Thread Subject:

sparse matrix multiplication: bad in Linux v7.5, good in Windows v7.1

Dear all,

My Matlab v7 (2005) matlab manual tells me that "Multiplication and division are performed on only the nonzero elements of sparse matrices". Such a set-up is very convenient for me.

On my older Windows matlab 7.1.0.124 (R14) (August 2005), multiplication (division) works just like the manual says it should:

EDU>> sparse([1 0 0 0 0])./0
ans =
   (1,1) Inf

But on a newer Linux matlab 7.5.0.338 (R2007b), the zero elements are also involved -- this is a big nuisance:

>> sparse([1 0 0 0 0])./0
ans =
   (1,1) Inf
   (1,2) NaN
   (1,3) NaN
   (1,4) NaN
   (1,5) NaN

How could I revert to the old version of multiplication? If that's not possible, how could I use spfun with two matrix inputs ... spfun(@times,[2 input matrices?]).

Many thanks for your help.

>
> how could I use spfun with two matrix inputs ... spfun(@times,[2 input matrices?]).
>

times(sa,sb) %???

Bruno

"Akim " <aaa@bbb.ccc> wrote in message
news:gl3j48$hc4$1@fred.mathworks.com...
> Dear all,
>
> My Matlab v7 (2005) matlab manual tells me that "Multiplication and
> division are performed on only the nonzero elements of sparse matrices".
> Such a set-up is very convenient for me.
>
> On my older Windows matlab 7.1.0.124 (R14) (August 2005), multiplication
> (division) works just like the manual says it should:
>
> EDU>> sparse([1 0 0 0 0])./0
> ans =
> (1,1) Inf
>
> But on a newer Linux matlab 7.5.0.338 (R2007b), the zero elements are also
> involved -- this is a big nuisance:
>
>>> sparse([1 0 0 0 0])./0
> ans =
> (1,1) Inf
> (1,2) NaN
> (1,3) NaN
> (1,4) NaN
> (1,5) NaN
>
> How could I revert to the old version of multiplication? If that's not
> possible, how could I use spfun with two matrix inputs ... spfun(@times,[2
> input matrices?]).

You can't revert to the old behavior, as it was a bug that was fixed in
MATLAB 7.2 (R2006a):

http://www.mathworks.com/support/bugreports/details.html?rp=245912

In order to do what you want with SPFUN, you'll need to write your own
function (which can be an anonymous function):

S = sparse([1 0 0 0 0]);
result = spfun(@(x) x./0, S)

--
Steve Lord
slord@mathworks.com

Or, you could you use "find" to extract the nonzeros, then do the scalar ./ or .* or whatever, and then rebuild the sparse matrix:

% note, not the best way to form a sparse matrix, but this is just a small example
% so it's OK:
A = sparse ([1 0 0 0 0]) ;

% now get the old behavior of ignoring the zeros in scalar A./s
s = 0 ;
[i j x] = find (A) ;
[m n] = size (A) ;
B = sparse (i, j, x./s, m, n) ;

Note the x./s, which operates on a dense matrix containing just the nonzeros of A,
which gives the result

B =

   (1,1) Inf

>> size (B)

ans =

     1 5

Thanks Tim and Steven, and Bruno for your help.

> % now get the old behavior of ignoring the zeros in scalar A./s
> s = 0 ;
> [i j x] = find (A) ;
> [m n] = size (A) ;
> B = sparse (i, j, x./s, m, n) ;

I think these code basically constitutes spfun, right?

In any case, it seems that given this bug fix, the simplest but somewhat crude solution is to divide by a very small, but non-zero number. E.g.

>> sparse([1 0 0 0 0])./1e-20
ans =
   (1,1) 1.0000e+20

For me this approximation is accurate enough.
Akim