Notes on coregistration of MOSES m = -1, 0, +1

2006 May 24 - 31

2006-May-24 CCK

First run of MOSES_ALIGN_ORDERS was yesterday (log060523.163229). The routine was intentionally limited to registration of just the m=-1 order, for just the first two exposures. Summary of results:

CHISQ_BADNESS: chisq =        1744892.8 chisq_reduced =       0.85008617
imx =       24.1310      25.6349      2076.54      2075.01
imy =      -3.62048      1019.20     -2.69529      1022.31
% COREGISTRATE: Alignment completed at Tue May 23 16:49:56 2006
ncalls =          555
....

CHISQ_BADNESS: chisq =        4656747.4 chisq_reduced =        2.2666494
imx =       26.4620      24.8723      2075.15      2075.08
imy =      -3.50844      1018.37     -1.30873      1024.03
% COREGISTRATE: Alignment completed at Tue May 23 17:09:52 2006
ncalls =          701
moses_align_orders took        39.142551 minutes to run

Comments:

The save file ended up in idlsave.dat because I forgot how to use SAVE. Amazingly, my complicated ALIGNMENT data structure seems to have worked.
Note that the above are not quite the final numbers, just the last ones reported by calls to CHISQ_BADNESS. Inspection of alignment[i].minus.cpoints.x[j] confirms that these are close to the final values, though. For the record, here are the final positions from the save file:

IDL> print, alignment[0].minus.cpoints.x
      24.1406      25.6308      2076.53      2074.99
IDL> print, alignment[0].minus.cpoints.y
     -3.60734      1019.23     -2.75141      1022.29
	  
IDL> print, alignment[1].minus.cpoints.x
      26.4619      24.8725      2075.15      2075.08
IDL> print, alignment[1].minus.cpoints.y
     -3.50831      1018.37     -1.30909      1024.03

I would expect the alignment to be the same for each exposure, and the hoped-for consistency is 0.1 pixel. This is what I in fact get:

IDL> print, alignment[0].minus.cpoints.x - alignment[1].minus.cpoints.x
     -2.32132     0.758270      1.38257   -0.0922852
IDL> print, alignment[0].minus.cpoints.y - alignment[1].minus.cpoints.y
   -0.0990317     0.858826     -1.44232     -1.74170

IDL> print, sqrt( (alignment[0].minus.cpoints.x - alignment[1].minus.cpoints.x)^2 + (alignment[0].minus.cpoints.y - alignment[1].minus.cpoints.y)^2 )
      2.32343      1.14567      1.99794      1.74414

The alignment values for the two exposures are only broadly consistent, with errors of 1.1-2.3 pixels. Admittedly, this is from analyzing the shortest exposures, so perhaps things will improve with deeper images.
Lots of function calls reported by amoeba. I should probably reuse imx and imy with subsequent calls to speed things up.
Value of reduced chi squared differs between these two alignment cases. I wonder why.
Looks like the complete run would take


		0.5 * 39min * 2 orders * 27 exposures = 17.6 hrs.

I've examined residuals from the save file, and conclude that the alignment is nowhere near done. I'm not sure what is wrong, but clearly this can be improved. Here is how I made the residuals, and then saved them as jpegs to look at later:

IDL> norm = median(cube_minus[*,*,1])/median(cube_plus[*,*,1])
IDL> print,norm
      2.21052
IDL> foo1=cube_minus[*,*,1]-cube_plus[*,*,1]*norm                  
IDL> foo2=cube_minus[*,*,0]-cube_plus[*,*,0]*norm
IDL> write_jpeg, '20060524_residuals_exp0.jpg',bytscl(foo1)     
% Loaded DLM: JPEG.
IDL> write_jpeg, '20060524_residuals_exp1.jpg',bytscl(foo2)
IDL> print, median(abs(foo1))
      55.2104
IDL> print, median(abs(foo2))
      14.5912

There is a silly mistake here, but it fortuitously cancels another silly mistake. See the notes from 2006-May-25, below.

