Your comments

Ok, thanks. I know it's running on CentOS 5, but I've asked for more details. I'll let you know as soon as possible.

Indeed, thanks for your input.
I tried M*1/norm0 which did not work, but no with the brackets.

However, it seems to me that max(M) with M sparse is really not optimized : for some reason, my M is initialized as a very large sparse matrice of size 4096*4096 but with only 16 non-zero elements. It takes almost 30 seconds to find the largest element, whereas if I put all the non-zero elements in a separate matrix, it takes a mere 0.04 seconds.