[Numpy-discussion] strange behavior of numpy.random.multivariate_normal, ticket:1842

[josef.pktd at gmail.com]

On Thu, Feb 16, 2012 at 4:44 AM, Pauli Virtanen <pav at iki.fi> wrote:
> Hi,
>
> 16.02.2012 06:09, josef.pktd at gmail.com kirjoitti:
> [clip]
>> numpy linalg.svd doesn't produce always the same results
>>
>> running this gives two different answers,
>> using scipy.linalg.svd I always get the same answer, which is one of
>> the numpy answers
>> (numpy random.multivariate_normal is collateral damage)
>
> Are you using a Windows binary for Numpy compiled with the Intel
> compilers, or maybe linked with Intel MKL?

This was with the official numpy installer, compiled with MingW

I just tried with 64 bit python 3.2 with MKL (Gohlke installer) and in
several runs I always get the same answer.

>
> If yes, one possibility is that the exact sequence of floating point
> operations in SVD or some other step in the calculation depends on the
> data alignment, which can affect rounding error.
>
> See http://www.nccs.nasa.gov/images/FloatingPoint_consistency.pdf
>
> That would explain why the pattern you see is quasi-deterministic. The
> other explanation would be using uninitialized memory at some point, but
> that seems quite unlikely.

Running the script on the commandline I always get several patterns,
but running the script in the same process didn't converge to a unique
pattern.

We had other cases of several patterns in quasi-deterministic linalg
before, but as far as I remember only in the final digits of
precision, where it didn't matter much except for reducing test
precision in my cases.

In the random multivariate normal case in the ticket the differences
are large, which makes them pretty unreliable and useless for
reproducability.

Josef

>
> --
> Pauli Virtanen
>
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion