A few (obvious) practical recommendations

Map size

Make an image twice the size of the primary beam (e.g. $2\times55''$ at 90 GHz and $2\times22''$ at 230 GHz for PdBI antenna) to ensure that all the area of the primary beam (inner quarter of the dirty map) will be cleaned whatever the deconvolution algorithm is used. However, avoid making a too large dirty image because the CLEAN algorithms will then try to deconvolve region outside the primary beam area where the noise dominates.

Support

Start your first deconvolution without any support to avoid biasing your clean image. If the source is spatially bounded, you can define a support around the source and restart the deconvolution with this a priori information. Be careful to check that there is no low signal-to-noise extended structure that could contain a large fraction of the source flux outside your support... Avoid defining a support too close to the natural edges of your source. Indeed, deconvolving noisy regions around your source is advisable because it ensures that you do not biased your deconvolution.

Stopping criterion

Choose the right stopping criterion.

• Estimate an empirical noise on your first deconvolved cleaned image with GO NOISE.
• If this empirical noise value is larger the value computed from the visibility weights (This noise value is one of the output of the UV_MAP command), your observation is probably dynamic limited, i.e. you have a bright source whose dirty sidelobes are much larger that the thermal noise. In this case use, set ARES to 0 and FRES to a fraction which depends on the sidelobe level of your dirty beam.
• Else you are in the noise limited case. Set FRES to 0 and ARES to a fraction of the empirical noise value (typically 0.5).

Convergence checks

Ensure that your deconvolution converge by checking that

• The cumulative flux as a function of the number of clean component has reached a roughly constant level (use \FLUX option of the deconvolution commands to see this curves).
• The residual map looks like noise where the source appears in the clean map.

Else change the values of the stopping criterion, in particular the number of clean components (NITER).

Deconvolution methods

If you want a robust result in all cases, start with HOGBOM. If you prefer obtaining a quick result, use CLARK but you then first need to check that the dirty sidelobes are not too large on the dirty beam. If you obtain stripes in your clean image:

• First check that there is no spurious visibilities that should be flagged.
• Then check that your deconvolution converged.
• If it is clear that you have an extended source structure, you should first ask yourself whether you are in the wide-field imaging case and act accordingly (see next chapter). Else you can try a CLEAN variant which better deals with cases that implies a large spatial dynamic. This is rare at PdBI.

Outside help

Always consult an expert until you become one.