AIPS HELP file for OMFIT in 31DEC24
As of Tue Apr 23 20:09:51 2024
OMFIT: Task to fit gaussian or spheres to UV data.
INPUTS
-----DATA specification------
INNAME Input UV file name (name)
INCLASS Input UV file name (class)
INSEQ Input UV file name (seq. #)
INDISK Input UV file disk unit #
SRCNAME Source Name (1 only)
SUBARRAY Subarray
TIMERANG Time: start day,hr,min,sec
stop day,hr,min,sec
ANTENNAS Antennas to include
UVRANGE Range of projected spacings
(thousands of wavelengths)
STOKES 'I', 'HALF', 'RR', and 'LL'
default = 'I'
FREQID Default FQ table entry (-1)
BCHAN First chan ( 0 => 1 )
ECHAN Last chan ( 0 => BCHAN!!!)
BIF First IF ( 0 => 1 )
EIF Last IF ( 0 => EIF !!!)
DOCALIB -1.0 101.0 > 0 calibrate data & weights
> 99 do NOT calibrate weights
GAINUSE CL (or SN) table to apply
DOPOL -1.0 10.0 If >0 correct polarization.
PDVER PD table to apply (DOPOL>0)
BLVER BL table to apply.
FLAGVER Flag table version
DOBAND -1.0 10.0 If >0 apply bandpass cal.
Method used depends on value
of DOBAND (see HELP file).
BPVER Bandpass table version
SMOOTH Smoothing function. See
HELP SMOOTH for details.
---- Input MODEL file -----
INFILE
DOFIT shortcut for model specs
--- Output MODEL file -----
OUTFILE
-------SCREEN output-------
PRTLEV recommended value = 3
+100 => explain noise
+10 => show intermed. values
---- Output DATA file -----
OUTPRINT
-- Output AIPS DISK file --
OUTNAME (name)
OUTCLASS (class)
OUTSEQ (seq. #)
OUTDISK (disk #)
OUTVERS (content type - see HELP
0: D 1: D-GS 2: S
3: GS 4: D/G 5: D/S
6: D/(GS) 7: D/|D|
NOISE (1) a priori noise [Jy]
(not already in weights)
(2) outlier threshold
[sigmas]
(3) low-amp cutoff threshold
[sigmas]
APARM Task Enrichment Parameters
Early Stopping Criteria:
( set <0 to turn off )
(1) percent change in Chi-squared
0 => 1 percent ( <0 to disable )
(2) greatest fractional
meaningful change in
parameters: 0 => 0.01
Misc. controls:
(3) LMM initial factor
(4) print covariance matrix
in OUTFILE (1=yes,2=no)
default=yes for a single
component model or if
matrix is small [<16x16]
NITER max # iterations ( <1 => 40 )
------Self-Calibration-------
SOLMODE self-calibration mode
['' => none requested]
traditional modes:
'A' => ampl. self-cal
'P' => phase self-cal
'M' => per comp self-cal
'T' => force closure
'B' => ignore closure
!! SOLINT and DPARM used only
if self-cal not requested !!!
SOLINT solution interval [ sec! ]
(=< 0 --> 100 years)
DPARM Specifies which parameters to
solve for in each iteration
(1,2) start/stop solving for
self-cal phases
(3,4) start/stop solving for
self-cal amplitudes
(5,6) start/stop solving for
non- self-cal
(source) parameters
[default = always solve at
at all iterations]
WEIGHTIT 0.0 3.0 Modify data weights function
BADDISK Disks not to use for scratch
HELP SECTION
OMFIT
Task: OMFIT fits a source model to the UV data directly.
Optionally, this task can simultaneously self-calibrate
the data using the model fit. There is no pre-compiled
limit to the number of input visibilities.
THIS IS DEVELOPMENT CODE written by Ketan Desai who no longer
works for NRAO. Its support is questionable.
Adverbs:[out-of-date, see inputs list above instead]
INNAME.....Input UV file name (name). Standard defaults.
INCLASS....Input UV file name (class). Standard defaults.
INSEQ......Input UV file name (seq. #). 0 => highest.
INDISK.....Disk drive # of input UV file. 0 => any.
