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

Pauli Virtanen pav@iki...
Thu Feb 16 08:08:46 CST 2012


16.02.2012 14:54, josef.pktd@gmail.com kirjoitti:
[clip]
> If I interpret you correctly, this should be a svd ticket, or an svd
> ticket as "duplicate" ?

I think it should be a multivariate normal ticket.

"Fixing" SVD is in my opinion not sensible: its only guarantee is that A
= U S V^H down to numerical precision and S are sorted. If the algorithm
assumes something extra, it is wrong. This sort of reproducibility
issues affect potentially all code (depends on the compiler and
libraries used), and trying to combat it at the linalg level is IMHO not
our business --- if someone really wants it, they should tell their C
compiler and all libraries to use a reproducible FP model.

However, we should ensure the algorithms we provide are stable against
rounding error. In this case, the random number generation is not, so it
should be fixed.

	Pauli



More information about the NumPy-Discussion mailing list