[Scipy-tickets] [SciPy] #869: stats.invnorm.cdf returns NaNs with low mu values

SciPy Trac scipy-tickets@scipy....
Fri Aug 19 21:51:50 CDT 2011


#869: stats.invnorm.cdf returns NaNs with low mu values
-------------------------+--------------------------------------------------
 Reporter:  jpaalasm     |       Owner:  somebody    
     Type:  defect       |      Status:  needs_review
 Priority:  normal       |   Milestone:  0.10.0      
Component:  scipy.stats  |     Version:  devel       
 Keywords:               |  
-------------------------+--------------------------------------------------

Comment(by josefpktd):

 I have no new ideas.
 If we had a numerically precise norm.logcdf then this could help, if
 someone finds a proof that the 0*inf term is actually zero, we could
 restrict to the first term.

 mu = 0.0028 seems a bit large for breaking down already, but the variance
 is already pretty small 0.0028**3=2.1952e-08 and so the pdf is quite
 spiky.

 R library VGAM: same breakdown and nans

 {{{
 > pinv.gaussian(0.003, 0.0029, 1)
 [1] 0.7443139
 > pinv.gaussian(0.003, 0.0028, 1)
 [1] NaN
 }}}

 Ralf: comment for docs looks good

 I would add reference to this ticket as code comment. Because mu=0.0028 is
 quite large and the pdf and veccdf still work, I think there is room for
 improvement. If we close the tickets, then code comments would give at
 least a hint that there is a problem and possible (approximate) work
 arounds.

-- 
Ticket URL: <http://projects.scipy.org/scipy/ticket/869#comment:6>
SciPy <http://www.scipy.org>
SciPy is open-source software for mathematics, science, and engineering.


More information about the Scipy-tickets mailing list