SRCNAME....Source name to be gridded. Must specify if input is
a multi-source data set, otherwise all sources are
included.
SUBARRAY...Subarray number to use. 0 => 1.
TIMERANG...Time range of the data to be used. In order:
Start day, hour, min. sec,
end day, hour, min. sec. Days relative to ref.
date.
ANTENNAS...List of antennas to use in selecting data. Only
baselines between antennas appearing on this list
are used. If the list is empty, all antennas are
allowed.
UVRANGE....Range of allowed baseline lengths in
kilowavelengths.
STOKES.....The desired Stokes type of the data:
'I','RR','LL', and 'HALF' are allowed
where ' ' defaults to 'I'
and 'HALF' processes 'LL' and 'RR' separately
FREQID.....Frequency identifier to select (you may determine
which is applicable from the OPTYPE='SCAN' listing
produced by LISTR).
BCHAN......Channel range to select.
ECHAN
BIF........IF range to select. Only one or all IFs
EIF can be selected.
DOCALIB....If true (>0), calibrate the data using information in the
specified Cal (CL) table for multi-source or SN table for
single-source data. Also calibrate the weights unless
DOCALIB > 99 (use this for old non-physical weights).
GAINUSE....version number of the CL table to apply to
multisource files or the SN table for single
source files. 0 => highest.
DOPOL......If > 0.5 then correct data for instrumental polarization
as represented in the AN or PD table. This correction is
only useful if PCAL has been run or feed polarization
parameters have been otherwise obtained. See HELP DOPOL
for available correction modes: 1 is normal, 2 and 3 are
for VLBI. 1-3 use a PD table if available; 6, 7, 8 are
the same but use the AN (continuum solution) even if a PD
table is present.
PDVER......PD table to apply if PCAL was run with SPECTRAL true and
0 < DOPOL < 6. <= 0 => highest.
BLVER......Version number of the baseline based calibration
(BL) table to appply. <0 => apply no BL table,
0 => highest.
FLAGVER....Specifies the version of the flagging table to be
applied. 0 => highest numbered table. <0 => no flagging
to be applied.
DOBAND.....If true (>0) then correct the data for the shape of the
antenna bandpasses using the BP table specified by BPVER.
The correction has five modes:
(a) if DOBAND=1 all entries for an antenna in the table
are averaged together before correcting the data.
(b) if DOBAND=2 the entry nearest in time (including
solution weights) is used to correct the data.
(c) if DOBAND=3 the table entries are interpolated in
time (using solution weights) and the data are then
corrected.
(d) if DOBAND=4 the entry nearest in time (ignoring
solution weights) is used to correct the data.
(e) if DOBAND=5 the table entries are interpolated in
time (ignoring solution weights) and the data are then
corrected.
BPVER......Specifies the version of the BP table to be applied.
<0 => no bandpass correction done.
SMOOTH.....Specifies the type of spectral smoothing to be applied to
a uv database . The default is not to apply any smoothing.
The elements of SMOOTH are as follows:
SMOOTH(1) = type of smoothing to apply: 0 => no smoothing
To smooth before applying bandpass calibration
1 => Hanning, 2 => Gaussian, 3 => Boxcar, 4 => Sinc
To smooth after applying bandpass calibration
5 => Hanning, 6 => Gaussian, 7 => Boxcar, 8 => Sinc
SMOOTH(2) = the "diameter" of the function, i.e. width
between first nulls of Hanning triangle and sinc
function, FWHM of Gaussian, width of Boxcar. Defaults
(if < 0.1) are 4, 2, 2 and 3 channels for SMOOTH(1) =
1 - 4 and 5 - 8, resp.
SMOOTH(3) = the diameter over which the convolving
function has value - in channels. Defaults: 1,3,1,4
times SMOOTH(2) used when input SMOOTH(3) < net
SMOOTH(2).
INFILE.....KEYIN format file containing additional control
parameters for the model-fit (see EXPLAIN OMFIT
for simple and detailed examples).
DOFIT......This is a new feature - and currently poorly
documented. Basically, this feature allows you
to enter a model at the POPS prompt instead of
specifying an INFILE.
OUTFILE....KEYIN format file to which to write final fitted model.
