SDCLN: Task to deconvolve image with CLEAN & Steer/Dewdney/Ito
INPUTS
Dirty image
INNAME Image name (name)
INCLASS Image name (class)
INSEQ 0.0 9999.0 Image name (seq. #)
INDISK 0.0 9.0 Image disk drive #
BLC 0.0 Used to specify plane.
Beam image
IN2NAME Image name (name)
IN2CLASS Image name (class)
IN2SEQ 0.0 9999.0 Image name (seq. #)
IN2DISK 0.0 9.0 Image disk drive #
Clean image
OUTNAME Image name (name)
OUTCLASS Image name (class)
OUTSEQ -1.0 9999.0 Image name (seq. #)
OUTDISK 0.0 9.0 Image disk drive #
INVERS -1.0 46655.0 CC file ver. no.
GAIN * 0.0 2.0 CLEAN loop gain
FLUX * 0.0 Min. residual in window (Jy)
PHAT 0.0 100.0 Prussian hat size.
NCOUNT * -1.0 # pixels > CUTOFF*peak to
switch from CLEAN to SDI
0 => no SDI method
CUTOFF 0.0 1.0 Cutoff for inclusion in SDI
STFACTOR * 0.0 1.0 SDI loop gain
NITER * Maximum # of CLEAN components
BITER Begin at BITER component
BMAJ * -999.9 FWHM(asec) maj. axis CLEAN
restoring beam.
BMIN * -999.9 FWHM(asec) min. axis CLEAN
restoring beam.
BPA * -360.0 360.0 CLEAN beam position angle
NBOXES 0.0 50.0 Number of boxes for CLEAN
CLBOX 0.0 4096.0 Four coordinates for each box
FACTOR * -5.0 5.0 Speedup factor see HELP SDCLN
MINPATCH * 0.0 1024.0 Min. beam halfwidth in AP.
MAXPIXEL * 0.0 500000.0 Maximum pixels searched in
each major cycle.
DOTV * -1.0 1.0 TV display of residuals?
BADDISK -1.0 1000.0 Disks to avoid for scratch.
HELP SECTION
SDCLN
Type: Task
Use: SDCLN does an AP-based CLEAN of the type devised by B. G.
Clark. CLEAN components are found by CLEANing the
brightest points in the residual map with a small area of
the dirty beam. More exact CLEANing is then done by
transforming the CLEAN components, multiplying by the
transform of the dirty beam, then transforming back to the
map plane and subtracting the result from the dirty map.
SDCLN may be restarted if further CLEANing is desired. The
dirty beam may be modified by adding a delta function of
amplitude PHAT centered on its peak. TJC claims that this
encourages smoothness in the clean map. SDCLN differs from
APCLN in that it offers the option to switch from the
standard CLEAN method of finding components to a Steer -
Dewdney - Ito (SDI) method. In the latter, all pixels
having residual flux > CUTOFF times the current peak
residual are taken as components. This method should avoid
the "ripples" produced by the standard CLEAN on extended
emission.
A single plane in a cube may be processed per run; the
resulting CLEAN (or residual) image will be put in the
corresponding plane of the output image. Other planes in
the output image will not be affected, i.e. are likely to
contain garbage. When several runs of SDCLN are made to
CLEAN a cube be sure to completely specify the output
image. A clean components table will be created with a
sequence number corresponding to the plane number. At the
end of each execution of SDCLN, the output components table
is compressed so that a pixel will appear only once in the
table. This avoids the necessity of running CCMRG.
SDCLN now partially supports assymmetric beams, see the
EXPLAIN section for more details. SDCLN supports TELL
including some change of parameters: see SHOW SDCLN.
Adverbs:
INNAME......The dirty map image name. Standard defaults.
INCLASS.....The dirty map image class. Standard defaults.
INSEQ.......The dirty map image sequence number. 0 => high
INDISK......The dirty map image disk drive no. 0 => any
BLC.........The third and higher dimensions are used to
specify the plane in the dirty and residual
images to be processed. NOTE: process in
increasing plane number.
IN2NAME.....The dirty beam image name. blank => actual
INNAME, otherwise standard defaults.
IN2CLASS....The dirty beam image class. Standard defaults
except blank => 'RBM001' if INCLASS = 'RIM001'
'LBM001' if INCLASS = 'LIM001'
'IBM001' if INCLASS = anything else
IN2SEQ......The dirty beam image sequence number. 0 => actual
INSEQIN2DISK.....The dirty beam image disk drive no. 0 => any
OUTNAME.....The clean map image name. blank => Standard
defaults based on INNAME.
