DSTEMR

[oamarques]

DSTEMR may return INFO>0. There was a bug in LAPACK 3.5.0 that was troublesome: a matrix of dimension 100 (a sort of perturbation of the identity matrix, provided by the developers of the R package), caused a disastrous infinite loop in one of the subroutines in the calling tree of DSTEMR (symmetric tridiagonal eigensolver based on the MRRR algorighm). That was because of NaNs that the safeguard mechanisms implemented therein failed to deal with. As a fix, the value of a logical flag that was related to the infinite loop was changed, starting with LAPACK 3.6.0. With that change the risk of an infinite loop has disappeared. However, we found out that that change to the logical variable could lead to unexpected consequences: DSTEMR may return INFO>0 (indicating an internal error in the auxiliary subroutine DLARRV) for some matrices that are not particularly “pathologic.” We are in the process of redesigning a few of the underpinnings of DSTEMR (including DLARRV) to make it more robust.

[oamarques]

We (Beresford Parlett and I) have implemented a new (and simpler) version of DLARRF, which is called by DLARRV, and which passed all our exhaustive tests. However, we would like to "torture" the code a little more before releasing it.

[uihsnv]

Is this going to be in ver 3.9?

[miroi]

Hello, any news concerning this issue ? Based on https://github.com/Reference-LAPACK/lapack/issues/151#issuecomment-640199933 fixing this can solve https://github.com/Reference-LAPACK/lapack/issues/151 .