This file can be used as an INFILE for another invocation
of OMFIT. If the last character in the OUTFILE string is
a '+' (plus) character, the model will be appended to
the output file (or create a new one) all without the
trailing plus sign. Without the plus, the OUTFILE will
be written only if it is a new file.
PRTLEV.....Output print level
0 => just the facts please - gimme the answer!
1 => basic info about the data (un)used in the fit
2 => some intermediary information
3 => more intermediary information - recommended level.
4 => too much intermediary information
10 => WAY TOO MUCH information
[good for debugging - intermediary parameters
are displayed in internal units ]
OUTPRINT...Text file containing model-fit residuals
OUTNAME....
OUTCLASS...
OUTSEQ.....
OUTVERS....Content to write to output AIPS data set
0: data
1: data - [gain model * source model]
2: source model
3: gain model * source model
4: data / gain model
5: data / source model
6: data / [ gain model * source model ]
NOISE(1)...Pre-fit amplitude (a priori) noise in Jy. 0 -> 1
Note - this enters into CHI-squared displayed at the end
and should be the additional noise not reflected in the
weights. Thus if the weights are actually 1/sigma^2 in
Jy already, set NOISE(1) = 1 (the NEW default). If the
wights should be 100 times bigger to be properly
calibrated, set NOISE(1)=0,1,
Estimated thermal noise per requested Stokes visibility.
If you set STOKES='I','RR','LL',or ' ', give here the
thermal noise per I visibility. If set 'HALF', give here
the thermal noise per 'RR' or 'LL' [assumed the same].
NOISE(2)...Threshold for editing data in sigmas (0 -> 10). Flag
data when residual exceeds NOISE(2)*NOISE(2)*Chisq.
Data editing parameter. Data that differ from the model
by more than APARM(2)*(model-fit residual estimated for
current model) will be discarded at each iteration.
NOISE(3)...Low threshold for editing data in sigmas (0 -> 0). Flag
data when amplitude less than NOISE(3) * SQRT(Chisq).
Data editing parameter. Data with amplitudes below
APARM(3)*(model-fit residual estimated for current
model) will be discarded for each iteration.
Early stopping criteria:
APARM(1)...Chi-squared must change less than this amount (in percent) to
stop. 0 -> 1 (set < 0 to turn this off)
APARM(2)...Greatest fraction change in any parameter must be less
than APARM(2) to stop. 0 -> 0.01. (set < 0 to turn
this off)
APARM(3)...LMM initial factor - determines how fast OMFIT moves,
bigger values mean slower movement.
(-60 < APARM(3) < 60)
APARM(4)...Covariance matrix information will be written to OUTFILE
at each iteration if OUTFILE is set and APARM(4) not = 2.
Set = 1 to see these data when the number of global
parameters exceeds 16.
****** The following APARMs are no longer used *******
APARM(5,6).Defaults = 0.01,0.05. APARM(5) is a threshold value
to which the fractional change in Chi-Squared over the
previous iteration is compared. APARM(6) is a
threshold value to which the ratio of each parameter's
change to its estimated error bar is compared. If
the changes to Chi-Squared and all the parameter
change/error-bar ratios fall below these thresholds,
the iteration will terminate early because the model
is not changing. Set either APARM(5) or APARM(6)<0
to force the full number of iterations to be performed.
APARM(7)...Default = 0. [allowed range = -60 ... 60]
2**APARM(6) is the initial value to use for the Levenberg-
Marquardt Scaling parameter: GAMMA.
Large Positive GAMMA initially ignores covariances
between parameters and proceeds as a Steepest Descent
algorithm.
Large Negative GAMMA initially accepts all covariances
between parameters and proceeds as an Inverse Hessian
algorithm.
In ENGLISH: GAMMA controls how much attention the
model-fitting procedure initially pays to the covariance
between the fitted parameters. If GAMMA is large, the
model-fitter starts by varying each parameter independentl
of all others. If the parameters are highly covariant,
i.e. if varying one has much the same effect as varying
another, this is not a good idea. A small GAMMA tells the
model-fitter to pay attention to these covariances from the
beginning.
The program increases or decreases GAMMA at each iteration
depending on how successful the last step was in reducing
Chi-Square; APARM(6) just sets the initial value.