OUTCLASS....The clean map image class. Standard behavior with
default = 'QCL001' if INCLASS = 'QIM001'
'UCL001' if INCLASS = 'UIM001'
'VCL001' if INCLASS = 'VIM001'
'RCL001' if INCLASS = 'RIM001'
'LCL001' if INCLASS = 'LIM001'
'ICL001' if INCLASS = anything else
OUTSEQ......The clean map image seq. no., 0=> highest unique
If >0; image will be created if new,
overwritten if image name exists.
OUTDISK.....Clean Map disk drive no., 0=> highest with space
INVERS......The CC file version number. 0=> highest
current one. -1 => create new one. If none
exist, one will be created in any case.
GAIN........The CLEAN loop gain. 0 => 0.10
Can be reset with TELL.
FLUX........Stop CLEAN when abs(resid. image max in CLEAN
window) < FLUX (Jy). Can be reset with TELL.
PHAT........The offset added to the pixel at the beam peak.
Should be about 0.1 to 0.5
NCOUNT......The CLEAN method is used until the number of
pixels included in a SDI iteration exceeds
NCOUNT. 0 => do not use the SDI method.
Can be reset with TELL.
CUTOFF......All pixels > CUTOFF times the current peak
residual are considered components. If the 2nd
highest point in the beam is B2, CUTOFF must be
> B2. 0 => 1.01 * B2.
STFACTOR....The loop gain (fraction of each pixel's flux taken
as a component flux) during the SDI phase of the
deconvolution. It must be <= the default.
0 => lower of 0.9 * (1-CUTOFF) and GAIN.
Can be reset to a lower value with TELL.
NITER.......CLEAN iteration limit. 0 => 200
If <0 then the CLEAN will stop at the first
negative component or ABS (NITER) whichever is
encountered first. Can be reset with TELL.
BITER.......Begin new CLEAN at iteration BITER.
If >0, you must completely specify the
OUTNAME, OUTCLASS and OUTSEQ of the clean map.
BMAJ........The FWHM (asec) major axis of the restoring beam.
If 0; value obtained from dirty image.
If <0; output will contain the residual image.
Can be reset with TELL.
BMIN........The FWHM (asec) minor axis of the restoring beam.
Can be reset with TELL.
BPA.........The position angle in the unrotated image of BMAJ.
Can be reset with TELL.
NBOXES......Number (<=50) of rectangular search boxes. 0 => 1
CLBOX.......A 4x50 array with the BLC and TRC of each box.
0 => use inner quarter of input map (first box)
0 => ignore box (boxes 2 - 50)
FACTOR......FACTOR>0 causes deeper CLEAN in each major cycle,
speeding CLEAN, maybe "eating" extended structure.
FACTOR=0 => the normal Clark CLEAN. FACTOR=-0.3 is
good for deep CLEANs of extended structure.
Can be reset with TELL.
MINPATCH....Minimum half width of the portion of the beam
which is used in the AP minor CLEAN. < 3 => 127.
Use 51 for deep CLEANs of extended sources.
Can be reset with TELL. Upper limit 1024.
MAXPIXEL....The maximum number of pixels that are searched for
components inside the ``AP'' in each major cycle.
< 1000 => 20050. This number affects the cpu
usage significantly. Too many causes the task to
search over many points it will never use. Too
few causes the task to do many more small major
cycles, also at great expense. Use this with
caution, but big wins are possible using larger
sizes on very large cleans. Can be changed by
TELL.
DOTV........Display residuals on TV channel 1. >= 0 => yes.
If true, you may stop the CLEANing with TV button
D after each residual map is displayed.
Can be reset with TELL.
BADDISK.....This array contains the numbers of disks on which
it is desired that scratch files not be located.
BADDISK has no effect on input and output maps.
EXPLAIN SECTION
SDCLN: Task which cleans an image using the BGC-AP algorithm
RELATED PROGRAMS: UVMAP,APCLN,IMAGR,CALIB,UVSUB,CCMOD,PRTCC
PURPOSE
SDCLN deconvolves a dirty beam from a dirty map image using
the CLEAN algorithm [Hogbom 1974] as modified to take advantage
of the Array Processor [Clark 1980]. At some stage, specified
by the user, SDCLN will switch to a Steer-Dewdney-Ito (SDI)
[1984] method for finding the model components.
CLEAN iteratively constructs discrete approximants F(n) to
a solution F of the convolution equation:
B^F = D (1)
where D denotes the discrete representation of the dirty map and
B of the dirty beam, the symbol ^ here denoting convolution. The
initial approximant is F(0)=0 everywhere. At the n'th iteration,
CLEAN searches for the extremum of the residual map R determined
at the (n-1)'th iteration:
R(n-1) = D - B^F(n-1) (2)
A delta-function "CLEAN component", centered at this extremum,
and of amplitude g (the loop GAIN) times its value, is added to
F(n-1) to yield F(n). The search over R is restricted to an
area A called the "CLEAN window". A is specified as a number
NBOXES of rectangular sub-areas, or CLBOXes.
