Much of the computation in `CLEAN' consists of shifting and scaling the dirty beam. As this is essentially a convolution it may, in some circumstances, be done more efficiently with 2-d FFTs. Clark's (1980) `CLEAN' algorithm does this, finding approximate positions and strengths of the point components using only a small patch of the dirty beam.

In detail, the Clark algorithm has two cycles, known as ``minor''
and ``major'' cycles. The *minor cycle* proceeds as follows:

- A beam patch (a segment of the discrete representation of the beam) is selected to include the highest exterior sidelobe.
- Points are selected from the dirty image if they have an intensity, as a fraction of the image peak, greater than the highest exterior sidelobe of the beam.
- A Högbom `CLEAN' is performed using the beam patch and the selected points of the dirty image. The stopping criterion for the `CLEAN' is roughly such that any remaining points would not be selected in step (2).

The algorithm then does a *major cycle* wherein the point source
model found by the minor cycle is transformed via an FFT, multiplied
by the weighted sampling function (inverse transform of the beam),
transformed back, and subtracted from the dirty image. Errors
introduced in a minor cycle by the beam patch approximation are, to
some extent, corrected in subsequent minor cycles.

1996 November 4

10:52:31 EST