Casa Cook Arvind

N. OOZEER, M. BIETENHOLZ & S.
GOEDHART
INTRODUCTION TO CASA:
A K AT- 7 D ATA R E D U C T I O N
GUIDE
2
. . . YO U C A N N O T A C Q U I R E E X P E R I E N C E B Y M A K I N G E X P E R I M E N T S . YO U C A N -
N O T C R E AT E E X P E R I E N C E . Y O U M U S T U N D E R G O I T.
ALBERT CAMUS.
Contents
1 Why do you want to learn CASA? 5

1.1 How to use this document 6
1.2 Conventions used 7
1.3 Methodology 8
2 Let us get serious 9

2.1 Overview of Calibration 9
2.2 STEP 1... Converting the hdf5 file to a MS 9
2.3 STEP 2... Loading and inspecting the data 10
2.4 STEP 3... Flagging 16
2.5 STEP 4... Calibration 22
2.6 STEP 5... Splitting the data 36
2.7 STEP 6... CLEAN aka Deconvolution 38
3 Spectral line calibration and imaging 43

3.1 Inspection and flagging 43
3.2 Set up models for the calibrators 45
3.3 Calibration 46
3.4 Velocity rest frames 53
3.5 Continuum subtraction 57
3.6 Finally, on to imaging 59
4
4 Troubleshooting and Support 63

4.1 Getting help 63
Bibliography 65
1
Why do you want to learn CASA?
For quite some decades, radio astronomy data reduction has been per-
formed using AIPS, Miriad and other packages. The coming online
of new and substantially enhanced radio interferometers like ALMA,
LOFAR and the JVLA has driven the development of a new software
package for interferometric data named CASA. CASA is better suited
to the complexity and data-volume from these instruments than previ-
ous existing packages. CASA was developed by NRAO, and was built
on the Aips++ codebase. It seems to be becoming the software suite of
choice for radio interferometric data. KAT-7 is a small interferometer
array and this tutorial presents a guide to reducing the Karoo Array
Telescope (KAT-7) data using CASA.
Producing an image from interferometer data is not straightfor-
ward: an interferometer does not form an image in the way an optical
telescope does, but rather the observations are made in a spatial fre-
quency space, called the u-v or visibility plane, which is essentially a
Fourier transform of the image plane. So, in order to make an image,
the interferometer measurements must be Fourier transformed.
Due to the nature of the interferometer, the (u, v) plane is not fully Figure 1.1: One of the KAT-7 dishes.
sampled. The presence of unsampled regions in the (u, v) causes ar-

tifacts to appear in the image plane. The actual measurements made
by the interferometer are called visibility measurements. In particu-
lar, the geometry of the telescopes in relation to the direction to the
source determines the instantaneous u-v coverage of an interferome-
ter. Often the source is observed for some time, and as the earth and
the telescopes on it rotate with respect to the source, the pattern of
(u, v) coverage also rotates, allowing more parts of the (u, v) plane to
be sampled. The Fourier transform of this sampling function is the
point-spread function of the instrument, called the “dirty beam” in
radio interferometry.
An interferometer is inherently a somewhat complex instrument,
and modern radio interferometers all produce multi-layered data sets
which have visibility measurements from all the different baselines
6 introduction to casa: a kat-7 data reduction guide
forming the interferometer, possibly of several different polarization

products, and in addition, they often have a large number of indepen-
dent frequency channels. Depending in the science aims, these inde-
pendent channels can be combined to form a single-channel image, or
they can be used to produce a multi-channel image cube.
However, before the visibility measurements can be sensibly Fourier
transformed to the image plane, they must be calibrated. Each tele-
scope has a complex gain which is not known a priori, but must be
determined from the observations. Often specific sources with known
properties, called calibrator sources, are observed with the express
purpose of determining these gain factors. The visibility measure-
ments on any baseline are affected by the complex gains of both tele-
scopes making up that baseline. The first stage of data processing in
aperture synthesis is to determine the amplitude gain and the phase-
shift due to each individual telescope (and the atmosphere above it),
which together form the complex gain for that telescope. Although
an interferometer is designed to keep these gains as stable as possible,
they do vary with time, in particular the phase part.
Data sets often also contain some bad or corrupted data, most com-
monly due to radio frequency interference (RFI1 ) or to malfunction of 1
http://pos.sissa.it/archive/
some part of the antenna electronics. This bad data should be re- conferences/107/001/RFI2010_001.
pdf
moved, and the process of removing them is known as flagging. If the
calibration is to be determined from observations of calibrator sources,
then any bad data for the calibrator sources should ideally be removed
before determining the calibration solutions. For other sources it is
generally easier to identify bad data after the calibration is applied.
Once the visibility data are properly calibrated and edited, they can
then be Fourier transformed to make an image. However, due to the
incomplete sampling in the image plane, the image so produced is the
convolution of the sky brightness with the point spread function of the
interferometer, which is called the “dirty beam”. The second part of
the data reduction is to try and correct the image, to the extent possi-
ble, for the effects of this incomplete sampling. Unlike data at some
other wavelengths, radio data can require hours to days of reduction to
achieve the highest dynamic range. Ideally, the image meets the con-
straints of physical consistency as well as plausibility. This document
will attempt to explain the various stages for getting radio images from
KAT-7 data.
1.1 How to use this document
This document assumes that you are conversant with Linux and Python.
If not, then the book “SAMS teach yourself Linux in 10 minutes”
will be a good start even though you will take more than 10 min-
why do you want to learn casa? 7
utes to read the whole book. Reading “Python for Dummies” will
also help. We will also not cover all the fundamentals of radio as-
tronomy, but from time to time we shall try to discuss some concepts
before proceeding so that the user is aware of what we are doing.
For a thorough introduction to the fundamentals of radio astronomy,
we recommend the NRAO Essential Radio Astronomy online course
(http://www.cv.nrao.edu/course/astr534/ERA.shtml).
The procedures consist of various steps and each step generally de-
pends on the previous ones. So avoid skipping to the next step if you
have warnings and errors, without clearing them. We will note those
steps that are independent of the preceding ones.
1.2 Conventions used
We have tried to use the following conventions through this document:
• A typewriter font is used for CASA tasknames and variables in

the main text, and for CASA input and outputs given in a red box,
for example:
CASA command:
prefix = ’CirX1’
msfile = prefix+’.ms’
• The Return key is synonymous with the ENTER key or ←-
Tips: lead to shortcuts that we find useful
Cautions: Avoid pitfalls
Comments: explain the relevant procedure

Pause: Time for a break
1.3 Methodology
In this document we will review the data reduction procedures that

we will use during the in-house workshop on CASA.
Figure 1.2: This mindmap shows the

routes for getting your data into CASA
Before you begin in this long journey of data reduction, make sure
you have enough time to spend. It can get very crazy; get lot of snacks,
coffee and put THE sign (Figure 1.3) in a place clearly visible to the
rest of the world. Basically we expect for a newbie to spend around
3-5 days to successfully reduce a KAT-7 dataset. However, this docu-
mentation will allow you to get a clean image in a day, the target is
within 10 hours of CONTINUOUS work.
It is not that hard but requires lot of attention and concentration.
The process will become quicker after you’ve worked with a few data
sets and have become familiar with it.
OK let us get serious now.
Python is case sensitive (as is Linux). Always ensure you using

the correct case.
Figure 1.3: THE sign

2
Let us get serious
2.1 Overview of Calibration
The visibility measurement made by the interferometer are complex

values, having either real and imaginary parts, or an amplitude and
a phase. The calibration process consists mostly of finding a set of
time and frequency dependent complex gain factors, by which our
visibilities need to be multiplied.
We develop our complex gain solution in stages. Some parts vary
with time but not so much with frequency, for example transmis-
sion through the atmosphere causes both phase shifts and attenuation.
However, the variation of the atmospheric part of the complex gain
across the observing band is usually small. Therefore, a combined so-
lution for all the different frequency channels of observations is usually
possible. The bandpass response of the instrument, by contrast, varies
across the frequency band, but is usually quite stable in time, so that
it needs to be solved for only once in an observing run.
So our whole set of complex gain factors is usually broken up into
several pieces, each of which is solved for separately. This division
is not necessarily unique, in the end all that matters is the complex
product of the individual parts of the calibration.
2.2 STEP 1... Converting the hdf5 file to a MS
Note that for the data set we are using for the tutorial, this step has
already been done, so you can skip ahead to the next section. If you
need to do this step, go to the KAT-7 archive to search for your file.
Then check the eLog to see if the telescope operator entered any infor-
mation during the observations, and take note of any if they did. Then,
proceed to converting the archive file to a Measurement Set (ms). An
ms is the form in which CASA will work on your data. Also remem-
ber at the end of the data reduction to add an entry in the eLog of
the results. Convert the hdf5 archive file to an ms using the in-house
script h5toms as follows:
linux command:
h5toms.py -f --no-auto -C 0 199,7990 --flags=0 all0 -r ant5 myfile.h5
The -f flag causes h5toms.py to produce a full-polarization output

file, while the no-auto causes h5toms to load only cross-correlations,
but not the auto-correlations. The -C 199,799 flag causes h5toms to
load only channels 199 to 799 (inclusive) of the original 1024 channels.
The ∼200 channels at either edge of the band are known to be bad,
so we do not load them. We have also applied the online flagging
using the --flags=’all’. The -r ant5 causes it to use antenna 5 as
the reference antenna (although you can change this later if required).
The output will be a directory myfile.full_pol.ms. This direc-
tory is your CASA measurement set. More options are available for
h5toms.py1 . Once the data has been successfully converted into an 1
h5toms.py -h lists the additional op-
ms, one can start CASA from the computer kat-imager2 (assuming tions e.g. for h5toms
you are running on the SKA Pinelands server), and then proceed with
the data reduction.
It’s probably a good idea to rename your ms to something a little
more descriptive at this stage.
linux command:
mv myfile.full_pol.ms School_data_av.ms
.
Of course if you’re looking at real data its probably a good idea to
use a name for the ms which reflects what the observations are.
2.3 STEP 2... Loading and inspecting the data
It is probably best to create a specific directory which will contain the

working CASA data. We will use the directory “my_working_directory”.
linux command:
mkdir my_working_directory
let us get serious 11
The first main step is to move the ms into your chosen working
directory, and then also to make the working directory your current
directory. For the workshop, you may have downloaded the file
School_data_av.ms.tar.gz, which you would unpack with the linux
command:
linux command:
tar -zxf School_data_av.ms.tar.gz
It should produce a directory called School_data_av.ms. This di-

rectory is your CASA measurement set.
Locate the ms directory from the previous step and move it into
the working directory.
mv myfile.full_pol.ms my_working_directory
cd my_working_directory
Although CASA measurements sets (ms) are ordinary linux di-

rectories, and can be renamed, moved, delected or copied like
any other directories, there are probably restrictions you should
observe:
• Do not move, delete etc. directories from the oper-

ating system while casapy is running, as casapy does
some ms caching, and will likely get confused. You
can remove an ms from within casapy by typeing e.g.,
rmtables(’crappy-data.ms’) and it will be fine, or you can
just exit casapy and then delete it.
• Do not alter the contents of the ms. While renaming or mov-

ing the ms directory is fine, its probably not a good idea to
add or delete any files to the ms, or to alter or rename any
of the files contined within it.
Most of the data reduction will be done from within casapy, which
provides an interactive shell (essentially the equivalent of interactive
python or ipython). From the Linux prompt you will therefore first
run casapy. You will then give the commands to run various CASA
programs from within casapy. In casapy you can also create vari-
ables, assign values to them, and then subsequently use them at any
time during the processing. It is generally convenient to define some
variables to hold the names of ms’s and tables. The sequence of cas-
apy commands can then be re-used for a different data set by merely
changing the values assigned to these variables.
Do NOT abort running programs by pressing Ctrl+C as CASA

has a bug and will screw up all your visibilities. Rather wait
for the command to be over and then undo the wrong steps.
In fact, CASA is designed to be highly script-able, and sequences of

commands could be placed in a script and run automatically. In this
session, however, we shall be using step-by-step commands.
Never delete the ms file unless you have finished splitting your
relevant data and are sure you do not need the ms anymore.
Below we define a variable to hold the ms (directory) name but