APARM(8)...Minimum SNR used to edit self-calibration solutions.
If the SNR of a self-calibration solution falls below
APARM(7), the weight of the SN table entry is set
negative -- it is NOT removed.
Default is not to reject any self-calibration solutions.
NB: currently, OMFIT's SNR calculation does not conform to
the rest of AIPS.
APARM(8) is currently disabled.
APARM(9)...APARM(9) controls one final criterion for baseline
selection and model refinement control. If self-cal is
selected: APARM(9)=0 forces baselines not part of a
triangle to be dropped while APARM(9)<>0 passes all
baselines. This behaviour is reversed when self-cal is
not selected: APARM(9)=0 passes all baselines, APARM(9)<>0
forces baselines to be a part of closing triangles.
SELF-CALIBRATION INPUTS:
NITER......Default = 40. Number of iterations to perform
SOLMODE....Self-Calibration mode selection string.
Default = ' ' (no self-calibration)
Allowed values:
'A' amplitude only self-cal
'P' phase only self-cal
'AP' full self-cal - amplitude and phase.
'MP' multiple phase-only self-cal
'MA' multiple amplitude-only self-cal
'MAP' multiple full - self-cal
"multiple" self-cal ==> independent self-cal for each
model component.
Additionally:
'T' explicitly requests that closure be forced
'B' explicitly requests that closure be ignored
default: closure is forced only if self-cal is selected
SOLINT.....Default = 36500*86400 [seconds]. Solution interval to use
if self-calibration is requested.
DPARM(1-2).specify at which iterations self-cal phases are solved for.
if self-cal phases not requested, phases are never solved
for.
default = 1,NITER IF phase self-cal requested via SOLMOD
DPARM(3-4).specify at which iterations self-cal amps are solved for.
if self-cal amps not requested, amps are never solved
for.
default = 1,NITER IF amp self-cal requested via SOLMOD
DPARM(5-6).specify at which iterations non-self-cal parameters are
solved for.
default = 1,NITER
WEIGHTIT...If > 0, change the data weights by a function of the
weights just before doing the solution. Choices are:
0 - no change weighting by 1/sigma**2
1 - sqrt (wt) weighting by 1/sigma may be more stable
2 - (wt)**0.25
3 - change all weights to 1.0
EXPLAIN SECTION
OMFIT: This task simultaneously self-calibrates and model-fits
UV data. The actual (non-linear) model-fitting is done
using the Levenberg-Marquardt method of least-squares
fitting (see, e.g., the _Numerical Recipes_ chapter on
modeling data, also Nash, Compact Numerical Methods for
Computers).
The matrix inversions are performed using LAPACK SVD code.
[Thank you LAPACK and NETLIB!!!!].
DOCUMENTOR: Michael Rupen, NRAO/Socorro.
PROGRAMMER: K. M. Desai, NRAO/SOC, UCSB, NRL/DC, NRAO/CV
RELATED TASKS: UVFIT, CALIB, [IMFIT, JMFIT, IMSTAT]?
PURPOSE
SOURCE MODELS
INITIAL GUESES FOR THE ANGULAR SIZE
WHY USE OMFIT
SIMPLE INFILE EXAMPLE
FULL DESCRIPTION OF INFILE FORMAT
COMPLICATED INFILE EXAMPLE
TERMINOLOGY and COMMENTS
FEATURES 'UNDER DEVELOPMENT'
PURPOSE
=======
OMFIT is a model-fitting program that fits the UV-data directly.
OMFIT can fit for different modeltypes simultaneously. Individual
models may span different ranges of frequency space with [possible]
overlap. Different Stokes parameters may be selected for fitting;
a few models can be fit to multiple STOKES parameters simultaneously.