Above are exposure 0 residuals.

Above are exposure 1 residuals.

MODIFICATIONS FOR NEXT RUN:

Do all 27 m = -1 images
Recycle imx, imy values to speed things up
tightened the convergence criterion for amoeba from 1e-4 to 1e-5, and now allow 2000 instead of 1000 calls.

Extracts from log060524.152418:

% Compiled module: $MAIN$.
% $MAIN$: Loading Level One MOSES data...
Performing coregistration for exposure        0
% Compiled module: COREGISTRATE.
% COREGISTRATE: Starting alignment at Wed May 24 15:26:55 2006
....
CHISQ_BADNESS: chisq =        1609928.1 chisq_reduced =       0.78340056
imx =       26.8515      25.6921      2074.50      2074.24
imy =      -4.44583      1019.93    -0.765763      1022.88
% COREGISTRATE: Alignment completed at Wed May 24 15:35:21 2006
ncalls =          291
Performing coregistration for exposure        1
....
CHISQ_BADNESS: chisq =        4598330.9 chisq_reduced =        2.2376981
imx =       26.8659      25.5338      2074.61      2074.26
imy =      -4.41270      1019.83    -0.887940      1022.93
% COREGISTRATE: Alignment completed at Wed May 24 15:42:41 2006
ncalls =          233
Performing coregistration for exposure        2
....
CHISQ_BADNESS: chisq =        9845535.5 chisq_reduced =        4.7908773
imx =       26.8523      25.4894      2074.61      2074.27
imy =      -4.46986      1019.79    -0.737358      1023.08
% COREGISTRATE: Alignment completed at Wed May 24 15:50:12 2006
ncalls =          239
Performing coregistration for exposure        3
% COREGISTRATE: Starting alignment at Wed May 24 15:50:13 2006
CHISQ_BADNESS: chisq =    1.4621517e+10 chisq_reduced =        7118.7966
imx =       26.8527      25.4890      2074.61      2074.27
imy =      -4.46948      1019.79    -0.737227      1023.08
....
CHISQ_BADNESS: chisq =    6.9963484e+09 chisq_reduced =        3406.7490
imx =       27.4948      25.2048      2077.28      2074.46
imy =      -4.16954      1020.40    0.0537870      1023.45
% COREGISTRATE: Alignment completed at Wed May 24 16:01:13 2006
ncalls =          354
Performing coregistration for exposure        4
% COREGISTRATE: Starting alignment at Wed May 24 16:01:14 2006
CHISQ_BADNESS: chisq =    2.1569528e+11 chisq_reduced =        105881.32
imx =       27.4948      25.2048      2077.28      2074.46
imy =      -4.16954      1020.40    0.0537912      1023.45
....
CHISQ_BADNESS: chisq =    1.9745229e+11 chisq_reduced =        96981.066
imx =       27.2258      25.5781      2077.50      2074.41
imy =      -5.01160      1023.81     0.441277      1023.85
CHISQ_BADNESS: chisq =    1.9745229e+11 chisq_reduced =        96981.066
imx =       27.2258      25.5781      2077.50      2074.41
imy =      -5.01160      1023.81     0.441277      1023.85
% COREGISTRATE: Alignment completed at Wed May 24 17:02:23 2006
ncalls =         2010
% COREGISTRATE: Amoeba failed to converge.
% Execution halted at: COREGISTRATE      322

Comments:

This time, the convergence is much finer, and the values from the last badness call are essentially the final alignments (to < .001 pixel).
Alignments are much more consistent this time, at least for exposures 0,1,2. Then things get really strange with Exposure 3 & 4.
What's with the huge chi squared? I think I understand the growth in chi squared for exposures 0,1,2: this is just because, with longer exposures, the residuals become larger compared to the noise (the image are not, after all, identical!). But the huge jump with exposurese 3 & 4 seems unnatural. And note how big an 'improvement' is made in chi squared by wandering off in the wrong direction during exposure 3.
Why didn't amoeba declare itself converged? Seems like the changes toward the end were at the hairy edge of floating point resolution. I was only asking for 1e-5 convergence. But I'm not precisely sure how the convergence criterion works...

