Ilhan Polat
Ilhan Polat
@xoviat Still using home-baked norms and givens rotations. We need to convert them into LAPACK calls. So again needs work. I don't mind but seems unnecessarily bloated.
@ashimb9 What kind of thresholds?
Hi @horta Thanks for taking the time to report it. Here when you use `(L, True)` that means the original matrix is simply `[[1.]]` hence you are solving a 1x1...
How interesting, I didn't realize this was possible. I'll check what we did in the LAPACK wrapper when I have a bit more time.
I lost track of this so if you folks are happy with it, I'm OK with it, @DietBru wanna take a look?
if you cast your A matrix as complex, then it works in both cases. So probably when the LHS is real it selects a real-typed solver and complains.
This is just an oversight/shortcoming to automatically discover the `dtype` for preperation to ARPACK send, probably because it came before `svds` without this mechanism. Is there a reason to remove...
I haven't checked but probably that's due to ARPACK limitation that it handles only float32, float64, complex64 and complex128 in fortran dtypes. So SciPy has to upcast to the suitable...
I looked into it again and the reason svds works is because it makes an additional rather costly operation to infer the dtype https://github.com/scipy/scipy/blob/a4b79f899296f26330d03bf912c97a2c0b075ba4/scipy/sparse/linalg/_eigen/_svds.py#L454-L457 Hence that's the reason why `svds`...
Oh my, I didn't know I have users for this library. You are right, typically this is not made for large arrays since most operations are `O(n**3)` or more (I...