without the “ms” extension. We will use this variable to refer to the
ms and to various associated tables. The ms we are using in the tuto-
rial is called Cirx1_school.ms, so we define the following variables in
casapy:
CASA command:
prefix = ’Cirx1_school’
We now define variables that hold the names of various tables that
will be used during the reduction. The table names could be specified
directly, but if you do it by means of variables, the subsequent com-
mands we use can be copied and pasted and re-used for some other
reduction with the only modification required being the re-definition
of these variables.
CASA command:
bandpass_table0 = prefix +’_spw0.B0’

gain_table0 = prefix+’_spw0.G0’
gain_table1 = prefix+’_spw0.G1’
flux_table1 = prefix+’_spw0.fluxscale1’
We chose antenna ant5 as our reference antenna when we first con-

verted the data-set to an ms, because it seems to have more shielding
for RFI from the pedestal. Since it is best to use a consistent refer-
ence antenna throughout, we again define a variable whose value is
the current reference antenna.
CASA command:
reference_antenna= ’ant5’
Next, we use the task listobs to print out a summary of the obser-
vations in the ms (similar to the AIPS task LISTR with OPTY=’SCAN’):
CASA command:
listobs(vis=msfile)
CASA output:
(we have removed the leading part of the lines of the listobs output for clarity.)
================================================================================
MeasurementSet Name: /home/michael/casa-workshop2012/CirX1.ms MS Version 2
================================================================================
Observer: lindsay Project: 20120701-0006
Observation: KAT-7
Data records: 71904 Total integration time = 56296.5 seconds
Observed from 01-Jul-2012/13:57:43.4 to 02-Jul-2012/05:35:59.9 (UTC)
ObservationID = 0 ArrayID = 0
Date Timerange (UTC) Scan FldId FieldName nRows Int(s) SpwIds
01-Jul-2012/13:57:43.4 - 13:59:27.9 1 0 PKS 1934-638 168 14.1 [0]
14:00:13.4 - 14:04:58.4 2 1 Circinus X-1 420 14.8 [0]
14:05:28.4 - 14:06:21.4 3 2 PKS 1613-586 105 11.6 [0]
14:06:44.5 - 14:11:30.5 4 1 Circinus X-1 420 14.8 [0]
14:12:00.5 - 14:12:45.5 5 2 PKS 1613-586 84 14.5 [0]
. . . (we omit the bulk of the detailed scan listing here)
05:29:53.8 - 05:30:38.9 328 2 PKS 1613-586 84 15 [0]

05:30:56.9 - 05:31:41.4 329 2 PKS 1613-586 84 14.5 [0]
05:31:58.9 - 05:32:43.9 330 2 PKS 1613-586 84 15 [0]
05:33:01.9 - 05:33:46.9 331 2 PKS 1613-586 84 14.8 [0]
05:34:04.9 - 05:34:49.4 332 2 PKS 1613-586 84 14.8 [0]
05:35:06.9 - 05:35:59.9 333 2 PKS 1613-586 105 12 [0]
(nRows = Total number of rows per scan)
Fields: 3
ID Code Name RA Decl Epoch nRows
0 T PKS 1934-638 19:39:25.03058 -63.42.45.6999 J2000 2709
1 T Circinus X-1 15:20:40.85208 -57.10.00.1048 J2000 52542
2 T PKS 1613-586 16:17:17.89966 -58.48.07.8890 J2000 16653
Spectral Windows: (1 unique spectral windows and 1 unique polarization setups)
SpwID #Chans Frame Ch1(MHz) ChanWid(kHz) TotBW(kHz) Corrs
0 19 REST 1938.21094 -12500 234765.625 XX XY YX YY
The SOURCE table is empty: see the FIELD table
Antennas: 7:
ID Name Station Diam. Long. Lat.
0 ant1 ant1 12.0 m +021.24.39.4 -30.33.10.2
1 ant2 ant2 12.0 m +021.24.41.9 -30.33.09.1
2 ant3 ant3 12.0 m +021.24.38.6 -30.33.09.1
3 ant4 ant4 12.0 m +021.24.37.7 -30.33.09.1
4 ant5 ant5 12.0 m +021.24.37.1 -30.33.10.0
5 ant6 ant6 12.0 m +021.24.36.2 -30.33.12.5
6 ant7 ant7 12.0 m +021.24.35.2 -30.33.07.5
It is a good idea to save the listobs output for future reference,

as it gives the basic structure of the observations, indicating for exam-
ple which source was observed at what time. You can either copy the
listobs output from the casapy log file, or rerun listobs with extra
parameter listfile=[insert the name of the textfile you want to save
the output here].
What can you deduce from this listing of the data set?
• The observations started at 14:00 on 01-Jul-2012 and continued till

04:15 on 02 Jul 2012 (UTC times)
• Three sources were observed, with field names “PKS 1934-638”,

“Circinus X-1” and “PKS 1613-586” and source IDs 0, 1, and 2 re-
spectively.
• The reference frequency 1938.21094 MHz; One spectral window,

SPW, was used, with SpwID = 0. This SPW had 19 frequency chan-
nels (note that this data-set has already been averaged in frequency,
raw data KAT-7 data sets typically have 1024 channels
• The number of antennas taking part was 7
The science target in these observations was Circinus X-1 (ID=1),

with PKS 1934-638 and PKS 1613-586 (ID 0 and 2 respectively) being
the calibrators. The source PKS 1934-638 will be used as a flux density
calibrator, while PKS 1613-586 will be used to calibrate the antenna
phases.
In CASA, it is the field parameter which is generally used to spec-
ify the source, and one can use either the source name (e.g., field =
’Circinus X-1’) or the Source ID (e.g., field = ’2’) to specify a par-
ticular source. Note that the value of field is a string, and therefore
enclosed in quotes (’ ’) even when it is specifying the numeric source
ID.
CASA can automatically set the flux densities for the flux-density
calibrator sources, but they need to have field names in the correct for-
mat for this to work. In this case, our flux-density calibrator source
has the form of the name recognized by CASA “PKS 1939-638”. How-
ever, if your flux-density calibrator source does not have a name of
the recognized form, for example, “3C 48”, then you need to use
browsetable to replace it with, e.g., “3C48”. In a future version of
the h5toms script this will be fixed automatically. browsetable will
produce a GUI where you can edit the ms tables. Check the link
http://casa.nrao.edu/docs/UserMan/UserMansu193.html for details.
It is good practice to plot the antenna positions and check that ev-
erything being read properly and your antennas are where they should
be. This is done using:
CASA command:
plotants(vis=msfile)
From the above antenna plots, we can see that ant4 was not used in
this observation and also that the plots show the expected positions of
the KAT-7 antennas.
2.4 STEP 3... Flagging
The majority of interferometric data sets contain some corrupted data,

with the most common reason being failures of an antenna or some of
its associated electronics and radio frequency interference (RFI). The
process of editing out this data is generally called “flagging”, and is
usually the most boring and tedious part of the data reduction process.
Flagging is somewhat of an art and there are no hard rules. However,
most modern instruments, KAT-7 in particular, produce fairly clean
data (at least at frequencies above 1 GHz), so once you have removed
the data from any malfunctioning antennas, you should not have to
flag any more than 10% of your data, and often much less than that.
If possible you should flag either all visibilities for some time interval,
or all by antenna, rather than flagging individual baselines. The rea-
son for this is twofold: firstly, its more efficient, and secondly its less
biased. Although different interactive flagging tools are available, for
the initial flagging the best approach is usually to first make several
passes through the data, using various plotting and listing tools, to
identify the bad data. Once you have identified some sections of bad
data, you then flag them explicitly using flagdata. This allows you
to keep a detailed record of what was flagged, and to easily redo the
flagging if required. Going through your visibility data set point by
point and picking out individual points is rarely productive.
The following tasks are available for flagging data in CASA:
• flagmanager: save and manage versions of the flagging entries in

the ms
• flagautocorr: noninteractive flagging of auto-correlations
• plotms: interactive XY plotting and flagging of visibility data
• plotxy: interactive XY plotting and flagging of visibility data - note

that plotxy is slower than plotms and will eventually be phased out
• viewer: can display (as a raster image) ms data, with some editing
capabilities
• flagdata: This is the new non-interactive flagging (and unflagging)

of specified data in CASA 3.4 onwards. It has other flagging options
such as antenna shadowing, elevation flagging, etc.
We will mostly use the task plotms to examine the data. It is a

very flexible tool that allows you to interactively plot your visibility
data in a wide variety of ways. We will use it to identify the bad
data. We will then use the task flagdata to actually flag it. Although
data can be flagged directly in plotms, we suggest that it is better to
actually identify which antennas, baselines and times are responsible
for bad data before flagging them until you have some experience.
Flagging data only on the basis of e.g., low visibility amplitude can
lead to serious biases: if you expect a flux density of 1 Jy and you
flag all the visibilities outside the range 0.9 to 1.1 Jy, you would likely
find the remaining visibilities suggested a source of 1 Jy. However,
you’re not learning about the sky, you’re learning about the flagging
you already did!
Calibrator sources are much easier to flag than program sources,
because we largely know what kind of source they are, and therefore
what to expect from the visibility measurements. In particular, calibra-
tors are generally point-like, which means the true visibility amplitude
is independent of baseline (and of time), and the phase should be 0◦ .
Unfortunately, all we have available at this point is the uncalibrated vis-
ibilities. However, for an instrument like KAT-7, where all the dishes,
and thus all the baselines have approximately equal sensitivity, even
for the uncalibrated data for a point-like source, we can say this: the
(uncalibrated) visibility amplitude on all baselines should be approxi-
mately the same, say within a factor of 2 or so, and the (uncalibrated)
visibility phases should be fairly stable, although they do not need to
be near 0◦ .
The two most common failures to identify are probably “dead” or
malfunctioning antennas, and channels which are corrupted by RFI,
which is usually very narrow bandwidth. Even though automatic flag-
ging has already been done by h5toms.py, and may well have removed
most of the strong RFI, you should nonetheless always examine your
data.
However, first let us remove some bits data which is known to be
bad. The KAT-7 correlator is known to produce some visibilities with
spurious amplitudes of exactly 0. We will use flagdata to delete these.
We will also use a second invocation of the flagdata to flag any data
taken at antenna elevations below 10◦ . Data taken at very low antenna
elevations is often unreliable and noisy, so unless much of your data is
taken at low elevations, its often easier to just delete the small fraction
with antenna elevations below about 10◦ .
CASA command:
flagdata(vis=msfile,mode=’clip’,field =’’,
clipzeros=True, flagbackup = False)
flagdata(vis=msfile,mode=’elevation’,
lowerlimit=10.0, flagbackup=True)
Note that in “elevation” mode, the elevation limits provided

are those of the good data, rather than those of the data to be
flagged, so in this case any data below an elevation of 10◦ will
be flagged. We do not need to set an upper elevation limit.
Now lets plot the remaining visibilities. Phase stability, both in fre-
quency and time, is probably the most important thing to check first:
unless the visibility phases for the calibrator sources are reasonably
stable in time and across the different channels, calibration will likely
prove impossible.
Each visibility point is associated with one baseline, that is some
pair of antennas. However, instrumental failures are almost always
antenna-dependent, that is, due to a failure of some particular antenna,
and therefore all the baselines involving the bad antenna will have bad
visibility data. When identifying data to flag, it is best to try first to
establish whether it is some antenna that is bad, and if so flag all the
data to that antenna. Some bad data, however, in particular that due
to RFI, may not be antenna-dependent, and may have to be flagged
individually per baseline.
For starters, let us look at the phase as a function of frequency
on a per-baseline basis. We will start with just one example scan of
our strongest calibrator, PKS 1934-638, pick one scan from the listing
above, somewhere near the middle of the run. In this case we will pick
scan ID = 57, occurring between times 17:01:16.3 and 17:03:01.3.
Before you carry on, do spend some time getting to know the GUI
behind plotms and how to use it:
http://casaguides.nrao.edu/index.php?title=Data_flagging_with_plotms.
For axis definition in plotms check this link :
http://casaguides.nrao.edu/index.php?title=What%27s_the_difference_
between_Antenna1_and_Antenna2%3F_Axis_definitions_in_plotms.
Now actually run plotms to bring up the plot window and get a first
look at part of our visibility data. Note that the iteraxis=’baseline’
tells plotms to plot each baseline individually. Pressing the green
arrow at the bottom of the plotms interface will allow you to step
through the individual baseline plots interactively.
CASA command:
plotms(vis=msfile, xaxis=’channel’, yaxis=’phase’,

correlation=’XX,YY’, scan=’57’, field=’PKS 1934-638’,
iteraxis=’baseline’, coloraxis = ’corr’,
plotrange=[0,0,-180,180])
Your screen will look similar to this:
Figure 2.1: The output of plotms show-

ing phase versus channel for one base-
The two different colours represent the two correlations or polar- line.
ization products (XX,YY) which we have plotted. Note that the purple
points, which correspond to XX, hang together well, sloping slightly
from channel 0 to channel 19. They are fine. The tan-coloured points
for the other polarization, however, are scattered randomly, and will
not be usable. This plot is for the baseline between Ant1 and Ant2.
Hitting the right arrow button at the bottom will show the plot for
the next baseline Ant1, Ant3, where both polarizations look fine. The
next plot, Ant1, Ant4 again shows the XX points are fine, while the YY
points show strong variation across the band.
These effects are almost certainly due to an antenna, rather than a
single baseline, and the trick is to identify which of the antennas is
the one that is affected. By stepping through the baseline you will see
that the YY phases for all baselines involving ant2 look quite random,
while those involving ant4 are somewhat less random, but nonetheless
show large variation from channel to channel. In other words, at least
for this scan, there is a problem with the phase stability for the YY
polarization of Ant2 and Ant4.