Plan:

I've modified moses_align_orders.pro to save a snapshot (mosesAlignOrders_snap.sav) after every amoeba return. Then I can restore and run various things (like COREGISTRATE) by hand, with X11 graphical diagnostics.

2006-May-25

Created automatic documentation system, docs, and reformatted the present notebook in HTML.

Looking at data from the last run made me realize several important things:

My calculated 'residuals', above, took minus - plus instead of minus - zero. Oops.
The above mistake was oddly fortuitous, because inspecting the data makes it quite obvious that cube_plus contains the m=0 images, which are easily identifiable by virtue of having about half the counts of the outboard orders. THIS IS THE FIRST THING TO FIX. It is clearly a problem with moses_prep. Is the problem perhaps inherited from movie0, and shared among the descendants of movie0?
The offsets show, in any case, what looks like a 25-pixel horizontal offset between the orders that I did align. Did I end up aligning Si in one image to He in another? Or if I aligned He to He correctly, could it be that these residuals are themselves the imprint of Si? Probably wishful thinking.

Sure enough, the error was in MOSES_PREP:

		;Store in cube.
		cube_minus[*,*,j] = minus
		cube_zero[*,*,j]  = plus
		cube_plus[*,*,j]  = zero

The problem is in MOSES_PREP only: the movie programs were not affected. I have now fixed this.

Running prep again, moses_prep.log060525.171016.

Running MOSES_ALIGN_ORDERS again, moses_align_orders.log060525.172141.

2006-May-30

Looking at residuals from the last run. The amoeba failed to converge on exposure 12. So exposures 1-11 should have the best alignments that MOSES_ALIGN_ORDERS can get, while 13-26 should be unmodified (as saved by MOSES_PREP).

I've written a simple procedure, EXAMINE_RESIDUALS, to step through the residuals and examine them visually. It turns out that none of the contents of cube_minus were modified by MOSES_ALIGN_ORDERS:

IDL> restore,'mosesAlignOrders_snap.sav
IDL> cube_minus_aligned = cube_minus
IDL> cube_zero_aligned = cube_zero
IDL> restore,'mosesLevelOne.sav
IDL> pmm, cube_minus[*,*,7] - cube_minus_aligned[*,*,7]
% Compiled module: PMM.
      0.00000      0.00000
IDL> pmm, cube_minus[*,*,*] - cube_minus_aligned[*,*,*]
      0.00000      0.00000

How strange! And by the way, cube_plus is also unmodified. I am running MOSES_ALIGN_ORDERS from the IDL command line to get some visual feedback. I have commented out set_plot, 'z' so that I'll be able to watch the residuals evolve.

OK, I've watched the residuals evolve through convergence of MOSES_ALIGN_ORDERS on exposure 0, using COREGISTRATE with CHISQ_BADNESS. The residuals look just the way I would expect, but somehow the modified m = -1 image does not get stuffed back into cube_minus.

Here is the call to COREGISTRATE from MOSES_ALIGN_ORDERS:

	coregistrate, cube_zero[0:Nx-1,0:Ny-1,i], cube_minus[0:Nx-1, 0:Ny-1, i], $
		1, refx, refy, imx, imy, missing = missing_data,                      $
		control_pts = control_pts,                                            $
		image2 = cube_minus[0:Nx-1, 0:Ny-1, i],                               $
		Kx = alignment[i].minus.Kx,                                           $
		Ky = alignment[i].minus.Ky,                                           $
		/verbose

The image2 keyword should result in modification of cube_minus, shouldn't it? In a word, no. Have a look at this test program.

