I spent 3 days in Köln with Alvaro Sanchez-Mongé, Peter Schilke, Fanyi Meng, and Anika Schmiedeke working on Sgr B2 data and specifically on trying to combine the single dish (total power) data with my merged ACA+12m data from ALMA program 2013.1.00269.S.
tl;dr: feathering appears to work fine, and many other methods work equally well (assuming they can be implemented, which is uniformly harder). Negative bowls persisting after feathering are an indication of a problem with the input data.
Fourier space combination of the single-dish and interferometric data is by far the most straightforward to implement and the fastest. However, in the HC3N data, it left strong negative bowls, which should not be possible.
On the Einstein image, the image quality from feathering was not great, but there were no negative bowls left. The dynamic range is lower in the Einstein data, but this still hints that there is a problem with the HC3N images. We're investigating the possibility that the 7m data are improperly calibrated or weighted.
We never managed to create UV data from the single dish data because of complaints about the pointing information.
Cleaning with a single dish image as an input model
We attempted to clean the data using the single dish image as an input model.
On the real data, this failed with both tclean and clean. With tclean, there was no error message, it just hung indefinitely. With clean, the error message is:
*** Error *** LatticeExprNode - coordinates of operands mismatch Scanned so far: modelos_0 + __temp_model2 2016-06-01 10:35:05 SEVERE clean:::: An error occurred running task clean.
Changing CDELT to 1 GHz, which solved a previous issue, had no effect here.
It turns out tclean will fail silently if it doesn't find the startmodel, which has to be specified as an image prefix for version <=4.6, and the image has to exist as a .model file. For higher versions, 4.6+, the model can be directly referenced (as in clean). Eventually, tclean seems to have completed, though the results indicate that it does not treat the units as I expected; the total power data seems to be over-weighted by a factor of 10^3+, probably by the ratio of the beam areas:
For the simulated images (einstein), we get the error:
2016-06-01 10:55:11 SEVERE SynthesisImager::defineImage (file /Users/rpmbuild/gradle/workdir/casasources/release-4_5/code/synthesis/ImagerObjects/SynthesisImager.cc, line 668) Error in adding Mapper : Error in createImStore : ::operator!= (const IPosition&, const IPosition&) - left and right operand do not conform 2016-06-01 10:55:11 SEVERE tclean::task_tclean:: Exception from task_tclean : 2016-06-01 10:55:11 SEVERE SynthesisImager::defineImage (file /Users/rpmbuild/gradle/workdir/casasources/release-4_5/code/synthesis/ImagerObjects/SynthesisImager.cc, line 668) Error in adding Mapper : Error in createImStore : ::operator!= (const IPosition&, const IPosition&) - left and right operand do not conform
This is probably an issue with regridding. imregrid doesn't like our data:
2016-06-01 11:00:41 SEVERE imregrid::image::regrid Exception Reported: Exception: The number of pixel axes in the output shape and Coordinate System must be the same. Shape has size 4. Output coordinate system has 3 axes. 2016-06-01 11:00:41 SEVERE imregrid::image::regrid+ ... thrown by SPIIF casa::ImageRegridder::_regrid() const at File: /Users/rpmbuild/gradle/workdir/casasources/release-4_5/code/imageanalysis/ImageAnalysis/ImageRegridder.cc, line: 138
Adding a frequency axis to the FITS data (which was missing before...) seems to have fixed this error for the Einstein data, but not for the real HC3N data.
Linear Combination in Image Space
Linear combination of the single dish and interferometer data in image space, followed by image-space deconvolution, has been used successfully on HI data. In principle, this is very straightforward, except for the need for deconvolution. CASA now has an image-space deconvolution program (deconvolve), so I was able to implement this approach. However, the deconvolver seems to only work on the inner 1/4 of the image in each dimension, which left incomplete images that were difficult to compare. Additionally, CASA does not (obviously) carry the appropriate machinery for computing the residual image and adding the convolved model back to the residual image. Still, this is a promising route forward as it is computationally relatively cheap and mostly easy to implement.
Linear combination is partly implemented now using the CASA deconvolve task; it hasn't been generalized but you can see the outline / single working case at this link.