Let us check for phase stability against time. We will only plot
one channel by specifying spw=’*:9’ for this purpose. (Seeing as the
phase will vary both with time and channel, its hard to determine the
phase stability by looking at all channels and times simultaneously,
that’s why we’ve taken the approach of first taking a “slice” in time,
and looking at only one scan but all channels. Now we take a “slice”
in frequency, and look at only one channel but at all times).
CASA command:
plotms(vis=msfile, xaxis=’time’, yaxis=’phase’,

correlation=’XX,YY’, spw=’*:9’, field=’PKS 1934-638’,
iteraxis=’baseline’, coloraxis = ’corr’,
plotrange=[0,0,-180,180])
Again you will see a similar pattern. The visibility phases for XX are
quite stable in time, showing only small variations from scan to scan
and within each scan (mostly, the ∼10 or so visibility points within
a scan are so close together that the whole scan appears as a single
point in the plot). However, again, the visibility phases for the YY
polarization (tan colour) for ant2 and ant4 are quite unstable.
Since we have shown that the visibility phases for YY, ant2 and
ant4 are quite unstable both in time and across channels, we will
delete these points. Note, however, that so far, we have only examined
the data for one source, namely PKS 1934-638. Perhaps the data for
the other sources are fine! Before throwing out all the data, its worth
checking. In the plotms window, change “field” to PKS 1613-586 and
hit “plot”. There are more data for this source, but you will see the
exact same pattern. Since the data for YY, ant2 and ant4 are bad for
both our calibrator sources, we have no choice but to flag all these
data, including that for our target source Circinus X-1. Very likely
the target source data would also be bad, but in any case it cannot be
calibrated so we cannot use it. You could plot also the visibility data
for Circinus X-1, however since its not known to be a compact source
we do not know what to expect from the visibility phases. They may
vary strongly because the source is complex. It is only for calibrator
sources, where the actually visibility phase should be constant that we
can judge stability by plotting the visibility phases as we have done.
At present CASA has the limitation that it cannot calibrate one of
the XX, YY polarizations when the second one is missing, so we will ac-
tually have to flag all the polarizations for ant2 and ant4, even though
the XX polarization data does not have any problem. (There are four
total correlations, XX, YY, XY, YX. We will deal only with first two,
whose sum represents the unpolarized intensity. The calibration of the
polarized parts, XY,YX involves extra steps, and is not yet implemented
in CASA for linearly polarized feeds like those of KAT-7). Having
identified the bad data, we once again turn to flagdata to actually
flag it. We do not specify either field or correlation so flagdata
flags the requested data for all fields and correlations.
CASA command:
flagdata(vis=msfile, antenna=’ant2,ant4’,
flagbackup=True)
Referring to an antenna can be kind of confusing in CASA.

Like sources, CASA can refer to antennas by their IDs or num-
bers, or by their names. The confusing thing is that the antenna
names often look very similar to IDs, with KAT antennas hav-
ing names like, e.g., “ant1”. However, the antenna IDs start
with 0, while the names start with the number 1. You have to
keep track of whether you are referring to the antenna by ID or
by name. In this case, the antenna ID=0 has the name “ant1”
(as can be seen from the listobs output). So it is antennas ant2
and ant4, also known as ID = 1 and 3, which have the bad data
in this case.
If you now repeat the above plotms commands, you will see that the
visibility phases are now nicely coherent for all the plotted baselines,
and the baselines we have flagged no longer show up.
Now lets turn to the visibility amplitudes. Again we make use of
the fact that for continuum calibrator sources, the expected values are
known and should be the same on all baselines, and vary only slightly
across the channels. This time we can plot all the baselines together,
rather than iterating through them as we did before:
CASA command:
plotms(vis=msfile, xaxis=’channel’, yaxis=’amp’,

correlation=’XX,YY’, field=’PKS 1934-638’,
coloraxis=’corr’)
You will see that for most channels there is only a relatively small
variation of the visibility amplitude with channel for PKS 1934-638
(note the amplitude scale on the left). This small variation will be
corrected later with bandpass. However, channel ID=6 shows a much
larger scatter than any of the others, which is indicative that there may
be RFI in this channel. Switching to the other calibrator source shows
the same pattern. It is sometimes diagnostic to also examine the XY,
YX polarizations here. For calibrator sources, we expect the polarized
flux density to be a relatively small fraction of the unpolarized (i.e.,
XX, YY) one, whereas RFI is often strongly polarized. We conclude
that channel ID=6 is likely affected by RFI, and therefore we will pro-
ceed to flag it also, once again using flagdata. Once again, we will
flag that channel also for the target source data, since even if the tar-
get source data were not also affected by the RFI, the data would be
uncalibratable.
CASA command:
flagdata(vis=msfile, spw=’0:6’, flagbackup=True)
We have now flagged most of the bad data affecting this run, in
particular the bad data for the calibrator sources, and can now proceed
to the actual calibration.
2.5 STEP 4... Calibration
In order to explain calibration, we must first explain briefly the nature

of the visibility measurements we are trying to calibrate. If the field of
view is small enough so that it can be represented as a plane (rather
than as a curved part of the celestial sphere), then the sky brightness
can be written as I (l, m), where l and m are the two orthogonal co-
ordinates (typically RA and decl). The visibility function measured
by an ideal interferometer, V, is given by a two-dimensional Fourier
transform of this sky brightness:
Z Z
V (u, v) = I (l, m)e j2π (ul +vm) dldm (2.1)
where u, v represent the coordinates in the Fourier transform plane.

At any given instant, each baseline, or pair of telescopes, in the in-
terferometer measures this visibility function at a point in the Fourier
transform plane whose u, v coordinates are determined by the length
and orientation of the baseline, as projected onto the sky plane. Since
the telescopes rotate with the earth, each baseline typically sweeps out
an elliptical path in the u, v plane as the earth rotates. Although I is
real, V (u, v) is generally complex, with non-zero values for both real
and imaginary parts. V (u, v) is often described as an amplitude and a
phase, rather than by its real and complex parts. Note that V (u, v) is
symmetric with V (u, v) = V (−u, −v): generally the Fourier transform
of a real function, like I, is complex but symmetric.
The correlator output differs from the V measured by an ideal in-
terferometer for a variety of reasons, to do with both instrumental
effects as well as propagation effects in the earth’s atmosphere and
ionosphere2 . The relationship between the observed and the true vis- 2
At low frequencies, such as those ob-
ibility on the baseline between two telescopes, i and j, can be written served by KAT-7, the effect of the iono-
sphere is usually the dominant one
in a very general way as:
Vijobserved = Jij Vijtrue (2.2)

Jij represents the accumulation of all corruptions affecting the visi-
bility measurement on baseline ij. Both V and J have complex values
and are generally functions of frequency, polarization and time, but we
have suppressed the additional indices pertaining to frequency, polar-
ization and time for clarity.
Calibration involves determining the values of Jij , in order to invert
the above equation and to determine the best approximation possible
of Vijtrue from the observed values of Vijobserved .
This task of determining Jij is made considerably easier by the fact
that most of the effects contained in Jij are antenna-based. In other
words, Jij can be factored into Ji ⊗ Jj∗ (where Jj∗ is the complex conju-
gate of Jj . For an array consisting of Nant antennas there are therefore
only Nant unknown values of Ji which need to be determined, rather
than the approximately Nant 2 values of J which would be required
ij
otherwise. Bear in mind that we have suppressed the time and fre-
quency dependence here, so Ji is not a single complex number, but
a function describing the complex response of antenna i in both time
and frequency over the time-range and bandwidth of the observations.
The factorization occurs because the bulk of Jij has its origin in ef-
fects which corrupt the radio signal as it travels through the atmo-
sphere and the antenna electronics, these effects are therefore due to
an individual antenna i, but do not depend on the other antenna j
which makes up that particular baseline. Although there are baseline-
dependent effects, they are mostly very small, and for most purposes,
can be ignored.
It is practical to further factor the Ji into different components, rep-
resenting different factors affecting the radio signal, each of which can
be determined independently. Some of the components, for exam-
ple, are time-invariant, and therefore need to be determined only once
for the observing run, while others are time-variable. In our calibra-
tion process, we will first factor Jij into the antenna-dependent Ji and
thenceforth into a series of different factors, and each factor will be de-
termined in a different step. Our approach is similar to the so-called
“Measurement Equation”, which is a more formal approach to factor-
ing Jij described in [Sault and Cornwell, 1999], but we factor Jij into
slightly different components.
So we can write:
Vijobserved = Jij Vijtrue = ( Ji ⊗ Jj∗ )Vijtrue (2.3)
where
Ji = Ai Bi Gi . . . (2.4)
For example, Ai might be corrections for some effect known a priori,

such as an error in one of the antenna positions used in the correlator,
and Bi might be to describe the bandpass response of the antennas, in
other words the variation in sensitivity and phase between the differ-
ent channels, which is taken to not vary with time, while Gi provides
the variation in amplitude and phase response as a function of time,
but taken not to be a function of frequency, i.e., determined for all
frequency channels simultaneously. Note that the separation of these
effects is not unique, however, we want to ensure that the final product
Ji must be correct.
Some parts of Ji can be calculated without any reference to the vis-
ibility data. However, most of Ji will be determined from the values
of V observed for calibrator sources, where we know what the values of
V true should be and can therefore invert the equation above. In par-
ticular, most of the parts of Ji due to the atmosphere and ionosphere
cannot be known in advance. The values of Ji (recall that Ji in gen-
eral depend on time, but we have suppressed this dependence in the
notation), determined from the observations of calibrator sources are
then interpolated (or extrapolated) to the times corresponding to the
observations of the intended target source(s). The various calibration
programs perform the function of estimating the antenna-based cali-
bration factors (e.g., Ai , Bi etc., from above) from the (baseline-based)
calibrator source visibility measurements.
For example, the antenna-based bandpass response, Bi is typically
treated as being independent of time (although it is by definition a
function of observing frequency). It is typically determined by exam-
ining the visibilities for a strong calibrator source of known spectrum
(the simplest case would be a source with a flat spectrum, with true
flux density independent of the particular frequency or channel under
consideration). Often this is the first step in the calibration, since once
this bandpass correction is known, we can use an average across the
channels to determine the remaining parts of Ji , rather than having to
do it for each channel independently.
In order to use the visibility data from calibrator sources to calculate

