Open BambOoxX opened 3 years ago
No, this is okay to leave open. It turned out that I couldn't leverage so much of this code to implement the Auretz paper, so it made more sense to just add it to SpecialPolynomials.
@jverzani All right. I will add the matrix implementations I got so far tomorrow. As I said earlier, numerical issues seem to occur faster for complex coefficients even for scalar polynomials it seems. I wonder if it is due to the default solver, or else...
Yes, I've noticed that too when implementing AMRVW. I assumed it was due to more computations creating more opportunity for floating point drift. But it might be something else.
The code in SpecialPolynomials in the PR is such that this might be reverse. The authors note that the single shift case (which is the one complex values use) has better properties than the double shift case.
I'm not sure that handles the block comrade matrices you ulitmately desire though.
This issue is based on the discussion initiated in this Discourse thread
Here are some toy implementations to compare different strategies to compute roots of polynomials expressed in terms of orthogonal or canonical polynomials. Theoretical background is provided in papers cited on the discourse thread (recalled here for convenience)
This only tests a comrade implementation versus existing roots evaluation. At the moment, I do not know which way would be best to