I have now fixed the bug in MOSES_ALIGN_ORDERS. Also restored the set_plot,'z' to allow it to run without an X display. Running again, moses_align_orders.log060530.134821

I have examined the residuals for the first bit of this run. Now they look right:

IDL> restore,'mosesAlignOrders_snap.sav'
IDL> examine_residuals, cube_minus, cube_zero

Above is the residual between orders m = -1 and m = 0 for exposure 0.

Some rationalization:

The idea behind this coregistration procedure is that the average doppler shift (over the whole image, or any large region) should be close to zero. This assumption may or may not be correct; however, it is common practice with many spectrometers to establish the 'line center' position based on an average over a quiet solar image. What we are doing is analogous. So I expect that if the global correlation is maximized as a function of alignment, then any remaining difference between any two orders will be attributable to:

Errors. At this stage, since flat fields are not yet applied, this includes the difference between the detector flat fields as well as photon and readout noise.
Difference in PSF between the orders. As I have said before, this is not insignificant. I am convinced that this effect can be eliminated, but it will require some clever analysis on all our parts.
Line profile information, with the leading term in the residuals presumably coming from the lowest relevant moment of the line profile, which is the doppler shift.

Looking at the residuals between m = -1 and m = 0 from my runs today (using examine_residuals.pro), the PSF differences are quite prominent. Above is a 2x2 binned image of residuals from the first 0.25s exposure, which is convenient to look at because there are no saturated pixels. I can tell that we are seeing mainly differences in PSF because the residuals look symmetric -- that is, they resemble a convolution of one of the images with a kernel that is approximately symmetric about its origin. Doppler shifts should produce an asymmetric signature. Earlier, I produced similar PSF-related artifacts in my coregistration test suite.

The following questions are still not resolved:

Why is chi squared so large for many of the exposures? I can rationalize chi squared going as the square of the exposure time, but not more than that.
Why does AMOEBA sometimes fail to converge?

Below is a detailed summary of the latest run, moses_align_orders.log060530.134821. It is useful to remember that the final alignment values for a particular exposure show up in the initial guess for the subsequent exposure.

