Besides the change to the introduction of section 5, I've found a couple
  of questions:
  
  These techniques are used to refine the choice of values for bs and csf.
  (1) and (3) can be performed fairly rapidly on a low resolution version
  of an image (roughly one minute and five minutes, respectively, on an
  11/785-class machine).  Unfortunately low resolution "working versions"
  can have color artifacts compared to the high resolution originals:
>
> as a result of averaging,
>
  colors can be inside the destination gamut for one and outside it for
  the other.  [These artifacts are not a severe problem because correct
  pixel values can easily be obtained by probing the full resolution image
  once problem colors have been located in the low resolution version of
  the image]
  
  The sentence in brackets is unclear.  Is the point that the low
  resolution pictures identify regions of the picture where there may be
  out of gamut colors?  Furthermore, you also have to probe to test the
  false positives, right?  Were false positives or false negatives more of
  an issue?
>
> What I was trying to say here is that when there are concavities in
> the output gamut, filtering can produce an average value outside the
> gamut for the average of two colors (from the original image) that
> lie inside the gamut.  In such cases one can go back to the high res
> image to see if there are really out of gamut colours.  This was never
> a big deal, but I felt it should be mentioned.
[The original version of section 5 had a more complete description of this 
process. Basically two things can happen: 1. averaging can remove extreme 
values by averaging an extreme value with less extreme neighbours, and 2. 
averaging can produce spurious extreme values. The first is obvious, leading 
to out of gamut colours in the high res version that are not present in the 
low res trial version. The second is more subtle. It occurs when the gamut is 
concave, in which case it is possible to average two in-gamut colours to get 
an out-of-gamut one. The the trial version requires more extreme gamut mapping 
than would the high res version. This error is more likely to slip through 
than is the error of the first type. How frequently it occurs with real gamuts 
and real images is open to question as is how serious the colour distortions 
it produces. My objective in the re-write was to put the issue into as few 
words as possible, at the cost of a little obscurity, since there seems to be 
a real possibility of writing a lot of words, and giving too much importance 
to something just becasue we haven't followed through to prove that it is a 
minor issue.]

NEW:  These techniques are used to refine the choice of values for bs and csf.  (1) and (3) can be performed fairly rapidly on a 
low resolution version of an image (roughly one minute and five minutes, respectively, on an 11/785-class machine).  Note that the 
low resolution version of the image does not have precisely the same gamut at the high one due to the effect of averaging adjacent 
pixels.  However, these differences can be handled by probing the full resolution image to produced the correct color values once 
problem areas have been located.


  The transformation is actually too general.  To avoid hue shifts
  (principle (4)), it is necessary to require that Rn lie along the vector
  from Ws to Rs  with similar constraints for Gn and Bn. This constraint
  can also be too confining. Overall chromatic contrast can sometimes be
  increased if irregularities in the image and destination gamuts are
  matched to one another. So doing reflects a decision that the
  relationship between the colors in the image is more important than
  their absolute chromaticity.  As an example, an image such as Figure
  5.11 benefits from moving the red and green phosphors towards blue. [The
  resulting hue shifts, most noticeable when pure blue becomes greenish,
  are less objectionable than the lower chromatic contrast needed in the
  absence of hue shifts. (This contradicts the previous sentence!!)] 
  [The preceding square brackets are not a comment from the present round.] 
>
> I don't remember actually writing this, but it looks like the sort
> of damage I would have done.  See comments for the next question.
  
  Whose is the parenthetical note, and what should we do about it? 
  [I added the parenthetical (round brackets) note during my reworking of the 
section. John added the outer stuff. The parenthetical note acknowledges the 
fact that if red and green go toward blue greens become bluish, not the 
reverse.]
>
> We could change "from moving the red ... towards blue" to something like
> "from a modest compression and shear (see Figure <the umbrella chromaticity
> diagram>)."
[Compression and shear are new terms at this point. It would be appropriate to 
introduce them only if we were willing to go into a more complete discussion 
of the various changes a linear transformation can do to a solid body, 
relating them to colour appearance in an image.]
  
  [Are we certain that the cyan results from the umbrella transformation,
  and not from uncorrected reproduction?  Plotting the chromaticities, it
  looks to me as if the large hue changes were for red and green, not
  blue.]
  [Not my comment.]
  
  I'm not sure what is meant here - I can check to see how the hues are
  moving, but whoever wrote this (Bill?) please clarify.
>
> I'm the guilty party here.  I raised the question after having plotted
> the actual chromaticity shifts we used for the umbrella transformation.
> It looked to me like the R and G phosphors had significantly more hue
> shift to them than did the B phosphor, in that they diverged more from
> the vector towards the white point.  Hence it seemed odd to me that,
> if the hue shift observed resulted from this transformation, it should
> primarily appear as a shift in blue.
>
> I suppose that the definitive test is to look at the chromalins for one
> of the Munsell figures, with and without the umbrella transform and
> verify that we are reporting what we see there, as distinct (perhaps)
> from an observation we made of a monitor.
>
> If the observation is correct, then I am confused about why the given
> chromaticity mapping produced the observed hue shift.  I suppose I could
> believe that this is a result of non-linearity in the chromaticity diagram,
> but would like to have an expert speak to the issue!
[Here the problem is clearly a deficiency of data. What did we observe about 
subjective hue shifts in the reproduction compared to the original? What was 
the effect of the umbrella transformation on the gamut? John has given 
observations on the latter (above) but suggests they should be checked. I am 
convinced the problem is an obsence of observation, and that some one needs to 
write down what the facts are before we go too hard into explaining them.]
  
  If you answer these questions and the proposal in the previous note by
  tomorrow, I will try to get a new version of section 5 back for weekend
  reading.  (always the optimistyou guys work at night, right?)
  
  Maureen