One can show that in a general sparse matrix, it is straightforward to preserver the "end" zeros |
000...000X000...000XXX000...000X000...000 shows structure of a row of A in 2D Laplace example |
The first and last zeros can be explicitly taken into account by setting range of for loops properly |
The inside zeros "fill" i.e. become nonzero during process of Gaussian elimination |
Thus if bandwidth B i.e. |
000...000X000...000XXX000...000X000...000 |
Then computational complexity is not N3 as for a full (no zeros accounted) but NB2 |