Mohamed Tarek

Well, if the whole `P` matrix disappears due to 0 eigenvalues, LOBPCG basically turns into a preconditioned gradient descent algorithm, as opposed to a preconditioned *conjugate* gradient algorithm, so the...

Perhaps both options can be provided once the constraint issue for QR is figured out, not that I am volunteering to implement both! If this turns out to be too...

The constrained case for QR seems simple to handle. We just need to check that any additional column of `Q` that we are including in the basis, for which the...

So it seems a similar method to the one proposed above was also proposed in https://arxiv.org/pdf/1704.07458.pdf with the name `svqbDrop`. I haven't read the paper in details but it seems...

I have my doubts about this "fix" since it doesn't really change the overall basis. I would be very surprised if it actually fixed all of @mfherbst 's test cases.

@mfherbst may you try this PR for your real application and see if it does the trick or not? With this, I am getting a 100% success rate for your...

This method seems to produce unexpected results elsewhere. I will implement the QR orthogonalization.

@mschauer if tests pass, please give this a second review.

I added some comments to the PR. In short, the role of orthogonalization (technically orthonormalization) in LOBPCG is to make sure that the matrix `X` is `B`-orthonormal. This necessarily means...

@lobpcg thanks for your advice. I will try the column normalization method. The nice thing about this approach is that no additional matrix multiplications would be required sine `A *...