[Rd] Recent changes to as.complex(NA_real_)

[Martin Maechler]

>>>>> Martin Maechler 
>>>>>     on Thu, 28 Sep 2023 12:11:27 +0200 writes:

>>>>> Gregory R Warnes 
>>>>>     on Sat, 23 Sep 2023 13:22:35 -0400 writes:

>     > It sounds like we need to add arguments (with sensible
>     > defaults) to complex(), Re(), Im(), is.na.complex() etc to
>     > allow the user to specify the desired behavior.
> 
> I don't think I'd like such extra flexibility for all these, 
> ... ;-)  and even much less I'd like to be part of the group who
> then has to *maintain* such behavior ;-)
> 
[..........]

> Currently, I'm actually tending to *simplify* things
> drastically, also because it means less surprises in the long
> term and much less code reading / debugging in formatting /
> printing and dealing with complex numbers.
> NB: there *is* the re-opened PR#16752,
>     https://bugs.r-project.org/show_bug.cgi?id=16752
> where the investigation of the (C-level) R source is a major reason
> for my current thinking ..
> 
> What if we decided to really treat complex numbers much more
> than currently  as pairs of real (i.e. "double") numbers,
> notably also when print()ing them?
> Consequently, Re() and Im() would continue to return what they
> do now (contrary to Hervé's original proposal) also in case of
> non-finite numbers.
> 
> Of course, *no* change in arithmetic or other Ops (such as '==')
> nor  is.na(), is.finite(), is.nan(), etc.
> 
> The current formatting and printing of complex numbers is
> complicated in some cases unnecessarily inaccurate and in other
> cases unnecessarily *ugly*.
> I believe that formatting, we should change to basically format
> the (vector of) real parts and imaginary parts separately.
> E.g., it is really unnecessarily ugly to switch to exponential
> format for both Re and Im, in a situation like this:
> 
> > (-1):2 + 1i*1e99
> [1] 0e+00+1e+99i 0e+00+1e+99i 0e+00+1e+99i 0e+00+1e+99i
> 
> It is very ugly to use  exponential/scientific format for the Re()
> even if we'd fix the confusing and inaccurate *joint* rounding
> of Re and Im.

and then, I end with

> ... and indeed (as discusses here previously:
> While it makes some sense to print  NA
> identically for logical, integer and double,
> it seems often confusing *not* to show  <Re> + <Im>i
> in the complex case; where that *does* happen for
> Inf and NaN:
> 
>  > complex(, NA, ((-1):2))
>  [1] NA NA NA NA
>  > complex(, NaN, ((-1):2))
>  [1] NaN-1i NaN+0i NaN+1i NaN+2i
>  > complex(, c(-Inf,Inf), ((-1):2))
>  [1] -Inf-1i  Inf+0i -Inf+1i  Inf+2i
>  > 
> 
> where the first of these *does* keep the finite imaginary
> values, but does not show them
> 
> > (cN <- complex(, NA, ((-1):2))); rbind(Re(cN), Im(cN))
> [1] NA NA NA NA
>      [,1] [,2] [,3] [,4]
> [1,]   NA   NA   NA   NA
> [2,]   -1    0    1    2
> >

where really, I think we should keep that behavior (*), at least
for now: Changing it as well *does* have a relatively large
impact, is not back-compatible with (the long history of) S and
R, *and* it complicates documentation and teaching unnecessarily.

Experts will now how to differentiate the different complex NAs,
e.g. by using a simple utilities such as {"format complex", "print complex"}

  fc <- function(z) paste0("(",Re(z), ",", Im(z),")")
  pc <- function(z) noquote(fc(z))

which I've used now for testing/"visualizing" different scenarios


Martin


---
*) simply printing 'NA' in cases where is.na(.) is true and is.nan(.) is false