Iterations continue until either the number of iterations
n reaches a preset limit N (=ABS(NITER)), or the absolute value
of the extremum in the CLEAN window of the residual map
decreases to a preset value FLUX. If NITER is negative then the
CLEAN will stop at the first negative component. This option
should be used with caution and never be used on the last CLEAN
of an image. CLEAN uses negative components to correct when
too much flux has been subtracted from a pixel and does not
necessarily mean that the noise has been reached.
To diminish any spurious high spatial frequency features in
the solution, F(N) is normally convolved with a "hypothetical"
Gaussian "Clean Beam" H to construct a final "Clean Map" C:
C = H^F(N) + R(N) (3)
The clean beam H may be specified by the user through the
parameters BMAJ, BMIN, BPA, or it may be defaulted to an
elliptical Gaussian fitted to the central region of the dirty
beam B. SDCLN writes the array of "Clean Components" F(N) to
the CC extension of the clean map image file.
The Clark algorithm speeds up the deconvolution process by
splitting it into "major" and "minor" iteration cycles. At the
beginning of the m'th major cycle, it loads into the AP a
RESTRICTED residual map R'(m) containing only the LARGEST
(positive and negative) values in the current residual map R(m).
It then performs a "minor" cycle of iterations wherein new CLEAN
components are sought with (a) the restricted residual map R'(m)
replacing the full residual map R and (b) the dirty beam B being
approximated by its values inside a small area (the "beam
PATCH") with zeroes outside.
A minor cycle is terminated at iteration n' when the peak
in the restricted residual map R'(n') falls to a given multiple
[Clark 1980] of the largest value that was ignored in R(m) when
R'(m) was passed to the the AP. At the end of the cycle of
minor iterations, the current clean component list F(n') is
Fourier transformed, multiplied by the transform of the beam B,
transformed back into the map plane, and then subtracted from
the dirty map D, thereby performing step (2) EXACTLY with the
components list F(n') obtained at the end of the minor cycle.
Errors introduced in the minor cycle through use of the
restricted beam patch are corrected to some extent at this step.
This ends the m'th major cycle, the (m+1)th beginning when the
new restricted residual map R'(m+1) is loaded into the AP.
CLEANing ends (with the transform steps used at the end of a
major cycle) when either the total number of minor iterations
reaches NITER, or the residual value being CLEANed at a minor
iteration reaches FLUX.
The SDI method of finding components differs from the
CLEAN method described above. When there are many pixels in
the residual map of intensity nearly equal to that of the
maximum residual, it no longer makes sense to select just one
pixel at a time. The SDI method selects instead all pixels
having residual intensity greater than some fraction of the peak
residual. For this method to work correctly on a point source
located at a pixel, this fraction must be greater than the 2nd
highest point in the beam and SDCLN forces the fraction (set by
the user as CUTOFF) to be in the range Bp2-0.98. Once a pixel
has been selected by this method, it should in most cases be
included in every major cycle. For this reason, the loop gain
must be < (1-CUTOFF). The actual loop gain at any major cycle
is smaller than this to account for the convolution of the other
selected pixels onto the pixel having the maximum value. In
SDCLN, each selected pixel is regarded as a clean component and
written to the clean component file. This selection is the
equivalent of the minor cycle iterations described above. SDCLN
then removes the components found in the minor cycles by the
same method used for CLEAN. Major cycles continue until the
number of components found is greater than or equal abs(NITER).
SDI components are restored to the output image by the same
convolution used for the CLEAN components. Because the same
pixels will appear many times in the output components file,
this file is sorted, merged and resorted to combine all
components at the same pixels. This much smaller CC file will
speed later processing by CALIB, UVSUB, and all other tasks which
must use the CC model data.
An alternative Maximum entropy deconvolution program is
available in AIPS as task VM but it is experimental and
slow.
COMMENTS
Don't Run SDCLN ...
===================
If there is so little signal in your map that no sidelobes
of any source in it exceed the thermal noise, then no sidelobe
deconvolution is necessary, and running SDCLN is a waste of your
time and of CPU cycles.
General - You can help SDCLN when you run UVMAP
===============================================
Other things being equal, the accuracy of the deconvolution
process is greatest when the shape of the dirty beam is well
sampled. When mapping complicated fields, it is often necessary
to compromise between cell size and field of view; if you are
going to CLEAN a map image, you should set up your parameters in
UVMAP so that there will be at least three or four cells across
the main lobe of the dirty beam.
It is also important to make the map size large enough that
no strong sources whose sidelobes will affect your map have been
aliased by the FFT (preventing SDCLN from handling them
correctly). Consider making a strongly tapered map of a wide
field around your source at low resolution to diagnose confusion
before running SDCLN on a high resolution map (especially when
processing snapshot data from the lower VLA frequencies).