The following types of models are currently implemented:
The Source Models
=================
The following models are currently supported:
Model Type #parameters parameter description parameter unit *1*
'DOT' Point 3 Flux density of point source [Jy]
source Offset east of phase center [mas]
Offset north of phase center [mas]
'POL' Point 4 Flux density of point source [Jy]
source Stokes polarization [AIPS stokes code]
Offset east of phase center [mas]
Offset north of phase center [mas]
'GAU' Gaussian 6 Flux density of Gaussian [Jy]
Offset East of phase center [mas]
Offset North of phase center [mas]
Major axis angular size [mas]
Axial ratio (minor/major) [dimensionless]
Position Angle of major axis [deg. East of
North on the sky]
'DISK' Disk 7 Flux density of Disk [Jy]
Offset East of phase center [mas]
Offset North of phase center [mas]
Major axis [mas]
Minor axis [mas]
Position Angle of major axis [deg. East of
North on the sky]
*2* Order of the bessel function [dimensionless]
'SPHR' Disk 5 Flux density of spherical shell [Jy]
Offset East of phase center [mas]
Offset North of phase center [mas]
*4* Average radius [mas]
*4* Shell f [dimensionless]
'MAS' Maser 10 Flux density of Gaussian [Jy]
Offset East of phase center [mas]
Offset North of phase center [mas]
Major axis [mas]
Axial ratio (minor/major) [dimensionless]
Position Angle of major axis [deg. East of
North on the sky]
Center of Gaussian spectral
line profile [channels]
FWHM of Gaussian spectral
line profile [channels]
Drift Eastwards per channel
of position from center
of spectral line profile [mas/channel]
Drift Northwards per channel
of position from center
of spectral line profile [mas/channel]
'GAINS' Gain #IF*#POL Antenna number [as in AN table]
station gains [dimensionless]
ordered as either
IF1 IF2 ... IFn
or
IF1RR IF1LL IF2RR ... IFnLL
*1* Any modeltype specified using a 'Y' prefix, eg 'YDOT', will have all its
angular size parameters (major axis, position offsets) interpreted in
arcseconds instead of milliarcseconds (mas).
*2* Suitable choices for this parameter allow the use of a ring or a thin
sphere model [see below].
*3* See full infile below for an implementation of GAINS. A
triplet consisting of antenna number, mean gain, and vary flag may be
specified for each antenna in the data set.
**NB** when amplitude self-calibration is specified, although specified
gain values are read in and applied, they are not solved for.
*4* The shell has inner radius r_in and outer radius r_out. The
fitted radius, r_geo, is the geometrical mean of r_in and r_out
(r_geo = sqrt(r_in*r_out)). Shell-f works this way:
r_out = r * (1 + shell-f/2.)
r_in = r / (1 + shell-f/2.)
shell-f is close to but not equal to (r_out - r_in)/r_geo Ie. if
r_out = 1.25 and r_in = 1.0, then r_geo is 1.118 and shell-f = 0.23607
Initial guesses: angular size:
==============================
If the angular size is specified in arcseconds and a fiducial baseline length
is determined using UVPLT, the product should equal a scale-factor which is
typical of each modeltype:
Modeltype Angular size def Fiducial baseline length Scale factor
GAU FWHM Vis fcn half power pt 91.1
DISK Uniform disk radius First null of vis fcn 240.0
DISK Ring radius 150.0
DISK Thin sphere radius 280.0
The DISK model generates a thin sphere when n = 1.
The DISK model generates a uniform ring when n = -2.
The DISK model generates a uniform disk when n = 0.
As may be apparent from the above list, different types of models are
being added and requests will be accepted.
WHY USE OMFIT?
==============
OMFIT has the following advantages over UVFIT:
+) OMFIT does not have any built-in limit on the number of visibilities
to be fit. The data do not have to be pre-averaged to fit within a
fixed maximum number of visibilities; nor is any time-averaging
done within the task itself.
+) OMFIT can simultaneously solve for different model types. This means
you can request that, e.g., an Elliptical and a Circular Gaussian be
fit simultaneously to the data.
+) OMFIT can self-calibrate the UV data while model-fitting.
+) OMFIT can also perform multiple self-calibration, i.e., solve
for different self-calibration solutions in different directions on
the sky. This may be useful for low-frequency work, or when using
the phased-VLA for VLBI work.
+) OMFIT can solve for antenna-based IF and polarization dependent
time-independent gain factors. This can be useful for transferring
a final amplitude scale from a calibrator to a target source.
OMFIT's most significant disadvantages are:
-) OMFIT runs more slowly than UVFIT. This is because...