Performing coregistration for exposure        0
% Compiled module: COREGISTRATE.
% COREGISTRATE: Starting alignment at Tue May 30 13:53:29 2006
% Compiled module: AMOEBA.
% Compiled module: POLYWARP.
CHISQ_BADNESS: chisq =        8438225.3 chisq_reduced =        4.1061324
imx =       26.4000      25.0000      2075.00      2075.00
imy =      -3.50000      1018.00     -1.30000      1024.00
....
CHISQ_BADNESS: chisq =        2156059.8 chisq_reduced =        1.0489861
imx =       21.1022      21.3940      2068.25      2068.36
imy =      -5.81271      1017.92     -5.23799      1018.68
CHISQ_BADNESS: chisq =        2156056.5 chisq_reduced =        1.0489845
imx =       21.1017      21.3935      2068.25      2068.36
imy =      -5.81227      1017.92     -5.23795      1018.68
% COREGISTRATE: Alignment completed at Tue May 30 14:05:13 2006
ncalls =          370
Performing coregistration for exposure        1
% COREGISTRATE: Starting alignment at Tue May 30 14:07:42 2006
CHISQ_BADNESS: chisq =        5435152.3 chisq_reduced =        2.6443604
imx =       21.1016      21.3944      2068.25      2068.36
imy =      -5.81253      1017.92     -5.23825      1018.68
....
CHISQ_BADNESS: chisq =        5428321.6 chisq_reduced =        2.6410371
imx =       20.9810      21.5303      2068.30      2068.44
imy =      -5.99142      1017.78     -5.12446      1018.74
% COREGISTRATE: Alignment completed at Tue May 30 14:15:18 2006
ncalls =          235
Performing coregistration for exposure        2
% COREGISTRATE: Starting alignment at Tue May 30 14:17:49 2006
CHISQ_BADNESS: chisq =        11390957. chisq_reduced =        5.5420333
imx =       20.9812      21.5304      2068.30      2068.44
imy =      -5.99094      1017.78     -5.12439      1018.74
....
CHISQ_BADNESS: chisq =        11388261. chisq_reduced =        5.5408296
imx =       20.9575      21.5363      2068.34      2068.42
imy =      -6.01806      1017.74     -5.02928      1018.82
CHISQ_BADNESS: chisq =        11388219. chisq_reduced =        5.5408093
imx =       20.9574      21.5364      2068.34      2068.42
imy =      -6.01818      1017.74     -5.02916      1018.82
% COREGISTRATE: Alignment completed at Tue May 30 14:26:28 2006
ncalls =          267
Performing coregistration for exposure        3
% COREGISTRATE: Starting alignment at Tue May 30 14:29:05 2006
CHISQ_BADNESS: chisq =    3.5015171e+10 chisq_reduced =        17045.018
imx =       20.9573      21.5366      2068.34      2068.42
imy =      -6.01847      1017.74     -5.02932      1018.82
....
CHISQ_BADNESS: chisq =    1.8526900e+10 chisq_reduced =        9035.9078
imx =       22.5858      23.9492      2069.63      2070.46
imy =      -7.74742      1019.25     -5.62356      1017.42
CHISQ_BADNESS: chisq =    1.8526350e+10 chisq_reduced =        9035.6440
imx =       22.5859      23.9492      2069.63      2070.46
imy =      -7.74775      1019.25     -5.62391      1017.42
% COREGISTRATE: Alignment completed at Tue May 30 14:36:53 2006
ncalls =          235
Performing coregistration for exposure        4
% COREGISTRATE: Starting alignment at Tue May 30 14:39:28 2006
CHISQ_BADNESS: chisq =    4.2707100e+11 chisq_reduced =        209790.01
imx =       22.5859      23.9492      2069.63      2070.46
imy =      -7.74758      1019.25     -5.62372      1017.42
....
CHISQ_BADNESS: chisq =    3.5517067e+11 chisq_reduced =        174465.54
imx =       24.5128      24.0929      2069.38      2070.52
imy =      -7.79466      1019.34     -5.48531      1019.22
CHISQ_BADNESS: chisq =    3.5517090e+11 chisq_reduced =        174465.65
imx =       24.5128      24.0929      2069.38      2070.52
imy =      -7.79466      1019.34     -5.48531      1019.22
% COREGISTRATE: Alignment completed at Tue May 30 14:50:35 2006
ncalls =          349
Performing coregistration for exposure        5
% COREGISTRATE: Starting alignment at Tue May 30 14:53:19 2006
CHISQ_BADNESS: chisq =    2.6854113e+12 chisq_reduced =        1472806.8
imx =       24.5128      24.0929      2069.38      2070.52
imy =      -7.79466      1019.34     -5.48531      1019.22
....
CHISQ_BADNESS: chisq =    2.2887503e+12 chisq_reduced =        1256278.8
imx =       25.2460      23.9374      2072.59      2071.33
imy =      -6.94355      1018.75     -5.91531      1019.10
CHISQ_BADNESS: chisq =    2.2887498e+12 chisq_reduced =        1256278.5
imx =       25.2460      23.9374      2072.59      2071.33
imy =      -6.94356      1018.75     -5.91531      1019.10
% COREGISTRATE: Alignment completed at Tue May 30 15:03:51 2006
ncalls =          336
Performing coregistration for exposure        6
% COREGISTRATE: Starting alignment at Tue May 30 15:06:34 2006
CHISQ_BADNESS: chisq =    5.3326029e+09 chisq_reduced =        2601.7253
