Hi, Now I think I have a correct solution. And I finally understood what the original code was doing; it would have been correct if there was an additional line in the adjacency triangular matrix. Indeed the iterator didn't try to stop on the first element outside the matrix but on the first line outside the matrix. So the patch I now submit transforms the edge iterator so that the end iterator doesn't go so far but stop at the end of the matrix. It involves a special casing in the increment of the iterator so that it doesn't try to reach far outside the matrix. In the middle of the patch also lies my correction for the const-ness problem of the get_edge function I was explaining in the previous mail. Regards, Guillaume PS: And so that there is no misunderstanding: **passed** ../../../bin/boost/libs/graph/test/adjacency_matrix_test.test/gcc-3.4/debug/adjacency_matrix_test.test **passed** ../../../bin/boost/libs/graph/test/adjacency_matrix_test.test/gcc/debug/adjacency_matrix_test.test **passed** ../../../bin/boost/libs/graph/test/adjacency_matrix_test.test/intel-linux/debug/adjacency_matrix_test.test :-) PPS: As a side note, I think the iterator could be made smaller by removing the m_inc variable and replacing it by m_targ since they are equivalent on the second half of the road and m_inc is not used on the first half. But it is a story for another time.