[SciPy-user] differences between numpy.linalg.cholesky and scipy.linalg.cholesky?

[Emanuele Olivetti]

I'm using cholesky decomposition a lot and trying both
numpy.linalgcholesky and scipy.linalg.cholesky on
hearmitean positive definite matrix. Sometimes I
get """<class 'numpy.linalg.linalg.LinAlgError'>:
Matrix is not positive definite -
Cholesky decomposition cannot be computed"""
when using numpy's cholesky. No problems with
scipy's cholesky. Why?

Thanks in advance,

Emanuele

_______________________________________________
SciPy-user mailing list
SciPy...@scipy.org
http://projects.scipy.org/mailman/listinfo/scipy-user

[Dominique Orban]

On 1/5/08, Emanuele Olivetti <eman...@relativita.com> wrote:
> I'm using cholesky decomposition a lot and trying both
> numpy.linalgcholesky and scipy.linalg.cholesky on
> hearmitean positive definite matrix. Sometimes I
> get """<class 'numpy.linalg.linalg.LinAlgError'>:
> Matrix is not positive definite -
> Cholesky decomposition cannot be computed"""
> when using numpy's cholesky. No problems with
> scipy's cholesky. Why?

Could you provide a small example?

Dominique

[Emanuele Olivetti]

Dominique Orban wrote:
> On 1/5/08, Emanuele Olivetti <eman...@relativita.com> wrote:
>
>> I'm using cholesky decomposition a lot and trying both
>> numpy.linalgcholesky and scipy.linalg.cholesky on
>> hearmitean positive definite matrix. Sometimes I
>> get """<class 'numpy.linalg.linalg.LinAlgError'>:
>> Matrix is not positive definite -
>> Cholesky decomposition cannot be computed"""
>> when using numpy's cholesky. No problems with
>> scipy's cholesky. Why?
>>
>
> Could you provide a small example?
>
>

Currently it is not so easy to provide a simple example. I'm using
Cholesky factorization on 1000x1000 matrix generated from
some large datasets and getting the error few times. But I'll try to
generate one small soon.

Looking in numpy/scipy source code I see that:
- numpy.linalg.cholesky wraps the fortran function "dpotrf"
- scipy.linalg.decomp wraps "potrf"

Which is the difference between 'dpotrf' and 'potrf' ?

Thanks,

Emanuele

[Robert Kern]

Emanuele Olivetti wrote:
> Looking in numpy/scipy source code I see that:
> - numpy.linalg.cholesky wraps the fortran function "dpotrf"
> - scipy.linalg.decomp wraps "potrf"
>
> Which is the difference between 'dpotrf' and 'potrf' ?

Most LAPACK subroutines have variants for each type specified by a single-letter
prefix in front of the subroutine name.

S = float32
D = float64
C = complex64
Z = complex128

One of the features of f2py, which is used to make scipy.linalg.decomp but not
numpy.linalg, is that it can have templated interface definitions so one
declaration will wrap all four variants.

If you call scipy.linalg.cholesky with a float64 array, it will ultimately call
DPOTRF, too.

--
Robert Kern

"I have come to believe that the whole world is an enigma, a harmless enigma
that is made terrible by our own mad attempt to interpret it as though it had
an underlying truth."
-- Umberto Eco