It is helpful to regard CLEAN as an attempt to interpolate
missing samples in the (u,v) plane. The accuracy of the
interpolation is greatest where the original sampling is dense
or where the visibility function varies slowly. The accuracy is
least where you ask CLEAN to EXTRApolate into poorly sampled or
unsampled regions of the (u,v) plane where the visibility
function changes rapidly.
One such region is the center of the (u,v) plane in any VLA
map made from data where all of the fringe visibilities were
less than the integrated flux density of the source. You can
help CLEAN to guess what may have happened in the center of the
(u,v) plane (and thus to guess what the more extended structure
on your map should look like) by including a zero-spacing flux
density when you make your map with UVMAP. This gives CLEAN
a datum to "aim at" in the center of the (u,v) plane. Extended
structure can often be reconstructed well by deep CLEANing when
the zero-spacing flux density is between 100 percent and 125 percent of the
average visibility amplitude at the shortest spacings. If your
data do not meet this criterion, there may be no RELIABLE way
for you to reconstruct the more extended structure. (Some cases
with higher ratios of zero-spacing flux density to maximum
visibility amplitude can be successfully CLEANed, but
success is difficult to predict). If you see an increase in the
visibility amplitudes on the few shortest baselines in your
data, but not to near the integrated flux density, you may
get better maps of the FINE structure from SDCLN by excluding
these innermost baselines when you run UVMAP.
Another unsampled region lurks in the outer (u,v) plane in
many VLA maps of sources at declinations south of +50, if the
source has complicated fine structure. To see why, consult the
plots of (u,v) coverage for the VLA in Section 4 of the "Green
Book" [Hjellming 1982]. At lower declinations, some sectors of
the outer (u,v) plane are left poorly sampled, or unsampled,
even by "full synthesis" mapping. (There are missing sectors in
the outer (u,v) plane in ANY snapshot map). If the visibility
function of your source has much structure in the unsampled
sectors, CLEAN may work poorly on a high resolution map unless
it gets good "clues" about the source structure from the
well-sampled domain. If the clues are weak, badly extrapolated
visibilities in the unsampled regions can cause high frequency
ripples on the CLEAN map. In such cases, CLEAN may give maps
with better dynamic range if you are not too resolution-greedy,
and restrict your data to the well-sampled "core" of the (u,v)
plane when you run UVMAP.
Before applying CLEAN, examine your (u,v) coverage and
think whether you will be asking the algorithm to guess what
happened in such unsampled regions.
Frailties, Foibles and Follies
==============================
There are excellent discussions of CLEAN's built-in
idiosyncracies by Schwarz (1978, 1979), by Clark (1982) and by
Cornwell (1982).
Another way of looking at CLEAN is to think of it as
attempting to answer the question "What is the distribution of
amplitudes at the CLEAN component positions [F(N)] which best
fits the visibility data, if we define the sky to be blank
everywhere else ?" The algorithm can then be thought of as
a "search" for places where F should be non-zero, and an
adjustment of the intensities in F(N) to obtain the "best"
agreement with the data.
The reconvolution of F(N) with the hypothetical "clean
beam" H produces a "clean map" C whose transform is no longer a
"best fit" to the data (due to differences between the
transforms of H and of the dirty beam B). The merit of the
reconvolution is that it produces maps whose noise properties
are pleasing to the eye. It may also be used to "cover up"
instabilities in CLEAN stemming from poor extrapolation into the
unsampled regions of the (u,v) plane, by making H significantly
wider than the main lobe of B.
Note also that step (3) of the standard CLEAN combines this
reconvolution with the residual map, which contains faint sky
features convolved with the DIRTY beam B. If there is
significant signal remaining in the residual map, the effective
resolution of the Clean Map C varies with brightness. You must
therefore be particularly careful when comparing Clean maps made
at different frequencies or in different polarizations; you
should CLEAN all such maps sufficiently deeply that the
integrated flux density in the CLEAN components F(N) is much
greater than that in the residual map R(N).
A recurrent question about CLEAN concerns the uniqueness of
the Clean Map. In the absence of noise, CLEAN could adjust the
amplitudes of the components in F(N) to minimize the rms
difference between the observed visibility function and the
transform of F(N) [Schwarz 1978, 1979]. If the number of
degrees of freedom in F(N) is less than the number in the data,
CLEAN can (and in many practical cases does) converge on a
solution that is sensibly independent of your input parameters.
Noise and approximations in the algorithms complicate this
[Cornwell 1982], but realise that the solution CANNOT be unique
if the number of positions at which you end up with CLEAN
components exceeds the number of independent (u,v) data points.
