LAPACKException(1) during SVD

[Steven Sagaert]

when doing an SVD of a large matrix I get

ERROR: LAPACKException(1)

 in gesdd! at linalg/lapack.jl:1046

 in svdfact! at linalg/factorization.jl:660

 in svdfact at linalg/factorization.jl:664

It's definitely something related to the data because it works on different matrices & the code has been working for months. Any idea what that error is about?

I'm working on Ubuntu 14.04, julia version 0.3.5

[Andreas Noack]

This is an issue with the underlying divide and conquer algorithm used by the LAPACK routine we are calling. It has been reported at Numpy's list as well.

I think we should have an option for choosing the algorithm in svdfact, such that you can easily switch to a QR based solver when the DnC solver fails. Please open an issue with this. For at short term solution, you can get your values and vectors by directly calling LAPACK.gesvd!('S','S',A) where A is your matrix.

[Tamas Papp]

I think that LAPACK error codes above 0 are about ill-conditioning
and/or non-convergence. Perhaps you could share the matrix (in a binary
format, since copy-pasting decimal floats can change the values and
thus the conditioning).

Best,

Tamas