In short, we wish to find a row- and column ordering, and a sequence of Givens Rotations, that together will produce a QR factorization of our system of equations with a minimal computational cost.
Now, there are
combinations of row- and column orderings, if
our system has
equations with
unknowns. And if we have
non-zeros in the system, we will need
Givens Transforms,
giving us
combinations there. All in all, we have
possible combinations. For the
system
with roughly 12% non-zeros, this amounts to a number of combinations
in the order of
.
Clearly, exhaustive search thru the solution space is not an option. We will have to apply some strategy, in the hope that we can somewhat limit out search to some small number of small sub-spaces of the solution space.