Be suspicious if your Clean Map contains structures which
resemble those of the dirty beam. This may mean either that you
have not CLEANed deeply enough, or that SDCLN has had
difficulty in some unsampled sector of the (u,v) plane in your
case. This test is particularly important in the case of
snapshot maps, for which the sidelobes of the dirty beam have a
pronounced star (snowflake) structure.
GAIN and NITER
The depth to which SDCLN carries out its deconvolution is
approximately measured by the product NITER*GAIN. The value of
NITER and the execution time needed to reach a given CLEANing
depth are minimized by setting GAIN = 1.0, but setting
GAIN > 0.5 is recommended only when removing the sidelobes of
a single bright unresolved component from surrounding fainter
structure.
When CLEANing diffuse emission, GAIN = 0.1 (the
default) will be much better, for the following reason. The
search step of the algorithm begins its work at the highest peak
(which in an extended source may be a random noise "spike").
After one iteration, the spike is replaced in R by a negative
beamshape, so the next highest peaks are more likely to be found
where the spike would have had its biggest negative sidelobes
[see the diagram on p.11 of Clark (1982)]. If GAIN is high,
subsequent iterations tend to populate F(n) around the negative
sidelobe regions of the highest peaks. This "feedback" can be
turned off by making GAIN small enough that the negative
sidelobes of the first peaks examined in an extended structure
are lost in the noise, i.e. GAIN * (worst negative sidelobe
level) < signal-to-noise on the extended structure. In practice
setting GAIN << 0.1 makes SDCLN unacceptably slow (NITER too
large for a given CLEANing depth) so a compromise is needed.
GAINs in the range 0.1 to 0.25 are most commonly used.
If the source has some very bright compact features
embedded in weaker diffuse emission, it is illuminating to
examine the Clean Map when only the brightest structure has been
CLEANed, to check whether subsequent CLEANing of weaker diffuse
emission causes it to "go lumpy" via the sidelobe feedback
effect. This could be done with GAIN = 0.3-0.5, using either
NITER or the CLBOX selection to ensure that the search does not
stray into the extended emission. Then SDCLN can be restarted
with lower GAIN, higher NITER and wider CLBOXes to tackle the
diffuse structure. If the weak emission "goes lumpy" you may
need to rerun SDCLN with different combinations of GAIN and
NITER to find the most effective one for your particular case.
Ultimately you will stop SDCLN when the new CLEAN component
intensities approach the noise level on your map. On a map of
Stokes I, the appropriate stopping point will be indicated by
comparable numbers of positive and negative components appearing
in the CC list. On maps of Stokes Q and U, which can and will
be legitimately negative, you need to know the expected
sensitivity level of your observation to judge how deep to go.
It is NEVER worth increasing NITER and restarting SDCLN
once many negative CLEAN components appear in the CC list of an
I map. When this occurs, you are using SDCLN to shuffle the
noise in the residual map, which is about as sensible as
rearranging the deck chairs on the Titanic after it hit the
iceberg.
FLUX
This provides an alternative to NITER for terminating the
iterations in a given run of SDCLN. In practice, most users
prefer to control SDCLN by limiting each execution with NITER.
FLUX should then be set to your expected rms noise level times
the dynamic range of the dirty beam (peak/worst sidelobe), to
ensure that you do not inadvertently waste time iterating to a
high value of NITER while in fact CLEANing emission whose
sidelobes are lost in the noise. The FLUX criterion is applied
only to the residuals inside the CLEAN window.
PHAT
When dealing with two-dimensional extended structures,
CLEAN can produce artifacts in the form of low-level high
frequency stripes running through the brighter structure. These
stripes derive from poor interpolations into unsampled or poorly
sampled regions of the (u,v) plane. [When dealing with quasi
one-dimensional sources (jets), the artifacts resemble knots
(which may not be so readily recognized as spurious)].
SDCLN can invoke a modification of CLEAN that is intended to
bias it towards generating smoother solutions to the
deconvolution problem while preserving the requirement that the
transform of the CLEAN components list fits the data. The
mechanism for introducing this bias is the addition to the dirty
beam of a delta-function (or "spike") of small amplitude
(PHAT) while searching for the CLEAN components. [The beam
used for the deconvolution thereby resembles the helmet worn by
German military officers in World War I, hence the name
"Prussian Helmet Clean"]. The theory underlying the algorithm
is given by Cornwell (1982, 1983), where it is described as the
Smoothness Stabilized CLEAN (SSC).
PHAT has given good results in some cases (e.g. Sag A at
20cm - see Cornwell 1983), reducing the amplitude of fringelike
artifacts and producing more believable images.
