This week we submitted a new paper to A&A where we suggest a method for improving CLEAN images in the context of RM synthesis. The method allows one to make lower resolution images while obtaining better results than when using CLEAN alone and moreover it makes the results much less dependent on the choice of pixelization.
What we and many others have found is that RMCLEAN doesn't do such a great job at reconstructing the locations or fluxes of sources, especially when there are several of them close together. When writing the 3D CLEAN algorithm for Faraday synthesis we noticed that RMCLEAN doesn't even do that great with a single source unless it is located directly in the middle of an image pixel. The dynamic range in a CLEANed image is well-known to be limited due to the fact that you can't exactly model the location of a source in a pixelized image. You can partially overcome the issue by making the image have very high resolution, but especially for 3D imaging this becomes expensive both computationally and in terms of storage.
So we looked into this in a bit more detail, and devised a method for improving the CLEAN generated model using maximum likelihood (ML) estimation. The method is similar to others that have been suggested for aperture synthesis imaging. but it seems to have even more impact in the case of RM synthesis. In our testing, we found that the ML method dramatically reduces the error in measurements of both source location and flux. Somewhat surprisingly, we also found that increasing resolution doesn't reduce the errors in normal CLEAN images in the case where two sources are nearby to one another. However, he ML algorithm was able to get accurate results in such a case even in a low resolution image.
Ultimately we don't think that CLEAN (or any method that assumes the signal to be diagonal in pixel space) is the right approach in all cases, and we're actively working on new methods that take advantage of correlated structures in the data to help constrain reconstructions. Nevertheless, CLEAN is undoubtedly a useful algorithm in some circumstances, and together with this new algorithm it can give rather good results while being easy to implement and (relatively) computationally inexpensive.
The pre-print is available now on the arXiv. Give it a read if you want to learn more. And if you're interested in implementing the method for your own analysis, then be sure to stay tuned because I plan to release the code for this relatively soon.