- OMFIT does not load the visibilities into memory
- self-calibration involves lots of overhead
- multiple self-calibration has even more overhead
Fortunately, self-cal and multiple self-cal are not usually necessary,
and turning them off recovers much of the speed of UVFIT. OMFIT now
uses dynamic memory allocation which should also help recover some of
the overhead. Gains do not cost much in terms of speed and should be
used instead of amplitude self-cal whenever possible.
SIMPLE INFILE EXAMPLE
=====================
The INFILE can be quite complicated, so this description begins with the
simplest possible file:
! ## MINIMAL INFILE
!------------------------------------------------------------------------
! This is a minimal INFILE
!
! Type ID Bchan Echan FlTot vary? Epos vary? Npos vary? Maj vary?
! Min/Maj vary? PA vary?
'GAU' 1 7 7 45.0 'T' 5.0000 'T' 7.0000 'T' 8.5000 'T'
0.90 'T' 9.0 'T' /
/
! DO NOT REMOVE
!------------------------------------------------------------------------
! ## END OF MINIMAL INFILE
This infile requests fitting one component, labelled #1 (Icomp) in I/O,
in channel 7 (Bchan, Echan). The initial total flux is 45 Jy (FlTot);
the source is located (5, 7) mas (East, North) of the phase-center
(Epos, Npos). The model is
to be an elliptical Gaussian (Type) with initial parameters:
major axis= 8.5mas (Maj), minor/major axis axial ratio= 0.9 (Min/Maj),
position angle for the major axis= 9 degrees East of North (PA). All
parameters are allowed to vary during the fitting process.
**** The comment line entitled "DO NOT REMOVE" is required for proper
operation of the KEYIN input subroutine used by OMFIT. Since
generation of a suitable INFILE is often frustrated by KEYIN's
fussiness, it is recommended that this line not be left out, not
even accidentally [thats not funny...] ****
A more complicated sample INFILE follows the full description of the
INFILE format.
FULL DESCRIPTION OF INFILE FORMAT
=================================
INFILE is used to provide basic control parameters and is divided into
logical lines.
Logical lines are delimited by the '/' character.
The '!' character serves as a comment character.
Comment lines CANNOT appear _within_ a single logical line, i.e.,
1 10 101 'DOT' ! oops
5 'T' 1.0 'T' 2.0 'T' /
is NOT legal, while
1 10 101 'DOT'
5 'T' 1.0 'T' 2.0 'T' / ! whew
is OK.
An INFILE may have multiple logical lines; it is terminated by a single
otherwise blank logical line.
Description of logical line format (model specification):
Each logical line corresponds to one component to be fit, and consists of
a number of entries separated by blanks (or carriage returns):
1st entry: MODELTYPE component type, eg 'DOT', see list below
2nd entry: ID source number for multi-source files;
used as an "identification number" for both
single and multi-source files.
3rd entry: BEG-CHAN first channel in which this component exists
4th entry: END-CHAN last channel in which this component exists
Sub. entries: Pairs of parameters for the model. Each pair
consists of a parameter value (initial guess),
followed by 'T' or 'F':
'T' -> find the best fit to that parameter
'F' -> keep that parameter fixed
Last entry (/): character indicating end of logical line
==============================================================================
COMPLICATED INFILE EXAMPLE
==========================
! ## BEGIN COMPLICATED SAMPLE 'INFILE'
!------------------------------------------------------------------------
! This is a ridiculously complicated infile
! Mtype ID Bch Ech Flux flag EastPos flag NorthPos flag ...
'YGAU' 469 1 1 3.25208 'T' 82.31136 'T' -76.53779 'T'
.69031 'T' .96562 'T' 39.51926 'T' /
'YGAU' 468 1 1 2.33739 'T' 81.62495 'T' -83.76611 'T'
1.02294 'T' .95342 'F' 86.40079 'T' /
'YDOT' 467 2 2 .11282 'T' 65.22981 'T' -76.88685 'T' /
'GAU' 465 3 3 .11964 'T' -89554.05 'T' 80003.69 'T'
1001.43 'T' .99732 'T' -12.58175 'T' /
'YDOT' 462 3 3 1.04904 'T' -66.15225 'T' 85.19714 'T' /
'GAINS'
2 2.0 'T' /
'GAINS'
0 1.1 'T' /
'GAINS'
4 1.5 'F' /
/ ! end inputs
! Required Blank Line - leave alone !