imx =       25.2459      23.9374      2072.59      2071.33
imy =      -6.94356      1018.75     -5.91531      1019.10
....
CHISQ_BADNESS: chisq =    2.0433226e+09 chisq_reduced =        996.29124
imx =       26.1697      26.2833      2072.23      2071.96
imy =      -5.80089      1018.85     -5.13735      1017.84
CHISQ_BADNESS: chisq =    2.0429819e+09 chisq_reduced =        996.12514
imx =       26.1699      26.2829      2072.23      2071.96
imy =      -5.80082      1018.85     -5.13719      1017.84
% COREGISTRATE: Alignment completed at Tue May 30 15:17:42 2006
ncalls =          347
Performing coregistration for exposure        7
% COREGISTRATE: Starting alignment at Tue May 30 15:20:11 2006
CHISQ_BADNESS: chisq =    3.2378295e+09 chisq_reduced =        1578.6982
imx =       26.1699      26.2830      2072.23      2071.96
imy =      -5.80085      1018.85     -5.13725      1017.84
....
CHISQ_BADNESS: chisq =    2.5994588e+09 chisq_reduced =        1266.0055
imx =       26.1625      26.2760      2072.30      2071.96
imy =      -3.83107      1018.80     -4.91849      1018.04
CHISQ_BADNESS: chisq =    2.5992067e+09 chisq_reduced =        1265.8828
imx =       26.1625      26.2760      2072.30      2071.96
imy =      -3.83107      1018.80     -4.91849      1018.05
% COREGISTRATE: Alignment completed at Tue May 30 15:30:20 2006
ncalls =          317
Performing coregistration for exposure        8
% COREGISTRATE: Starting alignment at Tue May 30 15:33:21 2006
CHISQ_BADNESS: chisq =    2.9760573e+09 chisq_reduced =        1449.4183
imx =       26.1625      26.2760      2072.30      2071.96
imy =      -3.83107      1018.80     -4.91849      1018.05
....
CHISQ_BADNESS: chisq =    2.9563196e+09 chisq_reduced =        1439.8659
imx =       26.1202      26.3516      2072.34      2071.87
imy =      -3.96789      1018.85     -4.74246      1018.39
CHISQ_BADNESS: chisq =    2.9553041e+09 chisq_reduced =        1439.3706
imx =       26.1202      26.3515      2072.34      2071.87
imy =      -3.96784      1018.85     -4.74266      1018.39
% COREGISTRATE: Alignment completed at Tue May 30 15:42:10 2006
ncalls =          273
Performing coregistration for exposure        9
% COREGISTRATE: Starting alignment at Tue May 30 15:44:44 2006
CHISQ_BADNESS: chisq =    3.0861590e+09 chisq_reduced =        1503.0905
imx =       26.1202      26.3516      2072.34      2071.87
imy =      -3.96796      1018.85     -4.74265      1018.39
....
CHISQ_BADNESS: chisq =    2.1939274e+09 chisq_reduced =        1067.3553
imx =       25.9750      26.5327      2072.51      2072.03
imy =      -2.26547      1018.44     -4.28019      1018.61
CHISQ_BADNESS: chisq =    2.1938828e+09 chisq_reduced =        1067.3335
imx =       25.9750      26.5327      2072.51      2072.03
imy =      -2.26547      1018.44     -4.28020      1018.61
% COREGISTRATE: Alignment completed at Tue May 30 15:55:11 2006
ncalls =          319
Performing coregistration for exposure       10
% COREGISTRATE: Starting alignment at Tue May 30 15:57:48 2006
CHISQ_BADNESS: chisq =    2.4845427e+09 chisq_reduced =        1208.7415
imx =       25.9750      26.5327      2072.51      2072.03
imy =      -2.26547      1018.44     -4.28019      1018.61
....
CHISQ_BADNESS: chisq =    1.5098930e+09 chisq_reduced =        733.59415
imx =       26.3855      26.1889      2072.22      2071.90
imy =     0.0127509      1018.70     -4.01597      1018.86
CHISQ_BADNESS: chisq =    1.5098923e+09 chisq_reduced =        733.59378
imx =       26.3855      26.1888      2072.22      2071.90
imy =     0.0127462      1018.70     -4.01598      1018.86
% COREGISTRATE: Alignment completed at Tue May 30 16:08:10 2006
ncalls =          324
Performing coregistration for exposure       11
% COREGISTRATE: Starting alignment at Tue May 30 16:10:39 2006
CHISQ_BADNESS: chisq =    2.0484360e+09 chisq_reduced =        995.22948
imx =       26.3855      26.1888      2072.22      2071.90
imy =     0.0127432      1018.70     -4.01597      1018.86
....
CHISQ_BADNESS: chisq =    1.1040684e+09 chisq_reduced =        536.13773
imx =       26.3855      26.1888      2072.22      2071.90
imy =     0.0793048      1021.61     -3.01601      1018.86
CHISQ_BADNESS: chisq =    1.1040678e+09 chisq_reduced =        536.13742
imx =       26.3855      26.1888      2072.22      2071.90
imy =     0.0793013      1021.61     -3.01599      1018.86
% COREGISTRATE: Alignment completed at Tue May 30 16:17:06 2006
ncalls =          200
Performing coregistration for exposure       12
% COREGISTRATE: Starting alignment at Tue May 30 16:19:32 2006
CHISQ_BADNESS: chisq =    2.0316974e+09 chisq_reduced =        986.60952
imx =       26.3855      26.1888      2072.22      2071.90
imy =     0.0793936      1021.61     -3.01597      1018.86
....

