[Rd] RFC: diag(x, n) not preserving integer and logical x

[Martin Maechler]

>>>>> Duncan Murdoch <murdoch.duncan at gmail.com>
>>>>>     on Fri, 8 Aug 2014 10:51:10 -0400 writes:

Thank you, Duncan (read inline) 

    > On 07/08/2014, 4:51 AM, Martin Maechler wrote:
    >> This is not at all something new(*). As maintainer of the
    >> Matrix package, I don't like this inconsistency of base
    >> R's diag().  We have had the following -- forever, almost
    >> surely inherited from S and S+ :
    >> 
    >> diag(x) preserves the storage mode of x for 'complex' and
    >> 'double' precision, but converts integer and logicals to
    >> double :
    >> 
    >> > storage.mode(x <- 1i + 1:7); storage.mode(diag(x)) [1]
    >> "complex" [1] "complex" > storage.mode(x <- 0 + 1:7);
    >> storage.mode(diag(x)) [1] "double" [1] "double"
    >> 
    >> > storage.mode(x <- 1:7); storage.mode(diag(x)) [1]
    >> "integer" [1] "double" > storage.mode(x <- 1:7 > 3);
    >> storage.mode(diag(x)) [1] "logical" [1] "double"
    >> 
    >> and so it is actually a bit cumbersome (and a memory
    >> waste in the case of large matrices) to create a diagonal
    >> integer or logical matrix.
    >> 
    >> The help page does not mention the current behavior,
    >> though you may say it alludes to the fact that logicals
    >> are treated as 0/1 implicitly (**)
    >> 

    > I think the change to preserve integer makes sense, but
    > preserving logical does not.  A diagonal matrix has zeros
    > off the diagonal, and they are not logical.  

That's true if you come from the usual mathematical thinking
about matrices; and the same would extend to triangular
matrices, too.

In (S and) R however, matrices and array()s have always been
more general than in applied math (and have allowed any atomic (and even
more general) content).

    > Having diag() sometimes return a matrix with FALSE off the diagonal just
    > looks wrong.

[... if you have never seen them ..]
OTOH, sparse matrices (in the Matrix package) and also sparse
and dense triangular matrices (all with their classes) have
always (in the Matrix package) had their logical counter parts,

  > Diagonal(5)
  5 x 5 diagonal matrix of class "ddiMatrix"
       [,1] [,2] [,3] [,4] [,5]
  [1,]    1    .    .    .    .
  [2,]    .    1    .    .    .
  [3,]    .    .    1    .    .
  [4,]    .    .    .    1    .
  [5,]    .    .    .    .    1
  > Diagonal(5) > 0
  5 x 5 diagonal matrix of class "ldiMatrix"
       [,1] [,2] [,3] [,4] [,5]
  [1,] TRUE    .    .    .    .
  [2,]    . TRUE    .    .    .
  [3,]    .    . TRUE    .    .
  [4,]    .    .    . TRUE    .
  [5,]    .    .    .    . TRUE
  > 

and if you consider any sparseness for logical matrices, typical
application cases would have the TRUE to be rare rather than the
FALSE... and then of course, we have had the  as.numeric(<logical>)
behavior of S and R, all this leading to 
the usual  FALSE <==> 0  mapping 


    >> If I change this behavior such that logical and integer x
    >> are preserved,
    >> 
    >> make check-all
    >> 
    >> which includes all checks, including those of all
    >> recommended packages (including Matrix!) successfully
    >> runs through; so at least base + Recommended R never
    >> relies on the current behavior, nor should any "well
    >> programmed" R code ...
    >> 
    >> Hence my proposal, somewhat tentative for now, to change
    >> this diag(.) behavior.
    >> 
    >> Martin Maechler
    >> 
    >> *) and possibly something we "can not" change in R,
    >> because too much code implicitely may be depending on it,
    >> but now I hope we can still...
    >> 
    >> **) BTW, also including the somewhat amusing case of
    >> diag(c("A","B")).
    >> 
    >> ______________________________________________
    >> R-devel at r-project.org mailing list
    >> https://stat.ethz.ch/mailman/listinfo/r-devel
    >>