! ## END COMPLICATED SAMPLE 'INFILE'
Note the following characteristics about this example:
- This file specifies a model for three channels.
- Each channel is modelled independently.
[there are no features that span more than one channel.]
- Channel 1 has two gaussian components, the axial ratio and
position angle component 468 will not be solved for.
- Channel 2 has only a single point source component.
- Channel 3 has both a point and a gaussian component.
- The size and position parameters of the gaussian component
in channel 3 are specified in milliarcseconds.
- The default gain information overrides all previously encountered
gain information. ***
- All antennas will start with gains of 1.1 except for antenna 4 which
will start with 1.5.
- The mean gain factors specified here will multiply the model visibility.
Eg, initially the model visibility on baseline 2-4 will be
multiplied by 1.1*1.5 = 1.65 .
- mean gain factors will be determined for all stations except station 4.
*** except *** if amplitude self-calibration was requested through
SOLMODE = 'A', 'AP', or 'MAP'; in that case, mean gains will
not be determined for any station.
- The order of the logical lines is arbitrary [except for the default gain
specification as described above].
============================================================================
TERMINOLOGY and COMMENTS
========================
[3apr98 - outdated section (kmd) ]
global parameters - these are all model parameters that are not self-cal
parameters, eg source flux, major axis size, minor axis size, position
angle, ra offset, dec offset, etc.
local parameters - these are all model parameters that are self-cal
related and come in two flavours, local phases which are self-cal phases,
and local amplitudes which are self-cal amplitudes.
DoF - The number of observed visibilities minus the number of model
parameters.
"Discarded ?? global and ??? local params" - the singular value decomposition
method used to invert the normal matrices may choose to discard some
singular values [effectively discarding linear combinations of model
parameters] based upon the input value of CUTOF. This message simply
warns the user that some linear combinations of his/her parameters are
being discarded in the model fitting procedure.
1-sigma error calculation method - The principal problem with calculating
error bars in AIPS is that the weights, while possibly well known in a
relative sense, are unknown by an overall scale factor. If this were not
so, SWSR would be the Chi-Square statistic, RSWSR would be the reduced
Chi-Square, and the goodness-of-fit of the model could be evaluated.
OMFIT reports two estimates of the 1-sigma error. The 'formal' 1-sigma
error is shown in the OUTFILE output and uses the user estimated thermal
noise per visibility [default value = 1 Jy]. The 'suggested' 1-sigma
error is printed to screen when OMFIT runs and also is sent to the OUTFILE.
This error assumes that the model is a perfect representation of the data
so that the post-model fit residual is also an estimate of the thermal noise
per visibility. This error bar is probably what you want since presumably
you do think that the model fits the data well.
OMFIT is perfectly happy to allow you to solve for self-calibration
parameters and source position offsets. It is important to realize that
these two types of parameters are somewhat degenerate - especially when
there is only one source component in the model.
============================================================================
FEATURES 'UNDER DEVELOPMENT'
============================
See the section immediately following the local includes in
$QPGOOP/OMFIT.FOR .
You may request additional features, or report bugs via email to
kdesai@nrao.edu .
============================================================================
Comments from an M. Bietenholz, an experienced user of OMFIT
It will probably never be particularly easy to use. Its a bit clunky
because in typical usage, you have to have your input text file open
in an editor outside of AIPS to make changes between runs of OMFIT,
but there is probably no reasonable way of moving all the possible
model parameters to adverbs. It is also sensitive to the input file
format.
I always use it on uv-files which have already had the vis. averaged
together across both freq. and IF, I can't remember whether this is
just to reduce number of points to fit or if OMFIT doesn't work for
multiple IFs.
A related but larger issue is that very often, the dominant errors in
the visibilities are in fact calibration errors not noise, which are
highly correlated from 1 vis to the next in the typical case of
vis.recs. being a few secs. Using such small vis records means that
OMFIT uses a huge number of degrees of freedom, and gets uncertainties
which are ridiculously small. I usually UVAVG to scans (for typical
phase-ref VLBI scan lengths of ~1-2 min) before feeding to OMFIT.
