When studying spectral index, Faraday rotation or depolarization properties of extended sources, you should attempt to minimize uncertainties arising from differing resolutions at the different frequencies. It is important to try to ``scale" the distributions of baseline lengths in wavelengths that are used for the different observations. Ideally, you would scale the detailed distribution of antenna spacings but it is often difficult to scale more than the inner and outer baseline limits and the tapering and weighting functions (especially when observing with fixed arrays, as in VLBI). You should also make the hour-angle ranges of the observations at different frequencies as similar as possible.
Note that although using such ``scaled arrays'' optimizes your chances of measuring frequency-dependent properties of a source accurately, it cannot guarantee success. Even scaled arrays may have significantly different relative sensitivities to different scales of structure if the structure changes radically with frequency, e.g., if there are large spectral index gradients in either the total or polarized emission. You must also be careful when interpreting the final images if the databases at different frequencies are differently affected by missing antennas or by bad data. In many cases, the reliability of inter-frequency comparisons depends on how well the deconvolution algorithm (Lecture 9) interpolates unsampled parts of the u-v plane, even when scaled arrays have been used.
Observing at several adjacent, but not necessarily contiguous,
frequency bands can improve the u-v coverage, and so help you to
reconstruct complex structures with high fidelity, by ``bandwidth
synthesis". For this application, observing frequencies should be
spaced to minimize the sizes of residual ``holes" in the u-v coverage.
For example, if the largest ``holes" 
in the
coverage at one frequency are about 10% of the corresponding radii
, data from two observing frequencies separated by 5%
might be combined.