calibration quantities, such as the bandpass response, one should edit
out the bad visibility data first. This is why we started above with
examining the data and flagging bad points or antennas.
The typical sequence is to first use observations of a strong cal-
ibrator, often the flux-density calibrator, to determine the bandpass
response. With that in hand, we no longer have to make separate so-
lutions for every channel, but can combine them. One then uses the
observations of calibrator sources to determine the variation of Ji with
time. Many calibrator sources are variable on timescales of weeks or
longer and their flux densities are therefore not accurately known. In
order to set the flux density scale correctly, one needs observations of
at least one flux density calibrator source, whose flux density is known.
Typically, one uses one or a few scans of a flux density calibrator source
for this purpose, and then the remainder of the Ji can be obtained from
other calibrator sources whose flux density is not known (although it
should not vary during the observing run), but which are known to be
point-like. These secondary calibrators are known as phase calibrators
or secondary amplitude calibrators.
So, lets begin the calibration. For safety and sanity, one should be-
gin by “clearing the calibration”. The data structure of CASA measure-
ment sets (ms) is that the observed data are stored in a DATA column of
the ms, while the corrected data (i.e., those multiplied by Jij ) are stored
in the CORRECTED_DATA column once they become available. The val-
ues from which Jij are constructed are stored in various calibration
tables. There is a further column of visibility data called MODEL_DATA
where model values, if available, such as the expected visibility val-
ues for a calibrator of known structure, are stored. The task clearcal
initializes the CORRECTED_DATA and MODEL_DATA columns and therefore
clears calibrated data from any previous attempts at calibration from
them.
CASA command:
clearcal(vis=msfile,field=’’)
Time to take a break and have some coffee... The steps after will
need more attention.
Figure 2.2: A schematic diagram of the

calibration process.
There are few sources whose flux density is known reliably which
can be used as primary amplitude calibrators. The task setjy knows
about the most common of these, and will insert the correct flux den-
sity for the frequency of the observations as well as entering the model
into the MODEL_DATA column for subsequent use.
These primary flux density calibrators are often somewhat re-

solved, and thus source models should be used to represent the
source structure. The task setjy can be invoked to list the avail-
able source models (with listmodels=True). However, the flux
density calibrator for our observations is PKS 1939-638, which
is one that is close to ideal in that it has no structure on scales
larger than 100 . The default point model is therefore adequate
and no further source model is required in our case.
We still have to enter the actual flux density for PKS 1939-638 for
the actual frequency of the observations. In general the flux density
varies with frequency, and setjy will calculate a slightly different
value for each channel from the known spectrum. (In the case that
a model image is used for the flux-density calibrator, setjy will also
scale the model image to the total flux density for each channel in the
ms, and then enter the Fourier transform of the scaled model image
into the MODEL_DATA column for that source and channel.) Setting
fluxdensity = -1 causes setjy to look up the flux density for flux

density calibrators and enter the standard value.
CASA command:
setjy(vis=msfile, field=’PKS 1934-638’,

scalebychan=True, standard=’Perley-Butler 2010’,
fluxdensity=-1)
setjy should produce the following output in the logger, from

which one can see that it has identified calibrator source PKS 1934-638
and filled in the known flux density of ∼13 Jy in I (the exact flux den-
sity will vary somewhat and be slightly different in the different chan-
nels, with the listed value being that for the channel at the reference
frequency of 1932 MHz). If you see the message that your source “is
not recognized by Perley-Butler 2010”, then setjy will have en-
tered a default flux density of 1.0 Jy for your flux calibrator and you
have to fix the problem otherwise your flux density scale could be
wrong by a large factor.
CASA output:
setjy::::casa ##### Begin Task: setjy #####

......
setjy::imager::setjy() Using channel dependent flux densities
setjy::imager::data selection Selected 2709 out of 71904 rows.
setjy::imager::setjy() PKS 1934-638 (fld ind 0) spw 0 [I=13.182, Q=0, U=0, V=0] Jy, (Perley-Butler 2010)
....
Once we are happy that the correct flux density has been entered
for the flux-density calibrator, we can proceed to the next step, which
is to solve for the bandpass response, in other words for the variation
of the complex antenna gains across the different channels. (Note that
setjy above did not yet solve for any part of Ji , it only entered the
flux density of our calibrator, essentially the value of V true into the
ms so that we can start solving for Ji in the next steps. The variation
of the gains across the observing band is largely due to instrumental
effects and usually quite stable, so we only need to solve for it once.
Once we have solved for the bandpass response, we can combine all
the channels in a single solution for the subsequent steps.
However, we have the problem that the phase part of Ji can be quite
rapidly time-variable even though the amplitude part is not. In order
to properly calculate the bandpass response, we therefore want to do
some temporary phase-calibration for that scan which we intend to
use for the bandpass calibration, to prevent decorrelation when vec-

tor averaging the data in computing the bandpass solutions. In other
words, to obtain the bandpass solution, we will average the complex
visibility data in time over a whole scan. If there is any variation on
phase over the scan, we want to calibrate that out before performing
the bandpass solution.
We do this phase calibration using gaincal, and using only a por-
tion of the band, in this case the six channels from 7 to 11. These
solutions for the antenna phases as a function of time can then be
used to calibrate out any variation in the phase during the scan for the
purposes of the subsequent bandpass solution. The rapid phase varia-
tion we are calibrating out here are due mostly to the atmosphere, and
are not expected to vary significantly between the channels. Therefore,
although these phase solutions are determined from only a central por-
tion of the band, they are applicable to the whole band (i.e., to all 19
channels). The choice of the particular channels (7 to 11) is relatively
arbitrary, but usually its best to use a few channels near the centre of
the band, which were found to be relatively free from interference in
the editing steps earlier.
For this next step, we need to specify a reference antenna. When we
ran h5toms we chose reference antenna 5. However, its possible that
antenna 5 was not present at all during the run, or that we removed
some of its data in flagging. You can use plotms to plot antenna1
against time for all fields to quickly identify which antennas were
present during which parts of the run. In this case you should see
that antenna ID=0 is missing during the first part of the run, because
we flagged it earlier. Antenna ID=5 seems to be present throughout,
and can therefore serve as our reference antenna for the rest of the
calibration. Note that it is possible to use different reference antennas
for different parts of the calibration, but if possible you should use the
same antenna throughout. It is also desirable to choose a reference
antenna which is known to be stable and fairly sensitive (in the case
of KAT-7 the sensitivities of the 7 antennas is very similar, so the only
real criterion for picking the reference antenna is stability).
CASA command:
gaincal(vis=msfile, caltable=gain_table0,
field=’PKS 1934-638’, refant=reference_antenna,
spw=’0:7~11’, calmode=’p’, solint=’int’,
minsnr=4, minblperant=4, solnorm=T, gaintable=’’)
Remember that we defined the variables refant and

gain_table0 earlier to hold the name of the reference antenna
and gain table, respectively, that we’re using for this particular
ms.
After having determined the solutions, we should examine them

using plotcal. The task plotcal can plot a variety of gain solutions
in a variety of ways, check help plotcal to find out more. We remind
you that gaincal has factored Jij into antenna-based components, so
you don’t see a set of values for each baseline, but rather only one for
each antenna (Ji ), and the CASA programs will calculate the full value
of Jij as needed when calibrating the visibility values.
CASA command:
plotcal(caltable=gain_table0, xaxis=’time’,
yaxis=’phase’, field=’PKS 1934-638’,
iteration=’antenna’, plotrange=[0,0,-180,180])
On the plotcal plot, you should see that the gain-table phases (an-
tenna phases, or the phase-component to Ji ) should be fairly stable
during the minute or two of the scan-length. Small variations are fine,
indeed the point of running gaincal was exactly to find these small
variations so that they can subsequently be calibrated out from the
visibilities. However, if the antenna phase appears random from one
point to the next then something has likely gone wrong.
Now we can run bandpass to solve for the complex gain as a func-
tion of channel across the observing band. Specifying gaintable=[gain_table0]
causes bandpass to apply the time-dependent antenna phase calibra-
tion we determined in the previous step before forming the bandpass
solutions.
CASA command:
bandpass(vis=msfile, caltable=bandpass_table0,
field=’PKS 1934-638’, refant=reference_antenna,
solnorm=True, combine=’scan’, solint=’inf’,
bandtype=’B’, gaintable=[gain_table0])
You will see the message:
Insufficient unflagged antennas to proceed with this solve.

(time=2012/07/01/22:08:07.6 field=0 spw=0 chan=6)
Since we flagged channel 6 or spw=0:6 earlier on, this is not some-

thing to worry about, its just telling us that there is no solution for
that particular channel, but that doesn’t matter because we have no
visibility left for that channel. A message like this about some channel
for which there was visibility data left would be a problem.
The task plotcal can also be used to display the bandpass solution:
CASA command:
plotcal(caltable=bandpass_table0, xaxis=’chan’,
yaxis=’amp’, subplot=222, iteration=’antenna’)
First note that plots for the antennas for which the data have been
flagged (ant2, ant4) are blank, which is as expected. For the others
the gain amplitude varies only by 10% to 20% across the channels,
which is in the expected range. Since we asked bandpass to normalize
the solution, the average gain for each antenna is 1.0 — at this point
we’re only after the variation across the band. If any of the gains were
to differ from unity by more than a factor of 2 or so, then may be more
bad data to flag. One would then have to flag that bad data, and then
re-start the calibrations from the clearcal() step.
The next step derive the main, time-variable, part of Ji , which we
term Gi . This is done for each source individually, by comparing the
observed visibilities Vijobserved to the expected visibilities Vijtrue . For the
flux density calibrator, we have determined the latter values, and we
can get the correct values for Gi . For the phase calibrator sources, we
assume a point source model, so we know the phase part of all the Vtrue
is 0. We assume for now that the unknown flux-density of the phase-
calibrator sources is exactly 1.0 Jy (typically they are of this order).
If we then solve for Gi , we will obtain complex values whose ampli-
tudes differ from the correct ones by a factor depending only on the
true source flux density (by ([1 Jy]/[true source flux density])0.5 ). We
can then compare the Gi obtained for the secondary calibrators, and
scale them so that they match as well as possible the correctly-scaled
amplitudes of Gi obtained for the primary flux-density calibrator. The
scaling required gives us the true flux density of the secondary cali-
brator source in terms of the originally assumed one, i.e., 1 Jy. This
process is known as “flux density bootstrapping”.
The reason for this complicated procedure is that there are only
a handful of flux-density calibrators, so it is generally impossible to
choose a flux density calibrator close to the target source on the sky.
There are many phase calibrator sources, however, so a phase cali-
brator can be chosen that is much closer on the sky (typically a few
degrees ◦ away). The flux density bootstrapping allows us to use the
phase and amplitude calibration (i.e. Gi ) from these much nearer cal-
ibrator sources, which will provide a much better calibration for the
target source, but to set the flux density scale accurately from obser-
vations of the flux density calibrator. The flux density scale is usually
stable for hours, so once we have transferred the flux density scale to
the phase calibrator sources by flux density bootstrapping, the solu-
tions we obtain for the phase calibrator sources are equivalent to those
which would have been obtained if we’d known the flux density of the
phase-calibrators to begin with.
First we determine the values of Gi , also called the complex gains,
for the flux-density calibrator:
CASA command:
field=’PKS 1934-638’, solint=’inf’,
refant=reference_antenna, gaintype=’G’,
calmode=’ap’, gaintable=[bandpass_table0])
Note that the a new solution is started for each scan, the
solint=’inf’ only means that the solution interval can be ar-
bitrarily long up to the next scan boundary.
Next we determine the complex gains for the phase calibrator (so
far under the assumption that it has a flux density of 1 Jy).
CASA command:
field=’PKS 1613-586’, solint=’inf’,
refant=reference_antenna, gaintype=’G’, calmode=’ap’,
append=True, gaintable=[bandpass_table0])
Note that the append=True causes gaincal to append to the

specified gain table, rather than over-writing it.
Plotting again to have a look at the solutions:
CASA command:
plotcal(caltable=gain_table1, xaxis=’time’,
yaxis=’amp’, iteration =’antenna’)
Notice that, for each antenna, there are two sets of points. The lower
set, in this case with gain amplitudes near ∼0.1 corresponds to our flux
density calibrator (PKS 1934-638). Since we were working with the
correct flux density for this source (∼13 Jy), these gains already have
the correct values. The other set of points, with values near ∼0.2, are
those for PKS 1613-586, which differ from the true values by a scale
factor. The next step is to derive this scale factor, so as to make the two
sets of points match as well as possible. The true flux density of PKS
1613-586 can be trivially determined from the derived scale factor.
The task fluxscale calculates the scale factor, thus performing the
bootstrapping, and will output a correctly scaled version of the cali-
bration table, which we can then use to calibrate the data, as well as
printing out the derived value for PKS 1613-586’s flux density.
CASA command:
fluxscale(vis=msfile, caltable=gain_table1,
fluxtable=flux_table1, reference=[’PKS 1934-638’],
transfer=[’PKS 1613-586’])
CASA output:
INFO fluxscale ##### Begin Task: fluxscale #####