If you think SDCLN has left CLEAN artifacts in what should
be a smooth source, you have nothing to lose, and possibly
something to gain, by using PHAT for comparison purposes. Tim
Cornwell recommends that lower loop gains than normal, and
deeper CLEANing than normal, be used when using PHAT. Use
GAIN = 0.05 where you would have used 0.1 in the regular SDCLN,
for example. He also notes that his original recommendation for
PHAT = **2/2***2
may not be optimal, and that values in the range 0.1 to 0.50
generally give the best results. Be prepared to experiment with
different settings and to judge the results (severity of
stripes, overall dynamic range) for yourself. PHAT will not
eliminate stripes demanded by (bad ?) data.
The frequency of, and reasons for, PHAT's successes and
failures are not yet well understood. Your description of any
results (good or bad) attending use of PHAT will therefore be
appreciated by the AIPS group and by Tim Cornwell.
BITER
The default (BITER=0) restarts the CLEAN from scratch.
Other values of BITER are used when SDCLN is to be restarted
from an intermediate step in the process. When set >0, it
specifies the number of components from a previous clean to be
subtracted from the dirty map to obtain the residual map for the
first new iteration.
Restarts are sometimes needed after the computer has
crashed during a long CLEAN. Under these circumstances, the
iteration number at the end of the last major cycle is stored in
the AIPS file header. Provided that the crash has not destroyed
relevant image files (or the CC extension file) on disk, the
CLEAN may be restarted by setting BITER equal to the number of
iterations shown in the image header for the CLEAN map - if this
disagrees with the number in the internal file header (as may
happen if the crash comes at an unfortunate point in a cycle),
AIPS will adjust BITER downwards in a "fail-safe" way. (PRTCC
might be run to check the state of the components list in such
cases).
When you set BITER>0, you must set the OUTNAME and OUTSEQ
parameters explicitly to those of the Clean Map whose CC
extension file contains the components which are to be
subtracted. This Clean Map file will be overwritten by the new
Clean Map when SDCLN finishes, so if you wish to preserve it you
should either write it to tape using FITTP or create a second
copy of it on disk using SUBIM.
A components file F(N) can be reconvolved with a different
Clean Beam H by restarting SDCLN with NITER=BITER=n. This is an
effective tool for making SMALL changes the resolution of a
Clean Map. Do NOT use it for making large (factors >2) changes
in resolution, e.g. to "find" extended components of a source.
If a structure has been resolved out over the longer baselines,
these baselines contribute only noise, not signal, to maps of
that structure, and should be excluded from any search for the
structure. CLEANing a noisy map and then convolving it to
much lower resolution makes no sense. In such cases, you should
remap with an appropriate taper and run SDCLN on a dirty map
with a more appropriate resolution.
BITER may be set <0, in which case SDCLN will not subtract
the specified components before starting, but WILL restore them.
This strategem is useful only when the input map is a RESIDUAL
map, rather than a Clean map (see BMAJ).
BMAJ, BMIN, BPA
The default values of 0 for these parameters invoke an
algorithm whereby the central portion of the dirty beam B is
fitted with an elliptical Gaussian function whose parameters are
then used to specify the Clean Beam H. The algorithm can be
"fooled" by positive or negative sidelobes near the main
lobe of B, and has been known to prescribe unsatisfactory forms
for H, particularly for snapshot maps.
It is normally preferable to specify BMAJ, BMIN and BPA
explicitly from your own examination of the dirty beam, or after
a trial fit using the default. The Clean Map C may be easier to
interpret if BMIN is set equal to BMAJ, so that H is a circular
Gaussian, and any elongated structures are therefore seen in
their correct orientation. The frailties of CLEAN's
deconvolution will be least apparent if both are set equal to
the LONGEST dimension of the dirty beam.
Attempts to "super resolve" the source by setting BMAJ and
BMIN to the SHORTEST dimension of the dirty beam (or shorter)
skate on the proverbial thin ice, the more so if the number of
clean components in F(N) is comparable to, or larger than, the
number of independent visibility data used to make the dirty
map.
Note that if BMAJ, BMIN and BPA differ greatly from those
of the main lobe of the dirty beam, the parts of the Clean Map
derived from F(N) and from R(N) at step (3) will have greatly
different resolutions. This is very dangerous if R(N) contains
significant residual emission.
If BMAJ is set <0, then the output map contains the
residual map R(N) instead of the clean map C. This option
allows you to display, take hard copy of, or back up, the
residual map while deciding further strategy, retaining the
ability to regenerate the Clean Map later by using BITER <0.
NBOXES, CLBOX (TVBOX)
These parameters specify the size and location of the
rectangular boxes comprising the "CLEAN Window" area A. You
make the best use of prior knowledge about the source to help
SDCLN do its job when you restrict A as closely as possible to
the true boundaries of its emission. Recall that CLEAN attempts
to optimize F(n) as a representation of the sky, with zeroes
everywhere else. The more information you provide about where
the "real" emission is, the more likely SDCLN is to converge on
a unique, credible solution.
