HSS. Leaving uncompressed blocks and changing diagonal values. about strumpack HOT 3 OPEN

Hi, apologies for the late reply.

We have support for the block low-rank format, which is a non-hierarchical format, and which allows to compress only certain (admissible) blocks. For HSS, all off-diagonal blocks are always compressed.

Yes, there is a routine
void shift (scalar_t sigma)

from strumpack.

Hi Pieter,

Thanks for the reply.

1. So not applicable for HSS. I could make a tree that puts the weird parts in a separate block.

2. shift multiplies with a scalar. I need to add a vector to the diagonal.
   I.e. I have to solve several of systems of equations where the matrix only differs in the diagonals. So I want to compress it once and only do factorization and solve for each system.

I don't see such function so I will try to write it myself.

Thanks again.

from strumpack.

Indeed, you can do a lot by changing the tree.

Yes, the shift routine adds a scalar times the identity matrix to the HSS matrix, which might be useful for things like shift and invert eigensolvers. We did not anticipate the use case of adding a vector to the diagonal. You can look at the implementation of the shift routine.
Let me know if you need help.

Pieter

from strumpack.