[Numpy-discussion] Problem in LinearAlgebra?

[Nadav Horesh]

Many unstable problems have a stable solution if you choose the right 
algorithm. The question is if somewhere the developers decided to switch 
the equations solver, or there is a real bug. I hope that one of the 
developers will reply in this forum. I will try to look at it also, 
since it is a core component in the linear algebra package. Also if you 
suspect that your work-around is useful --- please post it here!

   Nadav

Rob Hooft wrote:
> Nadav Horesh wrote:
> 
>> I tested the problem with:
>>      1. Numeric 23.1 under python 2.3.2
>>      2. numarray 0.8 (I made a copy of the Scientific package where all
>>         calls to Numeric were replaced to numarray), under python 2.3.2
>> There results where about the same -- high coefficients for the 5th
>> order polynomials.
>> I would expect reliable fit for a high order polynomials only under very
>> special circumstances, so this is not a big surprise. 
> 
> 
> Thanks for your efforts.
> 
> The polynomial we're trying to fit here is not extremely unstable. As I 
> said, with Numeric 17.1.2 my class of problems used to be stable up to 
> at least 9th order. I really suspect a bug was introduced here which is 
> difficult to pinpoint because everybody reacts the natural way: this is 
> an intrinsicly unstable problem, so it is not unexpected. Somehow it 
> could have been better, though!
> 
> I managed to work around the problem so far by using a different solver 
> also built in to Scientific Python, so I am saved so far.
> 
> Regards,
> 
> Rob
>