Hi,

I don't get the result I expect to get when assigning a sparse matrix to another.
Even:

SparseMatrix<double> L(5,5);
L(0,0) = 1;
L = -L;

results in L resets to zeros. It's as if the assignment operator works as resizeLike (non conservative).

why don't you test the following code.

L = -1 * L.eval();

cheers,
yeonchool

Okay, so you have this magic undocumented function, which performs an assignment as expected. But don't you think it's a bug that the assignment operator doesn't behave like an assignment operator?

EDIT: Okay, so it is in the docs, but what does this cryptic:
"the matrix or vector obtained by evaluating this expression."
means?

zoharl wrote:SparseMatrix<double> L(5,5);
L(0,0) = 1;
L = -L;

results in L resets to zeros. It's as if the assignment operator works as resizeLike (non conservative).

Perhaps you need some more explanation. This is because you have the same matrix on the left-hand side and on the right-hand side of the assignment operator (we call this "the aliasing problem"). If you write M = -L, then everything works as expected. So, what you need to do is something like:

Code: Select all

SparseMatrix<double> L(5,5), tmp;
L(0,0) = 1;
tmp = L;
L = -tmp;


fearbone's code with eval() is a shortcut for achieving this.

The reason for this unfortunate behaviour is that it is impossible to write code that is efficient for both the case "M = -L" and "L = -L" (actually, in this case it may be possible to fix it, but not in general). It is noted in the documentation, on tutorial page 2 at http://eigen.tuxfamily.org/dox/Tutorial ... metic.html and we have a whole page about it at http://eigen.tuxfamily.org/dox/TopicAliasing.html . However, that is only in the context of dense matrices. If L is a dense matrix, then "L = -L" will in fact do what you expect. I'm afraid the documentation of sparse matrices is lagging a bit behind.

Ho, I see. Somehow I remembered that it happens also for different variables assignment, else I would have immediately considered the aliasing problem.
Thanks

hm, this should really be fixed. Actually with sparse matrices we should alway evaluate into a temporary and then swap pointers. I'm working on a patch.