Closed Luke-Pratley closed 8 years ago
I quickly coded the power method as a function in the measurement operator, and it looks like my idea fixed the problem.
I think that’s what we used to do in the example scripts. Basically, we used the normalised visibilities:
y1 = y/sqrt(nu)
and the normalised operator:
Phi_1 = Phi/sqrt(nu)
where nu is the operator norm. This was done only to accelerate the convergence of the CG solver.
On 23 Feb 2016, at 12:53, Luke-Pratley notifications@github.com wrote:
I quickly coded the power method as a function in the measurement operator, and it looks like my idea fixed the problem.
— Reply to this email directly or view it on GitHub https://github.com/astro-informatics/purify/issues/27#issuecomment-187669869.
Cool, that sounds like what I have done. And yes, I have done it for the same reason. Good to know that it is consistent.
At the start of the conjugate gradient method, the residuals are quite large. However, this depends on the gridding kernel used in the measurement operator. This gives me the impression it is an issue of normalisation. So, we need to normalise A^T * A, probably using the power method.