The verb TVBOX may conveniently be used to set the BLC and
TRC limits of the boxes after NBOXES has been typed in.
Following a prompt on your terminal, position the TV cursor at
the BLC of the first CLBOX and press any trackball button. Then
position the cursor at the TRC of this box and press button B.
Repeat this for all desired boxes. This will fill the CLBOX array
for the SDCLN inputs. The terminal prompt will also give
instructions for resetting any previously set corners should
you need to do so.
A practical detail: when NBOXES=1 and CLBOX specifies an
area <= 127 by 127, the entire residual map can be loaded into
an AP with 64k, and SDCLN can proceed very efficiently. This
speeds up execution enormously. If NBOXES>1, even if the area
in the CLBOXes adds up to less than 127 by 127, this economy is
lost.
The default [NBOXES=1, CLBOX=0] specifies that the inner
quarter of the map is to be searched for CLEAN components. A
particularly large economy in run time is achieved when this
default is used with 256 by 256 (or smaller) maps. In the 256
by 256 case not only will the Clean Window default to 127 by 127
(it's crafty enough to prune down the nominal 128 by 128) but
the necessary FFTs can be done entirely within the AP; under
these circumstances SDCLN proceeds at a headlong gallop (about
4 minutes for 1000 iterations).
No more than the inner quarter of the map can be CLEANed
accurately (due to the limited area over which the dirty beam
is known). SDCLN will however try to CLEAN a bigger area if
you ask it to, and will succeed to a level of approximation
which gets worse as you approach the edge of the map. Use
this naivety of SDCLN sparingly and with caution - it will
never do well in the outer 10 percent of the dirty map.
FACTOR
This knob protrudes from the inner workings of the Clark
algorithm, enabling the user to vary the criterion [Clark 1980]
by which minor iteration cycles are ended.
[For those with an interest in the gory details - SDCLN
first notes the ratio between the brightest point which was
ignored when the residual map R'(m) was passed to the AP and the
maximum residual R'(n') at some later iteration n'; it then
uses the Clark criterion with this ratio raised to the power
FACTOR replacing unity in Clark's summation].
FACTOR = 0 (the default) recovers the Clark criterion.
FACTOR > 0 allows the minor cycles to go on longer,
speeding up the CLEAN slightly (about 20 percent for FACTOR = 1), but
allowing you to get closer to the situation where residuals R'
in the AP become smaller than values which were ignored when the
AP was loaded. The search for new components becomes less and
less accurate (compared with a Hogbom algorithm using all of R
in step (2)), and the representation of extended structure in
the final F(N) deteriorates.
FACTOR < 0 makes the termination criterion more
conservative and thus improves the accuracy of the CLEANing at
the expense of forcing more major cycles with a corresponding
overhead in the FFTs done at the end of each one.
It is recommended that experiments with FACTOR normally be
confined to the range -0.3 < FACTOR < +0.3, negative values
being used when CLEANing complex extended structures, and
positive values when CLEANing very simple compact structures.
MINPATCH
This parameter specifies the minimum half-width of the beam
patch that will be allowed in a minor cycle. A smaller beam
patch allows the CLEAN to go faster at the expense of less
accurate subtraction at each iteration. The inaccuracy can lead
to errors which will not be recovered completely at the end of
the major cycle, especially if high GAINs are used or the
source has complex structure. MINPATCH=51 is recommended for
CLEANing complicated sources, but larger sizes up to 1024 should
be used for complicated (poor) dirty beams (i.e. VLA snapshots,
VLBI).
DOTV
Use of DOTV > 0 is STRONGLY recommended when CLEANing a
source for the first time. It causes the residual map to be
displayed on the TV after each major cycle, allowing you to
monitor what emerges as the CLEAN progresses. DOTV > 0 produces
a 15-sec pause after each major cycle to give you time to
inspect the display and to assess whether it is reasonable to
proceed. Pressing trackball D during this pause terminates
SDCLN in an orderly fashion. Pressing buttons A,B,C (or allowing
the pause time to elapse) starts the next major cycle.
When CLEANing a very complicated source for the first time,
it is often worth going beyond this interactive approach, by
taking hard copy of various stages of the CLEAN to compare
carefully later. Consider setting NITER to a value in the range
50-200 at first (the default is 200). Then take hard copy of
the lightly CLEANed map for later reference, and restart SDCLN
with a higher value of NITER. Doing this increasing NITER by
factors of order 2 each time can be very instructive in showing
you what SDCLN is doing to the extended structures in your
source. Spectral line users may wish to CLEAN a typical channel
in this way before deciding how best to parameterize SDCLN for
their production runs. (Note that the trial channel should be
reanalysed using the final choice of parameters, for
consistency; SDCLN's final Clean maps depend in detail on the
relative numbers of major and minor cycles which have been
performed).