However, the issue of the real number of degrees of freedom and
getting usable uncertainties does not have an easy solution,
so probably some intelligence and diligence on the part of users
will be unavoidable.
I also think OMFIT does not recognize degenerate cases, such as
amplitude self-cal *and* fitting your source flux density. I don't
remember what happens when you do this, I think it happily converges
somewhere or other but doesn't necessarily let you know that the
uncertainty on your fitted flux density is actually infinite. There
is some code to try an prune degenerate parameters from the fit, but I
think it only works for single parameters that have no effect on the
fit, e.g., p.a. of a *circular* Gaussian. I don't know whether it
would be easy to check fo the most common cases of degenerate
combinations of parameters - ie. you need at least one source
flux-density or gain-amplitude parameter NOT to be free.
The other issue that will be difficult to avoid is that you need a
reasonable starting model. Its not uncommon for there to be local
minima in the Chi2 hypersurface. Getting the starting position of your
model wrong by more than a beamsize or so is likely to converge to
some local minimum in the Chi2 space that has nothing to do with the
best fit. My 2-dim picture is that the Chi2 surface has some
concerted dip near the best fit, where you have some hope of finding
the global min by crawling downslope, but far away from the best-fit
its probably just some random hilly surface, so crawling downslope
will get you to one of an infinite number of local minima.
I don't think there is much than can be done about this. The
parameter space is probably generally too big to allow a brute-force
grid search amongst different starting points (a la difmap), so
probably the only way to get somewhat reliable results out of OMFIT is
to a) have a pretty good idea of your starting points, especially
positions and b) to run it several times. My recipe is usually to
give it starting parameter vector [x1], let it converge (hopefully) to
[y], and then restart it with a starting point on the opposite side of
[y], ie. something like [2*y - x1]. With luck it will again converge
to [y] and you're good. If not, a few more runs are in order. And
that is just the sourc-model parameters, I've been mostly taking it on
faith that starting the selfcal parameters at amp=1,phase=0 is okay.
---- notes from another time ---------------
You probably want to set APARM(1) to something quite small, I usually
use ~1/(5*N_Vis), 1 percent is not good for lots of vis. To make my life
easier, I usually run OMFIT to get a reasonable fit first to figure
out the noise level (rms per vis or something like in OMFIT output),
and then set NOISE(1) to that level. This doesn't change anything in
the fit, it just scales the Chi2 values that get printed out, so that
they wind up close to Degrees_of_Freedom and its easier to judge the
significance of changes to the fit. A gotcha here is that I think
Ketan did some wierd scaling of Chi2 internally. As long as you don't
change the vis data set you are okay (ie. Chi2 is arbitrarily scaled
but comparisons of different Chi2 values for different fits are
useful). If, however, you change the vis data, then it's dangerous to
compare Chi2 values. Another one of my hacks is to print out the
unscaled Chi2 so you can estimate the significance of e.g., throwing
out an antenna.
Once you've got a fit, it is best is to re-run OMFIT with a few
different starting values, usually not too far from the fitted values,
but try and start it with starting values on on different sides of the
fit point to make sure you are finding a global, not a local minimum.
Ie, run it with a starting value of major=1.0mas; converges to 1.532
mas; so rerun it with a starting value of 0.5mas - do you still get
1.532 mas? Repeat w/ starting value of 3mas.
Total flux density is usually pretty robust against this, but for
things like disk size (if vis-fn has nulls in) this can be a gotcha.
If you are fitting antenna gains as well, convergence can be an issue,
I've resorted to multiple runs of OMFIT with a fixed e.g., disk radius
to build up manually a disk radius vs. chi2 curve. It uses log
format for e.g., disk radii, so best to use a non-zero but
small compared to resolution value.
Check also that it *has* converged; I use PRTLEV 3; in some cases even
with e.g., NITER=500 it will not formally converge. I think it is
usually close to the an actual min in Chi2 in these cases so the
final fitted values are probably okay, but - hey, you're in the world
of far too many free parameters.