INFO fluxscale Opening MS: CirX1.ms for calibration.
INFO fluxscale Initializing nominal selection to the whole MS.
INFO fluxscale Beginning fluxscale--(MSSelection version)-------
INFO fluxscale Found reference field(s): PKS 1934-638
INFO fluxscale Found transfer field(s): PKS 1613-586
INFO fluxscale Flux density for PKS 1613-586 in SpW=0 is: 4.6868 +/- 0.012777 (SNR = 366.816, N = 10)
INFO fluxscale Storing result in CirX1_spw0.fluxscale1
What is it we can deduce from this output? How do we know

everything is fine? First, we can examine the messages produced
by fluxscale. It has determined a flux density for PKS 1613-586 of
4.687 ± 0.013 Jy, making it a fairly strong source. For a strong source
like this, the uncertainty in the bootstrapped flux density should be
< 1%, which indeed it is. You can also check that the flux density
is within the range of values usually seen for your particular source
at this frequency. For example, our source, PKS 1613-586, is listed
in the ATCA calibrator list at (http://www.narrabri.atnf.csiro.au/
calibrators) as having flux density of 4.21 Jy at 1.75 GHz (the clos-
est of the listed frequencies to ours), so our value of 4.687 Jy is quite
reasonable, as its only 16% higher. Variations of a factor of 2 are not
uncommon, if you get a flux density more than a factor of 2 different
than the listed, you may want to double check.
If you use plotcal as above, but on the new version of the table pro-
duced by fluxscale, i.e. caltable=flux_table1), you will see that the
gain amplitudes for PKS 1913-638 are no longer readily distinguish-
able from those for PKS 1613-586, which was the goal (the gain of the
antennas should not depend on which source you are observing!). You
may also plot the gain phases with:
CASA command:
plotcal(caltable=flux_table1, xaxis=’time’,
yaxis=’phase’, iteration =’antenna’,
plotrange=[0,0,-180,180])
And you should see that the gain phases are in fact very stable in
time, and relatively consistent between our two calibrator sources. For
ant5 the gain phases will be exactly 0◦ as this was our reference an-
tenna, so it has a gain phase of 0◦ by definition. In order to calibrate
the visibility data for our target source, we have to interpolate be-
tween the calibrator gain values. Since the gain phases for our phase
calibrator source PKS 1613-586 form a nice smooth line (in fact the
individual points may not be distinguishable on your plot), the inter-
polation should yield an accurate value. If, on the other hand, the gain
phases (or amplitudes) differ drastically from one point to the next,
then interpolation to the intervening target-source observations will
be dubious.
Our calibration therefore is looking good. The next step is to ac-
tually use it to calibrate the visibility data and produce our estimates
of Vijtrue . Here is where we calculate the Vijtrue = Vijobserved Jij−1 . This is
where we want to be careful about how we interpolate our gain so-
lutions (Ji ). For example, on the plot of gain phase against time, you
could see the gain phases for PKS 1613-586 all lay on a fairly smooth
curve, while those for PKS 1934-638 did not quite lie on the curve.
Since PKS 1613-586 is much closer on the sky to our target source
(Circinus X-1) the gain values (i.e., values of Ji ) derived from PKS
1613-586 are better for interpolating to the target. However, we also
want to calibrate the visibility data for PKS 1934-638, and of course
here we’re much better of using the gain solutions derived from PKS
1934-638 itself. (Note that although at this point, we’re largely done
with our calibrator sources, and will not have any immediate further
use for the calibrator-source visibilities, its probably a good idea to
apply the calibration properly to the calibrator sources as well as the
target source).
The easiest way to keep it straight is to apply the calibration sep-
arately to each source. The task that applies the calibration is called
applycal. It combines the requested gain tables, interpolates or ex-
trapolates as required to determine the full antenna-based Ji , calculates
the baseline-based Jij values as needed for each visibility measure-
ment, and then writes out the properly calibrated visibility data (our
estimate of Vijtrue ) into the CORRECTED_DATA column of the ms. You can
specify the field, which determines which source’s visibility data get
corrected, and gainfield, which specifies from which source’s gain
table entries the required correction is determined. The best calibra-
tion is usually when the field entry is the same as the gainfield
entry, but usually one does not have calibration solutions for the tar-
get sources, so for the target source, gainfield must necessarily be
different than field. Note that as applycal can simultaneously apply
several calibration tables, gainfield can have multiple entries, one for
each calibration table to be applied. The two calibration tables we want
to apply are the bandpass table, bandpass_table0, and the gain table
with the correct flux density scaling for PKS 1613-586 or flux_table1.
Since there are two tables to apply, we need two entries in gainfield.
Lets just start with the first source: PKS 1934-638. We want to
calibrate its visibility data and therefore set field to PKS 1934-638.
Since the bandpass_table0 contains only entries for PKS 1934-638

anyway, there is no choice and we can just leave the correspond-
ing gainfield entry blank – this will apply all sources in the table.
For the flux_table, we want to use only the entries derived from
PKS 1934-638 itself, so the second gain_field entry should be PKS
1934-638. The parameter interp specifies the nature of the interpola-
tion used in gain table in a manner analogous to gainfield (see help
applycal for details). In this case we will just choose the entry nearest
in time to the particular visibility to be calibrated in bandpass_table0
and interpolate linearly in flux_table1.
CASA command:
applycal(vis=msfile,
gaintable=[bandpass_table0,flux_table1],
field=’PKS 1934-638’,
gainfield=[’’, ’PKS 1934-638’],
interp=[’nearest’,’’])
Next we turn to our phase calibrator. field should now be set

to PKS 1613-586. Again we can leave the first entry in gainfield
blank, since we again want to use any solutions in the bandpass table.
However, this time we want the second gainfield entry to be PKS
1613-586.
CASA command:
field=’PKS 1613-586’,
gainfield=[’’, ’PKS 1613-586’],
Finally, we apply the calibration to the target field. Now field

becomes Circinus X-1. The first entry of gainfield again can stay
blank (we’re again using the bandpass solution from PKS 1934-634).
For the second entry in gainfield, however, we cannot use the same
value as field like we did above, because we have no gain solu-
tions for Circinus X-1. Since PKS 1613-586 is much closer on the
sky, we want to use its entries from flux_table1, but not those from
PKS 1934-634, so the second gainfield entry should again be PKS
1934-634.
CASA command:
field=’Circinus X-1’,
gainfield=[’’,’PKS 1613-586’],
2.6 STEP 5... Splitting the data
We are now almost ready to split the calibrated visibility data for our
target source into separate ms file. Note that this step is convenient,
rather than necessary. One could just as well make images (and even
proceed to self-calibration) using the original ms, but often it is con-
venient to split out the data of interest at this point. This particular
data set has already been averaged in frequency to reduce the size for
the purposes of the workshop. These days most raw data sets you
encounter will likely have far more than 19 channels, and data sets
from interferometers with more telescopes than KAT-7, such as JVLA,
ALMA, LOFAR and MeerKAT once it comes online, will have far more
visibility measurements at each time-stamp. In such cases one might
average down in frequency to reduce the size of the data set and make
for faster processing. It is not necessary in this case, so our split
output file will have the same number of channels as our original ms.
However, before you do any averaging, it is a good idea to briefly ex-
amine the calibrated visibility data for the target source in case there is
more bad data we need to flag. Recall that so far we have only exam-
ined the data for our calibrator sources, but not yet that for Circinus
X-1. Lets turn again to plotms. We will now plot correlated flux den-
sity against baseline length in wavelengths (uvwave). Such plots are
commonly made to diagnose the nature of the source structure. We
now have to specify that we want the CORRECTED_DATA column: we’ve
gone to all this trouble to determine how to calibrate our visibilities,
so we want to make sure we’re enjoying the fruits of our labours.
CASA command:
plotms(vis=msfile, field=’Circinus X-1’,

xaxis=’uvwave’, yaxis=’amp’,
correlation=’XX,YY’, ydatacolumn=’corrected’,
coloraxis = ’corr’)
You should see something like this:
Figure 2.3: Plot of visibility amplitude

This plot mostly shows that Circinus X-1 is not a point source, against baseline-length for our target
source, Circinus X-1.
since the correlated flux density seems to vary systematically with
baseline-length (uv-distance or uvwave). There are, however, a few out-
lier points still, with anomalously high amplitudes above 10 Jy, such
as the spikes at UVwave ∼ 140, 290 λ. These are likely due to RFI. If
we re-plot amplitude against channel, you will see that all the high
amplitude point are in a single channel: channel 18. To further nar-
row it down, select spw = ’*:18’ and xaxis=’time’. Not only do the
high-amplitude points occur in one channel, they occur at a specific
time: near 25:40. By zooming in on the plot using the “zoom” tool
(select the magnifying glass icon, and use the mouse to draw a box on
the screen showing the region you want to zoom in on. You can zoom
back out using the “plot” button), to see that the bad data occurs be-
tween times of 25:32:30 and the end of the scan at 25:34:00. Although
it is possible to flag graphically directly in plotms, I encourage you to
again use flagdata. Firstly, it is better to positively identify the bad
data to be flagged first, rather than zapping points at random on the
screen. Secondly, using the flagdata command means that you can, if
needed, precisely and accurately repeat the operation. Note also, that
we could do more work to locate the bad data, it might for example
be only one antenna or baseline which is bad. However, since its such
a small amount of data (under 1.5 minutes in one channel only), its
more expedient just to flag it all. Note also that plotms labels time
axes in a slightly non-standard way with hours having values > 24.
For flagdata these values have to be translated into a more formal
timerange.
CASA command:
flagdata(vis=msfile,
timerange=’2012/07/02/01:32:30~2012/07/02/01:34:00’,
field=’Circinus X-1’, spw=’0:18’, flagbackup=True)
So now we’ve removed those outliers, we can split the data for our
source of interest, Circinus X-1.
CASA command:
split(vis=msfile, outputvis=’CirX1_split.ms’,
datacolumn=’corrected’, field=’Circinus X-1’)
2.7 STEP 6... CLEAN aka Deconvolution
For interferometric data, the (u, v) coverage is virtually never com-

plete, and therefore the instrumental response, called the dirty beam
or the point spread function, PSF, has some unwanted secondary re-
sponses, known as side lobes. The Fourier transform of the visibilities
represents the convolution of the true image with the dirty beam.
There are various techniques to reduce the effects of the sidelobes,
in other words to effect a deconvolution. In radio astronomy, decon-
volution has been dominated by two non-linear algorithms: CLEAN
and the Maximum Entropy Method (MEM).
The original formulation of CLEAN was by Högbom [1974], and
worked purely in the image plane. We explain this method here as
it is easier to understand. Most modern implementations of CLEAN
actually work partly in the (u, v) plane to produce somewhat more
accurate results than Högbom’s original image-plane formulation, but
are conceptually similar. CLEAN works iteratively, transferring flux
density from the dirty image to a set of point-like “clean components”,
which represent the true distribution of emission on the sky or the
deconvolved image. After each step, a residual image is formed which
represents the original dirty image after subtraction of the CLEAN
components so far. Initially, before the first CLEAN component is
found and subtracted, the residual image is just the dirty image.
The Högbom CLEAN algorithm proceeds using these steps:
1. Find strength and position of the brightest point in the current

