I gave a look at the paper and at Higham's code pnorm.m, but I don't see any
random initial guess. On the other hand, if look at this run
octave:1> format long e
octave:2> S=sprand(5,5,0.5);
octave:3> norm(S)
ans = 1.25336637903553e+00
octave:4> norm(S)
ans = 1.25336637903553e+00
octave:5> norm(S)
ans = 1.25336638124303e+00
octave:6> norm(S)
ans = 1.25336638124303e+00
octave:7> norm(S)
ans = 1.25336638124303e+00
octave:8> norm(S)
ans = 1.25336638124303e+00
octave:9> norm(S)
ans = 1.25336638124303e+00
octave:10> norm(S)
ans = 1.25336638124303e+00
you see that the norm changes at the third computation and then it remains
constant. This is not a random behaviour. Can anybody reproduce it? I can
*almost* always reproduce it.
Yes, I was just able to reproduce that, and you're right, there's no
randomisation in the algorithm. This may point to a deeper error,
perhaps UB. I'm investigating with the debugger now...