Summary

Determines the amplitude and phase as a function of frequency for each spectral window containing more than one channel. Strong sources (or many observations of moderately strong sources) are needed to obtain accurate bandpass functions. The two solution choices are: individual antenna/based channel solutions 'B'; and a polynomial fit over the channels 'BPOLY'. The 'B' solutions can be determined at any specified time interval, and is recommended in most applications.

Introduction

For channelized data, it is usually desirable to solve for the gain variations in frequency as well as in time. Variation in frequency arises as a result of non-uniform filter passbands or other frequency-dependent effects in signal transmission. It is usually the case that these frequency-dependent effects vary on timescales much longer than the time-dependent effects handled by gaincal. Thus, it makes sense to solve for them as a separate term, using the bandpass task.

It is usually best to solve for the bandpass in channelized data before solving for the gain as a function of time. However, if the gains during the bandpass calibrator observations are fluctuating over the timerange of those observations, then it can be helpful to first solve for those time-dependent gains of that source with gaincal, and input these to bandpass via gaintable. See the examples section for more on how to do this.

See "Solving for Calibration" for more information on the task parameters bandpass shares with all solving tasks, including data selection, general solving properties and arrange prior calibration. Below we describe parameters unique to bandpass, and those common parameters with unique properties.

Bandpass types: bandtype

The bandtype parameter selects the type of solution used for the bandpass. The choices are 'B' and 'BPOLY'.

bandtype='B'

Use of bandtype='B' in bandpass differs from gaintype='G' in gaincal only in that it is determined for each channel in each spectral window. It is possible to solve for it as a function of time, but it is most efficient to keep the B solving timescale as long as possible, and use gaincal for frequency-independent rapid time-scale variations.

Do not use combine='spw' with bandtype='B', as this will generate a solution for all spws overlaid in channel coordinates, and for which it is not yet possible to apply to all spws in frequency coordinates.

The B solutions are limited by the signal-to-noise ratio available per channel, which may be limited. It is therefore important that the data be optimally coherent over the time-range of the B solutions. As a result, B solutions are almost always preceded by an initial, provisional gaincal solution. In turn, if the B solution improves the frequency domain coherence significantly, subsequent gaincal solutions using it will be better than the original. The SNR per bandpass channel can also be boosted by using a non-trivial frequency solint to partially average the MS visibility frequency channels for the solution. However, for accuracy, it is important to use a frequency solint that doesn't obscure actual systematic bandpass structure. If adequate SNR is unachievable by these means with the available data, use of bandtype='BPOLY' can be considered.

bandtype='BPOLY'

For some observations, it may be the case that the SNR per channel is insufficient to obtain a usable per-channel B solution. In this case it is desirable to solve instead for a best-fit functional form for each antenna using the BPOLY solver. The BPOLY solver fits (Chebychev) polynomials to the amplitude and phase of the calibrator visibilities as a function of frequency. Use of combine='spw' will cause a single common BPOLY solution to be determined in frequency space for all selected spectral windows in aggregate (plots of such solutions with plotcal will only show the evaluated polynomial for the first spw used in the solve). It is usually most meaningful to do per-spw solutions, unless groups of adjacent spectral windows are known a priori to share a single continuous bandpass response over their combined frequency range.

The BPOLY solver requires a number of unique sub-parameters (default values are given below):

The degamp and degphase parameters indicate the polynomial degree desired for the amplitude and phase solutions. The maskcenter parameter is used to indicate the number of channels in the center of the band to avoid passing to the solution (e.g., to avoid Gibbs ringing in central channels for PdBI data). The maskedge parameter drops beginning and end channels. The visnorm parameter turns on normalization of the visibilities before the solution is obtained (rather than after as for solnorm).

The combine parameter can be used to combine data across spectral windows, scans, and fields.

Note that bandpass will allow you to use multiple fields, and can determine a single solution for all specified fields using combine='field'. If you want to use more than one field in the solution, it is prudent to use an initial gaincal using proper flux densities for all sources (not just 1 Jy) and use this table as an input to bandpass because in general the phase towards two (widely separated) sources will not be sufficiently similar to combine them, and you want the same amplitude scale. If you do not include amplitude in the initial gaincal, you probably want to set visnorm=True also to take out the amplitude normalization change. Note also in the case of multiple fields, that the BPOLY solution will be labeled with the field ID of the first field used in the BPOLY solution.

Bandpass calibration considerations

Bandpass normalization (solnorm)

The solnorm parameter requires more explanation in the context of the bandpass. Most users are used to seeing a normalized bandpass, where the mean amplitude is unity and fiducial phase is zero. Use of solnorm=True allows this. However, the parts of the bandpass solution normalized away will be still left in any data to which it is applied, and thus you should not use solnorm=True if the bandpass calibration is the end of your calibration sequence (e.g. you have already done all the gain calibration you want to).

You can safely use solnorm=True if you do the bandpass first (perhaps using a throw-away initial gaincal calibration) as we suggest above, as later gaincal calibration stages will deal with this remaining calibration term. This does have the benefit of isolating the overall (channel independent) gains to the following gaincal stage. It is also recommended for the case where you have multiple scans on possibly different bandpass calibrators. It may also be preferred when applying the bandpass before doing gaincal and then fluxscale, as significant variation of bandpass among antennas could otherwise enter the gain solution and make (probably subtle) adjustments to the flux scale.

We finally note that solnorm=False at the bandpass step in the calibration chain will still in the end produce the correct results. It only means that there will be a part of what we usually think of the gain calibration inside the bandpass solution, particularly if bandpass is run as the first step.

What if the bandpass calibrator has a significant spectral variation?

The bandpass calibrator may have a spectral slope that will change the spectral properties of the solutions if a flat-spectrum model is used. If the slope is significant, the best remedy is to estimate the spectral shape and store that model in the bandpass calibrator MS. To do so, go through the normal steps of bandpass and the gaincal runs on the bandpass and flux calibrators, followed by setjy of the flux calibrator. The next step would be to use fluxscale on the bandpass calibrator to derive its spectral index. fluxscale can store this information in a python dictionary which is subsequently fed into a second setjy run, this time using the bandpass calibrator as the source and the derived spectrum (the python dictionary) as input. This step will create a source model with the correct overall spectral slope for the bandpass calibrator. Finally, rerun bandpass and all other calibration steps again, making use of the newly created internal bandpass model.