residual image. CLEAN makes the plausible assumption that this
peak is mainly due to a real signal and only a minor part comes
from the sidelobe responses to other real emission farther away in
the image. Note that one should specify one or more “clean win-
dows” in which to search for real emission.
2. Subtract a point source at this location. The brightness of the point

source is taken to be the brightness of the brightest point multi-
plied by a factor γ(≤ 1) called the loop gain. This point source is
the CLEAN component, and it is added to the list of CLEAN com-
ponents. The response of the instrument to a point source is the
dirty beam pattern. Therefore, to subtract this point source from
our dirty image, CLEAN subtracts a dirty beam pattern, scaled so
that its peak is the same as that of the point source, and centred on
the location of the brightest point, from the current residual image.
3. Go to step 1 unless any remaining point sources are below some

level specified for convergence, or the maximum number of CLEAN
components is reached.
4. Convolve the accumulated CLEAN components, which represent

the model of the deconvolved emission, with an idealized “clean”
beam (usually an elliptical Gaussian fitted to the central lobe of the
dirty beam).
5. Add the final residual image to the convolution from the previous
step. The result is what is called the CLEAN image. Note that for
purposes of calibration, it is usually the CLEAN components that
are used rather than the CLEAN image.
The CLEAN algorithm was designed for the case where the true
brightness distribution contains only a few unresolved sources. It
works well for cases where the true emission occurs only within small,
well-separated, regions across the field of view, in other words if the
field of view is mostly empty. The following excerpt has been taken
from Robert Reid’s Thesis [?]. “Physics and filled aperture measure-
ments give us several properties that images should have. Non-negativity:
Astronomical objects, even black holes, cannot steal light away from
our receivers, but raw interferometric images abound with negative
patches. Requiring the image to be non-negative is a strong constraint
on what should be used to fill in the gaps. Locality: Usually the range
of directions from which light can be received is limited by the an-
tennas being used. The rough location and extent of the source may
also be known from observations at other wavelengths. Raw images
always have ripples going off to infinity that should be quenched. If
nowhere else, nulls in the antenna reception pattern are invalid places
to receive light from. For many frequencies and parts of the sky, bright
radio sources are also quite rare and confined (as known from filled
aperture measurements), so images are expected be mostly empty.
Smoothness: Sources are expected to be continuous, and usually to
have continuous derivatives. Also, it is best to minimize the effect of
the unmeasured part of the sampling plane on the final image. Per-
fectly sharp edges depend on the entire sampling plane, far beyond
where interferometers can make measurements. Smoothing the image
concentrates its region of dependence on the finite area that interfer-
ometers sample. Agreement with theory: At first this seems to throw
out the baby with the bathwater, since true discoveries must be novel.
True discoveries, however, must rest on true observations. Since decon-
volution ideally agrees with all the measurements, we should not let
non-measurements mislead us. Another way to see this is to think of
the model as assisting us in separating reasonable interpolations from
unreasonable ones for the image, but the parameters of the model it-
self are determined by the measurements. For example a circle has an
infinite number of points on its circumference, but three points on its
locus are enough to specify it, if we are sure that it is a circle. In any
case, images tend to be constructed before models, thus this principle
is not always applicable.”
Two main inputs to clean: the image size (imsize) and cell size
(cell), which correspond to the size in pixels of the map, and the
size in arcseconds of a pixel, respectively, must be carefully chosen.
Here we show how to calculate these two parameters. By knowing the
frequency of our observations we can calculate the primary beam (i.e.,
the area of the sky observed with the telescope).
First we choose the size of the pixels or cells. This choice is deter-
mined by the resolution of the interferometer, which is given by the
synthesized beam. The value that is usually given is the full width at
half-maximum (FWHM). Although the exact value for FWHM of the
synthesized beam will depend on the details of the u, v-coverage and
weighting for each particular observations, and approximate value for
a given set of observations depends largely on the maximum baseline,
Bmax , measured in wavelengths (λ) and is given by:
1
FWHMsynthesizedbeam ' radians, or (2.5)
( Bmax /λ)
3438
FWHMsynthesizedbeam ' arcminutes (2.6)
Bmax /λ
You can determine Bmax in λ by using plotms to plot the visibility
amplitude as a function of the baseline length in λ or uvwave.
In order not to be limited by the pixel size, the central lobe of the
dirty beam needs to be well resolved, and therefore a value of ap-
proximately 41 of the above FWHM of the synthesized beam should
be chosen for the cell size. When CLEAN is run, one should check
that the smaller, or minor axis of the elliptical fitted CLEAN beam is
at least 3× the size of the cells. If not, CLEAN should be stopped and
restarted with better-suited values for cell.
In our case, we get an estimate of 3.70 for the resolution and thus
0.920 for the pixel size, which we will round down to 0.90 (note that
since the requirement is to have pixels small enough to resolve the
beam, rounding down is okay but not rounding up).
The other choice the user must make is the size of the region to
be imaged. The maximum size of the region that it is sensible to im-
age is determined by the field of view of the interferometer, which is
just determined by that of individual antennas, and is independent of
the array configuration. The FWHM of primary beam of a uniformly
illuminated antenna is given by Napier [1999]:
1.02 × ( νc )
FWHMprimarybeam = radians (2.7)
(dishdiameter
In our case, the KAT-7 dishes have a diameter of 12 m, so at our
observing frequency of 1822 MHz, the FWHM of the primary beam
is ∼480 . Of course, sources outside this field can still be seen by the
interferometer, but at reduced amplitude. In particular, a source at a
distance of FWHM/2 (remember that the FWHM is a diameter, not a
radius) will be attenuated by 50% due to the primary beam response.
For the relatively coarse resolution offered by KAT-7, one can afford
to image a relatively large region, and a reasonable choice is to image
a region with 3× the diameter of primary beam FWHM. In our case
this is equal to (3 × 480 = 1440 . Since we have determined our pixel
1440 )
size to be 0.900 , we need to image a region of = 160. It is common
0.900
to choose image sizes which are powers of 2, so we shall round this up
to 256, which is the value we use for imsize.
When do I stop cleaning?

Stop cleaning when the residuals are noise like, and/or the clean
has converged (the cleaned flux is no longer increasing)! Here are
some of the parameters of clean which determine when it stops:
niter - Number of CLEAN iterations to do. This can be useful
when you are doing tests, but this parameter has NO physical mean-
ing. Instead set to large number and let either threshold or do interac-
tive to stop the cleaning. Please also ensure you are not over cleaning
your image. Have an eye on the casapy log that the total fluxes are not
decreasing drastically.
threshold - Stop cleaning when peak residual has this value, give
units (i.e. mJy). Ideally, one would like the image to have a noise
approaching the theoretical one, in which case a threshold of ∼ 3×
the theoretical rms noise is appropriate. Note: to reach this limit the
data must be well calibrated/flagged and suffer from no serious arti-
facts (resolved out extended structure/negative bowls, poor PSF/(u, v)
coverage, limited dynamic range etc).
CLEANing deeply into the noise, that is with threshold set to a
value below the image rms level is dangerous if you have a large
CLEAN window. If the area of your CLEAN window is small, fewer
than say 10 times the area of the CLEAN beam, then it is fairly safe
to set threshold to some low value and just clean deeply. If the area
of your CLEAN window is larger than overcleaning can be a prob-
lem. Basically, once you reach rms (whether close to theoretical or
not), you are just picking noise up one place and putting it down in
another. With a small CLEAN window, there is not much freedom
here so nothing much can go wrong by moving a bit of noise around
within your small window. With a large window strange artefacts can
arise. Remember also, that rms in an otherwise blank region of your
image represents only a lower limit to the uncertainty of the parts of
your CLEAN image which have emission.
Starting the cleaning
Before you proceed to make an image, you will use the CASA
viewer to look and perform various tasks. Viewer3 is a GUI for im- 3
Follow the demo from: http:
age manipulation. //casa.nrao.edu/CasaViewerDemo/
casaViewerDemo.html
CASA command:
clean(vis=’CirX1_split.ms’, imagename=’CirX1_split.im’,
niter=5000, threshold=’5.0mJy’, psfmode=’hogbom’,
interactive= True, imsize=[256,256],
cell=[’0.5arcmin’ ,’0.5arcmin’], stokes=’I’,
weighting=’briggs’, robust=0.0)
3
Spectral line calibration and imaging
Spectral line calibration is pretty similar to that for continuum. The

real differences come in when imaging the data. Instead of averaging
together all the data before imaging, we make one image per frequency
channel, producing a data cube. In this tutorial we assume that you
have worked through Chapter 3 and understand the basics of CASA
and calibration.
The KAT-7 spectral line modes produce 4096 channels of data. The
outer 500 channels are usually discarded since the gain tapers off
rapidly as part of the guard bands. For this tutorial we will use a
subsection of data to speed things up. (The original data file is 34 GB).
We will be looking at a massive star forming cloud in the Galactic
Plane. At L-band we see a number of evolved HII regions (continuum)
and hydroxyl masers (spectral lines). The msfile is called G330_OH.ms.
3.1 Inspection and flagging
First, let us set up all our filenames:

CASA command:
prefix = ’G330_OH’
gtable0 = prefix + ’.G0’
btable0 = prefix + ’.B0’
gtable1 = prefix + ’.G1’
ftable1 = prefix + ’.fluxscale1’
Lets have a look at what we have here:

CASA command:
listobs(msfile)
CASA output:
##########################################
##### Begin Task: listobs #####
listobs(vis="G330_OH.ms",selectdata=True,spw="",field="",
antenna="",uvrange="",timerange="",correlation="",scan="",
intent="",feed="",array="",observation="",verbose=True,
listfile="",listunfl=False,cachesize=50)
================================================================================
MeasurementSet Name: /home/sharmila/spec_line_tut_OH/G330_OH.ms MS Version 2
================================================================================
Observer: sharmila Project: 20130621-0006
Observation: KAT-7
Data records: 12327 Total integration time = 40778.8 seconds
Observed from 21-Jun-2013/14:08:01.0 to 22-Jun-2013/01:27:39.8 (UTC)
ObservationID = 0 ArrayID = 0
Date Timerange (UTC) Scan FldId FieldName nRows SpwIds AveInts
21-Jun-2013/14:07:45.9 - 14:12:17.7 1 0 1934-638 189 [0] [30.2]
14:25:02.7 - 14:25:32.9 4 1 1613-586 21 [0] [30.2]
14:26:13.2 - 14:30:45.0 5 2 G330.89-0.36 189 [0] [30.2]
14:36:37.3 - 14:37:07.5 7 1 1613-586 21 [0] [30.2]
14:37:47.8 - 14:42:19.6 8 2 G330.89-0.36 189 [0] [30.2]
14:48:11.9 - 14:48:42.1 10 1 1613-586 21 [0] [30.2]
14:49:22.3 - 14:53:54.1 11 2 G330.89-0.36 189 [0] [30.2]
...........
...........
...........
01:10:33.0 - 01:11:33.4 159 1 1613-586 42 [0] [30.2]
01:11:48.5 - 01:16:20.3 160 2 G330.89-0.36 189 [0] [30.2]
01:22:07.6 - 01:22:37.8 162 1 1613-586 21 [0] [30.2]
01:23:23.1 - 01:27:54.9 163 2 G330.89-0.36 189 [0] [30.2]
(nRows = Total number of rows per scan)
Fields: 3
ID Code Name RA Decl Epoch nRows
0 T 1934-638 19:39:25.017468 -63.42.45.60158 J2000 2079
1 T 1613-586 16:17:17.893107 -58.48.07.88902 J2000 1176
2 T G330.89-0.36 16:10:20.541228 -52.06.14.90063 J2000 9072
Spectral Windows: (1 unique spectral windows and 1 unique polarization setups)
SpwID Name #Chans Frame Ch1(MHz) ChanWid(kHz) TotBW(kHz) Corrs
0 none 1001 TOPO 1665.985 0.381 381.9 XX YY
The SOURCE table is empty: see the FIELD table
Antennas: 7:
ID Name Station Diam. Long. Lat.
0 ant1 ant1 12.0 m +021.24.39.4 -30.33.10.2

