Radim Řehůřek

I can reproduce this in 0.1.9, thanks for the report y. @alepulver this error doesn't happen in 0.1.8 -- any idea where the regression's coming from?

0.1.8 returns ``` (array([[ 0.32372233, 0.4780341 , 0.79283656, 0.19518564], [-0.19518564, 0.79283656, -0.4780341 , 0.32372233]]), array([ 1.30002747, 1.30002747]), array([[ 0.36771077, 0.60121131, 0.36249528, 0.60986139], [ 0.60986139, -0.36249528, 0.60121131, -0.36771077]])) ``` (same machine)

Heh, might well be connected to the "fewer factors than requested" discussed elsewhere... @alepulver Is the output deterministic for you (0.1.8 works, 0.1.9 doesn't)? Or does it fail intermittently even...

Aha, ok, thanks. I'll have to investigate more thoroughly then :) I'll check to see if vanilla SVDLIBC behaves the same way too (no Python), and then it's debugging time...

I debugged, and the problem seems to be somewhere deep inside LAS2. The error bounds (`bnd` array coming out of `lanso`) are rubbish in face of algebraic multiplicity, although the...

I sent an email to Doug Rohde, hopefully he'll shed some light.

Doug's reply: "I haven't worked on SVDLIBC in years. And it was a port of someone else's code, so I never fully understood the majority of it. Mostly I was...

Divisi2 uses SVDLIBC too, and therefore suffers from the same problems. I already created and maintain a package for large-scale sparse SVD (`gensim`), so this is not a big issue...

Or let's put it differently -- the person who can successfully resolve this issue will be crowned the king of `sparsesvd` :)

I don't think that's a good idea @alepulver . Partly because we don't even know what "those cases" are. I'll try to delve into the details and fix this, but...