The Clark CLEAN

The Clark CLEAN algorithm (Clark 1980) (A&A 26, 89) is a variant of the Högbom CLEAN, and aims at improved speed for large images. The original Högbom algorithm spends a lot of time searching residuals which are negligible, and shifting and scaling the dirty beam. The Clark algorithm avoids these by searching a list of only the largest residuals, working with only a sub-region of the beam for much of the time, and by using FFTs to perform convolutions when it has to resort to using the full beam. This leads to an algorithm which is more efficient for large images (although a good deal more complicated). For small images though, the Högbom CLEAN may still be faster.

The Clark CLEAN plays a few tricks which you should know about. It has what are termed major and minor cycles. In the minor cycle, CLEAN components are searched for, and subtracted from the image, in a fashion very similar to that of the Högbom CLEAN. The difference is that only the central portion of the dirty beam is used for the subtraction, and a list of the largest residuals are searched (not the full region being CLEANed). The rationale is that for dirty beams that have a fairly good sidelobe pattern, this will be good enough to find the CLEAN components. At some designated time, this minor cycle is terminated and a major cycle computed. This means that the list of CLEAN components from the current cycle are `FFTed', and subtracted from the FFT of the residual image that resulted from the previous major cycle. In this way, errors that might have accumulated in the subtraction phase of the minor cycle with only a part of the beam are largely set to rights. It is also at this stage that the list of the largest residuals is re-determined.

This use of a small beam patch in the Clark CLEAN is a potential danger if the beam is poor. Images made from ATCA snapshot data, even if there are a few cuts, often have a beam in which the first sidelobe response is quite strong and outside of the beam patch that the Clark CLEAN typically uses.