Convergence issue of GMRES

My issue is about a convergence issue when using scipy.sparse.linalg.gmres routine. When using GMRES with a preconditioner, it seems that the convergence is tested by looking at the residual not the preconditioned residual. So, although the preconditioned residual is smaller than the absolute tolerance I ask, the iterations do no stop if the (not preconditioned) residual is greater than the tolerance. In order to fix this, I need to go into iterative.py and add the following lines (under the if conditional elif(ijob==3): if presid < atol and np.linalg.norm(x) > 0: resid = 0.0 break

This kind of hard coding is not something I prefer. I will appreciate if you can take a look at this issue and inform me about it.

Thanks!

Reproducing code example:

Sample code to reproduce the problem

Error message:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  ...

Scipy/Numpy/Python version information:

[zhaog6]

@rgommers Yes, I am rewriting the restarted GMRES. The original GMRES did use the residual instead of the preconditioned residual. @gokberkkabacaoglu you can trace the PR #15149

[zhaog6]

Edit: I found your repo and used the fastgmres branch. The GMRES iterations are indeed much faster (though partly from improvements to numpy).

Yes, there are also some improvements of data storage way based on Python and other mathematical optimizations.