I am running the
FEAST method option for
Eigenvalues on sparse matrices of dimension around 500,000. I look for about 50 eigenvalues. I have noticed a massive sensitivity of performance on what the entries are.
For example, I can have two sparse matrices with exactly the same entries non-zero. I simply modify the complex phase of some entries. This can result in runtimes of 5x longer or more.
Has anyone encountered this problem or any possible solutions?