[SciPy-dev] Problem with flapack

Fernando Pérez fperez at pizero.colorado.edu
Fri Feb 15 15:15:30 CST 2002


> 
> Sure, but I agree with Heiko that tiny zeros should be detected internally and
> cleaned up automatically.

Careful. I can't think off the top of my head of a concrete example, but I can
impgine easily cases where you really have an unusual mix of roots with
significant real parts and tiny imaginary components. In fact, in some of my
work something similar happens as certain parameters change and eigenvalues
for a certain operator start migrating off the real axis.

So I would rather have the numerical algorithms spit out whatever answer they
get and let *me* do any cosmetic cleanup I want to after the fact. If we want
convenience, a keyword parameter, *off* by default, could control this cleanup
to be done internally. Call me old-fashioned, but I prefer numerics to spit
out whatever they get, without trying to get too smart.

Of course, in cases where one has a bulletproof argument for cutting off
spurious values, fine. But those cases are rare enough that I'd rather see the
garbage and clean it myself (or at least have that option until I'm convinced
it's really garbage, and then I can turn  the 'auto_cleanup' on).

Just a thought,

f.




More information about the Scipy-dev mailing list