Skip to content

Unexpected errors in sparse Matrix arithmetic with zero-length dimensions

2 messages · Aaron Lun, Martin Maechler

Aaron Lun
# 
Dear list,

The Matrix package exhibits some unexpected behaviour in its arithmetic
methods for the edge case of a sparse matrix with a dimension of zero
length. The example below is the most illustrative, where changing the
contents of the vector causes the subtraction to fail for a sparse
matrix with no columns:?
????
Error in .Ops.recycle.ind(e1, len = l2) :?
? vector too long in Matrix - vector operation

This is presumably because Matrix recognizes that subtraction of zero
preserves sparsity and thus uses a different method in the second case.
However, I would have expected subtraction of a zero vector to work if
subtraction of a general vector is permissible. This is accompanied by
a host of related errors for sparsity-preserving arithmetic:
Error in .Ops.recycle.ind(e1, len = l2) :?
? vector too long in Matrix - vector operation
Error in .Ops.recycle.ind(e1, len = l2) :?
? vector too long in Matrix - vector operation
??????
A different error is raised for a sparse matrix with no rows:
Error in y - numeric(0) : <Matrix> - numeric(0) is undefined

I would have expected to just get 'y' back, given that the same code
works fine for other Matrix classes:
Correct behaviour of zero-dimension sparse matrices is practically
important to me; I develop a number of packages that rely on Matrix
classes, and in those packages, I do a lot of unit testing with zero-
dimension inputs. This ensures that my functions return sensible
results or fail gracefully in edge cases that might be encountered by
users. The current behaviour of sparse Matrix arithmetic causes my unit
tests to fail for no (obvious) good reason.

Best,

Aaron Lun

Research Associate
CRUK Cambridge Institute
University of Cambridge
R Under development (unstable) (2019-01-14 r75992)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 16.04.5 LTS

Matrix products: default
BLAS: /home/cri.camres.org/lun01/Software/R/trunk/lib/libRblas.so
LAPACK: /home/cri.camres.org/lun01/Software/R/trunk/lib/libRlapack.so

locale:
?[1] LC_CTYPE=en_GB.UTF-8???????LC_NUMERIC=C??????????????
?[3] LC_TIME=en_GB.UTF-8????????LC_COLLATE=en_GB.UTF-8????
?[5] LC_MONETARY=en_GB.UTF-8????LC_MESSAGES=en_GB.UTF-8???
?[7] LC_PAPER=en_GB.UTF-8???????LC_NAME=C?????????????????
?[9] LC_ADDRESS=C???????????????LC_TELEPHONE=C????????????
[11] LC_MEASUREMENT=en_GB.UTF-8 LC_IDENTIFICATION=C???????

attached base packages:
[1] stats?????graphics??grDevices
utils?????datasets??methods???base?????

other attached packages:
[1] Matrix_1.2-15

loaded via a namespace (and not attached):
[1] compiler_3.6.0??grid_3.6.0??????lattice_0.20-38
4 days later
Martin Maechler
# 
> Dear list,
    > The Matrix package exhibits some unexpected behaviour in its arithmetic
    > methods for the edge case of a sparse matrix with a dimension of zero
    > length. The example below is the most illustrative, where changing the
    > contents of the vector causes the subtraction to fail for a sparse
    > matrix with no columns:?
    > ????
    >> library(Matrix)
    >> x <- rsparsematrix(10, 0, density=0.1)
    >> ?
    >> x - rep(1, nrow(x)) # OK?
    >> x - rep(0, nrow(x)) # fails
    > Error in .Ops.recycle.ind(e1, len = l2) :?
    > ? vector too long in Matrix - vector operation

This is indeed clearly a lapsus of us / mine  as well as the
next examples:  Will all be fixed  "around" .Ops.recycle.ind() 

    > This is presumably because Matrix recognizes that subtraction of zero
    > preserves sparsity and thus uses a different method in the second case.
    > However, I would have expected subtraction of a zero vector to work if
    > subtraction of a general vector is permissible. This is accompanied by
    > a host of related errors for sparsity-preserving arithmetic:

    >> x / 1 # OK
    >> x / rep(1, nrow(x)) # fails?
    > Error in .Ops.recycle.ind(e1, len = l2) :?
    > ? vector too long in Matrix - vector operation
    >> ?
    >> x * 1 # OK
    >> x * rep(1, nrow(x)) # fails
    > Error in .Ops.recycle.ind(e1, len = l2) :?
    > ? vector too long in Matrix - vector operation
    > ??????
    > A different error is raised for a sparse matrix with no rows:

    >> y <- rsparsematrix(0, 10, density=0.1)
    >> ?
    >> y - numeric(1) # OK
    >> y - numeric(0) # fails
    > Error in y - numeric(0) : <Matrix> - numeric(0) is undefined

Thank you, that's another lapsus, I will fix before the next
release of Matrix.

    > I would have expected to just get 'y' back, given that the same code
    > works fine for other Matrix classes:

    >> z <- as(y, "dgeMatrix")
    >> z - numeric(0) # OK

sure.

    > Correct behaviour of zero-dimension sparse matrices is practically
    > important to me; I develop a number of packages that rely on Matrix
    > classes, and in those packages, I do a lot of unit testing with zero-
    > dimension inputs. This ensures that my functions return sensible
    > results or fail gracefully in edge cases that might be encountered by
    > users. The current behaviour of sparse Matrix arithmetic causes my unit
    > tests to fail for no (obvious) good reason.

Interesting that you need 0-dim sparse matrices.  I agree they
should work, too... and will fix
(but it seems they haven't been used much by others;  else I
would have expected these cases to have been reported long ago).

Further note that in R,

  maintainer("Matrix")

gives a nice address to send such findings.
(similarly for all other R packages !)

Thank you very much once more!
Martin

    > Best,
    > Aaron Lun

    > Research Associate
    > CRUK Cambridge Institute
    > University of Cambridge

--
Martin <Maechler at ....>   
Seminar f?r Statistik, ETH Z?rich   HG G 16         R?mistrasse 101
CH-8092 Zurich, SWITZERLAND