1 ant2 ant2 12.0 m +021.24.41.9 -30.33.09.1
2 ant3 ant3 12.0 m +021.24.38.6 -30.33.09.1
3 ant4 ant4 12.0 m +021.24.37.7 -30.33.09.1
4 ant5 ant5 12.0 m +021.24.37.1 -30.33.10.0
5 ant6 ant6 12.0 m +021.24.36.2 -30.33.12.5
6 ant7 ant7 12.0 m +021.24.35.2 -30.33.07.5
##### End Task: listobs #####
##########################################
spectral line calibration and imaging 45
We have averaged to 30 second integrations and there are 601 chan-

nels in this dataset to reduce the size of the file.
Now we set up our calibrator variables. We use PKS 1934-638 as the
flux and bandpass calibrator and PKS 1613-586 as the gain calibrator.
Antenna 6 is usually very stable, so we will use it as our reference
antenna.
CASA command:
f_cal = ’1934-638’
b_cal = ’1934-638’
g_cal = ’1613-586’
ref_ant = ’ant6’
Before we start calibration, it is vital to make sure that our calibra-

tors are free of RFI. First we check along the frequency axis. This is a
very narrow-band observation (1.1 MHz) so we don’t usually expect
to see much RFI. However, we have found self-generated RFI on some
antennas, so we will look at each baseline independently.
CASA command:
plotms(vis = msfile,
field = f_cal,
xaxis = ’channel’,
yaxis = ’amp’,
iteraxis = ’baseline’,
yselfscale = True)
Did you spot the RFI? Identify the channel and use flagdata to flag
the channel for the affected antennas.
CASA command:
flagdata(vis = msfile,
spw = ’0:548’,
antenna = ’ant5,ant6,ant7’)
3.2 Set up models for the calibrators
Now lets prepare the file for calibration. This step will clear all pre-
vious calibrations or, if it is a freshly made ms file, will add in the
MODEL and CORRECTED DATA columns.

CASA command:
clearcal(msfile)
Now we are going to set up models for our calibrators. First, we set
up our flux calibrator:
CASA command:
setjy(vis = msfile,
field = f_cal,
fluxdensity = -1,
standard = ’Perley-Taylor 99’)
PKS 1613-586 is not a very good calibrator, but it is difficult to find

isolated sources close to the Galactic Plane. Instead, we use a model of
the field, found by imaging the calibrator. You should have the direc-
tory 1613-586.model in your working directory. This is a CASA model
file, which is produced by CLEAN. This will populate the MODEL col-
umn for the source with the appropriate visibilities.
CASA command:
setjy(vis = msfile,
field = ’1613-586’,
fluxdensity = [0,0,0,0],
modimage = ’1613-586.model’)
3.3 Calibration
We first do a preliminary time-dependent phase calibration over a sub-

set of the band. We have to find a small section of the band that is
reasonably flat, and have enough bandwidth to achieve a reasonable
signal-to-noise. The number of channels is going to depend on the
total bandwidth of the observation. In this case, we are using 1/4 of
the total channels in the 1.5 MHz mode and don’t expect to see much
change across the bandpass.
CASA command:
plotms(vis=msfile,
field = f_cal,
yaxis = ’phase’,
avgtime = ’1e8’,
correlation = ’XX’,
antenna = ’ant6’,
coloraxis = ’antenna2’)
You will have noticed in plotms that the bandpass appears to be

relatively flat across this channel range, so we can in fact use all of the
channels. Note that this is not the case for the wider HI modes. You
may find solutions failing if you use too large a channel range.
CASA command:
ref_chans = ’0:1~600’
gaincal(vis = msfile,
field = b_cal,
caltable = gtable0,
refant=ref_ant,
spw = ref_chans,
calmode = ’p’,
solint = ’inf’,
minsnr = 5,
solnorm = True,
interp = ’nearest’)
Inspect the solutions using plotcal.
CASA command:
plotcal(caltable = gtable0,
xaxis = ’time’,
markersize=3,
plotsymbol=’.’,
iteration=’antenna’,
subplot=421,
fontsize=8)
All looks good, so lets continue on to the bandpass calibration. We

have found that in general, we do not need to do a time-dependent
bandpass calibration for KAT-7, even though we visit the bandpass
calibrator every hour in these observations (just in case), so we aver-
age over all of the bandpass scans to increase our signal-to-noise ratio.
To start with, we are going to use the normal ’per channel’ bandpass
solution ie bandtype=’B’. However, if the solutions turn out to be
very noisy, it may be better to use a polynomial fit to the bandpass
bandtype=’BPOLY’
CASA command:
bandpass(vis = msfile,
caltable = btable0,
field = b_cal,
refant = ref_ant,
solnorm = True,
combine = ’scan’,
solint = ’inf’,
bandtype = ’B’,
minsnr = 5,
gaintable = [gtable0],
interp = [’nearest’])
If you are using CASA 4.1 and above, there is a new task to plot
bandpasses. This will enable you to see both the amplitude and phase,
and overplot polynomial solutions (Figure 3.1).
CASA command:
plotbandpass(caltable = btable0,
xaxis = ’freq’,
yaxis = ’both’,
subplot = 42)
The problem with such narrow band observations is that we have

increased noise. Note that the peak-to-peak amplitude variation is
about 10% and the phase variation is about 8 degrees. However, there
is not much variation in the bandpass, so we can use a polynomial-
based bandpass calibration. Note below that we change bandtype to
’BPOLY’. We can specify different orders of polynomial fit for the phase
and amplitude if necessary (Figure 3.2).
Figure 3.1: The output of plotbandpass

for a few antennas.
CASA command:
btable1 = prefix +’.B1’

bandpass(vis=msfile,
caltable = btable1,
field = b_cal,
refant =ref_ant,
solnorm = True,
combine=’scan’,
solint = ’inf’,
bandtype = ’BPOLY’,
degamp = 10,
degphase = 10,
minsnr = 5,
gaintable = [gtable0],
Why is the bandpass calibration so important? In spectral line ob-

servations we are interested in source structure as a function of fre-
quency, so we want to remove any spurious variations due to the in-
strumental response. We may also want to subtract off continuum
emission in order to simplify our imaging and in this case we don’t
want there to be any residual continuum emission. How do we know
if we have sufficient signal to noise in our bandpass calibrator? We
typically want the rms noise from the calibrator to be much less than
the noise in the source. We use calibrators with reasonably high flux
density, but spend less time on them than we do on our target sources,
Figure 3.2: The output of plotbandpass

with B and BPOLY solutions overlaid.
so its good to review whether your observations satisfy the criterion

√ p
that Ical tcal should be significantly greater than Itarget ttarget . In the
case of this observation, the target actually has quite a high contin-
uum flux, and it will not be possible to meet this requirement without
spending an impractical amount of time on the bandpass calibrator.
Now we calibrate the gain on our flux calibrator, which in this ob-
servation happens to also be our bandpass calibrator.
CASA command:
caltable = gtable1,
field = f_cal,
solint = ’inf’,
refant = ref_ant,
gaintype = ’G’,
calmode = ’ap’,
solnorm = False,
minsnr = 5,
gaintable = [btable1],
Now check that there are no bad outliers in the gain calibration.
CASA command:
field = f_cal,
xaxis = ’time’,
yaxis = ’amp’,
markersize=3,
plotsymbol=’.’,
subplot=421,
fontsize=8)
field = f_cal,
xaxis = ’time’,
markersize=3,
plotsymbol=’.’,
subplot=421,
fontsize=8)
Calibrate the gains on the phase calibrator. Note that we append

the solutions to the table containing the gains on the flux calibrator.
CASA command:
caltable = gtable1,
field = g_cal,
solint = ’inf’,
refant = ref_ant,
gaintype = ’G’,
calmode = ’ap’,
append = True,
solnorm = False,
minsnr = 5,
gaintable = [btable1],
As usual, check the gain solutions and make sure that there’s noth-
ing nasty going on. Next, we transfer the flux scale from our flux
calibrator to the gain calibrator. This will automatically transfer the
flux calibration to the target when we apply the calibrations.
CASA command:
fluxscale(vis=msfile,
caltable = gtable1,
fluxtable = ftable1,
reference = f_cal,
transfer = g_cal)
CASA output:
##########################################
##### Begin Task: fluxscale #####
fluxscale(vis="G330_OH.ms",caltable="G330_OH.G1",
fluxtable="G330_OH.fluxscale1",reference="1934-638",
transfer="1613-586",listfile="",append=False,refspwmap=[-1],
incremental=False, fitorder=1)
Opening MS: G330_OH.ms for calibration.
Initializing nominal selection to the whole MS.
Beginning fluxscale--(MSSelection version)-------
Found reference field(s): 1934-638
Found transfer field(s): 1613-586
Flux density for 1613-586 in SpW=0 is: 1.01027 +/- 0.00741467 (SNR = 136.252, N = 14)
Storing result in G330_OH.fluxscale1
Writing solutions to table: G330_OH.fluxscale1
##### End Task: fluxscale #####
##########################################
Check that the derived flux is close to what is expected. Now we

are ready to apply the calibration and get ready for imaging.
CASA command:
target = ’G330.89-0.36’
field = target,
gaintable = [btable1, ftable1],
gainfield = [b_cal, g_cal],
interp = [’nearest’, ’linear’])
CASA command:
split_outputvis = target+’.split.ms’
split(vis = msfile,
outputvis = split_outputvis,
field = target,
datacolumn = ’corrected’)
3.4 Velocity rest frames
There are a few things that you need to do before you can dive in and
start cleaning your image. Since spectral line observations give us a
lot of information about velocities of the objects that we are observing,
it is important to be clear on what reference frames we are using. The
apparent frequency of a source, also known as the ’sky frequency’,
is influenced by Earth’s rotation and revolution around the Sun, the
motion of our Solar System around the Galaxy, the movement of our
Galaxy within the Local Group and there is also motion against the
cosmic microwave background. The various rest frames are listed in
Table 3.1. At L-band, the sky frequency change by up to 0.15 MHz
during the course of a year.
The KAT-7 system does not do any Doppler corrections, so its rest
frame is topocentric. The actual sky frequency of your source will
vary from day to day and if you are observing with high frequency
resolution, you may even see your spectral line shifting during the
course of a single day’s observation. This is definitely the case for
these hydroxyl maser observations. We have a velocity resolution of
68 m/s so we do see the source Doppler-shifted by Earth’s rotation.
You will usually want to have your final data cubes output into
a different rest frame. The velocity frame to use depends on your
science. In the case of this dataset, we are looking at a Galactic star
formation region. It is meaningfull in this case to work in the Local
Standard of Rest (LSR) frame. There are actually two definitions of
LSR, the kinematic LSR (LSRK) and dynamic LSR (LSRD). Most of the
time LSRK is used and is generally synomynous with LSR. The other
commonly used rest frame is barycentric, and has mostly replaced the
old heliocentric standard.
There are two ways to change rest frames in CASA. The CLEAN
task can change the reference frame of the spectral axis which is fine if
you don’t have to worry about doppler shifts during the course of the
observation. CVEL is a more general spectral regridding tool, which
enables you to correct for Doppler shifts, or change the channelisation
on your data prior to imaging. Doing the spectral regridding before

running CLEAN can save you time if you think you are going to be
trying different imaging settings.
Table 3.1: Velocity rest frames

Rest Frame Name Rest frame Corrects for Max
(km/s)
Topocentric Telescope Nothing 0
Geocentric Earth Centre Earth Rotation 0.5
Earth-Moon Earth-Moon Motion about 0.013

Barycentric centre of mass Earth + Moon
centre of mass
Heliocentric Centre of Sun Earth orbital motion 30
Barycentric Earth+Sun Earth+Sun 0.012

centre of mass centre of mass
Local standard of rest Centre of mass Solar motion rel 20

