I think I finally understand why software packages like PETSc has an option for an operator when doing something like conjugate gradient. Why isn’t having a matrix good enough for everyone?
Well turns out that while all linear operators can be translated to a matrix, it may not be the best way to represent the operator. As an example, consider a basis transformation from Bernstein polynomials to Jacobi (or vice versa). It’s certainly possible to find and construct a matrix which does the operation, but it’s ugly.
On the other hand, it’s not that bad to write a code which utilizes the properties of the polynomials and convert it within in O(n^2) time. The key is that a Jacobi polynomial is a sum of Bernsteins, and Bernsteins can be degree raised or dropped at will.
This function will outperform the matrix in many sense. For one, there’s no need to construct a matrix, which will take n^2 operations in the first place. Next, matrix multiplication will take a n^3 operation, so if we optimize enough, we will always beat it. Finally, it’s really less painful to code, because each line of the function serves a visible purpose.
Anyways, I’m sold.
(I’ll eventually publish the code in the summer)