometimes unstable when using a PSF that is less than twice the size of the image in both dimensions. I suspect this is due to the use of PCG algorithm which is very fast but not always stable. Need to investigate ways to stabilize the inversion in this case. Possible avenues to investigate:
The latter two, which would improve the Hessian approximation, will also make the forward step more accurate when imaging a wide field of view.