Both selfcal and gpscal assume that the gains are independent of frequency. This allows them to use data from all channels in forming gain solutions. However, as mentioned above, they do not apply bandpass corrections. Consequently, if you have multi-channel data and models, you will probably want to form a bandpass-corrected visibility dataset before doing self-calibration.
The assumption that the gains are independent of frequency should be treated with some caution. If all the visibilities for a given baseline go through one signal path, then the instrumental phase for the different visibilities should be equal (assuming the signal path is well equalised). However, atmospheric phase should vary linearly with frequency. Provided the fractional bandwidth is small, it is still a reasonable approximation that the gains are independent of frequency. Indeed this is the normal approximation used in deriving calibration solutions from a `channel 0' dataset.
If the visibilities for a given baseline go through multiple signal paths (e.g. the 2 IF system for the ATCA), then the instrumental phases for the different paths could be significantly different. Tentative results for the 2 IF system of the ATCA show that, apart from an offset, the phase of the two signal paths track each other as the ratio of the frequencies (i.e. the path length difference is constant). The offset between them can be eliminated either by adjusting the phases of the two paths to zero at the start of the observation, or by the primary calibration process. Provided the offset has been eliminated, and provided the frequencies are not too different, it may be useful to approximate the gains as independent of frequency.
If the phases vary significantly between the frequencies, then the different frequencies need to be self-calibrated individually. Ideally there would be a self-calibration task which knows that the phase errors at different frequencies tracked as the ratio of the frequencies.
When self-calibrating multi-channel visibility datasets, there are two distinct modes of calculating the variation of the model with frequency, depending on whether the model contains multiple planes, or whether multi-frequency techniques were used. We discuss these two possibilities in turn.