The log file has become too big to grok by inspection. Am developing a procedure to visualize the consistency (or lack thereof) in the alignments: ALIGN_SUMMARY. Alignment summary image align_summary2006May31113325.jpg, produced for the latest run, is below:

The above plots show how the positioning of the four control points varied through today's last run of MOSES_ALIGN_ORDERS. Only exposures 0-11 are included, since the script crashed during exposure 12. Note that each plot is ISO scaled (x scale and y scale the same). Note: The original image has been replaced by a slightly nicer one made May 31.

2006-May-31

Last night's summary image shows two symptoms of disease:

Inconsistency (large sdev in x and y alignment) and
Systematic drift from exposure to exposure, rather than random error.

Ideas (in the ordeer thay occurred to me, not in order of priority):

Check that the reference (first argument, m = 0 image for ith exposure) is updated for each call to COREGISTRATE in the loop in MOSES_ALIGN_ORDERS, and that Creference in the common block is also updated with each call to COREGISTRATE. Update: Looked at the source code of both routines and found no problem.
Check that normalization of reference in CHISQ_BADNESS is done over the (same) domain of good pixels in image and reference. Update: Looked at the source code and found no problem.
Use EXAMINE_RESIDUALS to look carefully at all exposures.
Align again using cross-correlation (CORR_BADNESS) instead of chi squared (CHISQ_BADNES)
Go back to coregistration test suite, and test its ability to cope with known issues in our data.
1. Make sure I use the latest COREGISTRATE. Migrate test suite here?
2. Modify test image for a larger shift with distortion.
3. Have some saturated pixels, marked missing, in one or both images.
Check number of missing pixels with each call to badness.

While checking 1 and 2, I added a few more comments to the code but did not change anything functionally.

Wrote SAVESCREEN to make documentation more convenient. This is the screen image saving code from ALIGN_SUMMARY.

Using EXAMINE_RESIDUALS, I see plausible alignments for exposures 0,1,2. In exposures 3 and 6, the minus order is apparently shifted left relative to zero. The apparent bipolar features are consistent throughout the residuals image.

It is difficult to examine the longer exposures, because data that should be marked missing in the image2 keyword output of COREGISTRATE isn't. This is the old POLY_2D bug (see modification notes in COREGISTRATE). I have now carried through the fix that was used for the badness procedures to the image2 keyword output. Time to run the alignment again, so that the data cubes are modified accordingly.

Re-running the alignment: moses_align_orders.log060531.134357. Oops, had to fix a bug in my mods to COREGISTRATE. Re-running. moses_align_orders.log060531.140206.