Prudence
If running SDCLN on a complicated source with low GAIN you
may need to work to large final values of NITER. As SDCLN is
the major consumer of CPU time in most map analyses, it is
prudent to preserve intermediate clean maps by writing them to a
FITS tape with FITTP. This allows you to recover from
disasters such as crashes, overCLEANing, etc. with minimal
impact on total execution time, your time, and on disk space.
Considerations for CLEANing Cubes
Spectral line cubes and other multidimensional images may
now be CLEANed one plane per run of SDCLN. The adverb BLC is
used to specify the desired plane. The output file will be the
same size as the input file and only the specified plane in the
output file will be modified. Thus, other planes will contain
garbage until they are processed. A separate components file
will be created for each plane (up to 46655) with sequence
number equal to the (1-rel) plane number. If the beam image is
also a cube then the corresponding plane for the beam will be
used otherwise the first (only) plane will be used.
NOTE: the planes should be processed in increasing plane
number in order to insure that the corresponding CLEAN
components file has the correct version number. If this gets
messed up, a restart using previous CLEAN components will pick
up the wrong components list.
Considerations for assymmetric beams
SDCLN partially supports the use of assymetric beams. The
routine that finds the components (minor cycles) uses only half
of the beam and assumes that it is symmetric. At the end of
major cycles the components are subtracted with the proper
beam. Thus, it is desirable that a CLEAN with an assymmetric
beam take several major cycles. For large images this is no
problem but for small images the use of a several CLEAN boxes
may be useful even if one is small and contains nothing of
interest. Also the adverb FACTOR can be used to reduce the
number of components per major cycle; a value of -1 may be a
good starting guess. More major cycles should be used for more
strongly assymmetric beams. Note, beam means the same thing as
point spread function.
Timing
======
As described above, SDCLN runs enormously faster when the
Clean Window is <= 127 by 127 and the map is <=256 by 256 than
when only fractions of the residual map can be stored in the AP
and large FFTs are required. CPU times for these small CLEANs
will be only a few minutes for several thousand iterations.
When bigger maps and Clean Windows are required, SDCLN's
run times become harder to predict. The steps involved in the
major cycles are the most consumptive, so the number of major
cycles required to do the CLEAN dominates the CPU time. The
number of major cycles required depends on the complexity of the
source, but CLEANs involving 10,000 components or so will often
chew through five or six of them. As SDCLN is the main CPU hog
in AIPS, you should not run SDCLN on large maps (>512) with more
than about 5000 iterations without estimating run times for your
case from more modest initial CLEANs. Also look at the
components file to check the need for such deep CLEANing before
turning SDCLN loose on this scale.
On a VAX 11-780 with an FPS 120B array processor the run
time in an otherwise empty machine is
3.0 * (M / 1024) **1.5 min to get maps
2.0 * (M / 1024) **2.0 min to find 1000 comp.
3.2 * (M / 1024) **1.5 min each major cycle
2.0 * (M / 1024) **2.0 min store maps
where M is the map size. For example cleaning the inner quarter
of a 1024 x 1024 map for 1000 iterations (assuming 4 major
cycles) is about 20 min run time. If the machine is busy the
run time may be substantially longer.
REFERENCES
Clark, B.G. (1980). "An Efficient Implementation of the
Algorithm "CLEAN", Astron.Ap., 89, 377-378
Clark, B.G. (1982). "Large Field Mapping", Lecture #10 in the
NRAO-VLA Workshop on "Synthesis Mapping"
Cornwell, T.J. (1982). "Image Restoration (and the CLEAN
Technique)", Lecture #9 in the NRAO-VLA Workshop on
"Synthesis Mapping"
Cornwell,T.J. (1982). "Can CLEAN be Improved ?", VLA Scientific
Memorandum No. 141.
Cornwell,T.J. (1983). "A Method of Stabilising the CLEAN
Algorithm", preprint.
Hjellming,R.M. (1982). "An Introduction to the NRAO Very Large
Array", Section 4.
Hogbom,J.A. (1974). "Aperture Synthesis with a Non-Regular
Distribution of Interferometer Baselines",Astron.Ap.Suppl.,
15, 417-426
Schwarz, U.J. (1978). "Mathematical-statistical Description of
the Iterative Beam Removing Technique", Astron.Ap.,65, 345.
Schwarz, U.J. (1979). "The Method "CLEAN" - Use, Misuse and
Variations", in Proc. IAU Colloq. No.49, "Image Formation
from Coherence Functions in Astronomy", ed. C. van
Schooneveld, (Dordrecht:Reidel), p. 261-275.
Steer, D.G., Dewdney, P.E., Ito, M.R. (1984). "Enhancements to
the Deconvolution Algorithm 'CLEAN'", Astron.Ap.,137, 159.
AIPS HELP file for SDCLN in 31DEC24