RFC: Kerning, postscript() and pdf()

Ei-ji Nakama has pointed out (from another Japanese user, I believe) that 
postscript() and pdf() have not been handling kerning correctly, and this 
is a request for opinions about how we should correct it.

Kerning is the adjustment of the spacing between letters from their 
natural width, so that for example 'Yo' is usually typeset with the o 
closer to the Y than 'Yl' would be.  Kerning is not very well 
standardized, so that for example R's default Helvetica and its URW clone 
(Nimbus Sans) have quite different ideas of the amount of kerning 
corrections for 'Yo'. This matters, because not many people actually see 
Helvetica when viewing R's PostScript or PDF output, but rather a similar 
face like Nimbus Sans or Arial, or in the case of Acrobat Reader, a not 
very similar face.  Kerning is only a feature of some proportionally 
spaced fonts and so not of Courier nor CJK fonts.

The current position (R <= 2.8.0) is that string widths have been 
computing using kerning from the Adobe Font Metric files for the nominal 
font, but the strings have been displayed without using kerning (at least 
in the viewers we are aware of, and the PostScript and PDF reference 
manuals mandate that behaviour, if rather obscurely).  This means that in 
strings such as 'You', the width used in the string placement differs from 
that actually displayed.

For postscript(), this doesn't have much impact, as centring or right 
justification ('hadj' in text()) is done by PostScript code and computes 
the width from the actual font used (and so copes well with font 
substitution).  It might affect the fine layout in plotmath, but using 
strings which would be kerned in annotations is not common.

For pdf() the effect is more commonly seen, as all text is set 
left-justified, and the computed width is used to centre/right-justify.

There are several things we could do:

A.  Do nothing, for back compatibility.  After all, this has been going on 
for years and no one has complained until last month.

B.  Ignore kerning, and hence change the string width computations to 
match the current display.  This is more attractive than it appears at 
first sight -- as far as I know all other devices ignore kerning, and we 
are increasingly used to seeing 'typeset' output without kerning.  It 
would be desirable when copying graphs by e.g. dev.copy2eps from devices 
that do not kern.

C.  Insert kerning corrections by splitting up strings, so e.g. 'You' is 
set as (Y)-140 kc(ou): this is what TeX engines do.

D.  Compute the position of each letter in the string and place them 
individually.

C and D would give visually identical output when the font used is exactly 
as specified, and hopefully also when a substitute font is using with the 
same glyph widths (as substituting Nimbus Sans for Helvetica, at least for 
some versions of each), but where the substitute is a poor match, C ought 
to look more elegant but line up less well.  D would produce much larger 
files than C.

We do have the option of not changing the output when there is no kerning. 
That would be by far the most common case except that some fonts 
(including Helvetica but not Nimbus Sans) kern between punctuation and a 
space, e.g. ', '.  I'm inclined to believe that most uses of ',' in R 
graphical output are not punctuation (certainly true of R's own examples), 
and also that we nowadays do not expect to see kerning involving spaces.

Ei-ji Nakama provided an implementation of C for pdf() and D for 
postscript() (thanks Ei-ji, and apologies that we did not have a chance to 
discuss the principles first).  I'm inclined to suggest that we should go 
forwards with at most two of these alternatives, and those two should be 
the same for postscript() and pdf() -- my own inclination is to B and C.

So questions:

1) Do people feel strongly that we should preserve graphical output from 
past versions of R, even when there are known bugs?  I can see the need to 
reproduce published figures, but normally this would also need using the 
same version of R.

2) Is kerning worth pursuing?

3) If so, is elegant looking output more important than exact layout?

4) If we allow kerning, should it be the default (or only) option?

To see that sometimes there can be a large effect, try in postscript() or 
pdf()

xx <- 'You You You You You You You You'
plot(0,0,xlim=c(0,1),ylim=c(0,1),type='n')
abline(v=0)
text(0, 0.5, xx, adj=0)
abline(v=strwidth(xx))
x2 <- strsplit(xx, "")
w <- sapply(x2, strwidth)
abline(v=sum(w))

The leftmost of the right pair of lines is the computed width, the 
rightmost the (normal) displayed width.

Unless there are cogent reasons to bring this forward to 2.8.1, any 
changes would be as from 2.9.0.

Brian Ripley