(LSR) of local stars to nearby stars
Galactocentric Centre of Milky Way 230

Milky Way rotation
Local Group Local Group Milky Way 100

Barycentric centre of mass motion
Virgocentric Centre of the Local Group 300

Local Virgo motion
Supercluster
Cosmic Microwave CMB Local Supercluster 600

Background motion
You can use the viewer to look at visibilities. Lets try it now.
Figure 3.3 shows a pretty severe shift over the course of the obser-
vations. If we had to image this dataset without correcting for the
Doppler shift, we would end up blending our spectral features to-
gether. We are going to be using CVEL to correct for the doppler shift,
and at the same time, change the velocity reference frame to the Local
Standard of Rest (VLSR ). Before we do that, we need to set the line rest
frequency in the measurement set. The KAT-7 observation framework

does not make provision for setting the rest frequency in the data files
so we shall have to do it manually. The rest frequency for this tran-
sition of OH is 1665.40184 MHz. We are going to open the spectral
window table and insert the rest frequency. The default rest frequency
in the file is usually the central observing frequency. Remember to
close the file before continuing because it will be locked until you do.
Figure 3.3: A raster plot of visibilities on

a single baseline. Note how the spectral
line is shifting through the channels over
time.
CASA command:
restfreq=1665.40184e6
tb.open(split_outputvis+’/SPECTRAL_WINDOW’,
nomodify=False)
tb.putcell(’REF_FREQUENCY’, 0, restfreq)
tb.close()
If run listobs now you should see that the reference frequency has
changed. Next we run cvel.
CASA command:
freq_string = str(restfreq/1e6)+’MHz’
cvel_outputvis = target+’.cvel.ms’
cvel(vis = split_outputvis,
outputvis = cvel_outputvis,
mode = ’velocity’,
interpolation = ’linear’,
outframe = ’LSRK’,
restfreq = freq_string)
Figure 3.4: A raster plot of visibilities

on a single baseline after doppler correc-
tion.
This can take some time for a large dataset. If you find yourself
running low on disk space, you can now delete the split file since
we will use the cvel file from now on. Have a look at the spectral
lines using the viewer and convince yourself that we have removed
the doppler shift. In Figure 3.4 you will see that the spectral lines now
stay in the same channel. Note that some of the edge channels are now
missing data. Make sure that you have split off sufficient channels to
account for this before you apply cvel. We will lose about 10 channels
in this process. When you clean your datacube, you should make sure
that these channels are not included. I generally avoid this problem by
setting the start velocity and number of channels for the output from
cvel such that I cover only the channels that I want to image. This trick
works well if you have observations spread over more than a month,
because the channel in which a particular velocity will be present will
vary over the year. This brings your data into a common reference
frame with consistent channel numbers. If you don’t know what these
parameters should be, you can run cvel on the full channel range,
then examine the resulting file and rerun cvel with a smaller output
range once you know which velocity range is of interest. Remember
to leave sufficient line-free channels on either side of your dataset for
continuum subtraction.
3.5 Continuum subtraction
This particular field is of a massive star formation region. The one

degree field of view includes several evolved HII regions as well as
younger regions containing masers. It will simplify the cleaning pro-
cess considerably if we can subtract the continuum and image it sepa-
rately from the line emission. Lets have a look at a spectrum and see if
we can establish the continuum baseline. In the case of these masers,
the emission is so strong that we can see it in individual integrations,
however you need to be careful because there are also weaker features
less than one Jy. In the case of weaker sources, we would need to
average together all of the data in order to spot the emission. This
works well only if the source is close to the phase centre and not too
extended. The best way is to make a dirty cube and examine this
through the viewer in order to determine line channels. In the case of
extremely weak line sources you may even need to do some cleaning
first before you can identify the lines.
Lets have a quick look at the averaged spectrum.
CASA command:
plotms(vis=cvel_outputvis,
ydatacolumn = ’data’,
yaxis = ’amp’,
avgtime=’1e8’,
avgscan=True)
There may be some emission around the last channels. It also turns
out that this plot does not show sufficient detail to figure out where
the line emission is. It will be much better to make a dirty cube first
and inspect the result to determine where we can place the continuum
windows. To make a dirty cube, we simple set the number of iterations
of CLEAN to zero.
CASA command:
clean(vis = cvel_outputvis,
imagename = ’dirty’,
mode = ’channel’,
restfreq = freq_string,
imsize = 256,
cell=’30arcsec’,
stokes = ’I’,
niter = 0,
psfmode = ’hogbom’,
interactive = False,
weighting = ’briggs’)
viewer(’dirty.image’)
We have a reduced channel range in this dataset in order to keep

the file sizes down, and as a result most of the channels contain line
emission. Continuum subtraction is a lot easier in the original 4096
channel dataset. Careful examination of the cube will show that the
only line-free channels are around 280 to 340. The continuum sub-
traction task will issue a warning about extrapolating over the outer
frequency ranges but will still do it.
CASA command:
cont_window = ’0:280~340’
uvcontsub(vis=cvel_outputvis,
fitspw = cont_window,
fitorder = 1,
want_cont = True)
Now you have two new measurement sets, one containing the con-
tinuum data, which you can image as usual, and one containing the
spectral data. We do not expect to see any absorption features in this
field, so if you do see negative features, it may mean that the con-
tinuum subtraction has failed. You can check this before going on
to the laborious task of deconvolution by making a dirty cube of the
continuum-subtracted data.
3.6 Finally, on to imaging
Now we can start cleaning the spectral line cube. Masers typically
have angular sizes of the order of milli-arcseconds and therefore are
not resolved with KAT-7. However, there are five individual sources
of maser emission in this field of view. Lets start the cleaning process.
You will notice that we still specify the rest frequency even though we
have already set it in the measurement set. This is because of some
inconsistencies in the way CASA handles spectral velocities - for now
it is safer to always specify the rest frequencies and the required output
frame.
CASA command:
restfreq=1665.40184e6
freq_string = str(restfreq/1e6)+’MHz’
contsub_vis = target+’.cvel.ms.contsub’
cube_namebase = target+’.cube.clean’
clean(vis = contsub_vis,
imagename = cube_namebase,
spw = ’0:1~590’,
mode = ’channel’,
restfreq = freq_string,
imsize = 256,
stokes = ’I’,
threshold = ’0.15Jy’,
niter = 20000,
interactive = True,
If you find you are running out of time, try setting interactive to
False and let CASA find its own way.
Optional: Try cleaning the continuum. Below are the commands
to start the continuum cleaning. Note that I have dropped the last 11
channels to get rid of the empty channels that we were left with after
CVEL.
CASA command:
cont_vis = target+’.cvel.ms.cont’
cont_image = target+’.cont.clean’
clean(vis=cont_vis,
imagename = cont_image,
spw = ’0:0~590’,
imsize = 512,
stokes = ’I’,
threshold = ’10mJy’,
niter = 5000,
interactive = True,
You will notice that this field has a pretty complex structure, con-
sisting of point sources as well as more nebulous evolved HII regions.
It is far better to clean this image using multi-scale clean, which will
fit for several size-scales. To set up multi-scale, we need to specify the
scales, in number of pixels, where 0 indicates a point source, 5 pixels
is about one beam, and then we do a few multiples of the beam, going
up to 4 times the beam in this case. Note that CASA will complain if
the scales are too large and then drop them. If doing the cleaning non-
interactively, it is good to also tell CASA to stop when it starts creating
too many negative components, otherwise things can go wrong very
quickly.
Warning: multiscale clean is a lot slower than the conventional

clean because it repeats the clean cycle for every scale specified
and needs to re-scale the psf for every scale. When setting this
task to non-interactive mode, and trying to clean down to 5
times the expected rms noise, this takes almost 3 hours on a
high-end desktop machine. This is also why we try to subtract
the continuum off in the uv-plane, since it speeds up imaging
the OH masers considerably.
CASA command:
cont_image = target+’.cont.multiscale.clean’
clean(vis=cont_vis,
imagename = cont_image,
multiscale = [0, 5, 10, 15, 20],
negcomponent = 1,
spw = ’0:0~590’,
imsize = 512,
stokes = ’I’,
threshold = ’10mJy’,
niter = 5000,
interactive = True,
4
Troubleshooting and Support
4.1 Getting help
In casapy, you can type help [commandname] to get help in a particular

command.
If you’ve encountered a problem, have a question, or would like
to report a bug, please send an email to our mailing list or visit our
website.
Bibliography
J. A. Högbom. Aperture Synthesis with a Non-Regular Distribution

of Interferometer Baselines. Astron. Astrophys. Suppl. Ser., 15:417, June
1974.
P. J. Napier. The Primary Antenna Elements. In G. B. Taylor, C. L.

Carilli, and R. A. Perley, editors, Synthesis Imaging in Radio Astronomy
II, volume 180 of Astronomical Society of the Pacific Conference Series,
page 37, 1999.
R. J. Sault and T. J. Cornwell. The Hamaker-Bregman-Sault Measure-

ment Equation. In G. B. Taylor, C. L. Carilli, and R. A. Perley, editors,
ASP Conf. Ser. 180: Synthesis Imaging in Radio Astronomy II, pages 657–
669, 1999.

Casa Cook Arvind

Uploaded by

Copyright:

Available Formats

Casa Cook Arvind

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Casa Cook Arvind

Uploaded by

Copyright:

Available Formats

N. OOZEER, M. BIETENHOLZ & S.

1 Why do you want to learn CASA? 5

2 Let us get serious 9

3 Spectral line calibration and imaging 43

4 Troubleshooting and Support 63

sampled. The presence of unsampled regions in the (u, v) causes ar-

forming the interferometer, possibly of several different polarization

1.1 How to use this document

1.2 Conventions used

We have tried to use the following conventions through this document:

• A typewriter font is used for CASA tasknames and variables in

• The Return key is synonymous with the ENTER key or ←-

Tips: lead to shortcuts that we find useful

Cautions: Avoid pitfalls

Comments: explain the relevant procedure

Pause: Time for a break

In this document we will review the data reduction procedures that

Figure 1.2: This mindmap shows the

Python is case sensitive (as is Linux). Always ensure you using

Figure 1.3: THE sign

2.1 Overview of Calibration

The visibility measurement made by the interferometer are complex

2.2 STEP 1... Converting the hdf5 file to a MS

h5toms.py -f --no-auto -C 0 199,7990 --flags=0 all0 -r ant5 myfile.h5

The -f flag causes h5toms.py to produce a full-polarization output

2.3 STEP 2... Loading and inspecting the data

It is probably best to create a specific directory which will contain the

tar -zxf School_data_av.ms.tar.gz

It should produce a directory called School_data_av.ms. This di-

Although CASA measurements sets (ms) are ordinary linux di-

• Do not move, delete etc. directories from the oper-

• Do not alter the contents of the ms. While renaming or mov-

Do NOT abort running programs by pressing Ctrl+C as CASA

In fact, CASA is designed to be highly script-able, and sequences of

Below we define a variable to hold the ms (directory) name but

bandpass_table0 = prefix +’_spw0.B0’

We chose antenna ant5 as our reference antenna when we first con-

. . . (we omit the bulk of the detailed scan listing here)

05:29:53.8 - 05:30:38.9 328 2 PKS 1613-586 84 15 [0]

It is a good idea to save the listobs output for future reference,

• The observations started at 14:00 on 01-Jul-2012 and continued till

• Three sources were observed, with field names “PKS 1934-638”,

• The reference frequency 1938.21094 MHz; One spectral window,

• The number of antennas taking part was 7

The science target in these observations was Circinus X-1 (ID=1),

2.4 STEP 3... Flagging

The majority of interferometric data sets contain some corrupted data,

• flagmanager: save and manage versions of the flagging entries in

• flagautocorr: noninteractive flagging of auto-correlations

• plotms: interactive XY plotting and flagging of visibility data

• plotxy: interactive XY plotting and flagging of visibility data - note

• flagdata: This is the new non-interactive flagging (and unflagging)

We will mostly use the task plotms to examine the data. It is a

Note that in “elevation” mode, the elevation limits provided

plotms(vis=msfile, xaxis=’channel’, yaxis=’phase’,

Your screen will look similar to this:

Figure 2.1: The output of plotms show-

polarization of Ant2 and Ant4.

plotms(vis=msfile, xaxis=’time’, yaxis=’phase’,

Referring to an antenna can be kind of confusing in CASA.